PLoS Genetics

Li, Yi;Xiong, Ziyi;Zhang, Manfei;Hysi, Pirro G.;Qian, Yu;Adhikari, Kaustubh;Weng, Jun;Wu, Sijie;Du, Siyuan;Gonzalez-Jose, Rolando;Schuler-Faccini, Lavinia;Bortolini, Maria-Catira;Acuna-Alonzo, Victor;Canizales-Quinteros, Samuel;Gallo, Carla;Poletti, Giovanni;Bedoya, Gabriel;Rothhammer, Francisco;Wang, Jiucun;Tan, Jingze;Yuan, Ziyu;Jin, Li;Uitterlinden, André G.;Ghanbari, Mohsen;Ikram, M. Arfan;Nijsten, Tamar;Zhu, Xiangyu;Lei, Zhen;Jia, Peilin;Ruiz-Linares, Andres;Spector, Timothy D.;Wang, Sijia;Kayser, Manfred;Liu, Fan

doi: 10.1371/journal.pgen.1010786pmid: 37459304

Introduction Human ear morphology is a complex trait that is influenced by a multidimensional set of correlated and heritable phenotypes. Studies have estimated the heritability of ear morphology phenotypes to be between 29% and 61% [1], indicating that genetic variants play a significant role in shaping the ear’s anatomy. Understanding the genetic basis of human ear morphology has important implications for multiple scientific disciplines such as human genetics, developmental biology, medicine, evolutionary biology, anthropology, and forensics. This study aims to generate knowledge that can aid in understanding genetic disorders related to ear morphology, improve diagnosis and treatment of ear-related diseases and also help in forensics and anthropology to identify individuals based on their ear morphology. Previous genome-wide association studies (GWASs) on ear morphology have focused on a limited number of qualitative ear features and have not utilized quantitative ear traits. A GWAS in 5,062 Latin Americans identified significant associations at seven genetic loci for ear lobe size and attachment, folding of antihelix, helix rolling, ear protrusion and antitragus size, suggesting the involvement of the genes EDAR, CART1, SP5, MPRS22, LRBA, LOC153910, and LOC100287225 [1]. A multi-ethnic GWAS meta-analysis on earlobe attachment in 74,660 samples revealed a total of 49 genetic loci [2], further emphasizing the complexity of the genetic architecture of ear morphology. Due to the complex and multidimensional nature of ear morphology, focusing on a limited number of qualitative ear features, as done in previous GWASs, can result in missing important phenotypic information and genetic associations. Additionally, traditional methods such as human perception [1] and questionnaire-based [2] phenotyping are prone to bias and instability. Therefore, in this study, we developed a fully automated computer pipeline that utilizes a deep learning convolutional neural network (CNN) [3] to quantify a large number of ear phenotypes from high-resolution digital side-face photos. This approach allows for a more comprehensive and accurate assessment of ear morphology and can provide a more complete understanding of the genetic factors underlying this trait. High correlations between various ear traits may be influenced by genetic variants with multi-trait effects. However, traditional single-trait GWAS, as used in previous ear GWASs [1,2], has limitations in identifying these genetic loci. To overcome these limitations, here we use C-GWAS [4], a recently developed method for combining GWAS summary statistics of multiple potentially related traits, to integrate our GWAS results. C-GWAS has been shown to have increased statistical power compared to traditional methods such as minimal p-values of multiple single-trait GWASs (MinGWAS) and MTAG [5]. Furthermore, it has been successfully applied to the analysis of a large number of facial traits, resulting in the identification of novel genetic loci associated with facial variation that were not identified via traditional methods [4]. In this study, we used a sample of 14,921 multi-ethnic individuals from Europe (N = 4,740), Asia (N = 4,835), and Latin America (N = 5,346) to investigate the genetic factors that contribute to ear morphology. We quantified 136 quantitative ear phenotypes using deep learning analyses of digital face images, and conducted GWASs and GWAS meta-analyses. These results were then combined using a recently developed C-GWAS method [4]. We identified novel genetic loci and confirmed previously established loci that are involved in normal-range quantitative variation of ear morphology in humans. We performed functional follow-up studies in gene-edited mice to further understand the impact of these genetic loci on ear morphology. Results Performance of CNN in quantitative ear phenotyping In this study, we used a deep learning convolution neural network (CNN) approach [3] for comprehensive ear landmarking on high-resolution 2D digital side face photos. We focused on the 17 most anatomically meaningful landmarks (S1A Fig) out of the 55 that could be located by the CNN, and derived 136 pairwise inter-landmark distances as input phenotypes for subsequent genetic analyses. We found that the inter-rater correlations on the left ear were reasonably high (mean r = 0.47, sd = 0.24, S2A Fig) and the correlations between manual- and CNN-derived phenotypes on the same ear were reasonably high too (mean r = 0.34, sd = 0.18, S2B Fig). Utilizing the symmetric nature of the human face, it was obvious that the left and right ear correlations for CNN-derived phenotypes (r = 0.52, sd = 0.10, S2D Fig) were significantly higher, and with a smaller standard deviation, than those from the same human rater (r = 0.44, sd = 0.20, S2C Fig). This demonstrated that the deep learning approach on ear phenotyping that we used in this study on the full set of facial images had at least an equal performance in the accuracy of ear landmarking compared to human perception, while also being more efficient. Sample characteristics and phenotype results This study included 14,921 individuals from five different population cohorts from three continents: Europe, Asia, and Latin America (S1B Fig and S1 Table). These cohorts were: two European cohorts (the Rotterdam Study with 3,675 participants from the Netherlands and the TwinsUK study with 1,065 participants from the UK), two East Asian cohorts (the Taizhou Longitudinal Study, TZL, with 2,348 Han Chinese participants from China and the National Survey of Physical Traits study, NSPT, with 2,487 Han Chinese participants from China), and one cohort of mixed ancestry (the CANDELA study with 5,346 Latin Americans, who have an estimated genomic admixture of 48% European, 46% Native American, and 6% African). It is worth noting that, while CANDELA and TZL have been used in previous GWAS studies on qualitative ear features [1,2], quantitative ear phenotypes have not been studied in these cohorts before. The majority (82–100%) of the 136 ear phenotypes studied were normally distributed, as determined by the Kolmogorov-Smirnov normality test (P > 0.05 after multiple testing correction, as shown in S2 and S3 Tables). The lower proportion of normally distributed ear phenotypes in the Rotterdam Study (RS) can be attributed to a higher proportion of older individuals (with an average age more than 10 years older than in the other cohorts, as seen in S1 Table). The effects of sex (min P = 3.5e-145) and age (min P = 5.9e-99) on ear phenotypes were investigated in detail using individual-level data from RS. These factors together explained up to 16.9% of the variance for certain ear phenotypes (e.g. phenotype L7-L14) (S4 Table). Women had longer distances involving otobasion inferius (L7) compared to men, while men had longer distances involving otobasion superius (L1) compared to women (S3A Fig). Aging was associated with a significant increase in ear length and thinness (S3B Fig). Unsupervised hierarchical clustering of the 136 ear phenotypes resulted in two main clusters, mainly representing vertical and horizontal ear features (S4A Fig). The genetic correlation estimates from genome-wide SNPs were high and similar to phenotypic correlations (S4B Fig), indicating shared genetic contributors across multiple ear phenotypes. The twin-based heritability of the 136 ear phenotypes was estimated in TwinsUK to have a mean of 0.52 (sd = 0.03, 0.49–0.56, S5A Fig). These estimates were generally higher than those obtained from genome-wide SNPs in the same cohort (mean = 0.38, sd = 0.09, 0.20–0.60, S5B Fig). The genetic heritability estimates were largely consistent with those for qualitative ear features reported previously in the CANDELA study [1]. GWASs, Meta-analysis, and C-GWAS Single-trait GWASs of 136 ear phenotypes were conducted in each of the five cohorts separately (as shown in S1B Fig) and the results were then meta-analyzed, resulting in 136 sets of single-trait summary statistics. The inflation factors in these single-trait GWASs were all within an acceptable range (average λ = 1.04, with 1.02 < λ < 1.06). The minimal p-values of the 136 single-trait GWASs were adjusted empirically using simulations as implemented in C-GWAS, so that the adjusted minimal p-values follow a uniform distribution under the null. These adjusted minimal p-values are abbreviated as "MinGWAS" below. Similarly, the C-GWAS results were also adjusted using the embedded simulations, so that the adjusted p-values also closely follow a uniform distribution under the null (as shown in Fig 1A–1C). This means that MinGWAS and C-GWAS results can be directly compared with each other as well as with those from any standard single-trait GWAS. Therefore, the traditional genome-wide significance threshold (P ≤ 5e-8) can be considered as the study-wide significance threshold in this study. Additional information on the C-GWAS method can be found in other publications [4]. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 1. Outcomes from GWAS meta-analysis (MinGWAS) and combined GWAS (C-GWAS) for 136 quantitative ear morphology traits in 5 multi-ethnic cohorts (N = 14,921). A simulation analysis derived the null distributions of the crude p-values (gray). The p-values corrected using the getCoef function (orange, details of function in Materials and methods) are compared with the uniform distribution (black) for both the C-GWAS results (A) and the MinGWAS result (B). After multiple testing correction, the C-GWAS results (upper part of D) and the minimal p-values from 136 single ear trait GWASs (lower part of D) are plotted using a combined Manhattan plot (D) and a Q-Q plot (C). In the Manhattan plot (D), the study-wide significance threshold (P = 5e-8, all p-values of MinGWAS and C-GWAS have been corrected) is indicated using dashed lines. A total of 16 study-wide significant loci were identified, among which 9 were significant in both C-GWAS and MinGWAS (orange dots and circles with cross), 1 solely was significant in MinGWAS (purple circle with cross), and the remaining 6 were significant only in C-GWAS (red dots and circles with cross). (E) The closest genes to regional lead SNPs are listed (E) (orange for novel). (F) Venn diagram of 15 candidate genes in the 15 loci highlighted by C-GWAS, 10 candidate genes in the 10 loci highlighted by MinGWAS, 50 candidate genes in 50 loci from previous GWASs of categorical ear features, and 243 candidate genes in 243 loci from previous GWASs of facial shape variation. (G) A list of 59 previously established ear-associated SNPs was looked up in the C-GWAS and MinGWAS results. In RS, the ear variance explained by 15 lead SNPs from loci identified by C-GWAS (H) is compared with that explained by 10 lead SNPs from loci identified by MinGWAS (I). https://doi.org/10.1371/journal.pgen.1010786.g001 MinGWAS and C-GWAS together identified a total of 1,205 ear-associated SNPs at 16 distinct genetic loci that exceeded the study-wide significance (Fig 1D–1E and Table 1). Eight of these loci had not been previously associated with ear features in previous GWASs and were identified for the first time in this study (Fig 1F); four of them were significant in both MinGWAS and C-GWAS, and four were significant solely in C-GWAS (Table 1 and Fig 1F). Notably, all eight novel loci had highly consistent allele effects across all five cohorts, despite their different continental ancestries (S6A–S6H Fig). The most significant novel finding was rs7812632 at 8q24.13 (PC-GWAS = 1.2e-19), where the G-allele mainly led to an increased vertical length of the ear. This SNP was top-associated with L6-L15 (PMinGWAS = 4.7e-10) and showed nominally significant association with L6-L15 in all cohorts, with similar effect sizes (except for TwinsUK due to QC failure). This SNP is located approximately 200kbp upstream of the HAS2 gene, which encodes hyaluronan synthase 2, a protein that plays an important role in embryonic development of branchial arches [6] and cranial neural crest cells (CNCCs) [7]. Variants in or near HAS2 have been previously associated with face morphology [8], body height [9], and male pattern baldness [10]. Additionally, this locus is the only one among all the significant ones we found that shows borderline genome-wide significant association with ear landmark L3 (PL3-L15 = 2.7e-7, S6D Fig), which anatomically approximates Darwin’s Tubercle (OMIM:124300, S7 Fig) [11]. We also tested the association between rs7812632 and Darwin’s Tubercle in CANDELA, as it is the only cohort studied in which this phenotype has been manually obtained. This SNP was nominally associated with Darwin’s Tubercle (P = 0.03) and the A-allele was associated with an increased prevalence. Download: PPT PowerPoint slide PNG larger image TIFF original image Table 1. SNPs associated with quantitative ear morphology traits identified by MinGWAS and C-GWAS. https://doi.org/10.1371/journal.pgen.1010786.t001 The second most significant novel finding was rs57788627 at 4q28.1 (PC-GWAS = 2.5e-15), where the G-allele mainly led to an increased ear width. This SNP was top-associated with L15-L16 (PMinGWAS = 2.3e-9) and showed nominally significant association with L15-L16 in almost all cohorts, with similar effect sizes (4/5, except NSPT). This locus is remarkable as it represents a gene desert and the nearest gene, INTU, is 408kb away from the region’s lead SNP (S6B Fig). A recent face GWAS by Xiong et al. found an association between INTU and human facial variation and also confirmed the enhancer activity of this region through luciferase assay experiments [8]. Among all genes in this region, INTU showed the most preferential expression in CNCCs. Additionally, INTU is involved in both ciliogenesis and convergent extension and plays a role in embryonic development [12]. Recent studies have suggested an important role of Intu in mouse skeletal development [13]. Furthermore, exome sequencing of orofaciodigital syndrome patients with facial, oral, and ear abnormalities revealed disease-associated mutations in INTU [14]. We attained mouse mutants using one-step CRISPR/cas9 mediated gene editing experiments to assess the functional significance of Intu gene expression during ear mouse development (for details, see below). Among the eight loci that were already reported in the two previous single-trait GWASs of qualitative ear traits, five were significant in our quantitative study in both MinGWAS and C-GWAS, specifically 1p12 TBX15, 2q12.3 EDAR, 2q31.1 SP5, 3q23 MRPS22, and 6q24.2 HIVEP2. One locus, 6q21 PRDM1/ATG5, reached study-wide significance only in MinGWAS, and two, 4q31.3 LRBA and 7q21.3 DLX6, reached study-wide significance only in C-GWAS. The allele effects of all these eight loci were in the same direction in all five cohorts (S6L–S6S Fig). Notably, all loci except 6q21 showed orders of magnitude higher significance in C-GWAS than in MinGWAS (Fig 1B), with rs17034666 at 2q12.3 EDAR being the most extreme example (from PMinGWAS = 1.09e-11 to PC-GWAS = 1.38e-24). The East Asian-specific missense variant of EDAR (EDARV370A, rs3827760) had been previously associated with a wide range of endoderm-derived phenotypes, such as chin protrusion [15], hair shape [16], hair thickness [17], sweat glands [18], size of feminine breasts [18], and shovel incisors characteristic [19,20]. However, this SNP was removed from our C-GWAS due to its very low MAF in European samples. The boost of significance at 2q12.3 EDAR is an example of the increased power of C-GWAS in detecting multi-trait effects. A multivariable fitting analysis in the RS cohort based on individual-level data showed that the lead SNPs from C-GWAS-significant loci explained on average 1.5 times and up to 3.1 times more phenotypic ear variance than the SNPs from MinGWAS-significant loci did (Fig 1H–1I). Integration with previous literature knowledge The two previous single-trait GWASs on qualitative ear morphology [1,2] identified 58 autosomal SNPs at 50 loci mainly associated with earlobe features. We looked up these 58 SNPs in three of our cohorts (RS, TwinsUK, and NSPT) not considering CANDELA and TZL because these two cohorts were used in the previous GWASs. For 49 SNPs (nine SNPs were non-polymorphic in at least one dataset and thus got excluded from this analysis), both C-GWAS (P = 1.1e-8) and MinGWAS (P = 2e-5) p-values highly significantly deviated from the null, while C-GWAS p-values obviously deviated further from the null than MinGWAS p-values did (Fig 1G). Under the nominal significance level, C-GWAS re-identified a larger number of the previously established loci than MinGWAS did (16 vs. 13). In addition, all our nominally significant associations involving earlobe landmarks (S5 Table) were consistent with the findings from previous studies of earlobe features, i.e., they were previously associated with earlobe features and the most significant ear phenotype in our study involved earlobe landmarks too. These results further confirm the increased statistical power of C-GWAS compared to the traditionally used MinGWAS. Both the ear and face belong to craniofacial phenotypes for which shared genetic effects maybe expected. Therefore, we examined the 238 distinct genetic loci previously associated with human facial shape variation [8,15,21–30] in our C-GWAS results for quantitative ear morphology. Two of these face loci, which harbor TBX15 and INTU, respectively, (S6 Table) showed study-wide significant association in our ear C-GWAS. On the other hand, we saw that a substantial proportion of the ear-associated loci identified in the present ear study (6/16) showed genome-wide significant association with facial shape variation in previous face GWASs (Fig 1F). The finding of a larger proportion of ear-associated loci implicated in facial shape than face-associated loci implicated in ear morphology is consistent with the longer span of early development of the face (4th to 9th gestation week) compared to the outer ear (6th and 7th gestation week) [31,32]. Furthermore, more than half (9/16) of our C-GWAS-identified ear loci (1p12 TBX15, 2q12.3 EDAR, 2q31.1 SP5/MYO3B, 3q23 MRPS22, 6q21 ATG5, 8q24.13 HAS2, 9q33.1 ASTN2, 10q22.2 C10orf11, 20q11.22 UQCC1), which were particularly associated with earlobe phenotypes in our study, were previously reported to have association with male pattern baldness in the GWAS catalog [10], and two of our ear loci, including 2q12.3 EDAR and 6q21 ATG5, were previously associated with mono eyebrow [21] (S7 Table). Male pattern baldness and mono eyebrow also reflect surface ectoderm-derived phenotypes. These results suggest that surface ectoderm-derived phenotypes such as facial shape, ear morphology, male pattern baldness, and mono eyebrow share genetic factors, which may be expected, but has not been empirically demonstrated before. These findings provide further evidence of the genetic overlap between different craniofacial phenotypes and highlights the importance of studying multiple traits together to gain a more comprehensive understanding of the genetic basis of craniofacial variation. Functional annotations Several lines of evidence support the functional implications of the 16 discovered ear-associated loci, including 8 novel loci. A gene ontology enrichment analysis highlighted 6 ear development-related biological process terms (FDR < 0.01, Tables 1 and S8) including ear development, inner ear development, development and morphogenesis of the skeletal system, and embryonic organ development and morphogenesis. The majority of the lead SNPs (10/16) showed evidence of regulatory activity in the 3DSNP database [33] (S9 Table). Four out of the 16 loci showed positive enhancer activity in transgenic mice supported by the spatial pattern of expression located in ears/branchial arch/craniofacial [34]. The majority (14/16) of nearby genes or 3D interaction genes has been associated with abnormal ear/craniofacial phenotypes in various databases [35]. Most (15/16) of nearby genes or 3D interaction genes were expressed in branchial arch and embryo ectoderm in mice [36,37]. These lines of evidence strengthen the reliability of our novel loci and support that ear and cranial morphogenesis share a substantial proportion of genetic factors. These findings provide further support for the functional relevance of the identified loci and highlight the importance of studying multiple traits together to gain a more comprehensive understanding of the genetic basis of craniofacial development. Embryonic cranial neural crest cells (CNCCs) are temporary, migratory stem cells that play a crucial role in the formation of the ear during embryonic development. We compared the expression of 16 genes located near the 16 ear-associated SNPs with a random set of 16 genes selected near randomly chosen and frequency-matched 16 SNPs in CNCCs and 49 other cell types for 10000 replicates. Compared with the randomly selected gene sets, this set of 16 targeted genes near the ear-associated SNPs had significantly higher expression in 23 types of cells (P < 0.05, Fig 2A), including those involved in the formation of constituent of ear, such as CNCCs, articular chondrocyte, osteoblast, skin fibroblast, and adipose mesenchymal stem cell. In addition, we performed a heritability enrichment analysis using the S-LDSC method [38] with respect to active regulatory regions based on our C-GWAS results of Europeans (RS and TwinsUK). This analysis showed that the S-LDSC coefficient Z-scores in mesenchymal stem cells, osteoblast primary cells, adult dermal fibroblast primary cells, adipose derived mesenchymal stem cell cultured cells, mesenchymal stem cell derived chondrocyte cultured cells, all ranked within the top 5% in a total of 102 types of cells and CNCCs slightly behind them (ranked 9th) (Fig 2C). These cell types from the heritability enrichment analysis are consistent with those where the 16 genes showed preferential expression. These cell types play important roles in ear morphogenesis, supporting the reliability of our findings. In addition, 14 out of these 16 genes showed preferential expression in CNCCs compared to 49 other cell types (Fig 2B). The genes near the eight newly discovered loci showed preferential expression in CNCC, highly consistent with the preferential expression pattern of the genes near the 8 previously established ear-associated loci. Interestingly, several of these genes (EDAR, INTU, and HAS2) have previously been linked to facial morphology [8], supporting the idea that genetic effects shaping the ear and face originate during early embryogenesis. These findings provide a priority list for future in-vivo studies on genes involved in ear variation and embryo surface ectoderm-derived phenotypes. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 2. Differential expressions of 16 genes near 16 ear-associated loci in 50 cell types. (A) Boxplots of normalized RNA-seq VST values for the 16 genes (in orange, details of genes shown in lower B plot) and 16 median gene expression value as control (randomly matched 1e-4 times using SNPsnap, in blue). Expression differences between the control genes and the 16 genes were tested in each cell type using the unpaired Wilcoxon rank-sum test. The expression of the 16 genes in CNCCs was also iteratively compared with that in all other cell types using paired Wilcoxon rank-sum test. Statistical significance was indicated: *P < 0.05. (B) Normalized RNA-seq VST values in CNCCs were compared with those in other 49 types of cells (purple), in 20 tissue cell types (blue), in 20 primary cell types (green), and in nine embryonic stem cell types (orange), using one-sample Student’s t-test. Dotted line represents Bonferroni corrected significant threshold (P < 2.27e-3). Significant gene labels are depicted in color, non-significant gene labels in black. (C) Partitioned heritability enrichments based on cell-type-specific regulatory annotations (More details see in Methods). Heritability enrichment Z-scores, as estimated by stratified linkage disequilibrium score regression (S-LDSC) of the C-GWAS summary data for GWASs of 136 ear traits. Trait abbreviations as in S10 Table. https://doi.org/10.1371/journal.pgen.1010786.g002 Ear effects in Intu and Tbx15 mutant mice Here, we examined adult ear morphology in Intu and Tbx15 mouse mutants using one-step CRISPR/cas9 mediated gene editing experiments to assess the functional significance of Intu and Tbx15 expression during ear development. Our breeding experiments generated 18 F2 9-weeks sexually mature Intu mice i.e., 10 heterozygous Intu+/-, 8 wild-type WT (Intu+/+), while homozygous loss-of-function of Intu was lethal. Quantitative assessment of mouse ear shape (assessed by PC analysis of 21 three-dimensional landmark coordinates) revealed significant differences (linear regression PPC3 = 1.5e-3) (Fig 3C) between the heterozygous Intu littermates and WT mice. The Intu genotype significantly associated with D6_14 (P = 2.1e-3, Beta = 1.78) and D3_6 (P = 3.3e-3, Beta = 1.74) (Fig 3E and S11 Table) after FDR correction. Notably, in humans, the top INTU variant rs57788627 was nominally significantly associated with ear phenotypes L3-L6 and L3-L7, which respectively corresponds to D6_14 and D3_6 in mice (P = 1.9e-5, Beta = 0.07; P = 8.7e-5, Beta = 0.06) (Fig 3E and S11 Table). Overall, the ears of the heterozygous Intu mutant mice were shorter than those of the WT mice consistently show in the result of PC3 and distance phenotypes (Fig 3D–3F). Furthermore, the heterozygous Intu mutant mice showed a significant trend of reduction in body length and fore and hind limb length (S8 Fig), which was consistent with previous findings [13]. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 3. In-vivo mouse models of INTU deficiency. Intu mutant mice (Intu-+/-, n = 10, 9 weeks) vs. C57BL/6 wild type control mice (WT+/+, n = 8, 9 weeks) were compared for ear and body morphological differences (for details of Tbx15 mice see S9 Fig). (A) Regional association plot for the associated locus nearby INTU, and the other two SNPs at this locus, which have been recently reported the association with face morphology, were also marked. (B) The schematic diagram of the one-step CRISPR/Cas9 technology used in Intu knockout mice (see S10 Fig for details). (C) Example of left profile craniofacial photo of Intu+/- mutant mice with removed hair (left); and 21 craniofacial landmarks pattern in 3D mice image (right). (D) The principal component analysis for 21 landmarks of Intu+/- mutant mice and WT+/+ mice. The left layer shown the detailed contributed proportion of 21 landmarks to the first 5 principal components. The upper left shown the cumulative variance is explained by the first 10 PCs. The significant association between the genotype and PC3 shown in upper right (* P < 0.05, ** P < 0.01, *** P < 0.001). The bottom shown the maximum PC3-, minimum PC3- and mean ear shapes (the labels on the mouse ear were consistent with the FigC). (E) The pattern of genetic association in humans (left) and in mice (right). (F) Effect of Intu knock-out on ear phenotypes in mice (blue for effect of heterozygote mutant and red for wild type). The labels on three ears were consistent with Fig S1 (more details see in Methods). https://doi.org/10.1371/journal.pgen.1010786.g003 Our breeding experiment generated 37 F2 9-weeks adult Tbx15 mice i.e., 9 homozygous Tbx15-/-, 18 heterozygous Tbx15+/-, 10 wild-type WT (Tbx15+/+). Previous gene editing experiments have already shown that Tbx15-/- mutant mice demonstrate facial variation, loss of weight, shorter limbs and a distinct “droopy ear” feature [39]. In our study, besides the facial differences reported previously, ear differences were obvious. In addition to the distinct “droopy ear” feature, the Tbx15 genotype was significantly associated with the ear landmarks PC1 (P = 3.2e-3) and PC4 (P = 1.9e-3) (S9C Fig). Overall, the ears of the Tbx15 mutant mice were obviously longer and wider than those of the WT mice (S9E and S9F Fig). Performance of CNN in quantitative ear phenotyping In this study, we used a deep learning convolution neural network (CNN) approach [3] for comprehensive ear landmarking on high-resolution 2D digital side face photos. We focused on the 17 most anatomically meaningful landmarks (S1A Fig) out of the 55 that could be located by the CNN, and derived 136 pairwise inter-landmark distances as input phenotypes for subsequent genetic analyses. We found that the inter-rater correlations on the left ear were reasonably high (mean r = 0.47, sd = 0.24, S2A Fig) and the correlations between manual- and CNN-derived phenotypes on the same ear were reasonably high too (mean r = 0.34, sd = 0.18, S2B Fig). Utilizing the symmetric nature of the human face, it was obvious that the left and right ear correlations for CNN-derived phenotypes (r = 0.52, sd = 0.10, S2D Fig) were significantly higher, and with a smaller standard deviation, than those from the same human rater (r = 0.44, sd = 0.20, S2C Fig). This demonstrated that the deep learning approach on ear phenotyping that we used in this study on the full set of facial images had at least an equal performance in the accuracy of ear landmarking compared to human perception, while also being more efficient. Sample characteristics and phenotype results This study included 14,921 individuals from five different population cohorts from three continents: Europe, Asia, and Latin America (S1B Fig and S1 Table). These cohorts were: two European cohorts (the Rotterdam Study with 3,675 participants from the Netherlands and the TwinsUK study with 1,065 participants from the UK), two East Asian cohorts (the Taizhou Longitudinal Study, TZL, with 2,348 Han Chinese participants from China and the National Survey of Physical Traits study, NSPT, with 2,487 Han Chinese participants from China), and one cohort of mixed ancestry (the CANDELA study with 5,346 Latin Americans, who have an estimated genomic admixture of 48% European, 46% Native American, and 6% African). It is worth noting that, while CANDELA and TZL have been used in previous GWAS studies on qualitative ear features [1,2], quantitative ear phenotypes have not been studied in these cohorts before. The majority (82–100%) of the 136 ear phenotypes studied were normally distributed, as determined by the Kolmogorov-Smirnov normality test (P > 0.05 after multiple testing correction, as shown in S2 and S3 Tables). The lower proportion of normally distributed ear phenotypes in the Rotterdam Study (RS) can be attributed to a higher proportion of older individuals (with an average age more than 10 years older than in the other cohorts, as seen in S1 Table). The effects of sex (min P = 3.5e-145) and age (min P = 5.9e-99) on ear phenotypes were investigated in detail using individual-level data from RS. These factors together explained up to 16.9% of the variance for certain ear phenotypes (e.g. phenotype L7-L14) (S4 Table). Women had longer distances involving otobasion inferius (L7) compared to men, while men had longer distances involving otobasion superius (L1) compared to women (S3A Fig). Aging was associated with a significant increase in ear length and thinness (S3B Fig). Unsupervised hierarchical clustering of the 136 ear phenotypes resulted in two main clusters, mainly representing vertical and horizontal ear features (S4A Fig). The genetic correlation estimates from genome-wide SNPs were high and similar to phenotypic correlations (S4B Fig), indicating shared genetic contributors across multiple ear phenotypes. The twin-based heritability of the 136 ear phenotypes was estimated in TwinsUK to have a mean of 0.52 (sd = 0.03, 0.49–0.56, S5A Fig). These estimates were generally higher than those obtained from genome-wide SNPs in the same cohort (mean = 0.38, sd = 0.09, 0.20–0.60, S5B Fig). The genetic heritability estimates were largely consistent with those for qualitative ear features reported previously in the CANDELA study [1]. GWASs, Meta-analysis, and C-GWAS Single-trait GWASs of 136 ear phenotypes were conducted in each of the five cohorts separately (as shown in S1B Fig) and the results were then meta-analyzed, resulting in 136 sets of single-trait summary statistics. The inflation factors in these single-trait GWASs were all within an acceptable range (average λ = 1.04, with 1.02 < λ < 1.06). The minimal p-values of the 136 single-trait GWASs were adjusted empirically using simulations as implemented in C-GWAS, so that the adjusted minimal p-values follow a uniform distribution under the null. These adjusted minimal p-values are abbreviated as "MinGWAS" below. Similarly, the C-GWAS results were also adjusted using the embedded simulations, so that the adjusted p-values also closely follow a uniform distribution under the null (as shown in Fig 1A–1C). This means that MinGWAS and C-GWAS results can be directly compared with each other as well as with those from any standard single-trait GWAS. Therefore, the traditional genome-wide significance threshold (P ≤ 5e-8) can be considered as the study-wide significance threshold in this study. Additional information on the C-GWAS method can be found in other publications [4]. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 1. Outcomes from GWAS meta-analysis (MinGWAS) and combined GWAS (C-GWAS) for 136 quantitative ear morphology traits in 5 multi-ethnic cohorts (N = 14,921). A simulation analysis derived the null distributions of the crude p-values (gray). The p-values corrected using the getCoef function (orange, details of function in Materials and methods) are compared with the uniform distribution (black) for both the C-GWAS results (A) and the MinGWAS result (B). After multiple testing correction, the C-GWAS results (upper part of D) and the minimal p-values from 136 single ear trait GWASs (lower part of D) are plotted using a combined Manhattan plot (D) and a Q-Q plot (C). In the Manhattan plot (D), the study-wide significance threshold (P = 5e-8, all p-values of MinGWAS and C-GWAS have been corrected) is indicated using dashed lines. A total of 16 study-wide significant loci were identified, among which 9 were significant in both C-GWAS and MinGWAS (orange dots and circles with cross), 1 solely was significant in MinGWAS (purple circle with cross), and the remaining 6 were significant only in C-GWAS (red dots and circles with cross). (E) The closest genes to regional lead SNPs are listed (E) (orange for novel). (F) Venn diagram of 15 candidate genes in the 15 loci highlighted by C-GWAS, 10 candidate genes in the 10 loci highlighted by MinGWAS, 50 candidate genes in 50 loci from previous GWASs of categorical ear features, and 243 candidate genes in 243 loci from previous GWASs of facial shape variation. (G) A list of 59 previously established ear-associated SNPs was looked up in the C-GWAS and MinGWAS results. In RS, the ear variance explained by 15 lead SNPs from loci identified by C-GWAS (H) is compared with that explained by 10 lead SNPs from loci identified by MinGWAS (I). https://doi.org/10.1371/journal.pgen.1010786.g001 MinGWAS and C-GWAS together identified a total of 1,205 ear-associated SNPs at 16 distinct genetic loci that exceeded the study-wide significance (Fig 1D–1E and Table 1). Eight of these loci had not been previously associated with ear features in previous GWASs and were identified for the first time in this study (Fig 1F); four of them were significant in both MinGWAS and C-GWAS, and four were significant solely in C-GWAS (Table 1 and Fig 1F). Notably, all eight novel loci had highly consistent allele effects across all five cohorts, despite their different continental ancestries (S6A–S6H Fig). The most significant novel finding was rs7812632 at 8q24.13 (PC-GWAS = 1.2e-19), where the G-allele mainly led to an increased vertical length of the ear. This SNP was top-associated with L6-L15 (PMinGWAS = 4.7e-10) and showed nominally significant association with L6-L15 in all cohorts, with similar effect sizes (except for TwinsUK due to QC failure). This SNP is located approximately 200kbp upstream of the HAS2 gene, which encodes hyaluronan synthase 2, a protein that plays an important role in embryonic development of branchial arches [6] and cranial neural crest cells (CNCCs) [7]. Variants in or near HAS2 have been previously associated with face morphology [8], body height [9], and male pattern baldness [10]. Additionally, this locus is the only one among all the significant ones we found that shows borderline genome-wide significant association with ear landmark L3 (PL3-L15 = 2.7e-7, S6D Fig), which anatomically approximates Darwin’s Tubercle (OMIM:124300, S7 Fig) [11]. We also tested the association between rs7812632 and Darwin’s Tubercle in CANDELA, as it is the only cohort studied in which this phenotype has been manually obtained. This SNP was nominally associated with Darwin’s Tubercle (P = 0.03) and the A-allele was associated with an increased prevalence. Download: PPT PowerPoint slide PNG larger image TIFF original image Table 1. SNPs associated with quantitative ear morphology traits identified by MinGWAS and C-GWAS. https://doi.org/10.1371/journal.pgen.1010786.t001 The second most significant novel finding was rs57788627 at 4q28.1 (PC-GWAS = 2.5e-15), where the G-allele mainly led to an increased ear width. This SNP was top-associated with L15-L16 (PMinGWAS = 2.3e-9) and showed nominally significant association with L15-L16 in almost all cohorts, with similar effect sizes (4/5, except NSPT). This locus is remarkable as it represents a gene desert and the nearest gene, INTU, is 408kb away from the region’s lead SNP (S6B Fig). A recent face GWAS by Xiong et al. found an association between INTU and human facial variation and also confirmed the enhancer activity of this region through luciferase assay experiments [8]. Among all genes in this region, INTU showed the most preferential expression in CNCCs. Additionally, INTU is involved in both ciliogenesis and convergent extension and plays a role in embryonic development [12]. Recent studies have suggested an important role of Intu in mouse skeletal development [13]. Furthermore, exome sequencing of orofaciodigital syndrome patients with facial, oral, and ear abnormalities revealed disease-associated mutations in INTU [14]. We attained mouse mutants using one-step CRISPR/cas9 mediated gene editing experiments to assess the functional significance of Intu gene expression during ear mouse development (for details, see below). Among the eight loci that were already reported in the two previous single-trait GWASs of qualitative ear traits, five were significant in our quantitative study in both MinGWAS and C-GWAS, specifically 1p12 TBX15, 2q12.3 EDAR, 2q31.1 SP5, 3q23 MRPS22, and 6q24.2 HIVEP2. One locus, 6q21 PRDM1/ATG5, reached study-wide significance only in MinGWAS, and two, 4q31.3 LRBA and 7q21.3 DLX6, reached study-wide significance only in C-GWAS. The allele effects of all these eight loci were in the same direction in all five cohorts (S6L–S6S Fig). Notably, all loci except 6q21 showed orders of magnitude higher significance in C-GWAS than in MinGWAS (Fig 1B), with rs17034666 at 2q12.3 EDAR being the most extreme example (from PMinGWAS = 1.09e-11 to PC-GWAS = 1.38e-24). The East Asian-specific missense variant of EDAR (EDARV370A, rs3827760) had been previously associated with a wide range of endoderm-derived phenotypes, such as chin protrusion [15], hair shape [16], hair thickness [17], sweat glands [18], size of feminine breasts [18], and shovel incisors characteristic [19,20]. However, this SNP was removed from our C-GWAS due to its very low MAF in European samples. The boost of significance at 2q12.3 EDAR is an example of the increased power of C-GWAS in detecting multi-trait effects. A multivariable fitting analysis in the RS cohort based on individual-level data showed that the lead SNPs from C-GWAS-significant loci explained on average 1.5 times and up to 3.1 times more phenotypic ear variance than the SNPs from MinGWAS-significant loci did (Fig 1H–1I). Integration with previous literature knowledge The two previous single-trait GWASs on qualitative ear morphology [1,2] identified 58 autosomal SNPs at 50 loci mainly associated with earlobe features. We looked up these 58 SNPs in three of our cohorts (RS, TwinsUK, and NSPT) not considering CANDELA and TZL because these two cohorts were used in the previous GWASs. For 49 SNPs (nine SNPs were non-polymorphic in at least one dataset and thus got excluded from this analysis), both C-GWAS (P = 1.1e-8) and MinGWAS (P = 2e-5) p-values highly significantly deviated from the null, while C-GWAS p-values obviously deviated further from the null than MinGWAS p-values did (Fig 1G). Under the nominal significance level, C-GWAS re-identified a larger number of the previously established loci than MinGWAS did (16 vs. 13). In addition, all our nominally significant associations involving earlobe landmarks (S5 Table) were consistent with the findings from previous studies of earlobe features, i.e., they were previously associated with earlobe features and the most significant ear phenotype in our study involved earlobe landmarks too. These results further confirm the increased statistical power of C-GWAS compared to the traditionally used MinGWAS. Both the ear and face belong to craniofacial phenotypes for which shared genetic effects maybe expected. Therefore, we examined the 238 distinct genetic loci previously associated with human facial shape variation [8,15,21–30] in our C-GWAS results for quantitative ear morphology. Two of these face loci, which harbor TBX15 and INTU, respectively, (S6 Table) showed study-wide significant association in our ear C-GWAS. On the other hand, we saw that a substantial proportion of the ear-associated loci identified in the present ear study (6/16) showed genome-wide significant association with facial shape variation in previous face GWASs (Fig 1F). The finding of a larger proportion of ear-associated loci implicated in facial shape than face-associated loci implicated in ear morphology is consistent with the longer span of early development of the face (4th to 9th gestation week) compared to the outer ear (6th and 7th gestation week) [31,32]. Furthermore, more than half (9/16) of our C-GWAS-identified ear loci (1p12 TBX15, 2q12.3 EDAR, 2q31.1 SP5/MYO3B, 3q23 MRPS22, 6q21 ATG5, 8q24.13 HAS2, 9q33.1 ASTN2, 10q22.2 C10orf11, 20q11.22 UQCC1), which were particularly associated with earlobe phenotypes in our study, were previously reported to have association with male pattern baldness in the GWAS catalog [10], and two of our ear loci, including 2q12.3 EDAR and 6q21 ATG5, were previously associated with mono eyebrow [21] (S7 Table). Male pattern baldness and mono eyebrow also reflect surface ectoderm-derived phenotypes. These results suggest that surface ectoderm-derived phenotypes such as facial shape, ear morphology, male pattern baldness, and mono eyebrow share genetic factors, which may be expected, but has not been empirically demonstrated before. These findings provide further evidence of the genetic overlap between different craniofacial phenotypes and highlights the importance of studying multiple traits together to gain a more comprehensive understanding of the genetic basis of craniofacial variation. Functional annotations Several lines of evidence support the functional implications of the 16 discovered ear-associated loci, including 8 novel loci. A gene ontology enrichment analysis highlighted 6 ear development-related biological process terms (FDR < 0.01, Tables 1 and S8) including ear development, inner ear development, development and morphogenesis of the skeletal system, and embryonic organ development and morphogenesis. The majority of the lead SNPs (10/16) showed evidence of regulatory activity in the 3DSNP database [33] (S9 Table). Four out of the 16 loci showed positive enhancer activity in transgenic mice supported by the spatial pattern of expression located in ears/branchial arch/craniofacial [34]. The majority (14/16) of nearby genes or 3D interaction genes has been associated with abnormal ear/craniofacial phenotypes in various databases [35]. Most (15/16) of nearby genes or 3D interaction genes were expressed in branchial arch and embryo ectoderm in mice [36,37]. These lines of evidence strengthen the reliability of our novel loci and support that ear and cranial morphogenesis share a substantial proportion of genetic factors. These findings provide further support for the functional relevance of the identified loci and highlight the importance of studying multiple traits together to gain a more comprehensive understanding of the genetic basis of craniofacial development. Embryonic cranial neural crest cells (CNCCs) are temporary, migratory stem cells that play a crucial role in the formation of the ear during embryonic development. We compared the expression of 16 genes located near the 16 ear-associated SNPs with a random set of 16 genes selected near randomly chosen and frequency-matched 16 SNPs in CNCCs and 49 other cell types for 10000 replicates. Compared with the randomly selected gene sets, this set of 16 targeted genes near the ear-associated SNPs had significantly higher expression in 23 types of cells (P < 0.05, Fig 2A), including those involved in the formation of constituent of ear, such as CNCCs, articular chondrocyte, osteoblast, skin fibroblast, and adipose mesenchymal stem cell. In addition, we performed a heritability enrichment analysis using the S-LDSC method [38] with respect to active regulatory regions based on our C-GWAS results of Europeans (RS and TwinsUK). This analysis showed that the S-LDSC coefficient Z-scores in mesenchymal stem cells, osteoblast primary cells, adult dermal fibroblast primary cells, adipose derived mesenchymal stem cell cultured cells, mesenchymal stem cell derived chondrocyte cultured cells, all ranked within the top 5% in a total of 102 types of cells and CNCCs slightly behind them (ranked 9th) (Fig 2C). These cell types from the heritability enrichment analysis are consistent with those where the 16 genes showed preferential expression. These cell types play important roles in ear morphogenesis, supporting the reliability of our findings. In addition, 14 out of these 16 genes showed preferential expression in CNCCs compared to 49 other cell types (Fig 2B). The genes near the eight newly discovered loci showed preferential expression in CNCC, highly consistent with the preferential expression pattern of the genes near the 8 previously established ear-associated loci. Interestingly, several of these genes (EDAR, INTU, and HAS2) have previously been linked to facial morphology [8], supporting the idea that genetic effects shaping the ear and face originate during early embryogenesis. These findings provide a priority list for future in-vivo studies on genes involved in ear variation and embryo surface ectoderm-derived phenotypes. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 2. Differential expressions of 16 genes near 16 ear-associated loci in 50 cell types. (A) Boxplots of normalized RNA-seq VST values for the 16 genes (in orange, details of genes shown in lower B plot) and 16 median gene expression value as control (randomly matched 1e-4 times using SNPsnap, in blue). Expression differences between the control genes and the 16 genes were tested in each cell type using the unpaired Wilcoxon rank-sum test. The expression of the 16 genes in CNCCs was also iteratively compared with that in all other cell types using paired Wilcoxon rank-sum test. Statistical significance was indicated: *P < 0.05. (B) Normalized RNA-seq VST values in CNCCs were compared with those in other 49 types of cells (purple), in 20 tissue cell types (blue), in 20 primary cell types (green), and in nine embryonic stem cell types (orange), using one-sample Student’s t-test. Dotted line represents Bonferroni corrected significant threshold (P < 2.27e-3). Significant gene labels are depicted in color, non-significant gene labels in black. (C) Partitioned heritability enrichments based on cell-type-specific regulatory annotations (More details see in Methods). Heritability enrichment Z-scores, as estimated by stratified linkage disequilibrium score regression (S-LDSC) of the C-GWAS summary data for GWASs of 136 ear traits. Trait abbreviations as in S10 Table. https://doi.org/10.1371/journal.pgen.1010786.g002 Ear effects in Intu and Tbx15 mutant mice Here, we examined adult ear morphology in Intu and Tbx15 mouse mutants using one-step CRISPR/cas9 mediated gene editing experiments to assess the functional significance of Intu and Tbx15 expression during ear development. Our breeding experiments generated 18 F2 9-weeks sexually mature Intu mice i.e., 10 heterozygous Intu+/-, 8 wild-type WT (Intu+/+), while homozygous loss-of-function of Intu was lethal. Quantitative assessment of mouse ear shape (assessed by PC analysis of 21 three-dimensional landmark coordinates) revealed significant differences (linear regression PPC3 = 1.5e-3) (Fig 3C) between the heterozygous Intu littermates and WT mice. The Intu genotype significantly associated with D6_14 (P = 2.1e-3, Beta = 1.78) and D3_6 (P = 3.3e-3, Beta = 1.74) (Fig 3E and S11 Table) after FDR correction. Notably, in humans, the top INTU variant rs57788627 was nominally significantly associated with ear phenotypes L3-L6 and L3-L7, which respectively corresponds to D6_14 and D3_6 in mice (P = 1.9e-5, Beta = 0.07; P = 8.7e-5, Beta = 0.06) (Fig 3E and S11 Table). Overall, the ears of the heterozygous Intu mutant mice were shorter than those of the WT mice consistently show in the result of PC3 and distance phenotypes (Fig 3D–3F). Furthermore, the heterozygous Intu mutant mice showed a significant trend of reduction in body length and fore and hind limb length (S8 Fig), which was consistent with previous findings [13]. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 3. In-vivo mouse models of INTU deficiency. Intu mutant mice (Intu-+/-, n = 10, 9 weeks) vs. C57BL/6 wild type control mice (WT+/+, n = 8, 9 weeks) were compared for ear and body morphological differences (for details of Tbx15 mice see S9 Fig). (A) Regional association plot for the associated locus nearby INTU, and the other two SNPs at this locus, which have been recently reported the association with face morphology, were also marked. (B) The schematic diagram of the one-step CRISPR/Cas9 technology used in Intu knockout mice (see S10 Fig for details). (C) Example of left profile craniofacial photo of Intu+/- mutant mice with removed hair (left); and 21 craniofacial landmarks pattern in 3D mice image (right). (D) The principal component analysis for 21 landmarks of Intu+/- mutant mice and WT+/+ mice. The left layer shown the detailed contributed proportion of 21 landmarks to the first 5 principal components. The upper left shown the cumulative variance is explained by the first 10 PCs. The significant association between the genotype and PC3 shown in upper right (* P < 0.05, ** P < 0.01, *** P < 0.001). The bottom shown the maximum PC3-, minimum PC3- and mean ear shapes (the labels on the mouse ear were consistent with the FigC). (E) The pattern of genetic association in humans (left) and in mice (right). (F) Effect of Intu knock-out on ear phenotypes in mice (blue for effect of heterozygote mutant and red for wild type). The labels on three ears were consistent with Fig S1 (more details see in Methods). https://doi.org/10.1371/journal.pgen.1010786.g003 Our breeding experiment generated 37 F2 9-weeks adult Tbx15 mice i.e., 9 homozygous Tbx15-/-, 18 heterozygous Tbx15+/-, 10 wild-type WT (Tbx15+/+). Previous gene editing experiments have already shown that Tbx15-/- mutant mice demonstrate facial variation, loss of weight, shorter limbs and a distinct “droopy ear” feature [39]. In our study, besides the facial differences reported previously, ear differences were obvious. In addition to the distinct “droopy ear” feature, the Tbx15 genotype was significantly associated with the ear landmarks PC1 (P = 3.2e-3) and PC4 (P = 1.9e-3) (S9C Fig). Overall, the ears of the Tbx15 mutant mice were obviously longer and wider than those of the WT mice (S9E and S9F Fig). Discussion In this study of the genetic architecture of ear morphology, we developed a CNN-based deep learning model to automatically locate 17 anatomical ear landmarks, allowing us to quantify 136 ear traits in 14,921 individuals from three continents. We then used a recently developed C-GWAS method [4] to combine GWASs of ear traits. These efforts allowed us to identify 16 genetic loci associated with multiple ear traits, including 8 novel and 8 previously reported loci [1,2]. Bioinformatic analysis supports the functional role of the newly identified genes in ear morphogenesis, and gene editing experiments in mice showed that a gene desert near INTU has a functional impact on ear morphology. Our findings also highlight shared genetic contributors between some surface ectoderm-derived appearance phenotypes, with an emphasis on ear and facial morphology, male pattern baldness, and mono eyebrow. The novel ear-associations at 5 of the 8 loci we identified here i.e., 4q28.1 INTU, 5p15.1 ANKH, 8q24.13 HAS2/HAS2-AS1, 10q22.2 C10orf11, and 20q11.22 UQCC1/GDF5 are supported by other evidence, whereas for 3 novel ear genes (2q13 ACOXL, 9q33.1 ASTN2, and 21q21.3 N6AMT1) no such evidence is currently available. The strong evidence for INTU and HAS2 has been detailed above. The gene ANKH, which regulates inorganic pyrophosphate transport, is located near the lead SNP rs10062331 at 5p15.1. ANKH has been linked to diseases such as Craniometaphyseal Dysplasia (OMIM: 123000) and Chondrocalcinosis 2 (OMIM: 118600), both of which are characterized by abnormal development of bones and connective tissue. ANKH’s association with skeletal disorders suggests it may play a role in ear cartilage development. The lead SNP rs10824309 at 10q22.2 is an intronic variant of C10orf11, also known as LRMDA. This gene has been associated with craniofacial traits such as adolescent idiopathic scoliosis [40], heel bone mineral density [41], and chin shape [21]. This provides additional support for the novel association between 10q22.2 and ear development as craniofacial and ear development both stem from a common origin during early development. The lead SNP rs2378353 located at 20q11.22 is within the intron of UQCC1, which encodes a transmembrane protein involved in the assembly of the ubiquinol-cytochrome c reductase complex. Polymorphisms in this gene have been linked to human height [42] and osteoarthritis [43], and rs2378353 has been found to have a significant impact on the splicing of UQCC1 in muscle-skeletal tissue, as well as its expression in multiple tissues including fibroblasts, muscle-skeletal tissue, and skin [44]. UQCC1 is expressed in the branchial arches and head surface ectoderm during early embryonic development in mice [45]. These findings, combined with the differential expression of UQCC1 in cranial neural crest cells, suggest that rs2378353 may play a role in ear morphogenesis by regulating the expression of UQCC1 during early development. The functional information for the other 3 novel ear-associated loci identified in this study is limited. Eight of the 16 identified loci we identified here with association with quantitative ear traits have been previously reported to be associated with qualitative ear features (mainly earlobe features) in single-trait GWASs, including 1p12 TBX15 [2], 2q12.3 EDAR [1,2], 2q31.1 SP5/MYO3B [1,2], 3q23 PISRT1/MRPS22 [1,2], 4q31.3 LRBA [1,2], 6q21 PRDM1/ATG5 [2], 6q24.2 LOC153910/HIVEP2 [2], and 7q21.3 DLX6/SEM1 [2]. Two of these (4q31.3 LRBA and 7q21.3 DLX6/SEM1), which were solely identified by C-GWAS in our study, were previously identified in a single GWAS of earlobe attachment involving ~70,000 samples; hence, in a study that was five times larger than our current study. That besides the 5x smaller sample size, our study identified 8 novel loci, half of which were identified only by C-GWAS, confirms the power gain of C-GWAS in detecting small but multi-trait effects, as previously highlighted also for human face shape traits [4]. A significant portion of the 16 ear-related loci discussed here have been shown to have strong ties to human facial structure [8,15,21,22,30]. This indicates that both ear and face morphologies are influenced by common genetic factors, likely stemming from shared developmental processes in the craniofacial area. Additionally, a considerable number of the 16 ear-related loci were also previously linked to male pattern baldness, and two loci were linked to mono-eyebrow [21]. These findings suggest that not only do ears and face share genetic factors, but ears, baldness, and mono-eyebrow also share genetic factors, and it is possible that other craniofacial traits may also be influenced by similar genetic factors. Further investigation is needed to fully understand this. Our study sheds light on the genetic basis of the Darwin tubercle, a special ear phenotype that is equivalent to the ear tip of other mammals and represents a trace of human evolution [11,46]. Previous genetic studies on this trait have been limited and have not produced significant results. Our findings indicate that the HAS2 gene influences the L3-L15 phenotype, which roughly represents the Darwin tubercle, marking the beginning of understanding this intriguing aspect of human evolution. This study involved cohorts belonging to three continental populations from Europe, Asia, and America. For detecting cross-population allelic effects, C-GWAS demonstrated noticeably improved power than MinGWAS as evident by boosting the significance of known ear-associated loci while keeping the same study-wide type-I error rate as MinGWAS. However, because our approach required the a priori exclusion of SNPs not polymorphic in one of the studied cohorts, population-specific association signals, could not be detected. Another flip-side caveat of our study, in which we maximized the sample size used for discovery, is that no additional population samples were available for direct replication of our novel findings. However, that half of the loci we identified represent previously known ear loci that we re-discovered with our approach, puts confidence in our approach and thus also in the true-positive status of the newly identified loci. Additional confidence is provided by other evidence we accumulated from the literature for most of the novel loci as well as by direct functional evidence in mice for one novel gene (INTU) and one previously reported gene (TBX15). Nevertheless, future studies in independent population samples from different continental populations are warranted to further verify our novel findings. Our study represents a landmark in the field of genetics as it is the first to examine the genetic basis of quantitative ear morphology using a deep-learning-based CNN phenotyping approach. By combining the results of 136 single ear trait GWASs conducted in well-sized population samples of different ancestral origins, our C-GWAS method enabled us to identify eight new loci and confirm eight previously reported loci that play a role in ear morphology. These loci impact a wide range of ear phenotypes, demonstrating the increased power of C-GWAS in detecting multi-trait effects. Furthermore, our findings suggest that many facial and ear traits share a substantial proportion of genetic determinants, derived from the surface ectoderm. Our study significantly advances the understanding of the genetics of human ear morphology and provides a list of promising candidate genes for further functional studies. Materials and methods Ethics statement This study includes five cohorts, all cohorts were approved by the Ethics Committee and all participants provided written informed consent to participate in the study. Details in following: The Rotterdam Study (RS) has been approved by the Medical Ethics Committee of the Erasmus MC (registration number MEC 02.1015) and by the Dutch Ministry of Health, Welfare and Sport (Population Screening Act WBO, license number 1071272-159521-PG). The TwinsUK Study (TwinsUK) has been approved by the St. Thomas’ Hospital Local Research Ethics Committee. The Taizhou Longitudinal Study (TZL) has been approved by the Ethics Committee of Human Genetic Resources at the Shanghai institute of life Sciences, Chinese Academy of Sciences (ER-SIBS-261410). The National Survey of Physical Traits Study (NSPT) has been approved by the Ethics Committee of Human Genetic Resources of School of Life Sciences, Fudan University, Shanghai (14117). The Consortium for the Analysis of the Diversity and Evolution of Latin America (CANDELA) Study has been approved by ethical committee at universities in all samples countries: the Universidad Nacional Autonoma de Mexico (Mexico), the Universidad de Antioquia (Colombia), the Universidad Peruana Cayetano Heredia (Peru), the Universidad de Tarapaca (Chile), the Universidade Federal do Rio Grande do Sul (Brazil) as well as at the University College London (UK). Study cohorts This study used a total of 14,921 samples that passed quality control from five cohorts of multi-ethnic ancestries from Europe (the Rotterdam Study, RS, N = 3,675; the TwinsUK study, N = 1,065), Asia (the Taizhou Longitudinal Study, TZL, N = 2,348; the National Survey of Physical Traits study, NSPT, N = 2,487), and Latin America (the CANDELA study, N = 5,346). Details about these five cohorts are included below. The Rotterdam Study The Rotterdam Study (RS) is a population-based prospective study of individuals aged ≥ 45 years living in a suburb of Rotterdam, the Netherlands. Details regarding the cohort profile have been described elsewhere [47]. The Rotterdam Study has been approved by the Medical Ethics Committee of the Erasmus MC (registration number MEC 02.1015) and by the Dutch Ministry of Health, Welfare and Sport (Population Screening Act WBO, license number 1071272-159521-PG). The Rotterdam Study has been entered into the Netherlands National Trial Register (NTR; www.trialregister.nl) and into the WHO International Clinical Trials Registry Platform (ICTRP; www.who.int/ictrp/network/primary/en/) under shared catalogue number NTR6831. All participants provided written informed consent to participate in the study. A total of 5,604 participants not wearing make-up, cream nor jewelry, were photographed using a Premier 3dMD face3-plus UHD camera (3dMD, Atlanta, Georgia, USA). Each person’s 3D avatar is rendered from three 2D high resolution photos taken from predefined angles, i.e., from front-top, left-down, and right-down directions. Ears in the left-down and right-down photos were manually segmented and chopped for subsequent phenotyping. Genotyping was carried out using the Infinium II HumanHap 550K Genotyping BeadChip version 3 (Illumina, San Diego, California USA). All SNPs were imputed using MACH software (www.sph.umich.edu/csg/abecasis/MaCH/) based on the 1000-Genomes Project reference population information [48]. Genotype and individual quality controls have been described in detail previously [49]. In current study, after all phenotype and genotype quality controls, this study included 3,675 individuals of RS. TwinsUK Study The TwinsUK Study (TwinsUK) included 1,153 phenotyped participants (all female and all of European ancestry) within the TwinsUK adult twin registry based at St. Thomas’ Hospital in London. Volunteers gave written informed consent under a protocol reviewed by the St. Thomas’ Hospital Local Research Ethics Committee. Participants were photographed using a Premier 3dMD face3-plus UHD camera (3dMD, Atlanta, Georgia, USA), and 3D avatars were rendered from two 2D photos taken from the left and right directions. Ears in the 2D photos were manually segmented and chopped for subsequent phenotyping. Genotyping of the TwinsUK cohort was done with a combination of Illumina HumanHap300 and HumanHap610Q chips. Intensity data for each of the arrays were pooled separately and genotypes were called with the Illuminus32 calling algorithm, thresholding on a maximum posterior probability of 0.95 as previously described [50]. Imputation was performed using the IMPUTE 2.0 software package using haplotype information from the 1000-Genomes Project (Phase 1, integrated variant set across 1,092 individuals, v2, March 2012). After all phenotype and genotype quality controls, the current study included a total of 9.35 million autosomal SNPs (MAF > 0.01, imputation R2 > 0.3, SNP call rate > 0.97, HWE > 1e-6) and 1,065 individuals of TwinsUK. Taizhou Longitudinal Study The Taizhou Longitudinal Study (TZL) includes Han Chinese sampled in Taizhou, Jiangsu province in 2014 [51]. In total, 2,600 individuals were enrolled. The TZL was approved by the Ethics Committee of Human Genetic Resources at the Shanghai institute of life Sciences, Chinese Academy of Sciences (ER-SIBS-261410). All participants had provided written consent. Participants were photographed using a Canon EOS700D camera. Four digital photographs of the face: left side (−90°), frontal (0°)*2, right side (90°) were taken from ∼1.5 meters. Ears were cropped from the left side facial photographs. All samples were genotyped using the Illumina HumanOmniZhongHua-8 chips, which interrogates 894,517 SNPs. Individuals with more than 5% missing data, related individuals, and the ones that failed the X-chromosome sex concordance check or had ethnic information incompatible with their genetic information were excluded. SNPs with more than 2% missing data, with a minor allele frequency smaller than 1%, and the ones that failed the Hardy–Weinberg deviation test (P < 1e-5) were also excluded. After applying these filters, we obtained a dataset of 2,600 samples with 776,213 SNPs. The chip genotype data were firstly phased using SHAPEIT [52]. IMPUTE2 [53] was then used to impute genotypes at non-genotyped SNPs using the 1000 Genomes Phase 3 data as the reference panel. After all phenotype and genotype quality controls, the current study included a total of 7,057,720 imputed and genotyped SNPs and 2,348 individuals of TZL. The National Survey of Physical Traits Study The National Survey of Physical Traits Study (NSPT) is part of the National Science & Technology Basic Research Project, which contained four sub-cohorts collected from three different regions of China in different years: Taizhou, Jiangsu province in 2015 and 2019, Zhengzhou, Henan province in 2017, Nanning, Guangxi province in 2018. The NSPT project was approved by the Ethics Committee of Human Genetic Resources of School of Life Sciences, Fudan University, Shanghai (14117). All participants provided written informed consent. Participants were photographed using a Canon EOS700D camera. Ten digital photographs of the face: left side (−90°)*2, left angle (−45°)*2, frontal (0°)*2, right angle (45°)*2, right side (90°)*2 were taken from ∼1.5 meters. Photos in each direction include one with eyes open and one with eyes closed. Ears were cropped from the left side facial photographs with open eyes. DNA was extracted from blood samples using the MagPure Blood DNA KF Kit. The DNA samples were genotyped on the Illumina Infinium Global Screening Array that investigates 707,180 variants which is a fully custom array designed by WeGene (https://www.wegene.com/). Individuals with more than 5% missing data, related individuals, and the ones that failed the X-chromosome sex concordance check or had ethnic information incompatible with their genetic information were excluded. In total, 2,487 individuals were enrolled (15HanTZ: N = 404; 17HanZZ: N = 644; 18HanNN: N = 1,119; 19HanTZ: N = 320). The genotype data were phased using SHAPEIT [52] and imputed to the 1000 Genomes Project Phase 3 reference panel using IMPUTE2 [53]. Variants exclusion criteria included INFO < 0.8, certainty score < 0.9, MAF < 0.02, SNP-wise call rate > 5%, and deviation from Hardy-Weinberg equilibrium (P < 1×10−6). After all phenotype and genotype quality controls, the current study included a total of 8,018,212 imputed and genotyped SNPs and 2,487 individuals of NSPT. CANDELA Study The Consortium for the Analysis of the Diversity and Evolution of Latin America (CANDELA) Study [54] consists of 6,630 volunteers recruited in five Latin American countries (Brazil, Colombia, Chile, Mexico and Peru). Ethics approvals were obtained from ethical committee at universities in all samples countries: the Universidad Nacional Autonoma de Mexico (Mexico), the Universidad de Antioquia (Colombia), the Universidad Peruana Cayetano Heredia (Peru), the Universidad de Tarapaca (Chile), the Universidade Federal do Rio Grande do Sul (Brazil) as well as at the University College London (UK). All participants provided written informed consent. Five digital photographs of the face: left side (−90°), left angle (−45°), frontal (0°), right angle (45°), right side (90°) were taken from ~1.5 meters at eye level using a Nikon D90 camera fitted with a Nikkor 50 mm fixed focal length lens. Ears were cropped from the left side facial photographs. DNA samples were genotyped on Illumina’s Omni Express BeadChip. After applying quality control filters 669, 462 SNPs and 6,357 individuals were retained for further analyses (2,922 males, 3,435 females). Average admixture proportions for this sample were estimated as: 48% European, 46% Native American and 6% African, but with substantial inter-individual variation. After all genomic and phenotypic quality controls this study included 6,238 individuals. The genetic PCs were obtained from the LD-pruned dataset of 93,328 SNPs. These PCs were selected by inspecting the proportion of variance explained and checking scatter and scree plots. The final imputed dataset used in the GWAS analyses included genotypes for 9,143,600 SNPs using the 1000-Genomes Phase I reference panel. After all phenotype and genotype quality controls, the current study included a total of 9,143,600 imputed and genotyped SNPs and 5,346 individuals of CANDELA. Deep learning based ear phenotyping Our computer pipeline for quantitative ear phenotyping, which includes automated ear detection, segmentation, landmarking, and phenotype acquisition is available at https://github.com/Fun-Gene/EarPhenotyping. In all cohorts, we used the same pipeline to minimize potential variation caused by different methodology. Specifically, we used the previously established two-stage ear landmark detector consisting of two CNNs deep-learnt from ~15,000 labeled ear images [3]. Both CNNs had the same architecture consisting of alternating 3 convolutional layers and 3 max-pooling layers followed by 3 fully connected layers. The 1st CNN detects various orientations and sizes and rectifies all coordinates at a coarse-grained level. The 2nd CNN was trained in a more controlled scenario for accurate landmarking of 55 ear landmarks. The performance of this two-stage landmark detector has been verified in various scenarios of difficulty that represents the state-of-the-art in ear landmark detection. In our application, we double checked all resultant landmarks from all images by eye-balling and removed those low-quality images that failed in CNN-based landmarking. After obtaining the coordinates of 55 landmarks, Generalized Procrustes Analysis (GPA) was used to remove affine variations due to shifting, rotation, and scaling. We then focused on the 17 most anatomical meaningful landmarks covering the entire ear (S1A Fig and S12 Table) and from those, calculated 136 inter-landmark distances as ear phenotypes in the genetic studies. Outliers greater than 3 standard deviations were removed and Z-transformed values were used in all subsequent analyses. For quality control purposes, a trained rater carefully labeled the 17 landmarks on both the left and the right ear in 50 randomly selected and shuffled images. Pearson’s correlations were calculated between the left and the right ear phenotypes from this rater, which were compared with the left-right correlations from the CNN approach. Within the same ear (the left side only), Pearson’s correlations between the rater and the CNN approach were calculated. Genetic correlation, heritability, GWASs, meta-analyses Genetic correlations were estimated from all SNPs using GEMMA [55] based on a multivariate linear mixed model. Unsupervised hierarchical clustering analysis is conducted using 1-abs (correlation) as a dissimilarity matrix, and each iteration is updated using the Lance-Williams formula [56]. Twin heritability was estimated using phenotype correlations in monozygotic (MZ) and dizygotic twins (DZ), h2 = 2(r(MZ)-r(DZ)). SNP-based heritability was estimated from all GWAS SNPs using the restricted maximum likelihood estimation in GCTA [57]. GWASs were independently carried out in each cohort. GWASs in unrelated individuals (RS, TZL, NSPT, and CANDELA) were performed using linear models assuming an additive allele effect adjusted for covariates sex, age, BMI, and top 5–10 genomic PCs using PLINK 1.9 [58]. GWASs in TwinsUK (females only) were performed using GEMMA [55], which implements an LME model with an empirical genetic relatedness matrix to account for cryptic pedigree and population structure. All cohorts were aligned according to the genome-build GRCh37.p13. Meta-analyses of all cohorts were conducted using the inverse variance fixed-effect model in PLINK 1.9. All statistical analyses were conducted using the R Environment for Statistical Computing (version 3.5.2) unless otherwise specified. C-GWAS In the application of C-GWAS analysis on 136 ear meta-analyses, we focused on SNPs with an observable frequency (MAF > 0.01) in three different continental groups (European, East Asian, and Latin American). The C-GWAS is an R library that is freely available at https://github.com/Fun-Gene/CGWAS. In the current application, default parameters were used. The summary statistics of the 136 meta-analyses were used as the input of C-GWAS. Details of C-GWAS method has been described previously [4]. In brief, the null hypothesis (H0) for a SNP under testing is the absence of any allelic effect on all traits, and the alternative hypothesis (H1) is that its allelic effects deviate from 0 for at least one of the multiple traits. C-GWAS incorporate two different tests originated from either the effect based inversed covariance weighting or the truncated Wald test to maximize statistical power. All resultant P-values from C-GWAS are adjusted using the getCoef function implemented in C-GWAS, which performs simulations (n simulations = 1e8) to guarantee that the null distribution of C-GWAS follows the uniform distribution in all quantiles. This simulation analysis was also applied for adjusting for minimal p-values of meta-analysis of 136 traits, and the adjusted minimal P-values are abbreviated as MinGWAS. Therefore, C-GWAS and from MinGWAS are directly comparable with each other and with any standard single-trait GWAS, so that the traditional genome-wide significance threshold of 5e-8 corresponds to our study-wide significance threshold. C-GWAS completed the analysis of 136 traits and 4,803,785 SNPs within 2 hours with 16 threads in parallel and peak memory usage of 32 GB. Post-GWAS analyses SNPs and nearby genes were annotated using ANNOVAR [59]. Enrichment analysis of biological processes, molecular functions, and cellular components were conducted using Metascape [60] based on the Gene Ontology (GO) database. Regulatory activities of associated SNPs and 3D interacting genes were explored using the 3DSNP database [33]. Enhancer activities and embryonic expression patterns at the associated loci were examined in transgenic mice in the Vista Enhancer Browser database [34]. Potential functional links between nearby genes, 3D interacting genes and ear/craniofacial features were examined in the Harmonizome database [35] and gene expression patterns in the branchial arch and embryonic ectoderm were examined in the MGI database [36,37]. Gene expressions in cranial neural crest cells (CNCCs) were compared between genes of interest and background genes over the genome using RNA-seq data from Prescott et al. [61], GTEx [44], and ENCODE [62]. We attained all annotated nearest genes of the 1e-4 sets of matched SNPs for 16 lead SNPs using SNPsnap [63], the 16 genes with the median expression value separately in 50 cell types were as the correspondent control genes. Partitioned heritability enrichments based on CNCCs [61], 4-stages of embryonic craniofacial tissues [64] and 97 cell types form Roadmap Epigenomics resource [65,66]. The annotations for CNCCs, embryonic craniofacial tissues and other cell types refer the previous study [67] using S-LDSC [38]. For CNCCs, we downloaded and processed H3K27ac and ATAC data, and all data from different replicates. Peaks were called using MACS2 [68]. We iteratively obtained and combined the most reliable peaks based on a 50% peak overlap rate for all replicates from of H3K27ac and ATAC data using BEDtools. [69]. Based on the combined peaks, we used ROSE [70] to infer enhancers including super-enhancers and annotated all enhancer regions (S14 Table). For embryonic craniofacial tissues, we combined all regions with the following annotations from the 25-state chromHMM model: ‘Enh’, ‘TxReg’, ‘PromD1’, ‘PromD2’, ‘PromU’ and ‘TssA’. For other cell types, we combined all regions with the following annotations from the 15-state chromHMM model: ‘1_TssA’, ‘2_TssAFlnk’, ‘7_Enh’ and ‘6_EnhG’. Each annotation was individually added to the baseline LD model [38]. Also, we use the C-GWAS summary data for 136 meta-analysis in European populations including RS and TwinsUK. Pleiotropic associations were looked up in GWAScatalog based on the region (basepair position of lead SNP +- 500kb) and annotated gene [71]. CRISPR-Cas9-mediated gene editing in mouse We generated C57BL/6J Intu knockout mice using the one-step CRISPR/Cas9 method [72] and tightly followed the steps described in a previous study of Tbx15 and Pax1 genes [39]. In brief, fertilized eggs obtained from super-ovulated females (four weeks) mated with males (seven-eight weeks) were microinjected with mixtures of Cas9 mRNA and sgRNA (S10 Fig). The injected eggs were cultured to day two and transferred to female mice. We obtained only positive heterozygote mutants because homozygote mutants of Intu were fatal. Compared with the intact allele (402kb), the mutant allele had a reduced sequence (251kb) with a removal of exons from 2 to 15 (S10 Fig). Mice were raised in a pathogen-free environment and bred according to SPF animal breeding standards. After eight months breeding experiments, we obtained a total of 19 F2 sexually mature mice (9-weeks), including 8 wild-type and 11 heterozygote mutant mice. For Tbx15, we used previously generated mice [39], including 10 wild-type, 18 heterozygote, and 10 homozygote mutant mice. The use of laboratory animals (SYXK 2019–0022) was licensed by the Beijing Municipal Science and Technology Commission. Mouse pinna phenotyping F2 sexually mature mice were sacrificed by cervical dislocation. Hair was removed with a razor and Weiting depilatory cream. A HandySCAN BLACK scanner was used to obtain 3D models (resolution 0.3 mm, accuracy 0.03 mm). The coordinates of 9 anatomical landmarks (3 facial and 6 ear landmarks) and 12 pseudo-landmarks were obtained using Geomagic Wrap (Fig 2C and S13 Table). The 12 pseudo-landmarks were equally spaced from the four partial curves of the mouse auricle contour (including L1-L4, L4-L6, L6-L5, and L5-L3) for better covering the contour of the auricle. The R package ‘geomorph’ were used for series of analysis (https://github.com/geomorphR/geomorph). GPA was used to remove the effects of translation, rotation and scaling. After superimposition of the GPA-adjusted coordinates, only the shape component remained in the aligned specimens. Mice with the outlier in PC1 were excluded then the GPA and PCA were redone. The first 10 PCs were used to represent dominant but different dimensional variations of all landmarks. The ear shape visualization of shape variation based on the maximum and minimum PC compare to mean ear shape. The inter-landmark also distances also were visualized (S13 Table). Z-transformed variables were used for association analysis. The effect of per mutant allele on ear shape was tested using linear regression with covariate sex, weight and body length. Multiple testing was corrected using FDR method. Ethics statement This study includes five cohorts, all cohorts were approved by the Ethics Committee and all participants provided written informed consent to participate in the study. Details in following: The Rotterdam Study (RS) has been approved by the Medical Ethics Committee of the Erasmus MC (registration number MEC 02.1015) and by the Dutch Ministry of Health, Welfare and Sport (Population Screening Act WBO, license number 1071272-159521-PG). The TwinsUK Study (TwinsUK) has been approved by the St. Thomas’ Hospital Local Research Ethics Committee. The Taizhou Longitudinal Study (TZL) has been approved by the Ethics Committee of Human Genetic Resources at the Shanghai institute of life Sciences, Chinese Academy of Sciences (ER-SIBS-261410). The National Survey of Physical Traits Study (NSPT) has been approved by the Ethics Committee of Human Genetic Resources of School of Life Sciences, Fudan University, Shanghai (14117). The Consortium for the Analysis of the Diversity and Evolution of Latin America (CANDELA) Study has been approved by ethical committee at universities in all samples countries: the Universidad Nacional Autonoma de Mexico (Mexico), the Universidad de Antioquia (Colombia), the Universidad Peruana Cayetano Heredia (Peru), the Universidad de Tarapaca (Chile), the Universidade Federal do Rio Grande do Sul (Brazil) as well as at the University College London (UK). Study cohorts This study used a total of 14,921 samples that passed quality control from five cohorts of multi-ethnic ancestries from Europe (the Rotterdam Study, RS, N = 3,675; the TwinsUK study, N = 1,065), Asia (the Taizhou Longitudinal Study, TZL, N = 2,348; the National Survey of Physical Traits study, NSPT, N = 2,487), and Latin America (the CANDELA study, N = 5,346). Details about these five cohorts are included below. The Rotterdam Study The Rotterdam Study (RS) is a population-based prospective study of individuals aged ≥ 45 years living in a suburb of Rotterdam, the Netherlands. Details regarding the cohort profile have been described elsewhere [47]. The Rotterdam Study has been approved by the Medical Ethics Committee of the Erasmus MC (registration number MEC 02.1015) and by the Dutch Ministry of Health, Welfare and Sport (Population Screening Act WBO, license number 1071272-159521-PG). The Rotterdam Study has been entered into the Netherlands National Trial Register (NTR; www.trialregister.nl) and into the WHO International Clinical Trials Registry Platform (ICTRP; www.who.int/ictrp/network/primary/en/) under shared catalogue number NTR6831. All participants provided written informed consent to participate in the study. A total of 5,604 participants not wearing make-up, cream nor jewelry, were photographed using a Premier 3dMD face3-plus UHD camera (3dMD, Atlanta, Georgia, USA). Each person’s 3D avatar is rendered from three 2D high resolution photos taken from predefined angles, i.e., from front-top, left-down, and right-down directions. Ears in the left-down and right-down photos were manually segmented and chopped for subsequent phenotyping. Genotyping was carried out using the Infinium II HumanHap 550K Genotyping BeadChip version 3 (Illumina, San Diego, California USA). All SNPs were imputed using MACH software (www.sph.umich.edu/csg/abecasis/MaCH/) based on the 1000-Genomes Project reference population information [48]. Genotype and individual quality controls have been described in detail previously [49]. In current study, after all phenotype and genotype quality controls, this study included 3,675 individuals of RS. TwinsUK Study The TwinsUK Study (TwinsUK) included 1,153 phenotyped participants (all female and all of European ancestry) within the TwinsUK adult twin registry based at St. Thomas’ Hospital in London. Volunteers gave written informed consent under a protocol reviewed by the St. Thomas’ Hospital Local Research Ethics Committee. Participants were photographed using a Premier 3dMD face3-plus UHD camera (3dMD, Atlanta, Georgia, USA), and 3D avatars were rendered from two 2D photos taken from the left and right directions. Ears in the 2D photos were manually segmented and chopped for subsequent phenotyping. Genotyping of the TwinsUK cohort was done with a combination of Illumina HumanHap300 and HumanHap610Q chips. Intensity data for each of the arrays were pooled separately and genotypes were called with the Illuminus32 calling algorithm, thresholding on a maximum posterior probability of 0.95 as previously described [50]. Imputation was performed using the IMPUTE 2.0 software package using haplotype information from the 1000-Genomes Project (Phase 1, integrated variant set across 1,092 individuals, v2, March 2012). After all phenotype and genotype quality controls, the current study included a total of 9.35 million autosomal SNPs (MAF > 0.01, imputation R2 > 0.3, SNP call rate > 0.97, HWE > 1e-6) and 1,065 individuals of TwinsUK. Taizhou Longitudinal Study The Taizhou Longitudinal Study (TZL) includes Han Chinese sampled in Taizhou, Jiangsu province in 2014 [51]. In total, 2,600 individuals were enrolled. The TZL was approved by the Ethics Committee of Human Genetic Resources at the Shanghai institute of life Sciences, Chinese Academy of Sciences (ER-SIBS-261410). All participants had provided written consent. Participants were photographed using a Canon EOS700D camera. Four digital photographs of the face: left side (−90°), frontal (0°)*2, right side (90°) were taken from ∼1.5 meters. Ears were cropped from the left side facial photographs. All samples were genotyped using the Illumina HumanOmniZhongHua-8 chips, which interrogates 894,517 SNPs. Individuals with more than 5% missing data, related individuals, and the ones that failed the X-chromosome sex concordance check or had ethnic information incompatible with their genetic information were excluded. SNPs with more than 2% missing data, with a minor allele frequency smaller than 1%, and the ones that failed the Hardy–Weinberg deviation test (P < 1e-5) were also excluded. After applying these filters, we obtained a dataset of 2,600 samples with 776,213 SNPs. The chip genotype data were firstly phased using SHAPEIT [52]. IMPUTE2 [53] was then used to impute genotypes at non-genotyped SNPs using the 1000 Genomes Phase 3 data as the reference panel. After all phenotype and genotype quality controls, the current study included a total of 7,057,720 imputed and genotyped SNPs and 2,348 individuals of TZL. The National Survey of Physical Traits Study The National Survey of Physical Traits Study (NSPT) is part of the National Science & Technology Basic Research Project, which contained four sub-cohorts collected from three different regions of China in different years: Taizhou, Jiangsu province in 2015 and 2019, Zhengzhou, Henan province in 2017, Nanning, Guangxi province in 2018. The NSPT project was approved by the Ethics Committee of Human Genetic Resources of School of Life Sciences, Fudan University, Shanghai (14117). All participants provided written informed consent. Participants were photographed using a Canon EOS700D camera. Ten digital photographs of the face: left side (−90°)*2, left angle (−45°)*2, frontal (0°)*2, right angle (45°)*2, right side (90°)*2 were taken from ∼1.5 meters. Photos in each direction include one with eyes open and one with eyes closed. Ears were cropped from the left side facial photographs with open eyes. DNA was extracted from blood samples using the MagPure Blood DNA KF Kit. The DNA samples were genotyped on the Illumina Infinium Global Screening Array that investigates 707,180 variants which is a fully custom array designed by WeGene (https://www.wegene.com/). Individuals with more than 5% missing data, related individuals, and the ones that failed the X-chromosome sex concordance check or had ethnic information incompatible with their genetic information were excluded. In total, 2,487 individuals were enrolled (15HanTZ: N = 404; 17HanZZ: N = 644; 18HanNN: N = 1,119; 19HanTZ: N = 320). The genotype data were phased using SHAPEIT [52] and imputed to the 1000 Genomes Project Phase 3 reference panel using IMPUTE2 [53]. Variants exclusion criteria included INFO < 0.8, certainty score < 0.9, MAF < 0.02, SNP-wise call rate > 5%, and deviation from Hardy-Weinberg equilibrium (P < 1×10−6). After all phenotype and genotype quality controls, the current study included a total of 8,018,212 imputed and genotyped SNPs and 2,487 individuals of NSPT. CANDELA Study The Consortium for the Analysis of the Diversity and Evolution of Latin America (CANDELA) Study [54] consists of 6,630 volunteers recruited in five Latin American countries (Brazil, Colombia, Chile, Mexico and Peru). Ethics approvals were obtained from ethical committee at universities in all samples countries: the Universidad Nacional Autonoma de Mexico (Mexico), the Universidad de Antioquia (Colombia), the Universidad Peruana Cayetano Heredia (Peru), the Universidad de Tarapaca (Chile), the Universidade Federal do Rio Grande do Sul (Brazil) as well as at the University College London (UK). All participants provided written informed consent. Five digital photographs of the face: left side (−90°), left angle (−45°), frontal (0°), right angle (45°), right side (90°) were taken from ~1.5 meters at eye level using a Nikon D90 camera fitted with a Nikkor 50 mm fixed focal length lens. Ears were cropped from the left side facial photographs. DNA samples were genotyped on Illumina’s Omni Express BeadChip. After applying quality control filters 669, 462 SNPs and 6,357 individuals were retained for further analyses (2,922 males, 3,435 females). Average admixture proportions for this sample were estimated as: 48% European, 46% Native American and 6% African, but with substantial inter-individual variation. After all genomic and phenotypic quality controls this study included 6,238 individuals. The genetic PCs were obtained from the LD-pruned dataset of 93,328 SNPs. These PCs were selected by inspecting the proportion of variance explained and checking scatter and scree plots. The final imputed dataset used in the GWAS analyses included genotypes for 9,143,600 SNPs using the 1000-Genomes Phase I reference panel. After all phenotype and genotype quality controls, the current study included a total of 9,143,600 imputed and genotyped SNPs and 5,346 individuals of CANDELA. Deep learning based ear phenotyping Our computer pipeline for quantitative ear phenotyping, which includes automated ear detection, segmentation, landmarking, and phenotype acquisition is available at https://github.com/Fun-Gene/EarPhenotyping. In all cohorts, we used the same pipeline to minimize potential variation caused by different methodology. Specifically, we used the previously established two-stage ear landmark detector consisting of two CNNs deep-learnt from ~15,000 labeled ear images [3]. Both CNNs had the same architecture consisting of alternating 3 convolutional layers and 3 max-pooling layers followed by 3 fully connected layers. The 1st CNN detects various orientations and sizes and rectifies all coordinates at a coarse-grained level. The 2nd CNN was trained in a more controlled scenario for accurate landmarking of 55 ear landmarks. The performance of this two-stage landmark detector has been verified in various scenarios of difficulty that represents the state-of-the-art in ear landmark detection. In our application, we double checked all resultant landmarks from all images by eye-balling and removed those low-quality images that failed in CNN-based landmarking. After obtaining the coordinates of 55 landmarks, Generalized Procrustes Analysis (GPA) was used to remove affine variations due to shifting, rotation, and scaling. We then focused on the 17 most anatomical meaningful landmarks covering the entire ear (S1A Fig and S12 Table) and from those, calculated 136 inter-landmark distances as ear phenotypes in the genetic studies. Outliers greater than 3 standard deviations were removed and Z-transformed values were used in all subsequent analyses. For quality control purposes, a trained rater carefully labeled the 17 landmarks on both the left and the right ear in 50 randomly selected and shuffled images. Pearson’s correlations were calculated between the left and the right ear phenotypes from this rater, which were compared with the left-right correlations from the CNN approach. Within the same ear (the left side only), Pearson’s correlations between the rater and the CNN approach were calculated. Genetic correlation, heritability, GWASs, meta-analyses Genetic correlations were estimated from all SNPs using GEMMA [55] based on a multivariate linear mixed model. Unsupervised hierarchical clustering analysis is conducted using 1-abs (correlation) as a dissimilarity matrix, and each iteration is updated using the Lance-Williams formula [56]. Twin heritability was estimated using phenotype correlations in monozygotic (MZ) and dizygotic twins (DZ), h2 = 2(r(MZ)-r(DZ)). SNP-based heritability was estimated from all GWAS SNPs using the restricted maximum likelihood estimation in GCTA [57]. GWASs were independently carried out in each cohort. GWASs in unrelated individuals (RS, TZL, NSPT, and CANDELA) were performed using linear models assuming an additive allele effect adjusted for covariates sex, age, BMI, and top 5–10 genomic PCs using PLINK 1.9 [58]. GWASs in TwinsUK (females only) were performed using GEMMA [55], which implements an LME model with an empirical genetic relatedness matrix to account for cryptic pedigree and population structure. All cohorts were aligned according to the genome-build GRCh37.p13. Meta-analyses of all cohorts were conducted using the inverse variance fixed-effect model in PLINK 1.9. All statistical analyses were conducted using the R Environment for Statistical Computing (version 3.5.2) unless otherwise specified. C-GWAS In the application of C-GWAS analysis on 136 ear meta-analyses, we focused on SNPs with an observable frequency (MAF > 0.01) in three different continental groups (European, East Asian, and Latin American). The C-GWAS is an R library that is freely available at https://github.com/Fun-Gene/CGWAS. In the current application, default parameters were used. The summary statistics of the 136 meta-analyses were used as the input of C-GWAS. Details of C-GWAS method has been described previously [4]. In brief, the null hypothesis (H0) for a SNP under testing is the absence of any allelic effect on all traits, and the alternative hypothesis (H1) is that its allelic effects deviate from 0 for at least one of the multiple traits. C-GWAS incorporate two different tests originated from either the effect based inversed covariance weighting or the truncated Wald test to maximize statistical power. All resultant P-values from C-GWAS are adjusted using the getCoef function implemented in C-GWAS, which performs simulations (n simulations = 1e8) to guarantee that the null distribution of C-GWAS follows the uniform distribution in all quantiles. This simulation analysis was also applied for adjusting for minimal p-values of meta-analysis of 136 traits, and the adjusted minimal P-values are abbreviated as MinGWAS. Therefore, C-GWAS and from MinGWAS are directly comparable with each other and with any standard single-trait GWAS, so that the traditional genome-wide significance threshold of 5e-8 corresponds to our study-wide significance threshold. C-GWAS completed the analysis of 136 traits and 4,803,785 SNPs within 2 hours with 16 threads in parallel and peak memory usage of 32 GB. Post-GWAS analyses SNPs and nearby genes were annotated using ANNOVAR [59]. Enrichment analysis of biological processes, molecular functions, and cellular components were conducted using Metascape [60] based on the Gene Ontology (GO) database. Regulatory activities of associated SNPs and 3D interacting genes were explored using the 3DSNP database [33]. Enhancer activities and embryonic expression patterns at the associated loci were examined in transgenic mice in the Vista Enhancer Browser database [34]. Potential functional links between nearby genes, 3D interacting genes and ear/craniofacial features were examined in the Harmonizome database [35] and gene expression patterns in the branchial arch and embryonic ectoderm were examined in the MGI database [36,37]. Gene expressions in cranial neural crest cells (CNCCs) were compared between genes of interest and background genes over the genome using RNA-seq data from Prescott et al. [61], GTEx [44], and ENCODE [62]. We attained all annotated nearest genes of the 1e-4 sets of matched SNPs for 16 lead SNPs using SNPsnap [63], the 16 genes with the median expression value separately in 50 cell types were as the correspondent control genes. Partitioned heritability enrichments based on CNCCs [61], 4-stages of embryonic craniofacial tissues [64] and 97 cell types form Roadmap Epigenomics resource [65,66]. The annotations for CNCCs, embryonic craniofacial tissues and other cell types refer the previous study [67] using S-LDSC [38]. For CNCCs, we downloaded and processed H3K27ac and ATAC data, and all data from different replicates. Peaks were called using MACS2 [68]. We iteratively obtained and combined the most reliable peaks based on a 50% peak overlap rate for all replicates from of H3K27ac and ATAC data using BEDtools. [69]. Based on the combined peaks, we used ROSE [70] to infer enhancers including super-enhancers and annotated all enhancer regions (S14 Table). For embryonic craniofacial tissues, we combined all regions with the following annotations from the 25-state chromHMM model: ‘Enh’, ‘TxReg’, ‘PromD1’, ‘PromD2’, ‘PromU’ and ‘TssA’. For other cell types, we combined all regions with the following annotations from the 15-state chromHMM model: ‘1_TssA’, ‘2_TssAFlnk’, ‘7_Enh’ and ‘6_EnhG’. Each annotation was individually added to the baseline LD model [38]. Also, we use the C-GWAS summary data for 136 meta-analysis in European populations including RS and TwinsUK. Pleiotropic associations were looked up in GWAScatalog based on the region (basepair position of lead SNP +- 500kb) and annotated gene [71]. CRISPR-Cas9-mediated gene editing in mouse We generated C57BL/6J Intu knockout mice using the one-step CRISPR/Cas9 method [72] and tightly followed the steps described in a previous study of Tbx15 and Pax1 genes [39]. In brief, fertilized eggs obtained from super-ovulated females (four weeks) mated with males (seven-eight weeks) were microinjected with mixtures of Cas9 mRNA and sgRNA (S10 Fig). The injected eggs were cultured to day two and transferred to female mice. We obtained only positive heterozygote mutants because homozygote mutants of Intu were fatal. Compared with the intact allele (402kb), the mutant allele had a reduced sequence (251kb) with a removal of exons from 2 to 15 (S10 Fig). Mice were raised in a pathogen-free environment and bred according to SPF animal breeding standards. After eight months breeding experiments, we obtained a total of 19 F2 sexually mature mice (9-weeks), including 8 wild-type and 11 heterozygote mutant mice. For Tbx15, we used previously generated mice [39], including 10 wild-type, 18 heterozygote, and 10 homozygote mutant mice. The use of laboratory animals (SYXK 2019–0022) was licensed by the Beijing Municipal Science and Technology Commission. Mouse pinna phenotyping F2 sexually mature mice were sacrificed by cervical dislocation. Hair was removed with a razor and Weiting depilatory cream. A HandySCAN BLACK scanner was used to obtain 3D models (resolution 0.3 mm, accuracy 0.03 mm). The coordinates of 9 anatomical landmarks (3 facial and 6 ear landmarks) and 12 pseudo-landmarks were obtained using Geomagic Wrap (Fig 2C and S13 Table). The 12 pseudo-landmarks were equally spaced from the four partial curves of the mouse auricle contour (including L1-L4, L4-L6, L6-L5, and L5-L3) for better covering the contour of the auricle. The R package ‘geomorph’ were used for series of analysis (https://github.com/geomorphR/geomorph). GPA was used to remove the effects of translation, rotation and scaling. After superimposition of the GPA-adjusted coordinates, only the shape component remained in the aligned specimens. Mice with the outlier in PC1 were excluded then the GPA and PCA were redone. The first 10 PCs were used to represent dominant but different dimensional variations of all landmarks. The ear shape visualization of shape variation based on the maximum and minimum PC compare to mean ear shape. The inter-landmark also distances also were visualized (S13 Table). Z-transformed variables were used for association analysis. The effect of per mutant allele on ear shape was tested using linear regression with covariate sex, weight and body length. Multiple testing was corrected using FDR method. Supporting information S1 Fig. Study design. (A) The location of selected 17 ear landmarks. (B) Design of the current study. https://doi.org/10.1371/journal.pgen.1010786.s001 (TIF) S2 Fig. Pearson’s correlation coefficients for 136 ear phenotypes derived from different methods (auto pipeline and manual-landmarking) and different ears (left and right ears). (A) correlation between left ear phenotypes by manual-landmarking (expert 1 vs. expert 2). (B) correlation between right ear phenotypes by manual-landmarking and auto-landmarking (expert 1 vs. auto). (C) correlation between phenotypes from left and right ears by manual-landmarking. (D) correlation between phenotypes from left and right ears by auto-landmarking. https://doi.org/10.1371/journal.pgen.1010786.s002 (TIF) S3 Fig. Effects of sex (left) and age (right) on 136 ear phenotypes in RS. Please note the different figure legends in these two figures. https://doi.org/10.1371/journal.pgen.1010786.s003 (TIF) S4 Fig. Two distinct clusters of 136 ear phenotypes derived from unsupervised hierarchical clustering. (A) Two clusters for 136 phenotypes. (B) Phenotypic (right up) and genetic correlation matrix (left down) within and between the cluster. https://doi.org/10.1371/journal.pgen.1010786.s004 (TIF) S5 Fig. Heritability of 136 ear phenotypes estimated in TwinsUK. (A) Twin heritability. (B) SNP-based heritability. https://doi.org/10.1371/journal.pgen.1010786.s005 (TIF) S6 Fig. 16 ear-associated loci we identified in our current MinGWAS and C-GWAS. The first 8 (A-H) were novel loci, others 8 were previously reported loci (I-P). Each figure includes three figures, LocusZoom (up) shows regional association plots for the top-associated ear phenotype (p values in CGWAS, except for the 6q21 PRDM1/ATG5, which solely identified by meta-analysis) with candidate genes aligned below according to the chromosomal positions (GRCh37.p13) followed by the linkage disequilibrium (LD) patterns (r2) of European. Ear map (left lower) shows the association (p values in Meta-analysis) between all ear phenotypes (P < 1e-3) and top-SNP identified in our analysis. Effect plot (right lower) shows effect sizes for the effect allele of top-SNPs from the association with top-associated ear phenotype in all 5 GWASs and meta-analysis. https://doi.org/10.1371/journal.pgen.1010786.s006 (DOCX) S7 Fig. Various degrees of Darwin’s tubercle. https://doi.org/10.1371/journal.pgen.1010786.s007 (TIF) S8 Fig. Comparison of 4 phenotypes of Intu+/- mutant mice and WT+/+ mice including weight, body length, forelimb length and posterior limb length (* P < 0.05, ** P < 0.01, *** P < 0.001). https://doi.org/10.1371/journal.pgen.1010786.s008 (TIF) S9 Fig. In-vivo mouse models of TBX15 deficiency. Homozygous Tbx15-/- mutant mice (N = 9, 9 weeks), heterozygous Tbx15+/- (N = 18, 9 weeks) and C57BL/6 WT+/+ control mice (N = 10, 9 weeks) were compared for ear and body morphological differences. (A) The schematic diagram of the one-step CRISPR/Cas9 technology used in Tbx15 knockout mice. (B) Example of left profile craniofacial photo of Tbx15-/- mutant mice with removal hair. (C) The principal component analysis for 21 landmarks of Tbx15-/- mutant mice, heterozygous Tbx15+/- and WT+/+ mice. The upper layer shown the detailed contribution proportion of 21 landmarks to the first 10 principal components. The middle layer shown the screenplot of first 10 PCs, the significant association between the genotype and PCs which including PC1 and PC4 (* P < 0.05, ** P < 0.01, *** P < 0.001). The bottom shown the maximum PC1-, minimum PC1-,maximum PC4-, minimum PC4-, and mean ear shapes. (D) The pattern of genetic association in humans (left) and in mice (right). (E) Effect of Tbx15 knock-out on ear phenotypes in mice (blue for effect of heterozygote mutant and red for wildtype). https://doi.org/10.1371/journal.pgen.1010786.s009 (TIF) S10 Fig. The INTU knockout mice using one-step CRISPR/Cas9 technology. (A) The schematic diagram. (B) the sequence of gRNA. (C) Genotype was identified by PCR (Positive 251kb (loss-function), negative 402kb (wild type), heterozygous including 251kb and 402kb). https://doi.org/10.1371/journal.pgen.1010786.s010 (TIF) S1 Table. Characteristics of 5 cohorts. https://doi.org/10.1371/journal.pgen.1010786.s011 (XLSX) S2 Table. Characteristics of 136 ear phenotypes in 5 cohorts. https://doi.org/10.1371/journal.pgen.1010786.s012 (XLSX) S3 Table. Kolmogorov-Smirnov normality tests of 136 ear phenotypes in 5 cohorts. https://doi.org/10.1371/journal.pgen.1010786.s013 (XLSX) S4 Table. The effects of age and sex on 136 ear phenotypes in the RS cohort. https://doi.org/10.1371/journal.pgen.1010786.s014 (XLSX) S5 Table. Replication of previously ear-associated SNPs in MinGWAS and C-GWAS. https://doi.org/10.1371/journal.pgen.1010786.s015 (XLSX) S6 Table. Replication of previously face-associated SNPs in MinGWAS and C-GWAS. https://doi.org/10.1371/journal.pgen.1010786.s016 (XLSX) S7 Table. GWAS Catalog entries for significant SNPs in MinGWAS and C-GWAS. https://doi.org/10.1371/journal.pgen.1010786.s017 (XLSX) S8 Table. Enrichment of 31 genes in the 16 ear-associated loci. https://doi.org/10.1371/journal.pgen.1010786.s018 (XLSX) S9 Table. The evidence of the 16 ear-associated lead SNPs or nearby genes in four databases including 3DSNP, VISTA, Harmonizome, and MGI. https://doi.org/10.1371/journal.pgen.1010786.s019 (XLSX) S10 Table. Partitioned heritability enrichments based on cell-type-specific regulatory annotations. https://doi.org/10.1371/journal.pgen.1010786.s020 (XLSX) S11 Table. Effects of sex and genotypes on 28 ear phenotypes in mice. https://doi.org/10.1371/journal.pgen.1010786.s021 (XLSX) S12 Table. Definition of 17 human ear landmarks. https://doi.org/10.1371/journal.pgen.1010786.s022 (XLSX) S13 Table. Definition of 21 ear landmarks in mice. https://doi.org/10.1371/journal.pgen.1010786.s023 (XLSX) S14 Table. Enhancer regions estimated based on the H3K27ac and ATAC-seq using ROSE. https://doi.org/10.1371/journal.pgen.1010786.s024 (XLSX) Acknowledgments The authors thank all sample donors for their contribution to this project. We would like to express our gratitude to Winston Rojas-Montoya for his invaluable support at the Universidad de Antioquia during the difficult period following the unfortunate passing of Gabriel Bedoya.

Adams, Christopher J.;Conery, Mitchell;Auerbach, Benjamin J.;Jensen, Shane T.;Mathieson, Iain;Voight, Benjamin F.

doi: 10.1371/journal.pgen.1010807pmid: 37418489

Introduction Germline mutations are the primary source of genetic variation between and within species. Quantifying where, what type, and how frequently mutations arise is therefore of fundamental importance to population genetic inference and complex trait studies. Better estimates of mutation rates improve tools designed to quantify population divergence times [1], demographic history [2], and the effects of background selection [3]. Moreover, models for the underlying de novo mutation rate from which burden of mutations can be statistically assessed have enabled discovery of genes [4,5] and non-coding sequences [6,7] contributing to complex disease [4,5,8,9]. Our working hypothesis is that there exists an underlying structure to the context-dependent effects that shape the mutation rate. Here, we focus on polymorphism probabilities as a proxy for the mutation rate that we hypothesize share the same context-dependent architecture subject to genetic drift, demography, selection, biased gene conversion, or additional phenomenon that operate across population history. The frequency of polymorphisms varies widely across the genome [10] and correlates with several genomic features [11–13], with new mutations caused by both exogenous and endogenous sources [14]. There is considerable evidence to suggest that local nucleotide context directly relates to the probability that a nucleotide mutates. A classic example of this is the ~14-fold higher rate of C>T transitions at methylated CpG sites, owing to spontaneous deamination of 5-methylcytosine [15–17]. Long tracts of low-complexity DNA show elevated variability in mutation rates, which is hypothesized to be the result of slippage of DNA polymerase during replication [18]. This prior work suggests that local sequence context is integral to understanding variation in polymorphism rates across the genome, and that the most predictive models will be best positioned to guide elucidation of the underlying mutational mechanisms. Our previous work demonstrated that a sequence context window of seven nucleotides (i.e., ‘7-mer’) provided a superior model to explain patterns of genetic variation relative to smaller windows that are commonly used (e.g., 3-mers) [19]. While an advance, this model was fundamentally limited for three reasons: scalability, regularization, and uncertainty. First, the size of the model–which increases by a factor of four for each nucleotide included–presents intrinsic limits both computationally and in terms of statistical power. Second, while it is straightforward to assume that every sequence context is meaningful, a more parsimonious model–informed by biological intuition–might be that only a subset of contexts contributes meaningfully to the observed variation in data. This is particularly important for inference of somatic and de novo mutation rates or in other data-sparse situations (e.g., across species). Finally, while our previous model provided a point estimate of the mean polymorphism probability, it did not immediately emit uncertainty resulting from multinomial variance and heterogeneity in larger sequence contexts. As sequence context sizes are expanded, there is functionally less data and thus more uncertainty in estimates, making point estimates even more unreliable. Quantifying uncertainty is also required for detecting differences in probabilities across models, for example when comparing differences in rates across populations [20–22] or at functional genomic features [23]. Ideally, a method should scale the inferred context length proportional to the amount of data and the biological signal that may be present within that data while providing uncertainty in estimated parameters and underlying probabilities. Previous work has sought to address these challenges, though methods introduced to date do not address all limitations simultaneously. Sparsity and scalability have been tackled through a deep-learning framework [24] as well as an IUPAC-motif-based clustering approach [25] which modeled polymorphism probabilities up through 9-mers. Another method explored polymorphism probabilities up through 7-mers using DNA shape covariates to reduce the parameter space [26]. All three methods are robust and effective at measuring point estimates of polymorphism probabilities in expanded sequence contexts, however none explicitly estimates the uncertainty of these parameters. Finally, the CIPI model [27] is a Bayesian method that addresses these issues, but focuses on applications with smaller context-window motifs (5-mer) in variant settings with fewer mutation events (e.g., somatic mutations in cancer or mutations in viral genomes) and is not obviously scalable computationally to larger size context windows and sizes of contemporary population genomics data sets in humans (e.g., hundreds of millions of polymorphic sites). Here, we develop a method that addresses all three limitations embedded in a novel model. We construct a Bayesian tree-based method that integrates sequence context window size, handles sparse data, and captures uncertainty in estimates of mutation probability via the posterior distribution. We subsequently apply our approach in multiple ways. First, we quantify differences in polymorphism probabilities between continental populations and place bounds on the effect sizes of potential undescribed context-dependent differences in the 1000 Genomes dataset [28]. Second, we explore the use of polymorphism datasets to predict de novo mutations. We measure the effect of population history, variant age, and sequence context size on model performance with the aim of generating a meaningful proxy to estimate the germline mutation rate. Finally, we build models of different great ape species and assess the similarity to human polymorphism models. Description of the method A tree-based sequence-context model captures variation in polymorphism probabilities We began by developing a model to describe the hierarchical relationship of sequence context dependencies over increasing window sizes. We structured this as a rooted, tree-based graph, where each type of substitution class is represented distinctly (Fig 1A). Each level of the tree represents an increasing window size of sequence considered, alternating between incorporating nucleotides to the window on the 3′ end for even-sized contexts and on the 5′ end for odd-sized contexts. We fold over reverse complementary contexts to reduce the parameter count (S1 Text). To ease readability, we denote each mutation with the sequence context, the nucleotide in scope underlined, and the polymorphism indicated with an arrow (e.g. TCC>T represents the polymorphism where the underlined cytosine has become a thymine). Each non-root edge represents the log-transformed, multiplicative shift in polymorphism probability captured by expanding sequence context. The root edge corresponds to an estimated base polymorphism probability for a given mutation type. For a given sequence context, each node in the tree represents the probability of observing a polymorphic site in the central nucleotide (referred to hereafter as polymorphism probability), and is the product of all edges, starting from the root that leads to the node (Fig 1B). As our previous work has shown for a specific level of sequence context, the distribution of observed counts for each sequence context can be modelled via independent multinomial distributions [19] facilitating likelihood calculation. The resulting multinomial probability vector corresponds to the combination of individual polymorphism probability estimates across each mutation type tree for each sequence context (S1 Text). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 1. Hierarchical relationship of sequence contexts and key algorithmic elements of Baymer. (A) Each mutation type is represented by a separate sequence context tree, related by the shared ‘mer’ level parameters and joint multinomial likelihood distribution. Each sequence context tree has a nested structure where information is partially pooled across each shared parent. (B) Polymorphism probabilities are parameterized as the product of the series of edges that lead to the sequence context of interest. (C) Sequence context trees are regularized using a spike-and-slab prior distribution. https://doi.org/10.1371/journal.pgen.1010807.g001 Within the model, we incorporate two features essential for downstream applications when comparing the outputs of competing models. First, we employ a Bayesian formulation which generates posterior distributions for polymorphism probabilities (S1 Text). This approach naturally provides uncertainty around parameter estimates which is essential for comparison of rates across different tabulated models. Second, we incorporate regularization in the parameter estimation procedure for tree edges. Previous sequence context models estimated parameters (ϕ) for all edges of the tree, meaning that all values of were effectively non-zero. However, our previous work suggested that perhaps only a fraction of edges meaningfully contribute information [19]. Hypothesizing that only a subset of edges is informative for estimating mutation probabilities, we regularize our tree model by incorporating a spike-and-slab prior on the ϕ parameters [29]. Our approach estimates the fraction of posterior samples in the slab, implying a non-zero effect on polymorphism probabilities, and in the spike, which implies no effect. Thus, the probability of an edge being included in the slab is the equivalent of the posterior inclusion probability (PIP) for our model. We tune the model such that the slab is favored when the evidence suggests a multinomial probability shift greater than 10% for a given context level (Fig 1C). This value was chosen weighing the stability of model convergence with the goal of inferring the largest possible effects. Because the posterior distribution is not analytically tractable, we implemented an adaptive Metropolis-within-Gibbs Markov Chain Monte Carlo (MCMC) sampling scheme [30] to sample from and thereby estimate the posterior distribution of this model. To further aid in convergence and enforce intermediate nodes to have identifiable mutation probabilities, we estimated parameters of the model level-by-level rather than all simultaneously, leveraging the conditional dependency structure of the hierarchical tree. Under this set-up, the unseen higher-order layers are assigned ϕa,b = 0 edges until their level has been sampled. We embedded this model and sampling scheme into software (named Baymer) for further testing and applications. A tree-based sequence-context model captures variation in polymorphism probabilities We began by developing a model to describe the hierarchical relationship of sequence context dependencies over increasing window sizes. We structured this as a rooted, tree-based graph, where each type of substitution class is represented distinctly (Fig 1A). Each level of the tree represents an increasing window size of sequence considered, alternating between incorporating nucleotides to the window on the 3′ end for even-sized contexts and on the 5′ end for odd-sized contexts. We fold over reverse complementary contexts to reduce the parameter count (S1 Text). To ease readability, we denote each mutation with the sequence context, the nucleotide in scope underlined, and the polymorphism indicated with an arrow (e.g. TCC>T represents the polymorphism where the underlined cytosine has become a thymine). Each non-root edge represents the log-transformed, multiplicative shift in polymorphism probability captured by expanding sequence context. The root edge corresponds to an estimated base polymorphism probability for a given mutation type. For a given sequence context, each node in the tree represents the probability of observing a polymorphic site in the central nucleotide (referred to hereafter as polymorphism probability), and is the product of all edges, starting from the root that leads to the node (Fig 1B). As our previous work has shown for a specific level of sequence context, the distribution of observed counts for each sequence context can be modelled via independent multinomial distributions [19] facilitating likelihood calculation. The resulting multinomial probability vector corresponds to the combination of individual polymorphism probability estimates across each mutation type tree for each sequence context (S1 Text). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 1. Hierarchical relationship of sequence contexts and key algorithmic elements of Baymer. (A) Each mutation type is represented by a separate sequence context tree, related by the shared ‘mer’ level parameters and joint multinomial likelihood distribution. Each sequence context tree has a nested structure where information is partially pooled across each shared parent. (B) Polymorphism probabilities are parameterized as the product of the series of edges that lead to the sequence context of interest. (C) Sequence context trees are regularized using a spike-and-slab prior distribution. https://doi.org/10.1371/journal.pgen.1010807.g001 Within the model, we incorporate two features essential for downstream applications when comparing the outputs of competing models. First, we employ a Bayesian formulation which generates posterior distributions for polymorphism probabilities (S1 Text). This approach naturally provides uncertainty around parameter estimates which is essential for comparison of rates across different tabulated models. Second, we incorporate regularization in the parameter estimation procedure for tree edges. Previous sequence context models estimated parameters (ϕ) for all edges of the tree, meaning that all values of were effectively non-zero. However, our previous work suggested that perhaps only a fraction of edges meaningfully contribute information [19]. Hypothesizing that only a subset of edges is informative for estimating mutation probabilities, we regularize our tree model by incorporating a spike-and-slab prior on the ϕ parameters [29]. Our approach estimates the fraction of posterior samples in the slab, implying a non-zero effect on polymorphism probabilities, and in the spike, which implies no effect. Thus, the probability of an edge being included in the slab is the equivalent of the posterior inclusion probability (PIP) for our model. We tune the model such that the slab is favored when the evidence suggests a multinomial probability shift greater than 10% for a given context level (Fig 1C). This value was chosen weighing the stability of model convergence with the goal of inferring the largest possible effects. Because the posterior distribution is not analytically tractable, we implemented an adaptive Metropolis-within-Gibbs Markov Chain Monte Carlo (MCMC) sampling scheme [30] to sample from and thereby estimate the posterior distribution of this model. To further aid in convergence and enforce intermediate nodes to have identifiable mutation probabilities, we estimated parameters of the model level-by-level rather than all simultaneously, leveraging the conditional dependency structure of the hierarchical tree. Under this set-up, the unseen higher-order layers are assigned ϕa,b = 0 edges until their level has been sampled. We embedded this model and sampling scheme into software (named Baymer) for further testing and applications. Verification and comparison Asymmetric context expansion improves parsimony and model inference The hierarchical tree-based Baymer graphs are constructed such that the difference in length of flanking nucleotides on either side of the focal nucleotide is zero for odd-length contexts, and one for even-length contexts. It follows that these trees can be constructed in three ways: expanding sequence contexts by including even-length contexts (termed “asymmetric models”)–alternating expansions starting at either the (1) left, or 5’, end, or (2) starting with the right, or 3’, end–or (3) expanding by exclusively using odd-length contexts (termed “symmetric models”; Fig 2A). When expanding an odd-length context by two nucleotides (e.g., 1-mer to 3-mer), symmetric models require 16 edges per context as compared to 20 edges per context (16 edges + 4 intermediate edges) in the asymmetric model. Despite more total edges in the asymmetric model tree architecture, we hypothesized the intermediate edges would more efficiently capture signal and provide greater resolution to detect inflection points in the tree where the local sequence context results in mutability changes. We implemented and tested all three models on the same dataset. First, we observed that for our most challenging model (9-mers), rate estimates are very strongly correlated amongst all models (Fig 2B and 2C). Furthermore, we observed that despite more total edges in the tree-graph, asymmetric models include approximately 38% fewer overall edges with high confidence (Fig 2D), suggesting greater parsimony. Finally, asymmetric models produced models that better fit holdout data than the symmetric models (S1 Text and Fig 2E). This improvement arises specifically in situations where there is sufficient data to estimate 8-mer edges, but insufficient data to confidently estimate 9-mer rates. Given our folding scheme, we opted for the biologically pragmatic choice of an even-length context tree architectural model that initiates alternating context expansions with the right end (3’), as this captures CpG effect(s) as early as possible in the tree without distributing the effect across more than one edge. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 2. Exploration of different strategies to build out the Baymer sequence context hierarchical trees using gnomAD non-Finnish European (NFE) polymorphisms with derived allele count greater than or equal to two in non-coding accessible regions. (A) Sequence contexts can be built by starting to alternate adding nucleotides with the left, or 5’, end of each odd-length context (blue path), with the right, or 3’, end of each context (red path), or by only including odd-length contexts and adding a nucleotide to both sides of the growing context (black path) for each expansion. (B) Baymer mean posterior estimates of 9-mer polymorphism probabilities estimated using even base pair data with the right alternation pattern and the left alternation pattern (Spearman correlation = 0.997; p < 10−100). (C) Baymer mean posterior estimates of 9-mer polymorphism probabilities estimated using even base pair data with the right alternation pattern and a model exclusively using odd-length sequence contexts (Spearman correlation = 0.998; p < 10−100). (D) Absolute count of edges in each tree architecture model with a PIP > 0.95 in a 9-mer model. (E) Multinomial likelihoods for each model are calculated on odd base pair NFE test data using 7-mer and 9-mer models. Polymorphism probability estimates were linearly scaled to match the mean polymorphism probability of the holdout dataset. https://doi.org/10.1371/journal.pgen.1010807.g002 Evaluation of the model demonstrates robust inference of the underlying rates with uncertainty A key feature of Baymer is that it estimates posterior distributions for each parameter, allowing for uncertainty in the probabilities of polymorphism at each sequence context. To evaluate the coverage of the estimated posterior probabilities, we used simulations to assess how often our posterior distribution captures simulated values. Using a pre-specified polymorphism probability table, we tested how frequently polymorphism probabilities estimated by Baymer captured the true value for each sequence context (S1 Text). We found that across all sequence context sizes, 89%, 93%, and 97% of context simulations contained the true polymorphism probability in the 90%, 95%, and 99% credible intervals, respectively (S1 Text and S1 Table). A second important feature is that regularization is embedded into the method, allowing for the creation of parsimonious models that capture most of the information with the fewest non-zero parameters. This part is critical to address cases where the amount of data is not large and limits power, or when considering larger windows of sequence context that are rare and/or uninformative. If robustly calibrated, we would expect probabilities inferred in a holdout set to strongly correlate with those estimated during a test phase (i.e., minimal overfitting). To evaluate the robustness of the inferred rates, we partitioned the human genome reference into two sets–even and odd base pairs–and used SNPs of allele count 2 or greater observed in the gnomAD [31] non-Finnish European (NFE) collection to independently train models (S1 Text). We compared the concordance of probabilities for models with sequence context windows up to 4 flanking nucleotides on either side (i.e., a 9-mer model) using the maximum likelihood estimate approach [19] and Baymer (S1 Fig). For each comparison, in addition to the Spearman correlation, we also calculated the root mean squared perpendicular error (RMSPE) from each point to the x-y axis, as a measure of the tightness of the distribution from the true, shared value (S1 Text). The maximum likelihood estimates of polymorphism probabilities (Fig 3A, Spearman correlation ρ = 0.915; RMSPE = 0.117) were less correlated and considerably less tightly distributed than those for Baymer-derived models (Fig 3B, ρ = 0.990; RMSPE = 0.035). This result occurred even after omitting ~16,000 sequence contexts with zero mutations in either dataset (odd and even base pairs) from the maximum likelihood model comparison, rendering practical use of large swaths of the model useless due to substantial overfitting at the 9-mer level. If zero-mutation contexts omitted from the maximum likelihood model were included, the correlations would perform considerably worse (S1 Text and S1D Fig, ρ = 0.876; RMSPE = 0.744), as these polymorphism probabilities are exclusively determined by pseudo counts. Within the NFE dataset, Baymer inferences were also robust across allele frequency bins (S2 Fig). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 3. Baymer model validation, transferability, and regularization in gnomAD non-Finnish European (NFE) polymorphisms with derived allele count greater than or equal to two in non-coding accessible regions. (A) Empirical 9-mer polymorphism probabilities for context mutations with at least one occurrence in both datasets (15,910 omitted context mutations) are plotted against one another (Spearman correlation = 0.915; p < 10−100; RMSPE = 0.12). (B) Baymer mean posterior estimates for 9-mer polymorphism estimates in even and odd base pair datasets (Spearman correlation = 0.990; p < 10−100; RMSPE = 0.035). (C) Baymer mean posterior estimates for 9-mer polymorphism estimates in odd base pair non-Finnish European gnomAD data and even base pair NYGC 1KG phase three data, down-sampled to match total number of polymorphisms and site frequency spectrum (Spearman correlation = 0.984; p < 10−100; RMSPE = 0.045). (D) Fraction of edges in the NFE model with a PIP > 0.95 in each sequence context window layer. Absolute count of edges above bars. (E) For high-data contexts with at least 100,000 total instances in the non-coding genome and 50 total mutations, fraction of edges at each sequence context window size across PIP bins. (F) Proportion of high-data contexts within each mutation type at each sequence context window size with PIP>0.95. https://doi.org/10.1371/journal.pgen.1010807.g003 We next sought to evaluate the transferability of inferred models between experimental collections; while internally consistent, the above procedure could simply reflect data set specific biases [32]. For this, we compared non-admixed, non-Finnish European (EUR) samples obtained from the 1000 Genomes (1KG) Project (re-sequenced by the New York Genome Center) [33] with the gnomAD NFE sample described above. As before, we split the data into even and odd base pairs but also applied a variant down-sampling procedure to match total variant count and site-frequency spectrum between both sets (S1 Text). By comparing variants found in the even base pair genome of gnomAD with the odd base pair genome of 1KG, this strategy ensures no variation overlapped between data sets. We observed that the probabilities estimated from both sample sets were strongly correlated (ρ = 0.984; RMSPE = 0.045; Fig 3C) though were slightly weaker than the correlations from each internal comparison and fit less tightly (gnomAD ρ = 0.990; RMSPE = 0.035; Fig 3B; 1KG ρ = 0.986; RMSPE = 0.042; S3 Fig). This result demonstrates that while some additional between-sample variation may exist, Baymer infers probabilities of polymorphism that are broadly consistent with one another, supporting the notion of model transferability across different data sets. We next aimed to quantify how well the model selects meaningful context features. We expected more proximal bases to the focal site to have a greater impact on polymorphism probabilities for two reasons, (i) due to data richness, and (ii) that proximity to the polymorphic site would suggest more direct impacts on mutability, e.g., the CpG context. Consistent with expectation, the fraction of edges with a PIP > 0.95 monotonically decreases as the sequence context size is increased (Fig 3D). For any given 9-mer context, we find a median of 3 edges included with high confidence in the model (S4A Fig). The median window of context-dependence for each 9-mer was 5 base pairs wide, although this inference is limited by the sparsity of the model (S4B Fig). Larger contexts best explain patterns of variation genome-wide We note that over 61% of all edges with a PIP > 0.95 are found in the 8-mer and 9-mer levels of our model of polymorphism observed in the gnomAD NFE data. While fewer than 2% of 9-mer edges meaningfully impact the final estimates, they still account for the most total absolute edges (7189 total edges > 0.95 PIP) and are enriched for larger effect sizes (S4C Fig). This observation holds even after filters for data sparsity (S1 Text and Fig 3E). This implies a considerable impact on polymorphism probabilities in extended sequence contexts, consistent with previous work [19,23–25]. This general trend is similarly consistent across mutation types (Fig 3F), although with a variable degree of impact, most notably with less additional variability estimated in wider CpG>T edges (S4D Fig). We thus evaluated the overall improvement in likelihood by expanding window sizes up to 9-mers. Compared to lower context models (e.g., 3-mer, 5-mer, or 7-mer) on holdout data, 9-mer Baymer models substantially improved the likelihood and best fit to the data (S1 Text and S2 Table). Asymmetric context expansion improves parsimony and model inference The hierarchical tree-based Baymer graphs are constructed such that the difference in length of flanking nucleotides on either side of the focal nucleotide is zero for odd-length contexts, and one for even-length contexts. It follows that these trees can be constructed in three ways: expanding sequence contexts by including even-length contexts (termed “asymmetric models”)–alternating expansions starting at either the (1) left, or 5’, end, or (2) starting with the right, or 3’, end–or (3) expanding by exclusively using odd-length contexts (termed “symmetric models”; Fig 2A). When expanding an odd-length context by two nucleotides (e.g., 1-mer to 3-mer), symmetric models require 16 edges per context as compared to 20 edges per context (16 edges + 4 intermediate edges) in the asymmetric model. Despite more total edges in the asymmetric model tree architecture, we hypothesized the intermediate edges would more efficiently capture signal and provide greater resolution to detect inflection points in the tree where the local sequence context results in mutability changes. We implemented and tested all three models on the same dataset. First, we observed that for our most challenging model (9-mers), rate estimates are very strongly correlated amongst all models (Fig 2B and 2C). Furthermore, we observed that despite more total edges in the tree-graph, asymmetric models include approximately 38% fewer overall edges with high confidence (Fig 2D), suggesting greater parsimony. Finally, asymmetric models produced models that better fit holdout data than the symmetric models (S1 Text and Fig 2E). This improvement arises specifically in situations where there is sufficient data to estimate 8-mer edges, but insufficient data to confidently estimate 9-mer rates. Given our folding scheme, we opted for the biologically pragmatic choice of an even-length context tree architectural model that initiates alternating context expansions with the right end (3’), as this captures CpG effect(s) as early as possible in the tree without distributing the effect across more than one edge. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 2. Exploration of different strategies to build out the Baymer sequence context hierarchical trees using gnomAD non-Finnish European (NFE) polymorphisms with derived allele count greater than or equal to two in non-coding accessible regions. (A) Sequence contexts can be built by starting to alternate adding nucleotides with the left, or 5’, end of each odd-length context (blue path), with the right, or 3’, end of each context (red path), or by only including odd-length contexts and adding a nucleotide to both sides of the growing context (black path) for each expansion. (B) Baymer mean posterior estimates of 9-mer polymorphism probabilities estimated using even base pair data with the right alternation pattern and the left alternation pattern (Spearman correlation = 0.997; p < 10−100). (C) Baymer mean posterior estimates of 9-mer polymorphism probabilities estimated using even base pair data with the right alternation pattern and a model exclusively using odd-length sequence contexts (Spearman correlation = 0.998; p < 10−100). (D) Absolute count of edges in each tree architecture model with a PIP > 0.95 in a 9-mer model. (E) Multinomial likelihoods for each model are calculated on odd base pair NFE test data using 7-mer and 9-mer models. Polymorphism probability estimates were linearly scaled to match the mean polymorphism probability of the holdout dataset. https://doi.org/10.1371/journal.pgen.1010807.g002 Evaluation of the model demonstrates robust inference of the underlying rates with uncertainty A key feature of Baymer is that it estimates posterior distributions for each parameter, allowing for uncertainty in the probabilities of polymorphism at each sequence context. To evaluate the coverage of the estimated posterior probabilities, we used simulations to assess how often our posterior distribution captures simulated values. Using a pre-specified polymorphism probability table, we tested how frequently polymorphism probabilities estimated by Baymer captured the true value for each sequence context (S1 Text). We found that across all sequence context sizes, 89%, 93%, and 97% of context simulations contained the true polymorphism probability in the 90%, 95%, and 99% credible intervals, respectively (S1 Text and S1 Table). A second important feature is that regularization is embedded into the method, allowing for the creation of parsimonious models that capture most of the information with the fewest non-zero parameters. This part is critical to address cases where the amount of data is not large and limits power, or when considering larger windows of sequence context that are rare and/or uninformative. If robustly calibrated, we would expect probabilities inferred in a holdout set to strongly correlate with those estimated during a test phase (i.e., minimal overfitting). To evaluate the robustness of the inferred rates, we partitioned the human genome reference into two sets–even and odd base pairs–and used SNPs of allele count 2 or greater observed in the gnomAD [31] non-Finnish European (NFE) collection to independently train models (S1 Text). We compared the concordance of probabilities for models with sequence context windows up to 4 flanking nucleotides on either side (i.e., a 9-mer model) using the maximum likelihood estimate approach [19] and Baymer (S1 Fig). For each comparison, in addition to the Spearman correlation, we also calculated the root mean squared perpendicular error (RMSPE) from each point to the x-y axis, as a measure of the tightness of the distribution from the true, shared value (S1 Text). The maximum likelihood estimates of polymorphism probabilities (Fig 3A, Spearman correlation ρ = 0.915; RMSPE = 0.117) were less correlated and considerably less tightly distributed than those for Baymer-derived models (Fig 3B, ρ = 0.990; RMSPE = 0.035). This result occurred even after omitting ~16,000 sequence contexts with zero mutations in either dataset (odd and even base pairs) from the maximum likelihood model comparison, rendering practical use of large swaths of the model useless due to substantial overfitting at the 9-mer level. If zero-mutation contexts omitted from the maximum likelihood model were included, the correlations would perform considerably worse (S1 Text and S1D Fig, ρ = 0.876; RMSPE = 0.744), as these polymorphism probabilities are exclusively determined by pseudo counts. Within the NFE dataset, Baymer inferences were also robust across allele frequency bins (S2 Fig). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 3. Baymer model validation, transferability, and regularization in gnomAD non-Finnish European (NFE) polymorphisms with derived allele count greater than or equal to two in non-coding accessible regions. (A) Empirical 9-mer polymorphism probabilities for context mutations with at least one occurrence in both datasets (15,910 omitted context mutations) are plotted against one another (Spearman correlation = 0.915; p < 10−100; RMSPE = 0.12). (B) Baymer mean posterior estimates for 9-mer polymorphism estimates in even and odd base pair datasets (Spearman correlation = 0.990; p < 10−100; RMSPE = 0.035). (C) Baymer mean posterior estimates for 9-mer polymorphism estimates in odd base pair non-Finnish European gnomAD data and even base pair NYGC 1KG phase three data, down-sampled to match total number of polymorphisms and site frequency spectrum (Spearman correlation = 0.984; p < 10−100; RMSPE = 0.045). (D) Fraction of edges in the NFE model with a PIP > 0.95 in each sequence context window layer. Absolute count of edges above bars. (E) For high-data contexts with at least 100,000 total instances in the non-coding genome and 50 total mutations, fraction of edges at each sequence context window size across PIP bins. (F) Proportion of high-data contexts within each mutation type at each sequence context window size with PIP>0.95. https://doi.org/10.1371/journal.pgen.1010807.g003 We next sought to evaluate the transferability of inferred models between experimental collections; while internally consistent, the above procedure could simply reflect data set specific biases [32]. For this, we compared non-admixed, non-Finnish European (EUR) samples obtained from the 1000 Genomes (1KG) Project (re-sequenced by the New York Genome Center) [33] with the gnomAD NFE sample described above. As before, we split the data into even and odd base pairs but also applied a variant down-sampling procedure to match total variant count and site-frequency spectrum between both sets (S1 Text). By comparing variants found in the even base pair genome of gnomAD with the odd base pair genome of 1KG, this strategy ensures no variation overlapped between data sets. We observed that the probabilities estimated from both sample sets were strongly correlated (ρ = 0.984; RMSPE = 0.045; Fig 3C) though were slightly weaker than the correlations from each internal comparison and fit less tightly (gnomAD ρ = 0.990; RMSPE = 0.035; Fig 3B; 1KG ρ = 0.986; RMSPE = 0.042; S3 Fig). This result demonstrates that while some additional between-sample variation may exist, Baymer infers probabilities of polymorphism that are broadly consistent with one another, supporting the notion of model transferability across different data sets. We next aimed to quantify how well the model selects meaningful context features. We expected more proximal bases to the focal site to have a greater impact on polymorphism probabilities for two reasons, (i) due to data richness, and (ii) that proximity to the polymorphic site would suggest more direct impacts on mutability, e.g., the CpG context. Consistent with expectation, the fraction of edges with a PIP > 0.95 monotonically decreases as the sequence context size is increased (Fig 3D). For any given 9-mer context, we find a median of 3 edges included with high confidence in the model (S4A Fig). The median window of context-dependence for each 9-mer was 5 base pairs wide, although this inference is limited by the sparsity of the model (S4B Fig). Larger contexts best explain patterns of variation genome-wide We note that over 61% of all edges with a PIP > 0.95 are found in the 8-mer and 9-mer levels of our model of polymorphism observed in the gnomAD NFE data. While fewer than 2% of 9-mer edges meaningfully impact the final estimates, they still account for the most total absolute edges (7189 total edges > 0.95 PIP) and are enriched for larger effect sizes (S4C Fig). This observation holds even after filters for data sparsity (S1 Text and Fig 3E). This implies a considerable impact on polymorphism probabilities in extended sequence contexts, consistent with previous work [19,23–25]. This general trend is similarly consistent across mutation types (Fig 3F), although with a variable degree of impact, most notably with less additional variability estimated in wider CpG>T edges (S4D Fig). We thus evaluated the overall improvement in likelihood by expanding window sizes up to 9-mers. Compared to lower context models (e.g., 3-mer, 5-mer, or 7-mer) on holdout data, 9-mer Baymer models substantially improved the likelihood and best fit to the data (S1 Text and S2 Table). Applications Sequence context motifs are correlated with changes in polymorphism probability We next aimed to identify inflection points in the Baymer trees by examining the edges corresponding to the largest ϕs across each layer. Unsurprisingly, the CpG>T edge had the largest mean posterior ϕ magnitude (S3 Table). Consistent with our previous results [19], edges with the largest absolute mean posterior ϕ are largely localized at the intersection of poly-A repeat-rich sequences (lower rates of A>T substitutions) but particularly presented in 8-mer context by poly-A tract of length 4, where the mutation type extends one of the repeated patterns (e.g., CGCGAGAGA>C or CCCAAAA>C), and the CAATN motif which increases A>G mutability (2.35–2.99x increase, S3 Table). Next, we attempted to discover specific motifs that are enriched in the highest or lowest 1% of 9-mer polymorphism probabilities within each mutation type (S1 Text and S4 Table). We recapitulate almost all previously reported motifs [19,23,25]. Consistent with previous reports, we identify a preponderance of repeat-rich motifs, which is perhaps due to the impact of slippage in introducing mutations [18]. We discover numerous motifs with flanks extending 4 base pairs from the focal nucleotide that showed enrichment (21 motifs with p < 0.0001; S4 Table), emphasizing the utility of expanded sequence context windows for modeling mutability. Frequency of polymorphism across populations do not differ substantially across levels of sequence context Prior work has centered around evaluating whether mutation rates have changed over evolutionary time by evaluating differences in the proportions of sequence-context-dependent polymorphism between human populations [21,22,34–36]. To determine whether polymorphism probabilities differ across human populations, we analyzed individuals from the NYGC resequencing of 1000 Genomes Project (1KG) Phase III representing continental European, African, East Asian, and South Asian groups. We extracted variants private to these continental groups, down-sampling to match site-frequency spectra bins and overall sample sizes (S1 Text). We then applied Baymer to each individual dataset to model probabilities up to a 9-mer window of sequence context. We compared estimates of polymorphism probabilities in each population by assessing the degree to which the posterior distribution of each population’s model parameters overlapped. The fraction overlap of each distribution is a proxy for the probability that the underlying parameters are the same. Due to the implicit tree structure of sequence context models, polymorphism probability shifts in edges will affect all edges downstream of the context in question. Therefore, we identified contexts where both the estimated polymorphism probability and the immediate edge leading to that context were both considered very likely to be different between populations. Specifically, we identified contexts whose posterior estimates of polymorphism probabilities and edges both overlapped less than 1% in pairwise comparisons between the four populations (S5 Table). This included all the most notable previously reported 3-mer shifts across continental groups, including the increase in TCC>T mutations found in European relative to non-European ancestry populations [20–22,34,35]. We next focused on the remainder of 3-mer and wider extended sequence contexts (Table 1). While a handful of such sequence contexts have been implicated [34], these results are confounded by batch effects in the original 1KG sequencing data [37]. Since the data we use for our analysis is derived from the New York Genome Center resequencing project [33], we do not expect the same confounder. In our results, we observed the presence of nucleotide repeats, e.g., TA / CG dinucleotides; poly-C / poly-A in several of the divergent contexts, which could be explained by polymerase slippage [18]. Download: PPT PowerPoint slide PNG larger image TIFF original image Table 1. Baymer modeled 1KG private continental context mutations with extreme polymorphism probability differences. https://doi.org/10.1371/journal.pgen.1010807.t001 While the population-specific polymorphism probabilities estimated and polymorphism counts are identical between each pairwise comparison and thus correlated, we still note that 15/28 pairwise differences are specific to a single continental group. Of these, only the two canonical European context mutation differences (TCC>T and TCT>T) are in 3-mer contexts, otherwise all are found in 5-mer and greater window sizes. In South Asian samples, we find that the mean CTATA>T polymorphism probabilities are approximately 1.6 times higher than the remaining populations and in Africans TATATATC>G is approximately 1.9 times higher. The largest population-specific effect was discovered in East Asians where ATACCTC>A polymorphism probabilities are roughly 2.7 times higher than in European, African, or South Asian models. None of these effects have been explicitly documented before. Taken collectively, we observed relatively few instances of edges that were quantifiably different across continental groups, and those that were observed were largely confined to relatively small windows of context where we might have anticipated well-powered tests (e.g., 3- and 5-mers). To quantify the power of our specific analytic procedure for discovery and the sample size necessary to identify true differences in polymorphism probabilities, we performed simulations where true effect differences were ‘spiked-in’ between two populations over a range of weak to stronger effects and across a sampling of different sequence contexts (S1 Text). Differences in mutability between populations for this experiment are defined as the natural log of the polymorphism probabilities ratio (NLPPR) between each simulated population. This allowed us to construct credible sets of effects that we were reasonably well powered (>80%) to discover (Table 2). Unsurprisingly, the power scaled proportional to the number of context instances, simulated mutations in the dataset, and the size of the spiked-in differences (S5 Fig). Notably, extremely subtle shifts (NLPPR < = 0.01; 0.99–1.01-fold change) were not detectable at any sequence context size. On the opposite side of the spectrum, we found that we were reasonably powered to identify shift differences where NLPPR > 1.0 (fold decrease ≤ 0.37 or fold increase ≥ 2.72) up through 5-mers and in 6-mers with large sample sizes. For reference, the TCC>T polymorphism has an NLPPR = 0.291 (~1.34 fold increase)–the largest difference of any 3-mer by our calculation. Download: PPT PowerPoint slide PNG larger image TIFF original image Table 2. Power estimates for 1KG continental private polymorphism probabilities. https://doi.org/10.1371/journal.pgen.1010807.t002 In contrast, our experiment had essentially no power to discover 9-mer polymorphism probability changes and extremely limited power for 8-mers, even for large differences. Thus, there may exist large differences at these sizes that we could not reliably capture. These results are consistent with our comparisons in the real data (Table 1), as only differences within the detectable range at each mer-level were implicated. These power calculations suggest that, given the experiment we performed grouping all mutations together (agnostic to allele frequency or age, see Discussion), if any 3-mer differences greater than the TCC>T shift exist, we would have discovered these effects for a broad range of modest to very strong effects across a range of sequence contexts window sizes. This effectively sets bounds on the differences possible for this analysis scheme in this data. A sequence context model that captures variability in de novo mutational rates Given its formulation in handling data sparsity, we next sought to apply Baymer to develop a model that best captures rates of de novo mutations across the genome. We took advantage of a recent collection of 2,976 WGS Icelandic trios that identified 200,435 de novo events[38] and, analogous to the above, we partitioned de novo variants into even (for training) and odd (for testing) base pairs. We observed substantial improvement in the overall likelihood in the testing set for 5-mer size windows compared to 3-mers (3-mer vs 5-mer, ΔLL = 2,144), but only minimal improvement for increasing windows sizes further (5-mer vs 9-mer, ΔLL = 265, S6 Table). Indeed, Baymer did not select any sequence context feature beyond the 5-mer level with PIP > 0.95. This is not unexpected given our approach to regularization, as the number of events in larger sequence contexts is increasingly sparse, it is desirable to only include informative contexts to avoid overfitting. We next used Baymer to improve upon this baseline model. Previous work has demonstrated that inference of de novo mutational probabilities can be captured via rare variant polymorphism data obtained from population sets as a proxy [23]. We hypothesized that a partitioned set of polymorphism data based on: (i) larger sample sizes that (ii) closely matched the ancestry of the de novo set and (iii) focused on rare variants as a proxy to capture the most recent mutation events would generate the most transferrable model and robust rate estimates. To build variant partitions, we used variant call set data from gnomAD, focused on either a population-matched proxy (i.e., NFE, the non-Finnish European subset) or variant calls from all samples in gnomAD regardless of ancestry (i.e., ALL). For each of these, we created three partitions focused (i) exclusively on variants with one allele count (i.e., singletons; for NFE labeled NFE-1), (ii) exclusively on variants with two allele counts (i.e., doubletons; for NFE labeled NFE-2), and (iii) variants with allele count of two or greater (for NFE labeled NFE-2+). Beyond this, we also identified a set of putatively derived substitutions in the human lineage by comparing the GRCh38 human reference genome with ancestral sequences obtained from primates[39]. We applied Baymer to each variant set independently, comparing the likelihoods of each model to explain rates of de novo mutation in the test set after downscaling probabilities proportional to the sample size. First, we observed that for 3-mer sequence context models, the set of variants obtained from the de novo training set outperformed all other models despite 102 to 1,377 times fewer variants contributing to them than the polymorphism datasets (Fig 4A and S6 Table). In contrast, for larger windows of context (i.e., 7-mer and 9-mer), several of the polymorphism partitions explained the data better than one trained directly from de novo events. This result indicates that increased sample size is required to detect meaningful shifts in polymorphism probabilities in larger sequence context windows. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 4. Modeling de novo mutation probabilities using polymorphism datasets. Even base pair Halldorsson et al. de novo training data modeled by Baymer is compared to Baymer-modelled polymorphism datasets partitioned by allele count. (A) Multinomial likelihoods for each model are calculated on odd de novo base pair test data at various sequence context sizes. Polymorphism probability estimates were linearly scaled to match the mean polymorphism probability of the holdout dataset. (B) Polymorphism datasets were down-sampled to match the size of the even base pair de novo data (70,364 variants) and multinomial likelihoods were calculated on odd de novo base pair data. Each dataset was down-sampled using 5 different random seeds. The Log Likelihood of the 9-mer de novo training model is indicated with the blue dotted line. https://doi.org/10.1371/journal.pgen.1010807.g004 Despite evidence to suggest singleton datasets should best recapitulate de novo variation [4,23,31], we observed that models that trained exclusively on singletons and ALL-2 performed considerably worse than the rest across all windows of sequence context (Fig 4A and S6 Table). While our prior intuition that larger numbers of variants would have provided better rate estimates from increased power deeper in the context tree, rate models exclusively estimated with singletons suffer the most from the impact of recurrent mutations [20,40], especially at CpG sites, which include the highest polymorphism probability mutation type (CpG>T) (S6 Table). Alternatively, population concordance between training and test and/or the quality of variant calls used in training the model could also impact performance. As such, we next sought to explore the effect of noise in low allele count variants. Although we only used variants that passed gnomAD quality control checks, this filter still included a large proportion of variants with a negative log-odds ratio of being a true variant (AS_VQSLOD < 0; S6 Fig). This pattern was also evident for other variant allele counts but were most striking in singletons and the ALL-2 variant groups. Stricter quality filters (AS_VQSLOD > 5–10) considerably improved model performance, but still did not surpass the de novo training model at the 3-mer level (S6 Table). Our NFE singleton Baymer model trained on the strictest quality filter tested (AS_VQSLOD > 10) nearly equaled our best performing model, NFE-2+, with ~ 1/30th the number of variants, but came up just short. In summary, we observed that training from a population matched sample which excluded singletons, NFE-2+, best predicted rates of de novo mutations in 5-mer or larger contexts, better than models trained on de novo events directly. Next, we sought to determine which sample set best modelled the de novo test set adjusting for the total number of variants within the partition. To control sample size differences, we downsampled each partition to match the number of variants observed in the de novo training set (n = 70,364) five times. After down-sampling and when considering 9-mer context models, we observed that the partitions which included NFE exclusively (noted in green, Fig 4B) performed on average better than using the entirety of gnomAD, “ALL” (noted in orange in Fig 4B), which included a more diverse panel of individuals within Europe (e.g., Finnish) but also beyond Europe (e.g., East and South Asian, African and African American). This is consistent with prior belief that, after controlling for the total sample size, variants that derive from samples where ancestries more closely match are the most informative. A grafted tree approach provides superior estimates of de novo mutational probabilities Given the observations that de novo models only outperform polymorphism-based models when either small sequence contexts are used (Fig 4A) or the sample size is controlled (Fig 4B), we next sought to explore a transfer learning-inspired [41] strategy to improve upon our model performance. Transfer learning has previously been employed in a sequence context modelling setting [24]. We hypothesized that regularization means that de novo models have reduced performance with expanded sequence contexts due to low sample sizes. Indeed, our de novo model did not have the power necessary to confidently (PIP > 0.95) include any non-zero shifts in sequence contexts larger than 5-mers in the model (Fig 5A). The larger polymorphism datasets, however, were well-powered to detect shifts in every level of the tree (Fig 5A). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 5. Tree grafting strategy to share information between Baymer models. (A) For each de novo proxy model, we calculated the fraction of context polymorphism probability edges with a PIP > 0.95 in 2-mer through 9 mer-levels as a proxy for the degree of regularization in each model. (B) Edges in the de novo training model that are included with high-confidence (PIP>0.95) are very similar in magnitude and direction to their equivalents in the best-performing proxy model, NFE-2+, in 2-mer through 9-mer levels, implying a shared polymorphism probability shift structure. (C) Proposed tree-grafting schema for modeling de novo mutations that leverages mer-levels where de novo data is plentiful (1-mer through 3-mers) and uses polymorphism data to model the remainder of each model in larger mer-levels (4-mer through 9-mers) where the de novo model is underpowered. (D) The grafted tree method outperforms the previously best-performing model, NFE-2+. https://doi.org/10.1371/journal.pgen.1010807.g005 The nested tree structure of our polymorphism probability models provides a natural strategy where specific branches of the estimated trees can be interchanged, i.e., a “grafted” tree. We asked how similar estimates for edges in expanded sequence contexts are between our de novo model and the best-performing polymorphism model, NFE-2+. In edges in 2-mer and greater levels where the de novo training model is powered enough to detect shifts (PIP > 0.95), the mean posterior estimates of shifts are highly correlated (Fig 5B). This suggests a grafted tree approach is feasible, leveraging the polymorphism datasets for those edges the de novo model is incapable of estimating properly due to sparsity (Fig 5C). Therefore, we built a grafted tree model using 1- to 3-mer edges estimated in the de novo training data model, and 4- to 9-mer edges estimated using the NFE-2+ data model. The resulting combined model had a greater fit to the holdout de novo data than either the NFE-2+ model or de novo model alone (Fig 5D and S1 Text). Sequence context-dependent mutability is shared between closely related great ape species Finally, we examined how well human polymorphism models could capture variability in polymorphism levels observed in populations of great ape. Using polymorphism data from the Great Ape Genome Project [42], we built Baymer models of Pan troglodytes and Gorilla gorilla (S1 Text). We note broad agreement in estimated polymorphism probabilities between humans and chimpanzees (S7A Fig, Spearman correlation ρ = 0.950; RMSPE = 0.089) or gorillas (S7B Fig, Spearman correlation ρ = 0.942; RMSPE = 0.103). These results indicate that the rates of polymorphism at higher orders of sequences contexts are similar across closely related great ape species. As we were especially interested in how human polymorphism models compared with chimpanzee and gorilla models in predicting holdout data in each respective species, we then tested models on odd base pair data in each species, training models using even base pair data for the species in focus (S1 Text). For both chimpanzee (S7C Fig) and gorilla (S7D Fig) tests, species-matched 9-mer models outperformed all other models. While human-derived models are outperformed at the 9-mer level, it is notable that human 9-mer models are more likely than chimpanzee 7-mer models against chimpanzee data and gorilla 5-mer models against gorilla data (S7 Table). Taken collectively, these results suggest the rates of polymorphism at higher orders of sequences contexts are similar across closely related great ape species, with within-species models best capturing variability in observed polymorphism levels. Sequence context motifs are correlated with changes in polymorphism probability We next aimed to identify inflection points in the Baymer trees by examining the edges corresponding to the largest ϕs across each layer. Unsurprisingly, the CpG>T edge had the largest mean posterior ϕ magnitude (S3 Table). Consistent with our previous results [19], edges with the largest absolute mean posterior ϕ are largely localized at the intersection of poly-A repeat-rich sequences (lower rates of A>T substitutions) but particularly presented in 8-mer context by poly-A tract of length 4, where the mutation type extends one of the repeated patterns (e.g., CGCGAGAGA>C or CCCAAAA>C), and the CAATN motif which increases A>G mutability (2.35–2.99x increase, S3 Table). Next, we attempted to discover specific motifs that are enriched in the highest or lowest 1% of 9-mer polymorphism probabilities within each mutation type (S1 Text and S4 Table). We recapitulate almost all previously reported motifs [19,23,25]. Consistent with previous reports, we identify a preponderance of repeat-rich motifs, which is perhaps due to the impact of slippage in introducing mutations [18]. We discover numerous motifs with flanks extending 4 base pairs from the focal nucleotide that showed enrichment (21 motifs with p < 0.0001; S4 Table), emphasizing the utility of expanded sequence context windows for modeling mutability. Frequency of polymorphism across populations do not differ substantially across levels of sequence context Prior work has centered around evaluating whether mutation rates have changed over evolutionary time by evaluating differences in the proportions of sequence-context-dependent polymorphism between human populations [21,22,34–36]. To determine whether polymorphism probabilities differ across human populations, we analyzed individuals from the NYGC resequencing of 1000 Genomes Project (1KG) Phase III representing continental European, African, East Asian, and South Asian groups. We extracted variants private to these continental groups, down-sampling to match site-frequency spectra bins and overall sample sizes (S1 Text). We then applied Baymer to each individual dataset to model probabilities up to a 9-mer window of sequence context. We compared estimates of polymorphism probabilities in each population by assessing the degree to which the posterior distribution of each population’s model parameters overlapped. The fraction overlap of each distribution is a proxy for the probability that the underlying parameters are the same. Due to the implicit tree structure of sequence context models, polymorphism probability shifts in edges will affect all edges downstream of the context in question. Therefore, we identified contexts where both the estimated polymorphism probability and the immediate edge leading to that context were both considered very likely to be different between populations. Specifically, we identified contexts whose posterior estimates of polymorphism probabilities and edges both overlapped less than 1% in pairwise comparisons between the four populations (S5 Table). This included all the most notable previously reported 3-mer shifts across continental groups, including the increase in TCC>T mutations found in European relative to non-European ancestry populations [20–22,34,35]. We next focused on the remainder of 3-mer and wider extended sequence contexts (Table 1). While a handful of such sequence contexts have been implicated [34], these results are confounded by batch effects in the original 1KG sequencing data [37]. Since the data we use for our analysis is derived from the New York Genome Center resequencing project [33], we do not expect the same confounder. In our results, we observed the presence of nucleotide repeats, e.g., TA / CG dinucleotides; poly-C / poly-A in several of the divergent contexts, which could be explained by polymerase slippage [18]. Download: PPT PowerPoint slide PNG larger image TIFF original image Table 1. Baymer modeled 1KG private continental context mutations with extreme polymorphism probability differences. https://doi.org/10.1371/journal.pgen.1010807.t001 While the population-specific polymorphism probabilities estimated and polymorphism counts are identical between each pairwise comparison and thus correlated, we still note that 15/28 pairwise differences are specific to a single continental group. Of these, only the two canonical European context mutation differences (TCC>T and TCT>T) are in 3-mer contexts, otherwise all are found in 5-mer and greater window sizes. In South Asian samples, we find that the mean CTATA>T polymorphism probabilities are approximately 1.6 times higher than the remaining populations and in Africans TATATATC>G is approximately 1.9 times higher. The largest population-specific effect was discovered in East Asians where ATACCTC>A polymorphism probabilities are roughly 2.7 times higher than in European, African, or South Asian models. None of these effects have been explicitly documented before. Taken collectively, we observed relatively few instances of edges that were quantifiably different across continental groups, and those that were observed were largely confined to relatively small windows of context where we might have anticipated well-powered tests (e.g., 3- and 5-mers). To quantify the power of our specific analytic procedure for discovery and the sample size necessary to identify true differences in polymorphism probabilities, we performed simulations where true effect differences were ‘spiked-in’ between two populations over a range of weak to stronger effects and across a sampling of different sequence contexts (S1 Text). Differences in mutability between populations for this experiment are defined as the natural log of the polymorphism probabilities ratio (NLPPR) between each simulated population. This allowed us to construct credible sets of effects that we were reasonably well powered (>80%) to discover (Table 2). Unsurprisingly, the power scaled proportional to the number of context instances, simulated mutations in the dataset, and the size of the spiked-in differences (S5 Fig). Notably, extremely subtle shifts (NLPPR < = 0.01; 0.99–1.01-fold change) were not detectable at any sequence context size. On the opposite side of the spectrum, we found that we were reasonably powered to identify shift differences where NLPPR > 1.0 (fold decrease ≤ 0.37 or fold increase ≥ 2.72) up through 5-mers and in 6-mers with large sample sizes. For reference, the TCC>T polymorphism has an NLPPR = 0.291 (~1.34 fold increase)–the largest difference of any 3-mer by our calculation. Download: PPT PowerPoint slide PNG larger image TIFF original image Table 2. Power estimates for 1KG continental private polymorphism probabilities. https://doi.org/10.1371/journal.pgen.1010807.t002 In contrast, our experiment had essentially no power to discover 9-mer polymorphism probability changes and extremely limited power for 8-mers, even for large differences. Thus, there may exist large differences at these sizes that we could not reliably capture. These results are consistent with our comparisons in the real data (Table 1), as only differences within the detectable range at each mer-level were implicated. These power calculations suggest that, given the experiment we performed grouping all mutations together (agnostic to allele frequency or age, see Discussion), if any 3-mer differences greater than the TCC>T shift exist, we would have discovered these effects for a broad range of modest to very strong effects across a range of sequence contexts window sizes. This effectively sets bounds on the differences possible for this analysis scheme in this data. A sequence context model that captures variability in de novo mutational rates Given its formulation in handling data sparsity, we next sought to apply Baymer to develop a model that best captures rates of de novo mutations across the genome. We took advantage of a recent collection of 2,976 WGS Icelandic trios that identified 200,435 de novo events[38] and, analogous to the above, we partitioned de novo variants into even (for training) and odd (for testing) base pairs. We observed substantial improvement in the overall likelihood in the testing set for 5-mer size windows compared to 3-mers (3-mer vs 5-mer, ΔLL = 2,144), but only minimal improvement for increasing windows sizes further (5-mer vs 9-mer, ΔLL = 265, S6 Table). Indeed, Baymer did not select any sequence context feature beyond the 5-mer level with PIP > 0.95. This is not unexpected given our approach to regularization, as the number of events in larger sequence contexts is increasingly sparse, it is desirable to only include informative contexts to avoid overfitting. We next used Baymer to improve upon this baseline model. Previous work has demonstrated that inference of de novo mutational probabilities can be captured via rare variant polymorphism data obtained from population sets as a proxy [23]. We hypothesized that a partitioned set of polymorphism data based on: (i) larger sample sizes that (ii) closely matched the ancestry of the de novo set and (iii) focused on rare variants as a proxy to capture the most recent mutation events would generate the most transferrable model and robust rate estimates. To build variant partitions, we used variant call set data from gnomAD, focused on either a population-matched proxy (i.e., NFE, the non-Finnish European subset) or variant calls from all samples in gnomAD regardless of ancestry (i.e., ALL). For each of these, we created three partitions focused (i) exclusively on variants with one allele count (i.e., singletons; for NFE labeled NFE-1), (ii) exclusively on variants with two allele counts (i.e., doubletons; for NFE labeled NFE-2), and (iii) variants with allele count of two or greater (for NFE labeled NFE-2+). Beyond this, we also identified a set of putatively derived substitutions in the human lineage by comparing the GRCh38 human reference genome with ancestral sequences obtained from primates[39]. We applied Baymer to each variant set independently, comparing the likelihoods of each model to explain rates of de novo mutation in the test set after downscaling probabilities proportional to the sample size. First, we observed that for 3-mer sequence context models, the set of variants obtained from the de novo training set outperformed all other models despite 102 to 1,377 times fewer variants contributing to them than the polymorphism datasets (Fig 4A and S6 Table). In contrast, for larger windows of context (i.e., 7-mer and 9-mer), several of the polymorphism partitions explained the data better than one trained directly from de novo events. This result indicates that increased sample size is required to detect meaningful shifts in polymorphism probabilities in larger sequence context windows. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 4. Modeling de novo mutation probabilities using polymorphism datasets. Even base pair Halldorsson et al. de novo training data modeled by Baymer is compared to Baymer-modelled polymorphism datasets partitioned by allele count. (A) Multinomial likelihoods for each model are calculated on odd de novo base pair test data at various sequence context sizes. Polymorphism probability estimates were linearly scaled to match the mean polymorphism probability of the holdout dataset. (B) Polymorphism datasets were down-sampled to match the size of the even base pair de novo data (70,364 variants) and multinomial likelihoods were calculated on odd de novo base pair data. Each dataset was down-sampled using 5 different random seeds. The Log Likelihood of the 9-mer de novo training model is indicated with the blue dotted line. https://doi.org/10.1371/journal.pgen.1010807.g004 Despite evidence to suggest singleton datasets should best recapitulate de novo variation [4,23,31], we observed that models that trained exclusively on singletons and ALL-2 performed considerably worse than the rest across all windows of sequence context (Fig 4A and S6 Table). While our prior intuition that larger numbers of variants would have provided better rate estimates from increased power deeper in the context tree, rate models exclusively estimated with singletons suffer the most from the impact of recurrent mutations [20,40], especially at CpG sites, which include the highest polymorphism probability mutation type (CpG>T) (S6 Table). Alternatively, population concordance between training and test and/or the quality of variant calls used in training the model could also impact performance. As such, we next sought to explore the effect of noise in low allele count variants. Although we only used variants that passed gnomAD quality control checks, this filter still included a large proportion of variants with a negative log-odds ratio of being a true variant (AS_VQSLOD < 0; S6 Fig). This pattern was also evident for other variant allele counts but were most striking in singletons and the ALL-2 variant groups. Stricter quality filters (AS_VQSLOD > 5–10) considerably improved model performance, but still did not surpass the de novo training model at the 3-mer level (S6 Table). Our NFE singleton Baymer model trained on the strictest quality filter tested (AS_VQSLOD > 10) nearly equaled our best performing model, NFE-2+, with ~ 1/30th the number of variants, but came up just short. In summary, we observed that training from a population matched sample which excluded singletons, NFE-2+, best predicted rates of de novo mutations in 5-mer or larger contexts, better than models trained on de novo events directly. Next, we sought to determine which sample set best modelled the de novo test set adjusting for the total number of variants within the partition. To control sample size differences, we downsampled each partition to match the number of variants observed in the de novo training set (n = 70,364) five times. After down-sampling and when considering 9-mer context models, we observed that the partitions which included NFE exclusively (noted in green, Fig 4B) performed on average better than using the entirety of gnomAD, “ALL” (noted in orange in Fig 4B), which included a more diverse panel of individuals within Europe (e.g., Finnish) but also beyond Europe (e.g., East and South Asian, African and African American). This is consistent with prior belief that, after controlling for the total sample size, variants that derive from samples where ancestries more closely match are the most informative. A grafted tree approach provides superior estimates of de novo mutational probabilities Given the observations that de novo models only outperform polymorphism-based models when either small sequence contexts are used (Fig 4A) or the sample size is controlled (Fig 4B), we next sought to explore a transfer learning-inspired [41] strategy to improve upon our model performance. Transfer learning has previously been employed in a sequence context modelling setting [24]. We hypothesized that regularization means that de novo models have reduced performance with expanded sequence contexts due to low sample sizes. Indeed, our de novo model did not have the power necessary to confidently (PIP > 0.95) include any non-zero shifts in sequence contexts larger than 5-mers in the model (Fig 5A). The larger polymorphism datasets, however, were well-powered to detect shifts in every level of the tree (Fig 5A). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 5. Tree grafting strategy to share information between Baymer models. (A) For each de novo proxy model, we calculated the fraction of context polymorphism probability edges with a PIP > 0.95 in 2-mer through 9 mer-levels as a proxy for the degree of regularization in each model. (B) Edges in the de novo training model that are included with high-confidence (PIP>0.95) are very similar in magnitude and direction to their equivalents in the best-performing proxy model, NFE-2+, in 2-mer through 9-mer levels, implying a shared polymorphism probability shift structure. (C) Proposed tree-grafting schema for modeling de novo mutations that leverages mer-levels where de novo data is plentiful (1-mer through 3-mers) and uses polymorphism data to model the remainder of each model in larger mer-levels (4-mer through 9-mers) where the de novo model is underpowered. (D) The grafted tree method outperforms the previously best-performing model, NFE-2+. https://doi.org/10.1371/journal.pgen.1010807.g005 The nested tree structure of our polymorphism probability models provides a natural strategy where specific branches of the estimated trees can be interchanged, i.e., a “grafted” tree. We asked how similar estimates for edges in expanded sequence contexts are between our de novo model and the best-performing polymorphism model, NFE-2+. In edges in 2-mer and greater levels where the de novo training model is powered enough to detect shifts (PIP > 0.95), the mean posterior estimates of shifts are highly correlated (Fig 5B). This suggests a grafted tree approach is feasible, leveraging the polymorphism datasets for those edges the de novo model is incapable of estimating properly due to sparsity (Fig 5C). Therefore, we built a grafted tree model using 1- to 3-mer edges estimated in the de novo training data model, and 4- to 9-mer edges estimated using the NFE-2+ data model. The resulting combined model had a greater fit to the holdout de novo data than either the NFE-2+ model or de novo model alone (Fig 5D and S1 Text). Sequence context-dependent mutability is shared between closely related great ape species Finally, we examined how well human polymorphism models could capture variability in polymorphism levels observed in populations of great ape. Using polymorphism data from the Great Ape Genome Project [42], we built Baymer models of Pan troglodytes and Gorilla gorilla (S1 Text). We note broad agreement in estimated polymorphism probabilities between humans and chimpanzees (S7A Fig, Spearman correlation ρ = 0.950; RMSPE = 0.089) or gorillas (S7B Fig, Spearman correlation ρ = 0.942; RMSPE = 0.103). These results indicate that the rates of polymorphism at higher orders of sequences contexts are similar across closely related great ape species. As we were especially interested in how human polymorphism models compared with chimpanzee and gorilla models in predicting holdout data in each respective species, we then tested models on odd base pair data in each species, training models using even base pair data for the species in focus (S1 Text). For both chimpanzee (S7C Fig) and gorilla (S7D Fig) tests, species-matched 9-mer models outperformed all other models. While human-derived models are outperformed at the 9-mer level, it is notable that human 9-mer models are more likely than chimpanzee 7-mer models against chimpanzee data and gorilla 5-mer models against gorilla data (S7 Table). Taken collectively, these results suggest the rates of polymorphism at higher orders of sequences contexts are similar across closely related great ape species, with within-species models best capturing variability in observed polymorphism levels. Discussion Here, we present Baymer, a Bayesian method to model mutation rate variation that computationally scales to large windows of nucleotide sequence context (S2 Text and S8 Table), robustly manages sparse data through an efficient regularization strategy, and emits posterior probabilities that capture uncertainty in estimated probabilities. Consistent with previous studies [24–26], we show that expanded sequence context models in most current human datasets are overfit with classic empirical methods but considerably improve model performance when properly regularized. As a result, this method allows for renewed evaluation of experiments that originally were statistically limited to polymorphism probability models with small sequence context windows. We examined differences in polymorphism probabilities between the continental populations in the 1KG project. While differences in 3-mer polymorphism probabilities have been well-documented [20–22] and expansions up to 7-mers have been tested [34], both methods rely on empirical models with frequentist measures of uncertainty. Here, we expanded the search space out to 9-mer windows and leverage the uncertainty estimated in the model to directly quantify differences in these populations. We note that many of the differences discovered contain poly-nucleotide repeats. There is some prior literature on the mechanism of slippage in polymerases during replication of such sequences [18], so differential efficiencies of these enzymes across populations could conceivably result in these patterns. However, it is also very possible that artifacts from sequencing errors with differential effects across populations could explain the differences. Despite being well-powered to identify a large range of differences in 3-mer and smaller contexts, we identified very few contexts that differ with high probability between the populations tested. This implies that if large-scale population differences in the mutation spectrum do exist at these window context sizes, they may be comprised of numerous subtle shifts rather than a few large changes, in agreement with conclusions from prior work [22]. We also explicitly placed bounds on the magnitude of differences that could possibly exist in these data without being detected, quantifying what differences we can expect to be discovered given the way variants are grouped in this experiment. Even though the 1KG project is relatively small compared to current datasets, the number of sequence contexts available for modeling is dataset-independent and inherently limited by the sequence diversity of the human genome. Thus, while more polymorphism data could lead to the discovery of additional smaller shifts in the future, bigger datasets will not improve the power to detect larger shifts in this allele frequency agnostic setting. In fact, for very large samples, polymorphisms in some contexts can become saturated [43], reducing the information content in a similar manner as overly sparse data. Thus, both to increase power and to improve modeling resolution, it will become necessary to partition the data (e.g., by allele frequency or variant age [36], or other genomic features). It remains a challenge to disentangle the contribution of demography [20,35,44] versus changes in the underlying mutation rate on the mutation spectrum. Here, we control for the site frequency spectrum of variants included, but the next stage of this model will need to incorporate more sophisticated demographic features. Integrating Baymer-derived trees with a joint mutation spectrum and demographic history method, such as mushi [35], is a promising future direction. While this work focuses on modeling mutability in regions minimally affected by background selection, constraint could also bias estimates. Given prior work [19] we do not expect the underlying sequence-context-mediated mutability to behave any differently in constrained regions, suggesting future Baymer-estimated codon-aware models to explicitly model expected variation in coding regions. We also aimed to address the degree to which polymorphism datasets could be used to approximate the de novo mutation rate. Currently, true de novo mutation datasets are limited in size, which place bounds on the scope of inference for adequate sequence context modeling. We demonstrate that polymorphism datasets are accurate proxies for de novo mutation models and largely share the same context-dependent mutability shifts, though in contrast to reports in the literature [4,23,31], the focus exclusively on singleton variants (at least, using gnomAD calls) performed poorly relative to all other considered models. Indeed, our experiment indicates that it is preferable to use germline mutation models based on large polymorphism datasets that can estimate shifts through the 9-mer level than it is to use the largest 3-mer de novo dataset, as is commonly used in the literature [4,5,31]. Including exclusively variants from either polymorphism data or de novo data was also suboptimal, however, as the best possible model we built for estimating de novo mutation rates used de novo mutations in concert with polymorphism datasets. The success of this experiment implies a general context-dependent mutability architecture that underlies the human mutation spectrum. The similarity of the derived dataset, which in theory represents the oldest subset of variants tested, to the de novo variation further strengthens this argument. We note this dataset could in theory be biased towards European samples given the history of the Human Genome Project [45], and as such, refinements will need to be made as more diverse representations of the human genome are created. Overall, this work suggests that although there have been some well-documented small changes in context-dependent mutation rates, the general architecture remains largely conserved during modern human history. Our experiments modeling great ape variation suggests this general architecture might be more pervasive across the tree of life. While some specific mutation spectra differences have been documented [22,46], we note broad agreement amongst closely related species as well as similar signals in extended sequence contexts. For those non-human species with WGS datasets, cohort-sizes are usually very small (< 100), however, Baymer is well-suited to handle these sparse data situations. Furthermore, for those species with very little data, this work opens the exciting possibility to leverage closely related species’ models as priors for estimating variation in less well-characterized relatives. Further work is necessary to model species across the tree of life to determine the extent that sequence context-dependent mutability is shared and how transferable 9-mer models can be. One limitation of the model is the treatment of multi-allelic sites. Currently, multi-allelic sites are treated as separate polymorphisms which violates assumptions of the multinomial model, where only one outcome is possible for each locus. When we excluded multi-allelic sites, we observed biases in the rates of CpG>A and CpG>G mutations, which are disproportionately filtered as a side-effect of sharing the same sequence contexts with CpG>T mutations. A more nuanced approach that models multiallelic and biallelic sites separately and then integrates jointly would deal with this issue, though multiple mutations at the same nucleotide position with the same allele change would require additional effort [40]. Finally, although we can identify regions of the tree where polymorphism probabilities diverge and thus infer critical points in the tree, this model is tailored towards polymorphism probability estimation rather than explicitly for motif discovery [27]. Our objective is to estimate polymorphism probabilities rather than finding those contexts with the largest effect sizes. Although including even-length contexts yields better-performing models, the current tree architecture only explicitly captures the effects of half of such contexts. While adding one nucleotide at a time pseudo-symmetrically for tree generation reduces the computational sampling load, it makes for more difficult interpretation of the resulting mono-nucleotide impacts. Baymer’s formulation also does not model the mutability of target contexts independent of mutation type, which currently requires post-hoc analysis to identify motifs that have non-specific mutability signatures. Future work will therefore need to integrate all possible paths through the sequence context tree and share information across contexts between mutation type trees. In all our experiments, we focused on the entirety of the non-coding genome that is accessible to sequencing. That said, Baymer can easily be applied to any genomic features of interest for both polymorphism probability estimates and comparisons of feature-dependent sequence context mutability changes. Our approach does not currently incorporate genomic features in the model, but given genomic area bounds, polymorphism probabilities can be tailored to a biological question of interest. Addressing questions regarding the impact of genomic features on observed polymorphisms will be enhanced with well-regularized models, as smaller genomic areas or specific variant conditions can induce considerable data sparsity by reducing the number of contexts and/or polymorphisms available. Therefore, Baymer paves the way for exciting possibilities to study the effects of genomic features, variant age, and smaller subpopulations on sequence context-dependent mutation rate variation. Supporting information S1 Fig. Empirical even odd polymorphism probability scatter plots for the NFE dataset including contexts with zero mutation variants. Baymer mean posterior estimates for (A) 3-mer models (Spearman correlation = 0.999; p < 10−100; RMSPE = 0.0009), (B) 5-mer models (Spearman correlation = 0.999; p < 10−100; RMSPE = 0.0063), (C) 7-mer models (Spearman correlation = 0.992; p < 10−100; RMSPE = 0.0459), and (D) 9-mer models (Spearman correlation = 0.876; p < 10−100; RMSPE = 0.7441) in even and odd base pair datasets. Polymorphism probabilities in the bottom two and top left quadrants correspond to those contexts where no mutations are present for the given mutation type in the respective datasets. These polymorphism probabilities are exclusively calculated using pseudocounts. https://doi.org/10.1371/journal.pgen.1010807.s001 (EPS) S2 Fig. Comparison of Baymer mean posterior estimates for differing allele frequency (AF) bins in the NFE dataset. (A-D) AF 0.02–0.05 compared against 0.05–0.15 AF, 0.15–0.30 AF, 0.30–0.50 AF, and 0.50–0.85 AF, respectively. (E-G) AF 0.05–0.15 AF compared against 0.15–0.30 AF, 0.30–0.50 AF, and 0.50–0.85 AF, respectively. (H-I) 0.15–0.30 AF compared against 0.30–0.50 AF and 0.50–0.85 AF, respectively. (J) 0.30–0.50 AF compared against 0.50–0.85 AF. https://doi.org/10.1371/journal.pgen.1010807.s002 (EPS) S3 Fig. Comparison of empirical and Baymer-derived 9-mer polymorphism probabilities in NYGC-resequenced 1000 Genomes Phase 3 non-admixed non-Finnish European (EUR) polymorphisms with derived AC ≥ two in non-coding accessible regions. (A) Empirical 9-mer polymorphism probabilities for context mutations with at least one occurrence in both datasets (102,875 omitted context mutations) are plotted against one another (Spearman correlation = 0.862; RMSPE = 0.175). (B) Baymer mean posterior estimates for 9-mer polymorphism estimates in even and odd base pair datasets (Spearman correlation = 0.986; RMSPE = 0.042). https://doi.org/10.1371/journal.pgen.1010807.s003 (EPS) S4 Fig. Overview of the characteristics of edge mutability change dynamics in Baymer models of the NFE dataset. (A) Histogram of the number of edges per 9-mer that were inferred to confidently change polymorphism probabilities (PIP > 0.95). (B) Histogram of the maximum edge size per each 9-mer that was inferred to confidently change polymorphism probabilities (PIP > 0.95). (C) Estimated distributions of phi for each mer size level. (D) The distribution of the fractional differences of each 9-mer mean posterior polymorphism probability with their respective nested 3-mer mean posterior polymorphism probability estimates, partitioned by mutation type. https://doi.org/10.1371/journal.pgen.1010807.s004 (EPS) S5 Fig. Fraction overlap of simulated datasets trained by Baymer at varying sequence contexts and log changes to the null polymorphism probability. https://doi.org/10.1371/journal.pgen.1010807.s005 (EPS) S6 Fig. Variant Quality Scores reported in gnomAD by allele count. Distribution of gnomAD AS_VQSLOD quality scores in non-Finnish European samples (“NFE”; A-C) and in all populations (“ALL”; D-F), separated into singletons (A,D), doubletons (B,E), and variants with allele count greater than or equal to 3 (C,F). https://doi.org/10.1371/journal.pgen.1010807.s006 (EPS) S7 Fig. Comparison of Homo sapiens Baymer model (NFE-2+ model) estimates with Pan troglodytes and Gorilla gorilla great ape species. (A) Mean polymorphism estimates of Homo sapiens model plotted against mean polymorphism estimates of Pan troglodytes model (Spearman correlation = 0.957; RMSPE = 0.088). (B) Mean polymorphism estimates of Homo sapiens model plotted against mean polymorphism estimates of Gorilla gorilla model (Spearman correlation = 0.950; RMSPE = 0.097). (C) Multinomial likelihoods for each model are calculated on odd base pair Pan troglodytes test data at various sequence context sizes. Pan troglodytes model is trained on even base pair data only. (D) Multinomial likelihoods for each model are calculated on odd base pair gorilla gorilla test data at various sequence context sizes. Gorilla gorilla model is trained on even base pair data only. Polymorphism probability estimates were linearly scaled to match the mean polymorphism probability of the holdout dataset. https://doi.org/10.1371/journal.pgen.1010807.s007 (EPS) S1 Table. Calibration of credible sets across mer levels by measuring number of simulations capturing the truth value. https://doi.org/10.1371/journal.pgen.1010807.s008 (XLSX) S2 Table. Likelihood of even base pair NFE Baymer models on 9-mer odd base pair holdout NFE data. https://doi.org/10.1371/journal.pgen.1010807.s009 (XLSX) S3 Table. Top 5 largest mean posterior phi estimates for each context size. https://doi.org/10.1371/journal.pgen.1010807.s010 (XLSX) S4 Table. Motifs tested for enrichment in the top or bottom 1% of polymorphism probabilities for each mutation type. https://doi.org/10.1371/journal.pgen.1010807.s011 (XLSX) S5 Table. Baymer modeled 1KG private continental context mutations with extreme polymorphism probability differences. https://doi.org/10.1371/journal.pgen.1010807.s012 (XLSX) S6 Table. Sample sizes and test likelihoods for each de novo comparison dataset. https://doi.org/10.1371/journal.pgen.1010807.s013 (XLSX) S7 Table. Sample sizes and test likelihoods for each great ape comparison test. https://doi.org/10.1371/journal.pgen.1010807.s014 (XLSX) S8 Table. Run-times for models with increasing sequence context windows. https://doi.org/10.1371/journal.pgen.1010807.s015 (XLSX) S1 Text. Supplementary Methods. https://doi.org/10.1371/journal.pgen.1010807.s016 (DOCX) S2 Text. Computational Considerations. https://doi.org/10.1371/journal.pgen.1010807.s017 (DOCX) S3 Text. Data Availability. https://doi.org/10.1371/journal.pgen.1010807.s018 (DOCX) Acknowledgments We would like to thank Dr. Ziyue Gao for her very helpful feedback on the manuscript and in the development process.

Wagner, Ellen R.;Nightingale, Nicole M.;Jen, Annie;Overmyer, Katherine A.;McGee, Mick;Coon, Joshua J.;Gasch, Audrey P.

doi: 10.1371/journal.pgen.1010593pmid: 37410771

Introduction Many physiological processes are essential for growth, but so too is the coordination of those processes to form an integrated cellular system. Actively dividing cells must coordinate metabolism and division with the synthesis and segregation of DNA, proteins, organelles, and other macromolecules, all within a precisely timed cell cycle. Failure to coordinate these processes can jeopardize fitness due to suboptimal cellular composition and energy expenditures. Mechanistically, much remains unknown about how cells coordinate cellular processes. One of the best studied examples is the intimate control of successive cell cycle phases, which depends on interconnected transcriptional and post-translational controls regulated by dispersed checkpoints along the way [1–8]. The cell cycle is also coordinated with metabolism; cell-cycle regulators coordinate metabolic flux with cell-cycle phases, which may be related to cell size checkpoints since cells must reach a critical size before a new cell cycle is initiated [2,3,6–8]. A critical feature of integrated cellular systems is thus balancing energy demands with division and replication. Knowing how cells coordinate growth and division with other physiological processes is important for understanding how cells function on a fundamental level, but it also has practical applications. Microbes can be engineered to produce a variety of commodity chemicals and biofuels with high yields to maximize economic returns. Microbial design strategies have considered how cells allocate resources so as to redirect cellular energy toward making compounds of interest [9–15]. Redirecting resources away from other processes can improve cellular product yields and thus decrease costs [16,17]. However, an added complication is that many industrial processes are stressful for engineered microbes, which mount stress-defense systems that further deplete cellular resources from product formation. The interplay of growth, metabolism, division, and stress defense remain murky, limiting engineering efforts [18,19]. Here, we studied how growth, division, and metabolism are normally coupled in cells by investigating a strain in which these processes have been decoupled. We previously characterized a series of Saccharomyces cerevisiae strains engineered to produce biofuel products from xylose, a pentose sugar abundant in plant biomass but not recognized by S. cerevisiae as a fermentable sugar [17,20–22]. A major goal for sustainable biofuel production is to utilize xylose and other carbon sources to maximize biomass conversion to products. Past work in our center found that engineering S. cerevisiae to ferment xylose anaerobically requires core xylose metabolism genes (encoding xylose isomerase, xylulokinase, and transaldolase [23,24]); however, introducing these genes is not enough to enable fermentation. Many groups have combined strain engineering with adaptive laboratory evolution to evolve xylose fermentation [25–30]. Cells require additional null mutations in oxidoreductase GRE3, iron-sulfur (Fe-S) chaperone ISU1, and RAS signaling inhibitor IRA2 [27,31]. We previously showed that these mutations help to rewire cellular signaling to unnaturally upregulate the growth-promoting Protein Kinase A (PKA) pathway in conjunction with Snf1 that usually responds to poor carbon sources [32]. We proposed that activating PKA and Snf1 promotes growth in the context of an otherwise unrecognized carbon source [32]. Coordinated induction of PKA and Snf1 allows cells to recognize xylose as a fermentable carbon source while enhancing growth and metabolism signals. Although PKA activation is critical for anaerobic xylose growth and metabolism, during that study we made a surprising discovery: the mechanism of PKA up-regulation influences how growth and metabolism are coordinated. PKA can be activated by RAS activity, which stimulates adenylate cyclase to produce the allosteric regulator cAMP that binds and dissociates the PKA regulatory subunit Bcy1 (Fig 1A) [33]. In engineered yeast, activating PKA by IRA2 deletion, thus increasing RAS activity, enables rapid anaerobic xylose fermentation and growth on xylose as the sole carbon source. However, activating PKA by deleting the PKA regulatory subunit BCY1 allows rapid anaerobic xylose fermentation but with little to no growth (Fig 1B and 1C) [32]. In both strains, the effect is due to PKA upregulation since inhibition of PKA activity blocked both metabolism and growth [32]. Thus, deleting BCY1 in this strain background decouples growth and xylose metabolism for reasons that are not known. Importantly, other uncoupled biological processes related to PKA function have also been described [34], implying PKA’s central role in process coupling. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 1. Activation of PKA is needed for xylose fermentation. A. A brief overview of the PKA signaling pathway. B-C. Average (n = 6 biological replicates) growth (OD600, optical density) and (B) xylose concentration in the medium over time of parental Y184, ira2Δ, bcy1Δ, and ira2Δbcy1Δ strains grown anaerobically on rich medium containing xylose as a carbon source. Asterisks denote significant differences in profiles (p < 0.05, ANOVA); n.s. indicates ‘not significant’ (p > 0.05). D. Growth and fermentation capabilities of strains shown in B-C. https://doi.org/10.1371/journal.pgen.1010593.g001 Here, we explored phenotypic consequences of IRA2 and BCY1 deletions to elucidate how cells normally coordinate growth and metabolism. We integrated transcriptomic, phospho-proteomic, and lipidomic analysis across a suite of strains with different mutations and growth/metabolism phenotypes. The results implicated the importance of lipid metabolism as a linchpin in the coordination of growth with metabolism: cells lacking BCY1 show unique transcriptomic and lipidomic responses that point to defects in lipid regulation. To uncover causal genes, we also performed adaptive evolution to re-evolve growth coordination in the bcy1Δ strain. Remarkably, the evolved strain acquired mutations in a PKA catalytic subunit TPK1 and phospholipid biosynthesis regulator OPI1, among other genes, and Opi1 was required for the growth-metabolism coupling in the evolved strain. These results suggest that PKA-dependent regulation of lipid metabolism is critical for growth, perhaps to coordinate membrane biogenesis and signaling with other cellular processes. Results We began by characterizing a suite of strains with different anaerobic xylose growth and fermentation capabilities. Parental strain Y184 harbors the xylose-metabolism gene cassette along with mutations in ISU1 and GRE3 but cannot grow on or metabolize xylose anaerobically (Fig 1B and 1C). Deleting IRA2 from this strain allows cells to grow on and metabolize xylose anaerobically. In contrast, deletion of BCY1 from Y184 permits rapid anaerobic xylose fermentation but with only minimal growth (Fig 1B, 1C, and 1E). Previous work studying growth over 90 hours validated limited growth of the bcy1Δ strain even after long periods [35]. We also investigated an ira2Δbcy1Δ double mutant. The double mutant was phenotypically similar to the bcy1Δ, although its xylose fermentation capabilities were highly variable across replicates, for reasons we do not understand but could pertain to extremely high PKA activity. As such, the double mutant was not statistically different from either the Y184 parent or the bcy1Δ strain. Nonetheless, we used it to investigate the genetics of PKA signaling through these different branches. The three strains grow indistinguishably on glucose (p> 0.05, S1A Fig) with similar glucose consumption (p > 0.05, S1B Fig) and ethanol production (p > 0.05, S1C Fig), indicating that these phenotypes are specific to anaerobic xylose conditions. We started by comparing transcriptomic responses to identify transcripts whose abundance across the strain panel correlates with growth or anaerobic xylose metabolism. Cells were grown in an anaerobic chamber to mid-log phase on rich medium with glucose as a carbon source (YPD) then switched to rich medium containing only xylose (YPX) for three hours, long enough for the ira2Δ strain to resume growing (Fig 2A). We performed short-read sequencing to measure changes in transcript abundance after the glucose-to-xylose shift. To understand strain responses, we compared transcript abundances across strains grown under each condition; we also compared the fold change in transcript abundance within each strain responding to carbon shift. There were major differences in expression comparing the strains growing on xylose, whereas only mild expression differences were observed comparing strains grown on glucose (see Fig 2C, right panel). Correspondingly, strains do not differ substantially in their ability to grow anaerobically on glucose (S1 Fig) [35]. Thus, the differences in the fold-change expression response to the carbon shift are driven by differences in the xylose condition. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 2. Few transcriptomic patterns correlate with anaerobic xylose growth. A. Experimental overview. Strains were grown anaerobically in rich glucose medium to early/mid-log phase, then switched to anaerobic rich xylose medium for three hours. B. Expression of 65 genes whose log2(fold change) upon glucose to xylose shift is different (FDR < 0.05) in at least one of the three non-growing strains (Y184, bcy1Δ, ira2Δbcy1Δ) compared to ira2Δ. Genes (rows) were organized by hierarchical clustering across biological triplicates measured for each strain (columns). Genes discussed in the text are annotated on the figure. C. Hierarchical clustering of 292 genes whose log2(fold change) upon glucose to xylose shift is different (FDR < 0.05) between the non-fermenting Y184 strain and the two robustly xylose fermenting strains (ira2Δ, bcy1Δ). The blue-yellow heatmap on the left represents the log2(fold change) in expression upon glucose to xylose shift across biological triplicates (columns). The purple-green heatmap on the right represents the abundance of each transcript (rows) in each strain grown on glucose (G) or xylose (X), relative to the average (n = 3) abundance of that transcript measured in the Y184 YPD sample. Clusters I and II are described in the text. D. Expression of 15 genes from C) that have annotations linked to glycolysis, gluconeogenesis, TCA cycle, and carbohydrate storage. https://doi.org/10.1371/journal.pgen.1010593.g002 Our expectation at the outset was two-fold. On the one hand, we expected to find expression changes common to the xylose fermenting ira2Δ and bcy1Δ mutants, but discordant in Y184 cells–these expression patterns may relate to xylose metabolism, since the Y184 strain is incapable of xylose metabolism [31,32]. On the other hand, expression patterns unique to the ira2Δ strain–the only strain capable of growing anaerobically on xylose–may reflect expression patterns related to growth. Few gene expression patterns correlate strictly with growth phenotypes Somewhat surprisingly, there were few genes whose expression correlated strictly with growth phenotypes. Only two genes showed xylose-responsive expression changes that were specific to ira2Δ cells compared to the other three strains analyzed as a group in the statistical model (FDR < 0.05; see Methods): daughter-cell-specific glucanase DSE4 and L-homoserine-O-acetyltransferase MET2. In fact, hierarchical clustering of all genes with a transcriptomic change in response to the carbon shift showed that the ira2Δ strain’s response to xylose shift was most similar to that of Y184 cells, even though one strain can grow on and anaerobically ferment xylose and the other cannot (S2A Fig and S1 Table). We next performed pairwise comparisons of the glucose-to-xylose fold-change responses between each strain and the ira2Δ strain, then combined the lists of genes identified in all three comparisons. This method identified 65 genes; however, investigating the expression patterns once again indicated that the ira2Δ response was most similar to Y184 cells but with weaker magnitudes of change (Fig 2B and S2 Table). This set of 65 genes was enriched for genes induced in the environmental stress response (iESR genes [36], p = 2x10-7, hypergeometric test). Many genes induced in the Y184 and ira2Δ strains, but largely not in bcy1Δ strains, included genes related to metabolism, including several in the mitochondrial TCA cycle and peroxisomal fatty-acid oxidation pathway, which may reflect that bcy1Δ strains are more likely to recognize xylose as a fermentable carbon source. We specifically investigated the set of 65 genes for those that encode cell-cycle regulators and kinases, since these may be involved in growth kinetics; however only three, six, or eight genes within these categories were differentially expressed in Y184, bcy1Δ, or ira2Δbcy1Δ cells, respectively, compared to ira2Δ cells in response to xylose shift (FDR < 0.05). The ira2Δ strain showed weakly lower expression of cyclin CLN2 and anaphase-promoting complex CDC20, whereas other strains showed strong reduction in expression (S3 Table). Transcript abundance for these genes is known to fluctuate during the cell cycle, thus while it is possible their expression influences growth arrest, it is likely that the expression of these genes reflects the expected difference between cycling (ira2Δ) and non-cycling (Y184, bcy1Δ, ira2Δbcy2Δ) cells. Expression of cell-cycle genes did not implicate arrest in a particular cell-cycle stage, consistent with early transcriptomic studies that showed that gene expression during cell-cycle arrest does not parallel expression of cells cycling through those phases [37]. Previous chemostat studies reported that repression of ribosomal protein (RP) and ribosome biogenesis (RiBi) genes is correlated with decreased growth, and these studies proposed that expression of these genes can predict cellular growth rate [38–41]. However, here we saw no correlation of RP and RiBi transcript abundance or response with growth phenotypes. The Y184 strain strongly repressed RP and RiBi genes upon xylose shift, which might be expected for a strain that arrests its growth, but so too did the ira2Δ strain, albeit with weaker magnitude of repression. Surprisingly, bcy1Δ and ira2Δbcy1Δ cells, whose growth is largely arrested after the xylose shift, showed little change in RP and RiBi transcripts compared to glucose-dependent growth (FDR <0.05, S2B Fig and S4 Table). These results reinforce past work from our lab that the expression of ribosome-associated genes does not necessarily parallel growth rate [42]. They further suggest that bcy1Δ strains cultured in xylose are unlikely limited by the abundance of RP and RiBi transcripts. Overall, while the non-growing strains have stronger repression of a few cell-cycle regulators when compared to the ira2Δ strain, there was not a clear gene expression pattern to describe why ira2Δ cells grow and bcy1Δ strains do not. Few gene expression patterns correlate strictly with metabolism phenotypes We next investigated shared gene expression changes related to robust anaerobic xylose fermentation. We compared expression in the Y184 strain responding to the xylose shift to the ira2Δ and bcy1Δ strains analyzed as a single group in the statistical model (we excluded the ira2Δbcy1Δ strain due to the variability of its fermentation phenotype, although it is capable of anaerobic xylose fermentation). This identified 292 differentially expressed genes (FDR < 0.05; S5 Table). Hierarchical clustering revealed that these genes typically had larger expression changes in Y184 and that those expression changes were progressively weaker across the strain series; once again, Y184 and the ira2Δ strain were more similar to one another than they were to the bcy1Δ strain (Fig 2C and S5 Table). Collectively, these genes were heavily enriched for genes in the ESR (p = 3.624x10-10, hypergeometric test). Deeper interrogation revealed several small gene clusters of interest. Cluster I contained 19 genes induced in the ira2Δ and bcy1Δ strains but repressed in the Y184 strain. This group did not contain any functional enrichments; however, proteins encoded by several of these genes localize to the endoplasmic reticulum. Cluster II contained 31 genes induced in Y184 and either unchanged or repressed in both ira2Δ and bcy1Δ strains. This group was enriched for genes involved in protein folding (p = 9.49x10-6, hypergeometric test) and included the Hsp90 chaperone and cochaperone genes HSP82, STI1, and AHA1, as well as the mitochondrial matrix protein chaperone HSP10. Hsp90 can act as a signal transducer for alternative carbon source metabolism [43], again suggesting that Y184 does not recognize xylose as a fermentable carbon source. We specifically interrogated the 292 genes for those involved in glycolysis, gluconeogenesis, TCA cycle, and carbohydrate storage, predicting that differences in expression would relate to altered xylose metabolism capabilities. This identified 15 genes with functional annotations linked to at least one of these processes (Fig 2D and S6 Table). The xylose fermenting strains shared expression at several hallmark genes. For example, Y184 cells strongly induced hexose transporter HXT5, normally induced by non-fermentable carbon sources, whereas the three mutant strains showed a weaker induction (FDR = 4.25x10-5). The glucose-repressed aldehyde dehydrogenease ALD2 was induced in Y184 and either did not change (ira2Δ cells) or was repressed (bcy1Δ cells) upon the switch to xylose (FDR = 4.32x107). Furthermore, glucose-induced transcriptional repressor MIG2 showed stronger induction in the xylose fermenting strains, and especially bcy1Δ strains, compared to Y184 (FDR = 0.047). These data are all consistent with the hypothesis that the xylose fermenters recognize xylose as a fermentable carbon, whereas Y184 activates a carbon-starvation response. Regulatory analysis reveals strain-specific differences in carbon, iron, and lipid gene control We next focused on understanding how growth and metabolism are decoupled in the bcy1Δ strain, and we thus directly compared its expression to that in ira2Δ cells. We focused on genes whose expression changes in response to the xylose shift were in opposing directions to implicate processes involved in decoupling growth and metabolism (Fig 3A and S7 Table, see Methods). Among the identified genes, we scored enrichment of functional terms as well as known targets of transcriptional regulators (S8 Table). We also used motif analysis to discover shared sequence motifs upstream of genes uniquely induced or repressed in the bcy1Δ strain, and then matched those to known transcription factor binding sites (see Methods). We identified 654 genes differentially expressed in bcy1Δ cells and with a fold-change in the opposite direction as ira2Δ cells upon the glucose-to-xylose shift (Fig 3A and S7 Table). Importantly, only 82 genes (12.5%) showed significant differences in basal gene expression when cells were grown on glucose (S3A Fig), indicating that the majority of genes are identified due to differences in response to xylose shift. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 3. Genes uniquely expressed in the bcy1Δ strain implicate an integrated response to xylose metabolism and growth coupling. A. Expression of 654 genes whose log2(fold change) upon glucose to xylose shift is different (FDR < 0.05) between the ira2Δ and bcy1Δ strains and whose expression change is in the opposite direction across the two strains (see Methods for details). Significant functional enrichments are annotated next to the two main clusters (p < 10−4, hypergeometric test). Bar graph inset represents the log2(fold change) of the two phosphatidic acid biosynthesis enzymes in this group, see text for details. B. Regulatory relationships between transcription factors whose targets or known binding sites were enriched in (A). Documented PKA-dependent phosphorylation is indicated by a P. See text for details. https://doi.org/10.1371/journal.pgen.1010593.g003 The results implicated several regulators, some with prior connections to anaerobic xylose fermentation. 318 genes induced in the bcy1Δ strain shifted to xylose, but repressed in the ira2Δ cells, were enriched for amino acid and sphingolipid biosynthesis genes, as well as targets of the carbon-responsive Azf1 transcription factor (p < 10−4, hypergeometric test). Previous work from our lab implicated Azf1 in anaerobic xylose fermentation, and indeed, we showed that the over-expression of AZF1 in an ira2Δ strain enhances the rate of anaerobic xylose utilization [32]. Additionally, PKA has been implicated in Azf1 phosphorylation [44]; together with the fact that the AZF1 gene is uniquely induced in the bcy1Δ strain suggest its functional importance in xylose metabolism (see Discussion). In contrast, several regulators were implicated by the 336 genes uniquely repressed in the bcy1Δ strain. These included genes harboring upstream binding sites of the iron-responsive Aft1/2 transcription factors (S3B Fig) and known targets of transcriptional activator Ino4 that responds to inositol for phospholipid biosynthesis (Fig 3A; see more below). Iron is an important cofactor of many enzymes, including those involved in mitochondrial respiration, lipid biogenesis, and amino acid biosynthesis, all of whose genes were among the differentially regulated genes studied here. Additionally, Aft1/2 regulation and the iron regulon have been linked with PKA activity; however, direct interactions remain to be identified [45]. Interestingly, Aft1/2 and Azf1 both are both connected to the regulator Mga2, which controls lipid and hypoxia genes and that we previously showed enhances anaerobic xylose fermentation when over-expressed in ira2Δ cells [32,46] (see Discussion). Targets of the Ino2/4 regulators that respond to inositol for phospholipid biosynthesis were also present in this gene set; while a majority of the targets identified here were repressed in the bcy1Δ strain, some of the known targets were repressed in the ira2Δ strain but induced in the bcy1Δ mutant (S3C Fig and S9 Table). This may reflect the complexities of the genes’ regulation by other factors. Nonetheless, Ino2/4 targets were enriched among the genes oppositely regulated in the bcy1Δ versus ira2Δ strain. Overall, these results provide an interesting link between PKA signaling, carbon and iron responses, and lipid metabolism (Fig 3B). The presence of many lipid biosynthesis genes in this gene set and the highly regulated role of lipids in cell growth and proliferation prompted a deeper investigation of lipid metabolism genes. The bcy1Δ strain repressed genes involved in ergosterol biosynthesis and some targets of Ino2/4 that are involved in phospholipid metabolism (Figs 3A and S3C). This response is consistent with the model that Ino2/4 activity is reduced. However, the bcy1Δ strain also induced some genes involved in the synthesis of phosphatidic acid (PA) (Fig 3A inset), which normally promotes Ino4 activity by sequestering Ino4’s inhibitor Opi1 to the ER membrane [47]. This response suggests that some connection between PA, Opi1, and Ino4 is disrupted in the absence of BCY1. PKA is known to regulate the Ino2/4 pathway through direct phosphorylation of Opi1 to increase its inhibitory activity [48]. Together, these results raised the possibility that the bcy1Δ strain has important differences in lipid metabolism and perhaps composition, which could be modulated by differences in PKA activity in this strain. Lipidomic and phosphoproteomic analyses show disrupted phospholipid metabolism in bcy1Δ cells Since the transcriptomic responses implicated differences in lipid metabolism, we investigated the lipidomic composition of our strains. Strains were grown in a similar design as the transcriptomic analysis, where anaerobically glucose-grown Y184, ira2Δ, bcy1Δ, and ira2Δbcy1Δ cells were shifted to anaerobic xylose media for three hours before lipids were analyzed by mass spectrometry (see Methods). We detected over 4000 lipid species including 239 that were confidently assigned to a particular lipid class (S10 Table). All detected lipid species were included in the statistical analysis to obtain a wholistic understanding of lipidome differences between the strains. We again sought to find lipidomic profiles correlated with xylose metabolism and growth, xylose metabolism but no growth, and no xylose metabolism or growth. We compared the Y184 strain to the three strains with upregulated PKA activity and identified 18 lipids whose change in abundance upon a shift to xylose significantly differed in Y184 cells. This group included phosphatidylserine (PS) species (Fig 4A and S11 Table). Interestingly, all three mutants increased the abundance of these PS species when shifted to xylose, whereas Y184 cells decreased the abundance of one and failed to induce the other to the same degree as the mutants. The gene encoding the PS synthase CHO1 was strongly induced in Y184 cells, indicating that the decrease in PS in Y184 cells is unlikely due to decreased CHO1 expression. Instead, we analyzed previous phosphoproteomic data from our lab and discovered that Cho1 was phosphorylated to a much higher degree in the Y184 strain on serine 46 (|log2FC| > 1, Table 1), a known PKA site that inhibits Cho1 activity [49]. Together, these results indicate PKA-dependent inhibition of PS synthesis in Y184 cells. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 4. bcy1Δ strains show altered phospholipids after anaerobic xylose shift. A-B. Abundance of lipids (rows) with a significant difference in log2(fold change) upon anaerobic glucose-to-xylose shift in (A) Y184 compared to PKA pathway mutants (ira2Δ, bcy1Δ, ira2Δbcy1Δ) analyzed as a group in the statistical model or (B) ira2Δ cells compared to ira2Δbcy1Δ cells. Lipids of interest are annotated. C. Partial phospholipid biosynthesis pathway with transcriptomic and lipidomic data represented. Yellow-blue boxes next to each enzyme name represent the average log2(fold change) in transcript abundance upon glucose-to-xylose shift for each strain, as outlined in the key. Significant differences compared to the ira2Δ strain (FDR < 0.05) are represented in sharp, bolded boxes, whereas insignificant differences are translucent. Colorized pathway arrows (yellow: induced, blue: repressed) represent the predominant transcript patterns for that enzymatic step when comparing the bcy1Δ and ira2Δ strains. Lipids whose fold-change in abundance is different in specific strains are according to the key. Lipid abbreviations: FFA–free fatty acids; PA–phosphatidic acid; DG–diacylglycerol; TG–triacylglycerol; PI–phosphatidylinositol; PS–phosphatidylserine; PE–phosphatidylethanolamine; PMME–monomethyl-phosphatidylethanolamine; PDME–dimethyl-phosphatidylethanolamine; PC–phosphatidylcholine; CL–cardiolipin. D. Average (n = 4) change in OD600 of ira2Δ and bcy1Δ grown anaerobically in rich xylose medium either in the absence (solid lines) or presence (dashed lines, IC) of inositol (75 μM) and choline (10 mM) (* indicates p = 2.4 x 10−6, ANOVA). https://doi.org/10.1371/journal.pgen.1010593.g004 Download: PPT PowerPoint slide PNG larger image TIFF original image Table 1. Phosphorylation changes of phospholipid biosynthetic enzymes. https://doi.org/10.1371/journal.pgen.1010593.t001 We next compared lipidomic profiles in the growing ira2Δ strain shifted to xylose to the bcy1Δ and ira2Δbcy1Δ strains that do not grow. Due to limited statistical power (caused by replicate variation in one of the three bcy1Δ strain replicates), we compared the ira2Δ response to ira2Δbcy1Δ cells, whose response was highly similar to two out of the three bcy1Δ strain replicates. One caveat of this analysis is that the ira2Δbcy1Δ strain displays a variable anaerobic-xylose fermentation profile; nonetheless, given the similarity to bcy1Δ phenotypes, the high reproducibility of the double mutant’s transcriptomic and lipidomic profiles suggests a good representation of hyper-active PKA signaling. It is possible that this analysis may miss some lipidomic changes related to the variation in ira2Δbcy1Δ metabolism profiles. Even so, we identified 67 lipids whose fold-change was significantly different in ira2Δbcy1Δ cells upon xylose shift versus ira2Δ cells (FDR < 0.05, Fig 4B and S12 Table). The analysis confidently classified six of the lipids, including phosphatidylethanolamines (PE), phosphatidyl dimethylethanolamines (PDME), and cardiolipins (CL). PE and multiple PDME species were more abundant in the ira2Δbcy1Δ strain exposed to the shift compared to ira2Δ cells (FDR < 0.05, Fig 4B). These differences were particularly interesting because PE is further metabolized to PDME and then to phosphatidylcholine (PC), the most abundant phospholipid in the cell, through three consecutive methylation reactions by Cho2 and Opi3, respectively (Fig 4C) [50]. While the CHO2 transcript was not differentially expressed between ira2Δ and bcy1Δ strains, OPI3 was: ira2Δ cells shifted to xylose induced OPI3 expression, whereas bcy1Δ and ira2Δbcy1Δ cells repressed it (FDR = 2.45x10-12 and FDR = 6.22x10-13, respectively). Previous studies suggest that blocking PC synthesis through OPI3 deletion, but not CHO2 deletion, inhibits growth due to the accumulation of phosphatidyl monomethylethanolamine (PMME) and insufficient PC production [51]. To investigate effects on PC, we analyzed all PC lipid moieties in the dataset; PC lipids were reproducibly lower in abundance after the xylose shift in bcy1Δ cells when compared to ira2Δ cells (p = 0.000419, ANOVA; S4 Fig and S13 Table). We propose that the bcy1Δ strain experiences a bottleneck in that pathway leading to PC synthesis from PE, which may impact its ability to grow on xylose (see Discussion). Among other lipids whose abundance was influenced by BCY1 deletion and xylose shift was cardiolipin, a major component of mitochondrial membranes critical for a variety of functions including acetyl coA synthesis, TCA cycle, iron metabolism, arginine metabolism, and protein import [52]. Interestingly, cardiolipin abundance was reduced in the ira2Δbcy1Δ strain upon xylose shift compared to ira2Δ cells. The difference is underscored by transcriptomic differences, since several cardiolipin biosynthetic genes were induced in ira2Δ cells but repressed or induced to a weaker extent in bcy1Δ and ira2Δbcy1Δ strains (FDR < 0.05). Additionally, production of PS, PE, and PC is dependent on properly functioning mitochondrial membranes as PS is shuttled into the mitochondria and converted to PE by the phosphatidylserine decarboxylase Psd1, before PE is shuttled back to the ER. Thus, the effects of cardiolipin reduction in bcy1Δ strains are further compounded by impacting other branches of phospholipid biosynthesis. We expected to see differential abundance of PA in ira2Δbcy1Δ cells versus ira2Δ cells, since bcy1Δ and ira2Δbcy1Δ strains induced some PA biosynthesis genes whereas ira2Δ cells do not (Fig 3A). While there were no significant differences in PA moieties between the strains (FDR > 0.05), we did identify altered phosphorylation status of the PA phosphatase enzyme Pah1 (S823; Table 1). Pah1 converts PA to diacylglycerol, which is funneled into storage lipids [50]. Phosphorylation of serine 823 is significantly lower in the bcy1Δ and ira2Δbcy1Δ strains compared to the ira2Δ strain (log2(fold change) < -1). Interestingly, this serine has not been previously annotated as a phosphorylated residue (BioGRID version 4.4.213) [53], but it is within a potential PKA consensus site (RRxxS/T). PKA is known to phosphorylate Pah1 at another residue not captured in our dataset to inhibit its activity [54]. Our results raise the possibility that S823 regulates Pah1 activity in a manner that affects PA in these strains. Overall, the differences seen in PE, PDME, and PC abundances, as well as differences in transcript abundance and phosphorylation status of phospholipid biosynthesis enzymes, suggest a bottleneck in the pathway in the bcy1Δ strains that may inhibit their ability to proliferate on xylose (see Discussion). Supplementation with phospholipid precursors only modestly improves growth We questioned if supplementing xylose medium with phospholipid precursors, particularly inositol and choline that can be funneled into phospholipid biosynthesis via the Kennedy Pathway, may bypass a possible bottleneck and thus rescue the bcy1Δ strain’s growth. We therefore grew bcy1Δ and ira2Δ strains anaerobically in xylose medium with and without choline and inositol supplementation (we included inositol since the INO1 gene is repressed in bcy1Δ cells) (Fig 4C and S5 Table). After 52 hours of growth in supplementation, bcy1Δ cells experienced a very modest but statistically significant growth improvement (p = 2.4 x 10−6, ANOVA; Fig 4D), whereas the ira2Δ strain did not. While the bcy1Δ strain’s inability to grow anaerobically on xylose cannot be fully explained by a deficiency in phospholipid precursors, the modest improvement implicates it as a contributing factor to the phenotype. Growth and metabolism can be genetically recoupled through directed evolution We took a second approach to identify pathways and processes responsible for growth coordination in bcy1Δ strains by conducting adaptive laboratory evolutions to recouple xylose-dependent growth and metabolism. The bcy1Δ strain was first grown anaerobically in rich medium supplemented with 2% glucose to accumulate mutations [55], then the culture was seeded into fresh anaerobic medium containing 2% xylose and 0.1% glucose and passaged periodically for ~35 generations until the culture showed robust changes in cellular density over time (see Methods). Single colonies were isolated and characterized for their growth and fermentation capabilities, and genetic changes were identified through whole genome sequencing (see Methods). Three independent evolutions were performed, and several colonies were selected at different stages of the evolutions. In all three experiments, we identified mutants with recoupled growth and metabolism despite the absence of BCY1, evident by their robust anaerobic growth on xylose medium compared to the ira2Δ strain (Figs 5A, S5 and S6). Interrogating the genome sequences identified multiple mutations in each strain, along with copy-number variations and aneuploidy in several of the evolved lines (Table 2). Only evolved mutations impacting the coding sequence of a gene were analyzed further. Interestingly, there was no genetic change common to all evolved strains, strongly suggesting multiple routes to recoupling growth and metabolism in the absence of BCY1. Four of the characterized strains from the three experiments regained growth rates comparable to and statistically indistinguishable from ira2Δ cells (p > 0.05, ANOVA), including EWY55 from the first culture, EWY87-1 and EWY87-3 from the second culture, and EWY89-3 from the third evolution culture (S6A and S6B, S6E Fig). Strains EWY89-1 and EWY89-2 showed modest growth on xylose but did not differ significantly from the bcy1Δ strain (p > 0.05, ANOVA; S6C and S6D Fig). Genetic changes for all evolved strains are listed in Table 2. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 5. Directed evolution recoupled growth and metabolism on xylose. A. Average (n = 3) change in OD600 of ira2Δ, bcy1Δ, and EWY55 strains grown anaerobically in rich xylose medium (*, p < 10−4, ANOVA; n.s., not significant). B. Change in OD600 (left panel) and xylose concentration (right panel) over 48 hours of EWY55 and EWY55 opi1Δ strains grown anaerobically on rich xylose medium. (*, p < 0.05, ANOVA). C. Expression of 233 genes whose transcript abundance during growth on xylose was significantly different in EWY55 and/or ira2Δ strains compared to the bcy1Δ strain (FDR < 0.05), visualized by hierarchical clustering. Data represent the log2 transcript abundance in each strain grown anaerobically in xylose compared to bcy1Δ strain. Cluster A (9 genes) and B (13 genes) are annotated, see text for details. D. Bar plot of the average and standard deviation log2(fold change) (n = 3) in lipid abundance of key lipids with reproducible differences 1.5-fold or greater in EWY55 compared to ira2Δ or bcy1Δ strains. Asterisks denote significant differences by ANOVA. https://doi.org/10.1371/journal.pgen.1010593.g005 Download: PPT PowerPoint slide PNG larger image TIFF original image Table 2. Genetic changes in evolved bcy1Δ strains. https://doi.org/10.1371/journal.pgen.1010593.t002 Strain EWY55 was particularly interesting. This strain harbored nonsynonymous mutations in several genes, including PKA catalytic subunit TPK1, the negative regulator of phospholipid genes, OPI1, described above, RIM8 that is required for anaerobic growth [56], and TFIIA large subunit TOA1 (Table 2). The OPI1 mutation was especially interesting because Opi1 was implicated in the phospholipid transcriptomic analysis above (Fig 3) and because the mutation changes a known phosphorylation site, serine 239, to alanine (Table 2). CKII has been reported to phosphorylate this site and was previously shown to activate Opi1 [57]. This poses the question of whether Opi1 is aberrantly regulated in the bcy1Δ strain, and if this is responsible for its lack of growth on xylose. To identify causal alleles responsible for recoupling growth and metabolism in the EWY55 strain, we performed single gene deletions and allele swaps in the bcy1Δ and EWY55 strains (see Methods). This strain background is derived from a wild isolate that is less genetically amenable than laboratory strains [58], and we were unable to recover TPK1 deletion in either strain despite many efforts. Deletion of RIM8 or TOA1 did not impact the growth of EWY55 cells, nor did substituting the parental alleles into the evolved strain (S5C Fig). However, deletion of the evolved opi1 gene partially but significantly reduced anaerobic xylose growth of the EWY55 strain in liquid medium (Fig 5B). Importantly, the strain retained robust xylose fermentation, indicating that Opi1 plays a role in the coupling of growth and metabolism (see Discussion). Complementation experiments to swap strain alleles were not successful, since introducing even the empty vector into this strain complemented anaerobic xylose growth on a plate for reasons that are not clear but may suggest that the cells grow differently during drug selection (S5C Fig). While we cannot be sure OPI1 is the causal gene, our results indicate that the genetics modulating this trait is complex and may result from different evolutionary paths, but at least in EWY55 is likely to include a role for evolved Opi1 function. Transcriptomic and lipidomic analysis in the evolved strain reveals altered phospholipids To further characterize the evolved EWY55 strain, we performed another transcriptomic and lipidomic experiment as described above (see Methods) with the main goal of identifying if the evolved EWY55 strain reverted its gene expression and lipid composition to that of the ira2Δ strain. Surprisingly, the EWY55 strain did not recapitulate the ira2Δ gene expression or lipid abundance profiles at most entities. We identified 297 transcripts less abundant in EWY55 growing anaerobically on xylose compared to the bcy1Δ strain (FDR < 0.05; S14 Table), and these were enriched for genes involved in mitochondrial functions, such as electron transport chain, oxidation-reduction, and targets of the HAP2/3/4/5 complex; genes involved in phospholipid metabolism; and genes involved in ergosterol synthesis (p < 0.05, hypergeometric test, see Methods). Many of these processes were significantly affected in our original comparison of the bcy1Δ and ira2Δ strains. Additionally, 93 genes with higher abundance in the EWY55 compared to the bcy1Δ cells (FDR < 0.05; S14 Table) were enriched for ribosomal protein genes and genes involved in translation and sulfate assimilation (p < 10−7, hypergeometric test), processes important for rapid growth. Since EWY55 cells recapitulated the xylose-dependent growth seen in ira2Δ cells, we next asked if its expression changes recapitulated ira2Δ patterns relative to bcy1Δ cells–surprisingly, most did not (Fig 5C). This indicates that the evolved EWY55 did not recouple growth and metabolism under anaerobic xylose conditions via reverting to the ira2Δ strain’s expression patterns. There were a few exceptions, including 22 transcripts of diverse functions (Fig 5C, Clusters A and B) in which expression differences in EWY55 recapitulated those seen in ira2Δ cells compared to the bcy1Δ strain (S15 Table; FDR < 0.05). While the role of these expression changes will require future study, it is intriguing that these clusters included several targets of the glucose-responsive transcription factor Rgt1 and the Sok2 regulator that responds to starvation and hypoxia; both genes have connections to PKA signaling [59–62]. We were particularly interested in phospholipid biosynthesis genes, given all the connections to this pathway throughout our studies. In general, EWY55 cells showed lower transcript abundances of phospholipid biosynthesis genes compared to the bcy1Δ strain grown anaerobically on xylose (S16 Table), making its expression even more divergent from the ira2Δ strain. The phospholipid composition further supports the unique changes of the EWY55 strain that permit recoupled growth and metabolism on xylose. The EWY55 strain showed significantly greater abundance of the storage lipid triacylglycerol (TG; Fig 5D and S17 Table; p < 10−4, ANOVA). Importantly, EWY55 had significantly lower levels of PDME and trended towards higher levels of PC compared to bcy1Δ cells, recapitulating the pattern in ira2Δ cells (Fig 5D and S17 Table). These results are consistent with the hypothesis that the evolved EWY55 strain altered the pathway compared to bcy1Δ cells. Together, our results underscore the complexity of responses to xylose growth and metabolism across ira2Δ cells, the parental bcy1Δ strain, and EWY55 (see Discussion). Few gene expression patterns correlate strictly with growth phenotypes Somewhat surprisingly, there were few genes whose expression correlated strictly with growth phenotypes. Only two genes showed xylose-responsive expression changes that were specific to ira2Δ cells compared to the other three strains analyzed as a group in the statistical model (FDR < 0.05; see Methods): daughter-cell-specific glucanase DSE4 and L-homoserine-O-acetyltransferase MET2. In fact, hierarchical clustering of all genes with a transcriptomic change in response to the carbon shift showed that the ira2Δ strain’s response to xylose shift was most similar to that of Y184 cells, even though one strain can grow on and anaerobically ferment xylose and the other cannot (S2A Fig and S1 Table). We next performed pairwise comparisons of the glucose-to-xylose fold-change responses between each strain and the ira2Δ strain, then combined the lists of genes identified in all three comparisons. This method identified 65 genes; however, investigating the expression patterns once again indicated that the ira2Δ response was most similar to Y184 cells but with weaker magnitudes of change (Fig 2B and S2 Table). This set of 65 genes was enriched for genes induced in the environmental stress response (iESR genes [36], p = 2x10-7, hypergeometric test). Many genes induced in the Y184 and ira2Δ strains, but largely not in bcy1Δ strains, included genes related to metabolism, including several in the mitochondrial TCA cycle and peroxisomal fatty-acid oxidation pathway, which may reflect that bcy1Δ strains are more likely to recognize xylose as a fermentable carbon source. We specifically investigated the set of 65 genes for those that encode cell-cycle regulators and kinases, since these may be involved in growth kinetics; however only three, six, or eight genes within these categories were differentially expressed in Y184, bcy1Δ, or ira2Δbcy1Δ cells, respectively, compared to ira2Δ cells in response to xylose shift (FDR < 0.05). The ira2Δ strain showed weakly lower expression of cyclin CLN2 and anaphase-promoting complex CDC20, whereas other strains showed strong reduction in expression (S3 Table). Transcript abundance for these genes is known to fluctuate during the cell cycle, thus while it is possible their expression influences growth arrest, it is likely that the expression of these genes reflects the expected difference between cycling (ira2Δ) and non-cycling (Y184, bcy1Δ, ira2Δbcy2Δ) cells. Expression of cell-cycle genes did not implicate arrest in a particular cell-cycle stage, consistent with early transcriptomic studies that showed that gene expression during cell-cycle arrest does not parallel expression of cells cycling through those phases [37]. Previous chemostat studies reported that repression of ribosomal protein (RP) and ribosome biogenesis (RiBi) genes is correlated with decreased growth, and these studies proposed that expression of these genes can predict cellular growth rate [38–41]. However, here we saw no correlation of RP and RiBi transcript abundance or response with growth phenotypes. The Y184 strain strongly repressed RP and RiBi genes upon xylose shift, which might be expected for a strain that arrests its growth, but so too did the ira2Δ strain, albeit with weaker magnitude of repression. Surprisingly, bcy1Δ and ira2Δbcy1Δ cells, whose growth is largely arrested after the xylose shift, showed little change in RP and RiBi transcripts compared to glucose-dependent growth (FDR <0.05, S2B Fig and S4 Table). These results reinforce past work from our lab that the expression of ribosome-associated genes does not necessarily parallel growth rate [42]. They further suggest that bcy1Δ strains cultured in xylose are unlikely limited by the abundance of RP and RiBi transcripts. Overall, while the non-growing strains have stronger repression of a few cell-cycle regulators when compared to the ira2Δ strain, there was not a clear gene expression pattern to describe why ira2Δ cells grow and bcy1Δ strains do not. Few gene expression patterns correlate strictly with metabolism phenotypes We next investigated shared gene expression changes related to robust anaerobic xylose fermentation. We compared expression in the Y184 strain responding to the xylose shift to the ira2Δ and bcy1Δ strains analyzed as a single group in the statistical model (we excluded the ira2Δbcy1Δ strain due to the variability of its fermentation phenotype, although it is capable of anaerobic xylose fermentation). This identified 292 differentially expressed genes (FDR < 0.05; S5 Table). Hierarchical clustering revealed that these genes typically had larger expression changes in Y184 and that those expression changes were progressively weaker across the strain series; once again, Y184 and the ira2Δ strain were more similar to one another than they were to the bcy1Δ strain (Fig 2C and S5 Table). Collectively, these genes were heavily enriched for genes in the ESR (p = 3.624x10-10, hypergeometric test). Deeper interrogation revealed several small gene clusters of interest. Cluster I contained 19 genes induced in the ira2Δ and bcy1Δ strains but repressed in the Y184 strain. This group did not contain any functional enrichments; however, proteins encoded by several of these genes localize to the endoplasmic reticulum. Cluster II contained 31 genes induced in Y184 and either unchanged or repressed in both ira2Δ and bcy1Δ strains. This group was enriched for genes involved in protein folding (p = 9.49x10-6, hypergeometric test) and included the Hsp90 chaperone and cochaperone genes HSP82, STI1, and AHA1, as well as the mitochondrial matrix protein chaperone HSP10. Hsp90 can act as a signal transducer for alternative carbon source metabolism [43], again suggesting that Y184 does not recognize xylose as a fermentable carbon source. We specifically interrogated the 292 genes for those involved in glycolysis, gluconeogenesis, TCA cycle, and carbohydrate storage, predicting that differences in expression would relate to altered xylose metabolism capabilities. This identified 15 genes with functional annotations linked to at least one of these processes (Fig 2D and S6 Table). The xylose fermenting strains shared expression at several hallmark genes. For example, Y184 cells strongly induced hexose transporter HXT5, normally induced by non-fermentable carbon sources, whereas the three mutant strains showed a weaker induction (FDR = 4.25x10-5). The glucose-repressed aldehyde dehydrogenease ALD2 was induced in Y184 and either did not change (ira2Δ cells) or was repressed (bcy1Δ cells) upon the switch to xylose (FDR = 4.32x107). Furthermore, glucose-induced transcriptional repressor MIG2 showed stronger induction in the xylose fermenting strains, and especially bcy1Δ strains, compared to Y184 (FDR = 0.047). These data are all consistent with the hypothesis that the xylose fermenters recognize xylose as a fermentable carbon, whereas Y184 activates a carbon-starvation response. Regulatory analysis reveals strain-specific differences in carbon, iron, and lipid gene control We next focused on understanding how growth and metabolism are decoupled in the bcy1Δ strain, and we thus directly compared its expression to that in ira2Δ cells. We focused on genes whose expression changes in response to the xylose shift were in opposing directions to implicate processes involved in decoupling growth and metabolism (Fig 3A and S7 Table, see Methods). Among the identified genes, we scored enrichment of functional terms as well as known targets of transcriptional regulators (S8 Table). We also used motif analysis to discover shared sequence motifs upstream of genes uniquely induced or repressed in the bcy1Δ strain, and then matched those to known transcription factor binding sites (see Methods). We identified 654 genes differentially expressed in bcy1Δ cells and with a fold-change in the opposite direction as ira2Δ cells upon the glucose-to-xylose shift (Fig 3A and S7 Table). Importantly, only 82 genes (12.5%) showed significant differences in basal gene expression when cells were grown on glucose (S3A Fig), indicating that the majority of genes are identified due to differences in response to xylose shift. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 3. Genes uniquely expressed in the bcy1Δ strain implicate an integrated response to xylose metabolism and growth coupling. A. Expression of 654 genes whose log2(fold change) upon glucose to xylose shift is different (FDR < 0.05) between the ira2Δ and bcy1Δ strains and whose expression change is in the opposite direction across the two strains (see Methods for details). Significant functional enrichments are annotated next to the two main clusters (p < 10−4, hypergeometric test). Bar graph inset represents the log2(fold change) of the two phosphatidic acid biosynthesis enzymes in this group, see text for details. B. Regulatory relationships between transcription factors whose targets or known binding sites were enriched in (A). Documented PKA-dependent phosphorylation is indicated by a P. See text for details. https://doi.org/10.1371/journal.pgen.1010593.g003 The results implicated several regulators, some with prior connections to anaerobic xylose fermentation. 318 genes induced in the bcy1Δ strain shifted to xylose, but repressed in the ira2Δ cells, were enriched for amino acid and sphingolipid biosynthesis genes, as well as targets of the carbon-responsive Azf1 transcription factor (p < 10−4, hypergeometric test). Previous work from our lab implicated Azf1 in anaerobic xylose fermentation, and indeed, we showed that the over-expression of AZF1 in an ira2Δ strain enhances the rate of anaerobic xylose utilization [32]. Additionally, PKA has been implicated in Azf1 phosphorylation [44]; together with the fact that the AZF1 gene is uniquely induced in the bcy1Δ strain suggest its functional importance in xylose metabolism (see Discussion). In contrast, several regulators were implicated by the 336 genes uniquely repressed in the bcy1Δ strain. These included genes harboring upstream binding sites of the iron-responsive Aft1/2 transcription factors (S3B Fig) and known targets of transcriptional activator Ino4 that responds to inositol for phospholipid biosynthesis (Fig 3A; see more below). Iron is an important cofactor of many enzymes, including those involved in mitochondrial respiration, lipid biogenesis, and amino acid biosynthesis, all of whose genes were among the differentially regulated genes studied here. Additionally, Aft1/2 regulation and the iron regulon have been linked with PKA activity; however, direct interactions remain to be identified [45]. Interestingly, Aft1/2 and Azf1 both are both connected to the regulator Mga2, which controls lipid and hypoxia genes and that we previously showed enhances anaerobic xylose fermentation when over-expressed in ira2Δ cells [32,46] (see Discussion). Targets of the Ino2/4 regulators that respond to inositol for phospholipid biosynthesis were also present in this gene set; while a majority of the targets identified here were repressed in the bcy1Δ strain, some of the known targets were repressed in the ira2Δ strain but induced in the bcy1Δ mutant (S3C Fig and S9 Table). This may reflect the complexities of the genes’ regulation by other factors. Nonetheless, Ino2/4 targets were enriched among the genes oppositely regulated in the bcy1Δ versus ira2Δ strain. Overall, these results provide an interesting link between PKA signaling, carbon and iron responses, and lipid metabolism (Fig 3B). The presence of many lipid biosynthesis genes in this gene set and the highly regulated role of lipids in cell growth and proliferation prompted a deeper investigation of lipid metabolism genes. The bcy1Δ strain repressed genes involved in ergosterol biosynthesis and some targets of Ino2/4 that are involved in phospholipid metabolism (Figs 3A and S3C). This response is consistent with the model that Ino2/4 activity is reduced. However, the bcy1Δ strain also induced some genes involved in the synthesis of phosphatidic acid (PA) (Fig 3A inset), which normally promotes Ino4 activity by sequestering Ino4’s inhibitor Opi1 to the ER membrane [47]. This response suggests that some connection between PA, Opi1, and Ino4 is disrupted in the absence of BCY1. PKA is known to regulate the Ino2/4 pathway through direct phosphorylation of Opi1 to increase its inhibitory activity [48]. Together, these results raised the possibility that the bcy1Δ strain has important differences in lipid metabolism and perhaps composition, which could be modulated by differences in PKA activity in this strain. Lipidomic and phosphoproteomic analyses show disrupted phospholipid metabolism in bcy1Δ cells Since the transcriptomic responses implicated differences in lipid metabolism, we investigated the lipidomic composition of our strains. Strains were grown in a similar design as the transcriptomic analysis, where anaerobically glucose-grown Y184, ira2Δ, bcy1Δ, and ira2Δbcy1Δ cells were shifted to anaerobic xylose media for three hours before lipids were analyzed by mass spectrometry (see Methods). We detected over 4000 lipid species including 239 that were confidently assigned to a particular lipid class (S10 Table). All detected lipid species were included in the statistical analysis to obtain a wholistic understanding of lipidome differences between the strains. We again sought to find lipidomic profiles correlated with xylose metabolism and growth, xylose metabolism but no growth, and no xylose metabolism or growth. We compared the Y184 strain to the three strains with upregulated PKA activity and identified 18 lipids whose change in abundance upon a shift to xylose significantly differed in Y184 cells. This group included phosphatidylserine (PS) species (Fig 4A and S11 Table). Interestingly, all three mutants increased the abundance of these PS species when shifted to xylose, whereas Y184 cells decreased the abundance of one and failed to induce the other to the same degree as the mutants. The gene encoding the PS synthase CHO1 was strongly induced in Y184 cells, indicating that the decrease in PS in Y184 cells is unlikely due to decreased CHO1 expression. Instead, we analyzed previous phosphoproteomic data from our lab and discovered that Cho1 was phosphorylated to a much higher degree in the Y184 strain on serine 46 (|log2FC| > 1, Table 1), a known PKA site that inhibits Cho1 activity [49]. Together, these results indicate PKA-dependent inhibition of PS synthesis in Y184 cells. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 4. bcy1Δ strains show altered phospholipids after anaerobic xylose shift. A-B. Abundance of lipids (rows) with a significant difference in log2(fold change) upon anaerobic glucose-to-xylose shift in (A) Y184 compared to PKA pathway mutants (ira2Δ, bcy1Δ, ira2Δbcy1Δ) analyzed as a group in the statistical model or (B) ira2Δ cells compared to ira2Δbcy1Δ cells. Lipids of interest are annotated. C. Partial phospholipid biosynthesis pathway with transcriptomic and lipidomic data represented. Yellow-blue boxes next to each enzyme name represent the average log2(fold change) in transcript abundance upon glucose-to-xylose shift for each strain, as outlined in the key. Significant differences compared to the ira2Δ strain (FDR < 0.05) are represented in sharp, bolded boxes, whereas insignificant differences are translucent. Colorized pathway arrows (yellow: induced, blue: repressed) represent the predominant transcript patterns for that enzymatic step when comparing the bcy1Δ and ira2Δ strains. Lipids whose fold-change in abundance is different in specific strains are according to the key. Lipid abbreviations: FFA–free fatty acids; PA–phosphatidic acid; DG–diacylglycerol; TG–triacylglycerol; PI–phosphatidylinositol; PS–phosphatidylserine; PE–phosphatidylethanolamine; PMME–monomethyl-phosphatidylethanolamine; PDME–dimethyl-phosphatidylethanolamine; PC–phosphatidylcholine; CL–cardiolipin. D. Average (n = 4) change in OD600 of ira2Δ and bcy1Δ grown anaerobically in rich xylose medium either in the absence (solid lines) or presence (dashed lines, IC) of inositol (75 μM) and choline (10 mM) (* indicates p = 2.4 x 10−6, ANOVA). https://doi.org/10.1371/journal.pgen.1010593.g004 Download: PPT PowerPoint slide PNG larger image TIFF original image Table 1. Phosphorylation changes of phospholipid biosynthetic enzymes. https://doi.org/10.1371/journal.pgen.1010593.t001 We next compared lipidomic profiles in the growing ira2Δ strain shifted to xylose to the bcy1Δ and ira2Δbcy1Δ strains that do not grow. Due to limited statistical power (caused by replicate variation in one of the three bcy1Δ strain replicates), we compared the ira2Δ response to ira2Δbcy1Δ cells, whose response was highly similar to two out of the three bcy1Δ strain replicates. One caveat of this analysis is that the ira2Δbcy1Δ strain displays a variable anaerobic-xylose fermentation profile; nonetheless, given the similarity to bcy1Δ phenotypes, the high reproducibility of the double mutant’s transcriptomic and lipidomic profiles suggests a good representation of hyper-active PKA signaling. It is possible that this analysis may miss some lipidomic changes related to the variation in ira2Δbcy1Δ metabolism profiles. Even so, we identified 67 lipids whose fold-change was significantly different in ira2Δbcy1Δ cells upon xylose shift versus ira2Δ cells (FDR < 0.05, Fig 4B and S12 Table). The analysis confidently classified six of the lipids, including phosphatidylethanolamines (PE), phosphatidyl dimethylethanolamines (PDME), and cardiolipins (CL). PE and multiple PDME species were more abundant in the ira2Δbcy1Δ strain exposed to the shift compared to ira2Δ cells (FDR < 0.05, Fig 4B). These differences were particularly interesting because PE is further metabolized to PDME and then to phosphatidylcholine (PC), the most abundant phospholipid in the cell, through three consecutive methylation reactions by Cho2 and Opi3, respectively (Fig 4C) [50]. While the CHO2 transcript was not differentially expressed between ira2Δ and bcy1Δ strains, OPI3 was: ira2Δ cells shifted to xylose induced OPI3 expression, whereas bcy1Δ and ira2Δbcy1Δ cells repressed it (FDR = 2.45x10-12 and FDR = 6.22x10-13, respectively). Previous studies suggest that blocking PC synthesis through OPI3 deletion, but not CHO2 deletion, inhibits growth due to the accumulation of phosphatidyl monomethylethanolamine (PMME) and insufficient PC production [51]. To investigate effects on PC, we analyzed all PC lipid moieties in the dataset; PC lipids were reproducibly lower in abundance after the xylose shift in bcy1Δ cells when compared to ira2Δ cells (p = 0.000419, ANOVA; S4 Fig and S13 Table). We propose that the bcy1Δ strain experiences a bottleneck in that pathway leading to PC synthesis from PE, which may impact its ability to grow on xylose (see Discussion). Among other lipids whose abundance was influenced by BCY1 deletion and xylose shift was cardiolipin, a major component of mitochondrial membranes critical for a variety of functions including acetyl coA synthesis, TCA cycle, iron metabolism, arginine metabolism, and protein import [52]. Interestingly, cardiolipin abundance was reduced in the ira2Δbcy1Δ strain upon xylose shift compared to ira2Δ cells. The difference is underscored by transcriptomic differences, since several cardiolipin biosynthetic genes were induced in ira2Δ cells but repressed or induced to a weaker extent in bcy1Δ and ira2Δbcy1Δ strains (FDR < 0.05). Additionally, production of PS, PE, and PC is dependent on properly functioning mitochondrial membranes as PS is shuttled into the mitochondria and converted to PE by the phosphatidylserine decarboxylase Psd1, before PE is shuttled back to the ER. Thus, the effects of cardiolipin reduction in bcy1Δ strains are further compounded by impacting other branches of phospholipid biosynthesis. We expected to see differential abundance of PA in ira2Δbcy1Δ cells versus ira2Δ cells, since bcy1Δ and ira2Δbcy1Δ strains induced some PA biosynthesis genes whereas ira2Δ cells do not (Fig 3A). While there were no significant differences in PA moieties between the strains (FDR > 0.05), we did identify altered phosphorylation status of the PA phosphatase enzyme Pah1 (S823; Table 1). Pah1 converts PA to diacylglycerol, which is funneled into storage lipids [50]. Phosphorylation of serine 823 is significantly lower in the bcy1Δ and ira2Δbcy1Δ strains compared to the ira2Δ strain (log2(fold change) < -1). Interestingly, this serine has not been previously annotated as a phosphorylated residue (BioGRID version 4.4.213) [53], but it is within a potential PKA consensus site (RRxxS/T). PKA is known to phosphorylate Pah1 at another residue not captured in our dataset to inhibit its activity [54]. Our results raise the possibility that S823 regulates Pah1 activity in a manner that affects PA in these strains. Overall, the differences seen in PE, PDME, and PC abundances, as well as differences in transcript abundance and phosphorylation status of phospholipid biosynthesis enzymes, suggest a bottleneck in the pathway in the bcy1Δ strains that may inhibit their ability to proliferate on xylose (see Discussion). Supplementation with phospholipid precursors only modestly improves growth We questioned if supplementing xylose medium with phospholipid precursors, particularly inositol and choline that can be funneled into phospholipid biosynthesis via the Kennedy Pathway, may bypass a possible bottleneck and thus rescue the bcy1Δ strain’s growth. We therefore grew bcy1Δ and ira2Δ strains anaerobically in xylose medium with and without choline and inositol supplementation (we included inositol since the INO1 gene is repressed in bcy1Δ cells) (Fig 4C and S5 Table). After 52 hours of growth in supplementation, bcy1Δ cells experienced a very modest but statistically significant growth improvement (p = 2.4 x 10−6, ANOVA; Fig 4D), whereas the ira2Δ strain did not. While the bcy1Δ strain’s inability to grow anaerobically on xylose cannot be fully explained by a deficiency in phospholipid precursors, the modest improvement implicates it as a contributing factor to the phenotype. Growth and metabolism can be genetically recoupled through directed evolution We took a second approach to identify pathways and processes responsible for growth coordination in bcy1Δ strains by conducting adaptive laboratory evolutions to recouple xylose-dependent growth and metabolism. The bcy1Δ strain was first grown anaerobically in rich medium supplemented with 2% glucose to accumulate mutations [55], then the culture was seeded into fresh anaerobic medium containing 2% xylose and 0.1% glucose and passaged periodically for ~35 generations until the culture showed robust changes in cellular density over time (see Methods). Single colonies were isolated and characterized for their growth and fermentation capabilities, and genetic changes were identified through whole genome sequencing (see Methods). Three independent evolutions were performed, and several colonies were selected at different stages of the evolutions. In all three experiments, we identified mutants with recoupled growth and metabolism despite the absence of BCY1, evident by their robust anaerobic growth on xylose medium compared to the ira2Δ strain (Figs 5A, S5 and S6). Interrogating the genome sequences identified multiple mutations in each strain, along with copy-number variations and aneuploidy in several of the evolved lines (Table 2). Only evolved mutations impacting the coding sequence of a gene were analyzed further. Interestingly, there was no genetic change common to all evolved strains, strongly suggesting multiple routes to recoupling growth and metabolism in the absence of BCY1. Four of the characterized strains from the three experiments regained growth rates comparable to and statistically indistinguishable from ira2Δ cells (p > 0.05, ANOVA), including EWY55 from the first culture, EWY87-1 and EWY87-3 from the second culture, and EWY89-3 from the third evolution culture (S6A and S6B, S6E Fig). Strains EWY89-1 and EWY89-2 showed modest growth on xylose but did not differ significantly from the bcy1Δ strain (p > 0.05, ANOVA; S6C and S6D Fig). Genetic changes for all evolved strains are listed in Table 2. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 5. Directed evolution recoupled growth and metabolism on xylose. A. Average (n = 3) change in OD600 of ira2Δ, bcy1Δ, and EWY55 strains grown anaerobically in rich xylose medium (*, p < 10−4, ANOVA; n.s., not significant). B. Change in OD600 (left panel) and xylose concentration (right panel) over 48 hours of EWY55 and EWY55 opi1Δ strains grown anaerobically on rich xylose medium. (*, p < 0.05, ANOVA). C. Expression of 233 genes whose transcript abundance during growth on xylose was significantly different in EWY55 and/or ira2Δ strains compared to the bcy1Δ strain (FDR < 0.05), visualized by hierarchical clustering. Data represent the log2 transcript abundance in each strain grown anaerobically in xylose compared to bcy1Δ strain. Cluster A (9 genes) and B (13 genes) are annotated, see text for details. D. Bar plot of the average and standard deviation log2(fold change) (n = 3) in lipid abundance of key lipids with reproducible differences 1.5-fold or greater in EWY55 compared to ira2Δ or bcy1Δ strains. Asterisks denote significant differences by ANOVA. https://doi.org/10.1371/journal.pgen.1010593.g005 Download: PPT PowerPoint slide PNG larger image TIFF original image Table 2. Genetic changes in evolved bcy1Δ strains. https://doi.org/10.1371/journal.pgen.1010593.t002 Strain EWY55 was particularly interesting. This strain harbored nonsynonymous mutations in several genes, including PKA catalytic subunit TPK1, the negative regulator of phospholipid genes, OPI1, described above, RIM8 that is required for anaerobic growth [56], and TFIIA large subunit TOA1 (Table 2). The OPI1 mutation was especially interesting because Opi1 was implicated in the phospholipid transcriptomic analysis above (Fig 3) and because the mutation changes a known phosphorylation site, serine 239, to alanine (Table 2). CKII has been reported to phosphorylate this site and was previously shown to activate Opi1 [57]. This poses the question of whether Opi1 is aberrantly regulated in the bcy1Δ strain, and if this is responsible for its lack of growth on xylose. To identify causal alleles responsible for recoupling growth and metabolism in the EWY55 strain, we performed single gene deletions and allele swaps in the bcy1Δ and EWY55 strains (see Methods). This strain background is derived from a wild isolate that is less genetically amenable than laboratory strains [58], and we were unable to recover TPK1 deletion in either strain despite many efforts. Deletion of RIM8 or TOA1 did not impact the growth of EWY55 cells, nor did substituting the parental alleles into the evolved strain (S5C Fig). However, deletion of the evolved opi1 gene partially but significantly reduced anaerobic xylose growth of the EWY55 strain in liquid medium (Fig 5B). Importantly, the strain retained robust xylose fermentation, indicating that Opi1 plays a role in the coupling of growth and metabolism (see Discussion). Complementation experiments to swap strain alleles were not successful, since introducing even the empty vector into this strain complemented anaerobic xylose growth on a plate for reasons that are not clear but may suggest that the cells grow differently during drug selection (S5C Fig). While we cannot be sure OPI1 is the causal gene, our results indicate that the genetics modulating this trait is complex and may result from different evolutionary paths, but at least in EWY55 is likely to include a role for evolved Opi1 function. Transcriptomic and lipidomic analysis in the evolved strain reveals altered phospholipids To further characterize the evolved EWY55 strain, we performed another transcriptomic and lipidomic experiment as described above (see Methods) with the main goal of identifying if the evolved EWY55 strain reverted its gene expression and lipid composition to that of the ira2Δ strain. Surprisingly, the EWY55 strain did not recapitulate the ira2Δ gene expression or lipid abundance profiles at most entities. We identified 297 transcripts less abundant in EWY55 growing anaerobically on xylose compared to the bcy1Δ strain (FDR < 0.05; S14 Table), and these were enriched for genes involved in mitochondrial functions, such as electron transport chain, oxidation-reduction, and targets of the HAP2/3/4/5 complex; genes involved in phospholipid metabolism; and genes involved in ergosterol synthesis (p < 0.05, hypergeometric test, see Methods). Many of these processes were significantly affected in our original comparison of the bcy1Δ and ira2Δ strains. Additionally, 93 genes with higher abundance in the EWY55 compared to the bcy1Δ cells (FDR < 0.05; S14 Table) were enriched for ribosomal protein genes and genes involved in translation and sulfate assimilation (p < 10−7, hypergeometric test), processes important for rapid growth. Since EWY55 cells recapitulated the xylose-dependent growth seen in ira2Δ cells, we next asked if its expression changes recapitulated ira2Δ patterns relative to bcy1Δ cells–surprisingly, most did not (Fig 5C). This indicates that the evolved EWY55 did not recouple growth and metabolism under anaerobic xylose conditions via reverting to the ira2Δ strain’s expression patterns. There were a few exceptions, including 22 transcripts of diverse functions (Fig 5C, Clusters A and B) in which expression differences in EWY55 recapitulated those seen in ira2Δ cells compared to the bcy1Δ strain (S15 Table; FDR < 0.05). While the role of these expression changes will require future study, it is intriguing that these clusters included several targets of the glucose-responsive transcription factor Rgt1 and the Sok2 regulator that responds to starvation and hypoxia; both genes have connections to PKA signaling [59–62]. We were particularly interested in phospholipid biosynthesis genes, given all the connections to this pathway throughout our studies. In general, EWY55 cells showed lower transcript abundances of phospholipid biosynthesis genes compared to the bcy1Δ strain grown anaerobically on xylose (S16 Table), making its expression even more divergent from the ira2Δ strain. The phospholipid composition further supports the unique changes of the EWY55 strain that permit recoupled growth and metabolism on xylose. The EWY55 strain showed significantly greater abundance of the storage lipid triacylglycerol (TG; Fig 5D and S17 Table; p < 10−4, ANOVA). Importantly, EWY55 had significantly lower levels of PDME and trended towards higher levels of PC compared to bcy1Δ cells, recapitulating the pattern in ira2Δ cells (Fig 5D and S17 Table). These results are consistent with the hypothesis that the evolved EWY55 strain altered the pathway compared to bcy1Δ cells. Together, our results underscore the complexity of responses to xylose growth and metabolism across ira2Δ cells, the parental bcy1Δ strain, and EWY55 (see Discussion). Discussion We began this work with two primary goals: to identify signatures of xylose-dependent growth and metabolism across a suite of strains with varying capabilities and to elucidate the mechanism through which growth and metabolism are decoupled in cells lacking BCY1. One key result from our work is that there is no obvious gene expression signature associated with the ability to grow anaerobically on xylose (Fig 2). While we did identify a handful of cell-cycle genes whose expression was consistent with cycling in the ira2Δ strain, there were no clear signatures correlated with growth. This was especially interesting in the case of ribosome-related genes, since there has been much debate about whether the level of RP transcripts underlies growth rate [38–42]. In chemostat experiments where growth is limited by nutrient restriction, the abundance of RP and RiBi genes correlates with growth rate, consistent with one set of long-standing models of growth limitations in bacteria [63–68]. However, other seminal studies focusing on stress conditions suggest that growth during stress is not limited by ribosome production [42,69–73]. Our results show clearly that expression of RP and RiBi genes is higher in the non-growing bcy1Δ strains than dividing ira2Δ cells (S2B Fig). In contrast, the EWY55 strain that recovers anaerobic growth on xylose shows higher expression of ribosome-related genes, perhaps supporting rapid division. Together, our results add to a growing body of work that shows that, although production of ribosome components is often correlated with growth rate, division dynamics cannot be universally predicted by RP and RiBi transcript abundances. However, transcriptomic patterns did implicate an interconnected network of expression differences specific to the bcy1Δ strain, and in turn the evolved EWY55 strain, connected to PKA signaling (Fig 3). The affected network implicates mitochondrial function, iron response, carbon metabolism, and phospholipids. We propose that these processes are normally coordinated by PKA signaling in a manner that requires the regulatory subunit Bcy1. Our results are consistent with prior implications that these processes are involved in anaerobic xylose fermentation. Targets of the carbon-responsive transcription factor Azf1 were altered in the bcy1Δ strain upon xylose shift compared to the ira2Δ strain (Fig 3A). We previously showed that altering expression of this transcription factor affects xylose fermentation rates and growth in an ira2Δ strain, and this was surprisingly connected to the ER-localized transcription factor Mga2 [32]. While Mga2 targets were not statistically enriched in comparisons here, 42% of the genes whose promoter is bound by Mga2 (11/26 genes) differed in expression between bcy1Δ and ira2Δ cells. Additionally, Aft1/2 activity and localization is dependent on Mga2 presence, thus adding another possible connection for Mga2 in our proposed regulatory network [74] (Fig 3B). Our results here strongly suggest that deletion of BCY1 naturally augments the transcription factors’ abundance and/or activity. These factors may indirectly alter mitochondrial and/or iron homeostasis. In fact, deletion of the iron-sulfur scaffold protein ISU1, an important sensor of iron availability, is required for anaerobic xylose metabolism [26,31]. Why ISU1 deletion is required for xylose fermentation remains unclear, but one possibility is that it aids in metabolic rewiring influenced by the iron regulon, the Aft1/2 transcription factors, and altered levels of PKA activity [45,75–77]. Remarkably, PKA is directly connected to all these processes. Past work implicated PKA in directly phosphorylating Azf1 [32,44]. While a direct link between PKA and Aft1/2 activity has yet to be identified, PKA catalytic subunit Tpk2 is required to repress the high-affinity iron uptake pathway under standard conditions [76]. Additionally, Ira2 can localize to mitochondria, suggesting that PKA can also localize to this organelle [78]. In fact, PKA is found at the mitochondria of higher eukaryotes [79], suggesting that yeast PKA may also localize to the mitochondria. Together, our results suggest that upregulated PKA activity is required for xylose fermentation and can occur via either IRA2 or BCY1 deletion, but deletion of BCY1 produces stronger effects that underscore its higher per-cell rate of xylose fermentation [32]. A fundamental aspect of BCY1 deletion is that cells can no longer grow robustly despite enhanced xylose metabolism. Our integrated analysis points to a defect in phospholipid flux or metabolism as a major contributor to this decoupling. First, the bcy1Δ strains showed altered gene expression, including Ino2/4 targets such as INO1 (Figs 3A, 4C and S3C), that pointed to differences in phospholipid metabolism. Second, we found that bcy1Δ strains grown anaerobically on xylose display an altered lipid profile that implicates altered metabolism in the PE-PDME-PC pathway, along with phosphorylation differences on key phospholipid enzymes (Table 1). Finally, re-evolving a coupling between anaerobic-xylose growth and metabolism in the bcy1Δ parent implicated mutations in PKA subunit TPK1 and the Ino2/4 repressor OPI1 (Table 2), which is known to be directly regulated by PKA phosphorylation [48]. Opi1 has complex roles in regulating phospholipids, including during the switch to invasive growth depending on nutrients [80], a process also regulated by the RAS/PKA pathway [81–85]. While we were unable to elucidate the exact role of these alleles, our results suggest that the OPI1 mutation may alter Opi1 regulation, especially given that the identified mutation in Opi1 occurs at a known CKII kinase site that regulates Opi1 activity [57]. Our past network inference across this panel of engineered strains revealed altered phosphorylation of CKII targets [32]. Finding that complete deletion of the mutated OPI1 allele reduced growth of the evolved EWY55 strain on xylose (Fig 5B) suggests altered Opi1 activity in the bcy1Δ strain is somehow resolved by mutation of this CKII site. We propose that an interplay between PKA and possibly CKII affect Opi1 regulation in the bcy1Δ strain, and that this interplay is important for growth coupling. Importantly, phospholipid metabolism is required for growth and division. Cells must generate enough phospholipids to support membrane biogenesis [4,86–91]. Furthermore, phospholipids function in inter-organelle communication, connecting the ER and mitochondria via the ER-mitochondria encounter structure (ERMES). Impairment of this structure and inter-organelle communication is known to cause diverse mitochondrial phenotypes and disrupt phospholipid biosynthesis [92,93], connecting phospholipid metabolism to mitochondrial functions, including xylose flux [31]. One possibility is that impaired regulation of Opi1 and Ino2/4 in bcy1Δ cells disrupt growth in the bcy1Δ strain due to insufficient levels of growth-supporting lipids (PC) via decreased production or increased recycling. But another possibility is that accumulation of methylated PE intermediates during the conversion to PC create a toxic buildup coupled with insufficient PC (Fig 4C). Ishiwata-Kimata et al. (2022) [51] found that accumulation of PMME leads to a growth defect by triggering the unfolded protein response and growth arrest. Accumulation of PDME in the bcy1Δ strain (Fig 4B and 4C) may also lead to ER stress, preventing growth paired with interfered ER-mitochondrial communication. Importantly, the evolved EWY55 strain does not share the bcy1Δ strain’s accumulation of PMDE, leading us to propose that EWY55 cells have overcome the possible bottleneck in PC synthesis (Fig 5D). Future studies analyzing pathway flux are needed to fully confirm the presence of and recovery from a bottleneck in phospholipid biosynthesis. A major remaining question is how deletion of BCY1, but not IRA2, decouples growth from metabolism specifically under the conditions studied here. One possibility is that BCY1 deletion upregulates PKA activity to a higher level than deletion of IRA2, whose activation of PKA is indirect via cAMP regulation [33]. PKA activity over some threshold could cause decoupling, as deletion of BCY1 is well characterized to sensitize cells to environmental stressors [94]. An alternate model is that localized cAMP production could influence when and where PKA is active in ira2Δ cells. cAMP exists in concentration gradients in cells to control the subcellular location of active PKA [95–97]. It is possible areas with low cAMP concentration locally inactivate PKA in the ira2Δ strain, whereas BCY1 deletion leads to wholesale activation of PKA throughout the cell. Fitting with this model, BCY1 deletion inhibits growth and metabolism on non-fermentable carbon sources, causing cell death during the diauxic shift and stationary phase, likely from uninhibited PKA [98,99]. However, a third possibility is that loss of BCY1 leads to misdirection of PKA activity. PKA can be directed to subcellular targets in higher eukaryotes via A-kinase anchoring proteins (AKAP) that bind to and direct localization of PKA [100]. While yeast do not possess orthologs of AKAPs, functional analogs have been proposed including Bcy1 itself [79,101,102]. Anaerobic xylose growth and metabolism may be decoupled in bcy1Δ strains via disrupted subcellular localization and substrate interactions of PKA that are coordinated by Bcy1. Additionally, Bcy1 is reported to interact with fatty acid synthases subunits (Fas1/2) [102], implying a direct, physical connection between PKA and lipid biosynthesis. While future studies of PKA localization and substrate interactions are needed to confirm this model, our results show that Bcy1 plays a special role in coordinating PKA activity. It is also possible that other signaling pathways, such as TORC1, may be involved in modulating growth and metabolism phenotypes under anaerobic xylose conditions [103], though our work thus far has not investigated a role for TORC1 in these phenotypes. Finally, this study has been solely focused on the RAS branch of PKA activation, but it is also possible that the Gpa2 branch possesses an important role in growth and metabolism coupling [104]. It is clear that more studies are required to obtain a complete understanding of this mechanism. It is evident from this and many other studies that cells have deeply intertwined the regulation of multiple processes, and disrupting one can have dramatic impacts on many others. Our results here and in previous work implicate the importance of regulatory rewiring in decoupling cellular processes. While engineering xylose metabolism pathways is essential to enable the process, anaerobic xylose fermentation is not enacted without rewiring the regulatory system to simultaneously activate Snf1 along with PKA [31,32]. Here, we propose roles for several regulators, including Opi1 and Bcy1, among downstream effectors like Azf1, Aft1/2, Mga2, and Ino2/4, in modulating growth and metabolism decoupling on anaerobic xylose. Our results strongly suggest that upstream regulatory tinkering rather than altering individual downstream effectors will be required to optimally engineer new cell functions. Methods Media and growth conditions Cells were grown in YP media (10 g/L yeast extract, 20 g/L peptone) with 20g/L of either glucose or xylose. Aerobic cultures were grown at 30°C with vigorous shaking. Anaerobic cultures were grown in a Coy anaerobic chamber (10% CO2, 10% H2, 80% N2) at 30°C with a metal stir bar for mixing. All cultures were inoculated with cells grown aerobically to saturation in YP-glucose and washed one time with the desired growth medium. Anaerobic cultures were inoculated into media incubated in the anaerobic chamber for >16 hours before inoculation. Cell density was monitored by optical density at 600 nm (OD600) with an Eppendorf Spectrophotometer. Sugar and ethanol concentrations were measured with HPLC-RID (Refractive Index Detector) analysis [27]. Growth on solid media (Fig S5C) was performed by collecting 1 OD worth of cells from a saturated YP-glucose culture, washing cells with YP-xylose, and plating serial dilutions onto solid YP medium with 2% xylose, with or without 100 μg/mL of nourseothricin. Plates were grown in a Coy anaerobic chamber for seven days before imaging. Strains and cloning Saccharomyces cerevisiae strains used in this study are described in Table 3. Gene knockouts were created by homologous recombination with either KanMX or Hph cassettes [105,106] and confirmed with diagnostic PCRs. The KanMX cassette was rescued from the bcy1Δ and EWY55 strains with CRISPR-Cas9 using a gRNA specific for KanMX and a repair template containing the flanking sequence. The bcy1Δ or EWY55 strain’s allele of OPI1, RIM8, or TOA1 was cloned into the pKI plasmid, carrying a nourseothricin [NAT] resistance marker, using standard cloning techniques. Plasmids were verified with Sanger sequencing, then transformed into the appropriate bcy1Δ or EWY55 KAN marker rescued strain using NTC selection. Download: PPT PowerPoint slide PNG larger image TIFF original image Table 3. Strains used in this study. https://doi.org/10.1371/journal.pgen.1010593.t003 RNA-seq sample collection, RNA extraction, library preparation, and sequencing Cells from saturated cultures of Y184, ira2Δ, bcy1Δ, and ira2Δbcy1Δ were used to inoculate anaerobic YPD cultures at OD600 0.05. Cultures grew for five hours to early/mid-log phase. 50 mL of the culture was collected, washed with YPX, then used to inoculate anaerobic YPX cultures as described above. Cold 5% phenol/95% ethanol was added to the remaining 50mL YPD cultures, which were harvested by centrifuging at 3000 RPM for 3 minutes and flash frozen in liquid nitrogen. Cell pellets were stored at -80°C until further processing. The YPX cultures grew for 3.5 hours, when the ira2Δ strain resumed growth. Cold phenol/ethanol was added to the 50 mL cultures, which were harvested, flash frozen, and stored at -80°C. Samples were collected from three independent replicates performed on different days. Total RNA was extracted using hot phenol lysis [107] and DNA was digested with Turbo-DNase (Life Technologies, Carlsbad, CA) for 30 minutes at 37°C. RNA was precipitated at 20°C in 2.5 M LiCl for 30 min. rRNA was depleted with EPiCenter Ribo-Zero Magnetic Gold Kit (Yeast) RevA kit (Illumina Inc, San Diego, CA), and the remaining RNA was purified using Agencourt RNACleanXP (Beckman Coulter, Indianapolis, IN) by following the manufacturers’ protocols. RNA-seq libraries were created with the Illumnia TruSeq stranded total RNA kit (Illumina) following the preparation guide (revision C), AMPure XP beads were used for PCR purification (Beckman Coulter, Indianapolis, IN), and cDNA generated with SuperScript II reverse transcriptase (Invitrogen, Carlsbad, CA) as described in the Illumina kit. Libraries were standardized to 2 μM and clusters were generated with standard Cluster kits (version 3) and the Illumina Cluster station. Paired-end 50-bp reads were generated using standard SBS chemistry (version 3) on an Illumina NovaSeq 6000 sequencer. RNA-seq data processing and analysis RNA-seq reads were processed with Trimmomatic version 0.3 [108] and mapped to the Y22-3 genome [58] using BWA-MEM version 0.7.17 with default settings. Read counts were calculated with HTSeq version 0.6.0 [109] using the Y22-3 gene annotations. All raw data were deposited in the NIH GEO database (GSE220465). Raw sequence counts were normalized using trimmed mean of M-values (TMM) method [110]. log2 fold changes (FC) between YP-xylose and YP-glucose samples for each strain and replicate were calculated, then hierarchical clustered using Gene Cluster 3.0 [111] and visualized with Java Treeview version 1.2.0 [112]. Differential expression was analyzed using linear modeling in edgeR version 4.0.3 [113] using pairwise and group comparisons, calling significance at < 0.05 Benjamini and Hochberg false discovery rate (FDR) [114]. Genes in Fig 2B were identified by pairwise comparisons between The Y184, bcy1Δ, and ira2Δbcy1Δ strains with the ira2Δ strain. Genes in Fig 2C were identified by comparing the ira2Δ, bcy1Δ, and ira2Δbcy1Δ strains as a group in the statistical model with the Y184 strain. Genes in Fig 3A were identified first by pairwise comparison between the ira2Δ and bcy1Δ strains, then subsequently further grouped by genes reproducibly expressed in opposing directions between the two strains (e.g. log2FC > 0 in bcy1Δ and log2FC < 0 in ira2Δ). Genes differentially expressed between EWY55 and bcy1Δ strains grown anaerobically on xylose were identified using edgeR version 4.0.3 [113] at FDR [114] < 0.05. Genes were median centered, the log2 YPX abundance of EWY55 or ira2Δ transcripts, relative to log2 bcy1Δ YPX abundance were calculated, then hierarchically clustered Gene Cluster 3.0 [111] and visualized in Java Treeview version 1.2.0 [112]. Functional gene ontology (GO) term and transcriptional regulator enrichment was performed using SetRank [115]; an FDR cutoff of 0.05 was used for transcription target analysis and a Bonferroni corrected p-value cutoff of 10−4 was used to assess overlapping GO categories. Targets of transcription factors were downloaded from YeasTract [116] using only targets with DNA binding evidence. Upstream regulatory motifs were identified with MEME suite version 5.4.1 [117] and associated transcription factors were implicated using Tomtom [118]. Lipidomics sample collection and preparation Cells were grown as described previously for the RNAseq collection, flash frozen in liquid nitrogen, then stored at -80°C. On the day of analysis, each sample was removed from -80°C and maintained on dry ice until time of extraction. 240 μL chilled methanol was added to cell pellet samples in their native tubes over dry ice. Native tubes were transferred to ice and then vortexed. Samples were then transferred to 2 mL microcentrifuge tubes over ice. Next 800 μL of chilled methyl tert-butyl ether (MTBE) was added to native tubes followed by vortexing; these samples were also transferred to the microcentrifuge tube. Microcentrifuge tubes were then vortexed for 10 seconds. A 1/32 teaspoon (0.15 mL) of 1,180 μm glass beads (16–25 US sieve) was added to each tube along with 200 μL LC-MS grade water. Tubes were vortexed for 10 seconds. All tubes were centrifuged at 4°C for 2 minutes at 5,000 x g to pellet cell debris. An extraction blank was prepared per sample preparation steps directly into a 2 mL microcentrifuge tube without yeast. 200 μL of the top (lipophilic) layer from each tube was aliquoted into a low volume amber borosilicate glass autosampler vial with tapered insert. For pooled YPD and pooled YPX samples, the 200 μL aliquot was performed in duplicate. Each vial was dried in a vacuum concentrator for approximately one hour. For pooled YPD and pooled YPX samples, resuspension was performed with 50 μL of a 9:1 MeOH:toluene solution on the first of two preparations (“1X”) while the second preparation was resuspended in 25 μL of 9:1 MeOH:toluene (“2X”). Remaining dried samples were resuspended in 50 μL of 9:1 MeOH:toluene. Each vial was vortexed vigorously for 10 seconds to ensure resuspension of the dried contents. Samples were placed in the instrument’s autosampler at 4°C to await injection. Lipidomics LC-MS analysis LC-MS/MS analysis was performed using an Acquity CSH C18 column (2.1 mm × 100 mm, 1.7 μm particle size, Waters) held at 50°C and a Vanquish Binary Pump (400 μL/mL flow rate; Thermo Scientific, Waltham, MA). Mobile phase A consisted of ACN:H2O (70:30, v/v) with 10 mM ammonium acetate and 0.025% acetic acid. Mobile phase B consisted of IPA:ACN (9:1, v/v) with 10 mM ammonium acetate and 0.025% acetic acid. Initially, mobile phase B was held at 2% for 2 min and increased to 30% over 3 min. In consecutive ramping steps, mobile phase B was increased to 50% over 1 minute, increased to 85% over 14 minutes, and increased to 99% over 1 minute. The gradient was held at 99% mobile phase B for 7 minutes, then decreased to 2% over 0.25 minutes. The column was equilibrated at 2% mobile phase B for 1.75 minutes before the next injection. 10 μL of each extract was injected by a Vanquish Split Sampler HT autosampler (Thermo Scientific, Waltham, MA) in a randomized order. The LC system was coupled to a Q Exactive HF Orbitrap mass spectrometer (MS) through a heated electrospray ionization (HESI II) source (Thermo Scientific, Waltham, MA). Source conditions were as follows: HESI II and capillary temperature at 350°C, sheath gas flow rate at 25 units, aux gas flow rate at 15 units, sweep gas flow rate at 5 units, spray voltage at |3.5 kV|, and S-lens RF at 60.0 units. The MS was operated in a polarity switching mode acquiring positive and negative full MS and MS2 spectra (Top2) within the same injection. Acquisition parameters for full MS scans in both modes were 30,000 resolution, 1 × 106 automatic gain control (AGC) target, 100 ms ion accumulation time (max IT), and 200 to 2000 m/z scan range. MS2 scans in both modes were then performed at 30,000 resolution, 1 × 105 AGC target, 50 ms max IT, 1.0 m/z isolation window, stepped normalized collision energy (NCE) at 20, 30, 40, and a 10.0 s dynamic exclusion. Lipidomics data analysis The resulting LC–MS data were processed using Compound Discoverer 3.1 (Thermo Scientific, Waltham, MA) and LipiDex, an in-house-developed software suite [119]. All peaks between 0.4 min and 21.0 min retention time and between100 Da and 5000 Da MS1 precursor mass were aggregated into compound groups using a 10-ppm mass, 0.2 min retention time tolerance, a minimum peak intensity of 1x10^5, a maximum peak-width of 0.75 min, and a signal-to-noise (S/N) ratio of 3. Features were required to be 5-fold greater intensity in samples than blanks. MS/MS spectra were searched against an in-silico generated lipid spectral library. Spectral matches were required to have a dot product score greater than 500 and a reverse dot product score greater than 700. Lipid MS/MS spectra which contain acyl-chain specific fragments and contained no significant interference (<75%) from co-eluting isobaric lipids were identified at molecular species level. If individual fatty acid substituents were unresolved, then identifications were made with the sum of the fatty acid substituents. Lipid features were further filtered based on 1) presence in a minimum of two raw files, 2) a median absolute retention time deviation of 3.5, and 3) average pooled relative standard deviations of less than 30%. Differential abundance of lipids was analyzed with linear modeling in edgeR version 4.0.3 using pairwise comparisons and a Benjamini and Hochberg [114] FDR < 0.05 to call significance. After the log2FC between YP-xylose and YP-glucose samples for each strain was calculated, lipids were hierarchically cluster in Gene Cluster 3.0 [111] and visualized in Java Treeview 1.2.0 [112]. For all phosphatidylcholine moieties, a paired ANOVA with a cutoff of p < 0.05 was performed between ira2Δ and bcy1Δ samples. All raw and processed lipidomics data files were deposited in MassIVE database under dataset number MSV000090868. For EWY55 lipidomics data, differential abundance of lipids was analyzed by calculating the log2(fold change) ratio between YPX and YPD samples for each strain and replicate. The paired log2(fold change) differences between EWY55 and ira2Δ or bcy1Δ samples were calculated, and an absolute value difference greater than 1.5 on a log2 scale was called significant. Average log2(fold change) of classified, significant lipid classes of interest were calculated along with standard error, and significant differences in lipid classes were called by a paired ANOVA. Phosphoproteomics data Phosphoproteomics data from Myers et al. (2019) [32] was analyzed to compare the phosphorylation of phospholipid biosynthesis enzymes. Reproducible pairwise comparisons between YP-xylose samples of strains with a log2 fold-change >2 were called significant. Inositol and choline supplementation YP-xylose medium was prepared as described above. Myo-inositol (Sigma, Burlington, MA) was added to a final concentration of 75 μM and choline (Thermo Scientific, Waltham, MA) to a concentration of 10 mM. Anaerobic cultures were inoculated from saturated overnight cultures to an OD600 of 0.1. Growth, xylose concentration, and ethanol concentration was monitored over 44 hours. A paired ANOVA between YP-xylose and YP-xylose-inositol-choline cultures was performed to determine significant differences between growth using a p value cutoff of 0.05. Adaptive laboratory evolutions bcy1Δ cells were inoculated in anaerobic YP-glucose medium at an OD600 of 0.01 and grown for ~21 generations. This was used to seed a fresh anaerobic YP-glucose culture at an OD600 of 0.01, which grew for ~7 generations. From this, a YP-2% xylose 0.1% glucose culture was seeded at an OD600 of 0.01, then grown for ~7 generations. This process was repeated four more times before plating the culture on YP-xylose and collecting single colonies capable of growing anaerobically on xylose. Evolutions were performed in three independent cultures. Evolved bcy1Δ strain genome sequencing and analysis Evolved bcy1Δ strains were grown aerobically in YP-glucose and genomic DNA was extracted using the Qiagen (Hilden, Germany) Genomic-tip 20/G kit following manufacturer’s protocol. Genomic DNA was fragmented into ~200 bp fragments using a sonifier with four minutes on and one minute off while incubating on ice, repeated for a total of four cycles. DNA libraries were made using the NEBNext Ultra II DNA Library Prep Kit for Illumina protocol, using the NEBNext Multiplex Oligos for Illumina (Dual Index Primers Set 1) (New England Biolabs, Ipswich, MA). Paired-end 300 bp reads were generated on an Illumina MiSeq. Variants in the parental bcy1Δ strain were identified with GATK version 4.2 (Broad Institute) and substituted into the Y22-3 reference genome as a mapping reference. Reads were mapped to the newly generated bcy1Δ strain genome, and variants were called using GATK version 4.2 and single nucleotide polymorphisms (SNPs) annotated with SnpEff version 5.0 and vcftools version v0.,1.12b. Only SNPs occurring in the coding region of a gene were considered for further analysis and verified with Sanger sequencing. Media and growth conditions Cells were grown in YP media (10 g/L yeast extract, 20 g/L peptone) with 20g/L of either glucose or xylose. Aerobic cultures were grown at 30°C with vigorous shaking. Anaerobic cultures were grown in a Coy anaerobic chamber (10% CO2, 10% H2, 80% N2) at 30°C with a metal stir bar for mixing. All cultures were inoculated with cells grown aerobically to saturation in YP-glucose and washed one time with the desired growth medium. Anaerobic cultures were inoculated into media incubated in the anaerobic chamber for >16 hours before inoculation. Cell density was monitored by optical density at 600 nm (OD600) with an Eppendorf Spectrophotometer. Sugar and ethanol concentrations were measured with HPLC-RID (Refractive Index Detector) analysis [27]. Growth on solid media (Fig S5C) was performed by collecting 1 OD worth of cells from a saturated YP-glucose culture, washing cells with YP-xylose, and plating serial dilutions onto solid YP medium with 2% xylose, with or without 100 μg/mL of nourseothricin. Plates were grown in a Coy anaerobic chamber for seven days before imaging. Strains and cloning Saccharomyces cerevisiae strains used in this study are described in Table 3. Gene knockouts were created by homologous recombination with either KanMX or Hph cassettes [105,106] and confirmed with diagnostic PCRs. The KanMX cassette was rescued from the bcy1Δ and EWY55 strains with CRISPR-Cas9 using a gRNA specific for KanMX and a repair template containing the flanking sequence. The bcy1Δ or EWY55 strain’s allele of OPI1, RIM8, or TOA1 was cloned into the pKI plasmid, carrying a nourseothricin [NAT] resistance marker, using standard cloning techniques. Plasmids were verified with Sanger sequencing, then transformed into the appropriate bcy1Δ or EWY55 KAN marker rescued strain using NTC selection. Download: PPT PowerPoint slide PNG larger image TIFF original image Table 3. Strains used in this study. https://doi.org/10.1371/journal.pgen.1010593.t003 RNA-seq sample collection, RNA extraction, library preparation, and sequencing Cells from saturated cultures of Y184, ira2Δ, bcy1Δ, and ira2Δbcy1Δ were used to inoculate anaerobic YPD cultures at OD600 0.05. Cultures grew for five hours to early/mid-log phase. 50 mL of the culture was collected, washed with YPX, then used to inoculate anaerobic YPX cultures as described above. Cold 5% phenol/95% ethanol was added to the remaining 50mL YPD cultures, which were harvested by centrifuging at 3000 RPM for 3 minutes and flash frozen in liquid nitrogen. Cell pellets were stored at -80°C until further processing. The YPX cultures grew for 3.5 hours, when the ira2Δ strain resumed growth. Cold phenol/ethanol was added to the 50 mL cultures, which were harvested, flash frozen, and stored at -80°C. Samples were collected from three independent replicates performed on different days. Total RNA was extracted using hot phenol lysis [107] and DNA was digested with Turbo-DNase (Life Technologies, Carlsbad, CA) for 30 minutes at 37°C. RNA was precipitated at 20°C in 2.5 M LiCl for 30 min. rRNA was depleted with EPiCenter Ribo-Zero Magnetic Gold Kit (Yeast) RevA kit (Illumina Inc, San Diego, CA), and the remaining RNA was purified using Agencourt RNACleanXP (Beckman Coulter, Indianapolis, IN) by following the manufacturers’ protocols. RNA-seq libraries were created with the Illumnia TruSeq stranded total RNA kit (Illumina) following the preparation guide (revision C), AMPure XP beads were used for PCR purification (Beckman Coulter, Indianapolis, IN), and cDNA generated with SuperScript II reverse transcriptase (Invitrogen, Carlsbad, CA) as described in the Illumina kit. Libraries were standardized to 2 μM and clusters were generated with standard Cluster kits (version 3) and the Illumina Cluster station. Paired-end 50-bp reads were generated using standard SBS chemistry (version 3) on an Illumina NovaSeq 6000 sequencer. RNA-seq data processing and analysis RNA-seq reads were processed with Trimmomatic version 0.3 [108] and mapped to the Y22-3 genome [58] using BWA-MEM version 0.7.17 with default settings. Read counts were calculated with HTSeq version 0.6.0 [109] using the Y22-3 gene annotations. All raw data were deposited in the NIH GEO database (GSE220465). Raw sequence counts were normalized using trimmed mean of M-values (TMM) method [110]. log2 fold changes (FC) between YP-xylose and YP-glucose samples for each strain and replicate were calculated, then hierarchical clustered using Gene Cluster 3.0 [111] and visualized with Java Treeview version 1.2.0 [112]. Differential expression was analyzed using linear modeling in edgeR version 4.0.3 [113] using pairwise and group comparisons, calling significance at < 0.05 Benjamini and Hochberg false discovery rate (FDR) [114]. Genes in Fig 2B were identified by pairwise comparisons between The Y184, bcy1Δ, and ira2Δbcy1Δ strains with the ira2Δ strain. Genes in Fig 2C were identified by comparing the ira2Δ, bcy1Δ, and ira2Δbcy1Δ strains as a group in the statistical model with the Y184 strain. Genes in Fig 3A were identified first by pairwise comparison between the ira2Δ and bcy1Δ strains, then subsequently further grouped by genes reproducibly expressed in opposing directions between the two strains (e.g. log2FC > 0 in bcy1Δ and log2FC < 0 in ira2Δ). Genes differentially expressed between EWY55 and bcy1Δ strains grown anaerobically on xylose were identified using edgeR version 4.0.3 [113] at FDR [114] < 0.05. Genes were median centered, the log2 YPX abundance of EWY55 or ira2Δ transcripts, relative to log2 bcy1Δ YPX abundance were calculated, then hierarchically clustered Gene Cluster 3.0 [111] and visualized in Java Treeview version 1.2.0 [112]. Functional gene ontology (GO) term and transcriptional regulator enrichment was performed using SetRank [115]; an FDR cutoff of 0.05 was used for transcription target analysis and a Bonferroni corrected p-value cutoff of 10−4 was used to assess overlapping GO categories. Targets of transcription factors were downloaded from YeasTract [116] using only targets with DNA binding evidence. Upstream regulatory motifs were identified with MEME suite version 5.4.1 [117] and associated transcription factors were implicated using Tomtom [118]. Lipidomics sample collection and preparation Cells were grown as described previously for the RNAseq collection, flash frozen in liquid nitrogen, then stored at -80°C. On the day of analysis, each sample was removed from -80°C and maintained on dry ice until time of extraction. 240 μL chilled methanol was added to cell pellet samples in their native tubes over dry ice. Native tubes were transferred to ice and then vortexed. Samples were then transferred to 2 mL microcentrifuge tubes over ice. Next 800 μL of chilled methyl tert-butyl ether (MTBE) was added to native tubes followed by vortexing; these samples were also transferred to the microcentrifuge tube. Microcentrifuge tubes were then vortexed for 10 seconds. A 1/32 teaspoon (0.15 mL) of 1,180 μm glass beads (16–25 US sieve) was added to each tube along with 200 μL LC-MS grade water. Tubes were vortexed for 10 seconds. All tubes were centrifuged at 4°C for 2 minutes at 5,000 x g to pellet cell debris. An extraction blank was prepared per sample preparation steps directly into a 2 mL microcentrifuge tube without yeast. 200 μL of the top (lipophilic) layer from each tube was aliquoted into a low volume amber borosilicate glass autosampler vial with tapered insert. For pooled YPD and pooled YPX samples, the 200 μL aliquot was performed in duplicate. Each vial was dried in a vacuum concentrator for approximately one hour. For pooled YPD and pooled YPX samples, resuspension was performed with 50 μL of a 9:1 MeOH:toluene solution on the first of two preparations (“1X”) while the second preparation was resuspended in 25 μL of 9:1 MeOH:toluene (“2X”). Remaining dried samples were resuspended in 50 μL of 9:1 MeOH:toluene. Each vial was vortexed vigorously for 10 seconds to ensure resuspension of the dried contents. Samples were placed in the instrument’s autosampler at 4°C to await injection. Lipidomics LC-MS analysis LC-MS/MS analysis was performed using an Acquity CSH C18 column (2.1 mm × 100 mm, 1.7 μm particle size, Waters) held at 50°C and a Vanquish Binary Pump (400 μL/mL flow rate; Thermo Scientific, Waltham, MA). Mobile phase A consisted of ACN:H2O (70:30, v/v) with 10 mM ammonium acetate and 0.025% acetic acid. Mobile phase B consisted of IPA:ACN (9:1, v/v) with 10 mM ammonium acetate and 0.025% acetic acid. Initially, mobile phase B was held at 2% for 2 min and increased to 30% over 3 min. In consecutive ramping steps, mobile phase B was increased to 50% over 1 minute, increased to 85% over 14 minutes, and increased to 99% over 1 minute. The gradient was held at 99% mobile phase B for 7 minutes, then decreased to 2% over 0.25 minutes. The column was equilibrated at 2% mobile phase B for 1.75 minutes before the next injection. 10 μL of each extract was injected by a Vanquish Split Sampler HT autosampler (Thermo Scientific, Waltham, MA) in a randomized order. The LC system was coupled to a Q Exactive HF Orbitrap mass spectrometer (MS) through a heated electrospray ionization (HESI II) source (Thermo Scientific, Waltham, MA). Source conditions were as follows: HESI II and capillary temperature at 350°C, sheath gas flow rate at 25 units, aux gas flow rate at 15 units, sweep gas flow rate at 5 units, spray voltage at |3.5 kV|, and S-lens RF at 60.0 units. The MS was operated in a polarity switching mode acquiring positive and negative full MS and MS2 spectra (Top2) within the same injection. Acquisition parameters for full MS scans in both modes were 30,000 resolution, 1 × 106 automatic gain control (AGC) target, 100 ms ion accumulation time (max IT), and 200 to 2000 m/z scan range. MS2 scans in both modes were then performed at 30,000 resolution, 1 × 105 AGC target, 50 ms max IT, 1.0 m/z isolation window, stepped normalized collision energy (NCE) at 20, 30, 40, and a 10.0 s dynamic exclusion. Lipidomics data analysis The resulting LC–MS data were processed using Compound Discoverer 3.1 (Thermo Scientific, Waltham, MA) and LipiDex, an in-house-developed software suite [119]. All peaks between 0.4 min and 21.0 min retention time and between100 Da and 5000 Da MS1 precursor mass were aggregated into compound groups using a 10-ppm mass, 0.2 min retention time tolerance, a minimum peak intensity of 1x10^5, a maximum peak-width of 0.75 min, and a signal-to-noise (S/N) ratio of 3. Features were required to be 5-fold greater intensity in samples than blanks. MS/MS spectra were searched against an in-silico generated lipid spectral library. Spectral matches were required to have a dot product score greater than 500 and a reverse dot product score greater than 700. Lipid MS/MS spectra which contain acyl-chain specific fragments and contained no significant interference (<75%) from co-eluting isobaric lipids were identified at molecular species level. If individual fatty acid substituents were unresolved, then identifications were made with the sum of the fatty acid substituents. Lipid features were further filtered based on 1) presence in a minimum of two raw files, 2) a median absolute retention time deviation of 3.5, and 3) average pooled relative standard deviations of less than 30%. Differential abundance of lipids was analyzed with linear modeling in edgeR version 4.0.3 using pairwise comparisons and a Benjamini and Hochberg [114] FDR < 0.05 to call significance. After the log2FC between YP-xylose and YP-glucose samples for each strain was calculated, lipids were hierarchically cluster in Gene Cluster 3.0 [111] and visualized in Java Treeview 1.2.0 [112]. For all phosphatidylcholine moieties, a paired ANOVA with a cutoff of p < 0.05 was performed between ira2Δ and bcy1Δ samples. All raw and processed lipidomics data files were deposited in MassIVE database under dataset number MSV000090868. For EWY55 lipidomics data, differential abundance of lipids was analyzed by calculating the log2(fold change) ratio between YPX and YPD samples for each strain and replicate. The paired log2(fold change) differences between EWY55 and ira2Δ or bcy1Δ samples were calculated, and an absolute value difference greater than 1.5 on a log2 scale was called significant. Average log2(fold change) of classified, significant lipid classes of interest were calculated along with standard error, and significant differences in lipid classes were called by a paired ANOVA. Phosphoproteomics data Phosphoproteomics data from Myers et al. (2019) [32] was analyzed to compare the phosphorylation of phospholipid biosynthesis enzymes. Reproducible pairwise comparisons between YP-xylose samples of strains with a log2 fold-change >2 were called significant. Inositol and choline supplementation YP-xylose medium was prepared as described above. Myo-inositol (Sigma, Burlington, MA) was added to a final concentration of 75 μM and choline (Thermo Scientific, Waltham, MA) to a concentration of 10 mM. Anaerobic cultures were inoculated from saturated overnight cultures to an OD600 of 0.1. Growth, xylose concentration, and ethanol concentration was monitored over 44 hours. A paired ANOVA between YP-xylose and YP-xylose-inositol-choline cultures was performed to determine significant differences between growth using a p value cutoff of 0.05. Adaptive laboratory evolutions bcy1Δ cells were inoculated in anaerobic YP-glucose medium at an OD600 of 0.01 and grown for ~21 generations. This was used to seed a fresh anaerobic YP-glucose culture at an OD600 of 0.01, which grew for ~7 generations. From this, a YP-2% xylose 0.1% glucose culture was seeded at an OD600 of 0.01, then grown for ~7 generations. This process was repeated four more times before plating the culture on YP-xylose and collecting single colonies capable of growing anaerobically on xylose. Evolutions were performed in three independent cultures. Evolved bcy1Δ strain genome sequencing and analysis Evolved bcy1Δ strains were grown aerobically in YP-glucose and genomic DNA was extracted using the Qiagen (Hilden, Germany) Genomic-tip 20/G kit following manufacturer’s protocol. Genomic DNA was fragmented into ~200 bp fragments using a sonifier with four minutes on and one minute off while incubating on ice, repeated for a total of four cycles. DNA libraries were made using the NEBNext Ultra II DNA Library Prep Kit for Illumina protocol, using the NEBNext Multiplex Oligos for Illumina (Dual Index Primers Set 1) (New England Biolabs, Ipswich, MA). Paired-end 300 bp reads were generated on an Illumina MiSeq. Variants in the parental bcy1Δ strain were identified with GATK version 4.2 (Broad Institute) and substituted into the Y22-3 reference genome as a mapping reference. Reads were mapped to the newly generated bcy1Δ strain genome, and variants were called using GATK version 4.2 and single nucleotide polymorphisms (SNPs) annotated with SnpEff version 5.0 and vcftools version v0.,1.12b. Only SNPs occurring in the coding region of a gene were considered for further analysis and verified with Sanger sequencing. Supporting information S1 Fig. PKA pathway mutants share similar growth and metabolism phenotypes when grown anaerobically on YPD. A-C. Average (n = 3 biological replicates) (A) growth (OD600, optical density), (B) glucose concentration, and (C) ethanol concentration of ira2Δ, bcy1Δ, and ira2Δbcy1Δ strains grown anaerobically on rich glucose medium (p > 0.05, ANOVA). https://doi.org/10.1371/journal.pgen.1010593.s001 (TIF) S2 Fig. Transcriptomic profile of strains with varying xylose utilization and growth capabilities show RP transcripts are not limiting. A. Expression of 5834 genes (rows) detected in all four strains (Y184, ira2Δ, bcy1Δ, ira2Δbcy1Δ), organized by hierarchical clustering of log2(fold change) upon glucose-to-xylose shift, as described in Fig 2. Each column represents one of three biological replicates of the denoted strain listed above. B. Expression of 135 ribosomal protein genes (rows) in all four strains (Y184, ira2Δ, bcy1Δ, ira2Δbcy1Δ), organized by hierarchical clustering of log2(fold change) upon glucose-to-xylose shift. The blue-yellow heatmap on the left represents the log2(fold change) in expression upon glucose to xylose shift across biological triplicates (columns). The purple-green heatmap on the right represents the abundance of each transcript (rows) in each strain grown on glucose (G) or xylose (X), relative to the average (n = 3) abundance of transcripts measured in the Y184 YPD sample. https://doi.org/10.1371/journal.pgen.1010593.s002 (TIF) S3 Fig. Genes expression changes specific to the bcy1Δ strain show changes in induction/repression and not basal mRNA abundances. A. Expression of 654 genes whose log2(fold change) upon glucose to xylose shift is different (FDR < 0.05) between the ira2Δ and bcy1Δ strains and whose change in expression is in the opposite direction (increased or decreased) across strains. (see Methods for details). The yellow-blue heatmap on the left represents the YPX/YPD log2(fold change). The green-purple heatmap on the right represents transcript (rows) abundance in anaerobic glucose (G) and anaerobic xylose (X) relative to the average (n = 3) abundance of transcript in the Y184 YPD sample. B. ~500 base pairs upstream of the ORF for genes repressed in the bcy1Δ strain upon shift to xylose were analyzed for enriched motifs (bottom motif; MEME Suite), analyzed for known transcription factor binding sites (TOMTOM), and identified the Aft1/2 consensus site (top motif; see Methods for details). C. Expression of 77 genes from A whose promoters are physically bound by Ino2 and/or Ino4, organized by hierarchical clustering. https://doi.org/10.1371/journal.pgen.1010593.s003 (TIF) S4 Fig. Phosphatidylcholine abundance upon shift to xylose is lower in bcy1Δ cells compared to ira2Δ cells. A. The majority of phosphatidylcholine species (rows) identified show significantly lower log2(fold change) upon the shift from glucose to xylose in the bcy1Δ and ira2Δbcy1Δ compared to the ira2Δ strain (p = 0.0015082, ANOVA). https://doi.org/10.1371/journal.pgen.1010593.s004 (TIF) S5 Fig. Evolved bcy1Δ strain recapitulates the ira2Δ strain’s phenotype but not transcriptome. A-B. Average (n = 3 biological replicates) of (A) xylose concentration or (B) ethanol concentration for ira2Δ, bcy1Δ, and EWY55 strains anaerobically grown in rich xylose medium (p > 0.05, ANOVA). C. Representatives of multiple replicates of EWY55 or EWY55 cells lacking OPI1 (top panels), RIM8 (middle panels), or TOA1 (bottom panels) and complemented with an empty vector or parental or evolved allele grown anaerobically on solid xylose (left) or glucose (right) medium with NTC selection. D. Average (n = 3 biological replicates) of ethanol concentration for EWY55 and EWY55 opi1Δ strains anaerobically grown in rich xylose medium (p > 0.05). https://doi.org/10.1371/journal.pgen.1010593.s005 (TIF) S6 Fig. Directed evolution on anaerobic xylose generated multiple evolved strains with varying growth rates. A-E. Average (n = 3 biological replicates) growth (OD600, optical density) of ira2Δ, bcy1Δ, EWY55, and (A) EWY87-1, (B) EWY87-3, (C) EWY89-1, (D) EWY89-2, or (E) EWY89-3 strains grown anaerobically on rich xylose medium (p < 10−4, ANOVA). https://doi.org/10.1371/journal.pgen.1010593.s006 (TIF) S1 Table. log2(fold change) (n = 3) for all transcripts in dataset. Companion table for S2A Fig. https://doi.org/10.1371/journal.pgen.1010593.s007 (XLSX) S2 Table. log2(fold change) (n = 3) for transcripts that significantly differ in fold change upon xylose shift in Y184, bcy1Δ, and ira2Δbcy1Δ cells compared to the ira2Δ strain. Companion table for Fig 2B. https://doi.org/10.1371/journal.pgen.1010593.s008 (XLSX) S3 Table. Average (n = 3) log2(fold change) of cell cycle kinase and cyclin transcripts. https://doi.org/10.1371/journal.pgen.1010593.s009 (XLSX) S4 Table. log2(fold change) (n = 3) for ribosomal protein transcripts. Companion table for S2B Fig. https://doi.org/10.1371/journal.pgen.1010593.s010 (XLSX) S5 Table. log2(fold change) (n = 3) for transcripts that significantly differ in fold change upon xylose shift in the xylose fermenting strains (ira2Δ, bcy1Δ) compared to the Y184 strain. Data for ira2Δbcy1Δ cells is also shown. Companion table for Fig 2C. https://doi.org/10.1371/journal.pgen.1010593.s011 (XLSX) S6 Table. log2(fold change) (n = 3) for transcripts that are annotated in central carbon metabolism and significantly differ in fold change upon xylose shift in ira2Δ and bcy1Δ cells compared to the Y184 strain. Data for ira2Δbcy1Δ cells is also shown. Companion table for Fig 2D. https://doi.org/10.1371/journal.pgen.1010593.s012 (XLSX) S7 Table. log2(fold change) (n = 3) for transcripts that significantly differ in fold change and directionality upon xylose shift in bcy1Δ cells compared to ira2Δ cells. Companion table for Fig 3A. https://doi.org/10.1371/journal.pgen.1010593.s013 (XLSX) S8 Table. Functional gene ontology enrichments for clusters in Fig 3A. https://doi.org/10.1371/journal.pgen.1010593.s014 (XLSX) S9 Table. log2(fold change) (n = 3) for Ino2/4 gene targets that significantly differ in fold change and directionality upon xylose shift in bcy1Δ cells compared to ira2Δ cells. Companion table for S3C Fig. https://doi.org/10.1371/journal.pgen.1010593.s015 (XLSX) S10 Table. log2(fold change) (n = 3) for all lipids in dataset. https://doi.org/10.1371/journal.pgen.1010593.s016 (XLSX) S11 Table. log2(fold change) in abundance (n = 3) for lipids that significantly differ in fold change upon xylose shift in ira2Δ, bcy1Δ, and ira2Δbcy1Δ cells compared to the Y184 strain. Companion table for Fig 4A. https://doi.org/10.1371/journal.pgen.1010593.s017 (XLSX) S12 Table. log2(fold change) in abundance (n = 3) for lipids that significantly differ in fold change upon xylose shift in ira2Δbcy1Δ cells compared to the ira2Δ strain. Companion table for Fig 4B. https://doi.org/10.1371/journal.pgen.1010593.s018 (XLSX) S13 Table. log2(fold change) in abundance (n = 3) for all phosphatidylcholine entities. Companion table for S4 Fig. https://doi.org/10.1371/journal.pgen.1010593.s019 (XLSX) S14 Table. Transcripts whose abundance significantly differs in EWY55 cells compared to the bcy1Δ strain in xylose. Transcript abundance differences (n = 3) of EWY55 and ira2Δ compared to bcy1Δ cells. https://doi.org/10.1371/journal.pgen.1010593.s020 (XLSX) S15 Table. Transcripts (n = 3) whose abundance significantly differs in EWY55 and ira2Δ cells compared to the bcy1Δ strain in xylose. Companion table for Fig 5C. https://doi.org/10.1371/journal.pgen.1010593.s021 (XLSX) S16 Table. Transcript abundance differences of phospholipid biosynthetic genes in EWY55 and ira2Δ cells compared to bcy1Δ cells on xylose. https://doi.org/10.1371/journal.pgen.1010593.s022 (XLSX) S17 Table. log2(fold change) in abundance (n = 3) for lipids that significantly differ in fold change upon xylose shift in EWY55 cells compared to the bcy1Δ strain. Companion table for Fig 5D. https://doi.org/10.1371/journal.pgen.1010593.s023 (XLSX) Acknowledgments We thank Mike Place for computational help with transcriptomic data analysis, James Hose and Venera Bouriakov for help with generating RNAseq libraries, Kevin Myers for phosphoproteomic data, and members of the Gasch lab for constructive discussions.

Lee, Rosemary S.;Geronimo, Carly L.;Liu, Liping;Twarowski, Jerzy M.;Malkova, Anna;Zakian, Virginia A.

doi: 10.1371/journal.pgen.1010853pmid: 37486934

Introduction Saccharomyces cerevisiae Pif1 is a multi-functional DNA helicase that is important for the integrity of both nuclear and mitochondrial (mt) DNA (reviewed in [1,2]). Although its exact role in mitochondria has not been determined, mtDNA is quickly lost in cells lacking this helicase. In addition, pif1Δ cells are defective in certain types of recombination between mt genomes [3]. Nuclear Pif1 has multiple functions. It negatively regulates telomerase at telomeres and double-strand breaks [4,5], participates in Okazaki fragment processing [6,7], suppresses genome instability at G-quadruplex motifs [8,9], contributes to fork arrest at the replication fork barrier within ribosomal DNA (rDNA) [10], is critical for Break-Induced Replication (BIR) repair of double strand breaks (DSBs) [11,12], promotes fork progression at tRNA genes [13,14] and centromeres [15], resolves converged replication forks [10,16], and affects the abundance of centromeric RNA [15]. Two isoforms of Pif1 are generated from the same mRNA by the use of alternative translational start sites (Fig 1A). The mitochondrial isoform is produced when translation starts at the first AUG (M1). When translation begins at the second AUG (M2), the nuclear isoform is made [4]. Pif1 localization into mitochondria is directed by a mitochondrial targeting signal (MTS) found between the two translation start sites [4]. Previously, we generated two separation of function alleles of PIF1 [4]. The pif1-m1 allele, in which the mitochondrial translation start site is mutated, confers normal nuclear Pif1 functions but fails to maintain mtDNA. Cells expressing pif1-m2, in which the nuclear translational start site is mutated, are mitochondrial proficient but deficient in nuclear Pif1 activities. However, the pif1-m2 allele is not a null for Pif1 nuclear functions; for example, rates of gross chromosomal rearrangements (GCR) are higher in pif1Δ cells than in pif1-m2 cells [8]. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 1. Structure and expression of wild type and mutant Pif1 proteins. (A) Structure of wild type Pif1. The 238 amino acid N-terminus is in purple; the 507 amino acid helicase domain is in white and the signature motif (SM) in pink; A, B and C are conserved regions of unknown functions; the 114 amino acid C-terminus is in blue. Numbers above diagram indicate amino acid number. Pif1 family helicases have three domains: an amino-terminal (purple), a helicase core (white), and a carboxyl-terminal (blue) domain. Translation from the first AUG (M1) creates the mitochondrial isoform and translation from the second AUG (M2) creates the nuclear isoform. Located between the two start sites is the mitochondrial targeting signal (MTS, green). A pif1-m1 allele generated by mutating M1 to alanine creates only the nuclear isoform, while a pif1-m2 allele generated by mutating M2 to alanine creates the mitochondrial isoform. (B) Structure of mutant proteins. Color scheme is the same as in A. Yellow bars mark the Pif1 NLS while red bar is the NLS from SV40 T-antigen. Lengths of each protein are shown to the right of diagrams. (C) Mutant Pif1 protein expression is comparable to that of wild type Pif1. Westerns analysis of wild type and mutant proteins using an anti-FLAG or anti-alpha tubulin (loading control) antibody on extracts from pif1Δ cells carrying an empty vector or the indicated allele of PIF1. https://doi.org/10.1371/journal.pgen.1010853.g001 Although the mitochondrial targeting sequence is known, the mechanism that targets Pif1 to the nucleus has not been characterized. Transport of proteins into the nucleus can be passive or active. Although even some large proteins can enter the nucleus via passive diffusion [17], the efficient transport of many proteins that are >~60 kDa is facilitated by their association with carrier proteins called importins [18]. Importins recognize a short amino acid sequence, known as a nuclear localization signal (NLS), which is found within the ORFs of a subset of proteins that are destined for the nucleus. Because Pif1 has a molecular weight of ~98 kDa, it likely contains an NLS. The goal of this study was to create a PIF1 allele that is defective in nuclear entry but retains wildtype mitochondrial function. We identified four basic amino acids (781KKRK784) in the carboxy-terminal portion of Pif1 as the core region of a candidate NLS. Deletion of these four amino acids did not affect mitochondrial function or protein abundance but resulted in defects in four of four tested nuclear functions. Insertion of a heterologous NLS from the simian virus 40 (SV40) T antigen suppressed the nuclear phenotypes of the mutant lacking these four amino acids (781KKRK784). We propose that this newly identified region of PIF1 is a functional NLS. Results The S. cerevisiae Pif1 is a Super Family I (SF1) DNA helicase. All SF1 helicases contain six motifs within a 400–500 amino acid helicase domain (orange bars marked by Roman numerals; Fig 1A). In addition, the helicase domain of Pif1 family helicases contains a 23 amino acid signature motif (SM) that is unique to this group of helicases (pink box, Fig 1A) and is essential for ATPase activity [19,20]. In contrast to the helicase domain, the amino (purple rectangle) and carboxyl (blue rectangle, Fig 1A) terminal regions that flank the helicase domain vary greatly in both length and sequence amongst Pif1 family helicases. As part of a strategy to identify separation-of-function PIF1 alleles, we carried out deletion analysis of the amino and carboxyl terminal regions of the protein. One of the findings from this analysis was that deletion of the terminal 79 amino acids disrupted Pif1 nuclear functions but retained mitochondrial function [21]. Given that a mutant lacking the same 79 amino acids had ATPase activity in vitro [22], we considered that this region might contain the Pif1 NLS. Mapping the position of the Pif1 NLS The cNLS Mapper program was used to search for candidate nuclear localization signals in the Pif1 open reading frame [23]. The program assigns prediction scores of up to 10 with higher scores indicating a higher probability of being an NLS. cNLS Mapper identified two candidate NLSs within Pif1: a monopartite NLS (777DEQVKKRKLDY787; prediction score of 8) and a bipartite NLS (417RQRGDVKFIDMLNRMRLGNIDDETEREFKKLSRP450; prediction score of 5). The bipartite NLS had a relatively low prediction score, and its sequence overlapped one of the canonical helicase motifs (motif IV) as well as another conserved region (motif A) (Fig 1A). Based on these findings, we focused our analysis on the monopartite NLS candidate sequence. Strains to test whether putative NLS is required for nuclear functions To determine if the predicted monopartite NLS (777DEQVKKRKLDY787) was important for Pif1 nuclear functions, we deleted the core of this sequence, 781KKRK784 to generate the pif1-NLSΔ (NLS deleted) allele (Fig 1B). Only the core of the putative NLS was deleted to reduce the possibility of disrupting sequences important for other Pif1 functions. In addition, these four amino acids match the consensus for a classical monopartite NLS motif: K-(K/R)-X-(K/R) [24]. As a control, we fused the NLS from the simian virus 40 (SV40) T antigen (126PKKKRKV132), which functions as an NLS in S. cerevisiae [25], to the end of the pif1-NLSΔ ORF to generate the pif1-NLSΔ+SV40 allele (Fig 1B). If our hypothesis is correct, pif1-NLSΔ cells should be defective in nuclear Pif1 functions, while still retaining mitochondrial activities. In addition, nuclear defects in pif1-NLSΔ cells should be reduced in pif1-NLSΔ+SV40 cells, even though the SV40 NLS is not in the same position as the predicted Pif1 NLS. We also generated a double mutant of Pif1, pif1-m2+NLSΔ, which deleted the candidate NLS from the partial loss of function pif1-m2 allele. Amino acids 781KKRK784 are not essential for protein stability or for maintenance of mitochondrial function Western blot analysis showed that the Pif1-NLSΔ protein was stably expressed (Fig 1C). In addition, attaching the SV40 NLS to the carboxyl-terminus of either WT Pif1 or Pif1-NLSΔ did not affect the abundance of either protein (Fig 1C). If 781KKRK784 is the core of a functional NLS, cells expressing pif1-NLSΔ should have impaired nuclear functions but normal mitochondrial activity. Yeast requires functioning mitochondria to grow on glycerol media. As expected, pif1-NLSΔ and pif1-NLSΔ+SV40 cells grew as well as WT cells on media containing glycerol as the sole carbon source (Fig 2). This finding indicates that deletion of 781KKRK784 did not affect Pif1 mitochondrial function. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 2. The Pif1 NLS is not required for its mitochondrial function(s). Ten-fold serial dilutions of pif1Δ cells carrying empty plasmid or plasmids with the indicated PIF1 allele were spotted on plates containing glucose (left) or glycerol (right). The genomic copy of PIF1 was replaced with NatMX. https://doi.org/10.1371/journal.pgen.1010853.g002 Amino acids 781KKRK784 of Pif1 are essential for suppressing telomere lengthening To determine if 781KKRK784 is a functional Pif1 NLS, we examined three nuclear phenotypes: telomere length, BIR, and Okazaki fragment processing in pif1-NLSΔ and pif1-NLSΔ+SV40 cells as well as Pif1 binding to specific sites in nuclear and mitochondrial DNA. Because Pif1 displaces telomerase from telomeres [26], pif1Δ cells have long telomeres [4]. Consistent with a lack of nuclear Pif1, pif1-NLSΔ cells had long telomeres that were similar in length to those in pif1Δ cells (Fig 3). This long telomere phenotype was suppressed in pif1-NLSΔ+SV40 cells. The ability of the SV40 NLS to suppress the telomere defects of pif1-NLSΔ cells supports our hypothesis that 781KKRK784 is the core of a functional NLS. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 3. The Pif1 NLS region is required to maintain wild type telomere length. Analysis of telomere lengths in two independent isolates of cells expressing PIF1 (WT) and various pif1 mutants. Isolates were streaked at least six times to allow the full impact of each gene on telomere length. Genomic DNA was digested with XhoI and analyzed by Southern hybridization using a probe specific to the Y’ sub-telomeric region. https://doi.org/10.1371/journal.pgen.1010853.g003 Pif1’s role in break-induced replication (BIR) requires amino acids 781KKRK784 Pif1 is critical for break-induced replication (BIR) [11,12]. Its proposed roles during BIR include unwinding the DNA duplex to allow the progression of BIR, unwinding the D-loop formed by the newly synthesized DNA, and stabilizing Polδ to enhance BIR processivity [12]. To test the effects of deleting 781KKRK784 on BIR [12,27,28], BIR efficiency was determined in derivatives of AM1003 [27], a strain that is well established for BIR assays (Fig 4A). AM1003 and its derivatives are otherwise haploid cells that are disomic for chromosome III and contain a galactose-inducible copy of the HO endonuclease (Fig 4A and S1 Table). One copy of chromosome III, the donor chromosome, is full-length but has a mutant HO recognition site, the MATα-inc allele, that is not cleaved by the HO endonuclease. The other copy of chromosome III (recipient) is truncated at the Z-region of MATa but contains an intact Ya-region and HO recognition site [27]. When galactose is present and HO is expressed, a DSB is induced at MATa on the truncated chromosome III. This DSB is repaired predominantly by BIR that is initiated by strand invasion into the full-length copy of chromosome III (donor). In this system, BIR copying must proceed for ~100 kb to produce a complete BIR outcome, which generates Ade+ Leu- cells (Fig 4A). This system also allows the identification of failed BIR events. In particular, failed DSB repair often leads to the loss of the truncated chromosome and formation of Ade-red Leu- colonies (chromosome loss, CL). Another phenotype of failed BIR is Ade-white Leu- which is indicative of half-crossover (HC). HCs result from interruption of BIR synthesis, which leads to the fusion between parts of the recipient and donor chromosomes and to the loss of other parts (Fig 4A). In WT (PIF1), the majority of DSBs (~ 75%) were repaired by BIR as monitored by the formation of Ade+Leu- colonies, while failed BIR events were less frequent (~9% of CL and ~4% of HC) (Fig 4B and S2 Table). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 4. The effects of pif1-NLSΔ on long-range BIR. (A) Schematic of the BIR genetic assay in a yeast disomic system (AM1003). Repair outcomes are distinguished by genetic markers. (B) The distribution of repair events after DSB induction in WT and indicated mutant strains. Asterisks represent statistically significant difference from WT (PIF1) as determined by contingency test (****: p<0.0001; **: p = 0.0016). (C) The fractions of GCR and no-GCR events among Ade+Leu- repair outcomes as determined by CHEF gel electrophoresis analysis. Asterisks indicates significant differences from WT and N/A = not applicable; NS = not significant (D) The distribution of all repair DSB events, including calculated % of three classes contributing to the formation of Ade+Leu- outcomes (from B), including BIR, half-crossover II (HC-II), and GCRs. The amount of GCRs was calculated by applying the fraction of GCR calculated from C to the total Ade+Leu- from B. The number of HC-II was assumed to be equal to HC-I based on B [27]. (E) An illustration demonstrating the formation of half-crossover events followed by two scenarios of chromatid segregation at mitosis leading to the formation of two half-crossover classes (HC-I and HC-II). In the scenarios shown, only one broken chromatid was repaired while the other was lost. Labels are provided for sister chromatids of both chromosome III homologs (1A, 1B, 2A, and 2B). See Materials and Methods and [27] for details. https://doi.org/10.1371/journal.pgen.1010853.g004 As shown previously [12,28], the fraction of Ade+Leu- outcomes in pif1Δ cells, was significantly reduced (Fig 4B and S2 Table) indicative of defective BIR. Meanwhile, DSB repair classes representing failed BIR (HC and CL), comprised ~40% of all cases, which was significantly higher than in WT cells (p<0.0001). Analyses of DSB repair in the pif1-NLSΔ mutant showed that the distribution of its DSB repair outcomes was similar to that of pif1Δ cells (Fig 4B and S2 Table). In particular, the fraction of Ade+Leu- cases was ~45% (significantly lower than in PIF1 (WT); p < 0.0001), while the fraction of failed BIR events (HC, and CL) was ~45%, significantly higher than in PIF1 (WT) (p < 0.0001). Importantly, addition of the SV40-NLS largely compensated for the BIR defects in pif1-NLSΔ cells (Fig 4B and S2 Table). These data show that 781KKRK784 is needed for efficient BIR completion, consistent with the hypothesis that it is the core of a functional NLS. In addition, another separation of function pif1-m2 mutation and the mutant containing two mutations pif1-m2 and pif1-NLSΔ were as defective in completion of BIR as pif1Δ and pif1-NLSΔ (Fig 4B and S2 Table). BIR in pif1-NLSΔ leads to chromosomal rearrangements We previously demonstrated that while the majority of Ade+Leu- colonies in WT cells result from completed BIR events, Ade+Leu- colonies in pif1Δ cells frequently represent gross-chromosomal rearrangements (GCRs) [28]. To determine the fraction of GCRs among Ade+Leu- colonies, we used contour-clamped homogeneous electric field (CHEF) gel electrophoresis (similar to described in [27]). As expected, we observed a significantly higher fraction of GCR events among Ade+Leu- colonies in pif1Δ cells (10/23, 43%) as compared to WT cells (1/30, 3%) (see Fig 4C for p-values). Likewise, ~36% of Ade+Leu- events were due to GCR events in pif1-m2 cells. Similar to pif1Δ, ~33% and ~31% of the Ade+Leu- colonies were due to GCR events in pif1-NLSΔ and pif1-m2+NLSΔ cells respectively, again, significantly higher than in WT (Fig 4C). In contrast, the frequency of GCR events in pif1-NLSΔ+SV40 cells was comparable to that of WT cells (Fig 4C). Moreover, we previously proposed [27] that many of even non-rearranged Ade+Leu- colonies in pif1 mutants likely resulted not from completed BIR events, but from the second type of half-crossovers that we call HC-II (see Fig 4D and schematics in Fig 4E for explanation). In brief, HC-II events result from mitotic segregation of half-crossover chromosomes with an intact copy of the donor chromosome, which leads to the formation of Ade+Leu- colonies. Assuming that the frequencies of Ade-white Leu- half-crossover (HC-I) and of HC-II classes are equal [27], we can estimate the fraction of HC-II (invisible) class among Ade+Leu- events for each strain background. Following the subtraction of GCR and HC-II from Ade+Leu- events, it appears that real BIR events are rarely completed in all pif1 mutants including pif1-NLSΔ, but successfully completed in pif1-NLSΔ+SV40 (similarly to WT) (Fig 4D). Together, we conclude that the pif1-NLSΔ and pif1Δ cells are similarly defective in long range BIR, which requires extensive DNA synthesis (up to 100kb) and which is detected by the BIR assay used here. pif1-NLSΔ disrupts BIR-associated mutagenesis To determine if pif1-NLSΔ cells are also defective at earlier stages of BIR, we monitored BIR progression within the first 16kb away from the double strand break by measuring the frequency of BIR-associated mutagenesis [28]. For these experiments, we used a lys2::A4 reporter gene inserted 16kb centromere-distal to MATα-inc. In this system (Fig 5A), Lys+ cells can be generated by frameshift mutations produced within the reporter gene, which are highly stimulated during BIR DNA synthesis [28,29]. Unlike the long-range BIR assay (Fig 4A), a positive signal in the mutagenesis assay (Fig 5A) requires that BIR proceed for only 16 kb. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 5. The effects of pif1-NLSΔ on BIR-associated frameshift mutagenesis. (A) Schematic of BIR-associated mutagenesis assay. The rate of mutagenesis was measured by the frequency of Lys- to Lys+ reversion using a lys2::A4 reporter: A -1bp frameshift within the lys2::A4 reporter generates a Lys+ phenotype. The reporter is inserted 16kb centromere distal to MATα-inc on the donor chromosome. Thus, BIR events that proceed at least 16 kb can generate Lys+ cells (I) and results in high rate of Lys+. When BIR does not reach 16kb, it cannot generate Lys+ cells, and the rate of Lys+ is low (II). (B) The rate of Lys+ events measured 7hrs after a galactose-induced DSB in cells expressing different PIF1 alleles. Asterisks (*) indicates statistically significant differences from WT (PIF1), determined by the Mann-Whitney test (****: p<0.0001; **: p = 0.0089). See also S3 and S4 Tables for the rates of Lys+ reversion prior to DSB and for the details of statistical analysis. https://doi.org/10.1371/journal.pgen.1010853.g005 As in earlier studies [28,29], BIR was associated with a high rate, 3.9x10-6 of Lys+ events in WT (PIF1) cells (Fig 5B and S3 and S4 Tables), and this high rate was galactose dependent (compare 7h versus 0h in S3 and S4 Tables). In contrast, pif1Δ cells had a 27x times lower BIR-associated Lys+ mutagenesis than PIF1 (WT) cells (See S3 and S4 Tables) [29]. Thus, in the absence of Pif1, BIR reaches the 16kb position in only ~4% of the events. Consistent with earlier observations [28], BIR-associated frameshifts at 16kb were 5x more frequent in pif1-m2 than in pif1Δ cells, and this level was 5x lower than in WT (Fig 5B and S3 and S4 Tables), confirming previous data that pif1-m2 is not a null allele in this assay [8,28,30]. The level of BIR-associated mutagenesis in pif1-NLSΔ cells was 2.3x lower than in WT (p<0.0001), and 2.2x higher than in pif1-m2 cells (p<0.0001) (Fig 5B and S3 and S4 Tables). Thus, like pif1-m2, pif1-NLSΔ was defective but not a null in the mutagenesis assay. Moreover, pif1-NLSΔ cells were not as deficient as pif1-m2 cells. In addition, the pif1-m2+NLSΔ double mutant was significantly more defective as compared to either single mutant (Fig 5B and S3 and S4 Tables). Nonetheless, even this double mutant was 3.6x less defective than pif1Δ cells (p< 0.0001 for all comparisons). Thus, there must be residual nuclear Pif1 even in pif1-m2+NLSΔ cells. Finally, the level of mutagenesis was restored in pif1-NLSΔ+SV40 compared to pif1-NLSΔ (p<0.0001), but not to the wild type level (p = 0.0089) (Fig 5B and S3 and S4 Tables). Together, we conclude that all four pif1 mutants, including pif1-NLSΔ, were highly BIR defective as seen by both long- and short-range BIR assays. However, results from the short-range BIR assay indicated that the three pif1 mutants are not equally defective for BIR. BIR-associated DNA synthesis is defective in pif1-NLSΔ cells Although both the long- and short-range BIR assays revealed that pif1-NLSΔ cells are BIR-defective, neither assay measures BIR synthesis directly. Therefore, we used the digital droplet PCR (ddPCR) based AMBER (Assay for monitoring BIR elongation rate) assay, which precisely measures the amount of DNA synthesized during individual BIR events [31,32] by monitoring copy number changes at different BIR positions and at multiple time points after galactose addition (Fig 6A, see legend for details). AMBER uses donor-specific primer sets for three separate locations within a 61 kb BIR region and then normalizes these values to the signals obtained at the ACT1 locus to determine relative copy number. Primer sets were located at 1.9kb (P2), 22.4kb (P3), and 61.4kb (P4) centromere distal to the DSB. Copy numbers at these positions are 1x before BIR and can increase to a maximum of 2x following BIR. Evidence of DNA synthesis was defined as an increase in donor DNA copy number of at least 1.1x. At 10hrs after DSB induction, WT PIF1 cells had a 1.7x copy number at P2 and P3 and 1.6x at P4 (See Fig 6B and 6C and S5 Table). In contrast, in pif1Δ cells, at 10 hrs, increased copy number was detected only at P2 and was only ~ 1.1x (Fig 6B and 6C and S5 Table), results similar to an earlier report [32]. In pif1-m2 cells at 10 hrs, a small increase was observed at both P2 and P3 (~1.1x; Fig 6B and 6C), but not at P4. At 10 hrs, pif1-NLSΔ cells had increased copy number at all three sites (~1.3x at P2 and ~1.1x at P3 and P4; Fig 6B and 6C). Since DNA synthesis in Pif1-defective mutants is often incomplete and therefore some newly synthesized DNA may degrade by 10 hours, we also compared the maximum copy number increases reached in different strains between 0 and 10 hrs (see Fig 6D) and came to the same conclusions. In addition, as expected, the copy number was higher in pif1-NLSΔ+SV40 (~1.5x at P2 and ~1.4 at P3 and P4) than in pif1-NLSΔ cells. The extent of BIR synthesis was similar in pif1-m2+NLSΔ and pif1Δ cells (~1.1x at P2; no increase detected at P3 and P4) (Fig 6B, 6C and 6D). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 6. The effects of pif1-NLSΔ on the extent of BIR synthesis detected by AMBER. (A) Schematic of AMBER analysis measuring the extent of BIR synthesis in a yeast disomic system (derivatives of AM1003). Colored triangles indicate locations of primer sets. (P1 is located at the left arm of chromosome III (~153kb away from MATα-inc) and used as a control. P2, P3, and P4 are placed 1.9kb, 22.4kb, and 61.4kb centromere distal from MATα-inc. (B) DNA synthesis measured by AMBER in WT (PIF1) and in strains containing various mutations of PIF1 during the first 10hrs following addition of galactose. The results are presented as Boltzmann sigmoidal curve. One out of three independent biological repeats that showed similar results is shown in each panel. (See S5 Table for other repeats). Mean values of target to reference (ACT1) loci ratios were calculated by Poisson distribution based on 20,000 droplets with error bars representing upper and lower Poisson 95% CI. (C, D) The comparison between the extents of BIR synthesis in various pif1 mutants at 10hrs after DSB induction (C) or the maximum detected between 0h and 10hrs (D) as assessed by AMBER. Mean±s.d (n = 3, three independent biological repeats). Note: some values are higher in D than in C due to BIR interruption prior to 10h followed by degradation of newly synthesized DNA in pif1 defective strains. https://doi.org/10.1371/journal.pgen.1010853.g006 In conclusion, cells lacking 781KKRK784 were defective in BIR initiation and progression and this defective phenotype was largely suppressed in pif1-NLSΔ+SV40 cells. In agreement with the results of BIR-associated mutagenesis assay, these results showed a hierarchy in the BIR defects of the different pif1 mutants: pif1Δ ≈ pif1-m2+NLSΔ > pif1-m2> pif1-NLSΔ>pif1-NLSΔ+SV40>WT. 781KKRK784 is important for Pif1 enrichment at nuclear, but not mitochondrial DNA binding sites The defects in three different nuclear functions suggests that the nuclear abundance of the Pif1-NLSΔ protein was lower than in PIF1 cells but not as low as in pif1Δ cells. We used chromatin immunoprecipitation (ChIP) of Pif1 to its binding sites as a proxy for Pif1 abundance. For these experiments, PIF1 (WT) and three pif1 mutant alleles were tagged at their C-termini with 13 MYC epitopes [33]. By western blot, all tagged PIF1 alleles were stably expressed (Fig 7A). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 7. Diminished enrichment of Pif1 protein lacking 781KKRK784 in the nucleus. (A) Westerns blot analysis of Myc-tagged PIF1 (WT) and pif1 mutant alleles using an anti-MYC or anti-PGK1 (loading control) antibody. (B) Pif1 enrichment measured by ChIP in WT (PIF1) and indicated pif1 mutants at two different nuclear positions (MAT (BIR-specific) and CEN13) and at mitochondrial-specific site (see Materials and Methods for details). Asterisks represent statistically significant difference as determined by unpaired t-test (***: p = 0.0005, *:p<0.05). https://doi.org/10.1371/journal.pgen.1010853.g007 We analyzed the recruitment of Pif1 during BIR at MAT at 4hrs after DSB induction (see Materials and Methods and [12] for details). As expected, we observed a high enrichment of Pif1 at MAT, while Pif1-NLSΔ MAT binding level was ~4.6x lower than that for WT Pif1 (Fig 7B). We also measured Pif1 binding in cycling cells (no DSB induction) at CEN13, a known site of high Pif1 binding [15]. Again, Pif1-NLSΔ binding to CEN13 was ~4.9x lower than WT Pif1. As predicted, defective nuclear Pif1-NLSΔ binding was restored in pif1-NLSΔ+SV40 cells at both MAT and CEN13 sites. As expected, Pif1 binding to a mitochondrial location was indistinguishable in PIF1 (WT) and three pif1 mutant cells (pif1-NLSΔ, pif1-m2+ NLSΔ, and pif1-NLSΔ+SV40) (Fig 7B). Together, ChIP data support the hypothesis that pif1-NLSΔ leads to a decrease of Pif1 localization to the nucleus but does not affect its presence in mitochondria. Partial suppression of dna2Δ lethality by pif1-NLSΔ PIF1 deletion rescues the lethality of dna2Δ cells [6]. This rescue is thought to reflect the role of Pif1 in Okazaki fragment maturation. The current model suggests that Pif1 and DNA Polymerase δ together generate long Okazaki flaps that cannot be cleaved by flap endonuclease 1 (Fen1). Rather, the long flaps are cleaved by Dna2, an essential multi-functional helicase-nuclease [34,35]. According to this model, if Pif1 is absent, long flaps are not generated, and Dna2 is no longer essential. For Pif1 to promote formation of long flap Okazaki fragments, Pif1 must be nuclear-localized. To test whether pif1-NLSΔ rescued the lethality of dna2Δ, we first used a qualitative assay where colony formation of pif1Δ dna2Δ cells transformed with centromeric plasmids containing various pif1 alleles was measured. As expected, dna2Δ cells carrying either the PIF1(WT) vector or the vector with pif1-NLSΔ+SV40 were not viable. In contrast, empty vector alone or vector carrying pif1-NLSΔ suppressed the lethality of dna2Δ (Fig 8A; compare NLSΔ (bottom) to “Empty vector” (top)). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 8. pif1-NLSΔ can partially suppress lethality of dna2Δ. (A) Schematic of experimental system for the qualitative suppression test is shown on the left. Right: representative samples of viability of pif1Δ dna2Δ cells carrying empty plasmid or plasmids with the indicated PIF1 allele. (B, C) The number and phenotypes of viable spores (colonies) obtained following meiosis of two diploids: DNA2/dna2::URA3 PIF1/pif1::KANMX and DNA2/dna2::URA3 pif1-NLSΔ /pif1::KANMX. The meiotic outcomes were analyzed by tetrad analysis (B) and by random spore analysis (C). G418s = sensitive to G418 (KANMX absent); G418r = resistant to G418 (KANMX present). https://doi.org/10.1371/journal.pgen.1010853.g008 To investigate suppression of dna2Δ in a more quantitative assay, we created two diploid strains that were both DNA2/dna2::URA3 heterozygotes. One of these strains was also PIF1/pif1::KANMX, while the other was pif1-NLSΔ/pif1::KANMX. Both diploid strains were sporulated and dissected. As expected, in the strain heterozygous for PIF1 (WT), there were no viable Ura+ G418S (dna2ΔPIF1) spores (out of 36 dissected tetrads; Fig 8B). Dissection of 35 tetrads from the pif1-NLSΔ/pif1::KANMX strain produced three viable Ura+G418S (dna2Δpif1-NLSΔ) spores while 24 Ura+ spores were G418R (dna2Δ pif1Δ). This result is indicative of a partial suppression of dna2Δ lethality by pif1-NLSΔ. As expected, in both strains, an equal number of G418R and G418S was observed among Ura- (DNA2 wt) colonies (Fig 8B). Our conclusion was further supported by the results of random spore analysis in the same diploids where among 42 Ura+ clones generated after sporulation of the pif1-NLSΔ/pif1::KANMX diploid, eight were G418S (dna2Δpif1-NLSΔ) while 34 were G418R (dna2Δpif1Δ) (Fig 8C). As expected, no G418S Ura+ (dna2ΔPIF1) colonies were observed in the strain heterozygous for PIF1 (WT) (Fig 8C). Therefore, we conclude that the deletion of PIF1 781KKRK784 partially suppresses the lethality of dna2Δ cells. Mapping the position of the Pif1 NLS The cNLS Mapper program was used to search for candidate nuclear localization signals in the Pif1 open reading frame [23]. The program assigns prediction scores of up to 10 with higher scores indicating a higher probability of being an NLS. cNLS Mapper identified two candidate NLSs within Pif1: a monopartite NLS (777DEQVKKRKLDY787; prediction score of 8) and a bipartite NLS (417RQRGDVKFIDMLNRMRLGNIDDETEREFKKLSRP450; prediction score of 5). The bipartite NLS had a relatively low prediction score, and its sequence overlapped one of the canonical helicase motifs (motif IV) as well as another conserved region (motif A) (Fig 1A). Based on these findings, we focused our analysis on the monopartite NLS candidate sequence. Strains to test whether putative NLS is required for nuclear functions To determine if the predicted monopartite NLS (777DEQVKKRKLDY787) was important for Pif1 nuclear functions, we deleted the core of this sequence, 781KKRK784 to generate the pif1-NLSΔ (NLS deleted) allele (Fig 1B). Only the core of the putative NLS was deleted to reduce the possibility of disrupting sequences important for other Pif1 functions. In addition, these four amino acids match the consensus for a classical monopartite NLS motif: K-(K/R)-X-(K/R) [24]. As a control, we fused the NLS from the simian virus 40 (SV40) T antigen (126PKKKRKV132), which functions as an NLS in S. cerevisiae [25], to the end of the pif1-NLSΔ ORF to generate the pif1-NLSΔ+SV40 allele (Fig 1B). If our hypothesis is correct, pif1-NLSΔ cells should be defective in nuclear Pif1 functions, while still retaining mitochondrial activities. In addition, nuclear defects in pif1-NLSΔ cells should be reduced in pif1-NLSΔ+SV40 cells, even though the SV40 NLS is not in the same position as the predicted Pif1 NLS. We also generated a double mutant of Pif1, pif1-m2+NLSΔ, which deleted the candidate NLS from the partial loss of function pif1-m2 allele. Amino acids 781KKRK784 are not essential for protein stability or for maintenance of mitochondrial function Western blot analysis showed that the Pif1-NLSΔ protein was stably expressed (Fig 1C). In addition, attaching the SV40 NLS to the carboxyl-terminus of either WT Pif1 or Pif1-NLSΔ did not affect the abundance of either protein (Fig 1C). If 781KKRK784 is the core of a functional NLS, cells expressing pif1-NLSΔ should have impaired nuclear functions but normal mitochondrial activity. Yeast requires functioning mitochondria to grow on glycerol media. As expected, pif1-NLSΔ and pif1-NLSΔ+SV40 cells grew as well as WT cells on media containing glycerol as the sole carbon source (Fig 2). This finding indicates that deletion of 781KKRK784 did not affect Pif1 mitochondrial function. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 2. The Pif1 NLS is not required for its mitochondrial function(s). Ten-fold serial dilutions of pif1Δ cells carrying empty plasmid or plasmids with the indicated PIF1 allele were spotted on plates containing glucose (left) or glycerol (right). The genomic copy of PIF1 was replaced with NatMX. https://doi.org/10.1371/journal.pgen.1010853.g002 Amino acids 781KKRK784 of Pif1 are essential for suppressing telomere lengthening To determine if 781KKRK784 is a functional Pif1 NLS, we examined three nuclear phenotypes: telomere length, BIR, and Okazaki fragment processing in pif1-NLSΔ and pif1-NLSΔ+SV40 cells as well as Pif1 binding to specific sites in nuclear and mitochondrial DNA. Because Pif1 displaces telomerase from telomeres [26], pif1Δ cells have long telomeres [4]. Consistent with a lack of nuclear Pif1, pif1-NLSΔ cells had long telomeres that were similar in length to those in pif1Δ cells (Fig 3). This long telomere phenotype was suppressed in pif1-NLSΔ+SV40 cells. The ability of the SV40 NLS to suppress the telomere defects of pif1-NLSΔ cells supports our hypothesis that 781KKRK784 is the core of a functional NLS. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 3. The Pif1 NLS region is required to maintain wild type telomere length. Analysis of telomere lengths in two independent isolates of cells expressing PIF1 (WT) and various pif1 mutants. Isolates were streaked at least six times to allow the full impact of each gene on telomere length. Genomic DNA was digested with XhoI and analyzed by Southern hybridization using a probe specific to the Y’ sub-telomeric region. https://doi.org/10.1371/journal.pgen.1010853.g003 Pif1’s role in break-induced replication (BIR) requires amino acids 781KKRK784 Pif1 is critical for break-induced replication (BIR) [11,12]. Its proposed roles during BIR include unwinding the DNA duplex to allow the progression of BIR, unwinding the D-loop formed by the newly synthesized DNA, and stabilizing Polδ to enhance BIR processivity [12]. To test the effects of deleting 781KKRK784 on BIR [12,27,28], BIR efficiency was determined in derivatives of AM1003 [27], a strain that is well established for BIR assays (Fig 4A). AM1003 and its derivatives are otherwise haploid cells that are disomic for chromosome III and contain a galactose-inducible copy of the HO endonuclease (Fig 4A and S1 Table). One copy of chromosome III, the donor chromosome, is full-length but has a mutant HO recognition site, the MATα-inc allele, that is not cleaved by the HO endonuclease. The other copy of chromosome III (recipient) is truncated at the Z-region of MATa but contains an intact Ya-region and HO recognition site [27]. When galactose is present and HO is expressed, a DSB is induced at MATa on the truncated chromosome III. This DSB is repaired predominantly by BIR that is initiated by strand invasion into the full-length copy of chromosome III (donor). In this system, BIR copying must proceed for ~100 kb to produce a complete BIR outcome, which generates Ade+ Leu- cells (Fig 4A). This system also allows the identification of failed BIR events. In particular, failed DSB repair often leads to the loss of the truncated chromosome and formation of Ade-red Leu- colonies (chromosome loss, CL). Another phenotype of failed BIR is Ade-white Leu- which is indicative of half-crossover (HC). HCs result from interruption of BIR synthesis, which leads to the fusion between parts of the recipient and donor chromosomes and to the loss of other parts (Fig 4A). In WT (PIF1), the majority of DSBs (~ 75%) were repaired by BIR as monitored by the formation of Ade+Leu- colonies, while failed BIR events were less frequent (~9% of CL and ~4% of HC) (Fig 4B and S2 Table). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 4. The effects of pif1-NLSΔ on long-range BIR. (A) Schematic of the BIR genetic assay in a yeast disomic system (AM1003). Repair outcomes are distinguished by genetic markers. (B) The distribution of repair events after DSB induction in WT and indicated mutant strains. Asterisks represent statistically significant difference from WT (PIF1) as determined by contingency test (****: p<0.0001; **: p = 0.0016). (C) The fractions of GCR and no-GCR events among Ade+Leu- repair outcomes as determined by CHEF gel electrophoresis analysis. Asterisks indicates significant differences from WT and N/A = not applicable; NS = not significant (D) The distribution of all repair DSB events, including calculated % of three classes contributing to the formation of Ade+Leu- outcomes (from B), including BIR, half-crossover II (HC-II), and GCRs. The amount of GCRs was calculated by applying the fraction of GCR calculated from C to the total Ade+Leu- from B. The number of HC-II was assumed to be equal to HC-I based on B [27]. (E) An illustration demonstrating the formation of half-crossover events followed by two scenarios of chromatid segregation at mitosis leading to the formation of two half-crossover classes (HC-I and HC-II). In the scenarios shown, only one broken chromatid was repaired while the other was lost. Labels are provided for sister chromatids of both chromosome III homologs (1A, 1B, 2A, and 2B). See Materials and Methods and [27] for details. https://doi.org/10.1371/journal.pgen.1010853.g004 As shown previously [12,28], the fraction of Ade+Leu- outcomes in pif1Δ cells, was significantly reduced (Fig 4B and S2 Table) indicative of defective BIR. Meanwhile, DSB repair classes representing failed BIR (HC and CL), comprised ~40% of all cases, which was significantly higher than in WT cells (p<0.0001). Analyses of DSB repair in the pif1-NLSΔ mutant showed that the distribution of its DSB repair outcomes was similar to that of pif1Δ cells (Fig 4B and S2 Table). In particular, the fraction of Ade+Leu- cases was ~45% (significantly lower than in PIF1 (WT); p < 0.0001), while the fraction of failed BIR events (HC, and CL) was ~45%, significantly higher than in PIF1 (WT) (p < 0.0001). Importantly, addition of the SV40-NLS largely compensated for the BIR defects in pif1-NLSΔ cells (Fig 4B and S2 Table). These data show that 781KKRK784 is needed for efficient BIR completion, consistent with the hypothesis that it is the core of a functional NLS. In addition, another separation of function pif1-m2 mutation and the mutant containing two mutations pif1-m2 and pif1-NLSΔ were as defective in completion of BIR as pif1Δ and pif1-NLSΔ (Fig 4B and S2 Table). BIR in pif1-NLSΔ leads to chromosomal rearrangements We previously demonstrated that while the majority of Ade+Leu- colonies in WT cells result from completed BIR events, Ade+Leu- colonies in pif1Δ cells frequently represent gross-chromosomal rearrangements (GCRs) [28]. To determine the fraction of GCRs among Ade+Leu- colonies, we used contour-clamped homogeneous electric field (CHEF) gel electrophoresis (similar to described in [27]). As expected, we observed a significantly higher fraction of GCR events among Ade+Leu- colonies in pif1Δ cells (10/23, 43%) as compared to WT cells (1/30, 3%) (see Fig 4C for p-values). Likewise, ~36% of Ade+Leu- events were due to GCR events in pif1-m2 cells. Similar to pif1Δ, ~33% and ~31% of the Ade+Leu- colonies were due to GCR events in pif1-NLSΔ and pif1-m2+NLSΔ cells respectively, again, significantly higher than in WT (Fig 4C). In contrast, the frequency of GCR events in pif1-NLSΔ+SV40 cells was comparable to that of WT cells (Fig 4C). Moreover, we previously proposed [27] that many of even non-rearranged Ade+Leu- colonies in pif1 mutants likely resulted not from completed BIR events, but from the second type of half-crossovers that we call HC-II (see Fig 4D and schematics in Fig 4E for explanation). In brief, HC-II events result from mitotic segregation of half-crossover chromosomes with an intact copy of the donor chromosome, which leads to the formation of Ade+Leu- colonies. Assuming that the frequencies of Ade-white Leu- half-crossover (HC-I) and of HC-II classes are equal [27], we can estimate the fraction of HC-II (invisible) class among Ade+Leu- events for each strain background. Following the subtraction of GCR and HC-II from Ade+Leu- events, it appears that real BIR events are rarely completed in all pif1 mutants including pif1-NLSΔ, but successfully completed in pif1-NLSΔ+SV40 (similarly to WT) (Fig 4D). Together, we conclude that the pif1-NLSΔ and pif1Δ cells are similarly defective in long range BIR, which requires extensive DNA synthesis (up to 100kb) and which is detected by the BIR assay used here. pif1-NLSΔ disrupts BIR-associated mutagenesis To determine if pif1-NLSΔ cells are also defective at earlier stages of BIR, we monitored BIR progression within the first 16kb away from the double strand break by measuring the frequency of BIR-associated mutagenesis [28]. For these experiments, we used a lys2::A4 reporter gene inserted 16kb centromere-distal to MATα-inc. In this system (Fig 5A), Lys+ cells can be generated by frameshift mutations produced within the reporter gene, which are highly stimulated during BIR DNA synthesis [28,29]. Unlike the long-range BIR assay (Fig 4A), a positive signal in the mutagenesis assay (Fig 5A) requires that BIR proceed for only 16 kb. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 5. The effects of pif1-NLSΔ on BIR-associated frameshift mutagenesis. (A) Schematic of BIR-associated mutagenesis assay. The rate of mutagenesis was measured by the frequency of Lys- to Lys+ reversion using a lys2::A4 reporter: A -1bp frameshift within the lys2::A4 reporter generates a Lys+ phenotype. The reporter is inserted 16kb centromere distal to MATα-inc on the donor chromosome. Thus, BIR events that proceed at least 16 kb can generate Lys+ cells (I) and results in high rate of Lys+. When BIR does not reach 16kb, it cannot generate Lys+ cells, and the rate of Lys+ is low (II). (B) The rate of Lys+ events measured 7hrs after a galactose-induced DSB in cells expressing different PIF1 alleles. Asterisks (*) indicates statistically significant differences from WT (PIF1), determined by the Mann-Whitney test (****: p<0.0001; **: p = 0.0089). See also S3 and S4 Tables for the rates of Lys+ reversion prior to DSB and for the details of statistical analysis. https://doi.org/10.1371/journal.pgen.1010853.g005 As in earlier studies [28,29], BIR was associated with a high rate, 3.9x10-6 of Lys+ events in WT (PIF1) cells (Fig 5B and S3 and S4 Tables), and this high rate was galactose dependent (compare 7h versus 0h in S3 and S4 Tables). In contrast, pif1Δ cells had a 27x times lower BIR-associated Lys+ mutagenesis than PIF1 (WT) cells (See S3 and S4 Tables) [29]. Thus, in the absence of Pif1, BIR reaches the 16kb position in only ~4% of the events. Consistent with earlier observations [28], BIR-associated frameshifts at 16kb were 5x more frequent in pif1-m2 than in pif1Δ cells, and this level was 5x lower than in WT (Fig 5B and S3 and S4 Tables), confirming previous data that pif1-m2 is not a null allele in this assay [8,28,30]. The level of BIR-associated mutagenesis in pif1-NLSΔ cells was 2.3x lower than in WT (p<0.0001), and 2.2x higher than in pif1-m2 cells (p<0.0001) (Fig 5B and S3 and S4 Tables). Thus, like pif1-m2, pif1-NLSΔ was defective but not a null in the mutagenesis assay. Moreover, pif1-NLSΔ cells were not as deficient as pif1-m2 cells. In addition, the pif1-m2+NLSΔ double mutant was significantly more defective as compared to either single mutant (Fig 5B and S3 and S4 Tables). Nonetheless, even this double mutant was 3.6x less defective than pif1Δ cells (p< 0.0001 for all comparisons). Thus, there must be residual nuclear Pif1 even in pif1-m2+NLSΔ cells. Finally, the level of mutagenesis was restored in pif1-NLSΔ+SV40 compared to pif1-NLSΔ (p<0.0001), but not to the wild type level (p = 0.0089) (Fig 5B and S3 and S4 Tables). Together, we conclude that all four pif1 mutants, including pif1-NLSΔ, were highly BIR defective as seen by both long- and short-range BIR assays. However, results from the short-range BIR assay indicated that the three pif1 mutants are not equally defective for BIR. BIR-associated DNA synthesis is defective in pif1-NLSΔ cells Although both the long- and short-range BIR assays revealed that pif1-NLSΔ cells are BIR-defective, neither assay measures BIR synthesis directly. Therefore, we used the digital droplet PCR (ddPCR) based AMBER (Assay for monitoring BIR elongation rate) assay, which precisely measures the amount of DNA synthesized during individual BIR events [31,32] by monitoring copy number changes at different BIR positions and at multiple time points after galactose addition (Fig 6A, see legend for details). AMBER uses donor-specific primer sets for three separate locations within a 61 kb BIR region and then normalizes these values to the signals obtained at the ACT1 locus to determine relative copy number. Primer sets were located at 1.9kb (P2), 22.4kb (P3), and 61.4kb (P4) centromere distal to the DSB. Copy numbers at these positions are 1x before BIR and can increase to a maximum of 2x following BIR. Evidence of DNA synthesis was defined as an increase in donor DNA copy number of at least 1.1x. At 10hrs after DSB induction, WT PIF1 cells had a 1.7x copy number at P2 and P3 and 1.6x at P4 (See Fig 6B and 6C and S5 Table). In contrast, in pif1Δ cells, at 10 hrs, increased copy number was detected only at P2 and was only ~ 1.1x (Fig 6B and 6C and S5 Table), results similar to an earlier report [32]. In pif1-m2 cells at 10 hrs, a small increase was observed at both P2 and P3 (~1.1x; Fig 6B and 6C), but not at P4. At 10 hrs, pif1-NLSΔ cells had increased copy number at all three sites (~1.3x at P2 and ~1.1x at P3 and P4; Fig 6B and 6C). Since DNA synthesis in Pif1-defective mutants is often incomplete and therefore some newly synthesized DNA may degrade by 10 hours, we also compared the maximum copy number increases reached in different strains between 0 and 10 hrs (see Fig 6D) and came to the same conclusions. In addition, as expected, the copy number was higher in pif1-NLSΔ+SV40 (~1.5x at P2 and ~1.4 at P3 and P4) than in pif1-NLSΔ cells. The extent of BIR synthesis was similar in pif1-m2+NLSΔ and pif1Δ cells (~1.1x at P2; no increase detected at P3 and P4) (Fig 6B, 6C and 6D). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 6. The effects of pif1-NLSΔ on the extent of BIR synthesis detected by AMBER. (A) Schematic of AMBER analysis measuring the extent of BIR synthesis in a yeast disomic system (derivatives of AM1003). Colored triangles indicate locations of primer sets. (P1 is located at the left arm of chromosome III (~153kb away from MATα-inc) and used as a control. P2, P3, and P4 are placed 1.9kb, 22.4kb, and 61.4kb centromere distal from MATα-inc. (B) DNA synthesis measured by AMBER in WT (PIF1) and in strains containing various mutations of PIF1 during the first 10hrs following addition of galactose. The results are presented as Boltzmann sigmoidal curve. One out of three independent biological repeats that showed similar results is shown in each panel. (See S5 Table for other repeats). Mean values of target to reference (ACT1) loci ratios were calculated by Poisson distribution based on 20,000 droplets with error bars representing upper and lower Poisson 95% CI. (C, D) The comparison between the extents of BIR synthesis in various pif1 mutants at 10hrs after DSB induction (C) or the maximum detected between 0h and 10hrs (D) as assessed by AMBER. Mean±s.d (n = 3, three independent biological repeats). Note: some values are higher in D than in C due to BIR interruption prior to 10h followed by degradation of newly synthesized DNA in pif1 defective strains. https://doi.org/10.1371/journal.pgen.1010853.g006 In conclusion, cells lacking 781KKRK784 were defective in BIR initiation and progression and this defective phenotype was largely suppressed in pif1-NLSΔ+SV40 cells. In agreement with the results of BIR-associated mutagenesis assay, these results showed a hierarchy in the BIR defects of the different pif1 mutants: pif1Δ ≈ pif1-m2+NLSΔ > pif1-m2> pif1-NLSΔ>pif1-NLSΔ+SV40>WT. 781KKRK784 is important for Pif1 enrichment at nuclear, but not mitochondrial DNA binding sites The defects in three different nuclear functions suggests that the nuclear abundance of the Pif1-NLSΔ protein was lower than in PIF1 cells but not as low as in pif1Δ cells. We used chromatin immunoprecipitation (ChIP) of Pif1 to its binding sites as a proxy for Pif1 abundance. For these experiments, PIF1 (WT) and three pif1 mutant alleles were tagged at their C-termini with 13 MYC epitopes [33]. By western blot, all tagged PIF1 alleles were stably expressed (Fig 7A). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 7. Diminished enrichment of Pif1 protein lacking 781KKRK784 in the nucleus. (A) Westerns blot analysis of Myc-tagged PIF1 (WT) and pif1 mutant alleles using an anti-MYC or anti-PGK1 (loading control) antibody. (B) Pif1 enrichment measured by ChIP in WT (PIF1) and indicated pif1 mutants at two different nuclear positions (MAT (BIR-specific) and CEN13) and at mitochondrial-specific site (see Materials and Methods for details). Asterisks represent statistically significant difference as determined by unpaired t-test (***: p = 0.0005, *:p<0.05). https://doi.org/10.1371/journal.pgen.1010853.g007 We analyzed the recruitment of Pif1 during BIR at MAT at 4hrs after DSB induction (see Materials and Methods and [12] for details). As expected, we observed a high enrichment of Pif1 at MAT, while Pif1-NLSΔ MAT binding level was ~4.6x lower than that for WT Pif1 (Fig 7B). We also measured Pif1 binding in cycling cells (no DSB induction) at CEN13, a known site of high Pif1 binding [15]. Again, Pif1-NLSΔ binding to CEN13 was ~4.9x lower than WT Pif1. As predicted, defective nuclear Pif1-NLSΔ binding was restored in pif1-NLSΔ+SV40 cells at both MAT and CEN13 sites. As expected, Pif1 binding to a mitochondrial location was indistinguishable in PIF1 (WT) and three pif1 mutant cells (pif1-NLSΔ, pif1-m2+ NLSΔ, and pif1-NLSΔ+SV40) (Fig 7B). Together, ChIP data support the hypothesis that pif1-NLSΔ leads to a decrease of Pif1 localization to the nucleus but does not affect its presence in mitochondria. Partial suppression of dna2Δ lethality by pif1-NLSΔ PIF1 deletion rescues the lethality of dna2Δ cells [6]. This rescue is thought to reflect the role of Pif1 in Okazaki fragment maturation. The current model suggests that Pif1 and DNA Polymerase δ together generate long Okazaki flaps that cannot be cleaved by flap endonuclease 1 (Fen1). Rather, the long flaps are cleaved by Dna2, an essential multi-functional helicase-nuclease [34,35]. According to this model, if Pif1 is absent, long flaps are not generated, and Dna2 is no longer essential. For Pif1 to promote formation of long flap Okazaki fragments, Pif1 must be nuclear-localized. To test whether pif1-NLSΔ rescued the lethality of dna2Δ, we first used a qualitative assay where colony formation of pif1Δ dna2Δ cells transformed with centromeric plasmids containing various pif1 alleles was measured. As expected, dna2Δ cells carrying either the PIF1(WT) vector or the vector with pif1-NLSΔ+SV40 were not viable. In contrast, empty vector alone or vector carrying pif1-NLSΔ suppressed the lethality of dna2Δ (Fig 8A; compare NLSΔ (bottom) to “Empty vector” (top)). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 8. pif1-NLSΔ can partially suppress lethality of dna2Δ. (A) Schematic of experimental system for the qualitative suppression test is shown on the left. Right: representative samples of viability of pif1Δ dna2Δ cells carrying empty plasmid or plasmids with the indicated PIF1 allele. (B, C) The number and phenotypes of viable spores (colonies) obtained following meiosis of two diploids: DNA2/dna2::URA3 PIF1/pif1::KANMX and DNA2/dna2::URA3 pif1-NLSΔ /pif1::KANMX. The meiotic outcomes were analyzed by tetrad analysis (B) and by random spore analysis (C). G418s = sensitive to G418 (KANMX absent); G418r = resistant to G418 (KANMX present). https://doi.org/10.1371/journal.pgen.1010853.g008 To investigate suppression of dna2Δ in a more quantitative assay, we created two diploid strains that were both DNA2/dna2::URA3 heterozygotes. One of these strains was also PIF1/pif1::KANMX, while the other was pif1-NLSΔ/pif1::KANMX. Both diploid strains were sporulated and dissected. As expected, in the strain heterozygous for PIF1 (WT), there were no viable Ura+ G418S (dna2ΔPIF1) spores (out of 36 dissected tetrads; Fig 8B). Dissection of 35 tetrads from the pif1-NLSΔ/pif1::KANMX strain produced three viable Ura+G418S (dna2Δpif1-NLSΔ) spores while 24 Ura+ spores were G418R (dna2Δ pif1Δ). This result is indicative of a partial suppression of dna2Δ lethality by pif1-NLSΔ. As expected, in both strains, an equal number of G418R and G418S was observed among Ura- (DNA2 wt) colonies (Fig 8B). Our conclusion was further supported by the results of random spore analysis in the same diploids where among 42 Ura+ clones generated after sporulation of the pif1-NLSΔ/pif1::KANMX diploid, eight were G418S (dna2Δpif1-NLSΔ) while 34 were G418R (dna2Δpif1Δ) (Fig 8C). As expected, no G418S Ura+ (dna2ΔPIF1) colonies were observed in the strain heterozygous for PIF1 (WT) (Fig 8C). Therefore, we conclude that the deletion of PIF1 781KKRK784 partially suppresses the lethality of dna2Δ cells. Discussion The multifunctional Pif1 helicase is critical for maintenance of both nuclear and mitochondrial DNA. Because pif1Δ cells grow slowly, pif1Δ strains are not ideal for studying nuclear Pif1 functions. Therefore, we sought to identify separation of function alleles that only affect nuclear Pif1. Using cNLS Mapper to identify a putative NLS in Pif1, we generated the pif1-NLSΔ allele. As shown here, pif1-NLSΔ is a separation of function allele, defective in all tested nuclear functions but having normal localization and function in mitochondria. Together, these data make a strong argument that 781KKRK784 is the core of a functional NLS. The interpretation that pif1-NLSΔ cells are defective in nuclear localization of Pif1 is supported by multiple findings. First, cNLS Mapper identified the deleted sequence as having a high probability of being an NLS. Moreover, the core of the predicted NLS, 781KKRK784, matches the consensus for a classical monopartite NLS motif: K-(K/R)-X-(K/R) [24]. Second, deletion of 781KKRK784 impaired four distinct nuclear functions: telomere length regulation (Fig 3), short- and long-range BIR (Figs 4,5 and 6), binding to nuclear target sites (Fig 7), and Okazaki fragment maturation (Fig 8). Third, all four of the nuclear defects documented in pif1-NLSΔ cells were suppressed by addition of the NLS from the SV40 T antigen, which was inserted at a different site within Pif1 than that of the endogenous 781KKRK784 motif (Fig 1B). Suppression of defects by a known NLS inserted at a different site in the protein is a classical criterion for identification of an NLS as established by studies in diverse organisms. Fourth, Pif1 localization and activity in mitochondria was not impaired in pif1-NLSΔ cells (Figs 2 and 7), as expected if the mutation affects only nuclear localization of Pif1. Fifth, Pif1 binding to nuclear but not mitochondrial binding sites was significantly reduced in pif1-NLSΔ cells (Fig 7). Taken together, these experiments make a strong argument that 781KKRK784 is a functional NLS. One could still argue that rather than being an NLS, 781KKRK784 might be important for a function other than nuclear localization. Although we cannot completely exclude this possibility, deleting 781KKRK784 does not compromise the known biochemical activity of Pif1. A purified mutant Pif1 lacking its terminal 79 amino acids, including 781KKRK784, has normal ATPase activity in vitro [22]. Therefore, the key biochemical activity of the Pif1-NLSΔ protein is almost surely uncompromised in pif1-NLSΔ cells. This interpretation is supported by our finding that pif1-NLSΔ cells are mitochondrial proficient (Fig 2), while ATPase-dead pif1-K264A cells are mitochondrial (as well as nuclear) defective [5]. Another possibility is that 781KKRK784 is required for interaction with a protein(s) required for all four of the nuclear functions assessed in our experiments. Again, we cannot rule out this possibility completely, but we consider it unlikely. First and most critically, we observed that all four nuclear functions were restored by attaching the SV40 NLS, 126PKKKRKV132, to the Pif1-NLSΔ protein. Following earlier studies on NLS characterization, the inserted SV40 NLS was placed at a different position from that of the deleted 781KKRK784 motif (Fig 1). In this new position, the amino acids flanking the SV40 NLS are different from those flanking 781KKRK784 in WT Pif1. Thus, the immediate amino acid context of a hypothetical protein-protein interaction site would be different in WT Pif1 versus Pif1-NLSΔ+SV40. Indeed, the reason we deleted only four amino acids (781KKRK784) rather than the entire 11 amino acid (777DEQVKKRKLDY787), computer-predicted NLS, was to minimize undesired changes in the engineered protein. The most parsimonious explanation for suppression of the nuclear defects of pif1-NLSΔ cells by insertion of the SV40 NLS at a heterologous site is that the deleted 781KKRK784 is required for the nuclear localization of Pif1. Second, multiple Pif1-interaction experiments, including genetic and biochemical approaches, have been done by us and others (see, for example, [36–38]). To our knowledge, none of these experiments identified a Pif1 interacting protein that is important for all four of the assessed nuclear Pif1 functions. Specifically, it is unlikely that 781KKRK784 is required for Pif1 interaction with PCNA as Pif1 interacts with PCNA via a non-canonical PCNA-binding site that is located distal to and is totally unlike the 781KKRK784 motif [39]. In addition, inhibition of telomerase by Pif1 in vitro requires just Pif1 and telomerase [26], while telomerase has no known role in BIR or Okazaki fragment maturation. In addition, this hypothetical Pif1 binding protein would be required for Pif1 binding to its nuclear targets (Fig 7B). We consider it unlikely that 781KKRK784 region is required for interaction with a protein required for all four tested nuclear functions. Therefore, we conclude that the 781KKRK784 motif is unlikely to be a protein-protein interaction site for a protein other than an importin. We still cannot fully exclude that this motif is required for the interaction with another NLS-containing protein, but even if it is true, it would still support our overall conclusion that 781KKRK784 is required for the nuclear import of Pif1. The pif1-NLSΔ allele is not a null for nuclear functions of Pif1 pif1-NLSΔ cells were defective in all tested nuclear functions. In two of these assays, telomere length regulation and long-range BIR, pif1-NLSΔ and pif1Δ cells were similarly defective (Fig 3 and 4). However, in other more sensitive assays, pif1-NLSΔ cells was not a null allele. For example, the very sensitive mutagenesis (Fig 5) and AMBER (Fig 6) assays revealed that both pif1-NLSΔ and pif1-m2 were more active than pif1Δ, consistent with earlier findings that pif1-m2 is not a null allele [8,28,30]. Specifically, the mutagenesis and AMBER assays showed that, in some cells, BIR proceeded for up to 22kb in both pif1-NLSΔ and pif1-m2 cells. In addition, using the AMBER assay, BIR was detected as far as 62 kb from the DSB in pif1-NLSΔ cells, but not in pif1-m2 cells (Fig 6). Even the double mutant pif1-m2+NLSΔ had residual BIR activity in the mutagenesis (but not the AMBER) assay. It is not surprising that the mutagenesis and AMBER assays did not show identical dependencies: the mutagenesis assay detects all completed DSB repair events with BIR synthesis within 16 kb of the DSB, even those that occur with a long delay, as long as they generate viable cells [28]. In contrast, the AMBER assay detects BIR synthesis only during the first 8–10 hours after the DSB [31,32]. On the other hand, the AMBER assay is a direct assay for BIR synthesis and detects not only finished, but also incomplete events [32] while the mutagenesis assay is indirect. As with the mutagenesis and AMBER assays, pif1-NLSΔ cells were not as defective as pif1Δ for suppression of dna2Δ lethality (Fig 8). Lethality of dna2Δ is thought to result from the persistence of long flaps produced by Pif1 during Okazaki fragment maturation. We propose that even a few long flaps resulting from the nuclear localization of small amounts of Pif1-NLSΔ protein is sufficient to kill cells lacking Dna2. Nevertheless, formation of some viable pif1-NLSΔ dna2Δ spores suggests that at least some of these cells had no nuclear Pif1-NLSΔ protein. Nuclear entry of Pif1 in the absence of the 781KKRK784 motif Given that residual levels of both BIR and Okazaki fragment maturation occurred in pif1-NLSΔ cells (Figs 5,6, and 8), nuclear entry of some Pif1 must occur even in the absence of the 781KKRK784 motif. Likewise, ChIP assays detected Pif1-NLSΔ protein bound to two of two known nuclear target sites, although at both sites, Pif1 bindings was ~5-fold lower than in WT cells (Fig 7). In contrast, Pif1-NLSΔ binding to a mitochondrial site was not reduced (Fig 7). We envision at least three possibilities for how Pif1 enters the nucleus in pif1-NLSΔ cells. The most likely mechanism is that a fraction of the 98kDa Pif1-NLSΔ diffuses into the nucleus, as there is no sharp molecular weight cut off for passive protein import [17]. Another possibility is that small amounts of Pif1 can enter the nucleus via its association with another protein that has its own NLS. Finally, a less active NLS (for example the bipartite NLS identified by cMapper; see Results) may function in a subset of cells. Nonetheless, our data suggest that nuclear entry of Pif1 is heavily reliant on the 781KKRK784 motif. We note that highly sensitive genetic assays are likely better able to detect low levels of nuclear entry compared to cytological or physical assays (such as telomere length). Moreover, small amounts of an enzyme are more likely than small levels of a structural protein to provide detectable function. New separation of function pif1 alleles with impaired nuclear function A pif1Δ strain is not convenient for analysis of nuclear Pif1 functions because of its slow growth. In addition, the defective mitochondrial genomes in different pif1Δ strains are not necessarily identical, and impaired mitochondrial function may result in secondary phenotypes that are not due directly to lack of Pif1. Hence, a separation-of-function form of Pif1 that is defective for nuclear, but proficient for mitochondrial functions, would be valuable. As shown here, in two sensitive BIR assays, pif1-NLSΔ cells had even more residual BIR activity than pif1-m2 (Figs 5 and 6). The fact that the doubly mutant pif1-m2+NLSΔ allele was more defective than either single mutant in sensitive BIR assays is consistent with the two mutations acting at different steps in Pif1 biogenesis. That is, the pif1-m2 allele results in reduced translation specifically of the nuclear form of Pif1 [4], while the NLSΔ allele reduces nuclear entry of Pif1. Even though there is residual nuclear protein by the most sensitive assays even in pif1-m2+NLSΔ cells, unlike pif1Δ, all three of the partial loss of function pif1 alleles had WT or near WT growth rates and maintained mitochondrial function. Moreover, even though the pif1-m2+ NLSΔ allele was not a null in the most sensitive BIR assays, it was more defective than either single mutation. Thus, a pif1-m2+ NLSΔ strain is currently the best one for studying nuclear functions of Pif1. The pif1-NLSΔ allele is not a null for nuclear functions of Pif1 pif1-NLSΔ cells were defective in all tested nuclear functions. In two of these assays, telomere length regulation and long-range BIR, pif1-NLSΔ and pif1Δ cells were similarly defective (Fig 3 and 4). However, in other more sensitive assays, pif1-NLSΔ cells was not a null allele. For example, the very sensitive mutagenesis (Fig 5) and AMBER (Fig 6) assays revealed that both pif1-NLSΔ and pif1-m2 were more active than pif1Δ, consistent with earlier findings that pif1-m2 is not a null allele [8,28,30]. Specifically, the mutagenesis and AMBER assays showed that, in some cells, BIR proceeded for up to 22kb in both pif1-NLSΔ and pif1-m2 cells. In addition, using the AMBER assay, BIR was detected as far as 62 kb from the DSB in pif1-NLSΔ cells, but not in pif1-m2 cells (Fig 6). Even the double mutant pif1-m2+NLSΔ had residual BIR activity in the mutagenesis (but not the AMBER) assay. It is not surprising that the mutagenesis and AMBER assays did not show identical dependencies: the mutagenesis assay detects all completed DSB repair events with BIR synthesis within 16 kb of the DSB, even those that occur with a long delay, as long as they generate viable cells [28]. In contrast, the AMBER assay detects BIR synthesis only during the first 8–10 hours after the DSB [31,32]. On the other hand, the AMBER assay is a direct assay for BIR synthesis and detects not only finished, but also incomplete events [32] while the mutagenesis assay is indirect. As with the mutagenesis and AMBER assays, pif1-NLSΔ cells were not as defective as pif1Δ for suppression of dna2Δ lethality (Fig 8). Lethality of dna2Δ is thought to result from the persistence of long flaps produced by Pif1 during Okazaki fragment maturation. We propose that even a few long flaps resulting from the nuclear localization of small amounts of Pif1-NLSΔ protein is sufficient to kill cells lacking Dna2. Nevertheless, formation of some viable pif1-NLSΔ dna2Δ spores suggests that at least some of these cells had no nuclear Pif1-NLSΔ protein. Nuclear entry of Pif1 in the absence of the 781KKRK784 motif Given that residual levels of both BIR and Okazaki fragment maturation occurred in pif1-NLSΔ cells (Figs 5,6, and 8), nuclear entry of some Pif1 must occur even in the absence of the 781KKRK784 motif. Likewise, ChIP assays detected Pif1-NLSΔ protein bound to two of two known nuclear target sites, although at both sites, Pif1 bindings was ~5-fold lower than in WT cells (Fig 7). In contrast, Pif1-NLSΔ binding to a mitochondrial site was not reduced (Fig 7). We envision at least three possibilities for how Pif1 enters the nucleus in pif1-NLSΔ cells. The most likely mechanism is that a fraction of the 98kDa Pif1-NLSΔ diffuses into the nucleus, as there is no sharp molecular weight cut off for passive protein import [17]. Another possibility is that small amounts of Pif1 can enter the nucleus via its association with another protein that has its own NLS. Finally, a less active NLS (for example the bipartite NLS identified by cMapper; see Results) may function in a subset of cells. Nonetheless, our data suggest that nuclear entry of Pif1 is heavily reliant on the 781KKRK784 motif. We note that highly sensitive genetic assays are likely better able to detect low levels of nuclear entry compared to cytological or physical assays (such as telomere length). Moreover, small amounts of an enzyme are more likely than small levels of a structural protein to provide detectable function. New separation of function pif1 alleles with impaired nuclear function A pif1Δ strain is not convenient for analysis of nuclear Pif1 functions because of its slow growth. In addition, the defective mitochondrial genomes in different pif1Δ strains are not necessarily identical, and impaired mitochondrial function may result in secondary phenotypes that are not due directly to lack of Pif1. Hence, a separation-of-function form of Pif1 that is defective for nuclear, but proficient for mitochondrial functions, would be valuable. As shown here, in two sensitive BIR assays, pif1-NLSΔ cells had even more residual BIR activity than pif1-m2 (Figs 5 and 6). The fact that the doubly mutant pif1-m2+NLSΔ allele was more defective than either single mutant in sensitive BIR assays is consistent with the two mutations acting at different steps in Pif1 biogenesis. That is, the pif1-m2 allele results in reduced translation specifically of the nuclear form of Pif1 [4], while the NLSΔ allele reduces nuclear entry of Pif1. Even though there is residual nuclear protein by the most sensitive assays even in pif1-m2+NLSΔ cells, unlike pif1Δ, all three of the partial loss of function pif1 alleles had WT or near WT growth rates and maintained mitochondrial function. Moreover, even though the pif1-m2+ NLSΔ allele was not a null in the most sensitive BIR assays, it was more defective than either single mutation. Thus, a pif1-m2+ NLSΔ strain is currently the best one for studying nuclear functions of Pif1. Materials and methods Yeast strains and growth conditions To assess the effects of mutant PIF1 alleles on dna2Δ lethality and on mitochondrial proficiency, we used two yeast diploid strains YCG57 or YCG59 (both derived from W303 [40]) that had the following genotypes: YCG57 (MATa/MATa leu2-3,112/leu2-3,112, trp1-1/trp1-1, can1-100/can1-100, ura3-1/ura3-1, ade2-1/ade2-1, his3-11/his3-11, PIF1/pif1::NatMX6), YCG59 (MATa/MATa leu2-3,112/leu2-3,112, trp1-1/trp1-1, can1-100/can1-100, ura3-1/ura3-1, ade2-1/ade2-1, his3-11/his3-11, PIF1/pif1::NatMX6, DNA2/dna2::KanMX6) [19] (see S1 Table for all genotypes of yeast strains). The efficiency of BIR and telomere lengths were determined using the AM1003 [27] and its derivatives. These strains are disomic for chromosome III and express the HO endonuclease under the control of a galactose-inducible promoter. The rate of BIR-associated mutagenesis was assessed using derivatives of AM1003 containing a frameshift reporter gene (lys2::A4) inserted into the donor chromosome 16kb centromere-distal to MATα-inc (AM1291 and its derivatives) [29]. This mutant allele of LYS2 gene has an insertion of 61bp, leading to a Lys- phenotype due to +1-bp frameshift. A Lys+ phenotype is usually restored by a 1bp deletion in the reporter. The construction of pif1-m2 and pif1Δ derivatives of AM1291 (AM2061 and AM2191, respectively) was described previously [28]. To insert other pif1 mutant alleles, we first replaced the hygromycin-resistance gene (HPH) located at HMR in AM1003, AM1291, and AM2061 with KANMX (G418 resistance marker) to create AM5640, AM5978, and AM6121 respectively. To introduce pif1-NLSΔ, these strains were co-transformed with an oligo OL5029 (S6 Table) and plasmid pRL10 (see S7 Table; [41]) that contains a region for targeting the NLS region that is complementary to the 50bp 5’ of the Pif1 NLS and the 50bp 3’ of the NLS region but lacks the NLS itself. Deletion of the PIF1 NLS region in AM6007 (AM1003 derivative), AM6032 (AM1291 derivative), and AM6201 (AM2061 derivative) was confirmed by PCR and sequencing. To insert the SV40-NLS into the strains containing pif1-NLSΔ, the SV40-NLS region was PCR amplified from pCG82 (see S7 Table) using the following primers: OL5032 (S6 Table) homologous to PIF1 region located centromere proximal to NLS, and OL5033 (S6 Table) where the small letters correspond to the SV40-NLS sequence of pCG82 and the capital letters represent homology to the region flanking the PIF1 gene (located centromere distal from PIF1). This PCR product was co-transformed with pRL11 (CRISPR-Cas9 plasmid that contains a region corresponding to 854 aa of Pif1 for CRISPR-Cas9 targeting; see S7 Table). The resulting strains containing the SV40-NLS are AM6057 (AM6007 derivative) and AM6060 (AM6032 derivative). For ChIP analysis, Pif1 alleles were tagged with 13xMyc epitopes that were amplified from the pFA6a-13XMyc-KanMX6 plasmid [33] by using the following primers: RL216 and RL217 (S6 Table). The resulting PCR product was co-transformed with pRL11 into AM6007, AM1003, and AM6201 to generate strains AM6737, AM6740, and AM7068, respectively (S1 Table). In addition, primers RL367 and RL217 were used to amplify SV40 NLS and 13xMyc together, and the PCR product was co-transformed with pRL11 into AM6007 to generate AM7077 (S1 Table). The constructs were confirmed by Sanger sequencing, and the level of Pif1 in these strains was analyzed by western blot. The construction of diploids used to analyze the suppression of dna2Δ lethality by pif1-NLSΔ involved the following steps: (i) haploid MATα-inc strains were produced by plating AM6007 and AM1003 on YEP-Gal media followed by selection of Ade-red Leu- cells to obtain AM6867 (pif1-NLSΔ) and AM6870 (PIF1) (see S1 Table). (ii) KANMX in AM6867 was replaced by HPH to obtain AM6872. (iii) NP265 (MATa pif1Δdna2Δ, see S1 Table) was a gift from Dr. Grzegorz Ira lab. AM6872 and AM6870 were crossed to NP265 to generate diploid strains (AM6902 and AM6900, respectively). Plasmids The empty vector used for the analyses of a mitochondrial function and of dna2Δ suppression was pRS414 plasmid (ARS CEN6 TRP1) [42]. pCG17 is a derivative of pRS414 that contains PIF1 under control of its endogenic promoter and that was constructed in the following steps: (i) the PsPX1/AgeI fragment containing PIF1 with a carboxy-terminal 3xFLAG tag under control of RRM3 promoter was cloned into pRS414 digested with PsPX1 and AgeI to produce pMB282 [9]; (ii) the RRM3 promoter in pMB282 was replaced with endogenous promoter of PIF1 (using digestion with PsPX1 and AgeI) to create pCG17 [19]. pCG80 (see S7 Table), which carries the pif1-NLSΔ allele, was generated from pCG17 by deleting the sequence corresponding to amino acids 781KKRK784 of Pif1. pCG82 was constructed from pCG80 by inserting the region of SV40 that encodes the SV40 NLS for T-antigen (amino acids 126PKKKRKV132) into the C-terminus of pif1-NLSΔ. Plasmids were introduced into yeast cells via lithium acetate transformation [43]. CRISPR-Cas9 plasmids (pRL10 and pRL11), that contain a target region for NLS of Pif1 and carboxy-terminal of Pif1, respectively, were constructed as described [41] and used in construction of pif1-NLSΔ (AM6007 and AM6032), pif1-m2+NLSΔ (AM6201), and pif1-NLSΔ+SV40 (AM6057 and AM6060) strains, respectively. pFA6a-13xMyc-KanMX6 plasmid [33] was used as a template to copy 13xMyc to tag various pif1 alleles for ChIP analysis. Media and growth conditions Depending on the experiment, cells were grown in rich medium YEPD (1% yeast extract, 2% peptone, and 2% glucose) or synthetic drop-out medium lacking tryptophan (Sc-Trp [44]) with glucose (2%) or raffinose (3%). For experiments assessing BIR efficiency or BIR-associated mutagenesis, strains were grown in synthetic drop-out medium lacking leucine with glucose (2%) (Sc-Leu [44]) in rich medium containing lactic acid (3%) instead of glucose (YEP-Lac [45]), followed by addition of galactose directly to the media (to a final concentration of 2%), or plated and on rich medium where glucose was replaced by galactose (2%) (YEP-Gal) [45]. The BIR outcomes were characterized by replica plating on synthetic drop-out media lacking leucine or adenine with glucose (2%) (Sc-Leu and Sc-Ade respectively) [45]. The BIR-associated mutagenesis was assessed by plating on synthetic drop-out medium lacking both adenine and lysine (Sc-Ade, Lys) [45]. To induce meiosis, diploid cells were sporulated by using pre-sporulation medium (YPA, similar to YEPD but potassium acetate was added to 2% instead of glucose) followed replica plating to sporulation medium (SM) that contained 2% potassium acetate [46]. Identification of the candidate Pif1 NLS motif and construction of Pif1 NLS alleles The cNLS Mapper (http://nls-mapper.iab.keio.ac.jp/) was used to predict putative nuclear localization signals (NLS) within Pif1 [23]. The core of the functional NLS was narrowed down to a stretch of four basic amino acid residues (781KKRK784). Deletion of these four amino acid motifs was generated in pCG17 using Quik Change Lightning Site-directed Mutagenesis (Agilent Technologies) to create pCG80. A heterologous NLS (126PKKKRKV132) from the SV40 T antigen [25] was introduced into the carboxy-terminal end of pCG17 and pCG80 using gBlocks gene fragments (Integrated DNA Technologies, IDT) and the Gibson Assembly® method (New England Biolabs, NEB). pif1-NLSΔ alleles were verified by both restriction enzyme digestion using AvaI and NotI (NEB) and DNA sequencing (GeneWiz). Immunoblot analysis Yeast total protein extracts were prepared using trichloroacetic acid (TCA) precipitation [47]. Briefly, 10 ml of mid-log phase cells were collected and resuspended in 20% TCA. The cells were lysed with glass beads and vortexing (three 1 min cycles, 4°C) using the FastPrep-24homogenizer (MP Biomedicals). Two different variations of the method were used after this point: Variation A (used for western blot analysis shown in Fig 1C). The glass beads were washed with 5% TCA and the precipitated proteins were collected by centrifugation. The protein was resuspended in 1X Laemmli buffer (150 mM Tris pH 6.8, 6% SDS, 30% glycerol, 0.3% bromophenol blue, 15% β-mercaptoethanol) and 1M Tris base, then boiled for 5 min. The proteins were resolved on a 7% SDS-PAGE and transferred to nitrocellulose membrane (GE Healthcare). FLAG-tagged Pif1 proteins were detected using mouse anti-FLAG M2 antibody (Sigma-Aldrich) and visualized with HRP-conjugated anti-mouse secondary antibody (Bio-Rad) using standard ECL detection reagents (GE Healthcare) and imaged using the AlphaImager HP system (Protein Simple). Variation B (used for western blot analysis shown in Fig 7A). The precipitated proteins were collected by centrifugation and washed by 0.5M Tris-HCl (pH 8.0). The protein was resuspended in 2x loading buffer (10mM Tris pH 6.8, 4% SDS, 20% glycerol, 0.2% bromphenol blue, 10% β-mercaptoethanol) and 0.5M Tris base (pH 8.0), then boiled for 5 min. The proteins were resolved on 4–20% SDS-PAGE and transferred to nitrocellulose membrane (Amersham #10600002). MYC-tagged Pif1 proteins were detected using antibody (Sigma M4439) and visualized using IRDye 800CW Goat anti-Mouse (#926–32210) from LI-COR. Blots were imaged using an LI-COR Odyssey-Fc Imager. Analyzing mitochondrial proficiency 100 ng of plasmid DNA was transformed into a heterozygous PIF1/pif1Δ::NatMX6 diploid strain YCG57 (see S1 Table). After sporulation and dissection, haploid pif1Δ spores carrying TRP1 plasmids were selected and grown to saturation in 5 ml Sc-Trp medium. As previously described [19], serial dilutions were spotted onto Sc-Trp medium with 2% glucose or 3% glycerol and incubated at 30°C for at least 3 days. Telomere length analysis Genomic DNA was purified from yeast strains as described in [45], and 1.0–1.5ug of this DNA was digested with XhoI restriction enzyme (NEB). Following XhoI digestion, DNA was separated by gel electrophoresis in 1.2% agarose gel in TBE buffer, transferred to a nylon membrane (GE Healthcare), and hybridized (as described [45]) to a 32P-labelled probe specific to Y’ pre-telomeric region [48] that was prepared as described in [45]. Blots were imaged using an Azure Sapphire Biomolecular Imager. All strains containing various alleles of PIF1 were propagated on YEPD plates for at least 100 generations to allow time for establishment of telomere lengths characteristic for each particular PIF1 allele. BIR efficiency assay Cells were grown in 5ml of liquid Sc-Leu media for approximately 24 hours at 30°C until saturation. Cells were then transferred to YEP-Lac media at ~1-3x106 cells/ml and grown for ~ 16 hours until cell density was about ~1-2x107 cells/mL. Then cells were plated on YEP-Gal and YEPD plates (at ~50 cells per plate). Plates were incubated for 5–7 days, and then replica plated onto Sc-Ade or Sc-Leu. The frequencies of BIR, gene conversion (GC), half-crossover (HC), and chromosome loss (CL) events were calculated based on respective fractions of Ade+ Leu–, Ade+ Leu+, Ade-white Leu-, Ade-red Leu- outcomes among all DSB repair events (as described in [45]). Since Ade+ Leu−can be generated by both BIR and GCRs, the fraction of GCRs among Ade+ Leu−was estimated by analyzing the representative number of these repair outcomes using CHEF gel electrophoresis (similar to [45]). Due to random segregation of chromosomes, half of the HC products segregate with an intact copy of full-length chromosome III leading to formation of repair products (HC-II) that are genetically indistinguishable from BIR [27]. Therefore, it was assumed that the number of HC-II events is equal to the number of HC-I (Ade-white Leu- events). Consequently, the number of BIR was adjusted by subtracting the number of HC-II. BIR mutagenesis assay Cells were grown at 30°C in 5ml of Sc-Leu for ~24 hours, diluted 20-fold with YEP-Lac, and then grown for ~16 hours until cell density reached ~1-2x107cells/mL. BIR was induced by addition of 20% galactose to a final concentration of 2%, and cells were incubated for 7 hours after galactose addition. The rate of mutagenesis that occurs during BIR-associated DNA replication was determined by plating cells on Sc-Ade, Lys and Sc-Ade (see [45] for the details of rate calculation). The resulting mutagenesis rate was compared to the rate of mutagenesis that occurs during a normal S phase, calculated by plating cells to the same media prior to galactose addition (see [45] for the details of calculation). AMBER (assay for monitoring BIR elongation rate) to detect BIR-associated DNA synthesis Cells were grown overnight at 30°C in 5ml of Sc-Leu, transferred (1:100 dilution) to YEP-Raffinose and incubated for 16hrs until cell density of ~5x106 cells/ml. Then, 50ml of cells were collected (0hr) and stored at -80°C. Galactose was added to the culture to a final concentration of 2% to induce HO endonuclease. During the following 10hrs of incubation at 30°C, ~30ml of cells were collected every hour and stored at -80°C. Genomic DNA was purified and quantified as described for AMBER assay in [31]. For droplet digital PCR (ddPCR) the reaction mix was assembled as follows: 6ul H2O, 1ul of each primer set (one set of primers for ACT1, and another–for target position), 10ul of ddPCR supermix (Biorad, #1863026) and 2ul DNA (0.1ng/uL) were added and mixed thoroughly by vortexting. ddPCR was conducted and analyzed as described in [31]. The results were presented as Boltzmann sigmoidal curve. Mean values of target to reference (ACT1) loci ratios were calculated by Poisson distribution based on 20,000 droplets with error bars representing upper and lower Poisson 95% CI. Assaying the effects of PIF1 mutants on the viability of dna2Δ cells A qualitative assay for characterization of dna2Δ lethality suppression by different alleles of pif1 has been previously described [19]. In particular 500 ng of plasmid DNA (plasmids included different alleles of PIF1) was transformed into a heterozygous PIF1/pif1Δ::NatMX6 DNA2/dna2Δ::KANMX6 diploid YCG59 (see S1 Table). After sporulation and dissection, haploid pif1Δdna2Δ spores carrying TRP1 plasmids were selected and grown in 5ml Sc-TRP medium. The cells were plated onto non-selective rich medium YEPD and onto Sc-TRP and grown at 30°C for 3–4 days. Growth on selective Sc-TRP medium indicates lack of Pif1 nuclear function, as PIF1 deficiency suppresses the lethality of dna2Δ cells [6]. For quantitative assay of lethality suppression, diploid cells that were DNA2/dna2::URA3 and also either PIF1/pif1::KANMX or Pif1-NLSΔ/pif1::KANMX were sporulated and the resulting spores were analyzed by either tetrad analysis (similar to [46]) or by random spore analysis (RSA). For RSA, diploid cells were grown in 5ml of liquid YEPD at 30°C overnight and transferred to 50ml of liquid YPA media (1:10 dilution ratio) and incubated for ~24hr at 30°C. Cells were collected and transferred into 50ml of liquid SM media and incubated for at least 72hrs at 30°C. Cells were spun down and resuspended in 5ml of sterile water. An equal volume of diethyl ether was added to samples, and the mixture was thoroughly mixed for at least 10 minutes at room temperature. Cells were centrifuged and washed by 30ml of sterile water four times, serially diluted, plated on YEPD plates, and incubated for 7days at 30°C. The phenotypes of meiotic outcomes were classified by replica plating onto Sc-Ade, Sc-Ura, and YEPD+G418 (0.5g/L). Chromatin Immunoprecipitation (ChIP) analysis To perform ChIP assay, 13 repeats of the Myc epitope were inserted at the C-terminus of both PIF1 and pif1-NLSΔ. The levels of Pif1 protein expressed in these two strains was determined by western blot analysis. ChIP analyses was performed as described previously [12] with several modifications. Specifically, yeast cells were grown to a density between 5.0x106 and 1x107 cells/ml in YEP-raffinose and HO-generated DSBs were induced by addition of 20% galactose (to a final concentration of 2%). 40ml of cells were collected before (0h) and after (4h) DSB induction and proteins were crosslinked by the addition of formaldehyde (Sigma #252549) to a final concentration of 1%. Cells were incubated at room temperature for 10 min followed by addition of glycine (125mM final concentration). Cells were incubated for 5 min, and samples were then stored at -80°C. Cells were lysed with glass beads, and DNA was sheard by sonication to an average size of 0.5kb. Extracts were divided into IP and input samples (9:1 ratio). IP samples were incubated with anti-Myc antibody (Sigma M4439) overnight at 4°C. Proteins were bound to Dynabeads Protein G (Invitrogen 1004D) and taken through a series of washes and then reverse crosslinked and eluted. Samples were treated with proteinase K followed by phenol DNA extraction and DNA precipitation. Protein enrichment was measured by quantitative-PCR (qPCR) analysis from the input (1:10 dilutions) and IP (undiluted) samples. The following primers were used for qPCR analysis (see S6 Table for details): RL220, RL221 (for MAT locus), RL251, RL252 (for CEN13), RL304, RL305 (for mtDNA). Both input and IP signals were normalized by using the following primers: RL218, RL219 (ACT1 locus) used for normalizing data for MAT, and RL196, RL197 (YBL028C locus) used for normalizing data for CEN13 and mtDNA. Yeast strains and growth conditions To assess the effects of mutant PIF1 alleles on dna2Δ lethality and on mitochondrial proficiency, we used two yeast diploid strains YCG57 or YCG59 (both derived from W303 [40]) that had the following genotypes: YCG57 (MATa/MATa leu2-3,112/leu2-3,112, trp1-1/trp1-1, can1-100/can1-100, ura3-1/ura3-1, ade2-1/ade2-1, his3-11/his3-11, PIF1/pif1::NatMX6), YCG59 (MATa/MATa leu2-3,112/leu2-3,112, trp1-1/trp1-1, can1-100/can1-100, ura3-1/ura3-1, ade2-1/ade2-1, his3-11/his3-11, PIF1/pif1::NatMX6, DNA2/dna2::KanMX6) [19] (see S1 Table for all genotypes of yeast strains). The efficiency of BIR and telomere lengths were determined using the AM1003 [27] and its derivatives. These strains are disomic for chromosome III and express the HO endonuclease under the control of a galactose-inducible promoter. The rate of BIR-associated mutagenesis was assessed using derivatives of AM1003 containing a frameshift reporter gene (lys2::A4) inserted into the donor chromosome 16kb centromere-distal to MATα-inc (AM1291 and its derivatives) [29]. This mutant allele of LYS2 gene has an insertion of 61bp, leading to a Lys- phenotype due to +1-bp frameshift. A Lys+ phenotype is usually restored by a 1bp deletion in the reporter. The construction of pif1-m2 and pif1Δ derivatives of AM1291 (AM2061 and AM2191, respectively) was described previously [28]. To insert other pif1 mutant alleles, we first replaced the hygromycin-resistance gene (HPH) located at HMR in AM1003, AM1291, and AM2061 with KANMX (G418 resistance marker) to create AM5640, AM5978, and AM6121 respectively. To introduce pif1-NLSΔ, these strains were co-transformed with an oligo OL5029 (S6 Table) and plasmid pRL10 (see S7 Table; [41]) that contains a region for targeting the NLS region that is complementary to the 50bp 5’ of the Pif1 NLS and the 50bp 3’ of the NLS region but lacks the NLS itself. Deletion of the PIF1 NLS region in AM6007 (AM1003 derivative), AM6032 (AM1291 derivative), and AM6201 (AM2061 derivative) was confirmed by PCR and sequencing. To insert the SV40-NLS into the strains containing pif1-NLSΔ, the SV40-NLS region was PCR amplified from pCG82 (see S7 Table) using the following primers: OL5032 (S6 Table) homologous to PIF1 region located centromere proximal to NLS, and OL5033 (S6 Table) where the small letters correspond to the SV40-NLS sequence of pCG82 and the capital letters represent homology to the region flanking the PIF1 gene (located centromere distal from PIF1). This PCR product was co-transformed with pRL11 (CRISPR-Cas9 plasmid that contains a region corresponding to 854 aa of Pif1 for CRISPR-Cas9 targeting; see S7 Table). The resulting strains containing the SV40-NLS are AM6057 (AM6007 derivative) and AM6060 (AM6032 derivative). For ChIP analysis, Pif1 alleles were tagged with 13xMyc epitopes that were amplified from the pFA6a-13XMyc-KanMX6 plasmid [33] by using the following primers: RL216 and RL217 (S6 Table). The resulting PCR product was co-transformed with pRL11 into AM6007, AM1003, and AM6201 to generate strains AM6737, AM6740, and AM7068, respectively (S1 Table). In addition, primers RL367 and RL217 were used to amplify SV40 NLS and 13xMyc together, and the PCR product was co-transformed with pRL11 into AM6007 to generate AM7077 (S1 Table). The constructs were confirmed by Sanger sequencing, and the level of Pif1 in these strains was analyzed by western blot. The construction of diploids used to analyze the suppression of dna2Δ lethality by pif1-NLSΔ involved the following steps: (i) haploid MATα-inc strains were produced by plating AM6007 and AM1003 on YEP-Gal media followed by selection of Ade-red Leu- cells to obtain AM6867 (pif1-NLSΔ) and AM6870 (PIF1) (see S1 Table). (ii) KANMX in AM6867 was replaced by HPH to obtain AM6872. (iii) NP265 (MATa pif1Δdna2Δ, see S1 Table) was a gift from Dr. Grzegorz Ira lab. AM6872 and AM6870 were crossed to NP265 to generate diploid strains (AM6902 and AM6900, respectively). Plasmids The empty vector used for the analyses of a mitochondrial function and of dna2Δ suppression was pRS414 plasmid (ARS CEN6 TRP1) [42]. pCG17 is a derivative of pRS414 that contains PIF1 under control of its endogenic promoter and that was constructed in the following steps: (i) the PsPX1/AgeI fragment containing PIF1 with a carboxy-terminal 3xFLAG tag under control of RRM3 promoter was cloned into pRS414 digested with PsPX1 and AgeI to produce pMB282 [9]; (ii) the RRM3 promoter in pMB282 was replaced with endogenous promoter of PIF1 (using digestion with PsPX1 and AgeI) to create pCG17 [19]. pCG80 (see S7 Table), which carries the pif1-NLSΔ allele, was generated from pCG17 by deleting the sequence corresponding to amino acids 781KKRK784 of Pif1. pCG82 was constructed from pCG80 by inserting the region of SV40 that encodes the SV40 NLS for T-antigen (amino acids 126PKKKRKV132) into the C-terminus of pif1-NLSΔ. Plasmids were introduced into yeast cells via lithium acetate transformation [43]. CRISPR-Cas9 plasmids (pRL10 and pRL11), that contain a target region for NLS of Pif1 and carboxy-terminal of Pif1, respectively, were constructed as described [41] and used in construction of pif1-NLSΔ (AM6007 and AM6032), pif1-m2+NLSΔ (AM6201), and pif1-NLSΔ+SV40 (AM6057 and AM6060) strains, respectively. pFA6a-13xMyc-KanMX6 plasmid [33] was used as a template to copy 13xMyc to tag various pif1 alleles for ChIP analysis. Media and growth conditions Depending on the experiment, cells were grown in rich medium YEPD (1% yeast extract, 2% peptone, and 2% glucose) or synthetic drop-out medium lacking tryptophan (Sc-Trp [44]) with glucose (2%) or raffinose (3%). For experiments assessing BIR efficiency or BIR-associated mutagenesis, strains were grown in synthetic drop-out medium lacking leucine with glucose (2%) (Sc-Leu [44]) in rich medium containing lactic acid (3%) instead of glucose (YEP-Lac [45]), followed by addition of galactose directly to the media (to a final concentration of 2%), or plated and on rich medium where glucose was replaced by galactose (2%) (YEP-Gal) [45]. The BIR outcomes were characterized by replica plating on synthetic drop-out media lacking leucine or adenine with glucose (2%) (Sc-Leu and Sc-Ade respectively) [45]. The BIR-associated mutagenesis was assessed by plating on synthetic drop-out medium lacking both adenine and lysine (Sc-Ade, Lys) [45]. To induce meiosis, diploid cells were sporulated by using pre-sporulation medium (YPA, similar to YEPD but potassium acetate was added to 2% instead of glucose) followed replica plating to sporulation medium (SM) that contained 2% potassium acetate [46]. Identification of the candidate Pif1 NLS motif and construction of Pif1 NLS alleles The cNLS Mapper (http://nls-mapper.iab.keio.ac.jp/) was used to predict putative nuclear localization signals (NLS) within Pif1 [23]. The core of the functional NLS was narrowed down to a stretch of four basic amino acid residues (781KKRK784). Deletion of these four amino acid motifs was generated in pCG17 using Quik Change Lightning Site-directed Mutagenesis (Agilent Technologies) to create pCG80. A heterologous NLS (126PKKKRKV132) from the SV40 T antigen [25] was introduced into the carboxy-terminal end of pCG17 and pCG80 using gBlocks gene fragments (Integrated DNA Technologies, IDT) and the Gibson Assembly® method (New England Biolabs, NEB). pif1-NLSΔ alleles were verified by both restriction enzyme digestion using AvaI and NotI (NEB) and DNA sequencing (GeneWiz). Immunoblot analysis Yeast total protein extracts were prepared using trichloroacetic acid (TCA) precipitation [47]. Briefly, 10 ml of mid-log phase cells were collected and resuspended in 20% TCA. The cells were lysed with glass beads and vortexing (three 1 min cycles, 4°C) using the FastPrep-24homogenizer (MP Biomedicals). Two different variations of the method were used after this point: Variation A (used for western blot analysis shown in Fig 1C). The glass beads were washed with 5% TCA and the precipitated proteins were collected by centrifugation. The protein was resuspended in 1X Laemmli buffer (150 mM Tris pH 6.8, 6% SDS, 30% glycerol, 0.3% bromophenol blue, 15% β-mercaptoethanol) and 1M Tris base, then boiled for 5 min. The proteins were resolved on a 7% SDS-PAGE and transferred to nitrocellulose membrane (GE Healthcare). FLAG-tagged Pif1 proteins were detected using mouse anti-FLAG M2 antibody (Sigma-Aldrich) and visualized with HRP-conjugated anti-mouse secondary antibody (Bio-Rad) using standard ECL detection reagents (GE Healthcare) and imaged using the AlphaImager HP system (Protein Simple). Variation B (used for western blot analysis shown in Fig 7A). The precipitated proteins were collected by centrifugation and washed by 0.5M Tris-HCl (pH 8.0). The protein was resuspended in 2x loading buffer (10mM Tris pH 6.8, 4% SDS, 20% glycerol, 0.2% bromphenol blue, 10% β-mercaptoethanol) and 0.5M Tris base (pH 8.0), then boiled for 5 min. The proteins were resolved on 4–20% SDS-PAGE and transferred to nitrocellulose membrane (Amersham #10600002). MYC-tagged Pif1 proteins were detected using antibody (Sigma M4439) and visualized using IRDye 800CW Goat anti-Mouse (#926–32210) from LI-COR. Blots were imaged using an LI-COR Odyssey-Fc Imager. Analyzing mitochondrial proficiency 100 ng of plasmid DNA was transformed into a heterozygous PIF1/pif1Δ::NatMX6 diploid strain YCG57 (see S1 Table). After sporulation and dissection, haploid pif1Δ spores carrying TRP1 plasmids were selected and grown to saturation in 5 ml Sc-Trp medium. As previously described [19], serial dilutions were spotted onto Sc-Trp medium with 2% glucose or 3% glycerol and incubated at 30°C for at least 3 days. Telomere length analysis Genomic DNA was purified from yeast strains as described in [45], and 1.0–1.5ug of this DNA was digested with XhoI restriction enzyme (NEB). Following XhoI digestion, DNA was separated by gel electrophoresis in 1.2% agarose gel in TBE buffer, transferred to a nylon membrane (GE Healthcare), and hybridized (as described [45]) to a 32P-labelled probe specific to Y’ pre-telomeric region [48] that was prepared as described in [45]. Blots were imaged using an Azure Sapphire Biomolecular Imager. All strains containing various alleles of PIF1 were propagated on YEPD plates for at least 100 generations to allow time for establishment of telomere lengths characteristic for each particular PIF1 allele. BIR efficiency assay Cells were grown in 5ml of liquid Sc-Leu media for approximately 24 hours at 30°C until saturation. Cells were then transferred to YEP-Lac media at ~1-3x106 cells/ml and grown for ~ 16 hours until cell density was about ~1-2x107 cells/mL. Then cells were plated on YEP-Gal and YEPD plates (at ~50 cells per plate). Plates were incubated for 5–7 days, and then replica plated onto Sc-Ade or Sc-Leu. The frequencies of BIR, gene conversion (GC), half-crossover (HC), and chromosome loss (CL) events were calculated based on respective fractions of Ade+ Leu–, Ade+ Leu+, Ade-white Leu-, Ade-red Leu- outcomes among all DSB repair events (as described in [45]). Since Ade+ Leu−can be generated by both BIR and GCRs, the fraction of GCRs among Ade+ Leu−was estimated by analyzing the representative number of these repair outcomes using CHEF gel electrophoresis (similar to [45]). Due to random segregation of chromosomes, half of the HC products segregate with an intact copy of full-length chromosome III leading to formation of repair products (HC-II) that are genetically indistinguishable from BIR [27]. Therefore, it was assumed that the number of HC-II events is equal to the number of HC-I (Ade-white Leu- events). Consequently, the number of BIR was adjusted by subtracting the number of HC-II. BIR mutagenesis assay Cells were grown at 30°C in 5ml of Sc-Leu for ~24 hours, diluted 20-fold with YEP-Lac, and then grown for ~16 hours until cell density reached ~1-2x107cells/mL. BIR was induced by addition of 20% galactose to a final concentration of 2%, and cells were incubated for 7 hours after galactose addition. The rate of mutagenesis that occurs during BIR-associated DNA replication was determined by plating cells on Sc-Ade, Lys and Sc-Ade (see [45] for the details of rate calculation). The resulting mutagenesis rate was compared to the rate of mutagenesis that occurs during a normal S phase, calculated by plating cells to the same media prior to galactose addition (see [45] for the details of calculation). AMBER (assay for monitoring BIR elongation rate) to detect BIR-associated DNA synthesis Cells were grown overnight at 30°C in 5ml of Sc-Leu, transferred (1:100 dilution) to YEP-Raffinose and incubated for 16hrs until cell density of ~5x106 cells/ml. Then, 50ml of cells were collected (0hr) and stored at -80°C. Galactose was added to the culture to a final concentration of 2% to induce HO endonuclease. During the following 10hrs of incubation at 30°C, ~30ml of cells were collected every hour and stored at -80°C. Genomic DNA was purified and quantified as described for AMBER assay in [31]. For droplet digital PCR (ddPCR) the reaction mix was assembled as follows: 6ul H2O, 1ul of each primer set (one set of primers for ACT1, and another–for target position), 10ul of ddPCR supermix (Biorad, #1863026) and 2ul DNA (0.1ng/uL) were added and mixed thoroughly by vortexting. ddPCR was conducted and analyzed as described in [31]. The results were presented as Boltzmann sigmoidal curve. Mean values of target to reference (ACT1) loci ratios were calculated by Poisson distribution based on 20,000 droplets with error bars representing upper and lower Poisson 95% CI. Assaying the effects of PIF1 mutants on the viability of dna2Δ cells A qualitative assay for characterization of dna2Δ lethality suppression by different alleles of pif1 has been previously described [19]. In particular 500 ng of plasmid DNA (plasmids included different alleles of PIF1) was transformed into a heterozygous PIF1/pif1Δ::NatMX6 DNA2/dna2Δ::KANMX6 diploid YCG59 (see S1 Table). After sporulation and dissection, haploid pif1Δdna2Δ spores carrying TRP1 plasmids were selected and grown in 5ml Sc-TRP medium. The cells were plated onto non-selective rich medium YEPD and onto Sc-TRP and grown at 30°C for 3–4 days. Growth on selective Sc-TRP medium indicates lack of Pif1 nuclear function, as PIF1 deficiency suppresses the lethality of dna2Δ cells [6]. For quantitative assay of lethality suppression, diploid cells that were DNA2/dna2::URA3 and also either PIF1/pif1::KANMX or Pif1-NLSΔ/pif1::KANMX were sporulated and the resulting spores were analyzed by either tetrad analysis (similar to [46]) or by random spore analysis (RSA). For RSA, diploid cells were grown in 5ml of liquid YEPD at 30°C overnight and transferred to 50ml of liquid YPA media (1:10 dilution ratio) and incubated for ~24hr at 30°C. Cells were collected and transferred into 50ml of liquid SM media and incubated for at least 72hrs at 30°C. Cells were spun down and resuspended in 5ml of sterile water. An equal volume of diethyl ether was added to samples, and the mixture was thoroughly mixed for at least 10 minutes at room temperature. Cells were centrifuged and washed by 30ml of sterile water four times, serially diluted, plated on YEPD plates, and incubated for 7days at 30°C. The phenotypes of meiotic outcomes were classified by replica plating onto Sc-Ade, Sc-Ura, and YEPD+G418 (0.5g/L). Chromatin Immunoprecipitation (ChIP) analysis To perform ChIP assay, 13 repeats of the Myc epitope were inserted at the C-terminus of both PIF1 and pif1-NLSΔ. The levels of Pif1 protein expressed in these two strains was determined by western blot analysis. ChIP analyses was performed as described previously [12] with several modifications. Specifically, yeast cells were grown to a density between 5.0x106 and 1x107 cells/ml in YEP-raffinose and HO-generated DSBs were induced by addition of 20% galactose (to a final concentration of 2%). 40ml of cells were collected before (0h) and after (4h) DSB induction and proteins were crosslinked by the addition of formaldehyde (Sigma #252549) to a final concentration of 1%. Cells were incubated at room temperature for 10 min followed by addition of glycine (125mM final concentration). Cells were incubated for 5 min, and samples were then stored at -80°C. Cells were lysed with glass beads, and DNA was sheard by sonication to an average size of 0.5kb. Extracts were divided into IP and input samples (9:1 ratio). IP samples were incubated with anti-Myc antibody (Sigma M4439) overnight at 4°C. Proteins were bound to Dynabeads Protein G (Invitrogen 1004D) and taken through a series of washes and then reverse crosslinked and eluted. Samples were treated with proteinase K followed by phenol DNA extraction and DNA precipitation. Protein enrichment was measured by quantitative-PCR (qPCR) analysis from the input (1:10 dilutions) and IP (undiluted) samples. The following primers were used for qPCR analysis (see S6 Table for details): RL220, RL221 (for MAT locus), RL251, RL252 (for CEN13), RL304, RL305 (for mtDNA). Both input and IP signals were normalized by using the following primers: RL218, RL219 (ACT1 locus) used for normalizing data for MAT, and RL196, RL197 (YBL028C locus) used for normalizing data for CEN13 and mtDNA. Supporting information S1 Table. Yeast Strains used in this study. https://doi.org/10.1371/journal.pgen.1010853.s001 (XLSX) S2 Table. The effect of pif1 mutations on the distribution of DSB repair outcomes. (Related to Fig 4) https://doi.org/10.1371/journal.pgen.1010853.s002 (XLSX) S3 Table. The rates of BIR-associated mutagenesis in various pif1-deficient mutants. (Related to Fig 5B). https://doi.org/10.1371/journal.pgen.1010853.s003 (XLSX) S4 Table. The calculation of BIR-associated mutagenesis in various pif1-dificient mutants. (Related to Fig 5B). https://doi.org/10.1371/journal.pgen.1010853.s004 (XLSX) S5 Table. AMBER analysis of all 3 independent biological repeats. (Related to Fig 6B) https://doi.org/10.1371/journal.pgen.1010853.s005 (XLSX) S6 Table. List of oligonucleotides used in this study. https://doi.org/10.1371/journal.pgen.1010853.s006 (XLSX) S7 Table. Plasmids used in this study. https://doi.org/10.1371/journal.pgen.1010853.s007 (XLSX) Acknowledgments We are thankful to Dr. Grzegorz Ira for the gift of dna2Δpif1Δ yeast strain and for help with ChIP.

Moehlman, Andrew T.;Kanfer, Gil;Youle, Richard J.

doi: 10.1371/journal.pgen.1010828pmid: 37440574

Introduction Mutations in PINK1 and Parkin lead to early onset Parkinson’s disease (PD). PINK1 is a kinase imported to mitochondria and degraded, unless shunted to the outer mitochondrial membrane when mitochondrial membrane potential is impaired [1–3]. Once stabilized on the outer mitochondrial membrane (OMM) PINK1 phosphorylates ubiquitin and the Parkin ubiquitin-like domain to recruit the E3 ligase Parkin to the mitochondria, which amplifies OMM protein ubiquitination [4–7]. This ubiquitination promotes recruitment of autophagy receptors and autophagy of damaged mitochondria [8–10]. Although the molecular mechanisms of PINK1 and Parkin are well-studied, how their absence leads to Parkinsonian phenotypes is less clear [11]. Mutations in PINK1 and Parkin do not lead to substantial or PD-related phenotypes in otherwise healthy mice [12,13]. However, Drosophila melanogaster mutants lacking either pink1 or parkin (park) have severe phenotypes [14–18]. Mutants in either park or pink1 lose flight muscle, undergo degeneration of certain dopaminergic neurons, and display locomotion and flight impairment. Notably, mitochondria in the indirect flight muscles are swollen and the elongated morphology is disrupted [14,18]. Depletion of mitochondria fusion genes or expression of genes regulating fission can rescue park-/- phenotypes, supporting a role for mitochondrial dynamics in the pathophysiology of these mutant phenotypes [19–22]. Unhealthy mitochondria activate innate immune pathways through release of damage associated molecular patterns (DAMPs) such as mitochondrial DNA (mtDNA) [23–25]. Human PD patients lacking PINK1 or Parkin exhibit increased inflammation and increased serum mtDNA [26], consistent with work showing that flies lacking Parkin express higher levels of genes implicated in oxidative stress and immune responses [27]. Parkin-dependent mitophagy has been proposed to limit mitochondrial DAMP release and subsequent activation of the cGAS/STING innate immune pathway, which was previously examined in PRKN-/- mice models [28]. However, this study utilized stress paradigms, as unstressed mice do not exhibit PD-like phenotypes, unlike flies. Thus, we explored the potential role of STING in parkin-null flies. During our study, a report indicated that loss of STING did not rescue Drosophila parkin mutant defects [29], which contrasted with our contemporaneous preliminary data. Herein we compared different strains of parkin mutant alleles and conclude that loss of STING activity suppresses the thorax muscle involution and the climbing defects of parkin-/- mutants. Surprisingly, loss of STING also improves the disrupted mitochondrial morphology in parkin-/- flight muscles, suggesting unexpected feedback of Drosophila STING on mitochondrial homeostasis. Results STING is necessary for muscle degeneration and climbing defects in parkin flies Thorax indention and bent wing phenotypes in parkin mutant flies are indicative of underlying indirect flight muscle (IFM) defects and attributed to mitochondria dysfunction inducing muscle apoptosis [14,19,30]. We generated flies harboring null alleles for parkin (park25) [14] and sting (stingΔRG5) [31]. Analysis of these double knockout (DKO) flies demonstrated that loss of sting rescued both the thorax and wing phenotypes of the parkin mutant flies (Fig 1A–1C). We obtained the independently derived park1 mutant and backcrossed this allele into the stingΔRG5 mutant background [15]. These flies also demonstrated reduced penetrance of the parkin phenotypes (Fig 1A–1C). For both backgrounds, the status of the sting and park-null alleles were scored based on the presence or absence of the balancer chromosomes and fly genotypes were routinely confirmed using PCR (S1 Fig). Both park25 and park1 homozygous flies demonstrate climbing defects, due to muscle degeneration and, later, age-dependent loss of dopaminergic neurons [14,15,32]. Using the negative geotaxis assay (Fig 1D), flies homozygous for stingΔRG5 were assayed for climbing ability in parkin wild-type, park25, and park1 backgrounds. Loss of sting alone had no effect on climbing ability in young (5–7 days-old) flies. For both parkin alleles, loss of sting suppressed the climbing defects of young parkin null adults (Fig 1D and S1 Video). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 1. STING mediates flight muscle degeneration in parkin-/- flies. (A) Representative images of the thoracic muscle indentation. The stingΔRG5 allele crossed to either null parkin allele (yellow arrows) rescues the thoracic defects of park25 and park1 mutants (red arrow). All flies were generated in or crossed to a wild type w1118 stock (B & C) Quantification of the thoracic indentations (B) or the downward bent wing posture (C) in the indicated genotypes. In all graphs, bars represent the percentage of flies displaying the indicated phenotype, numbers within or juxtaposed to the bars indicate the number of flies scored per genotype (n), and the error bars represent the 95% confidence interval for the population proportion. (D) Scatter plots of quantifications for negative geotaxis assays in the indicated genotypes. Each data point represents the mean of at least 3 technical replicate assays with a group of 15 to 20 flies. Horizontal bars indicate the mean of 5 independent biological replicas per genotype. Error bars display the standard deviation. Genotypes were tested for statistical significance with an 1-way ANOVA test with post-hoc multiple comparison testing with Bonferroni’s correction. (E) Example images for pink15, pink1B9, pink15; stingΔRG5 and pink1B9; stingΔRG5 male flies. Note that the loss of sting slightly affects the pink1-null phenotypes, in contrast to the strong level of suppression seen in parkin mutant combinations. (F & G) Quantification of the thorax indention and wing posture defect phenotypes in pink1-/y, or pink1-/y; stingΔRG5 flies. In all graphs, bars represent the percentage of flies displaying the indicated phenotype, numbers indicate the number of flies scored per genotype, and the error bars represent the 95% confidence interval for the population proportion. Significance was determined using Fisher’s Exact Test for differences between population proportions. Significant p-values are indicated on the graphs. https://doi.org/10.1371/journal.pgen.1010828.g001 We confirmed the veracity of the stingΔRG5 knockout allele by crossing stingΔRG5 flies with flies containing a sting deficiency chromosome in the park25 mutant background (S1E Fig). Resulting progeny harboring one copy of stingΔRG5 allele and the sting deletion displayed suppressed thorax and wing phenotypes in the homozygous park25 mutation (S1F–S1H Fig). These results support a necessary role for STING in progression of muscle degeneration of parkin-/- flies. We also observed that loss of sting in the pink15 or pink1B9 hemizygous mutant background rescued the severity of the thorax phenotypes only partially and to a lesser extent than in parkin-null flies (Fig 1E–1G). Pink1 has been reported to have multiple Parkin-independent interactions [21,33–35], and loss of STING may not affect these pathways, resulting in only minor suppression of the pink1 thoracic muscle phenotypes. To investigate our results on parkin that diverge from Lee et al. [29], we acquired the stingΔRG5; park25 line used in that study. We verified these animals with RT-qPCR and scored thorax and wing phenotypes in the homozygous stingΔRG5; park25 flies (S2 Fig). This line of stingΔRG5; park25 flies displayed minimal thorax indention phenotype but retained the park25 bent wing phenotype (S2C Fig). One explanation for the divergent results compared to Lee et al. [29] could be differences in the genetic background. Therefore the gifted stingΔRG5; park25 stock underwent eight generations of outcrossing to the w1118 stock followed by single-male fly crosses to a double balancer stock as described in detail in the Materials and Methods section. Resulting fly lines that retained the stingΔRG5 allele and the park25 allele, as tested with PCR, were self-crossed to test the resulting homozygous progeny. In the outcrossed stocks, loss of sting suppressed both the thorax indentation and bent wings of the parkin flies, compared to stingΔRG5/+(heterozygous); park25(homozygous) siblings (S2B and S2C Fig). Therefore, it appears that a yet-unknown background difference could contribute to the severity of the park25 phenotypes in stingΔRG5 mutant flies. STING influences the underlying mitochondria pathology in parkin flies Defects in parkin-null flies include disrupted mitochondrial morphology in indirect flight muscles (IFM) [14–17]. This has been linked to dysfunctional mitochondrial dynamics attributed both to blocking of mitochondrial autophagy [36,37] and to disruption of mitochondria fusion and fission dynamics [20,38]. To assess the mitochondrial health in the parkin-/- and sting-/- flies, we examined the IFM-associated mitochondria in thoraces of young flies (3–5 days post-eclosion) using Alexa Fluor488-labeled streptavidin to visualize mitochondria (Fig 2) [39,40]. As previously reported, park25 and park1 mutants possess disrupted morphology, with interrupted mitochondrial networks and the appearance of large swollen mitochondria aggregates (Fig 2C and 2E). These defects were substantially suppressed when the stingΔRG5 allele was crossed to either of the park-null alleles (Fig 2D and 2F), whereas loss of sting alone had no mitochondrial disruption compared to controls (Fig 2A and 2B). Blinded scoring of the IFM mitochondria integrity in randomized examples of ten thoraces per genotype reveals that although the mitochondria aggregation is partially suppressed, loss of sting does not completely restore mitochondria health (Fig 2G). These results suggest a role for STING function upstream or in parallel to the mitochondrial damage phenotypes in parkin-/- flies. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 2. Mitochondrial defects in parkin-/- flight muscles are suppressed by mutation of sting. Representative micrographs from indirect flight muscle tissue in w1118 (A) and stingΔRG5 (B) thoraces, and in flies homozygous for either of the parkin-null alleles (C and E). Loss of sting mitigates the swollen mitochondria defects in park25 and park1 muscles (D and F). Staining of mitochondria was performed with AlexaFluor488-conjugated streptavidin and actin bundles were visualized with iFluor647-conjugated phalloidin. Each image is a single 1μm confocal slice. All scale bars represent 10μM. Images were linearly adjusted for brightness and contrast to avoid obscuring morphology (A’–F’) 2X digital zoom of the corresponding mitochondria image, indicated with white dotted box. (G) Quantification of mitochondria morphology with blinded analysis from 10 examples per genotype, presented randomly. Data displays the percentage of thoraces in each category for each indicated genotype. (H) Summary of Fisher’s Exact Test’s from data presented in G. Key significant comparisons are highlighted in yellow. Adj p-value < 3.33E-03 was used for cutoff. https://doi.org/10.1371/journal.pgen.1010828.g002 Ubiquitous expression of STING reverts loss of STING but overexpression alone does not further exaggerate parkin mutant phenotypes To test specificity for loss of sting in suppressing the parkin-/- phenotypes, flies were generated to restore expression of STING in the stingΔRG5; park25 background. The park25 allele was recombined with a pAttB-UAS-STING-V5 transgene and with the ubiquitous driver Daughterless-Gal4 (Da.Gal4). Overexpression of STING with Da.Gal4 in parkin wild-type animals had no effect on parkin-related thorax phenotypes or mitochondria morphology (Fig 3B and 3D). These two chromosomes were moved into the stingΔRG5 mutant background, and then crossed together. The progeny expressing STING-V5 in a sting-/- and parkin-/- mutant background had high penetrance of thorax indentations and bent-down wings compared with sibling flies or progeny from a control cross to stingΔRG5; park25 with no Gal4 (Fig 3A–3C, 3G and 3H). The slightly increased proportion of bent wings and small disruptions in the mitochondria networks in the stingΔRG5;park25/25, UAS-STING flies are potentially due to “leaky” expression of the UAS-STING allele. Additionally, hs70-Gal4 driven expression of STING-V5 in park25 mutant flies did not affect the severity of the thorax indentations or mitochondria morphology (Fig 3B and 3F). Together, these results indicate that STING is involved in development of muscle degeneration of parkin-/- flies but increasing expression of STING is not sufficient to induce damage. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 3. Overexpression of STING reverts suppression of park phenotypes from deletion of sting without further increasing phenotype severity. (A) Examples of flies from crosses testing over-expression of a UAS-STING transgene. Re-expression of STING with the ubiquitous daughterless-Gal4 restored the parkin phenotypes in an otherwise sting-/-; park-/- background. (B & C) Quantification of thorax indentation and wing posture phenotypes in multiple UAS-STING overexpressing flies. Overexpression of UAS_Sting with da.Gal4 did not result in wing or thorax defects in wild-type flies (harboring two copies of parkin). Note that with the hs70-Gal4 driver, overexpression of Sting in sting wildtype but parkin mutant flies did not increase the severity or proportions of parkin mutant thorax defects. Error bars represent the 95% confidence interval for the population proportion and the numbers indicate number of flies scored. (D-H) Representative images of mitochondria morphology in IFM samples of the indicated genotypes. Samples were imaged and examined in a blinded manner. All scale bars represent 10μM and the images were linearly adjusted for brightness and contrast to avoid obscuring morphology. https://doi.org/10.1371/journal.pgen.1010828.g003 Apoptosis is reduced in sting; parkin flies, whereas phosphorylated Ub is elevated Apoptotic nuclei appear in the IFM of parkin mutant flies shortly following enclosing and cell death persists throughout adulthood [14,17]. To test whether loss of STING protects flies from muscle apoptosis, thoraces from flies aged one-day post-eclosion were dissected and TUNEL staining was performed to detect apoptotic nuclei (Fig 4A). A high number of TUNEL-positive nuclei were observed in the park25 mutant flies (Fig 4B). Loss of sting significantly suppressed the number of TUNEL-positive nuclei (Fig 4A and 4B), suggesting that STING is promoting apoptosis in the parkin mutant flies. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 4. Loss of STING suppresses cell death without preventing Ubiquitin phosphorylation. (A) Example images of TUNEL staining on thoracic muscle from the indicated genotypes. Images are max projection stacks of 10 slices at 1μm step size. Scale bars represent 50μm. (B) Quantification of relative number of TUNEL-Positive Nuclei. Data is graphed as number of positive nuclei per μm3 volume of muscle quantified from phalloidin staining. N = 3 biological replicas, Significance was determined based on the results of the Kruskal-Wallis test, with Dunn’s multiple comparisons test. (C) Normalized protein lysates from 5 adults flies of the indicated genotypes were subjected to SDS-PAGE followed by Western blotting for pSer65 Ubiquitin. (D) Quantification of lanes from pSer65-Ub western blots. N = 3 biological replicas, Significance was determined based on the results of an 1-way ANOVA, followed by Bonferroni’s multiple comparison’s testing. https://doi.org/10.1371/journal.pgen.1010828.g004 To assess whether activation of the PINK1/Parkin pathway was affected in STING-null flies, western blotting for phosphorylated-Serine65 of Ubiquitin was performed (Fig 4C). No change was detected in the amount of p-S65-Ub due to loss of sting alone (Fig 4D). Consistent with a previous report, park25 mutants display a high amount of p-S65-Ub, attributed to high basal PINK1 activity and decreased capacity to degrade ubiquitinated proteins via the proteasome or mitophagy [41]. We confirmed this result and replicated this in park1 mutants as well. Deletion of sting in either of the parkin mutant backgrounds further increased the amount of p-S65-Ub (Fig 4D). Western blots on protein samples isolated from dissected thoraxes, with the gut tract removed, confirmed that a similar increase in p-S65-Ub occurs in the thoracic wing muscle of stingΔRG5; park25 mutants (S3A and S3B Fig). As it was unclear whether the rescued mitochondrial morphology and decreased cell death contributes to the increase in the relative amount of p-Ubiquitin, we tested these samples already normalized for total protein levels for the mitochondrial respiratory Complex V subunit ATP5α (S3C Fig). ATP5α levels were slightly lowered in pink1 and both parkin mutants, and deletion of sting slightly increased amount of ATP5α (S3C and S3D Fig). When p-Ubiquitin is normalized to the relative amount of mitochondria protein, the difference in parkin flies and the sting; parkin double mutant flies is less severe, although still increased, than observed in the un-normalized data (Figs 4D and S3E). Together these results demonstrate that deletion of STING does not suppress phosphorylation of Ubiquitin at Ser65, and that this Pink1-mediated pathway remains activated. We also assessed the levels of p62 (dm: ref(2)p, hs: SQSTM), a major autophagy receptor in flies, which has been implicated in regulation of pink1/parkin-dependent mitophagy [42] and overexpression of p62 suppresses mitochondria dysfunction in muscles associated with aging [43]. Western blotting against p62 reveals an increase in p62 in the rescued stingΔRG5; park25 animals (S3F and S3G Fig), coinciding with the observed increase in pSer65-Ub and prevention of mitochondria turnover. Canonical STING signaling is not activated in young parkin flies STING is reported to act upstream of the NF-κB transcription factor Relish in Drosophila and regulate both anti-bacterial and anti-viral responsive genes [31,44–46]. We assayed STING-dependent response genes in parkin mutants and control flies and observed no aberrant activation of the STING-regulated anti-viral genes srg2 (CG42825) and srg3 (CG33926) in parkin-/- flies with RT-qPCR (S4A–S4D Fig). To test the hypothesis that decreased relish signaling is involved in the rescue of the parkin fly phenotype by deletion of sting, we generated fly lines with the park25 and relE20 null deletions recombined [47]. This combination of homozygous mutants results in lethality as among greater than 200 flies collected from three independent recombined lines, no homozygous park25, relE20 flies were observed (S4E Fig). We hypothesize that, since park-/- flies are hypersensitive to bacteria propagation [48,49], the combination of defects from loss of rel result in synthetic lethality, possibly distinct from the role of sting-mediated immune responses in park-/- flies. Further, an allele harboring null mutations of two cGAS-Like receptors, cGLR1ko and cGLR2ko [50] failed to completely replicate the loss of sting with regards to the parkin1 phenotypes (S4F Fig). In the cGLR1ko, cGLR2ko; parkin1 flies, only a minor decrease of the thorax phenotype penetrance was observed, and there was no effect on the severity of the wing posture defects. We then assayed levels of mtDNA, a putative, yet untested, cGLR-activating signal. From total column-purified DNA samples, the mtDNA copy number (normalized to nuclear DNA levels) was significantly lowered in parkin mutants, compared to the w1118 background controls. Loss of sting returned these mtDNA levels to approximately that of wild-type (S4G Fig). This supports the hypothesis that the disruption of mitochondrial homeostasis in parkin mutants is suppressed by deletion of sting (see also Fig 2). These findings and the evidence that loss of STING prevents the mitochondria morphology defects suggest that STING’s role in the parkin-/- flies may be separate from the reported function in anti-viral innate immunity. Transcriptomes of stingΔRG5;park25 flies implicate additional stress-response and innate immune pathways The unexpected result of the improved mitochondria morphology and the lack of an increase in STING-regulated expression of two anti-viral genes in the park mutant suggests more complex models for the suppression of park phenotypes when sting is mutated. Therefore, we performed RNA-sequencing to compare the transcriptomes of the stingΔRG5, park25, and stingΔRG5; park25 mutant flies, using the shared background stock w1118 as our wild-type control. Samples of total RNA from ten age-matched male flies (4–5 days post-eclosion) were prepared for RNA-sequencing. Following sequencing and preliminary analysis for quality control, at least two independent replicas per group were used for differential gene expression and gene set enrichment analysis. We verified that the expression levels of the STING-regulated genes srg2 (CG42825) and srg3 (CG33926) were significantly lower in both sting mutant groups and found that, in contrast, anti-viral srg1 is slightly increased in park25 mutant flies and not in stingΔRG5; park25 mutant flies (S4H Fig). Some of IMD/Relish mediated antimicrobial peptides, previously linked to STING activity following infection with Listeria monocytogenes [31], were shown to be elevated in the park mutants, which matches prior reports of AMP activity in parkin flies (S4H Fig). The most significantly upregulated GO term category in the park vs wildtype transcriptome comparison is antibacterial humoral response (Fig 5A). Thus, although it remains unclear if STING-mediated transcriptional responses are involved in the parkin fly phenotypes, if so, it would appear that antibacterial responses would be more important than anti-viral responses. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 5. Transcriptomic profiling reveals a role for pro-survival and stress responsive genes in suppressing parkin phenotypes. (A) Gene set enrichment analysis (GSEA) for RNA-sequencing results, comparing either (left) parkin to wild-type control, or (right) the stingΔRG5; park25 double knockout to parkin mutants. Enriched Gene Ontology (GO) terms are displayed as a dot plot, separating biological processes (bp) from cellular component sets (CC). Count indicates number of genes in the GO set, and p-values represent the adjusted p-value using the BH method. Note that in the double knockout vs parkin set the only significantly enriched sets were activated (higher in the double mutant). (B). Ridgeplots of the results from GSEA with the KEGG classifications network. Graphs indicate the distribution of expression levels for significantly enriched KEGG sets for analysis of parkin compared to wild-type samples, and stingΔRG5; park25 to parkin mutants. (C) Gene concept network after overrepresentation analysis (ORA) of the top differentially expressed genes between the stingΔRG5; park25 and parkin sets. Significance was determined with an adjusted FDR cutoff of 0.05 and Log2FC cutoff of 2. (D) Heatmap plot indicating expression levels of the top differentially expressed genes between the stingΔRG5; park25 double knockout and parkin mutants, graphed by gene and gene family set. (E) Activity assays for GST conjugation from thorax lysates of 10 flies of the specified genotype. Data is shown as percent activity compared to the positive purified GST control. Results from four biological replicas. Significance was determined based on the results of 1-way ANOVA, followed by Bonferroni’s multiple comparison’s test. https://doi.org/10.1371/journal.pgen.1010828.g005 Compared with the wild-type controls, parkin mutant samples have consistently lower expression of genes involved in mitochondrial respiration (Fig 5A, left panel), and expression of these genes was rescued in the double mutants (Fig 5A, right panel). One significantly enriched gene set in the double mutant flies is genes involved in glutathione metabolic processes (GO:0006749, KEGG: N00904) (Figs 5A and 5B and S5B). Compared to wild-type controls, parkin flies also have increased heat-responsive and humoral immune-response genes (Fig 5A, left panel). Examination of the highest enriched genes (Figs 5C and 5D and S5C) suggests that these changes come in part from higher expression of Turandot genes, a family of heat-response and oxidative stress-induced genes [51]. TotA and TotC are highly enriched in the sick parkin mutant flies, are significantly decreased in the double mutants, and lowest expressed in the two control groups (w1118 and sting-/-) (Figs 5D, S5C and S5D). Due to a previously established connection to parkin phenotypes, we hypothesized that increased GST activity could lessen the burden of toxic species in the stingΔRG5; park25 double mutant animals. The relative activity of GST enzymes in fly thorax protein extracts was investigated using a GST enzymatic assay. Loss of either parkin or sting had a slight nonsignificant increase in GST activity, however loss of both genes resulted in an increase of approximately 2-fold activity compared to the wild-type samples (Fig 5E). Given this relatively small increase in GST activity in the adult flies there may yet be additional undiscovered factors improving the health of these animals. STING is necessary for muscle degeneration and climbing defects in parkin flies Thorax indention and bent wing phenotypes in parkin mutant flies are indicative of underlying indirect flight muscle (IFM) defects and attributed to mitochondria dysfunction inducing muscle apoptosis [14,19,30]. We generated flies harboring null alleles for parkin (park25) [14] and sting (stingΔRG5) [31]. Analysis of these double knockout (DKO) flies demonstrated that loss of sting rescued both the thorax and wing phenotypes of the parkin mutant flies (Fig 1A–1C). We obtained the independently derived park1 mutant and backcrossed this allele into the stingΔRG5 mutant background [15]. These flies also demonstrated reduced penetrance of the parkin phenotypes (Fig 1A–1C). For both backgrounds, the status of the sting and park-null alleles were scored based on the presence or absence of the balancer chromosomes and fly genotypes were routinely confirmed using PCR (S1 Fig). Both park25 and park1 homozygous flies demonstrate climbing defects, due to muscle degeneration and, later, age-dependent loss of dopaminergic neurons [14,15,32]. Using the negative geotaxis assay (Fig 1D), flies homozygous for stingΔRG5 were assayed for climbing ability in parkin wild-type, park25, and park1 backgrounds. Loss of sting alone had no effect on climbing ability in young (5–7 days-old) flies. For both parkin alleles, loss of sting suppressed the climbing defects of young parkin null adults (Fig 1D and S1 Video). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 1. STING mediates flight muscle degeneration in parkin-/- flies. (A) Representative images of the thoracic muscle indentation. The stingΔRG5 allele crossed to either null parkin allele (yellow arrows) rescues the thoracic defects of park25 and park1 mutants (red arrow). All flies were generated in or crossed to a wild type w1118 stock (B & C) Quantification of the thoracic indentations (B) or the downward bent wing posture (C) in the indicated genotypes. In all graphs, bars represent the percentage of flies displaying the indicated phenotype, numbers within or juxtaposed to the bars indicate the number of flies scored per genotype (n), and the error bars represent the 95% confidence interval for the population proportion. (D) Scatter plots of quantifications for negative geotaxis assays in the indicated genotypes. Each data point represents the mean of at least 3 technical replicate assays with a group of 15 to 20 flies. Horizontal bars indicate the mean of 5 independent biological replicas per genotype. Error bars display the standard deviation. Genotypes were tested for statistical significance with an 1-way ANOVA test with post-hoc multiple comparison testing with Bonferroni’s correction. (E) Example images for pink15, pink1B9, pink15; stingΔRG5 and pink1B9; stingΔRG5 male flies. Note that the loss of sting slightly affects the pink1-null phenotypes, in contrast to the strong level of suppression seen in parkin mutant combinations. (F & G) Quantification of the thorax indention and wing posture defect phenotypes in pink1-/y, or pink1-/y; stingΔRG5 flies. In all graphs, bars represent the percentage of flies displaying the indicated phenotype, numbers indicate the number of flies scored per genotype, and the error bars represent the 95% confidence interval for the population proportion. Significance was determined using Fisher’s Exact Test for differences between population proportions. Significant p-values are indicated on the graphs. https://doi.org/10.1371/journal.pgen.1010828.g001 We confirmed the veracity of the stingΔRG5 knockout allele by crossing stingΔRG5 flies with flies containing a sting deficiency chromosome in the park25 mutant background (S1E Fig). Resulting progeny harboring one copy of stingΔRG5 allele and the sting deletion displayed suppressed thorax and wing phenotypes in the homozygous park25 mutation (S1F–S1H Fig). These results support a necessary role for STING in progression of muscle degeneration of parkin-/- flies. We also observed that loss of sting in the pink15 or pink1B9 hemizygous mutant background rescued the severity of the thorax phenotypes only partially and to a lesser extent than in parkin-null flies (Fig 1E–1G). Pink1 has been reported to have multiple Parkin-independent interactions [21,33–35], and loss of STING may not affect these pathways, resulting in only minor suppression of the pink1 thoracic muscle phenotypes. To investigate our results on parkin that diverge from Lee et al. [29], we acquired the stingΔRG5; park25 line used in that study. We verified these animals with RT-qPCR and scored thorax and wing phenotypes in the homozygous stingΔRG5; park25 flies (S2 Fig). This line of stingΔRG5; park25 flies displayed minimal thorax indention phenotype but retained the park25 bent wing phenotype (S2C Fig). One explanation for the divergent results compared to Lee et al. [29] could be differences in the genetic background. Therefore the gifted stingΔRG5; park25 stock underwent eight generations of outcrossing to the w1118 stock followed by single-male fly crosses to a double balancer stock as described in detail in the Materials and Methods section. Resulting fly lines that retained the stingΔRG5 allele and the park25 allele, as tested with PCR, were self-crossed to test the resulting homozygous progeny. In the outcrossed stocks, loss of sting suppressed both the thorax indentation and bent wings of the parkin flies, compared to stingΔRG5/+(heterozygous); park25(homozygous) siblings (S2B and S2C Fig). Therefore, it appears that a yet-unknown background difference could contribute to the severity of the park25 phenotypes in stingΔRG5 mutant flies. STING influences the underlying mitochondria pathology in parkin flies Defects in parkin-null flies include disrupted mitochondrial morphology in indirect flight muscles (IFM) [14–17]. This has been linked to dysfunctional mitochondrial dynamics attributed both to blocking of mitochondrial autophagy [36,37] and to disruption of mitochondria fusion and fission dynamics [20,38]. To assess the mitochondrial health in the parkin-/- and sting-/- flies, we examined the IFM-associated mitochondria in thoraces of young flies (3–5 days post-eclosion) using Alexa Fluor488-labeled streptavidin to visualize mitochondria (Fig 2) [39,40]. As previously reported, park25 and park1 mutants possess disrupted morphology, with interrupted mitochondrial networks and the appearance of large swollen mitochondria aggregates (Fig 2C and 2E). These defects were substantially suppressed when the stingΔRG5 allele was crossed to either of the park-null alleles (Fig 2D and 2F), whereas loss of sting alone had no mitochondrial disruption compared to controls (Fig 2A and 2B). Blinded scoring of the IFM mitochondria integrity in randomized examples of ten thoraces per genotype reveals that although the mitochondria aggregation is partially suppressed, loss of sting does not completely restore mitochondria health (Fig 2G). These results suggest a role for STING function upstream or in parallel to the mitochondrial damage phenotypes in parkin-/- flies. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 2. Mitochondrial defects in parkin-/- flight muscles are suppressed by mutation of sting. Representative micrographs from indirect flight muscle tissue in w1118 (A) and stingΔRG5 (B) thoraces, and in flies homozygous for either of the parkin-null alleles (C and E). Loss of sting mitigates the swollen mitochondria defects in park25 and park1 muscles (D and F). Staining of mitochondria was performed with AlexaFluor488-conjugated streptavidin and actin bundles were visualized with iFluor647-conjugated phalloidin. Each image is a single 1μm confocal slice. All scale bars represent 10μM. Images were linearly adjusted for brightness and contrast to avoid obscuring morphology (A’–F’) 2X digital zoom of the corresponding mitochondria image, indicated with white dotted box. (G) Quantification of mitochondria morphology with blinded analysis from 10 examples per genotype, presented randomly. Data displays the percentage of thoraces in each category for each indicated genotype. (H) Summary of Fisher’s Exact Test’s from data presented in G. Key significant comparisons are highlighted in yellow. Adj p-value < 3.33E-03 was used for cutoff. https://doi.org/10.1371/journal.pgen.1010828.g002 Ubiquitous expression of STING reverts loss of STING but overexpression alone does not further exaggerate parkin mutant phenotypes To test specificity for loss of sting in suppressing the parkin-/- phenotypes, flies were generated to restore expression of STING in the stingΔRG5; park25 background. The park25 allele was recombined with a pAttB-UAS-STING-V5 transgene and with the ubiquitous driver Daughterless-Gal4 (Da.Gal4). Overexpression of STING with Da.Gal4 in parkin wild-type animals had no effect on parkin-related thorax phenotypes or mitochondria morphology (Fig 3B and 3D). These two chromosomes were moved into the stingΔRG5 mutant background, and then crossed together. The progeny expressing STING-V5 in a sting-/- and parkin-/- mutant background had high penetrance of thorax indentations and bent-down wings compared with sibling flies or progeny from a control cross to stingΔRG5; park25 with no Gal4 (Fig 3A–3C, 3G and 3H). The slightly increased proportion of bent wings and small disruptions in the mitochondria networks in the stingΔRG5;park25/25, UAS-STING flies are potentially due to “leaky” expression of the UAS-STING allele. Additionally, hs70-Gal4 driven expression of STING-V5 in park25 mutant flies did not affect the severity of the thorax indentations or mitochondria morphology (Fig 3B and 3F). Together, these results indicate that STING is involved in development of muscle degeneration of parkin-/- flies but increasing expression of STING is not sufficient to induce damage. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 3. Overexpression of STING reverts suppression of park phenotypes from deletion of sting without further increasing phenotype severity. (A) Examples of flies from crosses testing over-expression of a UAS-STING transgene. Re-expression of STING with the ubiquitous daughterless-Gal4 restored the parkin phenotypes in an otherwise sting-/-; park-/- background. (B & C) Quantification of thorax indentation and wing posture phenotypes in multiple UAS-STING overexpressing flies. Overexpression of UAS_Sting with da.Gal4 did not result in wing or thorax defects in wild-type flies (harboring two copies of parkin). Note that with the hs70-Gal4 driver, overexpression of Sting in sting wildtype but parkin mutant flies did not increase the severity or proportions of parkin mutant thorax defects. Error bars represent the 95% confidence interval for the population proportion and the numbers indicate number of flies scored. (D-H) Representative images of mitochondria morphology in IFM samples of the indicated genotypes. Samples were imaged and examined in a blinded manner. All scale bars represent 10μM and the images were linearly adjusted for brightness and contrast to avoid obscuring morphology. https://doi.org/10.1371/journal.pgen.1010828.g003 Apoptosis is reduced in sting; parkin flies, whereas phosphorylated Ub is elevated Apoptotic nuclei appear in the IFM of parkin mutant flies shortly following enclosing and cell death persists throughout adulthood [14,17]. To test whether loss of STING protects flies from muscle apoptosis, thoraces from flies aged one-day post-eclosion were dissected and TUNEL staining was performed to detect apoptotic nuclei (Fig 4A). A high number of TUNEL-positive nuclei were observed in the park25 mutant flies (Fig 4B). Loss of sting significantly suppressed the number of TUNEL-positive nuclei (Fig 4A and 4B), suggesting that STING is promoting apoptosis in the parkin mutant flies. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 4. Loss of STING suppresses cell death without preventing Ubiquitin phosphorylation. (A) Example images of TUNEL staining on thoracic muscle from the indicated genotypes. Images are max projection stacks of 10 slices at 1μm step size. Scale bars represent 50μm. (B) Quantification of relative number of TUNEL-Positive Nuclei. Data is graphed as number of positive nuclei per μm3 volume of muscle quantified from phalloidin staining. N = 3 biological replicas, Significance was determined based on the results of the Kruskal-Wallis test, with Dunn’s multiple comparisons test. (C) Normalized protein lysates from 5 adults flies of the indicated genotypes were subjected to SDS-PAGE followed by Western blotting for pSer65 Ubiquitin. (D) Quantification of lanes from pSer65-Ub western blots. N = 3 biological replicas, Significance was determined based on the results of an 1-way ANOVA, followed by Bonferroni’s multiple comparison’s testing. https://doi.org/10.1371/journal.pgen.1010828.g004 To assess whether activation of the PINK1/Parkin pathway was affected in STING-null flies, western blotting for phosphorylated-Serine65 of Ubiquitin was performed (Fig 4C). No change was detected in the amount of p-S65-Ub due to loss of sting alone (Fig 4D). Consistent with a previous report, park25 mutants display a high amount of p-S65-Ub, attributed to high basal PINK1 activity and decreased capacity to degrade ubiquitinated proteins via the proteasome or mitophagy [41]. We confirmed this result and replicated this in park1 mutants as well. Deletion of sting in either of the parkin mutant backgrounds further increased the amount of p-S65-Ub (Fig 4D). Western blots on protein samples isolated from dissected thoraxes, with the gut tract removed, confirmed that a similar increase in p-S65-Ub occurs in the thoracic wing muscle of stingΔRG5; park25 mutants (S3A and S3B Fig). As it was unclear whether the rescued mitochondrial morphology and decreased cell death contributes to the increase in the relative amount of p-Ubiquitin, we tested these samples already normalized for total protein levels for the mitochondrial respiratory Complex V subunit ATP5α (S3C Fig). ATP5α levels were slightly lowered in pink1 and both parkin mutants, and deletion of sting slightly increased amount of ATP5α (S3C and S3D Fig). When p-Ubiquitin is normalized to the relative amount of mitochondria protein, the difference in parkin flies and the sting; parkin double mutant flies is less severe, although still increased, than observed in the un-normalized data (Figs 4D and S3E). Together these results demonstrate that deletion of STING does not suppress phosphorylation of Ubiquitin at Ser65, and that this Pink1-mediated pathway remains activated. We also assessed the levels of p62 (dm: ref(2)p, hs: SQSTM), a major autophagy receptor in flies, which has been implicated in regulation of pink1/parkin-dependent mitophagy [42] and overexpression of p62 suppresses mitochondria dysfunction in muscles associated with aging [43]. Western blotting against p62 reveals an increase in p62 in the rescued stingΔRG5; park25 animals (S3F and S3G Fig), coinciding with the observed increase in pSer65-Ub and prevention of mitochondria turnover. Canonical STING signaling is not activated in young parkin flies STING is reported to act upstream of the NF-κB transcription factor Relish in Drosophila and regulate both anti-bacterial and anti-viral responsive genes [31,44–46]. We assayed STING-dependent response genes in parkin mutants and control flies and observed no aberrant activation of the STING-regulated anti-viral genes srg2 (CG42825) and srg3 (CG33926) in parkin-/- flies with RT-qPCR (S4A–S4D Fig). To test the hypothesis that decreased relish signaling is involved in the rescue of the parkin fly phenotype by deletion of sting, we generated fly lines with the park25 and relE20 null deletions recombined [47]. This combination of homozygous mutants results in lethality as among greater than 200 flies collected from three independent recombined lines, no homozygous park25, relE20 flies were observed (S4E Fig). We hypothesize that, since park-/- flies are hypersensitive to bacteria propagation [48,49], the combination of defects from loss of rel result in synthetic lethality, possibly distinct from the role of sting-mediated immune responses in park-/- flies. Further, an allele harboring null mutations of two cGAS-Like receptors, cGLR1ko and cGLR2ko [50] failed to completely replicate the loss of sting with regards to the parkin1 phenotypes (S4F Fig). In the cGLR1ko, cGLR2ko; parkin1 flies, only a minor decrease of the thorax phenotype penetrance was observed, and there was no effect on the severity of the wing posture defects. We then assayed levels of mtDNA, a putative, yet untested, cGLR-activating signal. From total column-purified DNA samples, the mtDNA copy number (normalized to nuclear DNA levels) was significantly lowered in parkin mutants, compared to the w1118 background controls. Loss of sting returned these mtDNA levels to approximately that of wild-type (S4G Fig). This supports the hypothesis that the disruption of mitochondrial homeostasis in parkin mutants is suppressed by deletion of sting (see also Fig 2). These findings and the evidence that loss of STING prevents the mitochondria morphology defects suggest that STING’s role in the parkin-/- flies may be separate from the reported function in anti-viral innate immunity. Transcriptomes of stingΔRG5;park25 flies implicate additional stress-response and innate immune pathways The unexpected result of the improved mitochondria morphology and the lack of an increase in STING-regulated expression of two anti-viral genes in the park mutant suggests more complex models for the suppression of park phenotypes when sting is mutated. Therefore, we performed RNA-sequencing to compare the transcriptomes of the stingΔRG5, park25, and stingΔRG5; park25 mutant flies, using the shared background stock w1118 as our wild-type control. Samples of total RNA from ten age-matched male flies (4–5 days post-eclosion) were prepared for RNA-sequencing. Following sequencing and preliminary analysis for quality control, at least two independent replicas per group were used for differential gene expression and gene set enrichment analysis. We verified that the expression levels of the STING-regulated genes srg2 (CG42825) and srg3 (CG33926) were significantly lower in both sting mutant groups and found that, in contrast, anti-viral srg1 is slightly increased in park25 mutant flies and not in stingΔRG5; park25 mutant flies (S4H Fig). Some of IMD/Relish mediated antimicrobial peptides, previously linked to STING activity following infection with Listeria monocytogenes [31], were shown to be elevated in the park mutants, which matches prior reports of AMP activity in parkin flies (S4H Fig). The most significantly upregulated GO term category in the park vs wildtype transcriptome comparison is antibacterial humoral response (Fig 5A). Thus, although it remains unclear if STING-mediated transcriptional responses are involved in the parkin fly phenotypes, if so, it would appear that antibacterial responses would be more important than anti-viral responses. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 5. Transcriptomic profiling reveals a role for pro-survival and stress responsive genes in suppressing parkin phenotypes. (A) Gene set enrichment analysis (GSEA) for RNA-sequencing results, comparing either (left) parkin to wild-type control, or (right) the stingΔRG5; park25 double knockout to parkin mutants. Enriched Gene Ontology (GO) terms are displayed as a dot plot, separating biological processes (bp) from cellular component sets (CC). Count indicates number of genes in the GO set, and p-values represent the adjusted p-value using the BH method. Note that in the double knockout vs parkin set the only significantly enriched sets were activated (higher in the double mutant). (B). Ridgeplots of the results from GSEA with the KEGG classifications network. Graphs indicate the distribution of expression levels for significantly enriched KEGG sets for analysis of parkin compared to wild-type samples, and stingΔRG5; park25 to parkin mutants. (C) Gene concept network after overrepresentation analysis (ORA) of the top differentially expressed genes between the stingΔRG5; park25 and parkin sets. Significance was determined with an adjusted FDR cutoff of 0.05 and Log2FC cutoff of 2. (D) Heatmap plot indicating expression levels of the top differentially expressed genes between the stingΔRG5; park25 double knockout and parkin mutants, graphed by gene and gene family set. (E) Activity assays for GST conjugation from thorax lysates of 10 flies of the specified genotype. Data is shown as percent activity compared to the positive purified GST control. Results from four biological replicas. Significance was determined based on the results of 1-way ANOVA, followed by Bonferroni’s multiple comparison’s test. https://doi.org/10.1371/journal.pgen.1010828.g005 Compared with the wild-type controls, parkin mutant samples have consistently lower expression of genes involved in mitochondrial respiration (Fig 5A, left panel), and expression of these genes was rescued in the double mutants (Fig 5A, right panel). One significantly enriched gene set in the double mutant flies is genes involved in glutathione metabolic processes (GO:0006749, KEGG: N00904) (Figs 5A and 5B and S5B). Compared to wild-type controls, parkin flies also have increased heat-responsive and humoral immune-response genes (Fig 5A, left panel). Examination of the highest enriched genes (Figs 5C and 5D and S5C) suggests that these changes come in part from higher expression of Turandot genes, a family of heat-response and oxidative stress-induced genes [51]. TotA and TotC are highly enriched in the sick parkin mutant flies, are significantly decreased in the double mutants, and lowest expressed in the two control groups (w1118 and sting-/-) (Figs 5D, S5C and S5D). Due to a previously established connection to parkin phenotypes, we hypothesized that increased GST activity could lessen the burden of toxic species in the stingΔRG5; park25 double mutant animals. The relative activity of GST enzymes in fly thorax protein extracts was investigated using a GST enzymatic assay. Loss of either parkin or sting had a slight nonsignificant increase in GST activity, however loss of both genes resulted in an increase of approximately 2-fold activity compared to the wild-type samples (Fig 5E). Given this relatively small increase in GST activity in the adult flies there may yet be additional undiscovered factors improving the health of these animals. Discussion Together, our findings support a non-canonical role for Drosophila STING in the pathogenesis of mitochondria dysfunction in parkin-/- flies. Based on rescued mitochondria health and suppression of apoptosis, we propose that Drosophila STING is not responding solely to the presence of mitochondria-derived damage signals in the parkin mutants. These findings suggests that instead in flies there is an indirect role for STING or additional STING-induced genes in propagating upstream mitochondria-damage-induced signaling or indirectly promoting apoptosis. Additionally, there may also be a general dysregulation of autophagy or intraorganellar signaling in the STING-null mutants, as recent studies show that STING modulates autophagy [45,52,53] and lipid dependent starvation responses [54]. Boosting of autophagy through expression of ATG1 has also been shown to prevent mitochondria aggregation and rescue phenotypes of the parkin-/- flies [38]. We have identified a significant increase in p62 levels in the double mutant animals, suggesting either a block in autophagic turnover, or increased expression of p62. Increased amounts of p62 promotes longevity in flies [42,43] and promotes pro-survival NRF2 (CnC in flies) activity through an inhibitory interaction with the NRF2 regulator KEAP1 [55,56]. Previous work demonstrates that Drosophila STING’s function in innate immunity requires activation of the IMD (immune deficiency) pathway leading to increased Relish (NF-κB) signaling and this activation is partly dependent on the Drosophila IKKβ homologue [44,46]. Supporting this requirement for IKK signaling, the Drosophila IKKε homologue has been demonstrated to interact genetically with parkin mutations, as loss of IKKε suppressed the parkin wing and thorax phenotypes [57]. IMD/Rel-induced AMPs were previously upregulated in a transcriptomic study of parkin-/- mutants [27] and such AMPs are reported to promote neurodegeneration in aging animals [58]. A possible mechanism for our observations is that minor damage to mitochondria could signal through STING to induce antimicrobial gene expression that feeds back on mitochondria to cause unmitigable damage when Parkin is absent. The RNA-Seq analysis revealed a slightly increased expression of AMPs that are regulated by the canonical IMD/Rel pathway- including Attacin (AttA) and Diptericin (DptA)- or the MyD88/Toll pathway- Drosomycin (Drs) and Metchnikowin (Mtk)- in young parkin-/- flies. However, as tested with qPCR (S4A–S4D Fig), and corroborated with RNA-seq results (S4H Fig), the expression levels of two anti-viral STING-regulated genes were not consistently increased in the adult park-25 samples. Combination of a mutant allele lacking two cGAS-like Receptors cGLR1 and cGLR2 with the parkin1 mutant animal has a slight, but significant suppression of the parkin mutant thorax indention penetrance, indicating that cGLR1/cGLR2 may be dispensable for the role of STING in the fly parkin phenotype. There exist additional cGAS-like-receptors and knockout of all these simultaneously in parkin flies would be intriguing, yet technically challenging [50,59,60]. Recent evidence suggests that Drosophila STING possesses additional functions independent of the canonical activation of NF-kB innate immune signaling genes, such as regulation of autophagy or metabolism related pathways [45,53,54]. RNA-sequencing results and previously published microarray data from sting mutants [54] supports that there are additional cellular pathways dysregulated in flies lacking sting besides immune-related genes. The observed increase in Glutathione S-transferase enzyme expression could convey cytoprotective antioxidant buffering to the double mutant flies. Elevated expression of Glutathione S-transferases [32,61,62], toxic metal responsive genes such as MTF-1 [63], and increased activity of the antioxidative stress KEAP1/NRF2 pathway—which regulates GST gene expression—have each been demonstrated to suppress muscle and/or climbing phenotypes in parkin or pink1 mutant flies [32,61–64]. A previously published dataset (GEO accession #GSE167164) shows an upregulation in anti-toxin and anti-pesticide genes such as GstD1 and Cytochrome p450 family members in sting-/- mutants [54]. The transcriptomic analysis we performed did not reveal as strong of a change in these genes between the stingΔRG5 mutant and the wildtype control, however, we did observe a significant increase of GstE1, GstE11, and GstD2 in the stingΔRG5; park25 mutant samples. Furthermore, sting mutant flies were shown to have metabolic changes related to β-oxidation and lipid storage, which may influence mitochondria bioenergetics and promote antioxidant responses [54]. We propose that an increase in oxidative stress responses and GST activity could contribute to the improved outcomes of stingΔRG5; park25 animals, however there may yet be additional signaling factors during the developmental larval and pupal stages. Additionally, it remains unknown exactly why loss of sting fails to rescue pink1 at the same degree observed in parkin mutant alleles. We hypothesize that Parkin-independent components are contributing to the muscle degeneration in pink1 mutants, therefore loss of sting fails to suppress these phenotypes completely. A recent study on pink1 mutant flies implicate a different DNA-recognition receptor, EYA, as contributing to the severity of some neuronal and gut-based pink1 mutant phenotypes through regulation of Relish signaling [65]. This contribution may be similar or completely independent of STING’s function in parkin pathology. Additionally, our observed differences in the amount of Ubiquitin phosphorylation could reflect increased amounts of Pink1 activity, the Ub substrates, or a decrease in deubiquinating enzymes. The molecular details of phosphorylated-Ub regulation remains of high interest to the pink1/parkin field [41,66], reviewed recently in [67]. In summary, loss of sting in flies suppresses the severe phenotypes of parkin mutants, through a mechanism(s) independent of the canonical role in innate immunity signaling. The candidate pathways supported by our data includes anti-oxidative stress responses and activation of cell death pathways. These underlying changes to the transcriptional landscape in sting-/- flies necessitates further study to better understand the role of stress-responsive genes in mitigating mitochondrial and oxidative damage during fly development or disease. Materials and methods Experimental subject details Publicly available fly stocks (details in Table 1) were acquired from Bloomington Drosophila Stock Center (BDSC, Bloomington, IN). Experimental genotypes (see S1 Table for all genotypes) were made using classical genetics, utilizing the balancer chromosomes from w1118; wgSp-1/CyO; MKRS/TM6b, hu (BDSC stock #76357) when necessary. The null stingΔRG5 allele was gifted from Dr. Alan Goodman, Washington State University and was previously described [31]. The park25 allele was acquired from Dr. Alicia Pickrell, Virginia Tech University, and originally generated by Dr. Leo Pallanck, University of Washington [14]. All park25 mutant animals were maintained as heterozygous over the TM6b, Hu balancer and routinely checked with PCR for presence of the deletion. A second stock of stingΔRG5; park25/TM6b, hu flies were gifted to us from Dr. Alexander Whitworth. Male flies from these stocks were crossed to a w1118 background, then outcrossed for 6 further generations. After each other generation, single male flies used in crosses were checked for PCR after the cross was seeded, and only the ones carrying the park25 allele were selected. After 7 generations, single males were crossed to the double balanced stock for maintaining the outcrossed alleles, and again, PCR was used to confirm the presence of the park25 allele. Download: PPT PowerPoint slide PNG larger image TIFF original image Table 1. Materials and Critical Resources. https://doi.org/10.1371/journal.pgen.1010828.t001 pink1[5]/FM7 female flies were outcrossed to w[1118] males. After the first cross, freely recombining pink1[5]/w[1118] females were crossed to a FM7/y; CyO/+ male to ensure the X-chromosome pink1[5] allele was recovered. From there, multiple pink1[5]/FM7; +/CyO flies were crossed with the sting[RG5] allele to generate the pink1[5]/FM7; sting[RG5]/CyO candidate lines then PCR and phenotyping was used to confirm the pink1 genotype. pink1[B9]/FM7 females were crossed with a FM7; sting[RG5]/CyO male for two generations to generate the pink1[B9]/FM7; sting[RG5]/CyO stock. All flies used in experiments were raised on vials or bottles with Jazz Mix food (Thermo Fisher Scientific, P/N AS153), reconstituted in MilliQ water and prepared per recipe instructions. For experiments, flies were raised in a 25°C incubator on a standard L:D cycle with humidity control. Generating new UAS-STING-V5 alleles For generating the pUASTattB_UAS_STING insertion, the sting cDNA was PCR-amplified from LP14056 BDGP gold cDNA (DGRC stock #1064136, FlyBaseID: FBcl0189577, RRID:DGRC_1064136) and subcloned into a modified pUAST-attB (DGRC Stock #1419, RRID:DGRC_1419) with NEBuilder HiFi DNA Assembly Mix (New England Biolabs, P/N E2621). The C-terminal V5 sequence had been inserted with two annealed oligos ligated into the XhoI and XbaI sites on pUASTattB. After verifying with sequencing, plasmids were sent to BestGene (BestGene Inc. Chino Hills, CA) for injection services using 62E1 attP landing site flies, BDSC stock #9748. The Phi31C source was removed, and the mini-white positive progeny were used to established balanced lines. The UAS-STING allele was recombined with the park25 allele, and then assayed with PCR genotyping and verified with outcross and observation of the parkin homozygous phenotype. Wing muscle and thorax phenotyping Flies were anesthetized with CO2 and thoraces were examined under a dissecting microscope. Flies were scored for thorax shape and wing posture within the first 5 minutes of anesthesia. Blinding of genotypes to observer was performed when practical, including all of the initial assays involving the key genotypes in Fig 1. Geotaxis assays Male flies were collected at 0-1d post-eclosion and aged 5–7 days before testing. For testing, 15 to 20 flies were added to empty 10cm vials, labeled randomly, and a key was generated to preserve identity of tested stocks. The vials were placed in a plastic holder and the flies were manually tapped to the bottom of the vials. The flies were recorded for 20–30 seconds post disruption. Videos were scored using ImageJ to mark and annotate individual flies. Climbing Index was calculated as the percentage of flies in a vial that climbed greater than 6cm of the vial during the observed 20 seconds post disruption. Five independent trials, each with three technical repeats, were performed for a total of at least 75 total flies per genotype. Immunohistochemistry Flies (3–5 days old) were anesthetized with CO2 and thoraces were dissected away from the head and abdomen in cold phosphate buffered saline (PBS). The hemithoraces were bisected along the median plane, using a pair of microscissors (Fine Science Tools P/N 15006–09). Hemithoraces were fixing in 4% paraformaldehyde (Electron Microscopy Sciences P/N 15710) for 20 minutes at room temperature. After fixation, tissues were washed twice in PBS and then incubated twice for 10 minutes each in PBS with 0.1% TritonX (PBST). Tissues were then blocked in 5% goat serum diluted in PBST for 30 minutes, then incubated in AlexaFluor488-conjugated streptavadin (Jackson ImmunoResearch P/N 016-540-084) and iFluor647-conjugated phalloidin (Cayman Chemical Company P/N 20555) overnight at 4°C on a rotator. Samples were then washed three times with PBST and once with PBS. Thoraces and separated muscle pieces were then mounted directly on a 1.5 coverslip in Prolong Gold AntiFade (Thermo Fisher Scientific P/N P36930) media. Images of mitochondria morphology were acquired on a Zeiss LSM 880 Airyscan confocal with a 63X/1.4 objective Plan-Apochromat (Carl Zeiss) at 2X digital zoom and a 34-channel GAsP detector. Airyscan processing was performed in ZEN Black software (Zeiss). Images were analyzed in ImageJ and adjusted linearly for contrast and brightness. Western blotting For each genotype, 5 flies (3–5 days old) were anesthetized with CO2 and thoraces were isolated. Thoraces were homogenized in RIPA buffer supplemented with cOmplete EDTA-free Protease Inhibitor (Sigma-Aldrich P/N 04693159001) and PhosSTOP- phosphatase inhibitor tablet (Roche P/N 04906845001), and then samples were incubated for 10 minutes on ice. Samples were centrifuged at 10,000xG to remove tissue remains. Protein levels were quantified using the Pierce BCA protein assay kit (Thermo Fisher Scientific P/N 23228). The protein samples were normalized and then reduced by adding Lithium Dodecyl Sulfate sample buffer (GenScript P/N M00676-250) and 0.1M DTT then heating for 5 minutes at 99°C. Protein samples were separated on 4–12% SurePAGE, Bis-Tris gels (GenScript P/N M00654) and transferred to .45μM nitrocellulose membrane (BioRad P/N 1620115). Transfer efficiency and total protein amount was visualized using Ponceau S Staining Solution (Thermo Fisher Scientific P/N A40000279). Total protein images for each blot were acquired with ChemiDoc Imaging System (Bio-Rad Laboratories). Membranes were blocked for one hour with 3% milk or with 3% bovine serum albumin (Fisher Scientific P/N BP1600) (for pS65-Ub) in Tris-Buffered Saline with 0.1% Tween-20 (TBST), and probed overnight with the indicated antibody: anti-pUb(Ser65) (1:1000, rabbit polyclonal CST P/N 62802S), anti-a-Tubulin (1:4000, mouse mAb clone B-1-5-2, Millipore Sigma, P/N T5168), anti p62/ref(2)p (purified rabbit polyclonal Ab, a gift from Dr. Helmut Kramer, UT Southwestern Medical Center), ATP5α (1:4000, mouse mAb- Abcam P/N: ab14748) or V5-Tag D3H8Q (1:1000, Rabbit mAb Cell Signaling Technologies P/N 13202). After overnight incubation, membranes were washed 3 times with TBST and incubated for 1 hour with appropriate secondary antibodies diluted 1:10,000 in TBST+3% milk: HRP-coupled Donkey-anti-Rabbit or Sheep-anti-mouse (GE Healthcare Life Sciences), Goat Anti-Rabbit IgG IRDye 800CW-Conjugated (LI-COR Biosciences P/N 926–32211), or Goat Anti-Mouse IgG Antibody IRDye 680RD-Conjugated (LI-COR Biosciences P/N 926–68070). Blots were washed three times with TBST before imaging. For p65Ser-Ub detection, HRP-conjugated secondaries were incubated with SuperSignal Femto ECL (Thermo Scientific P/N 34095) for 3 minutes and imaged with ChemiDoc Imaging System (Bio-Rad Laboratories). All other primary antibodies were visualized with HRP-conjugated secondaries and incubation with AMersham ECL (Cytiva, P/N RPN2232). IR-conjugated secondaries were incubated simultaneously for one hour, and the blots were visualized on a LICOR Odyssey Fc multichannel imager after three washes with TBST. All images were processed, adjusted linearly for brightness and contrast, and analyzed for lane densitometry in ImageJ. Quantifications for pSer65-Ub, ATP5α, and p62/ref(2)p were first normalized to the intensity of the lane’s total stain. Concerning ATP5α: differences in optimal exposure times between replicas led to the necessity to represent data as percent of wild-type control in each experiment replica (S3D Fig). pSer65 Ub quantifications were divided by the mean relative amount of ATP5α for each genotype to approximate the amount of pSer65 per mitochondria (S3E Fig). RNA isolations and RT-qPCR RNA from 5 male flies were isolated using the Direct-zol RNA Miniprep kit (Zymo Research P/N R2050). Briefly, samples were homogenized in 300μL of Tri Reagant (Zymo Research P/N R2050-1-200) and then processed using the Zymo instructions for tissue samples. On-column DNAse treatments were performed before eluting the samples in 50μL of DPEC treated RNAse-free H2O, according to the Zymo Direct-zol kit protocol (DNaseI supplied with Direct-Zol kit). Sample were tested for quantity and purity with Nanodrop. 500ng of RNA samples were used for reverse transcriptase reactions, using the High-Capacity cDNA Reverse Transcription kit (Thermo Fisher Scientific P/N 4368814). cDNA samples were diluted 1:5 before using in qPCR reactions. qPCR was performed using indicated qPCR primers (primer sequences can be found in Table 1) with PowerUp SybrGreen Master Mix (Applied Biosystems P/N A25742) using a BioRad CFX384 Touch Real-Time PCR Thermocycler. Raw data was exported from BioRad Manager then analyzed using the ddCT method with Excel. CT values were normalized to the housekeeping gene rpl32 (also referred to as rp49) and then normalized to the wild-type control sample. Data from two to three independent biological replicas with three technical replicas per sample are presented. mtDNA copy number assays Total DNA was extracted from 10 male flies with the Quick DNA Miniprep Plus kit from Zymogen, according to the provided instructions. Quantification of mtDNA was performed using a multiplex TaqMan assays using validated probes against the mitochondrial gene mt:CoI and the nuclear-encoded gene rpL32 for reference [68]. Approx. 7ng of template DNA was used for each reaction. Primer details can be found in Table 1. qPCR reactions were performed on the BioRad 384CX system according to information provided for iTaqMan Supermix (BioRad P/N 1725130) with annealing temps at 60°C. Data was analyzed using the following method in Microsoft Excel: 1. mtDNA and nucDNA CT values were averaged from triplicate reactions. 2. Mitochondrial DNA content was normalized to nuclear DNA in each sample using the following equations: ΔCT = (nucDNA CT–mtDNA CT) then relative mitochondrial DNA content = 2 × 2^ΔCT. Replica biological samples were collected and isolated on separate days. Technical replicas were performed in each qPCR reaction run. Data is presented relative to the average wild-type (w1118) mitochondrial copy number for each biological set. For statistical analysis, a 1-way ANOVA and multiple comparison testing between each of the experimental genotypes. GST assays Age matched flies were collected and raised 4–5 days under standard conditions. For the assay, thoraces were dissected from ten flies per genotype/treatment/per replica were collected and immediately put on ice. The thorax samples were homogenized in GST assay sample buffer (100mM buffered potassium phosphate solution, pH 7.0, with 2mM EDTA). Samples were centrifuged at 10,000 x g for 15 minutes at 4°C and supernatants were assayed for protein concentration using a Pierce BCA protein assay kit (Thermo Fisher Scientific P/N 23228). Samples were normalized and diluted to a protein concentration of 2μg/μL, and Glutathione S-transferase activity was measured using a GST Assay Kit (Caymen Chemicals p/n 703302). After initiating the reactions, A340 was measured every minute for ten minutes. Rates of change were calculated from the plots of A340 vs. time, the blank well absorbance was subtracted from each, and then the activity rate (A340/min) was converted to estimated GST activity with the formula: The resulting estimated activities for each technical replica (3 per biological sample) were averaged together. For each biological replica, the provided purified GST enzyme was used as a positive control, and the resulting activity for each sample is represented as percent activity compared to the purified control. Graphed data represent normalized results from four repeated experiments. Apoptosis detection staining Apoptosis assays were performed on 3–4 day old flies, using the In-Situ Cell Death Detection Kit, TMR red (Roche P/N 12156792910), according to the manufacturer’s instructions. Briefly, fly thoraces were dissected and bisected in freshly prepared PBS. Hemi-thoraces were fixed for 20 minutes in 4% Paraformaldehyde in PBS, pH 7.4, freshly prepared. Tissues were washed first in PBS and then in permeabilization solution (0.1% Triton X100 in 0.1% sodium citrate, freshly prepared) for 15 minutes. Samples were incubated in the TUNEL detection solution in a humidified atmosphere for 60 min at 37°C in the dark. Tissues were then washed 3 times in PBST and blocked in 5% goat serum and 3% BSA diluted in PBST for 30 minutes, then incubated in iFluor647-conjugated phalloidin (Cayman Chemical Company P/N 20555) overnight at 4°C on a rotator. Samples were washed 3 times with PBST and once with PBS. Thoraces and separated muscle pieces were then mounted directly on a 1.5 coverslip in Prolong Gold AntiFade (Thermo Fisher Scientific P/N P36930) media. Images of thoraces were acquired on a Zeiss LSM 880 Airyscan confocal at 20X magnification. Quantification was performed using an ImageJ macro. In brief, stacks of 10 confocal sliced were used for max intensity projections. For each projected image, thresholding was applied to detect the phalloidin-labeled actin The muscle area was measured and the thresholded region was saved as a R.O.I. The same R.O.I. was used to count for the number of TUNEL-positive stained nuclei. The number of nuclei was then divided by the total area, to approximate the number of nuclei per μm2. TUNEL experiments were repeated two times, and the presented data represents biological replicas of at least 10 thoraces per genotype. RNA sequencing and analysis Flies were collected upon eclosion and aged 4 days in identical conditions, no more than 20 animals per vial. RNA from 10 male flies, per genotype and replica, were isolated using the Direct-zol RNA Miniprep kit (Zymo Research P/N R2050). Briefly, samples were homogenized in 300μL of Tri Reagant (Zymo Research P/N R2050-1-200) and then processed using the kit instructions for tissue samples. On-column DNAse treatments were performed before eluting the samples in 50μL of DPEC treated RNAse-free H2O. RNA-Seq library preparation and next generation sequencing RNA-Seq services were provided by Zymo Research Services, using their Total-RNA-Seq protocol. RNA quality was assessed with the Agilent TapeStation System. Total RNA-Seq libraries were constructed from 100ng of total RNA. rRNA depletion was performed according to standard protocol. Libraries were prepared using the Zymo-Seq RiboFree Total RNA Library Prep Kit (Cat # R3000) according to the manufacturer’s instructions. RNA-Seq libraries were sequenced on an Illumina NovaSeq to a sequencing depth of at least 30 million read pairs (150 bp paired-end sequencing) per sample. RNA-Seq data bioinformatics analysis The Zymo Research RNA-Seq pipeline was originally adapted from nf-core/rnaseq pipeline v1.4.2 (https://github.com/nf-core/rnaseq). The pipelines were built using Nextflow (https://www.nextflow.io/).2). Briefly, quality control of reads was carried out using FastQC v0.11.9 (http://www.bioinformatics.babraham.ac.uk/projects/fastqc). Adapter and low-quality sequences were trimmed from raw reads using Trim Galore! v0.6.6 (https://www.bioinformatics.babraham.ac.uk/projects/trim_galore). Trimmed reads were aligned to the reference genome using STAR v2.6.1d (https://github.com/alexdobin/STAR) [69]. BAM file filtering and indexing was carried out using SAMtools v1.9 (https://github.com/samtools/samtools) [70]. RNAseq library quality control was implemented using RSeQC v4.0.0 (http://rseqc.sourceforge.net/) and QualiMap v2.2.2-dev (http://qualimap.conesalab.org/)) [71,72]. Duplicate reads were marked using Picard tools v2.23.9 (http://broadinstitute.github.io/picard/). Library complexity was estimated using Preseq v2.0.3 (https://github.com/smithlabcode/preseq). Duplication rate quality control was performed using dupRadar v1.18.0 (https://bioconductor.org/packages/dupRadar/) [73]. Reads overlapping with exons were assigned to genes using featureCounts v2.0.1 (http://bioinf.wehi.edu.au/featureCounts/). Classification of rRNA genes/exons and their reads were based on annotations and RepeatMasker rRNA tracks from UCSC genome browser when applicable. Differential gene expression analysis was completed using DESeq2 v1.28.0 (https://bioconductor.org/packages/DESeq2/) [74]. Quality control and analysis results plots were visualized using MultiQC v1.9 (https://github.com/ewels/MultiQC) [75]. Further analysis and visualizations on the processed data were performed in R and Bioconductor. ClusterProfiler v.4.6 and Enrichplot v.1.19.0.01 were used for gene set enrichment analysis (GSEA) and plotting [76,77]. Heatmaps with normalized counts of highly enriched genes (absolute value of Log2 fold change greater than 3 and adjusted p.value less than 0.005) were generated with pHeatmap v.1.0.12. For all plots, ggplot2 v.3.4 and ggrepel v.0.9.2 were used for annotation. For ClusterProfiler GSEA analysis, cutoffs were: minGSSize = 50, maxGSSize = 250, and Benjamini-Hochberg adjusted p.value <0.05. Quantification and statistical analysis Quantitative data was recorded, transcribed, and maintained in Microsoft Excel. Data set descriptions, exploration, statistics, and graphing was performed in Graphpad Prism v.9.3. Detailed data sets and all statistical test details are provided in S1 Data File. Details including data descriptors, sample size (n), and specific statistical tests can be found in the figure legends. Proportions of fly populations were tested with the Wilson-Brown method to determine 95% confidence intervals. For categorical data, such as mitochondria morphology scores, a Fisher’s exact test was used to test for statistical significance between genotypes. Other quantitative data was assessed for normality using the Shapiro–Wilk test. For normally-distributed data, p-values were calculated using a one-way ANOVA test followed by Bonferroni’s or Dunnett’s multiple comparison tests. The Kruskal-Wallis test, with Dunn’s multiple comparisons test, was used for non-parametric data sets. For multiple comparison tests, significance between groups was determined as adjusted p-value less than 0.05. For all experiments, no prior sample size estimation was performed. Sample sizes were determined from previous studies. For all experiments, the collection of subjects (flies) in each genotype was randomized, and no inclusion/exclusion was performed. When practical and necessary, blinding of genotypes to observer was performed. All data quantification was done in a blind or automated manner. RStudio v.2022.07.2 running R v.4.2.2 was used for processing of RNA-seq data and generating the resulting plots. Details of the analysis pipeline are available in the previous description of RNA Seq Analysis. Generally, significance was determined after Benjamini-Hochberg correction and at a level of adjusted p.value< 0.05. For microscopy experiments, raw Airyscan confocal images were acquired and processed in Zen Black (Zeiss). Images were analyzed in FIJI/ImageJ2 v.2.3.0 [78] and quantification was finished in Microsoft Excel and graphed with Graphpad Prism. Images and figures were arranged in either Microsoft Powerpoint or Inkspace (https://inkscape.org). Resource availability All unique/stable reagents and animal stocks generated in this study are available from the lead contact and will be made available on request. Experimental subject details Publicly available fly stocks (details in Table 1) were acquired from Bloomington Drosophila Stock Center (BDSC, Bloomington, IN). Experimental genotypes (see S1 Table for all genotypes) were made using classical genetics, utilizing the balancer chromosomes from w1118; wgSp-1/CyO; MKRS/TM6b, hu (BDSC stock #76357) when necessary. The null stingΔRG5 allele was gifted from Dr. Alan Goodman, Washington State University and was previously described [31]. The park25 allele was acquired from Dr. Alicia Pickrell, Virginia Tech University, and originally generated by Dr. Leo Pallanck, University of Washington [14]. All park25 mutant animals were maintained as heterozygous over the TM6b, Hu balancer and routinely checked with PCR for presence of the deletion. A second stock of stingΔRG5; park25/TM6b, hu flies were gifted to us from Dr. Alexander Whitworth. Male flies from these stocks were crossed to a w1118 background, then outcrossed for 6 further generations. After each other generation, single male flies used in crosses were checked for PCR after the cross was seeded, and only the ones carrying the park25 allele were selected. After 7 generations, single males were crossed to the double balanced stock for maintaining the outcrossed alleles, and again, PCR was used to confirm the presence of the park25 allele. Download: PPT PowerPoint slide PNG larger image TIFF original image Table 1. Materials and Critical Resources. https://doi.org/10.1371/journal.pgen.1010828.t001 pink1[5]/FM7 female flies were outcrossed to w[1118] males. After the first cross, freely recombining pink1[5]/w[1118] females were crossed to a FM7/y; CyO/+ male to ensure the X-chromosome pink1[5] allele was recovered. From there, multiple pink1[5]/FM7; +/CyO flies were crossed with the sting[RG5] allele to generate the pink1[5]/FM7; sting[RG5]/CyO candidate lines then PCR and phenotyping was used to confirm the pink1 genotype. pink1[B9]/FM7 females were crossed with a FM7; sting[RG5]/CyO male for two generations to generate the pink1[B9]/FM7; sting[RG5]/CyO stock. All flies used in experiments were raised on vials or bottles with Jazz Mix food (Thermo Fisher Scientific, P/N AS153), reconstituted in MilliQ water and prepared per recipe instructions. For experiments, flies were raised in a 25°C incubator on a standard L:D cycle with humidity control. Generating new UAS-STING-V5 alleles For generating the pUASTattB_UAS_STING insertion, the sting cDNA was PCR-amplified from LP14056 BDGP gold cDNA (DGRC stock #1064136, FlyBaseID: FBcl0189577, RRID:DGRC_1064136) and subcloned into a modified pUAST-attB (DGRC Stock #1419, RRID:DGRC_1419) with NEBuilder HiFi DNA Assembly Mix (New England Biolabs, P/N E2621). The C-terminal V5 sequence had been inserted with two annealed oligos ligated into the XhoI and XbaI sites on pUASTattB. After verifying with sequencing, plasmids were sent to BestGene (BestGene Inc. Chino Hills, CA) for injection services using 62E1 attP landing site flies, BDSC stock #9748. The Phi31C source was removed, and the mini-white positive progeny were used to established balanced lines. The UAS-STING allele was recombined with the park25 allele, and then assayed with PCR genotyping and verified with outcross and observation of the parkin homozygous phenotype. Wing muscle and thorax phenotyping Flies were anesthetized with CO2 and thoraces were examined under a dissecting microscope. Flies were scored for thorax shape and wing posture within the first 5 minutes of anesthesia. Blinding of genotypes to observer was performed when practical, including all of the initial assays involving the key genotypes in Fig 1. Geotaxis assays Male flies were collected at 0-1d post-eclosion and aged 5–7 days before testing. For testing, 15 to 20 flies were added to empty 10cm vials, labeled randomly, and a key was generated to preserve identity of tested stocks. The vials were placed in a plastic holder and the flies were manually tapped to the bottom of the vials. The flies were recorded for 20–30 seconds post disruption. Videos were scored using ImageJ to mark and annotate individual flies. Climbing Index was calculated as the percentage of flies in a vial that climbed greater than 6cm of the vial during the observed 20 seconds post disruption. Five independent trials, each with three technical repeats, were performed for a total of at least 75 total flies per genotype. Immunohistochemistry Flies (3–5 days old) were anesthetized with CO2 and thoraces were dissected away from the head and abdomen in cold phosphate buffered saline (PBS). The hemithoraces were bisected along the median plane, using a pair of microscissors (Fine Science Tools P/N 15006–09). Hemithoraces were fixing in 4% paraformaldehyde (Electron Microscopy Sciences P/N 15710) for 20 minutes at room temperature. After fixation, tissues were washed twice in PBS and then incubated twice for 10 minutes each in PBS with 0.1% TritonX (PBST). Tissues were then blocked in 5% goat serum diluted in PBST for 30 minutes, then incubated in AlexaFluor488-conjugated streptavadin (Jackson ImmunoResearch P/N 016-540-084) and iFluor647-conjugated phalloidin (Cayman Chemical Company P/N 20555) overnight at 4°C on a rotator. Samples were then washed three times with PBST and once with PBS. Thoraces and separated muscle pieces were then mounted directly on a 1.5 coverslip in Prolong Gold AntiFade (Thermo Fisher Scientific P/N P36930) media. Images of mitochondria morphology were acquired on a Zeiss LSM 880 Airyscan confocal with a 63X/1.4 objective Plan-Apochromat (Carl Zeiss) at 2X digital zoom and a 34-channel GAsP detector. Airyscan processing was performed in ZEN Black software (Zeiss). Images were analyzed in ImageJ and adjusted linearly for contrast and brightness. Western blotting For each genotype, 5 flies (3–5 days old) were anesthetized with CO2 and thoraces were isolated. Thoraces were homogenized in RIPA buffer supplemented with cOmplete EDTA-free Protease Inhibitor (Sigma-Aldrich P/N 04693159001) and PhosSTOP- phosphatase inhibitor tablet (Roche P/N 04906845001), and then samples were incubated for 10 minutes on ice. Samples were centrifuged at 10,000xG to remove tissue remains. Protein levels were quantified using the Pierce BCA protein assay kit (Thermo Fisher Scientific P/N 23228). The protein samples were normalized and then reduced by adding Lithium Dodecyl Sulfate sample buffer (GenScript P/N M00676-250) and 0.1M DTT then heating for 5 minutes at 99°C. Protein samples were separated on 4–12% SurePAGE, Bis-Tris gels (GenScript P/N M00654) and transferred to .45μM nitrocellulose membrane (BioRad P/N 1620115). Transfer efficiency and total protein amount was visualized using Ponceau S Staining Solution (Thermo Fisher Scientific P/N A40000279). Total protein images for each blot were acquired with ChemiDoc Imaging System (Bio-Rad Laboratories). Membranes were blocked for one hour with 3% milk or with 3% bovine serum albumin (Fisher Scientific P/N BP1600) (for pS65-Ub) in Tris-Buffered Saline with 0.1% Tween-20 (TBST), and probed overnight with the indicated antibody: anti-pUb(Ser65) (1:1000, rabbit polyclonal CST P/N 62802S), anti-a-Tubulin (1:4000, mouse mAb clone B-1-5-2, Millipore Sigma, P/N T5168), anti p62/ref(2)p (purified rabbit polyclonal Ab, a gift from Dr. Helmut Kramer, UT Southwestern Medical Center), ATP5α (1:4000, mouse mAb- Abcam P/N: ab14748) or V5-Tag D3H8Q (1:1000, Rabbit mAb Cell Signaling Technologies P/N 13202). After overnight incubation, membranes were washed 3 times with TBST and incubated for 1 hour with appropriate secondary antibodies diluted 1:10,000 in TBST+3% milk: HRP-coupled Donkey-anti-Rabbit or Sheep-anti-mouse (GE Healthcare Life Sciences), Goat Anti-Rabbit IgG IRDye 800CW-Conjugated (LI-COR Biosciences P/N 926–32211), or Goat Anti-Mouse IgG Antibody IRDye 680RD-Conjugated (LI-COR Biosciences P/N 926–68070). Blots were washed three times with TBST before imaging. For p65Ser-Ub detection, HRP-conjugated secondaries were incubated with SuperSignal Femto ECL (Thermo Scientific P/N 34095) for 3 minutes and imaged with ChemiDoc Imaging System (Bio-Rad Laboratories). All other primary antibodies were visualized with HRP-conjugated secondaries and incubation with AMersham ECL (Cytiva, P/N RPN2232). IR-conjugated secondaries were incubated simultaneously for one hour, and the blots were visualized on a LICOR Odyssey Fc multichannel imager after three washes with TBST. All images were processed, adjusted linearly for brightness and contrast, and analyzed for lane densitometry in ImageJ. Quantifications for pSer65-Ub, ATP5α, and p62/ref(2)p were first normalized to the intensity of the lane’s total stain. Concerning ATP5α: differences in optimal exposure times between replicas led to the necessity to represent data as percent of wild-type control in each experiment replica (S3D Fig). pSer65 Ub quantifications were divided by the mean relative amount of ATP5α for each genotype to approximate the amount of pSer65 per mitochondria (S3E Fig). RNA isolations and RT-qPCR RNA from 5 male flies were isolated using the Direct-zol RNA Miniprep kit (Zymo Research P/N R2050). Briefly, samples were homogenized in 300μL of Tri Reagant (Zymo Research P/N R2050-1-200) and then processed using the Zymo instructions for tissue samples. On-column DNAse treatments were performed before eluting the samples in 50μL of DPEC treated RNAse-free H2O, according to the Zymo Direct-zol kit protocol (DNaseI supplied with Direct-Zol kit). Sample were tested for quantity and purity with Nanodrop. 500ng of RNA samples were used for reverse transcriptase reactions, using the High-Capacity cDNA Reverse Transcription kit (Thermo Fisher Scientific P/N 4368814). cDNA samples were diluted 1:5 before using in qPCR reactions. qPCR was performed using indicated qPCR primers (primer sequences can be found in Table 1) with PowerUp SybrGreen Master Mix (Applied Biosystems P/N A25742) using a BioRad CFX384 Touch Real-Time PCR Thermocycler. Raw data was exported from BioRad Manager then analyzed using the ddCT method with Excel. CT values were normalized to the housekeeping gene rpl32 (also referred to as rp49) and then normalized to the wild-type control sample. Data from two to three independent biological replicas with three technical replicas per sample are presented. mtDNA copy number assays Total DNA was extracted from 10 male flies with the Quick DNA Miniprep Plus kit from Zymogen, according to the provided instructions. Quantification of mtDNA was performed using a multiplex TaqMan assays using validated probes against the mitochondrial gene mt:CoI and the nuclear-encoded gene rpL32 for reference [68]. Approx. 7ng of template DNA was used for each reaction. Primer details can be found in Table 1. qPCR reactions were performed on the BioRad 384CX system according to information provided for iTaqMan Supermix (BioRad P/N 1725130) with annealing temps at 60°C. Data was analyzed using the following method in Microsoft Excel: 1. mtDNA and nucDNA CT values were averaged from triplicate reactions. 2. Mitochondrial DNA content was normalized to nuclear DNA in each sample using the following equations: ΔCT = (nucDNA CT–mtDNA CT) then relative mitochondrial DNA content = 2 × 2^ΔCT. Replica biological samples were collected and isolated on separate days. Technical replicas were performed in each qPCR reaction run. Data is presented relative to the average wild-type (w1118) mitochondrial copy number for each biological set. For statistical analysis, a 1-way ANOVA and multiple comparison testing between each of the experimental genotypes. GST assays Age matched flies were collected and raised 4–5 days under standard conditions. For the assay, thoraces were dissected from ten flies per genotype/treatment/per replica were collected and immediately put on ice. The thorax samples were homogenized in GST assay sample buffer (100mM buffered potassium phosphate solution, pH 7.0, with 2mM EDTA). Samples were centrifuged at 10,000 x g for 15 minutes at 4°C and supernatants were assayed for protein concentration using a Pierce BCA protein assay kit (Thermo Fisher Scientific P/N 23228). Samples were normalized and diluted to a protein concentration of 2μg/μL, and Glutathione S-transferase activity was measured using a GST Assay Kit (Caymen Chemicals p/n 703302). After initiating the reactions, A340 was measured every minute for ten minutes. Rates of change were calculated from the plots of A340 vs. time, the blank well absorbance was subtracted from each, and then the activity rate (A340/min) was converted to estimated GST activity with the formula: The resulting estimated activities for each technical replica (3 per biological sample) were averaged together. For each biological replica, the provided purified GST enzyme was used as a positive control, and the resulting activity for each sample is represented as percent activity compared to the purified control. Graphed data represent normalized results from four repeated experiments. Apoptosis detection staining Apoptosis assays were performed on 3–4 day old flies, using the In-Situ Cell Death Detection Kit, TMR red (Roche P/N 12156792910), according to the manufacturer’s instructions. Briefly, fly thoraces were dissected and bisected in freshly prepared PBS. Hemi-thoraces were fixed for 20 minutes in 4% Paraformaldehyde in PBS, pH 7.4, freshly prepared. Tissues were washed first in PBS and then in permeabilization solution (0.1% Triton X100 in 0.1% sodium citrate, freshly prepared) for 15 minutes. Samples were incubated in the TUNEL detection solution in a humidified atmosphere for 60 min at 37°C in the dark. Tissues were then washed 3 times in PBST and blocked in 5% goat serum and 3% BSA diluted in PBST for 30 minutes, then incubated in iFluor647-conjugated phalloidin (Cayman Chemical Company P/N 20555) overnight at 4°C on a rotator. Samples were washed 3 times with PBST and once with PBS. Thoraces and separated muscle pieces were then mounted directly on a 1.5 coverslip in Prolong Gold AntiFade (Thermo Fisher Scientific P/N P36930) media. Images of thoraces were acquired on a Zeiss LSM 880 Airyscan confocal at 20X magnification. Quantification was performed using an ImageJ macro. In brief, stacks of 10 confocal sliced were used for max intensity projections. For each projected image, thresholding was applied to detect the phalloidin-labeled actin The muscle area was measured and the thresholded region was saved as a R.O.I. The same R.O.I. was used to count for the number of TUNEL-positive stained nuclei. The number of nuclei was then divided by the total area, to approximate the number of nuclei per μm2. TUNEL experiments were repeated two times, and the presented data represents biological replicas of at least 10 thoraces per genotype. RNA sequencing and analysis Flies were collected upon eclosion and aged 4 days in identical conditions, no more than 20 animals per vial. RNA from 10 male flies, per genotype and replica, were isolated using the Direct-zol RNA Miniprep kit (Zymo Research P/N R2050). Briefly, samples were homogenized in 300μL of Tri Reagant (Zymo Research P/N R2050-1-200) and then processed using the kit instructions for tissue samples. On-column DNAse treatments were performed before eluting the samples in 50μL of DPEC treated RNAse-free H2O. RNA-Seq library preparation and next generation sequencing RNA-Seq services were provided by Zymo Research Services, using their Total-RNA-Seq protocol. RNA quality was assessed with the Agilent TapeStation System. Total RNA-Seq libraries were constructed from 100ng of total RNA. rRNA depletion was performed according to standard protocol. Libraries were prepared using the Zymo-Seq RiboFree Total RNA Library Prep Kit (Cat # R3000) according to the manufacturer’s instructions. RNA-Seq libraries were sequenced on an Illumina NovaSeq to a sequencing depth of at least 30 million read pairs (150 bp paired-end sequencing) per sample. RNA-Seq data bioinformatics analysis The Zymo Research RNA-Seq pipeline was originally adapted from nf-core/rnaseq pipeline v1.4.2 (https://github.com/nf-core/rnaseq). The pipelines were built using Nextflow (https://www.nextflow.io/).2). Briefly, quality control of reads was carried out using FastQC v0.11.9 (http://www.bioinformatics.babraham.ac.uk/projects/fastqc). Adapter and low-quality sequences were trimmed from raw reads using Trim Galore! v0.6.6 (https://www.bioinformatics.babraham.ac.uk/projects/trim_galore). Trimmed reads were aligned to the reference genome using STAR v2.6.1d (https://github.com/alexdobin/STAR) [69]. BAM file filtering and indexing was carried out using SAMtools v1.9 (https://github.com/samtools/samtools) [70]. RNAseq library quality control was implemented using RSeQC v4.0.0 (http://rseqc.sourceforge.net/) and QualiMap v2.2.2-dev (http://qualimap.conesalab.org/)) [71,72]. Duplicate reads were marked using Picard tools v2.23.9 (http://broadinstitute.github.io/picard/). Library complexity was estimated using Preseq v2.0.3 (https://github.com/smithlabcode/preseq). Duplication rate quality control was performed using dupRadar v1.18.0 (https://bioconductor.org/packages/dupRadar/) [73]. Reads overlapping with exons were assigned to genes using featureCounts v2.0.1 (http://bioinf.wehi.edu.au/featureCounts/). Classification of rRNA genes/exons and their reads were based on annotations and RepeatMasker rRNA tracks from UCSC genome browser when applicable. Differential gene expression analysis was completed using DESeq2 v1.28.0 (https://bioconductor.org/packages/DESeq2/) [74]. Quality control and analysis results plots were visualized using MultiQC v1.9 (https://github.com/ewels/MultiQC) [75]. Further analysis and visualizations on the processed data were performed in R and Bioconductor. ClusterProfiler v.4.6 and Enrichplot v.1.19.0.01 were used for gene set enrichment analysis (GSEA) and plotting [76,77]. Heatmaps with normalized counts of highly enriched genes (absolute value of Log2 fold change greater than 3 and adjusted p.value less than 0.005) were generated with pHeatmap v.1.0.12. For all plots, ggplot2 v.3.4 and ggrepel v.0.9.2 were used for annotation. For ClusterProfiler GSEA analysis, cutoffs were: minGSSize = 50, maxGSSize = 250, and Benjamini-Hochberg adjusted p.value <0.05. Quantification and statistical analysis Quantitative data was recorded, transcribed, and maintained in Microsoft Excel. Data set descriptions, exploration, statistics, and graphing was performed in Graphpad Prism v.9.3. Detailed data sets and all statistical test details are provided in S1 Data File. Details including data descriptors, sample size (n), and specific statistical tests can be found in the figure legends. Proportions of fly populations were tested with the Wilson-Brown method to determine 95% confidence intervals. For categorical data, such as mitochondria morphology scores, a Fisher’s exact test was used to test for statistical significance between genotypes. Other quantitative data was assessed for normality using the Shapiro–Wilk test. For normally-distributed data, p-values were calculated using a one-way ANOVA test followed by Bonferroni’s or Dunnett’s multiple comparison tests. The Kruskal-Wallis test, with Dunn’s multiple comparisons test, was used for non-parametric data sets. For multiple comparison tests, significance between groups was determined as adjusted p-value less than 0.05. For all experiments, no prior sample size estimation was performed. Sample sizes were determined from previous studies. For all experiments, the collection of subjects (flies) in each genotype was randomized, and no inclusion/exclusion was performed. When practical and necessary, blinding of genotypes to observer was performed. All data quantification was done in a blind or automated manner. RStudio v.2022.07.2 running R v.4.2.2 was used for processing of RNA-seq data and generating the resulting plots. Details of the analysis pipeline are available in the previous description of RNA Seq Analysis. Generally, significance was determined after Benjamini-Hochberg correction and at a level of adjusted p.value< 0.05. For microscopy experiments, raw Airyscan confocal images were acquired and processed in Zen Black (Zeiss). Images were analyzed in FIJI/ImageJ2 v.2.3.0 [78] and quantification was finished in Microsoft Excel and graphed with Graphpad Prism. Images and figures were arranged in either Microsoft Powerpoint or Inkspace (https://inkscape.org). Resource availability All unique/stable reagents and animal stocks generated in this study are available from the lead contact and will be made available on request. Supporting information S1 Fig. Verification of parkin and sting mutant alleles. Related to Fig 1. https://doi.org/10.1371/journal.pgen.1010828.s001 (PDF) S2 Fig. Analysis and validation of an independent stingΔRG5;park25 stock. Related to Fig 1. https://doi.org/10.1371/journal.pgen.1010828.s002 (PDF) S3 Fig. Measurements of phosphorylated Ubiquitin and p62 from mutant thorax samples. Related to Fig 4. https://doi.org/10.1371/journal.pgen.1010828.s003 (PDF) S4 Fig. Analysis of Sting-regulated innate immunity in park mutants. Related to Fig 5. https://doi.org/10.1371/journal.pgen.1010828.s004 (PDF) S5 Fig. RNA-Seq experimental details and additional sample comparisons. Related to Fig 5. https://doi.org/10.1371/journal.pgen.1010828.s005 (PDF) S1 Table. All D. melanogaster genotypes, listed by figure. https://doi.org/10.1371/journal.pgen.1010828.s006 (DOCX) S1 Video. Geotaxis Assay Example—Used to quantify climbing activity in flies. Related to Fig 1. Genotypes from left to right (note: vials were randomly assigned and blinded): D: stingΔRG5; park1, C: stingΔRG5, B: park1, A: stingΔRG5; park25, E: w1118. https://doi.org/10.1371/journal.pgen.1010828.s007 (MOV) S1 Data File. Data for Figure Graphs.xlsx—Includes all plotted data and statistics. https://doi.org/10.1371/journal.pgen.1010828.s008 (XLSX) S2 Data File. Appended Results file for RNA Sequencing.xlsx—Related to Fig 5. https://doi.org/10.1371/journal.pgen.1010828.s009 (XLSX) Acknowledgments The authors thank Hong Xu (NHLBI, NIH), Rachel Cox (Uniformed Services University of the Health Sciences), and members of the Youle lab for feedback on the project and manuscript. Fly stocks were graciously provided by Alexander Whitworth (University of Cambridge), Leo Pallanck (University of Washington), Alicia Pickrell (Virginia Tech), Ed Giniger (NINDS, NIH), Jean-Luc Imler (Université de Strasbourg) and Alan Goodman (Washington State University). Dr. Helmut Krämer (UT Southwestern Medical Center) provided the polyclonal p62 antibody and valuable technical advice. Stocks obtained from the Bloomington Drosophila Stock Center (NIH P40OD018537) were used in this study. STING plasmids were obtained from the Drosophila Genomics Resource Center (NIH 2P40OD010949). The microscopy experiments were supported by the NINDS Intramural core Light Imaging Facility (LIF).

Duddy, Olivia P.;Silpe, Justin E.;Fei, Chenyi;Bassler, Bonnie L.

doi: 10.1371/journal.pgen.1010809pmid: 37523407

Zhao, Shuai;Mo, Li-Xiang;Li, Wen-Tong;Jiang, Lian-Li;Meng, Yi-Yuan;Ou, Jian-Feng;Liao, Lu-Sheng;Yan, Yu-Si;Luo, Xue-Mei;Feng, Jia-Xun

doi: 10.1371/journal.pgen.1010867pmid: 37523410

Introduction In nature, plant-polysaccharide-degrading enzymes (PPDE) produced by heterotrophic fungi can efficiently digest plant polysaccharides, including cellulose, xylan and starch, into monosaccharides, which provides a carbon source to support fungal growth and development [1]. However, regulation of PPDE production by fungi is very complex, remaining incompletely understood thus far. Transcription factors (TFs) are centrally involved in this regulation and function at the gene transcription level. The filamentous fungus Penicillium oxalicum is an excellent producer of PPDE that are used as biocatalysts for biorefining of renewable lignocellulosic biomass, to produce biologically-based chemicals, including biofuels. Several TFs have been identified as being involved in control of PPDE biosynthesis in P. oxalicum [1], including CxrA [2]. CxrA dynamically regulates the expression of major PPDE genes, such as cbh1, eg1, bgl1 and xyn11A, as well as regulatory genes, such as clrB, cxrC and cbh [3–5]. The minimal DNA binding domain, CxrAΔ1–16&59–733 was found to bind the core DNA sequences 5’-ATCAGATCCTCAAAGA-3’ and 5’-GCTGAGTCCTT-3’ in the promoters of cbh1 and clrB, respectively [4]. However, the mechanism of CxrA function remains to be fully elucidated, specifically its interaction partners and post-translational modification. Protein methylation is carried out by methyltransferases, which commonly methylate nitrogen atoms in the ε-amino group of lysine and the guanidino group of arginine, respectively, using S-adenosyl-L-methionine as cofactor [6]. Arginine methylation is implicated in fundamental cellular processes, including DNA transcription, splicing and repair, as well as cellular metabolism [7]. Arginine methylation is catalyzed by protein arginine methyltransferase (PRMT), which is generally classified into four types. Type I and II are responsible for biosynthesis of asymmetric and symmetric ω-NG,NG-dimethylarginine, respectively. Type III is responsible for biosynthesis of ω-NG monomethyl arginine only, and type IV for δ-NG monomethyl arginine; type IV is also specific to fungi [8]. In Aspergillus nidulans, Aspergillus flavus and Penicillium expansum, four PRMTs have been identified, i.e., PRMT1 (type I), PRMT3 (type I), PRMT5 (type II) and RMT2 (type IV). Of these, PRMT1, PRMT3 and PRMT5 are involved in fungal growth, development, stress responses and secondary metabolism [9–11]. However, the effects of these PRMTs on cellulase and xylanase biosynthesis in filamentous fungi have not been reported. In this study, the molecular mechanism of CxrA regulation was comprehensively elucidated. Notably, methylation of CxrA by PRMT was essential for the proper function in positively regulating the biosynthesis of cellulase and xylanase in P. oxalicum. Results N-terminal residues 1–206 are required for the proper function of full-length CxrA To identify the essential domain in CxrA, DNA sequences encoding a series of truncated CxrA peptides were introduced into the locus of gene pepA (POX_d05452) encoding an aspartic protease [12] in the P. oxalicum mutant ΔcxrA;G418R+, to generate the corresponding mutants, i.e., Δ61–733;bleR+, Δ207–733;bleR+, Δ592–733;bleR+, Δ1–16;bleR+, Δ1–60;bleR+ and Δ1–60&207–733;bleR+ (Fig 1A), and confirmed by PCR (S1 Fig) with specific primers (S1 Table). In the previous work, the mutant Δku70ΔpepA;G418R+ showed the same production of cellulase and xylanase relative to the Δku70;hphR+, as well as fungal growth on potato dextrose agar (PDA), suggesting that the pepA is not involved in the production of cellulase and xylanase, thereby being suitable for gene replacement by expression cassette. This also meant that the Δku70;hphR+ can represent the Δku70ΔpepA;G418R+ at least regarding the production of cellulase and xylanase [12]. When cultured in medium containing Avicel for 2–4 days after transfer from glucose, mutants Δ61–733;bleR+, Δ1–60;bleR+ and Δ1–60&207–733;bleR+ produced cellulase and xylanase at similar levels to ΔcxrA;G418R+, whereas production by mutant Δ592–733;bleR+ was comparable to that of the complementation strain CcxrA;bleR+. Mutant Δ207–733;bleR+ exhibited 25.8–26.4% reduced cellulase and xylanase production after 4 days compared with CcxrA;bleR+ (Fig 1B–1D). These results indicated that residues 1–591 of CxrA act like the wild-type CxrA for biosynthesis of cellulase and xylanase in P. oxalicum, whereas residues 1–206 were sufficient to obtain almost wild type-level CxrA activity (~75%). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 1. Influences of different regions (A–D) and arginine (R) 94 (E–I) of CxrA on cellulase and xylanase production in P. oxalicum. (A) Construction scheme of P. oxalicum mutants expressing sequences encoding different regions of CxrA. (B) Production of filter paper cellulase (FPase), carboxymethyl cellulase (CMCase) (C) and xylanase (D) by the constructed P. oxalicum mutants shown in panel A. (E) Effects of R94 on the production of FPase, CMCase (F) and xylanase (G). Mutant R94A;bleR+ contained a mutated CxrA in which R94 was mutated to alanine. (H) Effects of R94 on the CxrAΔ1–16 interacting with itself and full-length CxrA analyzed using the yeast two-hybrid system. In panels B–G, the uppercase and lowercase letters indicate p < 0.01 and p < 0.05, respectively. Different letters indicate significant differences accessed by one-way ANOVA. Each experiment was performed as three biological replicates. https://doi.org/10.1371/journal.pgen.1010867.g001 Mutant Δ1–16;bleR+ produced 18.7%–46.1% more cellulase and xylanase than CcxrA;bleR+, suggesting that the oligopeptide CxrAΔ17–733 suppressed production of cellulase and xylanase, especially on day 4 (Fig 1B–1D). The morphological phenotypes of the various P. oxalicum mutants on agar plates containing various carbon sources were examined. The results indicated that the tested strains exhibited more or less alteration as compared as the complementation strain CcxrA;bleR+. For example, mutant Δ1–60;bleR+ and Δ1–60&207–733;bleR+ showed different size and color on PDA plates. In addition, unlike the CcxrA;bleR+, Δ61–733;bleR+, Δ592–733;bleR+ and Δ1–16;bleR+ had less growth on Avicel (S2 Fig). Methylation of arginine (R) 94 modulates the biosynthesis of cellulase and xylanase and self-interaction of CxrAΔ1–60 To determine whether post-translational modification of CxrA occurred under Avicel induction, the overexpression strain OcxrA-his;G418R+, in which CxrA was tagged with 6×His at the N-terminus, was cultured for 24 h in the presence of Avicel, and total intracellular proteins were extracted. To construct the overexpression strain OcxrA-his;G418R+, as shown in S3A Fig, an overexpression cassette comprised of the coding sequence of cxrA with its own promoter, the G418 (antibiotic) resistance gene and both upstream and downstream flanking sequences of gene pepA encoding an aspartic protease [12], was introduced into the background strain Δku70;hphR+ via homologous recombination. In the OcxrA-his;G418R+, there are two copies of cxrA under their native promoters-one at the native locus and the other at the pepA locus (S3A Fig). The obtained transformants of OcxrA-his;G418R+ were confirmed by PCR (S3B Fig) with specific primers (S1 Table). Furthermore, the production of cellulase and xylanase by the overexpression strain OcxrA-his;G418R+ and background strain Δku70;hphR+ were compared when cultured in Avicel medium for 2–4 days after transfer from glucose. OcxrA-his;G418R+ showed increased production of all tested cellulases and xylanases by 28.5–68.5% (S3C–S3E Fig) compared with Δku70;hphR+. The expression of cxrA in OcxrA-his;G418R+ on Avicel was also determined; as expected, the transcriptional abundance of cxrA significantly increased in OcxrA-his;G418R+ over that in Δku70;hphR+, under Avicel induction for 24 h (S3F Fig). This indicates that cxrA was functionally overexpressed, thereby promoting the secretion of cellulase and xylanase on Avicel. Immunoprecipitation-mass spectrometry was employed to investigate the post-translational modification of CxrA and to detect modifications by acetylation at lysine (K) 30, methylation at R94 and R453, and phosphorylation at threonine (T) 443, T449, T456 and T457 (S4 Fig). As discussed above, the key peptide CxrAΔ207–733 contained the acetylated K30 and methylated R94. K30 is required for binding of CxrA to the promoter region of target genes, such as the cellobiohydrolase gene cbh1, through its DNA-binding domain CxrAΔ1–16Δ59–733 [4]. Therefore, the effects of R94 methylation were investigated on the production of cellulase and xylanase in P. oxalicum. The mutant R94A;bleR+ was constructed, in which the R was changed to alanine (A), and then confirmed by PCR (S1E Fig). When cultured in Avicel medium for 2−4 days after transfer from glucose, R94A;bleR+ had comparable cellulase and xylanase production to that of ΔcxrA;G418R+, and considerably lower than both the background strain Δku70;hphR+ and complementation strain CcxrA;bleR+ (Fig 1E–1G). To elucidate the function of R94 in the self-interaction of CxrA, yeast two-hybrid (Y2H) analysis was employed. Autoactivation experiments indicated that the full-length CxrA caused autoactivation in Saccharomyces cerevisiae (S5 Fig), whereas the peptide CxrAΔ1–60 could not (S6 Fig). Therefore, CxrAΔ1–60 was used as the bait for the Y2H assay, finding that CxrAΔ1–60 interacted with the whole CxrA and itself, whereas the mutated CxrAΔ1-60R94A lost the interaction ability (Fig 1H), suggesting that R94 is required for the self-interaction of CxrAΔ1–60. Nevertheless, it is possible that interaction of full-length CxrA and itself may be different from the results obtained using the nonfunctional truncation mutant CxrAΔ1–60 that may affect protein fold. Protein arginine N-methyltransferases PRMT2 and PRMT3 modulate the production of cellulase and xylanase in P. oxalicum To search for protein arginine N-methyltransferases able to methylate R94 of CxrA, the genomic database of P. oxalicum strain HP7-1 was screened, finding four annotated protein arginine N-methyltransferases, POX_b03080, POX_d05270, POX_f08428 and POX_e06662 [13]. Of these, three candidates were successfully deleted in P. oxalicum mutant Δku70;hphR+, to generate mutants ΔPOX_b03080;G418R+, ΔPOX_d05270;G418R+ and ΔPOX_e06662;G418R+ (S7 Fig). In addition, an overexpression strain, OPOX_f08428;G418R+, was constructed by replacing a protease gene pepA [12] in the background strain Δku70;hphR+ (S7 Fig), in which the transcription of the POX_f08428, was controlled by its own promoter. In the OPOX0_f08428;G418R+, there are two copies of POX_f08428 under their native promoters-one at the native locus and the other at the pepA locus (S7A Fig). When cultured on Avicel medium for 2–4 days after transfer from glucose, only OPOX_f08428;G418R+ and ΔPOX_b03080;G418R+ exhibited changes in cellulase and/or xylanase production (Figs 2, 3 and S8). For instance, compared with Δku70;hphR+, cellulase and xylanase production by OPOX_f08428;G418R+ increased by 41.4–95.1% at Day 4 (Fig 2A–2C) and that of ΔPOX_b03080;G418R+ increased by 67.0–149.7% (Fig 3A–3C). Furthermore, the expression of POX_f08428 under Avicel induction for 48 h, significantly increased (S9 Fig). The cellulase and xylanase production of complementation strain CPOX_b03080;bleR+ was restored almost to the levels of Δku70;hphR+ (Fig 3A–3C). This suggests that POX_f08428 and POX_b03080 positively and negatively modulated the production of cellulase and xylanase in P. oxalicum, respectively, so the two proteins POX_f08428 and POX_b03080 were selected for further study. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 2. Functional and sequence analysis of arginine N-methyltransferase PRMT2 (POX_f08428) from P. oxalicum. (A) Filter paper cellulase (FPase) production. (B) Carboxymethyl cellulase (CMCase) production. (C) Xylanase production. (D) Phylogenetic analysis of PRMT2. (E) Methyltransferase production. (F) Yeast two-hybrid assay of PRMT2 and CxrA interaction. (G) In vitro methylation assay of CxrAΔ1–60. (H) In vitro methylation assay of CxrAΔ1-60R94A. anti-MMA: mono methyl arginine antibody; anti-ADMA: asymmetric dimethyl arginine antibody; SAM: S-adenosyl-methionine. “+” and “−” indicate the presence or absence of the test protein. In panels A-C, and E, data values indicate means ± standard deviation. **p <0.01 and *p <0.05 indicate significant differences between the mutant and Δku70, calculated by Student’s t-test. In panel D, the phylogenetic trees were constructed by MEGA version 11, using the neighbor-joining method and Poisson model. Bootstrap values shown on the branches were calculated with 1000 replicates. Afi: Aspergillus fischeri; Afu: Aspergillus fumigatus; Ano: Aspergillus novofumigatus; Acl: Aspergillus clavatus; Ani: Aspergillus nidulans; Aor: Aspergillus oryzae; Anig: Aspergillus niger; Tma: Talaromyces marneffei; Tst: Talaromyces stipitatus; Pox: P. oxalicum; Psu: Penicillium subrubescens; Pdi: Penicillium digitatum; Pex: Penicillium expansum; Pru: Penicillium rubens Wisconsin; Teq: Trichophyton equinum; Bgi: Blastomyces gilchristii; Pbr: Paracoccidioides brasiliensis; Epa: Emergomyces pasteurianus; Cbe: Cucurbitaria berberidis; Ncr: Neurospora crassa; Fox: Fusarium oxysporum; Tre: Trichoderma reesei; Tha: Trichoderma harzianum; Sce: Saccharomyces cerevisiae. https://doi.org/10.1371/journal.pgen.1010867.g002 Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 3. Functional and sequence analysis of arginine N-methyltransferase PRMT3 (POX_d03080) from P. oxalicum. (A) Filter paper cellulase (FPase) production. (B) Carboxymethyl cellulase (CMCase) production. (C) Xylanase production. (D) Evolutionary analysis of PRMT3. (E) Methyltransferase production. (F) Yeast two-hybrid assay of PRMT3 and CxrAΔ1–60 interaction. (G) Subcellular localization of PRMT3 in P. oxalicum. In panels A-C, and E, data values indicate means ± standard deviation. The uppercase and lowercase letters indicate p < 0.01 and p < 0.05, respectively. Different letters indicate significant differences calculated by one-way ANOVA. In panel D, the phylogenetic trees were constructed by MEGA version 11, using the neighbor-joining method and Poisson model. Bootstrap values shown on the branches were calculated with 1000 replicates. Afl: Aspergillus flavus; Afu: Aspergillus fumigatus; Acl: Aspergillus clavatus; Ani: Aspergillus nidulans; Aor: Aspergillus oryzae; Anig: Aspergillus niger; Ave: Aspergillus versicolor; Tma: Talaromyces marneffei; Pox: P. oxalicum; Pdi: Penicillium digitatum; Pru: Penicillium rubens Wisconsin; Ncr: Neurospora crassa; Tce: Talaromyces cellulolyticus; Tre: Trichoderma reesei; Tha: Trichoderma harzianum; Sce: Saccharomyces cerevisiae; Spo: Schizosaccharomyces pombe; Has: Homo sapiens. In panel G, red arrows represent GFP-PRMT3. The expression of prmt3-gfp was driven by its own promotor, Pprmt3, in the overexpression strain Oprmt3-gfp;G418R+. DAPI: 4,6-Diamidino-2-phenylindole. GFP: Green fluorescent protein. Scale bar = 20 μm. https://doi.org/10.1371/journal.pgen.1010867.g003 POX_f08428 contained 429 amino acid residues, and an ankyrin repeat domain (IPR002110, residues 61−109). POX_f08428 shared 100%, 66.67% and 34.83% identity, respectively, with PDE_04339 (EPS29390.1) from P. oxalicum strain 114–2, arginine N-methyltransferase 2 (XP_755242.1) from Aspergillus fumigatus strain Af293 and protein-arginine N5-methyltransferase (NP_010753.1) from S. cerevisiae S288C. POX_f08428 is conserved in eukaryotes and most closely related to its homologs in Aspergillus (Fig 2D). For convenience, POX_ f08428 was renamed PRMT2. POX_b03080 contained 546 amino acid residues and a methyltransfer_25 domain (IPR041698, residues 249 to 346). POX_b03080 shared 99.63%, 40.16% and 36.28% identity, respectively, with PDE_04847 (EPS29897.1) from P. oxalicum strain 114–2, ribosomal protein arginine N-methyltransferase Rmt3 (NP_595572.1) from S. pombe 972h- and protein arginine N-methyltransferase 3 (AAC39837.1) from H. sapiens. POX_b03080 was conserved in eukaryotes and most closely related to its homologs in Aspergillus (Fig 3D). For convenience, POX_b03080 was renamed PRMT3. To investigate the contribution of PRMT2 and PRMT3 to methyltransferase activity in P. oxalicum, methyltransferase production by overexpression strain Oprmt2;G418R+ and mutant Δprmt3;G418R+ was measured after culture in Avicel medium for 24 h. Methyltransferase production by Oprmt2;G418R+ significantly increased (Fig 2E), whereas that of Δprmt3;G418R+ significantly decreased, compared with Δku70;hphR+ and Cprmt3;bleR+, and there was no significant difference between Δku70;hphR+ and Cprmt3;bleR+ (Fig 3E), implying that PRMT2 and PRMT3 have methyltransferase activity. PRMT2 is responsible for R94-methylation of CxrAΔ1–60 To determin which of PRMT2 and PRMT3 responsible for methylation of R94 in CxrA, the Y2H assay was firstly employed. The results indicated that Y2HGold cells carrying BD-PRMT2 and AD-CxrA (Figs 2F and S10A), as well as BD-CxrAΔ1–60 and AD-PRMT3, grew well on SD/-Leu/-Trp/-His/-Ade agar plates (Fig 3F), indicating that the PRMT2 and PRMT3 interact with CxrA and CxrAΔ1–60, respectively. However, as unexpected, yeast cells carrying BD-PRMT3 and AD-CxrA were unable to growth on SD/-Leu/-Trp/-His/-Ade agar plates, suggesting that PRMT3 could not interact with the full-length CxrA (S10B Fig). Therefore, this difference maybe result from the use of a nonfunctional CxrAΔ1–60 lacking the DNA binding domain. CxrAΔ1–60 may affect protein folding. Furthermore, in vitro methylation experiments were performed. As unexpected, DNA sequence encoding a full-length CxrA fused with a glutathione-S-transferase (GST) tag failed to be recombinantly expressed in E. coli BL21. Therefore, the DNA fragments encoding CxrAΔ1–60 fused with a GST tag had to be expressed in E. coli BL21, as well as PRMT2/PRMT3 tagged with His, then the recombinant GST-CxrAΔ1–60 and His-PRMT2 or His-PRMT3 were purified. After reaction of GST-CxrAΔ1–60 and His-PRMT3 or His-PRMT2, with S-adenosyl-L-methionine as methyl donor, Western blotting was used to detect methylated CxrAΔ1–60 using specific antibodies including anti-mono methyl arginine, anti-symmetric di-methyl arginine, and anti-asymmetric di-methyl arginine. Bands corresponding to methylated GST-CxrAΔ1–60 appeared when recombinant His-PRMT2 and GST-CxrAΔ1–60 were treated with antibodies anti-mono methyl arginine and anti-asymmetric di-methyl arginine (Fig 2G), respectively, whereas no band was found for the mutated CxrAΔ1-60R94A (Fig 2H), showing that PRMT2 both mono-methylated and asymmetrically di-methylated CxrAΔ1–60 at R94. In contrast, the band corresponding to methylated GST-CxrAΔ1–60 did not appear when recombinant His-PRMT3 and GST-CxrAΔ1–60 were treated with specific antibodies (S11 Fig), indicating that PRMT3 is not responsible for CxrAΔ1–60 methylation. Nevertheless, it is possible that in vitro methylation modification of the full-length CxrA by PRMT2 or PRMT3 may be different from the results obtained using the nonfunctional truncation mutant CxrAΔ1–60 that may affect protein fold. PRMT3 appears to be a nuclear arginine N-methyltransferase To investigate the subcellular localization of PRMT3 in P. oxalicum hyphae, the overexpression strain Oprmt3-gfp;G418R+, carrying a GFP reporter was constructed, in which the fused gene prmt3-gfp was controlled by its own promoter, Pprmt3 (S7 Fig). In the Oprmt3-gfp;G418R+, there are two copies of prmt3 under their native promoters-one at the native locus and the other at the pepA locus (S7D Fig). After culture in Avicel medium for 24 h, the hyphae of Oprmt3-gfp;G418R+ and background strain Δku70;hphR+ were observed under a fluorescence microscope. The green fluorescence of the fusion protein, PRMT3-GFP, in Oprmt3-gfp;G418R+ appeared to overlap with blue fluorescence signals from 4,6-diamidino-2-phenylindole (DAPI), which specifically stains the nucleus, whereas there was no green fluorescence from Δku70;hphR+ (Fig 3G). PRMT3 mediates the expression of genes related to cellulase and xylanase biosynthesis by P. oxalicum To elucidate the effects of PRMT3 on the expression of genes related to cellulase and xylanase production by P. oxalicum, RNA-sequencing and a real-time quantitative reverse transcription PCR (RT-qPCR) assay were employed. Both Δku70;hphR+ and Δprmt3;G418R+ were cultured in Avicel medium for 24 h, after transfer from glucose, and their total RNA was extracted for RNA-sequencing. The sequencing data from three biological replicates were analyzed statistically and generated a high Pearson correlation coefficient (> 0.97) (S12 Fig), suggesting that the transcriptomic data were suitable for the subsequent analysis. With thresholds of |log2 fold change| > 1 and false discovery rate (FDR) < 0.05, there were 997 differentially expressed genes (DEGs) detected in Δprmt3;G418R+ compared with Δku70;hphR+, consisting of 321 up-regulated and 676 down-regulated genes (Fig 4A and S1 Dataset). The Kyoto Encyclopedia of Genes and Genomes (KEGG) annotation showed that these DEGs were mainly involved in metabolism, especially carbohydrate metabolism (75 genes) and amino acid metabolism (58 genes) (Fig 4B). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 4. Comparative analyses of transcriptomes from P. oxalicum mutant Δprmt3;G418R+ and background strain Δku70;hphR+ (A-E) and real-time reverse transcription quantitative PCR (RT-qPCR) confirmation (F). (A) Volcano plot of differentially expressed genes (DEGs). DEGs were selected with thresholds of |log2 fold change| > 1.0 and p-value < 0.05. (B) KEGG annotations of the DEGs modulated by PRMT3. (C) DEGs encoding cellulases and xylanases. (D) DEGs encoding putative transcription factors. (E) DEGs encoding putative sugar transporters. In panels D and E, the three columns of data corresponding to each fungal strain represent the three biological replicates for each strain. (F) The transcriptional levels of genes encoding major cellulases including two cellobiohydrolase genes cbh1 and cbh2, and an endo-β-1,4-glucanase gene eg1 in the P. oxalicum mutant Δprmt3;G418R+ relative to the background strain Δku70;hphR+. In panel F, the P. oxalicum strains were cultured for 4–48 h in the presence of Avicel. Gene expression in Δprmt3;G418R+ was normalized to the level of Δku70;hphR+. Data points show mean ± standard deviation. **p <0.01 and *p <0.05 indicate significant differences between Δprmt3;G418R+ and Δku70;hphR+, calculated by Student’s t-test. https://doi.org/10.1371/journal.pgen.1010867.g004 Among the 997 DEGs, nine cellulase/xylanase genes were found, namely one endo-β-1,4-glucanase gene (POX_d04883/eg1), five β-glucosidase genes (bgls, POX_a00284, POX_c04019, POX_e06772/bgl1, POX_e06434 and POX_f08346) and three endo-β-1,4-xylanase genes (POX_a01871/xyn11A, POX_c03490 and POX_f07706). Of these, POX_b02315/cel61A, POX_e06772/bgl1 and POX_e06434 were down-regulated (-2.4 < log2 fold change < -1.4), and the others were up-regulated (1.0 < log2 fold change < 1.6) in Δprmt3;G418R+ compared with Δku70;hphR+ (Fig 4C). The effects of PRMT3 on the expression of genes encoding putative TFs was then investigated by comparative analysis of transcriptomes, which identified 30 TF-encoding DEGs. The transcript abundances of 12 of them increased by 2.03–52.6-fold in Δprmt3;G418R+, whereas the other 18 genes decreased by 1.2–499-fold. Most of these TFs contained zinc-finger domains (i.e., six C2H2-type, 11 Zn2Cys6-type and one CCHC-type). Notably, expression of a key transcription repressor gene, cxrC in Δprmt3;G418R+ increased 2.46-fold, whereas transcripts of sporulation-regulated abaA, brlA and flbD decreased by 59.4–95.9% (Fig 4D). In addition, 25 DEGs encoding predicted sugar transporters were found, of which 13 were up-regulated (1.0 < log2 fold change < 2.9) and 12 down-regulated (-4.2 < log2 fold change < -1.3) in Δprmt3;G418R+. Of these, two major cellodextrin transporter genes, cdtC and cdtD were up-regulated 2.14- and 2.30-fold, respectively (Fig 4E). Three cellulase genes, cbh1, cbh2, and eg1 were subjected to RT-qPCR analysis of expression variation with induction duration and confirmation of RNA-sequencing data. The expression of eg1 after 24 h of induction was up-regulated in Δprmt3;G418R+ compared with Δku70;hphR+, whereas cbh1 and cbh2 showed no significant change, in agreement with the RNA-sequencing results. Transcriptional abundances of all three genes decreased up to 12 h, except for eg1 at 12 h, which markedly increased. At 48 h, the expression of cbh1 was up-regulated, whereas eg1 was down-regulated. The transcriptional level of cbh2 did not change (Fig 4F). PRMT3 was required for positive regulation of cellulase and xylanase production by cxrA As described above, PRMT3 was unable to methylate CxrA, but deletion of prmt3 from P. oxalicum increased cellulase and xylanase production by 67.0–149.7%, compared with the background strain Δku70;hphR+ after 4 days of Avicel induction (Fig 3). To investigate the effects of PRMT3 on CxrA action, the cxrA-overexpression strain OcxrA-gfp;G418R+ and mutant Δprmt3;cxrA-gfp; G418R+;bleR+ was sequentially constructed and confirmed by PCR (S7 Fig). In OcxrA-gfp;G418R+, the fusion gene cxrA-gfp was controlled by the predicted natural promoter of cxrA. In the OcxrA-gfp;G418R+ and Δprmt3;cxrA-gfp;G418R+;bleR+, there are two copies of cxrA under their native promoters- one at the native locus and the other at the pepA locus (S7E and S7F Fig). RT-qPCR confirmed that the expression of cxrA in OcxrA-gfp;G418R+ significantly increased, compared with that in Δku70;hphR+ after 48 h of Avicel induction (S9 Fig). Measurement of enzymatic activity revealed that cellulase and xylanase production by Δprmt3;cxrA-gfp;G418R+;bleR+ was 11.3–62.8% lower than that of the overexpression strain OcxrA-gfp;G418R+, when cultured in Avicel medium for 2–4 d (Fig 5A–5C), indicating that PRMT3 is required for positive regulation of cellulase and xylanase production by cxrA. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 5. Effects of PRMT3 on the regulatory functions of CxrA. (A) Filter paper cellulase (FPase) production. (B) Carboxymethyl cellulase (CMCase) production. (C) Xylanase production. (D) Subcellular localization of CxrA in P. oxalicum. In panels A-C, the fungal strains were pre-grown in glucose medium for 24 h, then transferred into Avicel medium for 2–4 d. The uppercase and lowercase letters indicate p < 0.01 and p < 0.05, respectively. Different letters indicate significant differences assessed by one-way ANOVA. All tested samples were normally distributed. In panel D, 4,6-diamidino-2-phenylindole (DAPI) was employed to stain mycelial nuclei. Subcellular localization of CxrA was determined with GFP signals observed by a fluorescence microscope. Scale bar = 20 μm. https://doi.org/10.1371/journal.pgen.1010867.g005 PRMT3 apparently assists the entry of CxrA into the nucleus To further elucidate the interactions between CxrA and PRMT3, Δprmt3;cxrA-gfp;G418R+;bleR+ and OcxrA-gfp;G418R+ were cultured in Avicel medium for 24 h, then their hyphae were collected for microscopic examination. The green fluorescent signals from CxrA-GFP in OcxrA-gfp;G418R+ appeared to overlap with the blue fluorescence signals from DAPI, indicating that CxrA was localized in the nucleus, whereas the green fluorescence signals from CxrA-GFP in Δprmt3;cxrA-gfp;G418R+;bleR+ were evenly distributed in the whole hypha. This seems that deletion of prmt3 partially obstructed nuclear translocation of CxrA protein, in other words, it appears that PRMT3 assists the entry of CxrA into the nucleus (Fig 5D). Co-expression analysis of genes regulated by CxrA and PRMT3 Co-expression analysis of genes regulated by PRMT3 and CxrA was performed by RNA-sequencing of Δprmt3;G418R+ and ΔcxrA;G418R+, compared with Δku70;hphR+ cultured in Avicel medium for 24 h, after transfer from glucose. The resulting data from three biological replicates had a high Pearson’s correlation coefficient (> 0.97) (S12 Fig). With thresholds of |log2 fold change| > 1 and FDR < 0.05, there were 2,552 DEGs in ΔcxrA;G418R+ compared with Δku70;hphR+, of which 1,253 were up-regulated and 1,299 down-regulated (S2 Dataset). These included 27 key cellulase and xylanase genes, 128 putative TF-encoding genes and 28 putative sugar transporter-encoding genes (S2 Dataset), in agreement with a previous report [4]. Comparative analysis identified 657 DEGs co-regulated by PRMT3 and CxrA under Avicel induction, of which 113 were up-regulated and 448 down-regulated in Δprmt3;G418R+ and ΔcxrA;G418R+, relative to Δku70;hphR+ (Fig 6A). These co-regulated genes were mainly involved in metabolism, especially carbohydrate and amino acid metabolism (Fig 6B). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 6. Co-regulation between CxrA and PRMT3 in P. oxalicum. (A) Numbers of differentially expressed genes (DEGs) in regulons between cxrA and prmt3. DEG regulons were determined by comparison of transcriptomic data from Δprmt3;G418R+ and ΔcxrA;G418R+ with that of the background strain Δku70;hphR+. “Up” and “Down” indicate up- and down-regulation. (B) KEGG annotations of the 657 genes co-regulated by cxrA and prmt3. (C) Heatmap illustrating the expression of DEGs encoding plant-cell-wall-degrading enzymes. (D) Heatmap showing the expression of DEGs encoding putative transcription factors. (E) Heatmap displaying the expression of DEGs encoding sugar transporters. In panels C and D, the three columns of data corresponding to each fungal strain represent the three biological replicates for each strain. https://doi.org/10.1371/journal.pgen.1010867.g006 The co-regulated DEG set included four BGL genes (bgl1, POX_a00284, POX_f08346 and POX_e06434) and two xylanase genes (xyn11A and POX_c03490). The mRNA/transcription levels of these genes were up-regulated in Δprmt3;G418R+, but down-regulated in ΔcxrA;G418R+, compared with Δku70;hphR+, except for bgl1 and POX_e06434, which were down-regulated in both mutants (Fig 6C). The expression of most co-regulated DEGs encoding TFs and sugar transporters decreased in Δprmt3;G418R+ and ΔcxrA;G418R+, compared with Δku70;hphR+ (Fig 6D and 6E), however, the cellodextrin transporter genes cdtC and cdtD were downregulated in ΔcxrA;G418R+ and upregulated in Δprmt3;G418R+, compared with Δku70;hphR+ (Fig 6E). Notably, gene prmt2 was included in the CxrA regulon, rather than in the prmt3 regulon. In ΔcxrA;G418R+, the expression of prmt2 increased by 1.2-fold, compared with Δku70;hphR+ (S2 Dataset), indicating that CxrA inhibits the expression of prmt2 on Avicel. N-terminal residues 1–206 are required for the proper function of full-length CxrA To identify the essential domain in CxrA, DNA sequences encoding a series of truncated CxrA peptides were introduced into the locus of gene pepA (POX_d05452) encoding an aspartic protease [12] in the P. oxalicum mutant ΔcxrA;G418R+, to generate the corresponding mutants, i.e., Δ61–733;bleR+, Δ207–733;bleR+, Δ592–733;bleR+, Δ1–16;bleR+, Δ1–60;bleR+ and Δ1–60&207–733;bleR+ (Fig 1A), and confirmed by PCR (S1 Fig) with specific primers (S1 Table). In the previous work, the mutant Δku70ΔpepA;G418R+ showed the same production of cellulase and xylanase relative to the Δku70;hphR+, as well as fungal growth on potato dextrose agar (PDA), suggesting that the pepA is not involved in the production of cellulase and xylanase, thereby being suitable for gene replacement by expression cassette. This also meant that the Δku70;hphR+ can represent the Δku70ΔpepA;G418R+ at least regarding the production of cellulase and xylanase [12]. When cultured in medium containing Avicel for 2–4 days after transfer from glucose, mutants Δ61–733;bleR+, Δ1–60;bleR+ and Δ1–60&207–733;bleR+ produced cellulase and xylanase at similar levels to ΔcxrA;G418R+, whereas production by mutant Δ592–733;bleR+ was comparable to that of the complementation strain CcxrA;bleR+. Mutant Δ207–733;bleR+ exhibited 25.8–26.4% reduced cellulase and xylanase production after 4 days compared with CcxrA;bleR+ (Fig 1B–1D). These results indicated that residues 1–591 of CxrA act like the wild-type CxrA for biosynthesis of cellulase and xylanase in P. oxalicum, whereas residues 1–206 were sufficient to obtain almost wild type-level CxrA activity (~75%). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 1. Influences of different regions (A–D) and arginine (R) 94 (E–I) of CxrA on cellulase and xylanase production in P. oxalicum. (A) Construction scheme of P. oxalicum mutants expressing sequences encoding different regions of CxrA. (B) Production of filter paper cellulase (FPase), carboxymethyl cellulase (CMCase) (C) and xylanase (D) by the constructed P. oxalicum mutants shown in panel A. (E) Effects of R94 on the production of FPase, CMCase (F) and xylanase (G). Mutant R94A;bleR+ contained a mutated CxrA in which R94 was mutated to alanine. (H) Effects of R94 on the CxrAΔ1–16 interacting with itself and full-length CxrA analyzed using the yeast two-hybrid system. In panels B–G, the uppercase and lowercase letters indicate p < 0.01 and p < 0.05, respectively. Different letters indicate significant differences accessed by one-way ANOVA. Each experiment was performed as three biological replicates. https://doi.org/10.1371/journal.pgen.1010867.g001 Mutant Δ1–16;bleR+ produced 18.7%–46.1% more cellulase and xylanase than CcxrA;bleR+, suggesting that the oligopeptide CxrAΔ17–733 suppressed production of cellulase and xylanase, especially on day 4 (Fig 1B–1D). The morphological phenotypes of the various P. oxalicum mutants on agar plates containing various carbon sources were examined. The results indicated that the tested strains exhibited more or less alteration as compared as the complementation strain CcxrA;bleR+. For example, mutant Δ1–60;bleR+ and Δ1–60&207–733;bleR+ showed different size and color on PDA plates. In addition, unlike the CcxrA;bleR+, Δ61–733;bleR+, Δ592–733;bleR+ and Δ1–16;bleR+ had less growth on Avicel (S2 Fig). Methylation of arginine (R) 94 modulates the biosynthesis of cellulase and xylanase and self-interaction of CxrAΔ1–60 To determine whether post-translational modification of CxrA occurred under Avicel induction, the overexpression strain OcxrA-his;G418R+, in which CxrA was tagged with 6×His at the N-terminus, was cultured for 24 h in the presence of Avicel, and total intracellular proteins were extracted. To construct the overexpression strain OcxrA-his;G418R+, as shown in S3A Fig, an overexpression cassette comprised of the coding sequence of cxrA with its own promoter, the G418 (antibiotic) resistance gene and both upstream and downstream flanking sequences of gene pepA encoding an aspartic protease [12], was introduced into the background strain Δku70;hphR+ via homologous recombination. In the OcxrA-his;G418R+, there are two copies of cxrA under their native promoters-one at the native locus and the other at the pepA locus (S3A Fig). The obtained transformants of OcxrA-his;G418R+ were confirmed by PCR (S3B Fig) with specific primers (S1 Table). Furthermore, the production of cellulase and xylanase by the overexpression strain OcxrA-his;G418R+ and background strain Δku70;hphR+ were compared when cultured in Avicel medium for 2–4 days after transfer from glucose. OcxrA-his;G418R+ showed increased production of all tested cellulases and xylanases by 28.5–68.5% (S3C–S3E Fig) compared with Δku70;hphR+. The expression of cxrA in OcxrA-his;G418R+ on Avicel was also determined; as expected, the transcriptional abundance of cxrA significantly increased in OcxrA-his;G418R+ over that in Δku70;hphR+, under Avicel induction for 24 h (S3F Fig). This indicates that cxrA was functionally overexpressed, thereby promoting the secretion of cellulase and xylanase on Avicel. Immunoprecipitation-mass spectrometry was employed to investigate the post-translational modification of CxrA and to detect modifications by acetylation at lysine (K) 30, methylation at R94 and R453, and phosphorylation at threonine (T) 443, T449, T456 and T457 (S4 Fig). As discussed above, the key peptide CxrAΔ207–733 contained the acetylated K30 and methylated R94. K30 is required for binding of CxrA to the promoter region of target genes, such as the cellobiohydrolase gene cbh1, through its DNA-binding domain CxrAΔ1–16Δ59–733 [4]. Therefore, the effects of R94 methylation were investigated on the production of cellulase and xylanase in P. oxalicum. The mutant R94A;bleR+ was constructed, in which the R was changed to alanine (A), and then confirmed by PCR (S1E Fig). When cultured in Avicel medium for 2−4 days after transfer from glucose, R94A;bleR+ had comparable cellulase and xylanase production to that of ΔcxrA;G418R+, and considerably lower than both the background strain Δku70;hphR+ and complementation strain CcxrA;bleR+ (Fig 1E–1G). To elucidate the function of R94 in the self-interaction of CxrA, yeast two-hybrid (Y2H) analysis was employed. Autoactivation experiments indicated that the full-length CxrA caused autoactivation in Saccharomyces cerevisiae (S5 Fig), whereas the peptide CxrAΔ1–60 could not (S6 Fig). Therefore, CxrAΔ1–60 was used as the bait for the Y2H assay, finding that CxrAΔ1–60 interacted with the whole CxrA and itself, whereas the mutated CxrAΔ1-60R94A lost the interaction ability (Fig 1H), suggesting that R94 is required for the self-interaction of CxrAΔ1–60. Nevertheless, it is possible that interaction of full-length CxrA and itself may be different from the results obtained using the nonfunctional truncation mutant CxrAΔ1–60 that may affect protein fold. Protein arginine N-methyltransferases PRMT2 and PRMT3 modulate the production of cellulase and xylanase in P. oxalicum To search for protein arginine N-methyltransferases able to methylate R94 of CxrA, the genomic database of P. oxalicum strain HP7-1 was screened, finding four annotated protein arginine N-methyltransferases, POX_b03080, POX_d05270, POX_f08428 and POX_e06662 [13]. Of these, three candidates were successfully deleted in P. oxalicum mutant Δku70;hphR+, to generate mutants ΔPOX_b03080;G418R+, ΔPOX_d05270;G418R+ and ΔPOX_e06662;G418R+ (S7 Fig). In addition, an overexpression strain, OPOX_f08428;G418R+, was constructed by replacing a protease gene pepA [12] in the background strain Δku70;hphR+ (S7 Fig), in which the transcription of the POX_f08428, was controlled by its own promoter. In the OPOX0_f08428;G418R+, there are two copies of POX_f08428 under their native promoters-one at the native locus and the other at the pepA locus (S7A Fig). When cultured on Avicel medium for 2–4 days after transfer from glucose, only OPOX_f08428;G418R+ and ΔPOX_b03080;G418R+ exhibited changes in cellulase and/or xylanase production (Figs 2, 3 and S8). For instance, compared with Δku70;hphR+, cellulase and xylanase production by OPOX_f08428;G418R+ increased by 41.4–95.1% at Day 4 (Fig 2A–2C) and that of ΔPOX_b03080;G418R+ increased by 67.0–149.7% (Fig 3A–3C). Furthermore, the expression of POX_f08428 under Avicel induction for 48 h, significantly increased (S9 Fig). The cellulase and xylanase production of complementation strain CPOX_b03080;bleR+ was restored almost to the levels of Δku70;hphR+ (Fig 3A–3C). This suggests that POX_f08428 and POX_b03080 positively and negatively modulated the production of cellulase and xylanase in P. oxalicum, respectively, so the two proteins POX_f08428 and POX_b03080 were selected for further study. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 2. Functional and sequence analysis of arginine N-methyltransferase PRMT2 (POX_f08428) from P. oxalicum. (A) Filter paper cellulase (FPase) production. (B) Carboxymethyl cellulase (CMCase) production. (C) Xylanase production. (D) Phylogenetic analysis of PRMT2. (E) Methyltransferase production. (F) Yeast two-hybrid assay of PRMT2 and CxrA interaction. (G) In vitro methylation assay of CxrAΔ1–60. (H) In vitro methylation assay of CxrAΔ1-60R94A. anti-MMA: mono methyl arginine antibody; anti-ADMA: asymmetric dimethyl arginine antibody; SAM: S-adenosyl-methionine. “+” and “−” indicate the presence or absence of the test protein. In panels A-C, and E, data values indicate means ± standard deviation. **p <0.01 and *p <0.05 indicate significant differences between the mutant and Δku70, calculated by Student’s t-test. In panel D, the phylogenetic trees were constructed by MEGA version 11, using the neighbor-joining method and Poisson model. Bootstrap values shown on the branches were calculated with 1000 replicates. Afi: Aspergillus fischeri; Afu: Aspergillus fumigatus; Ano: Aspergillus novofumigatus; Acl: Aspergillus clavatus; Ani: Aspergillus nidulans; Aor: Aspergillus oryzae; Anig: Aspergillus niger; Tma: Talaromyces marneffei; Tst: Talaromyces stipitatus; Pox: P. oxalicum; Psu: Penicillium subrubescens; Pdi: Penicillium digitatum; Pex: Penicillium expansum; Pru: Penicillium rubens Wisconsin; Teq: Trichophyton equinum; Bgi: Blastomyces gilchristii; Pbr: Paracoccidioides brasiliensis; Epa: Emergomyces pasteurianus; Cbe: Cucurbitaria berberidis; Ncr: Neurospora crassa; Fox: Fusarium oxysporum; Tre: Trichoderma reesei; Tha: Trichoderma harzianum; Sce: Saccharomyces cerevisiae. https://doi.org/10.1371/journal.pgen.1010867.g002 Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 3. Functional and sequence analysis of arginine N-methyltransferase PRMT3 (POX_d03080) from P. oxalicum. (A) Filter paper cellulase (FPase) production. (B) Carboxymethyl cellulase (CMCase) production. (C) Xylanase production. (D) Evolutionary analysis of PRMT3. (E) Methyltransferase production. (F) Yeast two-hybrid assay of PRMT3 and CxrAΔ1–60 interaction. (G) Subcellular localization of PRMT3 in P. oxalicum. In panels A-C, and E, data values indicate means ± standard deviation. The uppercase and lowercase letters indicate p < 0.01 and p < 0.05, respectively. Different letters indicate significant differences calculated by one-way ANOVA. In panel D, the phylogenetic trees were constructed by MEGA version 11, using the neighbor-joining method and Poisson model. Bootstrap values shown on the branches were calculated with 1000 replicates. Afl: Aspergillus flavus; Afu: Aspergillus fumigatus; Acl: Aspergillus clavatus; Ani: Aspergillus nidulans; Aor: Aspergillus oryzae; Anig: Aspergillus niger; Ave: Aspergillus versicolor; Tma: Talaromyces marneffei; Pox: P. oxalicum; Pdi: Penicillium digitatum; Pru: Penicillium rubens Wisconsin; Ncr: Neurospora crassa; Tce: Talaromyces cellulolyticus; Tre: Trichoderma reesei; Tha: Trichoderma harzianum; Sce: Saccharomyces cerevisiae; Spo: Schizosaccharomyces pombe; Has: Homo sapiens. In panel G, red arrows represent GFP-PRMT3. The expression of prmt3-gfp was driven by its own promotor, Pprmt3, in the overexpression strain Oprmt3-gfp;G418R+. DAPI: 4,6-Diamidino-2-phenylindole. GFP: Green fluorescent protein. Scale bar = 20 μm. https://doi.org/10.1371/journal.pgen.1010867.g003 POX_f08428 contained 429 amino acid residues, and an ankyrin repeat domain (IPR002110, residues 61−109). POX_f08428 shared 100%, 66.67% and 34.83% identity, respectively, with PDE_04339 (EPS29390.1) from P. oxalicum strain 114–2, arginine N-methyltransferase 2 (XP_755242.1) from Aspergillus fumigatus strain Af293 and protein-arginine N5-methyltransferase (NP_010753.1) from S. cerevisiae S288C. POX_f08428 is conserved in eukaryotes and most closely related to its homologs in Aspergillus (Fig 2D). For convenience, POX_ f08428 was renamed PRMT2. POX_b03080 contained 546 amino acid residues and a methyltransfer_25 domain (IPR041698, residues 249 to 346). POX_b03080 shared 99.63%, 40.16% and 36.28% identity, respectively, with PDE_04847 (EPS29897.1) from P. oxalicum strain 114–2, ribosomal protein arginine N-methyltransferase Rmt3 (NP_595572.1) from S. pombe 972h- and protein arginine N-methyltransferase 3 (AAC39837.1) from H. sapiens. POX_b03080 was conserved in eukaryotes and most closely related to its homologs in Aspergillus (Fig 3D). For convenience, POX_b03080 was renamed PRMT3. To investigate the contribution of PRMT2 and PRMT3 to methyltransferase activity in P. oxalicum, methyltransferase production by overexpression strain Oprmt2;G418R+ and mutant Δprmt3;G418R+ was measured after culture in Avicel medium for 24 h. Methyltransferase production by Oprmt2;G418R+ significantly increased (Fig 2E), whereas that of Δprmt3;G418R+ significantly decreased, compared with Δku70;hphR+ and Cprmt3;bleR+, and there was no significant difference between Δku70;hphR+ and Cprmt3;bleR+ (Fig 3E), implying that PRMT2 and PRMT3 have methyltransferase activity. PRMT2 is responsible for R94-methylation of CxrAΔ1–60 To determin which of PRMT2 and PRMT3 responsible for methylation of R94 in CxrA, the Y2H assay was firstly employed. The results indicated that Y2HGold cells carrying BD-PRMT2 and AD-CxrA (Figs 2F and S10A), as well as BD-CxrAΔ1–60 and AD-PRMT3, grew well on SD/-Leu/-Trp/-His/-Ade agar plates (Fig 3F), indicating that the PRMT2 and PRMT3 interact with CxrA and CxrAΔ1–60, respectively. However, as unexpected, yeast cells carrying BD-PRMT3 and AD-CxrA were unable to growth on SD/-Leu/-Trp/-His/-Ade agar plates, suggesting that PRMT3 could not interact with the full-length CxrA (S10B Fig). Therefore, this difference maybe result from the use of a nonfunctional CxrAΔ1–60 lacking the DNA binding domain. CxrAΔ1–60 may affect protein folding. Furthermore, in vitro methylation experiments were performed. As unexpected, DNA sequence encoding a full-length CxrA fused with a glutathione-S-transferase (GST) tag failed to be recombinantly expressed in E. coli BL21. Therefore, the DNA fragments encoding CxrAΔ1–60 fused with a GST tag had to be expressed in E. coli BL21, as well as PRMT2/PRMT3 tagged with His, then the recombinant GST-CxrAΔ1–60 and His-PRMT2 or His-PRMT3 were purified. After reaction of GST-CxrAΔ1–60 and His-PRMT3 or His-PRMT2, with S-adenosyl-L-methionine as methyl donor, Western blotting was used to detect methylated CxrAΔ1–60 using specific antibodies including anti-mono methyl arginine, anti-symmetric di-methyl arginine, and anti-asymmetric di-methyl arginine. Bands corresponding to methylated GST-CxrAΔ1–60 appeared when recombinant His-PRMT2 and GST-CxrAΔ1–60 were treated with antibodies anti-mono methyl arginine and anti-asymmetric di-methyl arginine (Fig 2G), respectively, whereas no band was found for the mutated CxrAΔ1-60R94A (Fig 2H), showing that PRMT2 both mono-methylated and asymmetrically di-methylated CxrAΔ1–60 at R94. In contrast, the band corresponding to methylated GST-CxrAΔ1–60 did not appear when recombinant His-PRMT3 and GST-CxrAΔ1–60 were treated with specific antibodies (S11 Fig), indicating that PRMT3 is not responsible for CxrAΔ1–60 methylation. Nevertheless, it is possible that in vitro methylation modification of the full-length CxrA by PRMT2 or PRMT3 may be different from the results obtained using the nonfunctional truncation mutant CxrAΔ1–60 that may affect protein fold. PRMT3 appears to be a nuclear arginine N-methyltransferase To investigate the subcellular localization of PRMT3 in P. oxalicum hyphae, the overexpression strain Oprmt3-gfp;G418R+, carrying a GFP reporter was constructed, in which the fused gene prmt3-gfp was controlled by its own promoter, Pprmt3 (S7 Fig). In the Oprmt3-gfp;G418R+, there are two copies of prmt3 under their native promoters-one at the native locus and the other at the pepA locus (S7D Fig). After culture in Avicel medium for 24 h, the hyphae of Oprmt3-gfp;G418R+ and background strain Δku70;hphR+ were observed under a fluorescence microscope. The green fluorescence of the fusion protein, PRMT3-GFP, in Oprmt3-gfp;G418R+ appeared to overlap with blue fluorescence signals from 4,6-diamidino-2-phenylindole (DAPI), which specifically stains the nucleus, whereas there was no green fluorescence from Δku70;hphR+ (Fig 3G). PRMT3 mediates the expression of genes related to cellulase and xylanase biosynthesis by P. oxalicum To elucidate the effects of PRMT3 on the expression of genes related to cellulase and xylanase production by P. oxalicum, RNA-sequencing and a real-time quantitative reverse transcription PCR (RT-qPCR) assay were employed. Both Δku70;hphR+ and Δprmt3;G418R+ were cultured in Avicel medium for 24 h, after transfer from glucose, and their total RNA was extracted for RNA-sequencing. The sequencing data from three biological replicates were analyzed statistically and generated a high Pearson correlation coefficient (> 0.97) (S12 Fig), suggesting that the transcriptomic data were suitable for the subsequent analysis. With thresholds of |log2 fold change| > 1 and false discovery rate (FDR) < 0.05, there were 997 differentially expressed genes (DEGs) detected in Δprmt3;G418R+ compared with Δku70;hphR+, consisting of 321 up-regulated and 676 down-regulated genes (Fig 4A and S1 Dataset). The Kyoto Encyclopedia of Genes and Genomes (KEGG) annotation showed that these DEGs were mainly involved in metabolism, especially carbohydrate metabolism (75 genes) and amino acid metabolism (58 genes) (Fig 4B). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 4. Comparative analyses of transcriptomes from P. oxalicum mutant Δprmt3;G418R+ and background strain Δku70;hphR+ (A-E) and real-time reverse transcription quantitative PCR (RT-qPCR) confirmation (F). (A) Volcano plot of differentially expressed genes (DEGs). DEGs were selected with thresholds of |log2 fold change| > 1.0 and p-value < 0.05. (B) KEGG annotations of the DEGs modulated by PRMT3. (C) DEGs encoding cellulases and xylanases. (D) DEGs encoding putative transcription factors. (E) DEGs encoding putative sugar transporters. In panels D and E, the three columns of data corresponding to each fungal strain represent the three biological replicates for each strain. (F) The transcriptional levels of genes encoding major cellulases including two cellobiohydrolase genes cbh1 and cbh2, and an endo-β-1,4-glucanase gene eg1 in the P. oxalicum mutant Δprmt3;G418R+ relative to the background strain Δku70;hphR+. In panel F, the P. oxalicum strains were cultured for 4–48 h in the presence of Avicel. Gene expression in Δprmt3;G418R+ was normalized to the level of Δku70;hphR+. Data points show mean ± standard deviation. **p <0.01 and *p <0.05 indicate significant differences between Δprmt3;G418R+ and Δku70;hphR+, calculated by Student’s t-test. https://doi.org/10.1371/journal.pgen.1010867.g004 Among the 997 DEGs, nine cellulase/xylanase genes were found, namely one endo-β-1,4-glucanase gene (POX_d04883/eg1), five β-glucosidase genes (bgls, POX_a00284, POX_c04019, POX_e06772/bgl1, POX_e06434 and POX_f08346) and three endo-β-1,4-xylanase genes (POX_a01871/xyn11A, POX_c03490 and POX_f07706). Of these, POX_b02315/cel61A, POX_e06772/bgl1 and POX_e06434 were down-regulated (-2.4 < log2 fold change < -1.4), and the others were up-regulated (1.0 < log2 fold change < 1.6) in Δprmt3;G418R+ compared with Δku70;hphR+ (Fig 4C). The effects of PRMT3 on the expression of genes encoding putative TFs was then investigated by comparative analysis of transcriptomes, which identified 30 TF-encoding DEGs. The transcript abundances of 12 of them increased by 2.03–52.6-fold in Δprmt3;G418R+, whereas the other 18 genes decreased by 1.2–499-fold. Most of these TFs contained zinc-finger domains (i.e., six C2H2-type, 11 Zn2Cys6-type and one CCHC-type). Notably, expression of a key transcription repressor gene, cxrC in Δprmt3;G418R+ increased 2.46-fold, whereas transcripts of sporulation-regulated abaA, brlA and flbD decreased by 59.4–95.9% (Fig 4D). In addition, 25 DEGs encoding predicted sugar transporters were found, of which 13 were up-regulated (1.0 < log2 fold change < 2.9) and 12 down-regulated (-4.2 < log2 fold change < -1.3) in Δprmt3;G418R+. Of these, two major cellodextrin transporter genes, cdtC and cdtD were up-regulated 2.14- and 2.30-fold, respectively (Fig 4E). Three cellulase genes, cbh1, cbh2, and eg1 were subjected to RT-qPCR analysis of expression variation with induction duration and confirmation of RNA-sequencing data. The expression of eg1 after 24 h of induction was up-regulated in Δprmt3;G418R+ compared with Δku70;hphR+, whereas cbh1 and cbh2 showed no significant change, in agreement with the RNA-sequencing results. Transcriptional abundances of all three genes decreased up to 12 h, except for eg1 at 12 h, which markedly increased. At 48 h, the expression of cbh1 was up-regulated, whereas eg1 was down-regulated. The transcriptional level of cbh2 did not change (Fig 4F). PRMT3 was required for positive regulation of cellulase and xylanase production by cxrA As described above, PRMT3 was unable to methylate CxrA, but deletion of prmt3 from P. oxalicum increased cellulase and xylanase production by 67.0–149.7%, compared with the background strain Δku70;hphR+ after 4 days of Avicel induction (Fig 3). To investigate the effects of PRMT3 on CxrA action, the cxrA-overexpression strain OcxrA-gfp;G418R+ and mutant Δprmt3;cxrA-gfp; G418R+;bleR+ was sequentially constructed and confirmed by PCR (S7 Fig). In OcxrA-gfp;G418R+, the fusion gene cxrA-gfp was controlled by the predicted natural promoter of cxrA. In the OcxrA-gfp;G418R+ and Δprmt3;cxrA-gfp;G418R+;bleR+, there are two copies of cxrA under their native promoters- one at the native locus and the other at the pepA locus (S7E and S7F Fig). RT-qPCR confirmed that the expression of cxrA in OcxrA-gfp;G418R+ significantly increased, compared with that in Δku70;hphR+ after 48 h of Avicel induction (S9 Fig). Measurement of enzymatic activity revealed that cellulase and xylanase production by Δprmt3;cxrA-gfp;G418R+;bleR+ was 11.3–62.8% lower than that of the overexpression strain OcxrA-gfp;G418R+, when cultured in Avicel medium for 2–4 d (Fig 5A–5C), indicating that PRMT3 is required for positive regulation of cellulase and xylanase production by cxrA. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 5. Effects of PRMT3 on the regulatory functions of CxrA. (A) Filter paper cellulase (FPase) production. (B) Carboxymethyl cellulase (CMCase) production. (C) Xylanase production. (D) Subcellular localization of CxrA in P. oxalicum. In panels A-C, the fungal strains were pre-grown in glucose medium for 24 h, then transferred into Avicel medium for 2–4 d. The uppercase and lowercase letters indicate p < 0.01 and p < 0.05, respectively. Different letters indicate significant differences assessed by one-way ANOVA. All tested samples were normally distributed. In panel D, 4,6-diamidino-2-phenylindole (DAPI) was employed to stain mycelial nuclei. Subcellular localization of CxrA was determined with GFP signals observed by a fluorescence microscope. Scale bar = 20 μm. https://doi.org/10.1371/journal.pgen.1010867.g005 PRMT3 apparently assists the entry of CxrA into the nucleus To further elucidate the interactions between CxrA and PRMT3, Δprmt3;cxrA-gfp;G418R+;bleR+ and OcxrA-gfp;G418R+ were cultured in Avicel medium for 24 h, then their hyphae were collected for microscopic examination. The green fluorescent signals from CxrA-GFP in OcxrA-gfp;G418R+ appeared to overlap with the blue fluorescence signals from DAPI, indicating that CxrA was localized in the nucleus, whereas the green fluorescence signals from CxrA-GFP in Δprmt3;cxrA-gfp;G418R+;bleR+ were evenly distributed in the whole hypha. This seems that deletion of prmt3 partially obstructed nuclear translocation of CxrA protein, in other words, it appears that PRMT3 assists the entry of CxrA into the nucleus (Fig 5D). Co-expression analysis of genes regulated by CxrA and PRMT3 Co-expression analysis of genes regulated by PRMT3 and CxrA was performed by RNA-sequencing of Δprmt3;G418R+ and ΔcxrA;G418R+, compared with Δku70;hphR+ cultured in Avicel medium for 24 h, after transfer from glucose. The resulting data from three biological replicates had a high Pearson’s correlation coefficient (> 0.97) (S12 Fig). With thresholds of |log2 fold change| > 1 and FDR < 0.05, there were 2,552 DEGs in ΔcxrA;G418R+ compared with Δku70;hphR+, of which 1,253 were up-regulated and 1,299 down-regulated (S2 Dataset). These included 27 key cellulase and xylanase genes, 128 putative TF-encoding genes and 28 putative sugar transporter-encoding genes (S2 Dataset), in agreement with a previous report [4]. Comparative analysis identified 657 DEGs co-regulated by PRMT3 and CxrA under Avicel induction, of which 113 were up-regulated and 448 down-regulated in Δprmt3;G418R+ and ΔcxrA;G418R+, relative to Δku70;hphR+ (Fig 6A). These co-regulated genes were mainly involved in metabolism, especially carbohydrate and amino acid metabolism (Fig 6B). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 6. Co-regulation between CxrA and PRMT3 in P. oxalicum. (A) Numbers of differentially expressed genes (DEGs) in regulons between cxrA and prmt3. DEG regulons were determined by comparison of transcriptomic data from Δprmt3;G418R+ and ΔcxrA;G418R+ with that of the background strain Δku70;hphR+. “Up” and “Down” indicate up- and down-regulation. (B) KEGG annotations of the 657 genes co-regulated by cxrA and prmt3. (C) Heatmap illustrating the expression of DEGs encoding plant-cell-wall-degrading enzymes. (D) Heatmap showing the expression of DEGs encoding putative transcription factors. (E) Heatmap displaying the expression of DEGs encoding sugar transporters. In panels C and D, the three columns of data corresponding to each fungal strain represent the three biological replicates for each strain. https://doi.org/10.1371/journal.pgen.1010867.g006 The co-regulated DEG set included four BGL genes (bgl1, POX_a00284, POX_f08346 and POX_e06434) and two xylanase genes (xyn11A and POX_c03490). The mRNA/transcription levels of these genes were up-regulated in Δprmt3;G418R+, but down-regulated in ΔcxrA;G418R+, compared with Δku70;hphR+, except for bgl1 and POX_e06434, which were down-regulated in both mutants (Fig 6C). The expression of most co-regulated DEGs encoding TFs and sugar transporters decreased in Δprmt3;G418R+ and ΔcxrA;G418R+, compared with Δku70;hphR+ (Fig 6D and 6E), however, the cellodextrin transporter genes cdtC and cdtD were downregulated in ΔcxrA;G418R+ and upregulated in Δprmt3;G418R+, compared with Δku70;hphR+ (Fig 6E). Notably, gene prmt2 was included in the CxrA regulon, rather than in the prmt3 regulon. In ΔcxrA;G418R+, the expression of prmt2 increased by 1.2-fold, compared with Δku70;hphR+ (S2 Dataset), indicating that CxrA inhibits the expression of prmt2 on Avicel. Discussion Previous work demonstrated that an Zn2Cys6-type TF, CxrA, promotes the biosynthesis of cellulase and xylanase in P. oxalicum [2]. From this study, it appears that the N-terminal region, CxrAΔ207–733 is essential to the regulatory function of full-length CxrA, containing the DNA-binding domain (CxrAΔ1–16&Δ59–733) [4] and the methylation site, R94 under induction conditions. Nevertheless, in general, for many Zn2Cys6-type TFs such as ClrB, the C-terminal region is capable of transcriptional activation and the intermediate region participates in the regulation of TF activity [14]. In addition, the N-terminal CxrAΔ17–733 repressed the regulatory action of CxrA by an unknown mechanism. Moreover, CxrA appeared to interact with the truncation mutant CxrAΔ1–60 via Y2H assay, i.e., dimerize, in common with other TFs belonging to the zinc finger family, such as ACE3 [15], CLR-2, XLR-1 [16] and CxrC [5]. The truncation mutant CxrAΔ1–60 may affect protein fold, thereby leading to the different results from interaction between the full-length CxrA and itself. However, whether self-interaction of the full-length CxrA occurs actually in P. oxalicum merits further investigation, as well as whether this interaction is required for CxrA function. Post-translational modification generally modulates protein function, especially methylation, phosphorylation, and acetylation. Although the exact function of arginine methylation is still controversial, accumulated evidence indicates that it is involved in many cellular processes, such as transcription activation and repression, protein-protein interaction, DNA repair and pre-mRNA splicing [17,18]. However, these reports mainly focus on humans and other mammals [19–25] and to the best of our knowledge, there is only one previous report on arginine methylation in microorganisms [11]. This study found that R94 in CxrAΔ1–60 was methylated by protein arginine N-methyltransferase PRMT2. Nevertheless, it should be noted that the truncated CxrAΔ1–60 used for in vitro methylation experiments is nonfunctional and lacks the DNA binding domain, and thus, that it is possible that interactions between full-length CxrA and PRMT2 may be different due to changes in protein folding, etc. Moreover, R94 was required for the activation by CxrA of cellulase and xylanase biosynthesis (Fig 7), as well as the interaction between CxrAΔ1–60 and itself, which might be due to the methylation of R94. However, it did not exclude the possibility that the exchange of the amino acid maybe simply lead to misfolding, which requires further observation by either CD-spectroscopy or HDX-MS. How methylation modification influences the regulatory roles of CxrA also remains unclear. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 7. Proposed model of CxrA function in P. oxalicum grown on Avicel. CxrAΔ1–60 is methylated by an arginine N-methyltransferase PRMT2 at R94, and interacts with itself. In addition, the CxrAΔ1–60 interacts with another arginine N-methyltransferase, PRMT3, which apparently assists the nuclear translocation of CxrA. PRMT3 down-regulates the production of cellulase and xylanase in wild-type P. oxalicum, when growing on Avicel, by an unknown mechanism, but PRMT3 is also required for cxrA-mediated up-regulation of cellulase and xylanase production by P. oxalicum. Notably, the role for PRMT3 in regulating the expression of major cellulase and xylanase expression may change over time. The order of occurrence of the interaction between CxrAΔ1–60 and PRMT3, CxrAΔ1–60 self-interaction/dimerization, and CxrAΔ1–60 methylation remains unclear. Nevertheless, it should be noted that the CxrAΔ1–60 used for in vitro methylation experiments is nonfunctional and lacks the DNA binding domain, and thus, that it is possible that interactions between full-length CxrA and PRMT2 may be different due to changes in protein folding. Lines with arrows display activation, and barred lines show repression. Dashed lines indicate an unclear mechanism. https://doi.org/10.1371/journal.pgen.1010867.g007 Homologous alignment indicated that PRMT2 shared a low identity (34.83%) with type IV PRMT, RMT2 from yeast [26]. RMT2 catalyzes the formation of δ-NG monomethyl arginine [27]. Notably, in vitro methylation experiments confirmed that PRMT2 could catalyze the biosynthesis of the ω-NG monomethyl arginine and asymmetric ω-NG, NG-dimethylarginine [6]; this activity would normally be classified as a type I PRMT, so further study will be needed to clarify the correct classification of PRMT2. It should be noted that the it was not possible to generate the deletion mutant Δprmt2;G418R+, suggesting that PRMT2 is critical to cell-survival and that deletion results in Δprmt2;G418R+ becoming a non-viable mutant. Overexpression of prmt2 up-regulated cellulase and xylanase production under Avicel induction, and it is possible that increasing R94 methylation of CxrA strengthens its regulatory effect. Moreover, another arginine methyltransferase PRMT3 was also identified, which down-regulated cellulase and xylanase production at the late stage of Avicel induction in wild-type P. oxalicum, whereas PRMT3 was required for up-regulation of cellulase and xylanase production by cxrA. Notably, the role for PRMT3 in regulating cellulase and xylanase expression may change over time. In addition, unexpectedly, PRMT3 interacted with CxrAΔ1–60 but not with full-length CxrA via Y2H assay, whereas it did not methylate CxrAΔ1–60 (Fig 7). In addition, it appears that nuclear translocation of CxrA is facilitated by PRMT3, but that CxrA can also enter the nucleus independently of PRMT3. The different interactions between PRMT2 and the full-length CxrA, and nonfunctional CxrA might be due to changes in protein folding. PRMT3 is highly conserved in all eukaryotes [17] and is essential to a variety of cellular processes, for examples, PRMT3 influences the relative levels of small ribosomal subunits in yeast, by interaction with Rps2 (40S ribosomal protein S2), but not by methylation [19]; negatively regulates antiviral responses in Zebrafish [20]; stimulates tumorigenesis via controlling c-MYC stabilization in colorectal cancer [21]; represses retinoic acid signaling through interacting with retinal dehydrogenase 1 [22]. This study initially found that PRMT3 influenced the cellulase and xylanase biosynthesis in P. oxalicum through unknown mechanism. An important issue arising from the above findings is the apparent paradox that PRMT3 suppresses cellulase and xylanase biosynthesis in the wild-type P. oxalicum, but it is required for up-regulation of cellulase and xylanase production by cxrA. It appears that PRMT3 facilitates the entry of CxrA into the nucleus, repressed the expression of the important transcriptional repressor gene, cxrC [5] and dynamically regulated some cellulase and xylanase genes. This implies that PRMT3 repressed cellulase and xylanase production in P. oxalicum through a complex mechanism. To explain this mystery, the substrates methylated by PRMT3 in P. oxalicum should be identified. In addition, the regulatory effects of PRMT3 conflict with activation of CxrA, suggesting that genetic engineering involving their genes, to improve the yields of fungal cellulase and xylanase in synthetic biology, will be very challenging until their functions and interactions are much more clearly understood. Improved understanding of the functions of PRMTs is not only relevant to fungi, but also to human health. PRMT is considered a promising target for inhibition, with great potential for such inhibitors in cancer therapy [18] and treatment of other diseases. For example, the allosteric PRMT3 inhibitor SGC707 effectively reduced the extent of hepatic steatosis (fatty liver) in mice [28]. PRMT inhibition may also be possible by the alternative approach of down-regulating the expression of PRMT-encoding genes. Further research is needed to identify more regulatory genes and to improve understanding of cellulase and xylanase biosynthesis regulation in P. oxalicum. Materials and methods Microbial strains, medium and culture conditions All microbial strains used in this study are shown in Table 1. E. coli strains DH5α and DE3 were used for plasmid construction and protein heterologous expression, and were cultured in Luria-Bertani medium at 37°C. Download: PPT PowerPoint slide PNG larger image TIFF original image Table 1. Penicillium oxalicum strains used in this study. https://doi.org/10.1371/journal.pgen.1010867.t001 S. cerevisiae strains Y2HGold and Y187, used for the Y2H assay, were cultured at 30°C in yeast extract peptone dextrose medium (YPD) containing peptone (20.0 g/L), yeast extract (10.0 g/L), glucose (20.0 g/L) and adenine (4.0 g/L). Synthetic defined medium (SD Base) lacking tryptophan, leucine, histidine, and adenine (SD/–Trp/–Leu/–His/–Ade), with Aureobasidin A (AbA, 100 ng/mL) was used for screening of recombinant yeast. P. oxalicum strains, including the background strain Δku70;hphR+ [29] and a series of constructed mutants were routinely cultured at 28°C on PDA plates for 4−5 d. The conidia were collected with 0.2% Tween 80 (Sangon Biotech Co., Ltd., Shanghai, China) from the plates, and used for reproduction. To measure the production of cellulase and xylanase, P. oxalicum strains were pre-grown in modified minimal medium (MMM; g/L, (NH4)2SO4 4.0, CaCl2 0.6, KH2PO4 4.0, MgSO4·7H2O 0.6, FeSO4·7H2O 0.25, MnSO4·H2O 0.08, ZnCl2 0.085, CoCl2 0.1, 2 ml/L Tween 80, pH 5.5) containing 1% glucose as the sole carbon source for 24 h, then transferred into MMM containing 2% Avicel as the carbon source [5] for 2–4 d. P. oxalicum strains were cultured on PDA and MMM agar plates containing 1% glucose, 2% CMC, or 2% Avicel as carbon source, at 28°C for 4 d, for observation of colony phenotype. For RNA-sequencing and RT-qPCR assays, P. oxalicum strains were cultured at 28°C for 4–48 h, after a transfer from glucose (1%) medium, as described above. Yeast two-hybrid (Y2H) assay The Y2H assay was performed with the Matchmaker Gold Yeast Two-Hybrid System, following the manufacturer’s instructions (TaKaRa, Dalian, China). The DNA fragments encoding CxrAΔ1–60, CxrAΔ1-60R94A and PRMT2/3 used as the baits were amplified from the genomic DNA of HP7-1 and cloned into the plasmid pGBKT7 at EcoRI/BamHI sites, respectively, resulting in the recombinant pGBKT7-cxrAΔ1–60, pGBKT7-cxrAΔ1-60R94A, pGBKT7-prmt2 and pGBKT7-prmt3. The DNA fragments encoding PRMT2/3 and CxrA used as the preys, from P. oxalicum HP7-1 genomic DNA, were introduced into the plasmid pGADT7, respectively, to generate the pGADT7-cxrA, pGADT7-cxrAΔ1-60R94A, pGADT7-prmt2 and pGADT7-prmt3. Subsequently, these relevant recombinant plasmids were co-transformed into Y2HGold competent cells. The SD/-Trp/-Leu/-His/-Ade medium with 100 ng/mL AbA and 1 M X-α-Gal (chromogenic substrate) was used to screen for positive transformants. Y2HGold cells containing pGBKT7-p53 and pGADT7-T was the positive control, and Y2HGold containing pGBKT7-Lam and pGADT7-T was the negative control. GST-pulldown assay In vitro protein-protein interactions were investigated using the GST-pulldown assay, as described previously [5]. Western blotting was employed to test the target protein with anti-GST and anti-His antibodies (TransGen Biotech Co., Ltd.). Construction of recombinant P. oxalicum strains All mutants were constructed based on homologous recombination techniques, as described previously [2]. The primers used for mutant construction are shown in S1 Table. Colony phenotype observation P. oxalicum colonies on agar plates containing different carbon sources were photographed with a digital camera (EOS 6D; Canon Inc., Tokyo, Japan). RNA-Sequencing and RT-qPCR assays RNA-sequencing was performed by Frasergen (Wuhan, China) where the sequenced data were analyzed as described previously [30]. Briefly, the software packages SOAPnuke (v 2.1.0) [31] and HISAT2 (v 2.2.1) [32] were used for quality control of data and mapping to the P. oxalicum genome, respectively [13]. Gene expression was analyzed with both RSEM (v1.3.3) [33] and Bowtie2 (v2.3.5) [34], then visualized via the number of fragments per kilobase of exon per million mapped reads. Differentially expressed genes were searched with the DESeq2 tool [35], using false discovery rate (FDR) < 0.05 and |Log2 fold change| > 1.0 as thresholds. The RT-qPCR assay was performed with the ChamQ Universal SYBR qPCR Master Mix, following the manufacturer’s instructions (Vazyme Biotech, Nanjing, China). The internal reference gene was POX_d06005 (encoding actin), generated by PCR amplification (S1 Table). The relative expression of the tested gene in P. oxalicum mutants was normalized to that of the background strain Δku70;hphR+, calculated by the 2−ΔΔCT method [36]. All experiments were replicated at least three times. Enzyme activity assays The cellulase and xylanase activities of P. oxalicum strains were measured as described previously [2]. Filter paper cellulase and carboxymethyl cellulase were assayed with Whatman No. 1 filter paper (GE Healthcare Life Sciences, Little Chalfont, UK) and 1% carboxymethyl cellulose (Sigma-Aldrich, St. Louis, MO) as substrates, respectively. Xylanase activity was assayed with beechwood xylan (Megazyme International, Bray, Ireland) as substrate. An enzyme activity unit (U) was defined as the quantity of enzyme required to produce 1 μmoL of reducing sugar per min from the substrate. Cellulase and xylanase production by P. oxalicum was defined as units of enzyme activity per milliliter of crude culture or per gram of intracellular protein extracted from mycelia. Extraction of intracellular protein from P. oxalicum mycelia The harvested mycelia separated from MMM containing Avicel inoculated with P. oxalicum were ground into powder after adding liquid nitrogen. The powder was dissolved in protein extraction buffer comprising of 0.5 mM phenylmethylsulfonyl fluoride (PMSF), 8.5 g/L NaCl, 0.2 g/L NaH2PO4, 2.2 g/L Na2HPO4 and 0.4 g/L ethylene diamine tetraacetic acid (EDTA), then the supernatant containing mycelial proteins was collected by centrifugation. Protein concentration was measured by Bradford Assay Kit (Pierce Biotechnology, Rockford, IL, USA). Measurement of methyltransferase production by P. oxalicum Methyltransferase production of P. oxalicum mutant Δprmt3;G418R+, background strain Δku70;hphR+ and complementation strain Cprmt3;bleR+ was assayed using a histone methyltransferase ELISA assay kit (#YJ608911; mlBio, Shanghai, China), following the manufacturer’s instructions. In vitro methylation experiment In vitro methylation assay was performed as described previously [37]. In brief, the GST-tagged CxrA and His-tagged methyltransferase were expressed in E. coli and purified. The purified GST-CxrA and methyltransferase were mixed with S-adenosyl-L-methionine as methyl donor, and reacted for 2 h at 30°C. Western blotting was used to detect substrate methylation with the corresponding antibodies including anti-mono methyl arginine, anti-symmetric di-methyl arginine, and anti-asymmetric di-methyl arginine. Protein sequence analysis Protein sequences from P. oxalicum and the other related species were analyzed on and downloaded from, the NCBI website (https://www.ncbi.nlm.nih.gov/). Conserved domains were identified using the SMART database (http://smart.embl.de/). The cladogram of relative proteins was built using the Neighbor-joining method and the Poisson correction model in MEGA version X [38]. Alignment of protein sequences was performed with the multiple Sequence Alignment tool, ClustalW, on MUSCLE (https://www.ebi.ac.uk/Tools/msa/muscle/). Investigation of protein subcellular localization The subcellular localization of target proteins in the mycelia of P. oxalicum cultured in Avicel medium was determined with green fluorescent protein (GFP) as the reporter, as described previously [5]. LC-MS/MS analysis The LC-MS/MS assay to detect post-translational modification of CxrA was performed as described previously [5]. In brief, the fusion protein CxrA-GFP was precipitated with anti-GFP antibody (TransGen Biotech Co., Ltd.) and purified by BeaverBeads Proute from an A (or A/G) Immunoprecipitation Kit (Beaver Biomedical Engineering Co., Ltd. Suzhou, China). The isolated PoxCxrA-GFP was analyzed on a liquid chromatography system (Waters, Milford, MA, USA) coupled with a Thermo Scientific LTQ-Orbitrap Mass Spectrometer (Thermo Fisher Scientific, Bremen, Germany). Data statistical analysis The obtained data in this study was statistically analyzed using Microsoft Excel (Office 2019) (Microsoft, Redmond, WA, USA) and SPSS (IBM, Armonk, NY, USA) with Student’s t-test and one-way ANOVA, respectively. Microbial strains, medium and culture conditions All microbial strains used in this study are shown in Table 1. E. coli strains DH5α and DE3 were used for plasmid construction and protein heterologous expression, and were cultured in Luria-Bertani medium at 37°C. Download: PPT PowerPoint slide PNG larger image TIFF original image Table 1. Penicillium oxalicum strains used in this study. https://doi.org/10.1371/journal.pgen.1010867.t001 S. cerevisiae strains Y2HGold and Y187, used for the Y2H assay, were cultured at 30°C in yeast extract peptone dextrose medium (YPD) containing peptone (20.0 g/L), yeast extract (10.0 g/L), glucose (20.0 g/L) and adenine (4.0 g/L). Synthetic defined medium (SD Base) lacking tryptophan, leucine, histidine, and adenine (SD/–Trp/–Leu/–His/–Ade), with Aureobasidin A (AbA, 100 ng/mL) was used for screening of recombinant yeast. P. oxalicum strains, including the background strain Δku70;hphR+ [29] and a series of constructed mutants were routinely cultured at 28°C on PDA plates for 4−5 d. The conidia were collected with 0.2% Tween 80 (Sangon Biotech Co., Ltd., Shanghai, China) from the plates, and used for reproduction. To measure the production of cellulase and xylanase, P. oxalicum strains were pre-grown in modified minimal medium (MMM; g/L, (NH4)2SO4 4.0, CaCl2 0.6, KH2PO4 4.0, MgSO4·7H2O 0.6, FeSO4·7H2O 0.25, MnSO4·H2O 0.08, ZnCl2 0.085, CoCl2 0.1, 2 ml/L Tween 80, pH 5.5) containing 1% glucose as the sole carbon source for 24 h, then transferred into MMM containing 2% Avicel as the carbon source [5] for 2–4 d. P. oxalicum strains were cultured on PDA and MMM agar plates containing 1% glucose, 2% CMC, or 2% Avicel as carbon source, at 28°C for 4 d, for observation of colony phenotype. For RNA-sequencing and RT-qPCR assays, P. oxalicum strains were cultured at 28°C for 4–48 h, after a transfer from glucose (1%) medium, as described above. Yeast two-hybrid (Y2H) assay The Y2H assay was performed with the Matchmaker Gold Yeast Two-Hybrid System, following the manufacturer’s instructions (TaKaRa, Dalian, China). The DNA fragments encoding CxrAΔ1–60, CxrAΔ1-60R94A and PRMT2/3 used as the baits were amplified from the genomic DNA of HP7-1 and cloned into the plasmid pGBKT7 at EcoRI/BamHI sites, respectively, resulting in the recombinant pGBKT7-cxrAΔ1–60, pGBKT7-cxrAΔ1-60R94A, pGBKT7-prmt2 and pGBKT7-prmt3. The DNA fragments encoding PRMT2/3 and CxrA used as the preys, from P. oxalicum HP7-1 genomic DNA, were introduced into the plasmid pGADT7, respectively, to generate the pGADT7-cxrA, pGADT7-cxrAΔ1-60R94A, pGADT7-prmt2 and pGADT7-prmt3. Subsequently, these relevant recombinant plasmids were co-transformed into Y2HGold competent cells. The SD/-Trp/-Leu/-His/-Ade medium with 100 ng/mL AbA and 1 M X-α-Gal (chromogenic substrate) was used to screen for positive transformants. Y2HGold cells containing pGBKT7-p53 and pGADT7-T was the positive control, and Y2HGold containing pGBKT7-Lam and pGADT7-T was the negative control. GST-pulldown assay In vitro protein-protein interactions were investigated using the GST-pulldown assay, as described previously [5]. Western blotting was employed to test the target protein with anti-GST and anti-His antibodies (TransGen Biotech Co., Ltd.). Construction of recombinant P. oxalicum strains All mutants were constructed based on homologous recombination techniques, as described previously [2]. The primers used for mutant construction are shown in S1 Table. Colony phenotype observation P. oxalicum colonies on agar plates containing different carbon sources were photographed with a digital camera (EOS 6D; Canon Inc., Tokyo, Japan). RNA-Sequencing and RT-qPCR assays RNA-sequencing was performed by Frasergen (Wuhan, China) where the sequenced data were analyzed as described previously [30]. Briefly, the software packages SOAPnuke (v 2.1.0) [31] and HISAT2 (v 2.2.1) [32] were used for quality control of data and mapping to the P. oxalicum genome, respectively [13]. Gene expression was analyzed with both RSEM (v1.3.3) [33] and Bowtie2 (v2.3.5) [34], then visualized via the number of fragments per kilobase of exon per million mapped reads. Differentially expressed genes were searched with the DESeq2 tool [35], using false discovery rate (FDR) < 0.05 and |Log2 fold change| > 1.0 as thresholds. The RT-qPCR assay was performed with the ChamQ Universal SYBR qPCR Master Mix, following the manufacturer’s instructions (Vazyme Biotech, Nanjing, China). The internal reference gene was POX_d06005 (encoding actin), generated by PCR amplification (S1 Table). The relative expression of the tested gene in P. oxalicum mutants was normalized to that of the background strain Δku70;hphR+, calculated by the 2−ΔΔCT method [36]. All experiments were replicated at least three times. Enzyme activity assays The cellulase and xylanase activities of P. oxalicum strains were measured as described previously [2]. Filter paper cellulase and carboxymethyl cellulase were assayed with Whatman No. 1 filter paper (GE Healthcare Life Sciences, Little Chalfont, UK) and 1% carboxymethyl cellulose (Sigma-Aldrich, St. Louis, MO) as substrates, respectively. Xylanase activity was assayed with beechwood xylan (Megazyme International, Bray, Ireland) as substrate. An enzyme activity unit (U) was defined as the quantity of enzyme required to produce 1 μmoL of reducing sugar per min from the substrate. Cellulase and xylanase production by P. oxalicum was defined as units of enzyme activity per milliliter of crude culture or per gram of intracellular protein extracted from mycelia. Extraction of intracellular protein from P. oxalicum mycelia The harvested mycelia separated from MMM containing Avicel inoculated with P. oxalicum were ground into powder after adding liquid nitrogen. The powder was dissolved in protein extraction buffer comprising of 0.5 mM phenylmethylsulfonyl fluoride (PMSF), 8.5 g/L NaCl, 0.2 g/L NaH2PO4, 2.2 g/L Na2HPO4 and 0.4 g/L ethylene diamine tetraacetic acid (EDTA), then the supernatant containing mycelial proteins was collected by centrifugation. Protein concentration was measured by Bradford Assay Kit (Pierce Biotechnology, Rockford, IL, USA). Measurement of methyltransferase production by P. oxalicum Methyltransferase production of P. oxalicum mutant Δprmt3;G418R+, background strain Δku70;hphR+ and complementation strain Cprmt3;bleR+ was assayed using a histone methyltransferase ELISA assay kit (#YJ608911; mlBio, Shanghai, China), following the manufacturer’s instructions. In vitro methylation experiment In vitro methylation assay was performed as described previously [37]. In brief, the GST-tagged CxrA and His-tagged methyltransferase were expressed in E. coli and purified. The purified GST-CxrA and methyltransferase were mixed with S-adenosyl-L-methionine as methyl donor, and reacted for 2 h at 30°C. Western blotting was used to detect substrate methylation with the corresponding antibodies including anti-mono methyl arginine, anti-symmetric di-methyl arginine, and anti-asymmetric di-methyl arginine. Protein sequence analysis Protein sequences from P. oxalicum and the other related species were analyzed on and downloaded from, the NCBI website (https://www.ncbi.nlm.nih.gov/). Conserved domains were identified using the SMART database (http://smart.embl.de/). The cladogram of relative proteins was built using the Neighbor-joining method and the Poisson correction model in MEGA version X [38]. Alignment of protein sequences was performed with the multiple Sequence Alignment tool, ClustalW, on MUSCLE (https://www.ebi.ac.uk/Tools/msa/muscle/). Investigation of protein subcellular localization The subcellular localization of target proteins in the mycelia of P. oxalicum cultured in Avicel medium was determined with green fluorescent protein (GFP) as the reporter, as described previously [5]. LC-MS/MS analysis The LC-MS/MS assay to detect post-translational modification of CxrA was performed as described previously [5]. In brief, the fusion protein CxrA-GFP was precipitated with anti-GFP antibody (TransGen Biotech Co., Ltd.) and purified by BeaverBeads Proute from an A (or A/G) Immunoprecipitation Kit (Beaver Biomedical Engineering Co., Ltd. Suzhou, China). The isolated PoxCxrA-GFP was analyzed on a liquid chromatography system (Waters, Milford, MA, USA) coupled with a Thermo Scientific LTQ-Orbitrap Mass Spectrometer (Thermo Fisher Scientific, Bremen, Germany). Data statistical analysis The obtained data in this study was statistically analyzed using Microsoft Excel (Office 2019) (Microsoft, Redmond, WA, USA) and SPSS (IBM, Armonk, NY, USA) with Student’s t-test and one-way ANOVA, respectively. Supporting information S1 Fig. Construction strategy for P. oxalicum mutants involving introduction of truncated cxrA genes into ΔcxrA;G418R+ (A) and confirmation by PCR analysis (B-E). (B) PCR production of POX_d05452 with primers POX_d05452-F/POX_d05452-R. (C) DNA fragment with primers POX_d05452-LF/Ble-VR. (D) DNA fragment with primers Ble-VF/POX_d05452-LR. M: 1 kb DNA marker; 1: ddH2O; 2: Δku70;hphR+, 3: CcxrA;bleR+, 4: 1–60; 5: 1–206; 6:1–591; 7: 17–733; 8: 61–733; 9: 61–206. (E) PCR verification of mutant R94A;bleR+. M: 1 kb DNA marker; 1: ddH2O; 2: Δku70;hphR+, 3: R94A;bleR+. Left panel indicates amplification of DNA fragment with primers POX_d05452-F/POX_d05452-R; Middle panel shows PCR products with primers POX_d05452-LF/Ble-VR; Right panel shows PCR amplification of DNA fragment with primers Ble-VF/POX_d05452-LR. The bottom panel shows verification of DNA sequence. M: 1 kb DNA marker; 1–12: Transformants; +: Δku70;hphR+;–: ddH2O. https://doi.org/10.1371/journal.pgen.1010867.s001 (TIF) S2 Fig. Phenotypes of P. oxalicum strains on solid plates containing various carbon sources. These fungal strains were cultured for 4 d. PDA: Potato dextrose agar; CMC: Carboxymethyl cellulose. https://doi.org/10.1371/journal.pgen.1010867.s002 (TIF) S3 Fig. Construction of P. oxalicum overexpression strain OcxrA-his;G418R+ and the production of cellulase and xylanase by the constructed strains. (A) Schematic illustration showing construction strategy. (B) PCR confirmation. Upper panel shows PCR production of cxrA; Middle panel indicates DNA fragments with primers POX_d05452-F/G418-VR; Bottom panel presents DNA fragments with primers POX_d05452-F/G418-VR. (C) Filter paper cellulase (FPase) production. (D) Carboxymethyl cellulase (CMCase) production. (E) xylanase production. Fungal strains were pre-grown in glucose medium for 24 h, then transferred into Avicel medium for 2–4 d. Enzyme production is normalized to the intracellular proteins of mycelia representing fungal growth. (F) Relative expression of cxrA in both OcxrA-his;G418R+ and Δku70;hphR+. Total RNA as template was extracted from fungal mycelia harvested after culture on Avicel for 48 h. ** p < 0.01 and * p < 0.05 indicate significant differences between the overexpression strain and background strain, assessed by Student’s t-test. https://doi.org/10.1371/journal.pgen.1010867.s003 (TIF) S4 Fig. LC-MS/MS assay indicating post-translational modification of CxrA in the presence of Avicel. (A) Amino acid sequence; residues in yellow were modified. Red P, M and A represent phosphorylation, acetylation, and methylation. (B) Oligopeptides with red color were identified by LC-MS/MS. https://doi.org/10.1371/journal.pgen.1010867.s004 (PDF) S5 Fig. Autoactivation experiment on the full-length CxrA in Saccharomyces cerevisiae. Serial dilutions of yeast Y2HGold cells carrying pGBKT7-cxrA and pGBKT7, pGBKT7+pGADT7-p53, pGBKT7+pGADT7-Lam as controls were cultured on SD/-Trp, SD/-Trp/X-α-Gal and SD/-Trp/X-α-Gal/AbA at 30°C for 4 d. https://doi.org/10.1371/journal.pgen.1010867.s005 (TIF) S6 Fig. Autoactivation detection of the bait CxrAΔ1–60 (A) and determination of CxrAΔ1–60 toxicity to yeast cells (B). Serial dilutions of yeast Y2HGold cells carrying pGBKT7-cxrAΔ1–60 and pGBKT7 as control were cultured on SD/-Trp, SD/-Trp/X-α-Gal and SD/-Trp/X-α-Gal/AbA at 30°C for 4 d. https://doi.org/10.1371/journal.pgen.1010867.s006 (TIF) S7 Fig. Construction strategy (A–F) and PCR verification (G–Y) of P. oxalicum mutants used in this study. These strains include the overexpression strain OPOX_f08428;G418R+ (A; G–J), mutants ΔPOX_b03080;G418R+ (B; K–M), ΔPOX_d05270;G418R+ (B; N–P), ΔPOX_e06662;G418R+ (B; Q–S), complementation strain Cprmt3;bleR+ (C; T–V), Oprmt3-gfp;G418R+ (D; W), OcxrA-gfp;G418R+ (E; X) and Δprmt3;cxrA-gfp;G418R+;bleR+ (F; Y). (G; T) PCR amplification of POX_d05452 with primer pair POX_d05452-F/POX_d05452-R. (H), (L), (O) and (R) DNA fragments with primers Target-LF/G418-VR. (I) PCR production of G418 resistance gene. (J), (M), (P) and (S) DNA fragments with primers G418-VF/Target-RR. (K), (N) and (Q) PCR amplification of target genes. M: 1 kb DNA marker; 1–3: Three transformants; 4: Δku70;hphR+; 5: ddH2O. (T) PCR amplification of POX_d05452. (U) PCR products with primers POX_d05452-LF/Ble-VR. (V) PCR amplification of DNA fragment with primers Ble-VF/POX_d05452-LR. M: 1 kb DNA marker; 1: ddH2O; 2: Δku70;hphR+; 3: Cprmt3;bleR+. In panel W, X and Y, upper panel shows amplification of DNA fragment with primers POX_d05452-F/POX_d05452-R; Middle panel shows PCR products with primers POX_d05452-LF/Ble-VR; Bottom panel shows PCR amplification of DNA fragment with primers Ble-VF/POX_d05452-LR. M: 1 kb DNA marker; 1: ddH2O; 2: Δku70; 3: Oprmt3-gfp;G418R+, OcxrA-gfp;G418R+ or Δprmt3;cxrA-gfp;G418R+;bleR+. pepA (POX_d05452): aspartic protease gene; G418:; ble: bleomycin antibiotics gene; pepA-L: left-flanking sequence of gene pepA; pepA-R: right-flanking sequence of gene pepA; pepA-P: the promoter region of gene pepA; pepA-T: the terminus region of gene pepA; ORF: open reading frame; POX_f08428-P: the promoter region of gene POX_f08428; POX_f08428-T: the terminus region of gene POX_f08428; cxrA-P: the promoter region of gene cxrA; cxrA-T: the terminus region of gene cxrA. https://doi.org/10.1371/journal.pgen.1010867.s007 (TIF) S8 Fig. Filter paper cellulase production of P. oxalicum mutants ΔPOX_d05270;G418R+ and ΔPOX_e06662;G418R+. Fungal strains were cultivated on Avicel for 2–4 days after transfer from glucose. Data values indicate means ± standard deviation. https://doi.org/10.1371/journal.pgen.1010867.s008 (TIF) S9 Fig. Relative expression of the genes POX_f08428, prmt3 and cxrA in P. oxalicum overexpression strains OPOX_f08428;G418R+, Oprmt3-gfp;G418R+ and OcxrA-gfp;G418R+ as compared with background strain Δku70, respectively. P. oxalicum strains pre-grow in glucose medium for 24 h, and the harvested mycelia are transferred into Avicel medium and cultured for 48 h. Gene expression in the overexpression strain is normalized to the level of Δku70;hphR+. Data points show mean ± standard deviation. ** p < 0.01 and * p < 0.05 indicate significant differences between the overexpression strain and background strain, assessed by Student’s t-test. https://doi.org/10.1371/journal.pgen.1010867.s009 (TIF) S10 Fig. Autoactivation experiment of the PRMT2 in Saccharomyces cerevisiae. Serial dilutions of yeast Y2HGold cells carrying pGBKT7-prmt2 and pGBKT7, pGBKT7+pGADT7-p53, pGBKT7+pGADT7-Lam as controls were cultured on SD/-Trp, SD/-Trp/X-α-Gal and SD/-Trp/X-α-Gal/AbA at 30°C for 4 d. https://doi.org/10.1371/journal.pgen.1010867.s010 (TIF) S11 Fig. In vitro methylation assay of CxrAΔ1–60 by PRMT3. anti-MMA: mono methyl arginine antibody; anti-ADMA: asymmetric dimethyl arginine antibody; anti-DMA: dimethyl arginine antibody; SAM: S-adenosyl-methionine. “+” and “−” indicate the presence or absence of the test protein. https://doi.org/10.1371/journal.pgen.1010867.s011 (TIF) S12 Fig. Pearson’s correlation heatmap of Δprmt3;G418R+ and ΔcxrA;G418R+ transcriptomes compared with that of the background strain Δku70;hphR+, grown on Avicel. Total RNA was extracted from the mycelia of each strain after culture in Avicel medium for 24 h, following a transfer from glucose. https://doi.org/10.1371/journal.pgen.1010867.s012 (TIF) S1 Table. Primers used in this study. https://doi.org/10.1371/journal.pgen.1010867.s013 (DOCX) S1 Dataset. List of 997 differentially expressed genes in Δprmt3;G418R+ compared with the background strain Δku70;hphR+, grown on Avicel. https://doi.org/10.1371/journal.pgen.1010867.s014 (XLSX) S2 Dataset. List of 2552 differentially expressed genes in ΔcxrA;G418R+ compared with the background strain Δku70;hphR+, grown on Avicel. https://doi.org/10.1371/journal.pgen.1010867.s015 (XLSX) S3 Dataset. Data from Figs 1–5, S3, S8 and S9. https://doi.org/10.1371/journal.pgen.1010867.s016 (XLSX)

Kusmec, Aaron;Attigala, Lakshmi;Dai, Xiongtao;Srinivasan, Srikant;Yeh, Cheng-Ting “Eddy”;Schnable, Patrick S.

doi: 10.1371/journal.pgen.1010799pmid: 37410701

Introduction US maize yields increased substantially over the course of the 20th century [1] due to a combination of genetic improvement and improved agronomic practices that increased tolerance to higher planting densities [2]. Like other crops, maize grows best within specific temperature ranges. Exposure to excessive heat or cold alters normal physiological functions, leading to yield losses. Because global climate change increased and continues to increase the frequency and duration of extreme weather episodes [3–5], growth in global agricultural productivity has slowed since 1961 relative to a counterfactual model lacking climate change [6]. Furthermore, absent sufficient genetic adaptation to extreme temperatures and corresponding changes to agronomics, yield losses are expected to increase [7], leading to the possibility of synchronized shocks to global production systems [8]. Increases in global mean temperatures and alterations to precipitation patterns are expected to increase the incidence of two key, often co-occurring abiotic stresses: heat and drought. Together with soil properties, temperature and precipitation interact to determine water supply and demand dynamics throughout the growing season. Previous analyses have identified a strong, negative, non-linear effect of exposure to temperatures >29–30°C on maize [9–12]. Simulations using process-based crop growth models determined that temperature has a stronger effect on maize yields than precipitation [13] because higher temperatures drive increased plant and atmospheric water demand, overwhelming the positive or negative effects of altered precipitation on soil water availability. Subsequent work has demonstrated the importance of interactions between vapor pressure deficit (driven by temperature and humidity) and root-zone soil moisture for predicting maize yield anomalies [12,14,15]. Much attention has recently been paid to the adaptive potential of shifting planting dates to maintain historical crop cycle durations [16,17] coupled with irrigation and/or regional transfers of cultivars [18,19]. The goal of such a strategy is to leverage phenological variability to reduce exposure to heat and drought stress during critical developmental periods for maize, namely, anthesis and grain-fill. However, such a strategy would depend on the ability of plant breeders to leverage genetic variability for flowering time [explicitly recognized by [17]], which has a non-linear dependence on temperature [20]. Yield, as a measure of organismal fitness, is integrative [21], dependent on the contributions of multiple interacting biotic and abiotic factors that influence crop cycle duration in addition to other important, intermediate traits. While crop cycle duration adaptation will undoubtedly be a component of any future adaptative strategy, it cannot by itself insulate yields from the effects of higher average temperatures and the increased probability of extreme heat events. Such an adaptation is an example of an avoidance strategy whereby the stress is not encountered or reduced but does not necessarily address tolerance strategies whereby the plant endures encountered stress by reducing deleterious effects. Thus, improvements in the heat tolerance of maize hybrids should also be a target of future adaptive strategies. The possibility of realizing such adaptations depends on the presence of genetic variation for heat tolerance, which is a necessary condition for genetic gain in any trait, and the structure of the genetic covariances between different components of fitness. Concerns regarding the maintenance and exploitation of genetic variation have increased in the last two decades following the widespread adoption of genomic selection [22], which—depending on the mate-pairing schemes employed—alters the dynamics of genetic (co-)variances relative to phenotypic selection [23–26]. This is to say that the multivariate genetic architecture of present maize populations has been shaped by historical selection; it is from this base that breeding decisions will be made to adapt maize to a warming climate; and the dynamics of genetic (co-)variances have changed in the recent past. As such, it is important to understand the adaptive (or, possibly, maladaptive) trajectory of hybrid maize as a means for understanding the constraints within which we work at the present day. Despite annual increases in maize yields for the past 90 years, two studies suggest that concerns about the trajectory of hybrid maize heat tolerance are more than theoretical. The first study [27], using 100 years of yield reports for the US state of Indiana, documented temporal variation in the effect of cumulative degree days >29°C on maize yields with a peak near 1960, which coincides with the adoption of single-cross hybrids, followed by increasing susceptibility to heat stress. The second study [28] updated the models and data of [9] to predict maize yield anomalies in the eastern US during the 2012 heat wave using temperature effects estimated from 60 years of data. While one might expect that increases in maize yields have been partly due to increases in heat tolerance and that the models of [28] would therefore over-predict yield losses in 2012, they in fact under-predict actual yield losses, suggesting that modern maize hybrids are potentially less heat tolerant than historical varieties. However, both of these studies aggregate exposure to all temperatures >29°C, assuming that trends in heat tolerance are identical across a range of heat stress severities. Furthermore, the data used by these studies do not allow them to address the question of maladaptation of hybrid maize to high temperatures by separating the effects of genetics and agronomic practices. The studies cited above characterize the impacts of exposure to various temperatures on yields by combining national agricultural statistics (such as county-level yields reported by the United States Department of Agriculture, National Agricultural Statistics Service [USDA-NASS]) with econometric panel-data models [29,30]. Importantly, these aggregated data do not include information on the specific hybrids grown by individual farmers within a county, which are expected to be heterogeneous, and studies reliant on these data cannot, therefore, directly incorporate genetic variation into estimates of temperature responses as noted above. While the panel-data approach can capture the effects of global (i.e., genetic and agronomic) adaptation over time [31] [see also the long-differences approach of [32]], these studies are unable to specifically address the existence and nature of genetic adaptation. Attempts to estimate broad adaptive trends through temporal subsets [9] or weather-time interaction terms [12] typically fail to find statistically significant evidence for such trends. This failure to identify adaptive trends is probably due to the genetic heterogeneity of the observational units (typically counties) as studies that do incorporate explicit genetic information in wheat find evidence for genetic trends [33,34]. Stated differently, summary data measure the response at the level of the entire county without accounting for the effects of hybrids, which are treatments applied at the level of individual farmers’ fields, inappropriately conflating the observational and experimental units. Despite this, the panel data framework has several advantages that address important issues related to the control of spatial and temporal variation [35,36] that can impact the estimation of selection responses. Through the incorporation of location effects, this framework captures time-invariant but spatially variable unobserved effects such as soil quality. Through the incorporation of time effects (possibly stratified by larger geographical regions), this framework captures spatially invariant but temporally variable unobserved effects such as fertilizer usage or planting density. Inclusion of both dimensions (a two-way panel data model as opposed to cross-sectional [location only] or time-series [time only] analyses) is an effective means to reduce omitted variables bias and increase degrees of freedom for statistical analyses [30]. In the context of quantitative genetics, these are classified as common environmental effects and separated from genetic effects by the inclusion of additional random effects terms in the animal model [37]. Studies in wheat [33,34] and sorghum [38] have combined the panel-data approach with on-farm and variety testing data that includes information on cultivars to estimate cultivar-specific heat responses in these two crops, demonstrating either a trend toward increasing heat sensitivity (wheat) or no trend in heat sensitivity (sorghum). In maize, cross-sectional studies have identified recent increases in weather (particularly water availability) sensitivity [12,14] and long-term, complex trends in heat sensitivity, which has been increasing since 1960 [27], using aggregated data. The advantage of the wheat and sorghum studies compared to earlier panel data models is that the observational units are now identical with the experimental units: individual cultivars grown in experimental plots rather than genetically heterogeneous, geographical units. This permits the estimation of cultivar-specific coefficients on the weather variables that are conditioned on the time-(in)variant effects. However, these studies suffer from three limitations. First, the referenced studies rarely use data collected before 1985 [[27] is the sole exception] due to data availability [14,33,38] and/or the use of modern remote sensing technologies [12]. Thus, they fail to capture climate trends differing from those of the recent past [39], including the first half of the 20th century before climate change is estimated to have begun affecting agricultural productivity [6]. Second, the referenced studies use linear or piece-wise linear specifications for temperature response functions [with the exception of [14]]. Although some of these are more complex, admitting variable effects of heat at different periods of the growing season [33,34,38], they are still simplifications of complex physiological responses. The parametric assumptions imposed by these studies limit the flexibility and nature of potential non-linearities in the estimated temperature response functions, which can lead to misleading conclusions regarding genetic variation for such traits [40] and obscure non-intuitive responses to selection due to the existence of non-linear constraints [41,42]. Although [28] uses a cubic B-spline basis and temporally-variable temperature responses, their use of summary data does not allow the estimation of genetic trends. Third, the genetic variation introduced into estimates of the temperature response function in [33,34,38] is limited to responses to the most extreme temperatures. This provides a limited view of the genetic variation in temperature responses within the studied populations and limits inferences on selection for responses to all temperatures experienced during the growing season. Additionally, this can bias projections of future yield losses due to climate change by not allowing for the possibility that cultivars are differentially responsive to optimum temperatures. As done previously for wheat [33,34] and sorghum [38], we combine the two-way panel-data approach with records of crop variety testing programs operated by US land-grant universities. These programs began, for hybrid maize, in the 1930s and continue to the present day. They compare the yields of (pre-)commercial and public maize hybrids across multiple locations within individual US states to help farmers make informed decisions about which hybrids to plant based on performance over a short sample of the local climate. Records for these trials are publicly available; encompass both publicly and privately developed maize hybrids; include information on single-year yields for individual hybrids at each trial location; and cover multiple regimes of climate change across the 20th century. Here, we report on the collection and curation of public yield trial data and an associated climate dataset. The combined dataset covers 81 years of trials over four US states and includes 4,730 maize hybrids. In contrast to prior studies, we specify a functional linear mixed effects model using a B-spline basis that decreases constraints on the shape of the response and allows us to model hybrid-specific temperature response functions, capturing the genetic variance-covariance structure of responses across observed temperatures. This allows us to: (1) test for the effects of selection on temperature response functions over 81 years of hybrid maize breeding, and (2) investigate the modes of genetic variation in temperature responses. Results Historical data overview We collected trial results from four US states for 81 years (1934–2014; Fig 1A). Our dataset contains 172 unique trial locations comprising a total of 2,581 non-irrigated environments (location-year combinations); 4,730 hybrids; and 175,805 total observations. Average trial yields are generally higher than and moderately to highly correlated with the corresponding state averages (Fig 1B). Mean hybrid yields and standard deviations have increased over time within all states with the rate of increase of the former being greater, leading to a reduction in the coefficient of variation of yields over time in Illinois and Iowa and no trend in Kansas and Nebraska (S3 Fig). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 1. Characterization of the combined yield trial and climate datasets. (A) Geographical distribution of trials for 1934–2014. Counties are colored by 2021 average maize yields (bu/a) reported by USDA-NASS. White indicates no data. Trials are indicated by points, and the color indicates the number of trials occurring in that county (maximum 80 trials). Some trials fall within the bounds of an adjacent state but are not administered by that state. The map was generated using shapefiles distributed with the ‘maps’ R package (https://cran.r-project.org/package=maps). (B) Distribution of yields from all trials in each year stratified by state. Boxplots indicate the first and third quartiles and median yield in each year, and whiskers indicate 1.5 times the interquartile range. Solid lines indicate the average state yield for each year as reported by USDA-NASS. Pearson’s correlation (r2) between the average trial yield and state average yield within each state is reported in the insets. (C) Average exposure time for the growing season across all years and trials in 1°C temperature bins. The solid line indicates the average across all 2,581 trials, and dashed lines indicate the average for each state [colors correspond to panel (B)]. The background indicates exposure to temperatures below (cold stress), in, and above (heat stress) the optimal range for maize as typically defined for growing degree days (GDD, 8–30°C). (D) Cumulative distribution functions of average exposure time for the growing season as described in panel (C). Dots and horizontal lines indicate the proportion of the average growing season for which hybrids experience heat stress (>30°C). https://doi.org/10.1371/journal.pgen.1010799.g001 Trials were linked with historical weather records to derive the distribution of time exposed to temperatures in 1°C bins following [9,43] (S1 Fig). The average exposure distribution is slightly left-skewed with the majority of the growing season exposing maize to temperatures within the typical optimal range—8–30°C—as defined by GDD (Fig 1C; green region). During 1934–2014, these trials were exposed to 218±3.1 h/yr (mean ± s.e.; 6.38±0.09% of the growing season) of temperatures >30°C. Exposure to extreme temperatures varies by state (Fig 1C and 1D) and county (S4 Fig) with less exposure to heat stress in northern latitudes. Nevertheless, exposure to heat stress is well distributed across the various counties, years, and hybrids in our dataset with a minimum of 6,355 observations for the most extreme temperatures (>41°C) and at least 10,000 observations for all other temperatures (S5 Fig). Resistance to severe heat stress has declined We modeled the temperature responses of the 4,730 hybrids using a functional linear mixed effects regression model [44,45] where the temperature response function was composed of a fixed population mean function and random functions for each hybrid [see “Methods,” Eq (1); S2 Fig]. To reduce assumptions about the form of the temperature response functions, we used a cubic B-spline basis with three internal knots (S6 Fig) to model the temperature exposure distributions and temperature response functions [46] [see “Methods” for a description of an alternative parameterization using constant B-splines, equivalent to the step model of, e.g., [9,11]]. Potentially confounding factors for the estimation of temperature response functions are included as covariates in the model, including fixed effects of Julian planting date (S7 Fig) and season-total precipitation (S8 Fig); random effects of county to capture location-specific, time-invariant effects such as soil type (S9 Fig); and random regressions on year stratified by state to capture changes in agronomic practices (S10 Fig). We used block bootstrapping on years to control for spatial correlation within years and estimate 95% confidence bands (CBs) on the coefficient functions. The population mean coefficient function exhibits the same non-linear, negative effect of temperatures >30°C found by [9–11,27,28] but without specifying an optimal temperature range (Fig 2A and S11A Fig), and these negative effects are significant for temperatures ≥33°C (95% bootstrapped confidence level). The vertical distance between any two points on the curve indicates the percentage change in yield when substituting one hour of exposure to one temperature with one hour of exposure to another temperature. For example, substituting a one-hour exposure to 30°C with a one-hour exposure to 37°C predicts an approximately 0.21% reduction in yield. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 2. Quantification of directional selection on temperature response functions. Results shown in this figure are for exposure distributions and response functions parameterized as cubic B-splines (see “Methods” and S11 Fig for an alternative parameterization). Point estimates and confidence intervals/bands are based on 2,000 block bootstraps. (A) Population mean temperature response function for 4,730 maize hybrids. The solid line indicates the mean fixed effect coefficient function, and confidence bands at two confidence levels are shown. The vertical difference between any two points on the function indicates the percentage change in yield associated with substituting one hour of exposure at one temperature for another. (B) The left panel indicates the mean (dot) and 90% (thick line) and 95% (thin line) confidence intervals for selection on average yield (random hybrid intercepts, li). The right panel illustrates selection on the breeding values for the temperature response functions. Figure elements have the same meaning as in panel (A). (C) Mean correlation between average yield (random hybrid intercepts, li) and temperature response function coefficients (βih) for hybrids grouped by year of introduction. Labels indicate significance at the 90% (^) or 95% (*) confidence levels. (D) K-means clustering results for centered and scaled time series of weighted mean hybrid cohort temperature response function coefficients. Each curve represents the time series for response to a 1°C temperature bin smoothed by a cubic B-spline with seven internal knots. https://doi.org/10.1371/journal.pgen.1010799.g002 To reduce potential biases in the estimated strength and direction of selection [47], breeding values must be derived from a model that appropriately specifies common environmental effects [37], which can otherwise be absorbed into the estimated breeding values [47,48], leading to false positive genetic trends. Alternatively, common environmental effects can absorb genetic trends, reducing the power to detect selection [47]. Separation of these trends is further complicated when individuals are observed in a subset of years, leading to absorption of genetic by environmental trends [48]. The main output of the yield trials in our dataset are 2- and 3-year average yields to help farmers select stable, high-yielding hybrids. Consequently, 1,727/4,730 (39.5%) of hybrids are tested in only 2–3 years. However, this implies that the majority of hybrids are tested in at least four years and does not account for the fact that genetically-identical hybrids can be grown in multiple locations unlike the studies of selection in natural populations reviewed by [37,48]. The ability to separate genetic and environmental trends in such situations is dependent on the distribution of hybrids across years and the degree of genetic relatedness between pairs of hybrids [49] when both are considered as random effects. Although we lack pedigree and marker data for these hybrids, we predict the variances of contrasts between different hybrids or years under assumptions that would generate the largest theoretical variances [49,50] and show that these are, in general, small, indicating a good degree of connectivity even in the absence of pedigree data (S12 Fig). Directional selection on breeding values is then estimated by regressing cohort mean breeding values on time [51]. Additionally, tests of selection by regression are anticonservative due to the non-independence of breeding values in the mixed model [51]; thus, uncertainty in the estimated strength and direction of selection is assessed in this study using a block bootstrap as previously described. Our model estimates breeding values for both average yield (random hybrid intercepts) and temperature response function coefficients (see Eq (1), “Methods”) to capture the genetic components of increasing average maize yields and heat tolerance, respectively. To estimate selection on these breeding values, hybrids are assigned to the year in which they first appear in the dataset. A regression coefficient whose confidence interval excludes zero indicates that the mean breeding value has changed [47,51]. The estimated selection coefficients are depicted in Fig 2B (S11B Fig) for average yield (left sub-panel) and temperature response functions (right sub-panel) at the 90% and 95% confidence levels. First, we find evidence of directional selection on hybrid intercepts at a mean rate of 0.32% per year [95% CI (0.05, 0.71)], consistent with increases in per se yields during the 20th century [1]. Second, we find evidence at the 90% confidence level that selection has been acting to increase resistance to moderate heat stress caused by temperatures of 32–34°C over the previous 81 years. This agrees with numerous reports of improvements in drought-adaptive traits during the 20th century summarized by [2]; previous evidence for trends in heat tolerance using different datasets, models, and assumptions [27]; and recent, explicit selection for tolerance to these stresses in commercial breeding programs [52]. Third, this beneficial selection is counterbalanced by evidence for a decrease in resistance to severe heat stress caused by temperatures >38°C at a rate approximately one order of magnitude greater than the rate of improvement to moderate heat stress, consistent with the trends observed by [27] and the implications of the predictions made by [28]. The results using a constant B-spline basis are qualitatively similar, estimating trends in the same directions but with increased uncertainty (compare S11A, S11B and S13 Figs). Increases in sensitivity to severe heat stress despite favorable environmental trends Maize development is temperature-dependent [20], and the effects of different environmental factors on phenotype have been shown to vary across developmental periods in multiple species, including maize [28,53–59]. Thus, changes to phenology that shift the developmental exposure of heat stress—particularly to the susceptible flowering and grain-filling periods—may lead us to overestimate the negative effects of high temperatures and the strength and direction of any trends in breeding values for response coefficients, especially if exposure to stressful temperatures has also increased. While the trial records in our study lack flowering time data, we observe trends toward earlier planting dates in most of the 138/172 counties with at least four years of data across 1934–2014 (S14A Fig), which is associated with a small, but non-significant, negative effect on log-yield in both model specifications (S7 Fig). These trends, however, are associated with significant increases in exposure to optimal temperatures and decreases in exposure to stressful temperatures in most counties (S14B Fig). Despite rising temperatures at the global level [3], the US Corn Belt is at the center of a “warming bubble” [60,61], where average summer temperatures have decreased, leading to a (temporarily) more favorable growing environment for maize. We can examine the potential impacts of phenological shifts on the developmental timing of heat stress by using state-wide crop progress data collected by the USDA for 1982–2014. We used a logistic model following [62] to estimate the timing of developmental stages for each year’s crop and the method of [59] to estimate stage-specific exposure distributions for 74/172 counties with at least four years of data from 1982–2014. There were few significant trends toward increased heat stress exposure during the grain-filling period (S15 Fig) in agreement with the previous analysis. Overall, there is evidence that the growing environment has become more favorable to maize in our dataset. Because the tendency of the mixed model is to absorb genetic differences into environmental trends when individual genotypes are observed in a subset of years as is the case in our dataset [48], this strengthens the evidence for real genetic change underlying trends in average yields and moderate heat stress tolerance in addition to favorable environmental trends. Similarly, the observed genetic trend toward less severe heat stress tolerance is occurring in the presence of (a) favorable environmental trends and (b) increasing average yields, which may be an instance of cryptic evolution—genetic change masked by environmental change. (See Fig 3 of [47] for recommendations on the interpretation of such trends.) Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 3. Variance decomposition of hybrid trial yields. Results shown are based on 2,000 block bootstraps for the model where exposure distributions and response functions are parameterized by cubic B-splines (see “Methods” and S17 Fig for an alternative parameterization). (A) Percentage variance attributed to random effects. “Variety” indicates the genetic variance in hybrid intercepts; “Temperature,” the combined genetic variance for responses to all temperatures. Points indicate means and lines the 95% confidence intervals. (B) Genetic variance of breeding values for temperature response functions (N.B. the logarithmic scale). The solid line indicates the mean; dashed lines, the 95% confidence bands. (C) Genetic correlation function for the correlation between breeding values for temperature responses at different temperatures and the hybrid intercepts (bottom-most and leftmost row and column, respectively). “Int.” indicates the hybrid intercept. The bootstrapped mean function is depicted. https://doi.org/10.1371/journal.pgen.1010799.g003 Temperature responses are weakly and variably correlated with hybrid intercepts Because we observe evidence for selection in both hybrid intercepts and some regions of the temperature response function, a reasonable hypothesis is that change in the latter is the result of correlations to the former. We therefore calculated the correlation between the hybrid intercept and temperature coefficients within each hybrid cohort (13–123 hybrids/cohort; Fig 2C and S11C Fig). This revealed generally weak, non-significant, and fluctuating correlations across the time period of our dataset. This suggests that trends in temperature responses are unlikely to be an indirect response to selection on hybrid intercepts. Trends for responses to moderate and severe heat stress are complex but anticorrelated While we observe small directional trends across 1934–2014 for responses to temperatures >30°C, [27] previously observed a fluctuating trend for the combined effect of temperatures >29°C with a maximum near 1960. We focused on the region of the temperature response functions for heat stress (>30°C) and calculated weighted mean temperature response coefficients for each year. To identify common temporal patterns of selection across coefficients, we applied k-means clustering to the historical trajectories of these temperature-specific, mean coefficients. This identified three clusters (Fig 2D), two of which included non-zero domains of the selection function (Fig 2B). This revealed that temperature responses to 30–35°C and 37+ °C have been changing in anticorrelated fashion throughout the 20th century, leading to multiple local maxima and minima that alternate between the two primary clusters. Such a pattern suggests the influence of other evolutionary forces on heat stress responses in hybrid maize, including, but not limited to, genetic drift, changing selection pressures, indirect selection from other fitness components, and genetic tradeoffs internal to temperature response mechanisms. The trajectories in Fig 2D also suggested that the strength and/or direction of selection changed around 1975. We repeated the selection analyses, separating hybrids into those introduced before and after 1975 (S16 Fig). The estimated direction of selection did not change for the cubic B-spline model, although it did indicate a trend toward stronger selection post-1975. The constant B-spline model, however, now estimated statistically significant evidence for selection, especially for responses to temperatures >37°C, that switched direction after 1975. In particular, this model estimated a decrease in sensitivity to such temperature pre-1975 and an increase in sensitivity to such temperatures post-1975. This switch could provide a more fine-grained characterization of the trend observed by [27]. Sources of variation for yield and temperature response functions Considering the wide range of predicted outcomes for maize yields under climate change and ongoing plant breeding efforts to improve drought and heat tolerance, it is important to assess the potential for genetic adaptation to different temperatures by examining patterns of genetic covariation for temperature responses, which will constrain the directions in phenotypic space that are most genetically accessible. Variance for log-yield in this dataset is mostly attributable to the effects of changing agronomic practices over time and the interaction of genotype and environment captured by the functional regression on temperature exposure (Fig 3A). The amount of genetic variation for responses to different temperatures varies over three orders of magnitude in our historical dataset and is highest for temperatures >20°C (Fig 3B). This agrees with our earlier observations that directional selection on responses to extreme temperatures has been weak (Fig 2B, right panel) and likely not under significant indirect selection pressure from the much stronger selection on hybrid intercepts (Fig 2B [left panel] and C) during 1934–2014. Therefore, we would expect that if genetic variance for responses to stressful temperatures exists, it would likely be large relative to genetic variance for optimal temperatures, for example. However, we see that the amount of genetic variance for temperatures >20°C is relatively high, which suggests the presence of unused genetic variation for more beneficial responses to both optimal and stressful temperatures. However, selection acting on any continuous region of the temperature response function is also subject to genetic constraints between different regions of the function. The genetic correlation function describes the relative strength of the genetic covariance between responses to different temperature exposures and is the infinite-dimensional analogue of the genetic correlation matrix estimated in multivariate selection studies [40,63]. We observe strong, positive, local correlations between responses to similar temperatures along the diagonal (Fig 3C). Similar to the anticorrelation for time-series temperature response coefficients (Fig 2D), we also observe two negatively correlated regions corresponding to 33–39°C and 40+ °C (Fig 3C). These domains overlap—but are not identical with—the domains of the temperature response function experiencing different directions of selection (Fig 2B, right panel) and clusters of similar mean trajectories over time (Fig 2D) described above. Collectively, these suggest the existence of a genetic tradeoff for responses to moderate and severe heat stress (provisionally 30–35°C and 37+ °C, respectively). Although there are other, weaker possible tradeoffs suggested by the genetic correlation function (e.g., 0–4°C vs. 6–18°C, Fig 3C), the tradeoff suggested for extreme temperatures is quite strong. We do not believe that this is an artifact of the model. First, the alternative specification (constant B-spline basis) estimates correlations that are generally indistinguishable from zero except along the diagonal (S17C Fig). That pattern is most likely due to the lack of continuity conditions at the knots due to the low order of the basis functions. However, the strength of the negative correlations increases in regions corresponding to those also predicted to harbor a genetic tradeoff by the cubic B-spline basis. Second, analyses of multivariate genetic correlations typically identify a much smaller number of effective dimensions of genetic variation, indicating the presence of genetic constraints on the total genetic variance observed across all studied traits [42]. This pattern also seems to hold for function-valued (or infinite-dimensional) traits like the temperature response functions studied here [42]. Third, moderate-to-strong tradeoffs between regions of the domains of other function-valued traits have been observed in other species [40,41]. Complex modes of genetic variation for temperature response functions We used simple basis analysis (SBA) to obtain biologically interpretable modes of variation in the genetic covariance function [64,65] and the corresponding amounts of variance in these directions. Whereas principal functions analysis (PFA) identifies directions of maximal genetic variation, SBA optimizes a simplicity metric to calculate the genetic variation in predetermined directions in phenotypic space. Following the recommendations of [64], the first three SB functions for the genetic covariance function correspond to three classical biological modes of variation (Fig 4A) and capture 22.2% of the total genetic variance. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 4. Modes of genetic variation for temperature response functions. Results shown are based on 2,000 block bootstraps for the model where exposure distributions and response functions are parameterized by cubic B-splines (see “Methods” and S18 Fig for an alternative parameterization). (A) Loadings for the first three simple basis (SB) functions of the genetic covariance function for temperature responses with the genetic variance explained by each function. The form of each SB function is determined by the choice of a simplicity metric in contrast to principal function analysis (PFA), which identifies functional responses of greatest variance. Each SB function describes genetic variation in a biological interesting direction: SB1 quantifies overall performance (i.e., log-yield); SB2 quantifies the strength of a tradeoff between low and high temperatures; and SB3 quantifies a generalist-specialist tradeoff. The percentage of genetic variance associated with each SB function is the bootstrapped mean. (B) Loadings for the first three principal functions (PF) of the genetic covariance function. These three PFs account for at least 95% of the total genetic variance in temperature responses. Solid lines indicate the bootstrapped mean; dashed lines, the 95% confidence bands. The percentage of genetic variance associated with each PF is the bootstrapped mean. https://doi.org/10.1371/journal.pgen.1010799.g004 The first SB function is a horizontal line and measures variation for overall better performance (i.e., more positive responses to optimal temperatures and less negative responses to stressful temperatures) among hybrids. This direction accounts for 11.3% [(5.2, 18.0%), 95% bootstrapped confidence interval] of the total genetic variance. The second SB function measures a tradeoff between responses to temperatures less than or greater than 20°C—hybrids that are well-adapted to cold stress but not heat stress and vice versa. It captures 3.2% [(1.6, 5.6%), 95% CI] of the total genetic variance. Finally, the third SB function changes sign at 10 and 31°C, describing a generalist-specialist tradeoff, where hybrids that perform well at intermediate temperatures would be expected to perform poorly when exposed to extreme temperatures in either direction and vice versa. This direction accounts for 7.7% [(4.5, 11.5%), 95% CI] of the total genetic variance. SBA for the constant B-spline model accounts for a smaller proportion—12.8%—of the total genetic variance (S18A Fig). Overall, these three SB functions accounted for less than 25% of the total genetic variance. We therefore performed PFA to compare the directions of maximal genetic variance for temperature response functions with those calculated by SBA. The first three principal functions (PFs) of the genetic covariance function accounted for 96.5% of the total genetic variance with 82.7% [(65.6, 91.9%), 95% CI] of the total genetic variance in the direction of the first PF alone (Fig 4B). Unsurprisingly, the first PF captures the strong, alternating patterns of correlation between responses to temperatures >20°C (Fig 3C). The second and third PFs are less amenable to interpretation, especially because a large degree of uncertainty makes it difficult to draw conclusion about the relative signs of the loadings. But in any case, they account for only 13.7% of the total genetic variation. Historical data overview We collected trial results from four US states for 81 years (1934–2014; Fig 1A). Our dataset contains 172 unique trial locations comprising a total of 2,581 non-irrigated environments (location-year combinations); 4,730 hybrids; and 175,805 total observations. Average trial yields are generally higher than and moderately to highly correlated with the corresponding state averages (Fig 1B). Mean hybrid yields and standard deviations have increased over time within all states with the rate of increase of the former being greater, leading to a reduction in the coefficient of variation of yields over time in Illinois and Iowa and no trend in Kansas and Nebraska (S3 Fig). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 1. Characterization of the combined yield trial and climate datasets. (A) Geographical distribution of trials for 1934–2014. Counties are colored by 2021 average maize yields (bu/a) reported by USDA-NASS. White indicates no data. Trials are indicated by points, and the color indicates the number of trials occurring in that county (maximum 80 trials). Some trials fall within the bounds of an adjacent state but are not administered by that state. The map was generated using shapefiles distributed with the ‘maps’ R package (https://cran.r-project.org/package=maps). (B) Distribution of yields from all trials in each year stratified by state. Boxplots indicate the first and third quartiles and median yield in each year, and whiskers indicate 1.5 times the interquartile range. Solid lines indicate the average state yield for each year as reported by USDA-NASS. Pearson’s correlation (r2) between the average trial yield and state average yield within each state is reported in the insets. (C) Average exposure time for the growing season across all years and trials in 1°C temperature bins. The solid line indicates the average across all 2,581 trials, and dashed lines indicate the average for each state [colors correspond to panel (B)]. The background indicates exposure to temperatures below (cold stress), in, and above (heat stress) the optimal range for maize as typically defined for growing degree days (GDD, 8–30°C). (D) Cumulative distribution functions of average exposure time for the growing season as described in panel (C). Dots and horizontal lines indicate the proportion of the average growing season for which hybrids experience heat stress (>30°C). https://doi.org/10.1371/journal.pgen.1010799.g001 Trials were linked with historical weather records to derive the distribution of time exposed to temperatures in 1°C bins following [9,43] (S1 Fig). The average exposure distribution is slightly left-skewed with the majority of the growing season exposing maize to temperatures within the typical optimal range—8–30°C—as defined by GDD (Fig 1C; green region). During 1934–2014, these trials were exposed to 218±3.1 h/yr (mean ± s.e.; 6.38±0.09% of the growing season) of temperatures >30°C. Exposure to extreme temperatures varies by state (Fig 1C and 1D) and county (S4 Fig) with less exposure to heat stress in northern latitudes. Nevertheless, exposure to heat stress is well distributed across the various counties, years, and hybrids in our dataset with a minimum of 6,355 observations for the most extreme temperatures (>41°C) and at least 10,000 observations for all other temperatures (S5 Fig). Resistance to severe heat stress has declined We modeled the temperature responses of the 4,730 hybrids using a functional linear mixed effects regression model [44,45] where the temperature response function was composed of a fixed population mean function and random functions for each hybrid [see “Methods,” Eq (1); S2 Fig]. To reduce assumptions about the form of the temperature response functions, we used a cubic B-spline basis with three internal knots (S6 Fig) to model the temperature exposure distributions and temperature response functions [46] [see “Methods” for a description of an alternative parameterization using constant B-splines, equivalent to the step model of, e.g., [9,11]]. Potentially confounding factors for the estimation of temperature response functions are included as covariates in the model, including fixed effects of Julian planting date (S7 Fig) and season-total precipitation (S8 Fig); random effects of county to capture location-specific, time-invariant effects such as soil type (S9 Fig); and random regressions on year stratified by state to capture changes in agronomic practices (S10 Fig). We used block bootstrapping on years to control for spatial correlation within years and estimate 95% confidence bands (CBs) on the coefficient functions. The population mean coefficient function exhibits the same non-linear, negative effect of temperatures >30°C found by [9–11,27,28] but without specifying an optimal temperature range (Fig 2A and S11A Fig), and these negative effects are significant for temperatures ≥33°C (95% bootstrapped confidence level). The vertical distance between any two points on the curve indicates the percentage change in yield when substituting one hour of exposure to one temperature with one hour of exposure to another temperature. For example, substituting a one-hour exposure to 30°C with a one-hour exposure to 37°C predicts an approximately 0.21% reduction in yield. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 2. Quantification of directional selection on temperature response functions. Results shown in this figure are for exposure distributions and response functions parameterized as cubic B-splines (see “Methods” and S11 Fig for an alternative parameterization). Point estimates and confidence intervals/bands are based on 2,000 block bootstraps. (A) Population mean temperature response function for 4,730 maize hybrids. The solid line indicates the mean fixed effect coefficient function, and confidence bands at two confidence levels are shown. The vertical difference between any two points on the function indicates the percentage change in yield associated with substituting one hour of exposure at one temperature for another. (B) The left panel indicates the mean (dot) and 90% (thick line) and 95% (thin line) confidence intervals for selection on average yield (random hybrid intercepts, li). The right panel illustrates selection on the breeding values for the temperature response functions. Figure elements have the same meaning as in panel (A). (C) Mean correlation between average yield (random hybrid intercepts, li) and temperature response function coefficients (βih) for hybrids grouped by year of introduction. Labels indicate significance at the 90% (^) or 95% (*) confidence levels. (D) K-means clustering results for centered and scaled time series of weighted mean hybrid cohort temperature response function coefficients. Each curve represents the time series for response to a 1°C temperature bin smoothed by a cubic B-spline with seven internal knots. https://doi.org/10.1371/journal.pgen.1010799.g002 To reduce potential biases in the estimated strength and direction of selection [47], breeding values must be derived from a model that appropriately specifies common environmental effects [37], which can otherwise be absorbed into the estimated breeding values [47,48], leading to false positive genetic trends. Alternatively, common environmental effects can absorb genetic trends, reducing the power to detect selection [47]. Separation of these trends is further complicated when individuals are observed in a subset of years, leading to absorption of genetic by environmental trends [48]. The main output of the yield trials in our dataset are 2- and 3-year average yields to help farmers select stable, high-yielding hybrids. Consequently, 1,727/4,730 (39.5%) of hybrids are tested in only 2–3 years. However, this implies that the majority of hybrids are tested in at least four years and does not account for the fact that genetically-identical hybrids can be grown in multiple locations unlike the studies of selection in natural populations reviewed by [37,48]. The ability to separate genetic and environmental trends in such situations is dependent on the distribution of hybrids across years and the degree of genetic relatedness between pairs of hybrids [49] when both are considered as random effects. Although we lack pedigree and marker data for these hybrids, we predict the variances of contrasts between different hybrids or years under assumptions that would generate the largest theoretical variances [49,50] and show that these are, in general, small, indicating a good degree of connectivity even in the absence of pedigree data (S12 Fig). Directional selection on breeding values is then estimated by regressing cohort mean breeding values on time [51]. Additionally, tests of selection by regression are anticonservative due to the non-independence of breeding values in the mixed model [51]; thus, uncertainty in the estimated strength and direction of selection is assessed in this study using a block bootstrap as previously described. Our model estimates breeding values for both average yield (random hybrid intercepts) and temperature response function coefficients (see Eq (1), “Methods”) to capture the genetic components of increasing average maize yields and heat tolerance, respectively. To estimate selection on these breeding values, hybrids are assigned to the year in which they first appear in the dataset. A regression coefficient whose confidence interval excludes zero indicates that the mean breeding value has changed [47,51]. The estimated selection coefficients are depicted in Fig 2B (S11B Fig) for average yield (left sub-panel) and temperature response functions (right sub-panel) at the 90% and 95% confidence levels. First, we find evidence of directional selection on hybrid intercepts at a mean rate of 0.32% per year [95% CI (0.05, 0.71)], consistent with increases in per se yields during the 20th century [1]. Second, we find evidence at the 90% confidence level that selection has been acting to increase resistance to moderate heat stress caused by temperatures of 32–34°C over the previous 81 years. This agrees with numerous reports of improvements in drought-adaptive traits during the 20th century summarized by [2]; previous evidence for trends in heat tolerance using different datasets, models, and assumptions [27]; and recent, explicit selection for tolerance to these stresses in commercial breeding programs [52]. Third, this beneficial selection is counterbalanced by evidence for a decrease in resistance to severe heat stress caused by temperatures >38°C at a rate approximately one order of magnitude greater than the rate of improvement to moderate heat stress, consistent with the trends observed by [27] and the implications of the predictions made by [28]. The results using a constant B-spline basis are qualitatively similar, estimating trends in the same directions but with increased uncertainty (compare S11A, S11B and S13 Figs). Increases in sensitivity to severe heat stress despite favorable environmental trends Maize development is temperature-dependent [20], and the effects of different environmental factors on phenotype have been shown to vary across developmental periods in multiple species, including maize [28,53–59]. Thus, changes to phenology that shift the developmental exposure of heat stress—particularly to the susceptible flowering and grain-filling periods—may lead us to overestimate the negative effects of high temperatures and the strength and direction of any trends in breeding values for response coefficients, especially if exposure to stressful temperatures has also increased. While the trial records in our study lack flowering time data, we observe trends toward earlier planting dates in most of the 138/172 counties with at least four years of data across 1934–2014 (S14A Fig), which is associated with a small, but non-significant, negative effect on log-yield in both model specifications (S7 Fig). These trends, however, are associated with significant increases in exposure to optimal temperatures and decreases in exposure to stressful temperatures in most counties (S14B Fig). Despite rising temperatures at the global level [3], the US Corn Belt is at the center of a “warming bubble” [60,61], where average summer temperatures have decreased, leading to a (temporarily) more favorable growing environment for maize. We can examine the potential impacts of phenological shifts on the developmental timing of heat stress by using state-wide crop progress data collected by the USDA for 1982–2014. We used a logistic model following [62] to estimate the timing of developmental stages for each year’s crop and the method of [59] to estimate stage-specific exposure distributions for 74/172 counties with at least four years of data from 1982–2014. There were few significant trends toward increased heat stress exposure during the grain-filling period (S15 Fig) in agreement with the previous analysis. Overall, there is evidence that the growing environment has become more favorable to maize in our dataset. Because the tendency of the mixed model is to absorb genetic differences into environmental trends when individual genotypes are observed in a subset of years as is the case in our dataset [48], this strengthens the evidence for real genetic change underlying trends in average yields and moderate heat stress tolerance in addition to favorable environmental trends. Similarly, the observed genetic trend toward less severe heat stress tolerance is occurring in the presence of (a) favorable environmental trends and (b) increasing average yields, which may be an instance of cryptic evolution—genetic change masked by environmental change. (See Fig 3 of [47] for recommendations on the interpretation of such trends.) Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 3. Variance decomposition of hybrid trial yields. Results shown are based on 2,000 block bootstraps for the model where exposure distributions and response functions are parameterized by cubic B-splines (see “Methods” and S17 Fig for an alternative parameterization). (A) Percentage variance attributed to random effects. “Variety” indicates the genetic variance in hybrid intercepts; “Temperature,” the combined genetic variance for responses to all temperatures. Points indicate means and lines the 95% confidence intervals. (B) Genetic variance of breeding values for temperature response functions (N.B. the logarithmic scale). The solid line indicates the mean; dashed lines, the 95% confidence bands. (C) Genetic correlation function for the correlation between breeding values for temperature responses at different temperatures and the hybrid intercepts (bottom-most and leftmost row and column, respectively). “Int.” indicates the hybrid intercept. The bootstrapped mean function is depicted. https://doi.org/10.1371/journal.pgen.1010799.g003 Temperature responses are weakly and variably correlated with hybrid intercepts Because we observe evidence for selection in both hybrid intercepts and some regions of the temperature response function, a reasonable hypothesis is that change in the latter is the result of correlations to the former. We therefore calculated the correlation between the hybrid intercept and temperature coefficients within each hybrid cohort (13–123 hybrids/cohort; Fig 2C and S11C Fig). This revealed generally weak, non-significant, and fluctuating correlations across the time period of our dataset. This suggests that trends in temperature responses are unlikely to be an indirect response to selection on hybrid intercepts. Trends for responses to moderate and severe heat stress are complex but anticorrelated While we observe small directional trends across 1934–2014 for responses to temperatures >30°C, [27] previously observed a fluctuating trend for the combined effect of temperatures >29°C with a maximum near 1960. We focused on the region of the temperature response functions for heat stress (>30°C) and calculated weighted mean temperature response coefficients for each year. To identify common temporal patterns of selection across coefficients, we applied k-means clustering to the historical trajectories of these temperature-specific, mean coefficients. This identified three clusters (Fig 2D), two of which included non-zero domains of the selection function (Fig 2B). This revealed that temperature responses to 30–35°C and 37+ °C have been changing in anticorrelated fashion throughout the 20th century, leading to multiple local maxima and minima that alternate between the two primary clusters. Such a pattern suggests the influence of other evolutionary forces on heat stress responses in hybrid maize, including, but not limited to, genetic drift, changing selection pressures, indirect selection from other fitness components, and genetic tradeoffs internal to temperature response mechanisms. The trajectories in Fig 2D also suggested that the strength and/or direction of selection changed around 1975. We repeated the selection analyses, separating hybrids into those introduced before and after 1975 (S16 Fig). The estimated direction of selection did not change for the cubic B-spline model, although it did indicate a trend toward stronger selection post-1975. The constant B-spline model, however, now estimated statistically significant evidence for selection, especially for responses to temperatures >37°C, that switched direction after 1975. In particular, this model estimated a decrease in sensitivity to such temperature pre-1975 and an increase in sensitivity to such temperatures post-1975. This switch could provide a more fine-grained characterization of the trend observed by [27]. Sources of variation for yield and temperature response functions Considering the wide range of predicted outcomes for maize yields under climate change and ongoing plant breeding efforts to improve drought and heat tolerance, it is important to assess the potential for genetic adaptation to different temperatures by examining patterns of genetic covariation for temperature responses, which will constrain the directions in phenotypic space that are most genetically accessible. Variance for log-yield in this dataset is mostly attributable to the effects of changing agronomic practices over time and the interaction of genotype and environment captured by the functional regression on temperature exposure (Fig 3A). The amount of genetic variation for responses to different temperatures varies over three orders of magnitude in our historical dataset and is highest for temperatures >20°C (Fig 3B). This agrees with our earlier observations that directional selection on responses to extreme temperatures has been weak (Fig 2B, right panel) and likely not under significant indirect selection pressure from the much stronger selection on hybrid intercepts (Fig 2B [left panel] and C) during 1934–2014. Therefore, we would expect that if genetic variance for responses to stressful temperatures exists, it would likely be large relative to genetic variance for optimal temperatures, for example. However, we see that the amount of genetic variance for temperatures >20°C is relatively high, which suggests the presence of unused genetic variation for more beneficial responses to both optimal and stressful temperatures. However, selection acting on any continuous region of the temperature response function is also subject to genetic constraints between different regions of the function. The genetic correlation function describes the relative strength of the genetic covariance between responses to different temperature exposures and is the infinite-dimensional analogue of the genetic correlation matrix estimated in multivariate selection studies [40,63]. We observe strong, positive, local correlations between responses to similar temperatures along the diagonal (Fig 3C). Similar to the anticorrelation for time-series temperature response coefficients (Fig 2D), we also observe two negatively correlated regions corresponding to 33–39°C and 40+ °C (Fig 3C). These domains overlap—but are not identical with—the domains of the temperature response function experiencing different directions of selection (Fig 2B, right panel) and clusters of similar mean trajectories over time (Fig 2D) described above. Collectively, these suggest the existence of a genetic tradeoff for responses to moderate and severe heat stress (provisionally 30–35°C and 37+ °C, respectively). Although there are other, weaker possible tradeoffs suggested by the genetic correlation function (e.g., 0–4°C vs. 6–18°C, Fig 3C), the tradeoff suggested for extreme temperatures is quite strong. We do not believe that this is an artifact of the model. First, the alternative specification (constant B-spline basis) estimates correlations that are generally indistinguishable from zero except along the diagonal (S17C Fig). That pattern is most likely due to the lack of continuity conditions at the knots due to the low order of the basis functions. However, the strength of the negative correlations increases in regions corresponding to those also predicted to harbor a genetic tradeoff by the cubic B-spline basis. Second, analyses of multivariate genetic correlations typically identify a much smaller number of effective dimensions of genetic variation, indicating the presence of genetic constraints on the total genetic variance observed across all studied traits [42]. This pattern also seems to hold for function-valued (or infinite-dimensional) traits like the temperature response functions studied here [42]. Third, moderate-to-strong tradeoffs between regions of the domains of other function-valued traits have been observed in other species [40,41]. Complex modes of genetic variation for temperature response functions We used simple basis analysis (SBA) to obtain biologically interpretable modes of variation in the genetic covariance function [64,65] and the corresponding amounts of variance in these directions. Whereas principal functions analysis (PFA) identifies directions of maximal genetic variation, SBA optimizes a simplicity metric to calculate the genetic variation in predetermined directions in phenotypic space. Following the recommendations of [64], the first three SB functions for the genetic covariance function correspond to three classical biological modes of variation (Fig 4A) and capture 22.2% of the total genetic variance. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 4. Modes of genetic variation for temperature response functions. Results shown are based on 2,000 block bootstraps for the model where exposure distributions and response functions are parameterized by cubic B-splines (see “Methods” and S18 Fig for an alternative parameterization). (A) Loadings for the first three simple basis (SB) functions of the genetic covariance function for temperature responses with the genetic variance explained by each function. The form of each SB function is determined by the choice of a simplicity metric in contrast to principal function analysis (PFA), which identifies functional responses of greatest variance. Each SB function describes genetic variation in a biological interesting direction: SB1 quantifies overall performance (i.e., log-yield); SB2 quantifies the strength of a tradeoff between low and high temperatures; and SB3 quantifies a generalist-specialist tradeoff. The percentage of genetic variance associated with each SB function is the bootstrapped mean. (B) Loadings for the first three principal functions (PF) of the genetic covariance function. These three PFs account for at least 95% of the total genetic variance in temperature responses. Solid lines indicate the bootstrapped mean; dashed lines, the 95% confidence bands. The percentage of genetic variance associated with each PF is the bootstrapped mean. https://doi.org/10.1371/journal.pgen.1010799.g004 The first SB function is a horizontal line and measures variation for overall better performance (i.e., more positive responses to optimal temperatures and less negative responses to stressful temperatures) among hybrids. This direction accounts for 11.3% [(5.2, 18.0%), 95% bootstrapped confidence interval] of the total genetic variance. The second SB function measures a tradeoff between responses to temperatures less than or greater than 20°C—hybrids that are well-adapted to cold stress but not heat stress and vice versa. It captures 3.2% [(1.6, 5.6%), 95% CI] of the total genetic variance. Finally, the third SB function changes sign at 10 and 31°C, describing a generalist-specialist tradeoff, where hybrids that perform well at intermediate temperatures would be expected to perform poorly when exposed to extreme temperatures in either direction and vice versa. This direction accounts for 7.7% [(4.5, 11.5%), 95% CI] of the total genetic variance. SBA for the constant B-spline model accounts for a smaller proportion—12.8%—of the total genetic variance (S18A Fig). Overall, these three SB functions accounted for less than 25% of the total genetic variance. We therefore performed PFA to compare the directions of maximal genetic variance for temperature response functions with those calculated by SBA. The first three principal functions (PFs) of the genetic covariance function accounted for 96.5% of the total genetic variance with 82.7% [(65.6, 91.9%), 95% CI] of the total genetic variance in the direction of the first PF alone (Fig 4B). Unsurprisingly, the first PF captures the strong, alternating patterns of correlation between responses to temperatures >20°C (Fig 3C). The second and third PFs are less amenable to interpretation, especially because a large degree of uncertainty makes it difficult to draw conclusion about the relative signs of the loadings. But in any case, they account for only 13.7% of the total genetic variation. Discussion Maize yields have increased throughout the 20th century due to changes in both genetics and management practices [2]. Global climate change, however, threatens to impose severe yield penalties on maize over the course of the 21st century [9,11], leading to global food shortages [8]. Genetic adaptation to heat stress will be crucial to maintaining and increasing yields. Previous studies have often not addressed the question of genetic adaptation or found no evidence using proxy measures [9,27,31] [one exception is [66] but see the critique of [67]]. Large-scale panel studies in the US typically rely on aggregated county-level yields collected by USDA-NASS, where each data point represents a genetically heterogeneous mixture of hybrids that varies spatially and temporally. Thus, adaptive trends in genetics and management practices are confounded, leaving open the question of plant breeding’s realized and possible contributions to past and future temperature adaptation. As a step towards answering this question, we analyzed changes in temperature adaptation over 81 years of hybrid maize breeding (1934–2014) using data from public university extension service yield trials in the US Corn Belt and daily temperature records. In contrast to earlier studies, this dataset enables estimation of temperature response functions for individual hybrids. We extended the model of [9] to estimate temperature response functions for 4,730 maize hybrids. This two-way panel data model [36] separated genetic changes due to direct or indirect selection on temperature response (Fig 2B) from location effects common to all hybrids grown in a particular county (S9 Fig) and time-dependent trends in management (S10 Fig) common to all hybrids grown in a particular state and year [30,68], allowing us to profile the course of genetic adaptation to temperature [33,38]. We were able to show that maize has been, at least indirectly and partially, genetically adapted to heat stress. In particular, for temperatures 32–34°C, this selection was in an adaptive direction at the 90% confidence level. This result is consistent with evidence for improvements in general abiotic stress tolerance and leaf rolling under drought in newer hybrids [2]. However, we also found evidence for a maladaptive shift in responses to extremely stressful temperatures (>38°C), which can cause yield penalties through kernel abortion, dry matter partitioning changes, and oxidative stress among other processes [69,70]. This result agrees with the observations of [27], the implications of the predictions of [28], and could provide a partial explanation for the increasing sensitivity to vapor pressure deficit observed by [14]. Nevertheless, we found that genetic variation for extreme temperature responses is high, suggesting the potential for future genetic adaptation to such temperatures (Fig 3B). These shifts co-occurred with trends toward earlier planting dates and reduced exposure to heat stress (S14 and 15 Figs) as a consequence of the “warming bubble” over the US Midwest [60,61]. This phenomenon is attributed to agricultural intensification in the region during the twentieth century [71,72] and has produced a more favorable growing environment for maize [73]. Such a favorable environmental trend supports our conclusions of real and contrasting genetic trends for moderate and severe heat stress tolerance, following the recommendations of [47]. The maladaptive trend in severe heat stress tolerance may be a case of cryptic evolution, where environmental change masks genetic change, similar to the decrease in vapor pressure deficit tolerance observed in maize by [14]. This trend, in conjunction with the general increase in average maize yields, also suggests caution in the interpretation of the results of [17], who suggested that European maize yields may increase under climate change with appropriate phenological adaptation. However, as we have shown, phenological changes can mask genetic changes in other traits (here, heat tolerance) and lead to maladaptation under warmer environments despite “optimal” phenology. We also documented fluctuating patterns of selection on heat tolerance across time (Fig 2D), which supports the hypothesis that heat tolerance has been under, at best, indirect selection mediated by genetic correlations with other traits and/or other evolutionary pressures. This also suggested changes in the strength/direction of selection over time, and an analysis of selection on temperatures >30°C pre- and post-1975, suggests that this may indeed be the case (S16 Fig). However, our two parameterizations gave evidence for differing trends in this analysis and warrant further investigation. This inconsistent improvement in heat tolerance is counterintuitive when contrasted with the steady improvement in overall maize yields. However, this highlights the difference between yield per se and yield stability. While yield per se refers to the realized yield by a particular hybrid given some set of environmental conditions, yield stability refers to the variation around the per se value in a reference environment as environmental conditions vary. The heat tolerances (temperature response functions) that we estimated in this study are components of yield stability, not yield per se. Yield stability is a function of not only heat tolerance but resistance to drought, cold, pests, weeds, mineral availability, etc. Thus, as management practices change, shifts in limiting environmental factors also change selection pressures on their corresponding yield stability components, which can lead to counterintuitive trends such as increases in susceptibility to drought stress [12,14] or fluctuations in heat tolerance [27]. In short, [74] predicted such trends by noting that while modern maize hybrids are able to exploit favorable environmental conditions more fully, the accompanying greater year-to-year variability leads to larger decreases in yields (or, less yield stability) when environmental conditions are unfavorable. More recently, maize breeders have begun incorporating explicit selection on osmotic stress [52], which may modify these trends (similar to S16 Fig), although such hybrids were only widely introduced after 2014 (the final year of our dataset). The temperature response functions estimated here are function-valued traits. Such traits generally have strong genetic constraints [42], and prediction of selection responses is complicated because the function—and not the response to any individual temperature—is the target of selection [40]. SBA of the genetic covariance function reveals that, in aggregate, three classical-biological modes of variation [65] account for approximately 22% of the genetic variance in temperature response functions (Fig 4A). By contrast, PFA demonstrates that the majority of the genetic variance resides in a different genetic subspace that imposes more complex shapes and a major constraint on temperature response evolution (Fig 4B; cf. Fig 3C). Because the non-linearities common to function-valued traits can cause non-intuitive responses to selection [40], direct selection for heat tolerance would need to account for the highly constrained shape of the genetic covariance function along with the potential to alter responses to optimal temperatures in unfavorable directions as a consequence of those constraints. While we have leveraged a large dataset covering much spatial, temporal, and genetic variation, there are two avenues for further improvement. First, while we could estimate temperature response functions for individual hybrids, we lack pedigree or marker relationship data for these hybrids, which are generally unavailable. Incorporation of a relationship matrix into the estimation of the temperature response functions would increase the amount of information borrowing between them and help separate genetic from common environmental effects [37,48] by increasing the precision of the estimated differences between hybrids developed at different times [49]. This would be particularly useful for datasets like ours where each individual hybrid is observed in a subset of years (per the design of variety tests), which is a situation that increase the absorption of true genetic trends by environmental trends [48]. Second, we have assumed [following [9]] that the effects of temperature are identical over the growing season, which is manifestly false for maize. However, models making this assumption have proven robust in maize [9,27,43] and similar to estimates derived from process-based crop growth models [11]. Studies in wheat [33,34], sorghum [38], and maize [28,75] have demonstrated the utility of allowing time-varying temperature effects. Crucially, compared to our study, all of these studies reduce model complexity by some combination of constraining the functional form of the temperature response, admitting genetic variation for subsets of the temperature domain, or reducing the complexity of the exposure distribution. A model that combined the flexibility of ours with biologically appropriate, structural assumptions would be ideal but would also benefit from data on genetic relationships and observing each hybrid in a large number of environments. Ultimately, all of these studies draw on datasets collected as part of national statistics programs or designed for purposes other than estimation of heat tolerances per se and highlight the need for the development of datasets to study these questions explicitly. In conclusion, combined with reports on commercial breeding programs [52], our findings promote a tempered optimism for the capacity of plant breeders to improve heat adaptation in the 21st century. Efforts over the past 80 years have, at least indirectly, led to some adaptive trends but also reveal underlying genetic tradeoffs. Whether or not these tradeoffs can be overcome and further adaptation achieved fast enough to match or exceed the rate of climate change remains an open question in quantitative genetics terms. However, because a typical modern maize hybrid takes <7 years to develop and commercialize [76] and multiple emerging technologies may combine to further reduce generation times [77], it may be possible for breeders to adapt maize hybrids sufficiently rapidly to respond to changes in climate. Genetic analyses of heat [78] and drought [79] tolerance have identified numerous mechanisms by which plants can adaptively respond to heat and osmotic stress. Targeting these mechanisms has been shown to improve drought tolerance in maize hybrids [52]. This and continued research combined with emerging technologies are expected to contribute to further, more targeted adaptations in the future [52,80,81]. Materials and methods Yield data collection and curation Historical maize performance data were collected from print and online publications produced by university extension programs in the US states of Illinois (IL), Iowa (IA), Kansas (KS), and Nebraska (NE). More than 75 years of publications were collected for each of IL, IA, KS, and NE. Reports from IL, IA, and KS were converted to editable text files using optical character recognition software (ABBY Fine Reader, v12.1.3) and manually entered into Microsoft Excel spreadsheets. We used the Amazon Mechanical Turk (AMT) service to assist with conversion of the NE reports. Reports for 1948–2009 were obtained as low-quality scans of paper publications. A first campaign was run wherein AMT contractors were asked to enter data from the scans into Microsoft Excel spreadsheets manually. NE reports from 2010–2016 were of sufficient quality to be converted using ABBY Fine Reader and formatting errors were fixed by an automated Perl script. All NE spreadsheets were then submitted to a second AMT campaign to check the accuracy of data entry against the source PDFs. As a quality control measure, we introduced subtle changes to approximately 10% of the entries in a spreadsheet at random. A spreadsheet was considered checked if at least 70% of the introduced errors were identified and corrected. A bonus payment was disbursed if at least 90% of the introduced errors were identified. Checked spreadsheets from all states and years were then manually curated to maintain consistent formatting, spelling, and capitalization of brand and hybrid names. Reports from all years and states were combined and irrigated trials were removed. The remaining records were filtered (1) on the basis of the connectedness of counties and years and (2) to include varieties that had been extensively tested. For the connectedness filter, we required each county to be represented in at least two years of trials and each year to be represented in at least two counties. Note that this latter filter did not remove any data. We also required each hybrid to have a minimum of 15 records which had to have been collected from at least two counties and two years. Planting and harvesting dates were typically recorded for all trials. Trials missing either or both of these dates were excluded from further analysis. Due to regional variation in agronomic practices and weather, some trials were harvested as late as November or December. Therefore, calculations of temperature exposure and precipitation began on the reported planting date and ended on the earlier of the recorded harvest date or 30 September. Trial locations were reported by the US county containing the trial site and the town nearest to the trial site. Some more recent trials also include GPS coordinates for trial locations. For locations without GPS coordinates, county and town names were manually checked for accuracy. These names were used to query the geocoding service at openstreetmap.org to obtain approximate GPS coordinates for all trial locations. Weather data Weather data were collected and generated following [9,43] (S1 Fig). A dataset of daily predictions for minimum and maximum temperature and precipitation was constructed using the 4x4 km grid PRISM monthly dataset [82–84]. Only those PRISM grids that intersected with counties where a trial was located were retained. GPS coordinates for county boundaries were obtained from the US Census Bureau (https://www2.census.gov/geo/tiger/TIGER2017/COUNTY/tl_2017_us_county.zip). Monthly PRISM records were supplemented with denser, daily weather records from the US National Weather Service (NWS) Global Historical Climate Network (GHCN) stations. Stations were removed if they had moved more than 0.035° of latitude or longitude. PRISM grid cells were linked with the seven nearest GHCN stations having near continuous daily weather records for each year. A near continuous record is defined as a station having no more than three missing daily values for a given weather variable for the period January to September. Stations were linked separately for each variable such that a station could be linked for daily maximum temperature but excluded on the basis of excessive missing values for daily precipitation, for example. Missing daily values for GHCN stations were imputed from the daily values at the seven closest GHCN stations with non-missing values on that day and half-month fixed effects (e.g., 1–15 May, 16–31 May, 1–15 June, etc.). Ordinary least squares was used to predict missing values for minimum and maximum temperature, and Tobit regression for precipitation [implemented in the R package ‘VGAM’ [85]]. A type I Tobit model was used to account for the left-censoring at zero in precipitation records because negative precipitation is unrealistic. Regression models for each PRISM grid cell were constructed by regressing monthly PRISM values on monthly GHCN values derived from the seven closest stations in the imputed GHCN dataset for each variable as described above. These monthly relationships were then used to predict daily weather values at the resolution of the PRISM grid using the daily values from the seven closest GHCN stations. This produced a dataset of daily minimum and maximum temperature and precipitation covering the growing seasons of each trial at a 4x4 km resolution across all counties that contained trials. We then approximated the distribution of temperatures across each day using a sinusoidal model [86] and calculated the time in hours that each grid cell was exposed to 1°C temperature bins on each day throughout the growing season. The exposure time to each temperature bin for a trial was calculated by summing the time exposed to each temperature over the growing season in each PRISM grid cell and taking the average of all grid cells that intersected with the trial’s county. Precipitation was calculated in a similar fashion. Prior literature has used fixed growing seasons for estimating temperature exposure [e.g., [9,11]]. Because we have information on planting and harvesting dates for each trial in our dataset, we can more closely model the actual temperature distribution experienced by the hybrids [28]. Using a fixed growing season of April 24 (the date by which half of the trials were planted) to September 30 overestimates the length of the growing season by, on average, 374.4 h (15.6 d). Due to its greater biological relevance, we used a variable growing season and included a fixed regression coefficient on planting date in our model. The final weather dataset was determined by retaining trials experiencing 250–800 mm of precipitation (inclusive) across the growing season. The temperature exposure distributions were left-censored at -1°C and right-censored at 41°C. Statistical model To incorporate genetic variation into the estimation of the effect of temperature on maize yields, we extended the two-way panel data [36] model of [9,44] to a functional linear mixed effects model (FLMM; S2 Fig) [45]. In the following, we first present the common form of the FLMM before discussing two alternative specifications and the optimization of the representations for the functional coefficients and predictors. The common form of the FLMM considered here is (1) where yijst is the natural logarithm-transformed yield of the ith hybrid grown in the jth county of the sth state in the tth year; μ is the population mean; βd is the fixed effect of Julian planting date, djst, for the jth county of the sth state in the tth year; fμ (pjst) is the fixed effect coefficient function for the effect of precipitation in the jth county of the sth state in the tth year, where fμ(·) is modeled by a cubic B-spline basis with five basis functions; cj is the random effect of the jth county with distribution ; fs(t) is the random regression on time for the sth state evaluated in the tth year, where fs(·) is modeled by a second-degree, orthogonal polynomial basis and its coefficients have distribution ; li is the random hybrid effect with distribution ; gμ(h) is the fixed coefficient function for the effect of exposure to temperature h on yield; gi(h) is the random coefficient function for the response of the ith hybrid to temperature exposure (i.e., temperature response function) with distribution ; xjst(h) is a functional predictor returning the time (in hours) of exposure to temperature h for hybrids grown in the jth county of the sth state in the tth year; and εijst is an independent error term with distribution . The coefficient functions gμ(h) and gi(h) and the functional predictor xjst(h) were modeled by B-spline bases of equal dimension. Although temperature exposure distributions were left-censored at -1°C, the lower limit of integration in (1) is 0, reflecting the absorption of the lowest temperature bin by the intercept term. Alternative specifications of temperature functions. The integral in (1) captures potential non-linearities in the relationship between temperature exposure and log-yield. The representation(s) chosen for the functional components within the integral define a trade-off between the accuracy of the approximation of the integral and model complexity, particularly in the case of the gi(h)’s, where the variance-covariance matrix of the coefficients has M(M − 1)/2 unique entries and M is the dimension of the chosen representation. [9] introduced three alternative representations of the integral in (1) that have been used in subsequent literature to model this non-linear relationship. To facilitate comparisons between representations, we approximated the integral using B-spline bases of two different orders and dimensions: Step model: This model uses internal knots placed every 3°C to define a constant B-spline basis (order 1, M = 14). While this model is more flexible than the GDD model, its complexity is high. We use this dimension for consistency with prior work in maize using this model [9,11]. Polynomial model: This model uses a variable number of internal knots (see next section) to define a cubic B-spline basis (order 4). This is similar to the representation using Chebyshev polynomials by [9]. Depending on the dimension of the basis, it can express similar or greater flexibility to that of the step model but at a lower complexity. To the best of our knowledge, polynomial models have not been the preferred model specifications in past research. While the dimensions of the basis expansions for both models could be optimized, we consider that for the step model fixed by prior literature for comparative purposes. Optimization of the cubic B-spline basis. To choose a basis expansion for the functional predictor xjst(h) that optimizes the tradeoff between accurate representation of exposure times at different trials and subsequent model complexity, we devised a hybrid strategy [46,87]. For each trial, we sequentially fit regression splines using cubic B-splines as the basis functions and a number of internal knots, , where 43 is the number of discrete temperatures for which exposure times were estimated and subtraction of two accounts for the unique boundary knots. Addition of an internal knot at step s was accepted if the reduction in the mean squared error (MSE) of the fit met the relative criterion [87]: The parameter θ governs the tradeoff between goodness of fit and complexity. The mode of the optimal internal knot numbers for each trial was chosen as the optimum for the basis expansion of all trials (or median of the modes in the case of ties) [46]. This procedure selected three internal knots (M = 5) as optimal across the 2,581 trials, which was robust to variation in θ (S6 Fig). Model fitting and uncertainty estimation. To fit the FLMM, we used the ‘splines’ and ‘splines2’[88,89] packages to fit regression splines to the temperature exposure data and the ‘fda’ package [90] to compute the inner products of the functional coefficient and predictor bases. Basis expansions for fμ(·) were computed with the ‘splines’ package and for fs(·) using ‘poly()’. We fit the full model using the package ‘lme4’ [91]. To account for the effects of spatial covariation among trials within each year, we implemented a Bayesian block bootstrap using year as the blocking variable. We resampled the dataset 2,000 times and refit Eq (1) using both parameterizations described above. Point estimates in the text are the bootstrapped mean estimate, and confidence intervals/bands are the 95% bootstrapped confidence intervals/bands (unless otherwise stated) calculated using the ‘bca()’ function of the ‘coxed’ package [92]. Data connectivity When environmental effects are treated as random (Eq (1)), all possible contrasts between levels of a single factor are estimable, but the precision of the contrast depends on their distribution over levels of other factors [49]. We used the second suggested method of [49] to assess contrast variances between different common environmental effects in our model. This requires the estimation of a covariance matrix between the levels of each factor. We approximate these matrices as County: A spatial distance matrix using the Haversine distance between county centroids. Year: The expected covariance between years when modeled as a quadratic polynomial of time (as in Eq (1)). Trial (county-year): Kronecker product of the covariance matrices for county and year. Hybrid: A diagonal matrix with two along the diagonal. This simulates a variance-maximizing scenario, where the parents of each hybrid are assumed to be completely inbred [50], and hybrids are assumed to be unrelated. Quantifying selection Selection on temperature response functions was assessed by regressing the coefficients of the temperature response functions on calendar year for each resample to generate the bootstrap distribution for the selection function on temperature responses at 1°C intervals [51]. The temperature response function for each hybrid was assigned to the earliest year in which that hybrid appeared in the dataset. We used the number of observations for each hybrid as weights in the regression analysis to account for the unbalanced nature of the dataset. Confidence intervals that exclude zero are considered evidence of selection. To examine correlated changes over time in the responses to different temperatures, we calculated weighted mean coefficients for each hybrid cohort as defined above and performed k-means clustering on trajectories for responses to temperatures ≥30°C. The optimal number of clusters was chosen by plotting the between-cluster sum of squares for 2–6 clusters and applying the elbow method. Trends in phenology and exposure time To examine changes in phenology and temperature exposure over time, we took the subset of 138 counties with at least four years of data and regressed planting date (in Julian days), harvest date (in Julian days), growing season length (the difference between harvest and planting dates), and time exposed to each 1°C temperature bin (in days) on year. Because the effects of temperature vary with development, we also approximated stage-specific trends in temperature exposure using the methods of [59,62]. We obtained crop progress data for Illinois, Iowa, Kansas, and Nebraska in 1982–2014 from USDA-NASS (www.nass.usda.gov), removing years with fewer than five progress observations. Daily progress percentages were estimated by fitting a logistic model with the R function ‘nls()’[62]; if recorded data did not include 0% or 100%, we approximated the missing days using linear regression on the first (or last) recorded weekly interval. We could then estimate the fraction of the maize crop in a state-year combination at a particular stage by subtracting from its progress estimate the progress estimate of the following stage [59]. We then calculated weighted temperature exposure distributions for the 74 counties with at least four years of data from 1982–2014, where the weights were the fraction of the crop estimated to be in each stage on a given day. Quantification of trends was performed as described for the un-staged trial data. Decomposing the genetic covariance function The genetic covariance function describes genetic covariation between breeding values at different values of a functional covariate and is the infinite-dimensional analogue of the multivariate genetic covariance matrix [40,63]. For our study, the genetic covariance function describes the covariance between breeding values for the response to temperature h °C and breeding values for the response to temperature h′ °C. When h = h′, returns the variance of the breeding values for the response to the indicated temperature. Following [40,45,93], we calculated the genetic covariance function as where ϕ(·) are the B-spline basis functions for the appropriate specification evaluated at temperature h °C and h′ °C, and Dg is the estimated variance-covariance matrix of the temperature response functions as described in “Statistical model.” We performed simple basis analysis (SBA) [64] and principal functions analysis (PFA) using the ‘prinsimp’ package [94] following the recommendations of [65]. We examined the first three simple basis (SB) functions of the bootstrapped mean genetic covariance function using a simplicity metric based on first divided differences [equation 4.1 of [64]]. We also examined the minimum number of principal functions (PFs) required to explain ≥95% of the total genetic variance. We then performed SBA/PFA on each bootstrapped sample of the genetic covariance function to generate bootstrapped distributions for each function and its variance explained. Yield data collection and curation Historical maize performance data were collected from print and online publications produced by university extension programs in the US states of Illinois (IL), Iowa (IA), Kansas (KS), and Nebraska (NE). More than 75 years of publications were collected for each of IL, IA, KS, and NE. Reports from IL, IA, and KS were converted to editable text files using optical character recognition software (ABBY Fine Reader, v12.1.3) and manually entered into Microsoft Excel spreadsheets. We used the Amazon Mechanical Turk (AMT) service to assist with conversion of the NE reports. Reports for 1948–2009 were obtained as low-quality scans of paper publications. A first campaign was run wherein AMT contractors were asked to enter data from the scans into Microsoft Excel spreadsheets manually. NE reports from 2010–2016 were of sufficient quality to be converted using ABBY Fine Reader and formatting errors were fixed by an automated Perl script. All NE spreadsheets were then submitted to a second AMT campaign to check the accuracy of data entry against the source PDFs. As a quality control measure, we introduced subtle changes to approximately 10% of the entries in a spreadsheet at random. A spreadsheet was considered checked if at least 70% of the introduced errors were identified and corrected. A bonus payment was disbursed if at least 90% of the introduced errors were identified. Checked spreadsheets from all states and years were then manually curated to maintain consistent formatting, spelling, and capitalization of brand and hybrid names. Reports from all years and states were combined and irrigated trials were removed. The remaining records were filtered (1) on the basis of the connectedness of counties and years and (2) to include varieties that had been extensively tested. For the connectedness filter, we required each county to be represented in at least two years of trials and each year to be represented in at least two counties. Note that this latter filter did not remove any data. We also required each hybrid to have a minimum of 15 records which had to have been collected from at least two counties and two years. Planting and harvesting dates were typically recorded for all trials. Trials missing either or both of these dates were excluded from further analysis. Due to regional variation in agronomic practices and weather, some trials were harvested as late as November or December. Therefore, calculations of temperature exposure and precipitation began on the reported planting date and ended on the earlier of the recorded harvest date or 30 September. Trial locations were reported by the US county containing the trial site and the town nearest to the trial site. Some more recent trials also include GPS coordinates for trial locations. For locations without GPS coordinates, county and town names were manually checked for accuracy. These names were used to query the geocoding service at openstreetmap.org to obtain approximate GPS coordinates for all trial locations. Weather data Weather data were collected and generated following [9,43] (S1 Fig). A dataset of daily predictions for minimum and maximum temperature and precipitation was constructed using the 4x4 km grid PRISM monthly dataset [82–84]. Only those PRISM grids that intersected with counties where a trial was located were retained. GPS coordinates for county boundaries were obtained from the US Census Bureau (https://www2.census.gov/geo/tiger/TIGER2017/COUNTY/tl_2017_us_county.zip). Monthly PRISM records were supplemented with denser, daily weather records from the US National Weather Service (NWS) Global Historical Climate Network (GHCN) stations. Stations were removed if they had moved more than 0.035° of latitude or longitude. PRISM grid cells were linked with the seven nearest GHCN stations having near continuous daily weather records for each year. A near continuous record is defined as a station having no more than three missing daily values for a given weather variable for the period January to September. Stations were linked separately for each variable such that a station could be linked for daily maximum temperature but excluded on the basis of excessive missing values for daily precipitation, for example. Missing daily values for GHCN stations were imputed from the daily values at the seven closest GHCN stations with non-missing values on that day and half-month fixed effects (e.g., 1–15 May, 16–31 May, 1–15 June, etc.). Ordinary least squares was used to predict missing values for minimum and maximum temperature, and Tobit regression for precipitation [implemented in the R package ‘VGAM’ [85]]. A type I Tobit model was used to account for the left-censoring at zero in precipitation records because negative precipitation is unrealistic. Regression models for each PRISM grid cell were constructed by regressing monthly PRISM values on monthly GHCN values derived from the seven closest stations in the imputed GHCN dataset for each variable as described above. These monthly relationships were then used to predict daily weather values at the resolution of the PRISM grid using the daily values from the seven closest GHCN stations. This produced a dataset of daily minimum and maximum temperature and precipitation covering the growing seasons of each trial at a 4x4 km resolution across all counties that contained trials. We then approximated the distribution of temperatures across each day using a sinusoidal model [86] and calculated the time in hours that each grid cell was exposed to 1°C temperature bins on each day throughout the growing season. The exposure time to each temperature bin for a trial was calculated by summing the time exposed to each temperature over the growing season in each PRISM grid cell and taking the average of all grid cells that intersected with the trial’s county. Precipitation was calculated in a similar fashion. Prior literature has used fixed growing seasons for estimating temperature exposure [e.g., [9,11]]. Because we have information on planting and harvesting dates for each trial in our dataset, we can more closely model the actual temperature distribution experienced by the hybrids [28]. Using a fixed growing season of April 24 (the date by which half of the trials were planted) to September 30 overestimates the length of the growing season by, on average, 374.4 h (15.6 d). Due to its greater biological relevance, we used a variable growing season and included a fixed regression coefficient on planting date in our model. The final weather dataset was determined by retaining trials experiencing 250–800 mm of precipitation (inclusive) across the growing season. The temperature exposure distributions were left-censored at -1°C and right-censored at 41°C. Statistical model To incorporate genetic variation into the estimation of the effect of temperature on maize yields, we extended the two-way panel data [36] model of [9,44] to a functional linear mixed effects model (FLMM; S2 Fig) [45]. In the following, we first present the common form of the FLMM before discussing two alternative specifications and the optimization of the representations for the functional coefficients and predictors. The common form of the FLMM considered here is (1) where yijst is the natural logarithm-transformed yield of the ith hybrid grown in the jth county of the sth state in the tth year; μ is the population mean; βd is the fixed effect of Julian planting date, djst, for the jth county of the sth state in the tth year; fμ (pjst) is the fixed effect coefficient function for the effect of precipitation in the jth county of the sth state in the tth year, where fμ(·) is modeled by a cubic B-spline basis with five basis functions; cj is the random effect of the jth county with distribution ; fs(t) is the random regression on time for the sth state evaluated in the tth year, where fs(·) is modeled by a second-degree, orthogonal polynomial basis and its coefficients have distribution ; li is the random hybrid effect with distribution ; gμ(h) is the fixed coefficient function for the effect of exposure to temperature h on yield; gi(h) is the random coefficient function for the response of the ith hybrid to temperature exposure (i.e., temperature response function) with distribution ; xjst(h) is a functional predictor returning the time (in hours) of exposure to temperature h for hybrids grown in the jth county of the sth state in the tth year; and εijst is an independent error term with distribution . The coefficient functions gμ(h) and gi(h) and the functional predictor xjst(h) were modeled by B-spline bases of equal dimension. Although temperature exposure distributions were left-censored at -1°C, the lower limit of integration in (1) is 0, reflecting the absorption of the lowest temperature bin by the intercept term. Alternative specifications of temperature functions. The integral in (1) captures potential non-linearities in the relationship between temperature exposure and log-yield. The representation(s) chosen for the functional components within the integral define a trade-off between the accuracy of the approximation of the integral and model complexity, particularly in the case of the gi(h)’s, where the variance-covariance matrix of the coefficients has M(M − 1)/2 unique entries and M is the dimension of the chosen representation. [9] introduced three alternative representations of the integral in (1) that have been used in subsequent literature to model this non-linear relationship. To facilitate comparisons between representations, we approximated the integral using B-spline bases of two different orders and dimensions: Step model: This model uses internal knots placed every 3°C to define a constant B-spline basis (order 1, M = 14). While this model is more flexible than the GDD model, its complexity is high. We use this dimension for consistency with prior work in maize using this model [9,11]. Polynomial model: This model uses a variable number of internal knots (see next section) to define a cubic B-spline basis (order 4). This is similar to the representation using Chebyshev polynomials by [9]. Depending on the dimension of the basis, it can express similar or greater flexibility to that of the step model but at a lower complexity. To the best of our knowledge, polynomial models have not been the preferred model specifications in past research. While the dimensions of the basis expansions for both models could be optimized, we consider that for the step model fixed by prior literature for comparative purposes. Optimization of the cubic B-spline basis. To choose a basis expansion for the functional predictor xjst(h) that optimizes the tradeoff between accurate representation of exposure times at different trials and subsequent model complexity, we devised a hybrid strategy [46,87]. For each trial, we sequentially fit regression splines using cubic B-splines as the basis functions and a number of internal knots, , where 43 is the number of discrete temperatures for which exposure times were estimated and subtraction of two accounts for the unique boundary knots. Addition of an internal knot at step s was accepted if the reduction in the mean squared error (MSE) of the fit met the relative criterion [87]: The parameter θ governs the tradeoff between goodness of fit and complexity. The mode of the optimal internal knot numbers for each trial was chosen as the optimum for the basis expansion of all trials (or median of the modes in the case of ties) [46]. This procedure selected three internal knots (M = 5) as optimal across the 2,581 trials, which was robust to variation in θ (S6 Fig). Model fitting and uncertainty estimation. To fit the FLMM, we used the ‘splines’ and ‘splines2’[88,89] packages to fit regression splines to the temperature exposure data and the ‘fda’ package [90] to compute the inner products of the functional coefficient and predictor bases. Basis expansions for fμ(·) were computed with the ‘splines’ package and for fs(·) using ‘poly()’. We fit the full model using the package ‘lme4’ [91]. To account for the effects of spatial covariation among trials within each year, we implemented a Bayesian block bootstrap using year as the blocking variable. We resampled the dataset 2,000 times and refit Eq (1) using both parameterizations described above. Point estimates in the text are the bootstrapped mean estimate, and confidence intervals/bands are the 95% bootstrapped confidence intervals/bands (unless otherwise stated) calculated using the ‘bca()’ function of the ‘coxed’ package [92]. Alternative specifications of temperature functions. The integral in (1) captures potential non-linearities in the relationship between temperature exposure and log-yield. The representation(s) chosen for the functional components within the integral define a trade-off between the accuracy of the approximation of the integral and model complexity, particularly in the case of the gi(h)’s, where the variance-covariance matrix of the coefficients has M(M − 1)/2 unique entries and M is the dimension of the chosen representation. [9] introduced three alternative representations of the integral in (1) that have been used in subsequent literature to model this non-linear relationship. To facilitate comparisons between representations, we approximated the integral using B-spline bases of two different orders and dimensions: Step model: This model uses internal knots placed every 3°C to define a constant B-spline basis (order 1, M = 14). While this model is more flexible than the GDD model, its complexity is high. We use this dimension for consistency with prior work in maize using this model [9,11]. Polynomial model: This model uses a variable number of internal knots (see next section) to define a cubic B-spline basis (order 4). This is similar to the representation using Chebyshev polynomials by [9]. Depending on the dimension of the basis, it can express similar or greater flexibility to that of the step model but at a lower complexity. To the best of our knowledge, polynomial models have not been the preferred model specifications in past research. While the dimensions of the basis expansions for both models could be optimized, we consider that for the step model fixed by prior literature for comparative purposes. Optimization of the cubic B-spline basis. To choose a basis expansion for the functional predictor xjst(h) that optimizes the tradeoff between accurate representation of exposure times at different trials and subsequent model complexity, we devised a hybrid strategy [46,87]. For each trial, we sequentially fit regression splines using cubic B-splines as the basis functions and a number of internal knots, , where 43 is the number of discrete temperatures for which exposure times were estimated and subtraction of two accounts for the unique boundary knots. Addition of an internal knot at step s was accepted if the reduction in the mean squared error (MSE) of the fit met the relative criterion [87]: The parameter θ governs the tradeoff between goodness of fit and complexity. The mode of the optimal internal knot numbers for each trial was chosen as the optimum for the basis expansion of all trials (or median of the modes in the case of ties) [46]. This procedure selected three internal knots (M = 5) as optimal across the 2,581 trials, which was robust to variation in θ (S6 Fig). Model fitting and uncertainty estimation. To fit the FLMM, we used the ‘splines’ and ‘splines2’[88,89] packages to fit regression splines to the temperature exposure data and the ‘fda’ package [90] to compute the inner products of the functional coefficient and predictor bases. Basis expansions for fμ(·) were computed with the ‘splines’ package and for fs(·) using ‘poly()’. We fit the full model using the package ‘lme4’ [91]. To account for the effects of spatial covariation among trials within each year, we implemented a Bayesian block bootstrap using year as the blocking variable. We resampled the dataset 2,000 times and refit Eq (1) using both parameterizations described above. Point estimates in the text are the bootstrapped mean estimate, and confidence intervals/bands are the 95% bootstrapped confidence intervals/bands (unless otherwise stated) calculated using the ‘bca()’ function of the ‘coxed’ package [92]. Data connectivity When environmental effects are treated as random (Eq (1)), all possible contrasts between levels of a single factor are estimable, but the precision of the contrast depends on their distribution over levels of other factors [49]. We used the second suggested method of [49] to assess contrast variances between different common environmental effects in our model. This requires the estimation of a covariance matrix between the levels of each factor. We approximate these matrices as County: A spatial distance matrix using the Haversine distance between county centroids. Year: The expected covariance between years when modeled as a quadratic polynomial of time (as in Eq (1)). Trial (county-year): Kronecker product of the covariance matrices for county and year. Hybrid: A diagonal matrix with two along the diagonal. This simulates a variance-maximizing scenario, where the parents of each hybrid are assumed to be completely inbred [50], and hybrids are assumed to be unrelated. Quantifying selection Selection on temperature response functions was assessed by regressing the coefficients of the temperature response functions on calendar year for each resample to generate the bootstrap distribution for the selection function on temperature responses at 1°C intervals [51]. The temperature response function for each hybrid was assigned to the earliest year in which that hybrid appeared in the dataset. We used the number of observations for each hybrid as weights in the regression analysis to account for the unbalanced nature of the dataset. Confidence intervals that exclude zero are considered evidence of selection. To examine correlated changes over time in the responses to different temperatures, we calculated weighted mean coefficients for each hybrid cohort as defined above and performed k-means clustering on trajectories for responses to temperatures ≥30°C. The optimal number of clusters was chosen by plotting the between-cluster sum of squares for 2–6 clusters and applying the elbow method. Trends in phenology and exposure time To examine changes in phenology and temperature exposure over time, we took the subset of 138 counties with at least four years of data and regressed planting date (in Julian days), harvest date (in Julian days), growing season length (the difference between harvest and planting dates), and time exposed to each 1°C temperature bin (in days) on year. Because the effects of temperature vary with development, we also approximated stage-specific trends in temperature exposure using the methods of [59,62]. We obtained crop progress data for Illinois, Iowa, Kansas, and Nebraska in 1982–2014 from USDA-NASS (www.nass.usda.gov), removing years with fewer than five progress observations. Daily progress percentages were estimated by fitting a logistic model with the R function ‘nls()’[62]; if recorded data did not include 0% or 100%, we approximated the missing days using linear regression on the first (or last) recorded weekly interval. We could then estimate the fraction of the maize crop in a state-year combination at a particular stage by subtracting from its progress estimate the progress estimate of the following stage [59]. We then calculated weighted temperature exposure distributions for the 74 counties with at least four years of data from 1982–2014, where the weights were the fraction of the crop estimated to be in each stage on a given day. Quantification of trends was performed as described for the un-staged trial data. Decomposing the genetic covariance function The genetic covariance function describes genetic covariation between breeding values at different values of a functional covariate and is the infinite-dimensional analogue of the multivariate genetic covariance matrix [40,63]. For our study, the genetic covariance function describes the covariance between breeding values for the response to temperature h °C and breeding values for the response to temperature h′ °C. When h = h′, returns the variance of the breeding values for the response to the indicated temperature. Following [40,45,93], we calculated the genetic covariance function as where ϕ(·) are the B-spline basis functions for the appropriate specification evaluated at temperature h °C and h′ °C, and Dg is the estimated variance-covariance matrix of the temperature response functions as described in “Statistical model.” We performed simple basis analysis (SBA) [64] and principal functions analysis (PFA) using the ‘prinsimp’ package [94] following the recommendations of [65]. We examined the first three simple basis (SB) functions of the bootstrapped mean genetic covariance function using a simplicity metric based on first divided differences [equation 4.1 of [64]]. We also examined the minimum number of principal functions (PFs) required to explain ≥95% of the total genetic variance. We then performed SBA/PFA on each bootstrapped sample of the genetic covariance function to generate bootstrapped distributions for each function and its variance explained. Supporting information S1 Fig. Workflow for construction of the historical weather dataset. Time exposed to 1°C temperature bins and season-total precipitation were estimated for historical trials by combining daily weather records from the Global Historical Climate Network (GHCN) and PRISM project. Full details are described in subsection “Weather data” of the “Methods.” https://doi.org/10.1371/journal.pgen.1010799.s001 (PDF) S2 Fig. Modeling and analysis workflow. The relationships between variables in the data, the functional linear mixed effects model, and subsequent analyses are shown along with references to main text and supporting figures for results. For full details on the notation of the model, see “Statistical model” in “Methods.” Details of the different analyses can be found in the appropriate subsections of the “Methods.” https://doi.org/10.1371/journal.pgen.1010799.s002 (PDF) S3 Fig. Temporal trends in yield and yield variability by state for 1934–2014. Each point represents the indicated summary statistic of hybrid yields (bu/a) grown in all trials in that state and year. Best fit regression lines of the indicated statistic on time, 95% confidence bands, and regression equations are shown in each panel. Mean yields and yield standard deviations are increasing across time in all states. However, mean yields are increasing at least as quickly as yield standard deviations, leading to a decrease or no change in the coefficient of variation of yield across time. https://doi.org/10.1371/journal.pgen.1010799.s003 (PNG) S4 Fig. Average exposure to heat stress by county for 1934–2014. Average percentage of the total growing season (recorded planting date to the earlier of the recorded harvest date or 30 September) hybrids were exposed to temperatures >30°C in each county that contained at least two yield trials for 1934–2014. The map was generated using the ‘maps’ R package (https://cran.r-project.org/package=maps). https://doi.org/10.1371/journal.pgen.1010799.s004 (PNG) S5 Fig. Distribution of exposure to different temperatures across treatment factors. Each point indicates the number of levels of the factor indicated by the y-axis that were exposed to the temperature indicated on the x-axis. The maximum number of factor levels is given in the y-axis label. https://doi.org/10.1371/journal.pgen.1010799.s005 (PNG) S6 Fig. Optimization of the cubic B-spline basis for exposure distributions. The optimal number of knots (boundary plus internal) for a cubic B-spline basis was determined for each trial (n = 2,581) by iteratively adding knots placed at quantiles until the mean squared error was not reduced by a predetermined proportion (θ). “NA” indicates that the threshold was not reached by the maximum number of 20 knots. “*” indicates the mode of the optimal numbers of knots. https://doi.org/10.1371/journal.pgen.1010799.s006 (PNG) S7 Fig. Effect of planting date. Mean effect of Julian planting date on percentage yield for models with differing representations of temperature effects. Neither estimate is significantly different at the 95% (thick line) or 90% (thin line) confidence level. Estimates are based on 2,000 block bootstraps. https://doi.org/10.1371/journal.pgen.1010799.s007 (PNG) S8 Fig. Effect of precipitation on yield. Effects of precipitation on percentage yield for models with differing representations of temperature effects. Solid lines indicate the mean and shaded areas 95% confidence bands for 2,000 block bootstraps. https://doi.org/10.1371/journal.pgen.1010799.s008 (PNG) S9 Fig. Spatial distribution and magnitude of location (county) effects. (A) Mean percentage effect of each county from the cubic B-spline model. The map was generated using the ‘maps’ R package (https://cran.r-project.org/package=maps). (B) Comparison of mean location effects in models with differing representations of temperature effects. The dashed red line indicates equality. In both panels, the means of 2,000 block bootstraps are shown. https://doi.org/10.1371/journal.pgen.1010799.s009 (PNG) S10 Fig. Year effects on percentage yield. Each panel depicts the effect of year on percentage yield for the indicated state relative to 1981, the mean trial year in the dataset. Effects were estimated as the random regression of natural log-transformed yield (bu/a) on an orthogonal, quadratic polynomial of trial year. Solid lines indicate the mean and shaded regions the 95% confidence bands of 2,000 block bootstraps. https://doi.org/10.1371/journal.pgen.1010799.s010 (PNG) S11 Fig. Quantification of directional selection on temperature response functions. Results shown in this figure are for exposure distributions and response functions parameterized as constant B-splines (see “Methods” and Fig 2 for an alternative parameterization). Point estimates and confidence intervals/bands are based on 2,000 block bootstraps. (A) Population mean temperature response function for 4,730 maize hybrids. The solid line indicates the mean fixed effect coefficient function, and confidence bands at two confidence levels are shown. The vertical difference between any two points on the function indicates the percentage change in yield associated with substituting one-hour of exposure within one 3°C temperature bin for another. (B) The left panel indicates the mean (dot) and 90% (thick line) and 95% (thin line) confidence intervals for selection on the random hybrid intercepts. The right panel illustrates the selection function on breeding values for the temperature response functions. Figure elements have the same meaning as in panel (A). (C) Mean correlation between random hybrid intercepts (li) and temperature response function coefficients (βih) for hybrids grouped by year of introduction. Labels indicate significance at the 90% (^) or 95% (*) confidence levels. (D) Centered and scaled time series of weighted mean hybrid cohort temperature response function coefficients. Each curve represents the time series for response to a 3°C temperature bin smoothed by a cubic B-spline with seven internal knots. https://doi.org/10.1371/journal.pgen.1010799.s011 (PNG) S12 Fig. Predicted contrast variances for the difference between two levels of one factor when distributed over levels of a second factor. Each panel shows the distribution of predicted contrast variances for the difference between two levels of a factor when distributed over the levels of the other indicated factor. For example, the left side of panel A shows the contrast variances of counties when distributed over hybrids. The bin width is 0.05 in panels A-C and 0.1 in panel D. There are 172 levels of county; 81 levels of year; 4,730 levels of hybrid; and 2,581 levels of trial. See “Methods” for more details. https://doi.org/10.1371/journal.pgen.1010799.s012 (PNG) S13 Fig. Evidence for directional selection on temperature response functions at various confidence levels. For the estimation of directional selection, see “Methods”. The direction is the sign of the mean slope of 2,000 block bootstraps. Significance is defined as the exclusion of zero by the (100 × α)% confidence interval of 2,000 block bootstraps as indicated on the y-axis. https://doi.org/10.1371/journal.pgen.1010799.s013 (PNG) S14 Fig. Phenological and temperature exposure trends by state for 1934–2014. (A) Trends in Julian planting and harvest dates and growing season length by county. Only counties with at least four years of data are shown. Color indicates the sign of the slope of a linear regression of the indicated variable on year. Solid tiles indicate significance at an uncorrected p < 0.05 level. (B) Trends in temperature exposure time (in days) by county. Figure elements are the same as in panel A. Vertical dashed lines indicate the optimal temperature bounds for growing degree day calculations (10 and 30°C) and the moderate/severe heat stress breakpoint (36°C) identified in the main text. https://doi.org/10.1371/journal.pgen.1010799.s014 (PNG) S15 Fig. Trends in stage-wise temperature exposure by state for 1982–2014. Only counties with at least four years of data are shown. Color indicates the sign of the slope of a linear regression of the indicated variable on year. Solid tiles indicate significance at an uncorrected p < 0.05 level. Vertical dashed lines indicate the optimal temperature bounds for growing degree day calculations (10 and 30°C) and the moderate/severe heat stress breakpoint (36°C) identified in the main text. Stages are defined as: “Vegetative” = planting to silking; “Early grain fill” = silking to dough; “Late grain fill” = dough to maturity; “Drydown” = maturity to harvest. https://doi.org/10.1371/journal.pgen.1010799.s015 (PNG) S16 Fig. Directional selection on temperature response functions before and after 1975. The strength (cubic model) and direction (constant model) of selection differ between hybrids introduced before and after 1975. Results for the cubic B-spline parameterization are consistent in direction with those given in Fig 2B (right panel). Results for the constant B-spline parameterization show changes in direction for the different time periods and increased statistical significance compared with those given in S8B Fig (right panel). Solid lines indicate the mean strength and direction of selection, and dashed lines indicate the 95% confidence bands. Point estimates and confidence bands are based on 2,000 block bootstraps. https://doi.org/10.1371/journal.pgen.1010799.s016 (PNG) S17 Fig. Variance decomposition of hybrid trial yields. Results shown are based on 2,000 block bootstraps for the model where exposure distributions and response functions are parameterized by constant B-splines (see “Methods” and Fig 3 for an alternative parameterization). (A) Percentage variance attributed to random effects. “Variety” indicates the genetic variance in hybrid intercepts; “Temperature,” the combined genetic variance for responses to all temperatures. Points indicate means and lines the 95% confidence intervals. (B) Genetic variance of breeding values for temperature response functions (N.B. the logarithmic scale). The solid line indicates the mean; dashed lines, the 95% confidence bands. (C) Genetic correlation function for the correlation between breeding values for temperature responses at different temperatures and the hybrid intercepts (bottom-most and leftmost row and column, respectively). The bootstrapped mean function is depicted. https://doi.org/10.1371/journal.pgen.1010799.s017 (PNG) S18 Fig. Modes of genetic variation temperature response functions. Results shown are based on 2,000 block bootstraps for the model where exposure distributions and response functions are parameterized by constant B-splines (see “Methods” and Fig 4 for an alternative parameterization). (A) Loadings for the first three simple basis (SB) functions of the genetic covariance function for temperature responses with the genetic variance explained by each function. The form of each SB function is determined by the choice of a simplicity metric in contrast to principal function analysis (PFA), which identifies functional responses of greatest variance. Each SB function describes genetic variation in a biological interesting direction: SB1 quantifies overall performance (i.e., log-yield); SB2 quantifies the strength of a tradeoff between low and high temperatures; and SB3 quantifies a generalist-specialist tradeoff. The percentage of genetic variance associated with each SB function is the bootstrapped mean. (B) Loadings for the first three principal functions (PF) of the genetic covariance function. The first three PFs are shown, but the first nine PFs are required to account for at least 95% of the total genetic variance in temperature responses. Solid lines indicate the bootstrapped mean; dashed lines, the 95% confidence bands. The percentage of genetic variance associated with each PF is the bootstrapped mean. https://doi.org/10.1371/journal.pgen.1010799.s018 (PNG) S1 Table. Geographical distribution of trial locations. Data underlying Fig 1A. Columns include American National Standards Institute (ANSI) county code, longitude (decimal), latitude (decimal), and total number of trials conducted in that county. https://doi.org/10.1371/journal.pgen.1010799.s019 (CSV) S2 Table. Coefficients for the fixed effect of temperature on log-yield. Data underlying Fig 2A. Columns include temperature (°C), bootstrap, and estimate (% yield/h exposure). https://doi.org/10.1371/journal.pgen.1010799.s020 (CSV) S3 Table. Slopes for the regression of random hybrid intercepts on time. Data underlying Fig 2B, left panel. Columns include temperature (“Intercept” to distinguish it from the data in S4 Table), bootstrap, and regression coefficient (% yield/yr). https://doi.org/10.1371/journal.pgen.1010799.s021 (CSV) S4 Table. Slopes for the regression of random hybrid temperature coefficients on time. Data underlying Fig 2B, right panel. Columns include temperature (°C), bootstrap, and regression coefficient (% yield/h exposure/yr). https://doi.org/10.1371/journal.pgen.1010799.s022 (CSV) S5 Table. Correlations between hybrid intercepts and temperature coefficients stratified by year. Data underlying Fig 2C. Hybrids were assigned to the first year in which they appeared in the dataset. Columns include bootstrap, temperature (°C), year, and Pearson’s correlation coefficient. https://doi.org/10.1371/journal.pgen.1010799.s023 (GZ) S6 Table. k-means clustering of cohort mean temperature coefficients. Data underlying Fig 2D. Cohort means are the weighted average temperature coefficient for all hybrids that first appeared in the dataset in the indicated year. Only temperature bins 30 to >41°C inclusive are included. Columns include year, temperature (°C), cohort mean coefficient, and cluster. https://doi.org/10.1371/journal.pgen.1010799.s024 (CSV) S7 Table. Variance components. Data underlying Fig 3A. Columns include bootstrap, model component, and the estimated variance. https://doi.org/10.1371/journal.pgen.1010799.s025 (CSV) S8 Table. Temperature variance function coefficients. Data underlying Fig 3B. Columns include temperature (°C), bootstrap, and estimated variance. https://doi.org/10.1371/journal.pgen.1010799.s026 (CSV) S9 Table. Genetic correlation function for temperature. Data underlying Fig 3C. Columns include bootstrap, x-axis temperature (°C or “Intercept”), y-axis temperature (°C or “Intercept”), and Pearson’s correlation coefficient. https://doi.org/10.1371/journal.pgen.1010799.s027 (GZ) S10 Table. Variance explained by simple basis functions. Data underlying Fig 4A. Columns include bootstrap, simple basis function, and proportion of variance explained. https://doi.org/10.1371/journal.pgen.1010799.s028 (CSV) S11 Table. Principal function loadings. Data underlying Fig 4B. Columns include bootstrap, principal function, temperature (°C), and principal function loading. https://doi.org/10.1371/journal.pgen.1010799.s029 (CSV) S12 Table. Variance explained by principal functions. Data underlying Fig 4B. Columns include bootstrap, principal function, and proportion of variance explained. https://doi.org/10.1371/journal.pgen.1010799.s030 (CSV) Acknowledgments We gratefully acknowledge the efforts of the multiple university libraries that preserved the hybrid maize yield trial data for decades. Furthermore, this project would not have been possible without the combined efforts of numerous people who located, scanned, entered, and curated university extension service publications containing these data. We specifically thank Iowa State University Parks Memorial Library staff members Christopher Anderson, Michael Bobb, Olivia Garrison, Dawn Mick, Lorrie Pellack, and Kathy Thorson who helped search for reports and who, when necessary, contacted other university libraries to locate copies of these reports. We also thank Megan O’Donnell, Head of Research Data Services at Parks Memorial Library, for advice and assistance preparing the data for sharing and for navigating the complexities of copyrights. We thank Schnable Lab staff member Daniel Bade and undergraduate assistants who assisted with data entry and curation: Emily Adelizzi, Leeann Aguilar, David Aguirre, Sara Bazyn, Hannah Bellows, Abby Bravard, Chanelle Chimezie, Lauren Docherty, Alexander Donelson, Charity Elijah, Callie Feaker, Amal Hazura, Emma Kirkpatrick, Cameron Lahn, Stephanie Lee, Michael Ongie, Daniel Russel, Farshad Sadr, Kristina Sasse, Leo Savage, Abby Schmitt, Gillian Suhre, Tyler Sward, Simrita Varma, Trevor Weiss, and Zach Weiss. Our colleague Dan Nettleton provided valuable statistical advice.

Humphries, Ashley C.;Molina-Pelayo, Claudia;Sil, Parijat;Hazelett, C. Clayton;Devenport, Danelle;Mlodzik, Marek

doi: 10.1371/journal.pgen.1010849pmid: 37463168

Introduction Cellular polarization is critical for the morphogenesis and function of organs and most tissues during development, with perturbation of cellular polarity and tissue organization implicated in numerous diseases. In particular, epithelial cells can be polarized in two axes: the ubiquitous epithelial polarity in the apical-basal axis, and polarity in the plane of the epithelium, referred to as planar cell polarity (PCP) (reviewed in [1–7]). Although PCP was initially discovered in epithelia, the cellular mechanisms controlling PCP are detected in many other cell types, including migratory mesenchymal cells or neurons (reviewed in [3,8,9]. In this study we focus on PCP in epithelia. PCP establishment is mainly governed by members of the conserved Wnt/Frizzled-PCP pathway, referred to as the core PCP pathway (reviewed in [1–7]) The core PCP factors were all originally discovered in Drosophila but are highly conserved across metazoa. PCP complexes at the cell surface include the atypical seven-pass transmembrane (TM) cadherin Flamingo (Fmi; Celsr in mammals), the seven-pass TM protein Frizzled (Fz; Fzd in vertebrates with several family members), and the four-pass trans-membrane protein Vang (Vangl1 and Vangl2 in mammals; a.k.a. strabismus/stbm in Drosophila and Xenopus). These TM proteins recruit the core cytoplasmic PCP factors Dishevelled (Dsh; Dvl1-3 in mammals), Diego (Dgo; Inversin/Diversin in vertebrates), and Prickle (Pk1-3) (e.g. reviewed in [1–7]). The initial stage of PCP signaling results in asymmetric localizations of its core members. These become enriched into two complexes on opposite sides of a given cell, which generate an intracellular bridge to convey polarity across cell membranes from cell to cell across the tissue (see reviews above). The resulting complexes, anchored in the membrane as Fz-Fmi on one side and Vang-Fmi on the other, stabilize each other intercellularly (between cells) and antagonize each other intracellularly (within each cell). The intracellular antagonistic behavior of these complexes is mediated by the cytoplasmic PCP factors Dsh/Dvl, Dgo, and Pk. The final asymmetric localization is generated through a dynamic process, and thought to include transient intermediary subcomplexes (reviewed in [1–3,5–10]). Asymmetric core PCP complex localization then directs spatially restricted downstream signaling events through cell-type specific effectors, leading to cytoskeletal rearrangement, centriole/centrosome/ciliary positioning, migratory regulation, and/or nuclear read-outs (see reviews above). Asymmetric distribution of PCP complexes is a direct consequence of their interactions during polarity establishment. It is observable in several tissues, ranging from cells of the Drosophila wing, where these complexes align to the proximal-distal axis, to for example mouse skin, where they align in the antero-posterior axis [11–14], also reviewed in [1,3–5,15]. Molecular interactions promote the formation of stable complexes at proximal (Drosophila wing) or anterior (mouse skin) cell membranes (Fmi/Celsr1-Vang/Vangl-Pk) and the equivalent distal or posterior cell surfaces (Fmi/Celsr1-Fz/Fzd-Dsh/Dvl), with Fmi (Celsr1 in mammals) forming a homotypic interaction across cells, which stabilizes these complexes intercellularly. Simple epithelia like the Drosophila wing display not only such highly coordinated PCP complex localization logic, but also an obvious and simple PCP read-out, with the formation of a single-actin based hair at the distal vertex of each cell pointing distally. Disruption of PCP establishment in the wing is easily observable by misorientation of the cellular hairs within the field of cells, or the formation of multiple cellular hairs in a single cell (reviewed in [1,4,5,15]. Although the formation of PCP is still best studied in Drosophila, where the core factors were initially discovered and functionally dissected [1,4,5,15], the importance of PCP during vertebrate development and human disease has become widely recognized (reviewed in [2,3,6,7,10,16,17]). For example, PCP directs polarized ciliary beating to generate fluid flow in the embryonic node, trachea, oviduct and brain ventricles. Failure to generate coordinated fluid flow in PCP mutant mice leads to left-right patterning defects, defective mucociliary clearance, sterility and hydrocephalus [7,18–24]. Core PCP genes are essential for neural tube closure in mammals and pathological variants in PCP genes are strongly associated with neural tube defects in humans [25–27]. A notable example of PCP in mammals is the uniform alignment of body hairs across the skin surface, where the core PCP proteins direct polarized morphogenesis and tissue-wide alignment of hair follicles [14,28–30]. Among the core PCP factors the Vang/Vangl family proteins occupy a unique role, as they have been shown to physically interact with all other 5 core PCP factors (rev in [1,4,5,15]) and also the A/B-polarity protein Scribble (Scrib) [31,32], with Dsh/Dvl, Pk, Dgo, and Scrib binding to the cytoplasmic C-terminal tail of Vang/Vangl (e.g. see [12,33]). Similarly, Vang/Vangl proteins have been shown to associate with the transmembrane factors Fmi/Celsr in cis and Fz in trans [14,34–36]. Drosophila Vang was identified by its strong PCP loss-of-function defects in wing and eye screens, and was also demonstrated to cause a domineering non-autonomous phenotype affecting nearby wild-type cells through propagation of aberrant polarity from mutant cells outward [37,38]. Vang/Vangl genes encode four-pass transmembrane proteins with intracellular cytoplasmic amino- and carboxy- terminal regions. All its cytoplasmic interactions have thus far been physically mapped to the C-terminal tail. It requires Pk and Scrib interactions for the formation of its stable complex. While it can also interact with Dsh and Dgo, these latter interactions are thought to be more transient and antagonistic to stable complex formation, with for example Dsh/Dvl binding to Vang thought to antagonize mislocalized Vang protein and/or prevent a Vang-Pk interaction in a wrong cellular location [12,13,31–33]. Its mammalian homologs Vangl1-2 regulate all PCP processes studied in higher organisms (for example [32,39,40] reviewed in [41,42]). Mutations of both Vangl1 and Vangl2 have also been identified in human patients affected with spina bifida and craniorachischisis [43] and subsequently shown to affect PCP signaling and establishment, using the well established Drosophila wing PCP model [44]. It was previously suggested that phosphorylation of Vang/Vangl proteins might be an important regulatory mechanism of core PCP complex formation. Along these lines, it was shown that Vang/Vangl proteins are phosphorylated on serine/threonine residues in the N-terminal tail, which is thought to affect complex formation and stability [45–47]. In order to better define the mechanisms underlying Vang/Vangl regulation and its interactions with downstream PCP factors, we investigated how phosphorylation of Vang/Vangl might regulate its function. Here we describe a concerted effort using the Drosophila wing and mouse skin models to better define potential Vang/Vangl interactions regulated by tyrosine phosphorylation. We have identified specific tyrosine phosphorylation events via mass-spectroscopy analyses. Here, we focus on a conserved phosphorylated C-tail tyrosine, which resides within the broad region of Pk and Dsh/Dvl binding. We fine-mapped Pk and Dsh binding to identify the amino acids required and find that Pk and Dsh binding sites overlap. Strikingly, the defined phosphorylation site resides in this newly defined overlapping binding regions of Pk and Dsh, and it regulates Vang/Vangl interactions with these core factors. Pk binds preferentially to the phosphorylated state, and Dsh to the unphosphorylated region. Our functional in vivo rescue studies demonstrate that binding of Vang to both effectors is physiologically relevant, as all single point mutations, which allow selective binding to either Pk or Dsh fail to rescue the Vang null mutant phenotype, as evident both in pupal wings and in adult tissue. Not surprisingly based on previous reports [12,33,48], membrane association of Pk, albeit reduced, is not lost in the Pk binding Vang mutants, confirming a more complex Pk recruitment scenario. With our defined single point mutation for Dsh binding, we show the physiological relevance of the Dsh interaction with Vang for the first time. Overall, we have identified new regulatory means for binding to antagonistic effectors during PCP establishment via a phosphorylation switch at a conserved site. Results Vang/Vangl2 proteins display a multitude of phosphorylations in vivo Several studies have demonstrated that Drosophila Vang and mouse Vangl2 are phosphorylated on serine (S) and threonine (T) residues, and a functionally important S/T phosphorylation cluster has been defined within the N-terminus [45–47,49]. In addition, phosphorylation on a specific tyrosine (Y) has been shown to be important for correct Vangl2 trafficking in mammalian cells and Drosophila Vang in vivo, respectively [47,50]. Our initial analyses performed in the Kelly et al. (2016) study [45–47,49] strongly suggested that additional Y residues are also likely to be phosphorylated. To better define the role of tyrosine phosphorylation in Vang/Vangl function, we first verified that Drosophila Vang is tyrosine phosphorylated in vivo (Fig 1A). Immunoprecipitation of Vang-Flagx3 from larval wing discs, showed a positive phospho-tyrosine signal, which was reduced upon phosphatase treatment (Fig 1A; note that a down-shift in the mobility of Vang consistent with successful phosphatase treatment was observed). Second, Vang remained Y-phosphorylated in the Vang-Y341F mutant (Fig 1B). As this site has been shown to be phosphorylated, with Y341 (Y279/280 in mouse Vangl2) promoting correct membrane trafficking [47,50], these data suggested that other tyrosines must be phosphorylated in Vang proteins in vivo. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 1. Vang is tyrosine phosphorylated in vivo. (A-C) Western blots of lysates from pupal wing discs showing that Vang is tyrosine phosphorylated in vivo. Note the reduction in signal of anti-pY with PPase treatment (A, upper panel) and the corresponding loss of a band shift (lower panel). (B) Tyrosine phosphorylation is maintained in the Vang-Y341F protein, suggesting other tyrosines are phosphorylated; and tyrosine phosphorylation is independent of Fz, as anti-pY staining is not affected in homozygous mutants fzP21 null animals (C). (D) Immunoprecipitation of GFP-Vangl2 from E15.5 wild-type control (WT) and K14-GFP-Vangl2 mouse epidermal lysates. Western blot using anti-GFP antibodies detects an ~85KD band corresponding to the GFP-Vangl2 fusion protein (arrow). T = total protein input, IP = immunoprecipitate, Sup = supernatant. (E) Anti-GFP immunoprecipitates from wild-type control (WT) and K14-GFP-Vangl2 mouse epidermal lysates run on an SDS-PAGE gel stained with SPYRO-Ruby to detect total protein. The ~85KD band present in the K14-GFP-Vangl2 but not wild type control immunoprecipitate was excised and processed for mass spectrometry. (F) Schemtic cartoon of mouse Vangl2 and sequence alignment of mVangl2 with human, zebrafish, Xenopus, and Drosophila orthologs. The region spanning amino acids 269–312 in mVangl2 within the C-terminal cytoplasmic tail is shown. Mass spectrometry analysis detected phosphorylation at highly conserved tyrosine residues 279 and 308 of mouse Vangl2 (bold with asterisk). The position of these is also marked with red asterisks in the cartoon. https://doi.org/10.1371/journal.pgen.1010849.g001 We furthermore detected Y-phosphorylation of Vang in a fz- null mutant background (Fig 1C), indicating that at least some Vang Y-phosphorylation is Fz independent, unlike the N-terminal Vang/Vangl2 S/T-cluster phosphorylation. S/T cluster phosphorylation is Fz dependent and causes a detectable band shift on protein gels (see Fig 1C for loss of band shift; see also [46,47,49]). Taken together, these in vivo Drosophila data suggest that Vang is phosphorylated on tyrosine residue(s) outside the defined Y341 and that at least some of these additional phosphorylation events might be Fz independent. To gain insight into which tyrosines might be phosphorylated, we turned to a mass spectrometry-based approach. As it is technically challenging to obtain sufficient material from Drosophila pupal wings for such mass spectrometry studies [51], we used mouse Vangl2 from embryonic skin. Due to its abundance and accessibility, the skin epidermis is an excellent model for biochemical analyses of proteins in their native context. To identify post-translational modifications on epidermally expressed Vangl2, we performed an IP-MS analysis of GFP-Vangl2 protein purified from the skin of K14-GFP-Vangl2 transgenic embryos [52]. Protein lysates were prepared from skin samples dissected from embryos at E15.5, and GFP-Vangl2 was immunoprecipitated using GFP-antibodies (Fig 1D). A prominent band of ~85KD was present in immunoprecipitates from K14-GFP-Vangl2 embryos but not wild type controls. Gel fragments containing protein in the 85-90KD range were excised and processed for LC-MS/MS analysis (Fig 1E). Over 220 unique, high confidence peptides were recovered spanning >86% of the GFP-Vangl2 fusion protein. Importantly, phosphorylation modifications were detected at two different tyrosine residues Y279 and Y308 (Fig 1F), both of which are conserved across vertebrates and in Drosophila (Figs 1F, and S1 and S1 Table). Y279 is located within Vangl2’s TGN-sorting motif and is equivalent to Y341 in Drosophila, whose phosphorylation was previously implicated in Vangl2 trafficking [47,50], indicating that our methods were able to detect functionally relevant Vang/Vangl2 phospho-tyrosine modifications. Intriguingly, Y308 corresponds to residue Y374 in Drosophila, which is located within the previously defined Dsh and Pk binding region of Vang [53] (Figs 1F and 2A). The strong conservation of this tyrosine and its position within the Dsh/Dvl and Pk binding region (see below) suggested that modification at this site may be important for Vang/Vangl function. Vang binds Pk and Dsh at an overlapping region around the Y374/Y308 phospho-site Vang/Vangl proteins have been shown to physically interact in vitro with all other members of the core PCP factor group. The molecular interactions of Vang with the cytoplasmic core factors have previously been mapped to specific regions within the C-terminal tail of the protein (shaded C-tail stretch in Fig 2A): amino acids 363–447 are required for its interaction with Pk and Dsh [53], a region which contains a phosphorylated conserved tyrosine as identified in our mass spec study (Fig 1D–1F, schematic in Fig 2D). We thus aimed to refine the binding sites of Pk and Dsh within this region and define the potential role of the phosphorylated Y-residue in this context. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 2. Pk and Dsh bind to adjacent, partially overlapping regions in the C-terminal tail of Vang. (A) Schematic of Vang showing the previously mapped binding region of Pk and Dsh within its C-terminal tail (shaded in gray), residues 363–447 [53]. Y374 is indicated in red (this residue is equivalent to Y308 in mouse Vangl2, see alignment in Fig 1F). Note that Y374 is located within the shaded region. (B) Western blot showing binding between Myc-Pk and C-terminal truncations of Flagx3-Vang, using a series of C-terminal truncations within residues 363–447. Binding is retained up until the 1–363 truncation, defining residues 363–375 in Vang as critical for its binding to Pk. (C) Schematic of sequences of the Vang C-terminal truncations as used in (B) and (D), red box highlights amino acids required for Pk-binding. (D) Western blot showing binding between Dsh-GFP and C-terminal truncations of Flagx3-Vang. Note that binding is retained up until the 1–387 truncation and markedly reduced in the 1–375 truncation and shorter. (E) Sequence alignment showing the conservation of amino acids in the Drosophila Vang 364–387 region with mouse and human Vangl proteins. Colored asterisks highlight amino acids mutated in binding experiments shown in panels F, G and H. (F) Western blot showing binding between Myc-Pk and selected Vang-Flagx3 mutants as indicated. Colored asterisks refer to specific amino acids in sequence schematic in (E) mutated in the experiment. Note marked reduction in binding for the single mutant (Y374A) and almost complete loss of binding in the triple mutant FKYY371AAYA. (F’) Quantification, derived from 6 independent replicates, and statitistical analysis of binding differences. * p<0.05 as determined with Anova (and Tukey’s post test to compare all samples with each other). (G) Western blot showing binding between Dsh-GFP and the indicated Vang-Flagx3 mutants. Note the reduction in binding follows a similar pattern to what was observed with Pk (compare to panel F), the equivalent Y residue to Y374 is detected as phosphorylated in mouse Vangl2 (see Fig 1D–1G). (G’) Quantification, combined from 6–7 replicates, and statitistical analysis of binding differences. **** p <0.0001 as determined with Anova (and Tukey’s post test to compare all samples). (H) Western blot showing binding between Dsh-GFP and Vang-Flagx3 with the V376A mutant. Note a marked reduction in binding. V376 is at the junction to the Pk-binding region (blue asterisk in E), suggesting an overlap in binding regions between Pk and Dsh. It was the only single residue mutation in the 376–387 stretch to affect Dsh binding. (H’) Quantification, combined from 6 replicates, and statitistical analysis of binding differences. * p <0.05 as determined with Anova. https://doi.org/10.1371/journal.pgen.1010849.g002 We first made a series of C-terminal truncations of Drosophila Vang covering the known binding region [53], amino acids 363–447 in the C-terminal tail, and reduced this binding domain by a series of 12 amino acid truncations (Fig 2B and 2C). Pull-down experiments, performed in S2 cells, revealed that Pk could interact with all such C-terminal Vang truncations with the exception of 1–363, which suggested that amino acids 364–375 were required for Pk-binding (Fig 2B, and see 2C for sequence schematic of deletion constructs). Importantly, this amino acid stretch contains the Y residue, identified as phosphorylated in vivo in the mass spec studies (see above), and conserved surrounding residues. We next assessed Vang interactions with the other cytoplasmic effectors that were also previously shown to interact within the C-tail region [33,53]. Strikingly, Dsh also showed markedly reduced binding when residues 363–387 were deleted (Fig 2C and 2D). In the case of Dsh we observed binding loss with the 1–375 truncation, directly adjacent and overlapping to the Pk binding area, suggesting that some amino acids within 376–387 stretch are critical for the Vang-Dsh interaction (Fig 2C and 2D). Taken together, these data suggest that Pk and Dsh bind to regions immediately adjacent to each other, and likely overlapping. This region of Vang/Vangl2 contains the conserved phosphorylated tyrosine, Y374 in Drosophila and Y308 in mouse Vangl2 (Fig 2E), we thus tested whether this Y-residue and/or associated conserved residues affect interaction with either Pk or Dsh or both. For example, we wished to ask whether binding could be charge or aminoacid structure (aromatic ring) dependent. A Vang-Y374A substitution (which removes both structure and potential charge) caused a markedly diminished binding of both Pk and Dsh (Fig 2F and 2G, respectively, see quantification in 2F’ and 2G’). This effect was further enhanced in the triple mutant affecting the whole FKxY motif (FKYY371AAYA) with interaction of either Pk or Dsh further reduced (Fig 2F, 2F’, 2G and 2G’). It is worth noting that in control experiments with the core PCP effector Dgo or the apical-basal polarity protein Scribble (Scrib), which have both been shown to interact with the C-tail of Vang/Vangl2 [13,31–33], mutations in the conserved FKYY residue stretch of Vang did not affect their binding to Vang (S2 Fig). Furthermore, the specificity and importance of Y374 and the FKYY motif for the interaction with Pk and Dsh within the defined binding region was confirmed via additional point mutations in the same region. For example, mutations of K366, F371, or K372 did not demonstrate a detectable requirement for Pk binding (S2 Fig). Taken together, these data suggest that Pk and Dsh, two antagonistic cytoplasmic core PCP factors, share an overlapping binding site centered around a conserved phosphorylated tyrosine (Y374 in Drosophila and Y308 in mouse Vangl2) in the C-terminal tail region of Vang proteins with both Pk and Dsh requiring Y374 for binding within a conserved stretch of amino acids. To further refine the binding requirement(s) of the Vang-Dsh interaction within the neighboring residues 376–387 (necessary in the truncation series for Dsh binding to Vang, Fig 2D), we scanned the respective amino acids with single alanine substitutions. This revealed surprisingly that only one substitution to alanine (A), that of a partially conserved valine (V) at position 376 (see alignment in Fig 2E), V376A, led to a marked reduction in Dsh binding (Fig 2H and 2H’). Notably, this residue is directly adjacent to the region required for Pk binding, but we did not observe an impact of V376A on the Vang-Pk interaction (S2 Fig). Thus, in addition to the above conclusion regarding the Y374 motif, our data also suggested that we have defined a specific point mutant at V376, which is only partially conserved (Fig 2E), that specifically modulates Vang-Dsh binding (see also below). Y374 phosphorylation/charged state contributes to effector binding regulation As the key residue required for binding to both Pk and Dsh was a phosphorylated tyrosine, we thus asked whether phosphorylation—or charge associated with phosphorylation—at this site might regulate binding. We first performed substitution of Y374 to phenylalanine (F), with Y374F retaining the structural aromatic ring associated with tyrosine, but it remains uncharged as it cannot be phosphorylated. Conversely, Y374D or Y374E substitutions maintain charge, but they change the structure of the residue (loss of aromatic ring). Importantly, these substitutions also allowed comparison to the Y374A scenario, which diminished binding of both Pk and Dsh (Fig 2F–2G’). In these binding studies, Pk showed equally diminished binding to both Y374A and Y374F (Fig 3A, quantified in 3A’), while Dsh retained binding to Y374F (Fig 3B and 3B’; note Dsh-binding was reduced in the Y374A substitution in the same experiment as a control, cf. to Fig 2G and 2G’). These data suggested that Dsh binding requires the structure of the amino acid, an aromatic ring, shared between tyrosine and phenylalanine, but is not impacted by phosphorylation or charge of this site. To ask more directly whether a charge at Y374 can contribute to binding regulation, we tested a Y374D, Y373DD, and Y374E substitutions (with D and E, being charged residues, to mimic a partial phosphorylation charge). While Pk displayed significant binding to Y374D and Y373DD (Fig 3C, quantified in Fig 3C’, compare to Y374F and Y374A controls in same panels) as well as to Y374E (S3A Fig), Dsh bound poorly to these charged substitution mutations (Fig 3D and 3D’, compare to wt-control and Y374F in same panels; and S3B Fig). Taken together, these data suggested that Dsh requires an aromatic ring at the binding region centered on Y374 and its binding might be inhibited by charge, while the Vang-Pk interaction appeared to be dependent on the negative charge, consistent with the notion that it would be promoted by phosphorylation. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 3. Differential substitutions at Y374 reveal a role for charge/phosphorylation in binding regulation. (A) Western blot showing binding between Myc-Pk and the indicated Vang-Flagx3 mutants, Y374A and Y374F. Note binding is markedly reduced in both cases. (A’) Quantification, combined from 6 independent replicates, and statitistical analysis of binding differences. **p <0.005, and ***p<0.0005 as determined with Anova (and Tukey’s post test to compare all samples with each other). (B) Western blot showing binding between Dsh-GFP and the indicated Vang-Flagx3 mutants. Note binding is reduced with an alanine substitution, Y374A, but not with phenylalanine, Y374F, which retains the aromatic ring feature of tyrosine. (B’) Quantification, combined from 6 independent replicates, and statitistical analysis of binding differences: ****p<0.0001, determined with Anova (and Tukey’s post test). (C) Charged amino acid substitutions of Y374 partially restore Pk binding. Western blot showing binding between Myc-Pk and the indicated Vang-Flagx3 Y374 substitutions. Note that Y374D and YY373DD recover significant binding of Pk, as compared to Y374A and Y374F. (C’) Quantification, combined from 4–10 replicates, and statitistical analysis of binding differences. ***p<0.0005 as determined with Anova (and Tukey’s post test). (D) Charged amino acids interfere with the Dsh binding to Vang. Western blot showing binding between Dsh-GFP and the indicated Vang-Flagx3 substitutions. Note that Vang Y374F is by far the best interacting mutant, again confirming a requirement of an aromatic ring feature/structure for Dsh binding (see also panel B). (D’) Quantification, combined from 3–9 independent replicates, and statitistical analysis of binding differences. *p<0.05, **p <0.005, and ****p<0.0001 as determined with Anova (and Tukey’s post test). https://doi.org/10.1371/journal.pgen.1010849.g003 To confirm this hypothesis, we also generated the equivalent binding region as an in vitro peptide and phospho-peptide. The peptide generated encompasses the Vang region 367–381 that the cell based experiments predicted to be sufficient to bind to both Pk and Dsh, in its phosphorylated form at Y374, ISNSFKY[pY]EVDGVSN-amide (pY-peptide) and as a non-phosphorylated control (Y-peptide, see S4A Fig for peptide sequence). These peptide interaction assays could test whether the binding region defined above (Fig 2) is sufficient for effector binding, not just necessary, and whether phosphorylation influences the Pk interaction. The caveat of these purely in vitro association assays is however that these are rather artificial and mixing peptides with purified proteins in a test tube certainly misses the cellular physiological regulation. Nevertheles, and consistent with the co-IP studies, the (phosphorylated) pY-peptide preferentially interacted with Pk (Fig 4A). To further corroborate this, we assayed peptide interactions with Pk and Dsh in a competitive binding assay, revealing that Dsh-GFP protein when mixed with the pY peptide (coupled to beads) was readily outcompeted by addition of GFP-Pk to the solution (Fig 4B, quantifed in 4C; see also bottom panel in Fig 4B showing that Pk did not outcompete Dsh binding to the Y-peptide). While these interaction studies are consistent with the co-IPs shown above, the in vitro peptide assay displayed the above mentioned caveat, as for example the unphorphorylated peptide was bound equally by both Dsh and Pk, when assayed independently (S4B Fig). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 4. A phospho-peptide of Y374 surronding sequences binds preferentially to Pk. (A) Pk interacts better with/binds to the phosphorylated “pY” peptide (S4 Fig for peptide sequence). Compare bands between input and pY-bound in greay boxes for Pk vs. for Dsh (red boxes). (B) Binding of Dsh (Dsh-GFP) is outcompeted by Pk (myc-Pk) on the “p-Y” (phospho-peptide), but not the non-phosphorylated (“Y”) peptide (bottom panel). Western blot showing retention of Dsh-GFP on beads coupled with the respective peptide, either “pY” for “Y” among mixing with Pk. Upper blot shows stable Dsh-GFP input, TZ is an unrelated control protein. Note reduced binding of Dsh-GFP when mixed in solution with increasing levels of Pk in the context of pY peptide conjugated beads (middle panel), but not the Y peptide (bottom panel). (C) Quantification of binding competition as shown in (B), combined from 3 independent replicates. Increasing Pk levels in the solution (upper graph) caused a decrease in Dsh binding to the “pY” peptide (lower graph). Statitistical analysis of binding differences (to lane #2): **p <0.005, and ****p<0.0001 (determined with one-way Anova). https://doi.org/10.1371/journal.pgen.1010849.g004 In summary and taken together, the above data are consistent with the model that (i) the Vang region centered around Y374 is necessary and sufficient for binding to Pk and Dsh, (ii) Pk preferentially interacts with this region when Y374 is phosphorylated, and (iii) Dsh requires the structural features of an aromatic ring at Y374 in its binding behavior to Vang. Furthermore, these experiments defined single amino acid mutations, Y374F and V376A, that selectively abrogated binding to either Pk (Y374F) or Dsh (V376A). These observations allow for the first time the functional testing of a direct Vang-Dsh interaction requirement in vivo. Single point mutants show PCP defects To investigate the functional consequences in vivo of altering the interaction of Vang with Pk and Dsh, we performed rescue experiments using transgenes carrying the different single point mutants. We expressed Vang-Flagx3 WT, Y374A, Y374F and V376A, using a direct tubulin-driven expression in a Vang null mutant background (Vang6). The control wild-type construct tub-VangWT was able to fully rescue cellular hair orientation in Drosophila wings (Fig 5A–5C, see 5D for quantification and statistical analyses; see 5B for Vang6 null mutant control). Each of the single point mutants, Vang-Y374A, Vang-Y374F and Vang-V376A, rescued the Vang6 loss-of-function phenotype partially, displaying varying degrees of cellular hair orientation defects (Fig 5C, quantified in 5D). The phenotypes were quantified using FijiWingsPolarity [54], which revealed phenotypic differences from VangWT control in each case, and also among all individual point mutations (compare panels in Fig 5C and quantifications in 5D). Importantly, while the rescue was partial in all three cases, the phenotypes were significantly different from the “non-rescued” null allele and the rescued VangWT wings, consistent with the notion that all three point mutants retained partial function as expected. Also, they all displayed different phenotypes, which were visibly and statitistally different from each other (and the null allele; Fig 5C and 5D), suggesting differing partial function, consistent with their distinct biochemical behavior. Consistent with the molecular binding data, Vang-Y374A, which should interfere with direct binding of either Pk or Dsh showed the most severe defects (see cellular hair angle distribution in Fig 5C and 5D), closest to the Vang null phenotype but nonetheless different (see statistical analyses in Fig 5D). The point mutations affecting binding to a single effector (Y374F only binding Dsh, and V376A only interacting with Pk) showed a better rescue as compared to Y374A (Fig 5C and 5D; see also Discussion). Of note, the phenotypic differences were not due to different expression levels, as all genotypes showed similar protein levels in larval disc lysates (Fig 5E). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 5. Single Vang point mutants in the Pk and Dsh interaction region display PCP defects in vivo. (A) Drosophila wing (distal is right) and (A’) higher magnification of boxed area in wing on left. Cellular orientation is reflected by single hairs pointing distally, also visualized by colored arrows reflecting cellular orientation in lower panel. Angles were determined using plugin FijiWingsPolarity (ref: [54]). (B) Image of same region as in A’ of a Vang6 (null mutant) wing. Note misoriented cellular hairs forming waves and whorls, and schematic refelction of mispolarization by color coded arrows. See colored bar in A depicting color scheme. Wild-type (distal) orientation is associated with turquoise in the 160–180° range. (C) Sample images of cellular (mis)orientation in adult wings of Vang6 mutant flies (same region as in A’-B) upon rescue with tubulin-promoter expressed Vang-Flagx3 transgenes with WT (wild-type), Y374A, Y374F or V376A, respectively as indicated. Angles of cellular orientation are visualized through colored arrows in lower panels (see color bar/scale in panel A above). Note that while Vang-WT fully rescues the mutant phenotype (and appears like wild-type, compare to A’), all single point mutants fail to rescue the mutant defects. Note also that Vang-Y374A most resembles the Vang null phenotype, and the other point mutations, display different phenotypes. (D) Quantification and statistical analyses of the different behaviors of the individual Vang binding mutants in the in vivo rescue assay of the Vang-/- null allele, shown as violin plots displaying cellular hair orientation angles for the indicated genotypes, within the ROI area shown in panels A’-C from 3 independent wings were combined. Statitistical analysis was determined with one-way Anova (and Tukey’s post test to compare all samples with each other): **p <0.005, and ****p<0.0001. Note that all genotypes show statitistically different phenotypes from all other gentypes, with the exception of wild-type and the VangWT transgene rescue of the Vang-/- null allele. (E) Western blot of wing disc lysates from the indicated genotypes. Note comparable levels of expression in all transgenes. https://doi.org/10.1371/journal.pgen.1010849.g005 To expand on the different behavior of the individual Vang point mutants in vivo, we next tested their effects in a gain-of-function (GOF) assay in developing wings. A classic feature of core PCP factors is that their overexpression also causes PCP defects. Each core factor has thus a reproducible LOF phenotype and also a GOF phenotype, both reflected in stereotypic defects highly similar from wing to wing of any given genotype. As such the Vang-wt GOF (via nubbin-Gal4) induces similar misorientations from wing to wing (in any specific region of a wing; see wing ROI samples in S5A–S5B Fig; boxed areas in blue and red in S5A Fig correspond to left and right panels in S5B Fig, respectively). As such, Vang overexpression, via nubbin-Gal4, of either Vang-wt or any Vang-point mutant (see indicated genotypes in S5 Fig), caused cellular orientation defects with wings of the same genotype looking similar to each other, but with different appearances between the distinct Vang-mutant genotypes (S5 Fig). This confirmed that, for example, Vang-Y374F (binding only Dsh) overexpression caused distinct defects from Vang-WT or Vang-YY373DD (binding predominantly Pk) or Vang-Y374A (binding neither factor). Accordingly, all specific Vang mutations tested displayed distinct behavior (S5B Fig, see quantifications in angle distribution graphs on right of each panel; note expression levels of each mutant protein were comparable, S5C Fig). Together with the rescue data (Fig 5), these results are consistent with the notion that Vang interacts physiologically with both cytoplasmic core PCP factors, Pk and Dsh, during planar polarity establishment in vivo. Together with the biochemical dissection, these data thus suggest that Y374 phosphorylation is a critical node for physiological activity of Vang during PCP establishment and for the resolution of core PCP complexes into the Vang and Fz associated “antagonistic polarity domains” (see also Discussion). Vang point mutants affect core PCP factor localization during PCP establishment To get further functional insight into how the Vang point mutations affect PCP establishment and core PCP factor localization in vivo, we next tested the localization of these mutants and associated other core factors. Using the rescue transgenes (with a tub-Vang-Flagx3 backbone) in a Vang6 null background, we analyzed the localization of Vang-Flagx3 WT, Y374A, Y374F and V376A, Fmi, EGFP-Pk, and DshGFP with E-cad as cellular outline control. The polarity of these proteins was evaluated using the Packing Analyzer suite [55] (Figs 6 and S6). Vang-WT displayed wild-type localization, polarized along the proximo-distal axis (Fig 6A) with all other core PCP factors tested, Fmi, Pk, and Dsh, also showing wild-type polarized, membrane associated distribution within the proximo-distal axis in pupal wing tissue at 28h APF (Figs 6A and S6A). In contrast all three point mutants displayed largely apolar (or mispolarized) distribution (Fig 6B–6D, mispolarization is highlighted in the angle quantification as seen in the rosette diagrams in panels 6A’-6D’). Apolar PCP core factor distribution was also evident in analyses of Fmi distribution (S6 Fig). Pk and Dsh protein distribution was also apolar (Figs 6B–6D and S6B–S6D, see also below). However, both cytoplasmic core PCP factors, Dsh and Pk, maintained significant membrane association in all point mutant backgrounds (Figs 6B–6D and S6B–S6D, see also below). All three Vang point mutants appeared membrane associated similar to Vang-WT (Fig 6), suggesting that Vang trafficking is not affected in these point mutants (see Discussion). While it is possible that differences between the individual Vang point mutants exist, also based on past observations for example [56], a carefull side-by-side comparison of the respective point mutations would be needed. Importantly, none of the point mutations caused a complete loss of membrane associated Pk or Dsh. This is consistent with previous data demonstrating that even in Vang null clones Pk is still partially retained at cellular membranes, likely due to being farnesylated and requiring this modification for function [12,48]. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 6. Vang point mutants affect localization of core PC factors in vivo. (A-D) Confocal images of 28-30hr APF pupal wings of the genotypes indicated above panels. All pupal wing tissue is oriented horizinally with proximal side being left and distal right, stained for Vang-Flagx3 (anti-Flag, red), EGFP-Pk (anti-GFP, green), and E-cad (blue). (A) Vang-/-; tub-VangWT-Flagx3, (B) Vang-/-; tub-VangY374A-Flagx3, (C) Vang-/-; tub-VangY374F-Flagx3, (D) Vang-/-; tub-VangV376A-Flagx3. Grayscale single channel micrographs are shown in lower panels as indicated on left side. Note that Pk is still localized to membranes in all three point mutants, albeit less enriched. Scale bars: 20μm. (A’-D’) Quantification of polarity angles presented as rosette diagrams of VangFlagx3 protein in the respective genotypes indicated: (A’) Vang-/-; tub-VangWT-Flagx3, (B’) Vang-/-; tub-VangY374A-Flagx3, (C’) Vang-/-; tub-VangY374F-Flagx3, (D’) Vang-/-; tub-VangV376A-Flagx3. Note that the majority of cell orientation angles is in the proximo-distal axis (horizontal, highlighted in red in rosette diagram) in the Vang-/-; tub-VangWT background (A’, indistinguishable from wild-type), but that cellular orientation is largely randomized in all three point mutants (B’-D’, very similar to the Vang null allele). See also S6 Fig for additional core factor stainings and polarity analyses in these genetic backgrounds. https://doi.org/10.1371/journal.pgen.1010849.g006 In parallel to in vivo localization studies, we also tested whether Vang/Vangl variants might differently affect the stability of core PCP membrane complexes, which are anchored at cell junctions by the atypical cadherin Fmi (Celsr1 in mice). To do this, we transfected mouse primary keratinocytes with Celsr1-GFP and mCherry-Vangl2, which co-localize to the junctional interfaces between expressing cell pairs mediated by homotypic adhesion of Celsr1 (S7A Fig) [14]. We then used fluorescence recovery after photocbleaching (FRAP) to determine the mobility of Celsr1 and Vangl2 proteins located within the junctional membrane. Compared to Celsr1 and Vangl2 proteins located at free edges of the cell, Celsr1 and Vangl2 were largely immobile at cell junctions (~80% immobile fraction)(S7B–S7C Fig), as previously shown [36,57]. Mutation of Vangl2 Y308 to A (equivalent to Drosophila VangY374A) did not affect its ability to localize to cell junctions (S7A Fig), nor did the mutation impact Vangl2 mobility at cell junctions, or that of co-expressed Celsr1-GFP (S7B–S7C Fig). We conclude that although the Y308 residue in Vangl2 can be phosphorylated in mammalian epidermal cells, it is not required for Vangl2 recruitment to or stabilization at cell junctions. Taken together, the in vivo analyses of the Vang point mutations confirm that the direct interactions with Pk and Dsh are critical for correct PCP establishment and polarized core PCP complex segregation and thus Vang function. However, they also confirmed that Vang is not essential for membrane recruitment of either cytoplasmic PCP factor and also that the phosphorylation status of its Y374 residue (Y308 in mice) has no or minimal impact on junctional PCP core complex stability (see Discussion). Vang/Vangl2 proteins display a multitude of phosphorylations in vivo Several studies have demonstrated that Drosophila Vang and mouse Vangl2 are phosphorylated on serine (S) and threonine (T) residues, and a functionally important S/T phosphorylation cluster has been defined within the N-terminus [45–47,49]. In addition, phosphorylation on a specific tyrosine (Y) has been shown to be important for correct Vangl2 trafficking in mammalian cells and Drosophila Vang in vivo, respectively [47,50]. Our initial analyses performed in the Kelly et al. (2016) study [45–47,49] strongly suggested that additional Y residues are also likely to be phosphorylated. To better define the role of tyrosine phosphorylation in Vang/Vangl function, we first verified that Drosophila Vang is tyrosine phosphorylated in vivo (Fig 1A). Immunoprecipitation of Vang-Flagx3 from larval wing discs, showed a positive phospho-tyrosine signal, which was reduced upon phosphatase treatment (Fig 1A; note that a down-shift in the mobility of Vang consistent with successful phosphatase treatment was observed). Second, Vang remained Y-phosphorylated in the Vang-Y341F mutant (Fig 1B). As this site has been shown to be phosphorylated, with Y341 (Y279/280 in mouse Vangl2) promoting correct membrane trafficking [47,50], these data suggested that other tyrosines must be phosphorylated in Vang proteins in vivo. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 1. Vang is tyrosine phosphorylated in vivo. (A-C) Western blots of lysates from pupal wing discs showing that Vang is tyrosine phosphorylated in vivo. Note the reduction in signal of anti-pY with PPase treatment (A, upper panel) and the corresponding loss of a band shift (lower panel). (B) Tyrosine phosphorylation is maintained in the Vang-Y341F protein, suggesting other tyrosines are phosphorylated; and tyrosine phosphorylation is independent of Fz, as anti-pY staining is not affected in homozygous mutants fzP21 null animals (C). (D) Immunoprecipitation of GFP-Vangl2 from E15.5 wild-type control (WT) and K14-GFP-Vangl2 mouse epidermal lysates. Western blot using anti-GFP antibodies detects an ~85KD band corresponding to the GFP-Vangl2 fusion protein (arrow). T = total protein input, IP = immunoprecipitate, Sup = supernatant. (E) Anti-GFP immunoprecipitates from wild-type control (WT) and K14-GFP-Vangl2 mouse epidermal lysates run on an SDS-PAGE gel stained with SPYRO-Ruby to detect total protein. The ~85KD band present in the K14-GFP-Vangl2 but not wild type control immunoprecipitate was excised and processed for mass spectrometry. (F) Schemtic cartoon of mouse Vangl2 and sequence alignment of mVangl2 with human, zebrafish, Xenopus, and Drosophila orthologs. The region spanning amino acids 269–312 in mVangl2 within the C-terminal cytoplasmic tail is shown. Mass spectrometry analysis detected phosphorylation at highly conserved tyrosine residues 279 and 308 of mouse Vangl2 (bold with asterisk). The position of these is also marked with red asterisks in the cartoon. https://doi.org/10.1371/journal.pgen.1010849.g001 We furthermore detected Y-phosphorylation of Vang in a fz- null mutant background (Fig 1C), indicating that at least some Vang Y-phosphorylation is Fz independent, unlike the N-terminal Vang/Vangl2 S/T-cluster phosphorylation. S/T cluster phosphorylation is Fz dependent and causes a detectable band shift on protein gels (see Fig 1C for loss of band shift; see also [46,47,49]). Taken together, these in vivo Drosophila data suggest that Vang is phosphorylated on tyrosine residue(s) outside the defined Y341 and that at least some of these additional phosphorylation events might be Fz independent. To gain insight into which tyrosines might be phosphorylated, we turned to a mass spectrometry-based approach. As it is technically challenging to obtain sufficient material from Drosophila pupal wings for such mass spectrometry studies [51], we used mouse Vangl2 from embryonic skin. Due to its abundance and accessibility, the skin epidermis is an excellent model for biochemical analyses of proteins in their native context. To identify post-translational modifications on epidermally expressed Vangl2, we performed an IP-MS analysis of GFP-Vangl2 protein purified from the skin of K14-GFP-Vangl2 transgenic embryos [52]. Protein lysates were prepared from skin samples dissected from embryos at E15.5, and GFP-Vangl2 was immunoprecipitated using GFP-antibodies (Fig 1D). A prominent band of ~85KD was present in immunoprecipitates from K14-GFP-Vangl2 embryos but not wild type controls. Gel fragments containing protein in the 85-90KD range were excised and processed for LC-MS/MS analysis (Fig 1E). Over 220 unique, high confidence peptides were recovered spanning >86% of the GFP-Vangl2 fusion protein. Importantly, phosphorylation modifications were detected at two different tyrosine residues Y279 and Y308 (Fig 1F), both of which are conserved across vertebrates and in Drosophila (Figs 1F, and S1 and S1 Table). Y279 is located within Vangl2’s TGN-sorting motif and is equivalent to Y341 in Drosophila, whose phosphorylation was previously implicated in Vangl2 trafficking [47,50], indicating that our methods were able to detect functionally relevant Vang/Vangl2 phospho-tyrosine modifications. Intriguingly, Y308 corresponds to residue Y374 in Drosophila, which is located within the previously defined Dsh and Pk binding region of Vang [53] (Figs 1F and 2A). The strong conservation of this tyrosine and its position within the Dsh/Dvl and Pk binding region (see below) suggested that modification at this site may be important for Vang/Vangl function. Vang binds Pk and Dsh at an overlapping region around the Y374/Y308 phospho-site Vang/Vangl proteins have been shown to physically interact in vitro with all other members of the core PCP factor group. The molecular interactions of Vang with the cytoplasmic core factors have previously been mapped to specific regions within the C-terminal tail of the protein (shaded C-tail stretch in Fig 2A): amino acids 363–447 are required for its interaction with Pk and Dsh [53], a region which contains a phosphorylated conserved tyrosine as identified in our mass spec study (Fig 1D–1F, schematic in Fig 2D). We thus aimed to refine the binding sites of Pk and Dsh within this region and define the potential role of the phosphorylated Y-residue in this context. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 2. Pk and Dsh bind to adjacent, partially overlapping regions in the C-terminal tail of Vang. (A) Schematic of Vang showing the previously mapped binding region of Pk and Dsh within its C-terminal tail (shaded in gray), residues 363–447 [53]. Y374 is indicated in red (this residue is equivalent to Y308 in mouse Vangl2, see alignment in Fig 1F). Note that Y374 is located within the shaded region. (B) Western blot showing binding between Myc-Pk and C-terminal truncations of Flagx3-Vang, using a series of C-terminal truncations within residues 363–447. Binding is retained up until the 1–363 truncation, defining residues 363–375 in Vang as critical for its binding to Pk. (C) Schematic of sequences of the Vang C-terminal truncations as used in (B) and (D), red box highlights amino acids required for Pk-binding. (D) Western blot showing binding between Dsh-GFP and C-terminal truncations of Flagx3-Vang. Note that binding is retained up until the 1–387 truncation and markedly reduced in the 1–375 truncation and shorter. (E) Sequence alignment showing the conservation of amino acids in the Drosophila Vang 364–387 region with mouse and human Vangl proteins. Colored asterisks highlight amino acids mutated in binding experiments shown in panels F, G and H. (F) Western blot showing binding between Myc-Pk and selected Vang-Flagx3 mutants as indicated. Colored asterisks refer to specific amino acids in sequence schematic in (E) mutated in the experiment. Note marked reduction in binding for the single mutant (Y374A) and almost complete loss of binding in the triple mutant FKYY371AAYA. (F’) Quantification, derived from 6 independent replicates, and statitistical analysis of binding differences. * p<0.05 as determined with Anova (and Tukey’s post test to compare all samples with each other). (G) Western blot showing binding between Dsh-GFP and the indicated Vang-Flagx3 mutants. Note the reduction in binding follows a similar pattern to what was observed with Pk (compare to panel F), the equivalent Y residue to Y374 is detected as phosphorylated in mouse Vangl2 (see Fig 1D–1G). (G’) Quantification, combined from 6–7 replicates, and statitistical analysis of binding differences. **** p <0.0001 as determined with Anova (and Tukey’s post test to compare all samples). (H) Western blot showing binding between Dsh-GFP and Vang-Flagx3 with the V376A mutant. Note a marked reduction in binding. V376 is at the junction to the Pk-binding region (blue asterisk in E), suggesting an overlap in binding regions between Pk and Dsh. It was the only single residue mutation in the 376–387 stretch to affect Dsh binding. (H’) Quantification, combined from 6 replicates, and statitistical analysis of binding differences. * p <0.05 as determined with Anova. https://doi.org/10.1371/journal.pgen.1010849.g002 We first made a series of C-terminal truncations of Drosophila Vang covering the known binding region [53], amino acids 363–447 in the C-terminal tail, and reduced this binding domain by a series of 12 amino acid truncations (Fig 2B and 2C). Pull-down experiments, performed in S2 cells, revealed that Pk could interact with all such C-terminal Vang truncations with the exception of 1–363, which suggested that amino acids 364–375 were required for Pk-binding (Fig 2B, and see 2C for sequence schematic of deletion constructs). Importantly, this amino acid stretch contains the Y residue, identified as phosphorylated in vivo in the mass spec studies (see above), and conserved surrounding residues. We next assessed Vang interactions with the other cytoplasmic effectors that were also previously shown to interact within the C-tail region [33,53]. Strikingly, Dsh also showed markedly reduced binding when residues 363–387 were deleted (Fig 2C and 2D). In the case of Dsh we observed binding loss with the 1–375 truncation, directly adjacent and overlapping to the Pk binding area, suggesting that some amino acids within 376–387 stretch are critical for the Vang-Dsh interaction (Fig 2C and 2D). Taken together, these data suggest that Pk and Dsh bind to regions immediately adjacent to each other, and likely overlapping. This region of Vang/Vangl2 contains the conserved phosphorylated tyrosine, Y374 in Drosophila and Y308 in mouse Vangl2 (Fig 2E), we thus tested whether this Y-residue and/or associated conserved residues affect interaction with either Pk or Dsh or both. For example, we wished to ask whether binding could be charge or aminoacid structure (aromatic ring) dependent. A Vang-Y374A substitution (which removes both structure and potential charge) caused a markedly diminished binding of both Pk and Dsh (Fig 2F and 2G, respectively, see quantification in 2F’ and 2G’). This effect was further enhanced in the triple mutant affecting the whole FKxY motif (FKYY371AAYA) with interaction of either Pk or Dsh further reduced (Fig 2F, 2F’, 2G and 2G’). It is worth noting that in control experiments with the core PCP effector Dgo or the apical-basal polarity protein Scribble (Scrib), which have both been shown to interact with the C-tail of Vang/Vangl2 [13,31–33], mutations in the conserved FKYY residue stretch of Vang did not affect their binding to Vang (S2 Fig). Furthermore, the specificity and importance of Y374 and the FKYY motif for the interaction with Pk and Dsh within the defined binding region was confirmed via additional point mutations in the same region. For example, mutations of K366, F371, or K372 did not demonstrate a detectable requirement for Pk binding (S2 Fig). Taken together, these data suggest that Pk and Dsh, two antagonistic cytoplasmic core PCP factors, share an overlapping binding site centered around a conserved phosphorylated tyrosine (Y374 in Drosophila and Y308 in mouse Vangl2) in the C-terminal tail region of Vang proteins with both Pk and Dsh requiring Y374 for binding within a conserved stretch of amino acids. To further refine the binding requirement(s) of the Vang-Dsh interaction within the neighboring residues 376–387 (necessary in the truncation series for Dsh binding to Vang, Fig 2D), we scanned the respective amino acids with single alanine substitutions. This revealed surprisingly that only one substitution to alanine (A), that of a partially conserved valine (V) at position 376 (see alignment in Fig 2E), V376A, led to a marked reduction in Dsh binding (Fig 2H and 2H’). Notably, this residue is directly adjacent to the region required for Pk binding, but we did not observe an impact of V376A on the Vang-Pk interaction (S2 Fig). Thus, in addition to the above conclusion regarding the Y374 motif, our data also suggested that we have defined a specific point mutant at V376, which is only partially conserved (Fig 2E), that specifically modulates Vang-Dsh binding (see also below). Y374 phosphorylation/charged state contributes to effector binding regulation As the key residue required for binding to both Pk and Dsh was a phosphorylated tyrosine, we thus asked whether phosphorylation—or charge associated with phosphorylation—at this site might regulate binding. We first performed substitution of Y374 to phenylalanine (F), with Y374F retaining the structural aromatic ring associated with tyrosine, but it remains uncharged as it cannot be phosphorylated. Conversely, Y374D or Y374E substitutions maintain charge, but they change the structure of the residue (loss of aromatic ring). Importantly, these substitutions also allowed comparison to the Y374A scenario, which diminished binding of both Pk and Dsh (Fig 2F–2G’). In these binding studies, Pk showed equally diminished binding to both Y374A and Y374F (Fig 3A, quantified in 3A’), while Dsh retained binding to Y374F (Fig 3B and 3B’; note Dsh-binding was reduced in the Y374A substitution in the same experiment as a control, cf. to Fig 2G and 2G’). These data suggested that Dsh binding requires the structure of the amino acid, an aromatic ring, shared between tyrosine and phenylalanine, but is not impacted by phosphorylation or charge of this site. To ask more directly whether a charge at Y374 can contribute to binding regulation, we tested a Y374D, Y373DD, and Y374E substitutions (with D and E, being charged residues, to mimic a partial phosphorylation charge). While Pk displayed significant binding to Y374D and Y373DD (Fig 3C, quantified in Fig 3C’, compare to Y374F and Y374A controls in same panels) as well as to Y374E (S3A Fig), Dsh bound poorly to these charged substitution mutations (Fig 3D and 3D’, compare to wt-control and Y374F in same panels; and S3B Fig). Taken together, these data suggested that Dsh requires an aromatic ring at the binding region centered on Y374 and its binding might be inhibited by charge, while the Vang-Pk interaction appeared to be dependent on the negative charge, consistent with the notion that it would be promoted by phosphorylation. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 3. Differential substitutions at Y374 reveal a role for charge/phosphorylation in binding regulation. (A) Western blot showing binding between Myc-Pk and the indicated Vang-Flagx3 mutants, Y374A and Y374F. Note binding is markedly reduced in both cases. (A’) Quantification, combined from 6 independent replicates, and statitistical analysis of binding differences. **p <0.005, and ***p<0.0005 as determined with Anova (and Tukey’s post test to compare all samples with each other). (B) Western blot showing binding between Dsh-GFP and the indicated Vang-Flagx3 mutants. Note binding is reduced with an alanine substitution, Y374A, but not with phenylalanine, Y374F, which retains the aromatic ring feature of tyrosine. (B’) Quantification, combined from 6 independent replicates, and statitistical analysis of binding differences: ****p<0.0001, determined with Anova (and Tukey’s post test). (C) Charged amino acid substitutions of Y374 partially restore Pk binding. Western blot showing binding between Myc-Pk and the indicated Vang-Flagx3 Y374 substitutions. Note that Y374D and YY373DD recover significant binding of Pk, as compared to Y374A and Y374F. (C’) Quantification, combined from 4–10 replicates, and statitistical analysis of binding differences. ***p<0.0005 as determined with Anova (and Tukey’s post test). (D) Charged amino acids interfere with the Dsh binding to Vang. Western blot showing binding between Dsh-GFP and the indicated Vang-Flagx3 substitutions. Note that Vang Y374F is by far the best interacting mutant, again confirming a requirement of an aromatic ring feature/structure for Dsh binding (see also panel B). (D’) Quantification, combined from 3–9 independent replicates, and statitistical analysis of binding differences. *p<0.05, **p <0.005, and ****p<0.0001 as determined with Anova (and Tukey’s post test). https://doi.org/10.1371/journal.pgen.1010849.g003 To confirm this hypothesis, we also generated the equivalent binding region as an in vitro peptide and phospho-peptide. The peptide generated encompasses the Vang region 367–381 that the cell based experiments predicted to be sufficient to bind to both Pk and Dsh, in its phosphorylated form at Y374, ISNSFKY[pY]EVDGVSN-amide (pY-peptide) and as a non-phosphorylated control (Y-peptide, see S4A Fig for peptide sequence). These peptide interaction assays could test whether the binding region defined above (Fig 2) is sufficient for effector binding, not just necessary, and whether phosphorylation influences the Pk interaction. The caveat of these purely in vitro association assays is however that these are rather artificial and mixing peptides with purified proteins in a test tube certainly misses the cellular physiological regulation. Nevertheles, and consistent with the co-IP studies, the (phosphorylated) pY-peptide preferentially interacted with Pk (Fig 4A). To further corroborate this, we assayed peptide interactions with Pk and Dsh in a competitive binding assay, revealing that Dsh-GFP protein when mixed with the pY peptide (coupled to beads) was readily outcompeted by addition of GFP-Pk to the solution (Fig 4B, quantifed in 4C; see also bottom panel in Fig 4B showing that Pk did not outcompete Dsh binding to the Y-peptide). While these interaction studies are consistent with the co-IPs shown above, the in vitro peptide assay displayed the above mentioned caveat, as for example the unphorphorylated peptide was bound equally by both Dsh and Pk, when assayed independently (S4B Fig). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 4. A phospho-peptide of Y374 surronding sequences binds preferentially to Pk. (A) Pk interacts better with/binds to the phosphorylated “pY” peptide (S4 Fig for peptide sequence). Compare bands between input and pY-bound in greay boxes for Pk vs. for Dsh (red boxes). (B) Binding of Dsh (Dsh-GFP) is outcompeted by Pk (myc-Pk) on the “p-Y” (phospho-peptide), but not the non-phosphorylated (“Y”) peptide (bottom panel). Western blot showing retention of Dsh-GFP on beads coupled with the respective peptide, either “pY” for “Y” among mixing with Pk. Upper blot shows stable Dsh-GFP input, TZ is an unrelated control protein. Note reduced binding of Dsh-GFP when mixed in solution with increasing levels of Pk in the context of pY peptide conjugated beads (middle panel), but not the Y peptide (bottom panel). (C) Quantification of binding competition as shown in (B), combined from 3 independent replicates. Increasing Pk levels in the solution (upper graph) caused a decrease in Dsh binding to the “pY” peptide (lower graph). Statitistical analysis of binding differences (to lane #2): **p <0.005, and ****p<0.0001 (determined with one-way Anova). https://doi.org/10.1371/journal.pgen.1010849.g004 In summary and taken together, the above data are consistent with the model that (i) the Vang region centered around Y374 is necessary and sufficient for binding to Pk and Dsh, (ii) Pk preferentially interacts with this region when Y374 is phosphorylated, and (iii) Dsh requires the structural features of an aromatic ring at Y374 in its binding behavior to Vang. Furthermore, these experiments defined single amino acid mutations, Y374F and V376A, that selectively abrogated binding to either Pk (Y374F) or Dsh (V376A). These observations allow for the first time the functional testing of a direct Vang-Dsh interaction requirement in vivo. Single point mutants show PCP defects To investigate the functional consequences in vivo of altering the interaction of Vang with Pk and Dsh, we performed rescue experiments using transgenes carrying the different single point mutants. We expressed Vang-Flagx3 WT, Y374A, Y374F and V376A, using a direct tubulin-driven expression in a Vang null mutant background (Vang6). The control wild-type construct tub-VangWT was able to fully rescue cellular hair orientation in Drosophila wings (Fig 5A–5C, see 5D for quantification and statistical analyses; see 5B for Vang6 null mutant control). Each of the single point mutants, Vang-Y374A, Vang-Y374F and Vang-V376A, rescued the Vang6 loss-of-function phenotype partially, displaying varying degrees of cellular hair orientation defects (Fig 5C, quantified in 5D). The phenotypes were quantified using FijiWingsPolarity [54], which revealed phenotypic differences from VangWT control in each case, and also among all individual point mutations (compare panels in Fig 5C and quantifications in 5D). Importantly, while the rescue was partial in all three cases, the phenotypes were significantly different from the “non-rescued” null allele and the rescued VangWT wings, consistent with the notion that all three point mutants retained partial function as expected. Also, they all displayed different phenotypes, which were visibly and statitistally different from each other (and the null allele; Fig 5C and 5D), suggesting differing partial function, consistent with their distinct biochemical behavior. Consistent with the molecular binding data, Vang-Y374A, which should interfere with direct binding of either Pk or Dsh showed the most severe defects (see cellular hair angle distribution in Fig 5C and 5D), closest to the Vang null phenotype but nonetheless different (see statistical analyses in Fig 5D). The point mutations affecting binding to a single effector (Y374F only binding Dsh, and V376A only interacting with Pk) showed a better rescue as compared to Y374A (Fig 5C and 5D; see also Discussion). Of note, the phenotypic differences were not due to different expression levels, as all genotypes showed similar protein levels in larval disc lysates (Fig 5E). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 5. Single Vang point mutants in the Pk and Dsh interaction region display PCP defects in vivo. (A) Drosophila wing (distal is right) and (A’) higher magnification of boxed area in wing on left. Cellular orientation is reflected by single hairs pointing distally, also visualized by colored arrows reflecting cellular orientation in lower panel. Angles were determined using plugin FijiWingsPolarity (ref: [54]). (B) Image of same region as in A’ of a Vang6 (null mutant) wing. Note misoriented cellular hairs forming waves and whorls, and schematic refelction of mispolarization by color coded arrows. See colored bar in A depicting color scheme. Wild-type (distal) orientation is associated with turquoise in the 160–180° range. (C) Sample images of cellular (mis)orientation in adult wings of Vang6 mutant flies (same region as in A’-B) upon rescue with tubulin-promoter expressed Vang-Flagx3 transgenes with WT (wild-type), Y374A, Y374F or V376A, respectively as indicated. Angles of cellular orientation are visualized through colored arrows in lower panels (see color bar/scale in panel A above). Note that while Vang-WT fully rescues the mutant phenotype (and appears like wild-type, compare to A’), all single point mutants fail to rescue the mutant defects. Note also that Vang-Y374A most resembles the Vang null phenotype, and the other point mutations, display different phenotypes. (D) Quantification and statistical analyses of the different behaviors of the individual Vang binding mutants in the in vivo rescue assay of the Vang-/- null allele, shown as violin plots displaying cellular hair orientation angles for the indicated genotypes, within the ROI area shown in panels A’-C from 3 independent wings were combined. Statitistical analysis was determined with one-way Anova (and Tukey’s post test to compare all samples with each other): **p <0.005, and ****p<0.0001. Note that all genotypes show statitistically different phenotypes from all other gentypes, with the exception of wild-type and the VangWT transgene rescue of the Vang-/- null allele. (E) Western blot of wing disc lysates from the indicated genotypes. Note comparable levels of expression in all transgenes. https://doi.org/10.1371/journal.pgen.1010849.g005 To expand on the different behavior of the individual Vang point mutants in vivo, we next tested their effects in a gain-of-function (GOF) assay in developing wings. A classic feature of core PCP factors is that their overexpression also causes PCP defects. Each core factor has thus a reproducible LOF phenotype and also a GOF phenotype, both reflected in stereotypic defects highly similar from wing to wing of any given genotype. As such the Vang-wt GOF (via nubbin-Gal4) induces similar misorientations from wing to wing (in any specific region of a wing; see wing ROI samples in S5A–S5B Fig; boxed areas in blue and red in S5A Fig correspond to left and right panels in S5B Fig, respectively). As such, Vang overexpression, via nubbin-Gal4, of either Vang-wt or any Vang-point mutant (see indicated genotypes in S5 Fig), caused cellular orientation defects with wings of the same genotype looking similar to each other, but with different appearances between the distinct Vang-mutant genotypes (S5 Fig). This confirmed that, for example, Vang-Y374F (binding only Dsh) overexpression caused distinct defects from Vang-WT or Vang-YY373DD (binding predominantly Pk) or Vang-Y374A (binding neither factor). Accordingly, all specific Vang mutations tested displayed distinct behavior (S5B Fig, see quantifications in angle distribution graphs on right of each panel; note expression levels of each mutant protein were comparable, S5C Fig). Together with the rescue data (Fig 5), these results are consistent with the notion that Vang interacts physiologically with both cytoplasmic core PCP factors, Pk and Dsh, during planar polarity establishment in vivo. Together with the biochemical dissection, these data thus suggest that Y374 phosphorylation is a critical node for physiological activity of Vang during PCP establishment and for the resolution of core PCP complexes into the Vang and Fz associated “antagonistic polarity domains” (see also Discussion). Vang point mutants affect core PCP factor localization during PCP establishment To get further functional insight into how the Vang point mutations affect PCP establishment and core PCP factor localization in vivo, we next tested the localization of these mutants and associated other core factors. Using the rescue transgenes (with a tub-Vang-Flagx3 backbone) in a Vang6 null background, we analyzed the localization of Vang-Flagx3 WT, Y374A, Y374F and V376A, Fmi, EGFP-Pk, and DshGFP with E-cad as cellular outline control. The polarity of these proteins was evaluated using the Packing Analyzer suite [55] (Figs 6 and S6). Vang-WT displayed wild-type localization, polarized along the proximo-distal axis (Fig 6A) with all other core PCP factors tested, Fmi, Pk, and Dsh, also showing wild-type polarized, membrane associated distribution within the proximo-distal axis in pupal wing tissue at 28h APF (Figs 6A and S6A). In contrast all three point mutants displayed largely apolar (or mispolarized) distribution (Fig 6B–6D, mispolarization is highlighted in the angle quantification as seen in the rosette diagrams in panels 6A’-6D’). Apolar PCP core factor distribution was also evident in analyses of Fmi distribution (S6 Fig). Pk and Dsh protein distribution was also apolar (Figs 6B–6D and S6B–S6D, see also below). However, both cytoplasmic core PCP factors, Dsh and Pk, maintained significant membrane association in all point mutant backgrounds (Figs 6B–6D and S6B–S6D, see also below). All three Vang point mutants appeared membrane associated similar to Vang-WT (Fig 6), suggesting that Vang trafficking is not affected in these point mutants (see Discussion). While it is possible that differences between the individual Vang point mutants exist, also based on past observations for example [56], a carefull side-by-side comparison of the respective point mutations would be needed. Importantly, none of the point mutations caused a complete loss of membrane associated Pk or Dsh. This is consistent with previous data demonstrating that even in Vang null clones Pk is still partially retained at cellular membranes, likely due to being farnesylated and requiring this modification for function [12,48]. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 6. Vang point mutants affect localization of core PC factors in vivo. (A-D) Confocal images of 28-30hr APF pupal wings of the genotypes indicated above panels. All pupal wing tissue is oriented horizinally with proximal side being left and distal right, stained for Vang-Flagx3 (anti-Flag, red), EGFP-Pk (anti-GFP, green), and E-cad (blue). (A) Vang-/-; tub-VangWT-Flagx3, (B) Vang-/-; tub-VangY374A-Flagx3, (C) Vang-/-; tub-VangY374F-Flagx3, (D) Vang-/-; tub-VangV376A-Flagx3. Grayscale single channel micrographs are shown in lower panels as indicated on left side. Note that Pk is still localized to membranes in all three point mutants, albeit less enriched. Scale bars: 20μm. (A’-D’) Quantification of polarity angles presented as rosette diagrams of VangFlagx3 protein in the respective genotypes indicated: (A’) Vang-/-; tub-VangWT-Flagx3, (B’) Vang-/-; tub-VangY374A-Flagx3, (C’) Vang-/-; tub-VangY374F-Flagx3, (D’) Vang-/-; tub-VangV376A-Flagx3. Note that the majority of cell orientation angles is in the proximo-distal axis (horizontal, highlighted in red in rosette diagram) in the Vang-/-; tub-VangWT background (A’, indistinguishable from wild-type), but that cellular orientation is largely randomized in all three point mutants (B’-D’, very similar to the Vang null allele). See also S6 Fig for additional core factor stainings and polarity analyses in these genetic backgrounds. https://doi.org/10.1371/journal.pgen.1010849.g006 In parallel to in vivo localization studies, we also tested whether Vang/Vangl variants might differently affect the stability of core PCP membrane complexes, which are anchored at cell junctions by the atypical cadherin Fmi (Celsr1 in mice). To do this, we transfected mouse primary keratinocytes with Celsr1-GFP and mCherry-Vangl2, which co-localize to the junctional interfaces between expressing cell pairs mediated by homotypic adhesion of Celsr1 (S7A Fig) [14]. We then used fluorescence recovery after photocbleaching (FRAP) to determine the mobility of Celsr1 and Vangl2 proteins located within the junctional membrane. Compared to Celsr1 and Vangl2 proteins located at free edges of the cell, Celsr1 and Vangl2 were largely immobile at cell junctions (~80% immobile fraction)(S7B–S7C Fig), as previously shown [36,57]. Mutation of Vangl2 Y308 to A (equivalent to Drosophila VangY374A) did not affect its ability to localize to cell junctions (S7A Fig), nor did the mutation impact Vangl2 mobility at cell junctions, or that of co-expressed Celsr1-GFP (S7B–S7C Fig). We conclude that although the Y308 residue in Vangl2 can be phosphorylated in mammalian epidermal cells, it is not required for Vangl2 recruitment to or stabilization at cell junctions. Taken together, the in vivo analyses of the Vang point mutations confirm that the direct interactions with Pk and Dsh are critical for correct PCP establishment and polarized core PCP complex segregation and thus Vang function. However, they also confirmed that Vang is not essential for membrane recruitment of either cytoplasmic PCP factor and also that the phosphorylation status of its Y374 residue (Y308 in mice) has no or minimal impact on junctional PCP core complex stability (see Discussion). Discussion Here we identify through mass spectrometry analyses with mouse skin epidermis samples phosphorylation on mouse Vangl2 Y308 residue (equivalent to Y374 in Drosophila Vang). This tyrosine lies within with the overlapping binding regions of Pk and Dsh in Vang/Vangl, and, importantly, its charge/phosphorylation status regulates selective binding between Pk and Dsh, with phosphorylation tipping the balance towards Pk binding. We demonstrate in vivo that binding of Vang to both cytoplasmic core PCP factors is physiologically important (which is the first in vivo evidence for a Vang-Dsh binding requirement). Our study provides novel insight into the critical importance of Vang tyrosine phosphorylation and reveals mechanistic features of how regulation of the binding of antagonistic PCP factors to Vang/Vangl during the process of PCP complex segregation and polarity establishment is achieved. Phosphorylation of Y374/Y308 at the center of Vang/Vangl interactions with Pk and Dsh While previous work defined a broad region within the C-tail of Vang to interact with both Pk and Dsh [12,53,58–60], the mechanistic regulation and physiological significance of these interactions remained unresolved. Importantly, the defined region is conserved between Drosophila Vang and mammalian Vangl1/2 genes. Our data reveal that a small conserved stretch of amino acids within this broader region is both necessary (as shown in the whole Vang protein) and sufficient (as deduced from the in vitro peptide assays) to interact with both cytoplasmic core PCP factors. This region is well conserved between all Vang family members and centered on the tyrosine, which can be phosphorylated, as our mass spec data reveal. Mutational studies define that Pk binding is mediated by tyrosine phosphorylation and associated negative charge, while Dsh requires the aromatic ring found in tyrosine (and also phenylalanine) for its binding to Vang. It is worth noting that this Vang region, shared by both Pk and Dsh for binding, is specific for these two factors, as other Vang associated cytoplasmic PCP proteins, for example, Dgo and Scrib, are not affected by mutations within this domain. The importance of the Vang-Pk complex has been well documented in vivo and is also the core of one of the two stable PCP “core complexes” that result from PCP factor interactions and signaling (reviewed in [1–7]). In contrast, an interaction between Vang/Vangl and Dsh/Dvl family members has only been documented biochemically [12,53,58–60]. The dissection of binding requirements allowed us to generate single point mutations in Vang that separate binding to only one of the cytoplasmic factors, either Pk or Dsh. The associated in vivo rescue experiments provided the possibility for physiological testing of a functional requirement of the individual interactions between Vang and Dsh or Pk. While all three point mutations display partial rescue, their function is reduced and thus the respective amino acids are physiologically required. Interfering with binding of Vang to both factors (VangY374A) shows the weakest rescue, with the mutant displaying defects that are more similar to the Vang- null phenotype than the other point mutants. Nevertheless, as the point mutations affecting individual Vang-Pk or Vang-Dsh interactions also displayed only partial rescue of the Vang-/- defects, these data indicate that interactions with either cytoplasmic PCP factor, Pk and Dsh, are critical for in vivo Vang function in PCP core complex localization and hence PCP establishment. The phenotypic defects with the Vang-V376A mutant seen in adult wings suggest that it might behave as a mild neomorph (see Fig 5C and 5D), although no dominant effect was observed when heterozygous over Vang-WT. Importantly, this is the first physiological in vivo evidence demonstrating that Vang has a requirement to interact with Dsh during PCP complex segregation. Of all point mutants tested, Vang-Y374F, showed the strongest partial rescue (appeared closest to wild-type). This is consistent with the notion that Pk does not strictly require binding to Vang for membrane association [12,48], and that formation of a Vang-Pk complex is much more complcated than a single interaction between the two proteins. Of note, each single point mutant affects the formation of the stable polarized core PCP complexes, as evident in the protein localization studies in 28h APF pupal wings, suggesting that interfering with any interaction among the core PCP factors causes a significant disruption to the PCP interaction cascade and network, needed for normal asymmetric complex polarizations. How does Vang Y374/Y308 phosphorylation affect PCP complex resolution? It is intriguing to think about how phosphorylation, and lack thereof, affects the formation of stable core PCP complexes. Our data indicate that binding of Dsh/Dvl to Vang/Vangl is physiological, and yet in standard co-staining studies Vang and Dsh do not co-localize. How does Dsh binding to an unphosphorylated Y374 region affect core PCP complex formation? There are a few potential scenarios, and importantly Vang is not a major membrane recruiter of Pk, if at all [12,48] [and this work] and thus other factors likely contribute to this. First, a Vang-Pk association, which we assume is stable in wing cells in the proximal junctional membrane region, should likely not form in other areas of the cell membrane. As such a Vang-Dsh transient/intermediary interaction might serve a function to prevent Vang-Pk binding. If the kinase in question is asymmetrically localized or active, for example in the proximal area, then—and only then—a switch from Vang-Dsh to a Vang-Pk interaction would occur. If the kinase in question is not asymmetrically active or localized, Dsh binding to this Vang region might be required to prevent the kinase to act on Vang in cell membrane domains, where formation of a Vang-Pk complex should not form, for example the distal vertex of a wing cell. In such a mechanistic scenario Dsh would keep Vang/Vangl “flexible” to find the right cellular context, where/when the presence of the kinase would initiate the switch to a Vang-P at Y374 (Y308 in mVangl2) and thus support an interaction with Pk and its local effectors. While these are intriguing mechanistic models, they remain speculative. In general, the function of Vang/Vangl proteins in PCP establishment remains unresolved. While Vang family proteins are critical for the process and they can physically interact with all other core PCP factors, their contribution to the stability of the intercellular junctional complexes remains unclear, which seem to mainly require Fz-Fmi::Fmi interactions [34,61], although Vang/Vangl proteins are part of these asymmetric complexes and bind to Fz intercellularly [35]. Moreover, the non-stoichiometric manner in which the stable PCP complexes form [62], with for example one single Pk molecule per 6 Vang molecules, suggests complicated mechanistic scenarios that do not rely on one-to-one protein interactions. Importantly, the formation and maintenance of stable PCP complexes requires also Dgo [33] (Diversin in vertebrates; [63]) and extends beyond interactions among the core factors, including Scribble [31,32] and CK1ε [45–47,49] and many regulatory interactions are still to be discovered (see below). Complex in vivo experiments will be necessary to better understand the mechanistic sequence of events. What is the tyrosine kinase acting on Vang/Vangl? It is currently unclear which tyrosine kinase(s) act on Vang to mediate its phosphorylation on the Y374 residue (Y308 in mVangl2) and this is one of the regulatory interactions to be still discovered. Sequence motif searches in the Vang Y374 flaking region suggest that Src family kinases could be involved, with no other kinase family having a higher probability (by sequence alignment searches). It is however technically difficult to prove that Src kinases indeed act on this Vang residue in a physiological context (see below), and unfortunately in vitro kinase assays have proven uninformative, as most tyrosine kinases tested could phosphorylate Vang in such assays on multiple residues. Redundancy of Src kinases is an issue in in vivo studies in both our systems, mouse skin and Drosophila wing epithelia, as there are several Src family kinases in both Drosophila and mice, for example [64–68]. Moreover, in addition to cell survival requirements, many cellular functions are associated with Src family kinases (rev in e.g. [69–71]. For example, in Drosophila the two main Src family members are either viable with no overt developmental phenotypes in imaginal discs (Src64, redundant with Src42) or are largely cell lethal (Src42) when analyzed in vivo [64,68]. They have also been linked to a vast variety of cellular functions, ranging from cytoskeletal regulation and cell adhesion, to synaptic plasticity, proliferation, cell death, and others (reviewed in [69–71]. Src kinases remain nonetheless likely candidate(s), as we (i) observe GOF phenotypes consistent with a PCP function and (ii) genetic interactions with these Src GOF defects suggest that Vang is required in these contexts. However, again, a loss-of-function scenario to really demonstrate a Src function in PCP establishment remains elusive and should be the focus of future studies. Phosphorylation of Y374/Y308 at the center of Vang/Vangl interactions with Pk and Dsh While previous work defined a broad region within the C-tail of Vang to interact with both Pk and Dsh [12,53,58–60], the mechanistic regulation and physiological significance of these interactions remained unresolved. Importantly, the defined region is conserved between Drosophila Vang and mammalian Vangl1/2 genes. Our data reveal that a small conserved stretch of amino acids within this broader region is both necessary (as shown in the whole Vang protein) and sufficient (as deduced from the in vitro peptide assays) to interact with both cytoplasmic core PCP factors. This region is well conserved between all Vang family members and centered on the tyrosine, which can be phosphorylated, as our mass spec data reveal. Mutational studies define that Pk binding is mediated by tyrosine phosphorylation and associated negative charge, while Dsh requires the aromatic ring found in tyrosine (and also phenylalanine) for its binding to Vang. It is worth noting that this Vang region, shared by both Pk and Dsh for binding, is specific for these two factors, as other Vang associated cytoplasmic PCP proteins, for example, Dgo and Scrib, are not affected by mutations within this domain. The importance of the Vang-Pk complex has been well documented in vivo and is also the core of one of the two stable PCP “core complexes” that result from PCP factor interactions and signaling (reviewed in [1–7]). In contrast, an interaction between Vang/Vangl and Dsh/Dvl family members has only been documented biochemically [12,53,58–60]. The dissection of binding requirements allowed us to generate single point mutations in Vang that separate binding to only one of the cytoplasmic factors, either Pk or Dsh. The associated in vivo rescue experiments provided the possibility for physiological testing of a functional requirement of the individual interactions between Vang and Dsh or Pk. While all three point mutations display partial rescue, their function is reduced and thus the respective amino acids are physiologically required. Interfering with binding of Vang to both factors (VangY374A) shows the weakest rescue, with the mutant displaying defects that are more similar to the Vang- null phenotype than the other point mutants. Nevertheless, as the point mutations affecting individual Vang-Pk or Vang-Dsh interactions also displayed only partial rescue of the Vang-/- defects, these data indicate that interactions with either cytoplasmic PCP factor, Pk and Dsh, are critical for in vivo Vang function in PCP core complex localization and hence PCP establishment. The phenotypic defects with the Vang-V376A mutant seen in adult wings suggest that it might behave as a mild neomorph (see Fig 5C and 5D), although no dominant effect was observed when heterozygous over Vang-WT. Importantly, this is the first physiological in vivo evidence demonstrating that Vang has a requirement to interact with Dsh during PCP complex segregation. Of all point mutants tested, Vang-Y374F, showed the strongest partial rescue (appeared closest to wild-type). This is consistent with the notion that Pk does not strictly require binding to Vang for membrane association [12,48], and that formation of a Vang-Pk complex is much more complcated than a single interaction between the two proteins. Of note, each single point mutant affects the formation of the stable polarized core PCP complexes, as evident in the protein localization studies in 28h APF pupal wings, suggesting that interfering with any interaction among the core PCP factors causes a significant disruption to the PCP interaction cascade and network, needed for normal asymmetric complex polarizations. How does Vang Y374/Y308 phosphorylation affect PCP complex resolution? It is intriguing to think about how phosphorylation, and lack thereof, affects the formation of stable core PCP complexes. Our data indicate that binding of Dsh/Dvl to Vang/Vangl is physiological, and yet in standard co-staining studies Vang and Dsh do not co-localize. How does Dsh binding to an unphosphorylated Y374 region affect core PCP complex formation? There are a few potential scenarios, and importantly Vang is not a major membrane recruiter of Pk, if at all [12,48] [and this work] and thus other factors likely contribute to this. First, a Vang-Pk association, which we assume is stable in wing cells in the proximal junctional membrane region, should likely not form in other areas of the cell membrane. As such a Vang-Dsh transient/intermediary interaction might serve a function to prevent Vang-Pk binding. If the kinase in question is asymmetrically localized or active, for example in the proximal area, then—and only then—a switch from Vang-Dsh to a Vang-Pk interaction would occur. If the kinase in question is not asymmetrically active or localized, Dsh binding to this Vang region might be required to prevent the kinase to act on Vang in cell membrane domains, where formation of a Vang-Pk complex should not form, for example the distal vertex of a wing cell. In such a mechanistic scenario Dsh would keep Vang/Vangl “flexible” to find the right cellular context, where/when the presence of the kinase would initiate the switch to a Vang-P at Y374 (Y308 in mVangl2) and thus support an interaction with Pk and its local effectors. While these are intriguing mechanistic models, they remain speculative. In general, the function of Vang/Vangl proteins in PCP establishment remains unresolved. While Vang family proteins are critical for the process and they can physically interact with all other core PCP factors, their contribution to the stability of the intercellular junctional complexes remains unclear, which seem to mainly require Fz-Fmi::Fmi interactions [34,61], although Vang/Vangl proteins are part of these asymmetric complexes and bind to Fz intercellularly [35]. Moreover, the non-stoichiometric manner in which the stable PCP complexes form [62], with for example one single Pk molecule per 6 Vang molecules, suggests complicated mechanistic scenarios that do not rely on one-to-one protein interactions. Importantly, the formation and maintenance of stable PCP complexes requires also Dgo [33] (Diversin in vertebrates; [63]) and extends beyond interactions among the core factors, including Scribble [31,32] and CK1ε [45–47,49] and many regulatory interactions are still to be discovered (see below). Complex in vivo experiments will be necessary to better understand the mechanistic sequence of events. What is the tyrosine kinase acting on Vang/Vangl? It is currently unclear which tyrosine kinase(s) act on Vang to mediate its phosphorylation on the Y374 residue (Y308 in mVangl2) and this is one of the regulatory interactions to be still discovered. Sequence motif searches in the Vang Y374 flaking region suggest that Src family kinases could be involved, with no other kinase family having a higher probability (by sequence alignment searches). It is however technically difficult to prove that Src kinases indeed act on this Vang residue in a physiological context (see below), and unfortunately in vitro kinase assays have proven uninformative, as most tyrosine kinases tested could phosphorylate Vang in such assays on multiple residues. Redundancy of Src kinases is an issue in in vivo studies in both our systems, mouse skin and Drosophila wing epithelia, as there are several Src family kinases in both Drosophila and mice, for example [64–68]. Moreover, in addition to cell survival requirements, many cellular functions are associated with Src family kinases (rev in e.g. [69–71]. For example, in Drosophila the two main Src family members are either viable with no overt developmental phenotypes in imaginal discs (Src64, redundant with Src42) or are largely cell lethal (Src42) when analyzed in vivo [64,68]. They have also been linked to a vast variety of cellular functions, ranging from cytoskeletal regulation and cell adhesion, to synaptic plasticity, proliferation, cell death, and others (reviewed in [69–71]. Src kinases remain nonetheless likely candidate(s), as we (i) observe GOF phenotypes consistent with a PCP function and (ii) genetic interactions with these Src GOF defects suggest that Vang is required in these contexts. However, again, a loss-of-function scenario to really demonstrate a Src function in PCP establishment remains elusive and should be the focus of future studies. Materials and methods Mouse lines and embryonic skin harvesting Ethics statement. All mouse work in this study was approved by Princeton University’s Institutional Animal Care and Use Committee (IACUC) under protocol number 1867. K14-GFP-Vangl2 transgenic mice (FVB background, [52] were housed in an AAALAC-accredited facility following the Guide for the Care and Use of Laboratory Animals. E15.5 embryos were harvested from K14-Vangl2-GFP heterozygous dams in cold PBS and screened for GFP expression using a stereomicroscope equipped with epifluorescence. Full thickness backskins were dissected from both GFP+ and GFP- littermates embryos and flash frozen immediately in liquid nitrogen (LN2). LN2 was removed by evaporation and frozen backskins were stored for up to 3 months at -80C until cryolysis. Epidermal cryolysis and immunoprecipitation of GFP-Vangl2. Frozen skin samples pooled from four (for IP-Western) or eight (for IP-MS) GFP-Vangl2 and control backskins were processed via cryolysis using a CryoMill (Retch). Briefly, 2ml LN2-frozen lysis buffer droplets (Tissue Extraction Buffer, 1% Triton X-100, 10mM EDTA, 0.3mg/ml PMSF plus protease inhibitors in PBS) were mixed with frozen skin samples and processed by cryogenic grinding for 20 min using a ball mill cooled with LN2. Finely ground frozen epidermal-lysis buffer mixtures were lysed by thawing on ice for 1.5-2hrs. Lysates were cleared by adding 50ul Pansorbin and centrifuging at 14K rpm for 10min. Pre-cleared lysates were transferred to pre-washed, αGFP (rabbit αGFP, AbCam) antibody bound beads and incubated while rotating for 3 hours at 4 degrees C. Immunoprecipitates were washed twice with lysis buffer and eluted with 40ul 2X SDS sample buffer. For western blotting, 40ul total lysate and 40ul IP were run on a 10% SDS-PAGE gel and transferred to a PDVF membrane, blocked for 1hr at room temperature, incubated with primary, chicken αGFP (1:5000, Abcam) overnight at 4°C. After several washes in PBS-T, membrane was incubated for 45 min at room temperature with HRP α-chicken secondaries (1:2500), developed with BioRad Clarity ECL reagent and imaged on both film and with BioRad imager. For IP-MS samples, 20ul total lysate and 40ul IP were run on a 7.5% SDS-PAGE gel. Gel was fixed in 50% MeOH + 7% Acetic Acid for 30–60 minutes, rinsed with H20, stained with SPYRO Ruby overnight, and washed twice for 5 min each in 10% MeOH + 7% Acetic Acid. Gel imaging was performed on a Typhoon FLA-7000 (GE Healthcare). Proteomic sample preparation and mass spectrometry. IP Bands at ~85-90KD were excised then diced, and subjected to in-gel thiol reduction/alkylation and trypsin digestion using a method adapted from Shevchenko et al [72] to process samples for LC-MS/MS. Briefly, gel cubes were destained and washed extensively in 100 mM ammonium bicarbonate buffer, pH 8.8 (ABC), treated with 50 mM TCEP in ABC for 1 h at 55°C, washed, subjected to alkylation with 55 mM iodoacetamide in ABC for 30 min at room temperature in the dark, washed, and finally digested overnight with 1 ug Promega Trypsin Gold (Promega) per gel slice. Peptides from in-gel digest eluates were desalted using STAGE-Tips 9 prior to LC-MS analyses. LC-MS/MS analyses were performed on a high-resolution, high-mass-accuracy, reversed-phase nano-UPLC-MS platform, consisting of an Easy nLC Ultra 1000 nano-UPLC system coupled to an Orbi Elite mass spectrometer (ThermoFisher Scientific) equipped with a Flex Ion source (Proxeon Biosystems, Odense, Denmark). LC was conducted using a trapping capillary column (150 μm x ca. 40 mm, packed with 3 μm, 100 Å Magic AQ C18 resin, Michrom, Auburn, CA) at a flow rate of 5 μL/min for 4 min, followed by an analytical capillary column (75 μm x ca. 45 cm, packed with 3 μm, 100 Å Magic AQ C18 resin, Michrom) under a linear gradient of A and B solutions (solution A: 3% acetonitrile/ 0.1% formic acid; solution B: 97% acetonitrile/ 0.1% formic acid) from 5%-35% B over 90 at a flow rate of 300 nL/min. Nanospray was achieved using Picospray tips (New Objective, Woburn, MA) at a voltage of 2.4 kV, with the Elite heated capillary at 275°C. Full-scan (m/z 335–1800) positive-ion mass spectra were acquired in the Orbitrap at a resolution setting of 120,000. MS/MS spectra were simultaneously acquired using CID in the LTQ for the fifteen most abundant multiply charged species in the full-scan spectra, having signal intensities of >1000 NL. To aid in phosphosite mapping, Vangl2-positive slices were subjected LC-MS/MS over a 180 min gradient using the CID parameters above and also a second round of LC-MS/MS over a 180 min gradient during which MS/MS spectra were acquired by multistage activation (MSA) for the top 10 most abundant ions in the full-scan spectra, using excitation at the precursor m/z value as well as those corresponding to the neutral losses of phosphonic and phosphoric acids for ions of charge +2 and +3. Lockmass was employed, maintaining calibration to 2–3 ppm of accurate mass. Mass spectrometric data analysis. Resultant LC-MS/MS raw data files were processed using ProteomeDiscoverer (v. 1.4, ThermoFisher), to match MS/MS spectra against the UniProt Mus musculus database, or a GFP-Vangl2 fusion protein construct subset database using the Mascot search engine (v. 2.4, Matrix Science, London, UK.), allowing for a parent ion mass window of ±6 ppm, ≤ 3 missed trypsin cleavages, serine, threonine and tyrosine phosphorylation, methionine oxidation, asparagine and glutamine deamination and N-terminal protein acetylation as variable modifications, and carbamidomethylation of cysteines as a fixed modification. Peptide assignment cut-offs were specified at a high confidence level (<1% FDR). Phosphosite localization confidence scoring was achieved using the PhosphoRS 9 [73] (v. 3.1) node within the ProteomeDiscoverer framework. Relative abundance levels for proteins between experimental and controls were estimated using spectral counting. Raw mass spectra were visualized using Xcalibur (v. 2.2, ThermoFisher) and peptide spectral matches were visualized using ProteomeDiscoverer or Scaffold (v. 4.3.4, Proteome Software, Portland, OR). All phosphopeptide assignments were further validated by manual inspection. DNA constructs and S2 culture. Constructs used were as follows; pAc5.1-Flagx3, pAc5.1-Myc-Pk, pAc5.1-Dsh-GFP, pAc5.1-Scrib-PDZ-3-4-HA and pAc5.1-HA-Dgo (all gifts from Dr. Jenny, AECOM, USA) and pAc5.1-Vang-Flagx3 [44]. GFP-TZ was used as a control, and is the Drosophila ciliary protein Mks1. pAc5.1-Vang-Flagx3-Y374A, FKYY371AAYA, Y374F and V376A were generated though site-directed mutagenesis. C-terminal truncations of Vang were generated through PCR and cloned into pAc5.1-Flagx3 using NotI-XbaI restriction sites. All primers used are available upon request. Unless otherwise stated, lysates were prepared from S2 cells. S2 cells were maintained according to standard protocols, and were grown in Schneider’s Medium (Gibco) supplemented with 10% heat-inactivated Fetal Bovine Serum (Gibco). Cells were plated in 12 well plates at a dilution of 1.5x106 and were transfected with the indicated constructs using Effectene (Qiagen) according to manufacturer’s protocols. Cells were lysed ~48 hrs later in buffer containing 50mM Tris-HCl pH 7.5, 150mM NaCl, 1mM EDTA and 1% Triton-X-100. Pull-downs, immunoblotting and peptide interactions. For Flag pull-down experiments, lysates were extracted from S2 cells or from larval wing discs and were incubated at 4°C overnight with 10μl anti-Flag M2 affinity gel per sample. Washes were performed in buffer containing 50mM Tris-HCl pH 7.5, 350mM NaCl, 1mM EDTA, 0.1% SDS before elution in final sample buffer. For peptide binding, Vang peptides were conjugated to beads to enable assessment of direct binding. Lysates were generated in S2 cells, and the lysate was divided equally between tubes containing either 20μl of conjugated phospho-peptide or the unphosphorylated form and incubated at 4°C overnight. Washes were performed in buffer containing 10mM Tris-HCl pH7.5, 350mM NaCl, and 0.5mM EDTA before elution in final sample buffer. For GFP pull-downs. Samples were resolved by polyacrylamide gel electrophoresis according to standard protocols. The following primary antibodies were used for immunoblotting; Flag (Sigma M2 1:5000), Gamma-tubulin (Sigma GTU-88 1:1000), GFP (Roche 7.1&13.1 1:1000), Myc (SCBT 9E10 1:1000), Phospho-tyrosine (Millipore PY20 1:1000). Drosophila strains, dissections and phenotypic analyses. Flies were raised on standard medium, and maintained at 25°C unless otherwise stated. To generate UAS-Vang-Flagx3 mutant transgenic flies, site-directed mutagenesis was performed on vector pUAST-Vang-Flagx3-attB [44], before insertion into BDSC stock number 9750. To generate tub-Vang-Flagx3 mutant flies, site-directed mutagenesis was performed on vector pCaSpeR-tub-Vang-Flagx3 [47]. All transgenic strains were generated via Bestgene Inc. Wing discs were dissected from third instar larvae and prepared through incubation in final sample buffer at 95°C. For phosphatase treatment, wing discs were collected in PBS and transferred to lysis buffer containing 50mM Tris-HCl pH 7.5, 150mM NaCl, 1mM EDTA, and 1% Triton-X. Lysates were then incubated at 30°C for 30 minutes with lambda protein phosphatase (NEB). Adult wings were collected in PBS containing 0.1% Triton-X-100 (PBST) and incubated for 1 hr at room temperature before mounting in 80% glycerol in PBS. Hair orientation was quantified using the FijiWingsPolarity plugin [54]. Western blot quantification and statistical analyses. Western blot quantifications were made in ImageJ, measuring the intensity of each band and substracting the background. Each measurement was normalized to the loading control(s). To determine statistical differences between data sets of continuous data, we performed nonparametric ANOVA and Tukey’s post test in GraphPad. P-values lower than 0.05 were considered significant, and levels of significance are indicated by number of asterisks (see Figure legends for specific details). The error bars represent the SEM. Drosophila pupal wing fluorescence stainings and confocal analyses. White pupae were collected (0h APF/after puparium formation) and aged at 25°C until dissection at 28-30h APF. Dissections were performed as follows: in brief, pupae were immobilized on double-sided tape, removed from the pupal case, and placed into PBS, in which pupae were partially dissected to remove fat tissue, fixed in 4% paraformaldehyde in PBS for 45 min at RT or overnight at 4°C, and washed 3x in PBS and 0.1% Triton X-100. Wing membranes were removed, and immunostaining was performed by standard techniques. In brief, tissue was incubated in wash buffer containing 10% normal goat serum overnight for primary antibody (4°C), washed 3x with PBS, and incubated with the secondary antibody for at least 2 h (25°C) and fluorescent phalloidin for staining the actin cytoskeleton. Wings were washed 3x with PBS and mounted in Mowiol (see also [74] for more details). Pupal-wing images were acquired at RT using a confocal microscope, Leica SP8. Images were processed with ImageJ (National Institute of Health) and analyzed using the Packing Analyzed software Fiji plug-in [55]. Fluorescence recovery after photobleaching (FRAP) analysis. Approximately 150,000 wild type CD1 keratinocytes were seeded in no.1.5 glass bottom dishes (ibidi #81151) coated with fibronectin. 20–24 hours post-plating, cells were co-transfected with Celsr1-GFP and mCherry-Vangl2 or Celsr1-GFP and mCherry-Vangl2Y308A plasmids. 24 hours after transfection, cells were switched to E-media containing 1.5mM Ca2+ and incubated for an additional 20–24 hours for adequate Celsr1-GFP expression and keratinocytes to form stable cell-cell junctions. Before imaging, cells were switched to phenol-red free E-media with 1.5mM Ca2+. Cells were imaged using a 488nm and 561nm laser, 60X magnification objective (with additional zoom that rendered a pixel size of 110nm) on Nikon A1R-STED confocal microscope equipped with a stage-top Tokai-Hit incubation chamber to maintain 37 degrees and 5% CO2. Keeping magnification, laser power (both for bleach and acquisition), pixel dwell time and acquisition rate constant across all measurements, 1um diameter circular bleach ROIs and one ROI per junction(s) or cells edge(s) were created for bleaching and recovery. The FRAP acquisition sequence consisted of 3 reference pre-bleach images followed by bleach using 405nm laser and finally 70 frames with 4 seconds intervals to monitor fluorescence recovery at junctions. The acquired images in the time series were checked for Z-drift and corrected for presence of any XY drift in Fiji. A reference ROI was made in a non-bleached region to correct for overall bleaching during image acquisition. A background ROI was created outside the fluorescent cell in each image. The ROI values were extracted from drift corrected images in NIS elements software and subsequently processed in Microsoft excel and Graphpad Prism. Each image time series was background and bleach corrected (to be referred as corrected intensity henceforth) and thereafter the corrected intensity profile was normalized as (Ft−Fbleach)/ (Fini−Fbleach), where, Ft is the corrected intensity of the ROI at a given time point, Fbleach is the corrected intensity at the time point immediately after bleaching, Fini is the mean ROI intensity of the three pre-bleach frames. Each mean recovery curve was fitted to exponential one phase association equation in Graphpad Prism and the fitted Plateau and Y0 values were used to determine the immobile fraction = 1- {(Plateau-Y0)/(1- Y0)}. The averaged traces for each condition was fitted to the model with an r-squared value > 0.9. Data represented is pooled from four independent experiments. Cell culture and transfection. Primary wild type CD1 Keratinocytes were cultured using previously published protocol [75]. Keratinocytes were in E-Media prepared in the laboratory (see [75] for composition) supplemented with 50μM Calcium Chloride. For FRAP experiments, phenol-red free DMEM and HF-12 were used to prepare pigment-free imaging E-media and supplemented with 1.5mM Calcium Chloride. Cells were transfected using Effectene reagent following a modified manufacturer’s protocol. 400ng DNA comprising of Celsr1-GFP and mCherry-Vangl2 [14] or mCherry-Vangl2Y308A (this paper) in a ratio of 2:1 was used for co-transfection. Mouse lines and embryonic skin harvesting Ethics statement. All mouse work in this study was approved by Princeton University’s Institutional Animal Care and Use Committee (IACUC) under protocol number 1867. K14-GFP-Vangl2 transgenic mice (FVB background, [52] were housed in an AAALAC-accredited facility following the Guide for the Care and Use of Laboratory Animals. E15.5 embryos were harvested from K14-Vangl2-GFP heterozygous dams in cold PBS and screened for GFP expression using a stereomicroscope equipped with epifluorescence. Full thickness backskins were dissected from both GFP+ and GFP- littermates embryos and flash frozen immediately in liquid nitrogen (LN2). LN2 was removed by evaporation and frozen backskins were stored for up to 3 months at -80C until cryolysis. Epidermal cryolysis and immunoprecipitation of GFP-Vangl2. Frozen skin samples pooled from four (for IP-Western) or eight (for IP-MS) GFP-Vangl2 and control backskins were processed via cryolysis using a CryoMill (Retch). Briefly, 2ml LN2-frozen lysis buffer droplets (Tissue Extraction Buffer, 1% Triton X-100, 10mM EDTA, 0.3mg/ml PMSF plus protease inhibitors in PBS) were mixed with frozen skin samples and processed by cryogenic grinding for 20 min using a ball mill cooled with LN2. Finely ground frozen epidermal-lysis buffer mixtures were lysed by thawing on ice for 1.5-2hrs. Lysates were cleared by adding 50ul Pansorbin and centrifuging at 14K rpm for 10min. Pre-cleared lysates were transferred to pre-washed, αGFP (rabbit αGFP, AbCam) antibody bound beads and incubated while rotating for 3 hours at 4 degrees C. Immunoprecipitates were washed twice with lysis buffer and eluted with 40ul 2X SDS sample buffer. For western blotting, 40ul total lysate and 40ul IP were run on a 10% SDS-PAGE gel and transferred to a PDVF membrane, blocked for 1hr at room temperature, incubated with primary, chicken αGFP (1:5000, Abcam) overnight at 4°C. After several washes in PBS-T, membrane was incubated for 45 min at room temperature with HRP α-chicken secondaries (1:2500), developed with BioRad Clarity ECL reagent and imaged on both film and with BioRad imager. For IP-MS samples, 20ul total lysate and 40ul IP were run on a 7.5% SDS-PAGE gel. Gel was fixed in 50% MeOH + 7% Acetic Acid for 30–60 minutes, rinsed with H20, stained with SPYRO Ruby overnight, and washed twice for 5 min each in 10% MeOH + 7% Acetic Acid. Gel imaging was performed on a Typhoon FLA-7000 (GE Healthcare). Proteomic sample preparation and mass spectrometry. IP Bands at ~85-90KD were excised then diced, and subjected to in-gel thiol reduction/alkylation and trypsin digestion using a method adapted from Shevchenko et al [72] to process samples for LC-MS/MS. Briefly, gel cubes were destained and washed extensively in 100 mM ammonium bicarbonate buffer, pH 8.8 (ABC), treated with 50 mM TCEP in ABC for 1 h at 55°C, washed, subjected to alkylation with 55 mM iodoacetamide in ABC for 30 min at room temperature in the dark, washed, and finally digested overnight with 1 ug Promega Trypsin Gold (Promega) per gel slice. Peptides from in-gel digest eluates were desalted using STAGE-Tips 9 prior to LC-MS analyses. LC-MS/MS analyses were performed on a high-resolution, high-mass-accuracy, reversed-phase nano-UPLC-MS platform, consisting of an Easy nLC Ultra 1000 nano-UPLC system coupled to an Orbi Elite mass spectrometer (ThermoFisher Scientific) equipped with a Flex Ion source (Proxeon Biosystems, Odense, Denmark). LC was conducted using a trapping capillary column (150 μm x ca. 40 mm, packed with 3 μm, 100 Å Magic AQ C18 resin, Michrom, Auburn, CA) at a flow rate of 5 μL/min for 4 min, followed by an analytical capillary column (75 μm x ca. 45 cm, packed with 3 μm, 100 Å Magic AQ C18 resin, Michrom) under a linear gradient of A and B solutions (solution A: 3% acetonitrile/ 0.1% formic acid; solution B: 97% acetonitrile/ 0.1% formic acid) from 5%-35% B over 90 at a flow rate of 300 nL/min. Nanospray was achieved using Picospray tips (New Objective, Woburn, MA) at a voltage of 2.4 kV, with the Elite heated capillary at 275°C. Full-scan (m/z 335–1800) positive-ion mass spectra were acquired in the Orbitrap at a resolution setting of 120,000. MS/MS spectra were simultaneously acquired using CID in the LTQ for the fifteen most abundant multiply charged species in the full-scan spectra, having signal intensities of >1000 NL. To aid in phosphosite mapping, Vangl2-positive slices were subjected LC-MS/MS over a 180 min gradient using the CID parameters above and also a second round of LC-MS/MS over a 180 min gradient during which MS/MS spectra were acquired by multistage activation (MSA) for the top 10 most abundant ions in the full-scan spectra, using excitation at the precursor m/z value as well as those corresponding to the neutral losses of phosphonic and phosphoric acids for ions of charge +2 and +3. Lockmass was employed, maintaining calibration to 2–3 ppm of accurate mass. Mass spectrometric data analysis. Resultant LC-MS/MS raw data files were processed using ProteomeDiscoverer (v. 1.4, ThermoFisher), to match MS/MS spectra against the UniProt Mus musculus database, or a GFP-Vangl2 fusion protein construct subset database using the Mascot search engine (v. 2.4, Matrix Science, London, UK.), allowing for a parent ion mass window of ±6 ppm, ≤ 3 missed trypsin cleavages, serine, threonine and tyrosine phosphorylation, methionine oxidation, asparagine and glutamine deamination and N-terminal protein acetylation as variable modifications, and carbamidomethylation of cysteines as a fixed modification. Peptide assignment cut-offs were specified at a high confidence level (<1% FDR). Phosphosite localization confidence scoring was achieved using the PhosphoRS 9 [73] (v. 3.1) node within the ProteomeDiscoverer framework. Relative abundance levels for proteins between experimental and controls were estimated using spectral counting. Raw mass spectra were visualized using Xcalibur (v. 2.2, ThermoFisher) and peptide spectral matches were visualized using ProteomeDiscoverer or Scaffold (v. 4.3.4, Proteome Software, Portland, OR). All phosphopeptide assignments were further validated by manual inspection. DNA constructs and S2 culture. Constructs used were as follows; pAc5.1-Flagx3, pAc5.1-Myc-Pk, pAc5.1-Dsh-GFP, pAc5.1-Scrib-PDZ-3-4-HA and pAc5.1-HA-Dgo (all gifts from Dr. Jenny, AECOM, USA) and pAc5.1-Vang-Flagx3 [44]. GFP-TZ was used as a control, and is the Drosophila ciliary protein Mks1. pAc5.1-Vang-Flagx3-Y374A, FKYY371AAYA, Y374F and V376A were generated though site-directed mutagenesis. C-terminal truncations of Vang were generated through PCR and cloned into pAc5.1-Flagx3 using NotI-XbaI restriction sites. All primers used are available upon request. Unless otherwise stated, lysates were prepared from S2 cells. S2 cells were maintained according to standard protocols, and were grown in Schneider’s Medium (Gibco) supplemented with 10% heat-inactivated Fetal Bovine Serum (Gibco). Cells were plated in 12 well plates at a dilution of 1.5x106 and were transfected with the indicated constructs using Effectene (Qiagen) according to manufacturer’s protocols. Cells were lysed ~48 hrs later in buffer containing 50mM Tris-HCl pH 7.5, 150mM NaCl, 1mM EDTA and 1% Triton-X-100. Pull-downs, immunoblotting and peptide interactions. For Flag pull-down experiments, lysates were extracted from S2 cells or from larval wing discs and were incubated at 4°C overnight with 10μl anti-Flag M2 affinity gel per sample. Washes were performed in buffer containing 50mM Tris-HCl pH 7.5, 350mM NaCl, 1mM EDTA, 0.1% SDS before elution in final sample buffer. For peptide binding, Vang peptides were conjugated to beads to enable assessment of direct binding. Lysates were generated in S2 cells, and the lysate was divided equally between tubes containing either 20μl of conjugated phospho-peptide or the unphosphorylated form and incubated at 4°C overnight. Washes were performed in buffer containing 10mM Tris-HCl pH7.5, 350mM NaCl, and 0.5mM EDTA before elution in final sample buffer. For GFP pull-downs. Samples were resolved by polyacrylamide gel electrophoresis according to standard protocols. The following primary antibodies were used for immunoblotting; Flag (Sigma M2 1:5000), Gamma-tubulin (Sigma GTU-88 1:1000), GFP (Roche 7.1&13.1 1:1000), Myc (SCBT 9E10 1:1000), Phospho-tyrosine (Millipore PY20 1:1000). Drosophila strains, dissections and phenotypic analyses. Flies were raised on standard medium, and maintained at 25°C unless otherwise stated. To generate UAS-Vang-Flagx3 mutant transgenic flies, site-directed mutagenesis was performed on vector pUAST-Vang-Flagx3-attB [44], before insertion into BDSC stock number 9750. To generate tub-Vang-Flagx3 mutant flies, site-directed mutagenesis was performed on vector pCaSpeR-tub-Vang-Flagx3 [47]. All transgenic strains were generated via Bestgene Inc. Wing discs were dissected from third instar larvae and prepared through incubation in final sample buffer at 95°C. For phosphatase treatment, wing discs were collected in PBS and transferred to lysis buffer containing 50mM Tris-HCl pH 7.5, 150mM NaCl, 1mM EDTA, and 1% Triton-X. Lysates were then incubated at 30°C for 30 minutes with lambda protein phosphatase (NEB). Adult wings were collected in PBS containing 0.1% Triton-X-100 (PBST) and incubated for 1 hr at room temperature before mounting in 80% glycerol in PBS. Hair orientation was quantified using the FijiWingsPolarity plugin [54]. Western blot quantification and statistical analyses. Western blot quantifications were made in ImageJ, measuring the intensity of each band and substracting the background. Each measurement was normalized to the loading control(s). To determine statistical differences between data sets of continuous data, we performed nonparametric ANOVA and Tukey’s post test in GraphPad. P-values lower than 0.05 were considered significant, and levels of significance are indicated by number of asterisks (see Figure legends for specific details). The error bars represent the SEM. Drosophila pupal wing fluorescence stainings and confocal analyses. White pupae were collected (0h APF/after puparium formation) and aged at 25°C until dissection at 28-30h APF. Dissections were performed as follows: in brief, pupae were immobilized on double-sided tape, removed from the pupal case, and placed into PBS, in which pupae were partially dissected to remove fat tissue, fixed in 4% paraformaldehyde in PBS for 45 min at RT or overnight at 4°C, and washed 3x in PBS and 0.1% Triton X-100. Wing membranes were removed, and immunostaining was performed by standard techniques. In brief, tissue was incubated in wash buffer containing 10% normal goat serum overnight for primary antibody (4°C), washed 3x with PBS, and incubated with the secondary antibody for at least 2 h (25°C) and fluorescent phalloidin for staining the actin cytoskeleton. Wings were washed 3x with PBS and mounted in Mowiol (see also [74] for more details). Pupal-wing images were acquired at RT using a confocal microscope, Leica SP8. Images were processed with ImageJ (National Institute of Health) and analyzed using the Packing Analyzed software Fiji plug-in [55]. Fluorescence recovery after photobleaching (FRAP) analysis. Approximately 150,000 wild type CD1 keratinocytes were seeded in no.1.5 glass bottom dishes (ibidi #81151) coated with fibronectin. 20–24 hours post-plating, cells were co-transfected with Celsr1-GFP and mCherry-Vangl2 or Celsr1-GFP and mCherry-Vangl2Y308A plasmids. 24 hours after transfection, cells were switched to E-media containing 1.5mM Ca2+ and incubated for an additional 20–24 hours for adequate Celsr1-GFP expression and keratinocytes to form stable cell-cell junctions. Before imaging, cells were switched to phenol-red free E-media with 1.5mM Ca2+. Cells were imaged using a 488nm and 561nm laser, 60X magnification objective (with additional zoom that rendered a pixel size of 110nm) on Nikon A1R-STED confocal microscope equipped with a stage-top Tokai-Hit incubation chamber to maintain 37 degrees and 5% CO2. Keeping magnification, laser power (both for bleach and acquisition), pixel dwell time and acquisition rate constant across all measurements, 1um diameter circular bleach ROIs and one ROI per junction(s) or cells edge(s) were created for bleaching and recovery. The FRAP acquisition sequence consisted of 3 reference pre-bleach images followed by bleach using 405nm laser and finally 70 frames with 4 seconds intervals to monitor fluorescence recovery at junctions. The acquired images in the time series were checked for Z-drift and corrected for presence of any XY drift in Fiji. A reference ROI was made in a non-bleached region to correct for overall bleaching during image acquisition. A background ROI was created outside the fluorescent cell in each image. The ROI values were extracted from drift corrected images in NIS elements software and subsequently processed in Microsoft excel and Graphpad Prism. Each image time series was background and bleach corrected (to be referred as corrected intensity henceforth) and thereafter the corrected intensity profile was normalized as (Ft−Fbleach)/ (Fini−Fbleach), where, Ft is the corrected intensity of the ROI at a given time point, Fbleach is the corrected intensity at the time point immediately after bleaching, Fini is the mean ROI intensity of the three pre-bleach frames. Each mean recovery curve was fitted to exponential one phase association equation in Graphpad Prism and the fitted Plateau and Y0 values were used to determine the immobile fraction = 1- {(Plateau-Y0)/(1- Y0)}. The averaged traces for each condition was fitted to the model with an r-squared value > 0.9. Data represented is pooled from four independent experiments. Cell culture and transfection. Primary wild type CD1 Keratinocytes were cultured using previously published protocol [75]. Keratinocytes were in E-Media prepared in the laboratory (see [75] for composition) supplemented with 50μM Calcium Chloride. For FRAP experiments, phenol-red free DMEM and HF-12 were used to prepare pigment-free imaging E-media and supplemented with 1.5mM Calcium Chloride. Cells were transfected using Effectene reagent following a modified manufacturer’s protocol. 400ng DNA comprising of Celsr1-GFP and mCherry-Vangl2 [14] or mCherry-Vangl2Y308A (this paper) in a ratio of 2:1 was used for co-transfection. Ethics statement. All mouse work in this study was approved by Princeton University’s Institutional Animal Care and Use Committee (IACUC) under protocol number 1867. K14-GFP-Vangl2 transgenic mice (FVB background, [52] were housed in an AAALAC-accredited facility following the Guide for the Care and Use of Laboratory Animals. E15.5 embryos were harvested from K14-Vangl2-GFP heterozygous dams in cold PBS and screened for GFP expression using a stereomicroscope equipped with epifluorescence. Full thickness backskins were dissected from both GFP+ and GFP- littermates embryos and flash frozen immediately in liquid nitrogen (LN2). LN2 was removed by evaporation and frozen backskins were stored for up to 3 months at -80C until cryolysis. Epidermal cryolysis and immunoprecipitation of GFP-Vangl2. Frozen skin samples pooled from four (for IP-Western) or eight (for IP-MS) GFP-Vangl2 and control backskins were processed via cryolysis using a CryoMill (Retch). Briefly, 2ml LN2-frozen lysis buffer droplets (Tissue Extraction Buffer, 1% Triton X-100, 10mM EDTA, 0.3mg/ml PMSF plus protease inhibitors in PBS) were mixed with frozen skin samples and processed by cryogenic grinding for 20 min using a ball mill cooled with LN2. Finely ground frozen epidermal-lysis buffer mixtures were lysed by thawing on ice for 1.5-2hrs. Lysates were cleared by adding 50ul Pansorbin and centrifuging at 14K rpm for 10min. Pre-cleared lysates were transferred to pre-washed, αGFP (rabbit αGFP, AbCam) antibody bound beads and incubated while rotating for 3 hours at 4 degrees C. Immunoprecipitates were washed twice with lysis buffer and eluted with 40ul 2X SDS sample buffer. For western blotting, 40ul total lysate and 40ul IP were run on a 10% SDS-PAGE gel and transferred to a PDVF membrane, blocked for 1hr at room temperature, incubated with primary, chicken αGFP (1:5000, Abcam) overnight at 4°C. After several washes in PBS-T, membrane was incubated for 45 min at room temperature with HRP α-chicken secondaries (1:2500), developed with BioRad Clarity ECL reagent and imaged on both film and with BioRad imager. For IP-MS samples, 20ul total lysate and 40ul IP were run on a 7.5% SDS-PAGE gel. Gel was fixed in 50% MeOH + 7% Acetic Acid for 30–60 minutes, rinsed with H20, stained with SPYRO Ruby overnight, and washed twice for 5 min each in 10% MeOH + 7% Acetic Acid. Gel imaging was performed on a Typhoon FLA-7000 (GE Healthcare). Proteomic sample preparation and mass spectrometry. IP Bands at ~85-90KD were excised then diced, and subjected to in-gel thiol reduction/alkylation and trypsin digestion using a method adapted from Shevchenko et al [72] to process samples for LC-MS/MS. Briefly, gel cubes were destained and washed extensively in 100 mM ammonium bicarbonate buffer, pH 8.8 (ABC), treated with 50 mM TCEP in ABC for 1 h at 55°C, washed, subjected to alkylation with 55 mM iodoacetamide in ABC for 30 min at room temperature in the dark, washed, and finally digested overnight with 1 ug Promega Trypsin Gold (Promega) per gel slice. Peptides from in-gel digest eluates were desalted using STAGE-Tips 9 prior to LC-MS analyses. LC-MS/MS analyses were performed on a high-resolution, high-mass-accuracy, reversed-phase nano-UPLC-MS platform, consisting of an Easy nLC Ultra 1000 nano-UPLC system coupled to an Orbi Elite mass spectrometer (ThermoFisher Scientific) equipped with a Flex Ion source (Proxeon Biosystems, Odense, Denmark). LC was conducted using a trapping capillary column (150 μm x ca. 40 mm, packed with 3 μm, 100 Å Magic AQ C18 resin, Michrom, Auburn, CA) at a flow rate of 5 μL/min for 4 min, followed by an analytical capillary column (75 μm x ca. 45 cm, packed with 3 μm, 100 Å Magic AQ C18 resin, Michrom) under a linear gradient of A and B solutions (solution A: 3% acetonitrile/ 0.1% formic acid; solution B: 97% acetonitrile/ 0.1% formic acid) from 5%-35% B over 90 at a flow rate of 300 nL/min. Nanospray was achieved using Picospray tips (New Objective, Woburn, MA) at a voltage of 2.4 kV, with the Elite heated capillary at 275°C. Full-scan (m/z 335–1800) positive-ion mass spectra were acquired in the Orbitrap at a resolution setting of 120,000. MS/MS spectra were simultaneously acquired using CID in the LTQ for the fifteen most abundant multiply charged species in the full-scan spectra, having signal intensities of >1000 NL. To aid in phosphosite mapping, Vangl2-positive slices were subjected LC-MS/MS over a 180 min gradient using the CID parameters above and also a second round of LC-MS/MS over a 180 min gradient during which MS/MS spectra were acquired by multistage activation (MSA) for the top 10 most abundant ions in the full-scan spectra, using excitation at the precursor m/z value as well as those corresponding to the neutral losses of phosphonic and phosphoric acids for ions of charge +2 and +3. Lockmass was employed, maintaining calibration to 2–3 ppm of accurate mass. Mass spectrometric data analysis. Resultant LC-MS/MS raw data files were processed using ProteomeDiscoverer (v. 1.4, ThermoFisher), to match MS/MS spectra against the UniProt Mus musculus database, or a GFP-Vangl2 fusion protein construct subset database using the Mascot search engine (v. 2.4, Matrix Science, London, UK.), allowing for a parent ion mass window of ±6 ppm, ≤ 3 missed trypsin cleavages, serine, threonine and tyrosine phosphorylation, methionine oxidation, asparagine and glutamine deamination and N-terminal protein acetylation as variable modifications, and carbamidomethylation of cysteines as a fixed modification. Peptide assignment cut-offs were specified at a high confidence level (<1% FDR). Phosphosite localization confidence scoring was achieved using the PhosphoRS 9 [73] (v. 3.1) node within the ProteomeDiscoverer framework. Relative abundance levels for proteins between experimental and controls were estimated using spectral counting. Raw mass spectra were visualized using Xcalibur (v. 2.2, ThermoFisher) and peptide spectral matches were visualized using ProteomeDiscoverer or Scaffold (v. 4.3.4, Proteome Software, Portland, OR). All phosphopeptide assignments were further validated by manual inspection. DNA constructs and S2 culture. Constructs used were as follows; pAc5.1-Flagx3, pAc5.1-Myc-Pk, pAc5.1-Dsh-GFP, pAc5.1-Scrib-PDZ-3-4-HA and pAc5.1-HA-Dgo (all gifts from Dr. Jenny, AECOM, USA) and pAc5.1-Vang-Flagx3 [44]. GFP-TZ was used as a control, and is the Drosophila ciliary protein Mks1. pAc5.1-Vang-Flagx3-Y374A, FKYY371AAYA, Y374F and V376A were generated though site-directed mutagenesis. C-terminal truncations of Vang were generated through PCR and cloned into pAc5.1-Flagx3 using NotI-XbaI restriction sites. All primers used are available upon request. Unless otherwise stated, lysates were prepared from S2 cells. S2 cells were maintained according to standard protocols, and were grown in Schneider’s Medium (Gibco) supplemented with 10% heat-inactivated Fetal Bovine Serum (Gibco). Cells were plated in 12 well plates at a dilution of 1.5x106 and were transfected with the indicated constructs using Effectene (Qiagen) according to manufacturer’s protocols. Cells were lysed ~48 hrs later in buffer containing 50mM Tris-HCl pH 7.5, 150mM NaCl, 1mM EDTA and 1% Triton-X-100. Pull-downs, immunoblotting and peptide interactions. For Flag pull-down experiments, lysates were extracted from S2 cells or from larval wing discs and were incubated at 4°C overnight with 10μl anti-Flag M2 affinity gel per sample. Washes were performed in buffer containing 50mM Tris-HCl pH 7.5, 350mM NaCl, 1mM EDTA, 0.1% SDS before elution in final sample buffer. For peptide binding, Vang peptides were conjugated to beads to enable assessment of direct binding. Lysates were generated in S2 cells, and the lysate was divided equally between tubes containing either 20μl of conjugated phospho-peptide or the unphosphorylated form and incubated at 4°C overnight. Washes were performed in buffer containing 10mM Tris-HCl pH7.5, 350mM NaCl, and 0.5mM EDTA before elution in final sample buffer. For GFP pull-downs. Samples were resolved by polyacrylamide gel electrophoresis according to standard protocols. The following primary antibodies were used for immunoblotting; Flag (Sigma M2 1:5000), Gamma-tubulin (Sigma GTU-88 1:1000), GFP (Roche 7.1&13.1 1:1000), Myc (SCBT 9E10 1:1000), Phospho-tyrosine (Millipore PY20 1:1000). Drosophila strains, dissections and phenotypic analyses. Flies were raised on standard medium, and maintained at 25°C unless otherwise stated. To generate UAS-Vang-Flagx3 mutant transgenic flies, site-directed mutagenesis was performed on vector pUAST-Vang-Flagx3-attB [44], before insertion into BDSC stock number 9750. To generate tub-Vang-Flagx3 mutant flies, site-directed mutagenesis was performed on vector pCaSpeR-tub-Vang-Flagx3 [47]. All transgenic strains were generated via Bestgene Inc. Wing discs were dissected from third instar larvae and prepared through incubation in final sample buffer at 95°C. For phosphatase treatment, wing discs were collected in PBS and transferred to lysis buffer containing 50mM Tris-HCl pH 7.5, 150mM NaCl, 1mM EDTA, and 1% Triton-X. Lysates were then incubated at 30°C for 30 minutes with lambda protein phosphatase (NEB). Adult wings were collected in PBS containing 0.1% Triton-X-100 (PBST) and incubated for 1 hr at room temperature before mounting in 80% glycerol in PBS. Hair orientation was quantified using the FijiWingsPolarity plugin [54]. Western blot quantification and statistical analyses. Western blot quantifications were made in ImageJ, measuring the intensity of each band and substracting the background. Each measurement was normalized to the loading control(s). To determine statistical differences between data sets of continuous data, we performed nonparametric ANOVA and Tukey’s post test in GraphPad. P-values lower than 0.05 were considered significant, and levels of significance are indicated by number of asterisks (see Figure legends for specific details). The error bars represent the SEM. Drosophila pupal wing fluorescence stainings and confocal analyses. White pupae were collected (0h APF/after puparium formation) and aged at 25°C until dissection at 28-30h APF. Dissections were performed as follows: in brief, pupae were immobilized on double-sided tape, removed from the pupal case, and placed into PBS, in which pupae were partially dissected to remove fat tissue, fixed in 4% paraformaldehyde in PBS for 45 min at RT or overnight at 4°C, and washed 3x in PBS and 0.1% Triton X-100. Wing membranes were removed, and immunostaining was performed by standard techniques. In brief, tissue was incubated in wash buffer containing 10% normal goat serum overnight for primary antibody (4°C), washed 3x with PBS, and incubated with the secondary antibody for at least 2 h (25°C) and fluorescent phalloidin for staining the actin cytoskeleton. Wings were washed 3x with PBS and mounted in Mowiol (see also [74] for more details). Pupal-wing images were acquired at RT using a confocal microscope, Leica SP8. Images were processed with ImageJ (National Institute of Health) and analyzed using the Packing Analyzed software Fiji plug-in [55]. Fluorescence recovery after photobleaching (FRAP) analysis. Approximately 150,000 wild type CD1 keratinocytes were seeded in no.1.5 glass bottom dishes (ibidi #81151) coated with fibronectin. 20–24 hours post-plating, cells were co-transfected with Celsr1-GFP and mCherry-Vangl2 or Celsr1-GFP and mCherry-Vangl2Y308A plasmids. 24 hours after transfection, cells were switched to E-media containing 1.5mM Ca2+ and incubated for an additional 20–24 hours for adequate Celsr1-GFP expression and keratinocytes to form stable cell-cell junctions. Before imaging, cells were switched to phenol-red free E-media with 1.5mM Ca2+. Cells were imaged using a 488nm and 561nm laser, 60X magnification objective (with additional zoom that rendered a pixel size of 110nm) on Nikon A1R-STED confocal microscope equipped with a stage-top Tokai-Hit incubation chamber to maintain 37 degrees and 5% CO2. Keeping magnification, laser power (both for bleach and acquisition), pixel dwell time and acquisition rate constant across all measurements, 1um diameter circular bleach ROIs and one ROI per junction(s) or cells edge(s) were created for bleaching and recovery. The FRAP acquisition sequence consisted of 3 reference pre-bleach images followed by bleach using 405nm laser and finally 70 frames with 4 seconds intervals to monitor fluorescence recovery at junctions. The acquired images in the time series were checked for Z-drift and corrected for presence of any XY drift in Fiji. A reference ROI was made in a non-bleached region to correct for overall bleaching during image acquisition. A background ROI was created outside the fluorescent cell in each image. The ROI values were extracted from drift corrected images in NIS elements software and subsequently processed in Microsoft excel and Graphpad Prism. Each image time series was background and bleach corrected (to be referred as corrected intensity henceforth) and thereafter the corrected intensity profile was normalized as (Ft−Fbleach)/ (Fini−Fbleach), where, Ft is the corrected intensity of the ROI at a given time point, Fbleach is the corrected intensity at the time point immediately after bleaching, Fini is the mean ROI intensity of the three pre-bleach frames. Each mean recovery curve was fitted to exponential one phase association equation in Graphpad Prism and the fitted Plateau and Y0 values were used to determine the immobile fraction = 1- {(Plateau-Y0)/(1- Y0)}. The averaged traces for each condition was fitted to the model with an r-squared value > 0.9. Data represented is pooled from four independent experiments. Cell culture and transfection. Primary wild type CD1 Keratinocytes were cultured using previously published protocol [75]. Keratinocytes were in E-Media prepared in the laboratory (see [75] for composition) supplemented with 50μM Calcium Chloride. For FRAP experiments, phenol-red free DMEM and HF-12 were used to prepare pigment-free imaging E-media and supplemented with 1.5mM Calcium Chloride. Cells were transfected using Effectene reagent following a modified manufacturer’s protocol. 400ng DNA comprising of Celsr1-GFP and mCherry-Vangl2 [14] or mCherry-Vangl2Y308A (this paper) in a ratio of 2:1 was used for co-transfection. Supporting information S1 Fig. IP-MS approach from mouse skin to identify Vangl2 PTMs. https://doi.org/10.1371/journal.pgen.1010849.s001 (DOCX) S2 Fig. Specificity of Pk and Dsh binding to the Vang region 364–387. https://doi.org/10.1371/journal.pgen.1010849.s002 (DOCX) S3 Fig. Charged amino acids interfere with Dsh binding to Vang. https://doi.org/10.1371/journal.pgen.1010849.s003 (DOCX) S4 Fig. Y374 peptide and phospho-peptide binding to Pk and Dsh. https://doi.org/10.1371/journal.pgen.1010849.s004 (DOCX) S5 Fig. Distinct gain-of-function in vivo behavior of Vang and the respective Pk and Dsh binding mutants. https://doi.org/10.1371/journal.pgen.1010849.s005 (DOCX) S6 Fig. Vang point mutants affect localization of core PC factors in vivo. https://doi.org/10.1371/journal.pgen.1010849.s006 (DOCX) S7 Fig. FRAP analysis of junctional Celsr1 and Vangl2 in cultured keratinocytes. https://doi.org/10.1371/journal.pgen.1010849.s007 (DOCX) S1 Table. Vangl2 peptides and PTMs. https://doi.org/10.1371/journal.pgen.1010849.s008 (XLSX) Acknowledgments We are grateful to Andreas Jenny and Ursula Weber for plasmids, and we thank Giovanna Collu, and all Devenport and Mlodzik lab members, past and present, for helpful input and discussions. We thank Saw Kyin and David Perlman of the Molecular Biology Proteomics & Mass Spectrometry Core facility at Princeton University who provided technical support and analyzed the mass spectrometry data.

Belew, Mezmur D.;Chien, Emilie;Michael, W. Matthew

doi: 10.1371/journal.pgen.1010831pmid: 37478128

Introduction Cells can control gene expression at many levels, including individual loci, groups of genes that perform a common function, or at the level of the entire genome. Whole-genome control of transcription occurs in many different contexts. For example, cells undergoing quiescence silence their genomes [1,2], as do all proliferating cells as they prepare for mitosis [3–5]. In addition, during germline development, many animals silence the genome in germline progenitor cells until they have been specified as germline [6,7]. Furthermore, a common feature of gametogenesis is a genome silencing event as gametes complete meiotic prophase [8,9]. The genome can also be globally activated, and examples of this are also found in early development, where many organisms undergo a zygotic genome activation event as embryos transition from maternal to zygotic control of embryogenesis [10]. While the concept of whole-genome activation and silencing has been appreciated for a long time, it is only recently that the molecular mechanisms in play have begun to be elucidated. One particularly useful model system for the study of whole-genome control of transcription has been the roundworm C. elegans. Remarkably, during germline development in the worm, the genome is silenced and then reactivated on at least four distinct occasions. During early embryogenesis, in one- and two-cell embryos, transcription is globally repressed via binding of the OMA-1 and OMA-2 proteins to the basal transcription factor TAF-4, as this sequesters TAF-4 in the cytoplasm [11]. At the four-cell stage somatic genomes are activated. However, in germline precursor cells, the so-called P-lineage, transcription remains globally silenced [12,13]. Silencing in the P-lineage is controlled by PIE-1, a zinc-finger containing protein whose mechanism of action during silencing is not yet fully understood. Upon germline specification, PIE-1 is degraded and germline genome activation occurs in the Z2 and Z3 primordial germ cells (PGCs) [14]. PGCs remain transcriptionally active through hatching of the embryo into an L1 larva. Recent work from our group has shown that if embryos hatch in an environment lacking nutrients, then the PGC genome is silenced once again [15]. Starvation-induced genome silencing is triggered by the energy sensing kinase AMPK and requires the TOP-2/condensin II chromatin compaction pathway as well as components of the H3K9me/heterochromatin pathway [15]. In this system, TOP-2/condensin II promotes H3K9me2 and -me3 deposition on germline chromatin, and thus we named this pathway the Global Chromatin Compaction (GCC) to reflect the linear organization of the system components. We have also previously shown that when starved L1s are fed, then the germline genome is reactivated. This requires the induction of DNA double-strand breaks that serve to promote chromatin decompaction [16,17]. Germline chromatin remains transcriptionally active through the remainder of development and, in hermaphrodites, through oogenesis until the genome is silenced once again at the end of meiotic prophase [18,19]. Thus, in the nematode germline, there are four genome silencing events—in early embryos by OMA-1/2, in the P-lineage by PIE-1, in starved L1s by the GCC pathway, and in oocytes via an unknown mechanism. The global repression of transcription during late oogenesis is not unique to C. elegans. In Drosophila melanogaster, oocytes are developmentally arrested at prophase I and repress transcription between the fifth and eighth stages of oogenesis, just before their re-entry into meiosis [20]. Similarly, in mice, primary follicles that contain prophase I arrested oocytes have been shown to repress transcription just before the resumption of meiosis [21,8,9]. This repression persists throughout fertilization until minor ZGA occurs at first cleavage [22,23]. The conservation of this theme across organisms of different complexity suggests that global transcriptional repression in oocytes is an important feature of the oocyte-to-embryo transition. Despite this conservation, however, the molecular pathway(s) responsible for repression are still unknown. In this study, we address the problem of how oocytes in C. elegans shut down transcription as a function of completing meiotic prophase. The nematode is an ideal system to address this important question, as the events leading up to oocyte maturation and fertilization are well described. During germline development in hermaphrodites, animals first produce sperm and then switch to making oocytes [24]. Oocytes are produced in an assembly-line like process, where cells exit pachytene and then progress through the remainder of meiotic prophase within the proximal portion of the tube-shaped gonad [25]. Within the proximal gonad, oocytes can be clearly identified by their position relative to the spermatheca—the oocyte closest is named -1, and the next most proximal oocyte -2, et cetera. Previous work has shown that -4 and -3 oocytes are transcriptionally active, and then the genome is silenced at the -2 position (see Fig 1A) [19]. Other work has shown that, just prior to genome silencing, the condensin II complex is recruited to chromatin, where it helps to compact chromatin during the formation of bivalents, a unique chromosome structure that enables the subsequent meiotic divisions (see Fig 1A) [26]. By the -2 position, bivalents have formed and the genome is silenced. At the -1 position the oocyte receives a signal from sperm to initiate maturation, and the cells then enter meiotic M-phase (Fig 1A) [27]. Here, we show that genome silencing in oocytes is organized by cyclin-dependent kinase 1 (CDK-1 in C. elegans) and requires the known silencers TOP-2/condensin II, the H3K9me/heterochromatin pathway, and PIE-1. Loss of any one of these components results in aberrant RNA polymerase II (RNAPII) activity at the -2 position. Interestingly, we also report that in oocytes distal to the -2 position, PIE-1 is mainly localized in the nucleolus, and that at the -2 position the nucleolus dissolves in a TOP-2/condensin II dependent manner. Our data identify the molecular components for the oocyte genome silencing system and suggest a model where the nucleolar residency of PIE-1 prevents it from blocking RNAPII activity until the nucleolus dissolves at the -2 position. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 1. Timing and CDK-9 dependence of the phosphorylation of RNAPIISer2 in C. elegans proximal oocytes. A. Schematic summarizing transcriptional activity and chromatin compaction in the four most proximal oocytes. See Introduction for details. B. N2s were treated with either control or cdk-9 RNAi. Dissected gonads from these animals were fixed and stained for DNA (blue) and RNAPIIpSer2 (red). Depletion of CDK-9 results in the loss of RNAPIIpSer2 signal in proximal oocytes. Scale bar represents a length of 2 μm. C. Pachytene nuclei from the same animals in (B) were fixed and stained for DNA (blue) and RNAPIIpSer2 (red). Unlike proximal oocytes cdk-9 RNAi does not alter RNAPIIpSer2 signal in pachytene nuclei. Scale bar represents a length of 2 μm. https://doi.org/10.1371/journal.pgen.1010831.g001 Results RNAPIIpSer2 is dependent on CDK-9 in proximal oocytes The goal of this study was to analyze genome silencing during oogenesis in C. elegans. To monitor transcription, we used an antibody that recognizes the active and elongating form of RNAPII. This reagent is a rabbit polyclonal antibody termed ab5095, purchased from Abcam (Waltham, MA), and it detects a phospho-epitope on the second serine within the carboxy-terminal repeat domain (CTD) of RNAPII (RNAPIIpSer2) [28]. In previous work we had validated ab5095 for use in C. elegans via the demonstration that reactivity depends on the presence of both RNAPII and phosphate, and that ab5095 accurately labels transcriptionally active nuclei in worm embryos and in the PGCs of L1 larvae [15]. For the current study, we used it to examine the transcriptional status of oocytes obtained from gonads dissected from adult hermaphrodites. Prior work that examined proximal oocytes has shown that RNAPIIpSer2 signals decreased starting with the fourth most proximal oocyte (termed -4) and became undetectable through the two most proximal oocytes (-2 and -1) [19]. We, therefore, focused our analysis on the four most proximal oocytes and found that RNAPIIpSer2 signal was detected invariably on the chromatin of oocytes at the -4 position (Fig 1B, control RNAi panel). However, as we examined the three most proximal oocytes, we observed several different patterns for the RNAPIIpSer2 signal: (1) signal exclusively on chromatin, (2) no signal at all, (3) signal present both on and off chromatin, and (4) a nucleoplasmic RNAPIIpSer2 signal that mostly excludes chromatin (Figs 1B and S1). It thus appears that while in some oocytes RNAPIIpSer2 signal completely disappeared, in others, it was merely removed from the chromatin and was present in the nucleoplasm. This form of nucleoplasmic signal has also been observed in human cells as they approach mitosis and is explained by RNAPII coming off the chromatin while maintaining CTD phosphorylations [29]. In mammals, the kinase that phosphorylates serine 2 within the RNAPII CTD is P-TEFb, which is composed of the CDK9 kinase and cyclins T1 or T2 [30]. To determine if the corresponding kinase in C. elegans, CDK-9, plays a similar role in proximal oocytes, we used RNAi to deplete the protein and we then stained for RNAPIIpSer2. As shown in Figs 1B and S1, exposure to cdk-9 RNAi caused a reduction in all three forms of RNAPIIpSer2 signal (on chromatin, off chromatin, and both on and off chromatin). This shows that the off-chromatin signals are due to CDK-9 phosphorylated RNAPII that had been displaced from chromatin but not yet dephosphorylated. For the remainder of this study, we only considered nuclei with RNAPIIpSer2 on chromatin as being transcriptionally active, and the off-chromatin signals were ignored. Interestingly, we also noticed that in more distal regions of the gonad, for example the pachytene region, RNAPIIpSer2 signal was unchanged by cdk-9 RNAi (Fig 1C). This is consistent with previous work showing that CDK-12 is the major RNAPIIpSer2 kinase in the mitotic and pachytene regions of the gonad [31]. It thus appears that there is a handoff, from CDK-12 to CDK-9, for generating RNAPIIpSer2 during oogenesis in C. elegans. The TOP-2/condensin II axis controls genome silencing during meiotic prophase Our data, together with previous findings [19], show that genome silencing initiates at the -3 position and is largely complete by the -2 position. As detailed above, there are four genome silencing systems that have been identified thus far in C. elegans: the TOP-2/condensin II axis (in starved L1 PGCs) [15], the H3K9me pathway (in starved L1 PGCs) [15]; the OMA-1/2 proteins (in one- and two-cell embryos) [11], and the PIE-1 protein (in the P2, P3, and P4 embryonic blastomeres) [12,13]. Which, if any, of these systems might also be used for genome silencing in oocytes? We know that OMA-1/2 are dispensable for oocyte silencing, as previous work had directly examined this possibility [11]. Thus, we turned our attention to the remaining three systems and focused initially on TOP-2/condensin II. To do so we examined RNAPIIpSer2 in oocytes from animals exposed to either top-2, capg-2 (which encodes a protein specific for the condensin II complex), or double top-2/capg-2 RNAi. All three of these treatments impacted the RNAPIIpSer2 pattern in a similar manner: signal now persisted in -2 oocytes, indicating that genome silencing had failed (Fig 2A and 2B). To quantify these data, we measured RNAPIIpSer2 signal intensity present on chromatin within individual oocytes. We normalized these values against signal intensities present in the pachytene region, an irrelevant tissue, in order to account for potential differences in antibody penetration across the different sample sets (see Methods for a more detailed explanation). As shown in Fig 2B, in the control samples, there was a significant difference in RNAPIIpSer2 signal intensity between the -3 and -2 positions, reflecting the genome silencing that occurs in -2 oocytes under normal conditions. Also shown in Fig 2B is the difference in RNAPIIpSer2 signal intensity between -2 oocytes from control samples, relative to the samples where TOP-2 and/or CAPG-2 had been depleted. Loss of TOP-2/CAPG-2 caused a significant increase in RNAPIIpSer2 signal at the -2 position. Based on these data, we conclude that loss of TOP-2 and/or CAPG-2 impacts genome silencing in -2 oocytes. Since genome silencing normally occurs in late prophase, one possibility is that loss of TOP-2/CAPG-2 somehow delays progression through diakinesis, and thus the persistence of RNAPII transcription in -2 oocytes of top-2/capg-2 depleted samples is an indirect consequence of defects in prophase timing. This is not the case, however, as when we stained for the prophase marker phospho-serine 10 of histone H3 (H3pS10) [32] we saw no difference in the timing of H3pS10 deposition (S2 Fig). We note that H3pS10 signal intensity was seemingly increased after top-2/capg-2 RNAi, however in the absence of an independent source of H3pS10 signal that can be used to normalize signal intensity across different sample sets it is impossible to know if this is a physiological effect or an artifact of differential antibody penetration. Nonetheless, it is clear that depletion of TOP-2/CAPG-2 does not delay prophase timing. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 2. Aberrant RNAPIIpSer2 signal is observed in proximal oocytes when TOP-2 and condensin II are depleted. A. N2 animals were treated with control, top-2, capg-2, or top-2/capg-2 double RNAi. Gonads were dissected from these animals and were fixed and stained for DNA (blue) and RNAPIIpSer2 (red). Each panel series (-3, -2, -1, and pachytene) represent nuclei from the same gonad. Individual depletion of TOP-2, CAPG-2, and their co-depletion results in persistent RNAPIIpSer2 signal on the chromatin of -3 and -2 oocytes. Scale bar represents a length of 2 μm. B. Quantification of data presented in (A). RNAPIIpSer2 signal was measured using ImageJ. The average ratio of normalized raw integrated density is plotted on the y-axis. Between 10 and 22 gonads were analyzed per condition and data were collected across multiple independently performed replicates. For control RNAi 22 gonads across 5 replicates were analyzed, for top-2 RNAi 15 gonads were analyzed across 2 replicates, for capg-2 RNAi it was 20 gonads across 2 replicates and for top-2/capg-2 RNAi it was 10 gonads across a single replicate. Significance was measured using student’s t-test or Wilcoxon Rank Sum test. **p<0.01; *p<0.05. In control animals, the ratio of signal density decreases significantly between -3 and -2 oocyte positions. Loss of TOP-2 or CAPG-2 results in persistent RNAPIIpSer2 signal in -2 oocytes. Co-depletion of TOP-2 and CAPG-2 produces a stronger phenotype, with persistent RNAPIIpSer2 signal in -3 and -2 oocytes. https://doi.org/10.1371/journal.pgen.1010831.g002 CDK-1 acts upstream of TOP-2/condensin II to trigger genome silencing Our data identify a role for TOP-2 and condensin II in the silencing of transcription as oocytes prepare for maturation and the meiotic divisions. Silencing initiates at the -3 position and is complete by -2, and this is also when condensin II is recruited to oocyte chromatin [26]. Previous work has also shown that condensin II activity is dependent on phosphorylation by CDK1 [33], and thus it stands to reason that CDK-1 may also play a role in genome silencing, through regulation of condensin II. To test this hypothesis, we used RNAi to reduce CDK-1 activity in proximal oocytes and we assessed the impact on transcriptional activity. As shown in Fig 3A and 3B, cdk-1 RNAi triggered a significant increase in RNAPIIpSer2 signal intensity in -3, -2, and -1 oocytes, showing that CDK-1 does indeed control genome silencing. To pursue this further, we next asked the opposite question—how does elevating CDK-1 activity impact genome silencing? For this we targeted the CDK-1 inhibitory kinase WEE-1.3 for depletion, as previous work has shown that WEE-1.3 negatively regulates CDK-1 in nematode oocytes and, further, that CDK-1 is the sole target of WEE-1.3 [34]. Treatment with wee-1.3 RNAi resulted in a decrease in RNAPIIpSer2 signal intensity across all oocytes spanning the -6 to -2 positions, showing that genome silencing was happening much earlier than in the control condition (Fig 4A and 4B). These data posed an important question: is the precocious silencing observed after wee-1.3 RNAi also dependent on TOP-2/condensin II? To address this, we co-depleted WEE-1.3 and CAPG-2, and we found that transcriptional activity was now restored to proximal oocytes (Fig 4A and 4B). Thus, co-depletion with CAPG-2 reverses the effects of WEE-1.3 depletion alone, and this shows that condensin II is required for the precocious silencing observed when CDK-1 levels are increased via depletion of WEE-1.3. Based on these data, we conclude that CDK-1 acts upstream of TOP-2/condensin II, in a positive manner, to promote genome silencing. Furthermore, our data suggest that the CDK-1 can activate condensin II for genome compaction and silencing at an activity state below that needed to trigger nuclear envelope breakdown (NEB) and entry into meiotic M-phase. This, in turn, suggests that CDK-1 activity rises gradually in the proximal gonad, as opposed to an abrupt activation in -1 oocytes. This would be consistent with previous work in HeLa cells as well as frog egg extracts showing that CDK-1-cyclin B activity rises gradually during interphase and prophase [35,36]. To gain additional evidence for a gradual acquisition of the M-phase state in proximal oocytes we stained them with MPM-2, an antibody that recognizes mitotic phosphoproteins [37]. As shown in S3A Fig, MPM-2 antigens are present at low levels in -5 oocytes and they gradually accumulate as we move proximally in the gonad, such that -1 oocytes show a high level of MPM-2 reactivity and -3 and -2 oocytes a more intermediate level. These data are consistent with the idea that CDK-1 activity gradually increases, and that levels of CDK-1 that are proficient to trigger genome compaction and silencing in -3/-2 oocytes are not yet sufficient for NEB and entry into meiotic M-phase. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 3. Aberrant RNAPIIpSer2 signal is observed in proximal oocytes when CDK-1 is depleted. A. N2 animals were treated with control or cdk-1 RNAi. Gonads were dissected from these animals and were fixed and stained for DNA (blue) and RNAPIIpSer2 (red). Loss of CDK-1 results in RNAPIIpSer2 signal persisting into the -2 and -1 oocyte positions. B. Quantification of data presented in (A). RNAPIIpSer2 signal was measured using ImageJ. The average ratio of normalized raw integrated density is plotted on the y-axis. 19 samples were analyzed for each RNAi treatment over 2 independent replicates. Significance was measured using student’s t-test or Wilcoxon Rank Sum test. **p<0.01; *p<0.05. Depletion of CDK-1 resulted in a significant increase in on-chromatin RNAPIIpSer2 signal in the three most proximal oocytes. https://doi.org/10.1371/journal.pgen.1010831.g003 Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 4. Hyperactivation of CDK by the depletion of WEE-1.3 causes unscheduled transcriptional silencing. A. N2 animals were treated with either control, wee-1.3, or wee-1.3/capg-2 double RNAi. Gonads were dissected from these animals and were fixed and stained for DNA (blue) and RNAPIIpSer2 (red). Depletion of WEE-1.3 results in the loss of on-chromatin RNAPIIpSer2 signals in the more distal (-6 to -4 position) oocytes. Co-depletion using wee-1.3/capg-2 RNAi reverses the loss of RNAPIIpSer2 signal. Scale bars represent a length of 2 μm. B. Quantification of data presented in (A). RNAPIIpSer2 signal was measured using ImageJ. The average ratio of normalized raw integrated density is plotted on the y-axis. 20 samples were analyzed for each RNAi treatment over 2 independently performed replicates. Significance was measured using student’s t-test or Wilcoxon Rank Sum test. **p<0.01; *p<0.05. Depletion of WEE-1.3 resulted in a decrease in RNAPIIpSer2 signal from -6 to -2 oocyte positions. Co-depletion of WEE-1.3 and CAPG-2 resulted in restored RNAPIIpSer2 signal in -6 to -3 oocyte positions. https://doi.org/10.1371/journal.pgen.1010831.g004 The presence of RNAPIIpSer2 signal at -1 oocytes after cdk-1 RNAi but not after top-2/capg-2 RNAi also raises the idea that there is an additional transcriptional repression mechanism in play when CDK-1 activity levels are high enough to trigger meiotic M-phase entry and NEB. To test this idea, we performed phase-contrast microscopy to assess nuclear envelope integrity of -1 oocytes in cdk-1 RNAi treated samples. We saw that more cdk-1 RNAi samples had intact nuclear envelopes than control RNAi treated samples (S3B Fig) establishing a correlation between silencing at -1 oocytes and NEB. When we performed a similar analysis on wee-1.3 RNAi oocytes (that precociously silence their genome in a top-2/condensin II dependent manner) we did not see a change in the number of samples with intact nuclear membrane relative to control RNAi animals (S3C Fig). These data show that the silencing at -1 is tied to NEB progression while the TOP-2/condensin II dependent silencing is not and suggest the existence of an additional silencing mechanism at -1 oocytes that is distinct from the silencing mechanism discussed so far in this manuscript. H3K9me3 marks accumulate on oocyte chromatin during genome silencing We next examined a role for the H3K9me pathway in oocyte genome silencing. Our previous work showed that the Z2/Z3 PGCs accumulate H3K9me marks to a significant extent, relative to neighboring somatic cells, as they prepare for silencing [15]. To see if this also occurs on oocyte chromatin we used an antibody, termed ab176916 and purchased from Abcam (Waltham, MA), that recognizes the tri-methylated form of lysine 9 on histone H3. We have previously validated this antibody with the demonstration that reactivity is lost in Z2/Z3 of L1 larvae in a strain lacking SET-25, the major H3K9me3 methyltransferase in C. elegans [15]. We limited the current analysis to H3K9me3 because available antibodies against H3K9me2 do not recognize their target when the neighboring serine 10 is phosphorylated [38], and we have seen in S2 Fig that H3pS10 is prominent in proximal oocytes. As shown in Fig 5A, H3K9me3 first appears in -5 oocytes, and then gradually accumulates such that by the -2 and -1 positions the H3K9me3 and DNA signals overlap extensively. To examine the reproducibility of this pattern we quantified H3K9me3 signal intensity across 11 gonads and then compared the values obtained for the -5 to -2 positions to the value obtained for -1 position within a given gonad. As shown in Fig 5B, the pattern of gradual accumulation of H3K9me3 marks as oocytes moved from distal to proximal was indeed highly reproducible. In C. elegans, H3K9me3 is produced primarily by the SET-25 methyltransferase [39]. Indeed, when H3K9me3 was examined in set-25 mutants and compared to wild type samples, we observed that the H3K9me3 signals were attenuated (Fig 5A). Fig 5C shows additional examples of -1 and -2 oocytes for wild type and set-25 mutants at higher magnification, highlighting the differences in H3K9me3 signal intensity. No such attenuation was observed when met-2 mutants were compared to wild type (Fig 5A), consistent with the MET-2 methyltransferase primarily responsible for producing H3K9me1 and me2 [39]. Our findings are also consistent with those of Bessler et al, 2010, who observed that H3K9me3 levels in the pachytene region of the gonad do not change in met-2 mutants, relative to wild type [40]. We conclude that H3K9me3 marks accumulate dramatically on chromatin at the time that oocytes are silencing their genomes. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 5. H3K9me3 signals significantly increase as oocytes become more proximal. A. Wild-type, met-2 mutant, and set-25 mutant gonads were dissected, fixed, and stained for DNA (blue) and H3K9me3 (red). In N2s, an increase in H3K9me3 signal is observed at the -3 to -1 positions in comparison to more distal oocytes and pachytene nuclei from the same animal. Loss of MET-2 does not affect the H3K9me3 accumulation pattern. Depletion of SET-25 results in loss of H3K9me3 signal in proximal oocytes. Scale bar represents a length of 2 μm. B. Quantification of data for wild type samples in (A). H3K9me3 signals for each oocyte in -5 to -2 positions relative to the most proximal oocyte are presented. 20 samples were analyzed over 2 independent replicates. H3K9me3 signal accumulates on chromatin in the more proximal oocyte positions. C. Additional examples of -2 and -1 oocytes stained as in part (A). Either wild type (N2) or set-25 mutant oocytes are shown. Note that the set-25 samples have only trace amounts of H3K9me3 on the chromatin, relative to the wild-type sample. https://doi.org/10.1371/journal.pgen.1010831.g005 The accumulation of H3K9me3 in the oocytes was reminiscent of the H3K9me spreading we saw during L1 starvation which is dependent on TOP-2 and condensin II [15]. Given this, we wished to know if TOP-2 and condensin II also facilitated the spreading of H3K9me3 in maturing oocytes by staining gonads from top-2/capg-2 RNAi treated samples for H3K9me3. We saw that there was no difference in H3K9me3 signal between control and top-2/capg-2 RNAi treated samples (S4A and S4B Fig). Therefore, unlike in starved L1s, the accumulation of H3K9me3 in maturing oocytes does not require TOP-2 and condensin II. Both the SET-25 and MET-2 methyltransferases are required for genome silencing in oocytes Having observed SET-25 dependent accumulation of H3K9me3 marks in proximal oocytes we next asked if SET-25 plays a role in genome silencing. Staining of set-25 mutants for RNAPIIpSer2 revealed that -2 oocytes contained significantly more RNAPIIpSer2 signal than did the control samples (Fig 6A and 6B), showing that silencing was attenuated. Interestingly, we also observed a silencing defect in met-2 mutants (Fig 6A and 6B). Thus, not only do H3K9me marks accumulate during silencing, but the enzymes responsible for catalyzing these modifications are also important for genome silencing. We conclude that the H3K9me pathway globally silences transcription in developing oocytes, as it does in the PGCs of starved L1s. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 6. The H3K9 methyltransferases MET-2 and SET-25 are required for transcriptional repression in proximal oocytes. A. Gonads from wild-type, met-2 mutant, and set-25 mutant animals were dissected, fixed, and stained for DNA (blue) and RNAPIIpSer2 (red). Mutations of either met-2 or set-25 results in persistent RNAPIIpSer2 signal on chromatin, compared to wild-type animals. Scale bar represents a length of 2 μm. B. Quantification of data presented in (A). RNAPIIpSer2 signal was measured using ImageJ. The average ratio of normalized raw integrated density is plotted on the y-axis. 10 to 20 samples were analyzed for each condition over 2 independent replicates. 15 wild-type samples were analyzed while 10 met-2 mutant and 20 set-25 mutants were analyzed. Significance was measured using student’s t-test or Wilcoxon Rank Sum test. **p<0.01; *p<0.05. Loss of MET-2 resulted in a significant persistence of RNAPIIpSer2 signal on chromatin in -2 and -1 oocytes. Loss of SET-25 led to a significant persistence of RNAPIIpSer2 signal on chromatin in -2 oocytes. https://doi.org/10.1371/journal.pgen.1010831.g006 Genome silencing is coupled to chromatin compaction in oocytes Previous work has shown that condensin II loads on to oocyte chromosomes at the -3 position and is required for the intense chromatin compaction that occurs as bivalents are formed [26]. In starved L1s, genome silencing is coupled to chromatin compaction [15], and thus it was important to monitor compaction in the oocyte system. For this we turned to a previously utilized strain that carries a transgene encoding mCherry-tagged histone H2B, which marks chromatin [15,17]. Living hermaphrodites were immobilized and oocyte chromatin was imaged using confocal microscopy. We compared control samples to those that been exposed to top-2/capg-2 RNAi, and we looked at -2 oocytes, which is where genome silencing is occurring. As described in the Methods and S5 Fig, we measured the volume of the chromatin masses and found that depletion of TOP-2/CAPG-2 caused a significant increase in volume, consistent with a defect in compaction, and similar observations were made after set-25 or met-2 RNAi (Figs 7 and S5). Thus, as is the case in the PGCs of starved L1s, chromatin compaction in oocytes is driven by actions of the TOP-2/condensin II axis and components of the H3K9me pathway. We note that the effects observed here for bivalent compaction after top-2/capg-2 RNAi are less extreme than those reported by Chan and colleagues, and this is likely due to differences in how condensin II was inactivated [26]. In our experiments, we use a feeding RNAi treatment that targets capg-2. By contrast, Chan and colleagues combined RNAi with a temperature-sensitive allele, both targeting the hcp-6 subunit of condensin II, thereby using two forms of condensin II inactivation in the same experiment. As detailed by Chan and colleagues, the effect on bivalent compaction of the temperature-sensitive hcp-6 (hcp-6ts) allele alone, without RNAi, is rather modest. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 7. TOP-2, condensin II and the H3K9 methyltransferases MET-2 and SET-25 are all required for proper bivalent compaction. Living proximal oocytes, treated with control, top-2/capg-2, met-2, or set-25 RNAi, were imaged for chromatin compaction using a strain harboring mCherry-tagged histone H2B. Bivalent volume in the -2 oocyte was measured using ImageJ. The average bivalent volume is plotted on the y-axis. 5 oocyte nuclei (with an average of 4 bivalents per nucleus) were analyzed for each RNAi treatment. Significance was measured using student’s t-test or Wilcoxon Rank Sum test. **p<0.01; *p<0.05. Exposure to top-2/capg-2, met-2 or set-25 RNAi treatments results in significantly larger bivalents. https://doi.org/10.1371/journal.pgen.1010831.g007 PIE-1 is required for genome silencing during meiotic prophase in oocytes and localizes to the nucleolus prior to silencing In a final set of experiments, we examined a requirement for PIE-1 in oocyte silencing, as recent work has shown that PIE-1 is present in the adult gonad [41]. Fig 8A shows that RNAi against pie-1 causes a persistence of transcription in the -2 position, and quantification shows that RNAPIIpSer2 signal intensity is significantly higher in both -3 and -2 oocytes after pie-1 RNAi, relative to the control samples (Fig 8B). Interestingly, PIE-1 depletion had no effect on bivalent compaction (Fig 8C). Thus, like TOP-2/condensin II and the H3K9me pathway, PIE-1 is required to repress transcription in -3/-2 oocytes, but unlike these factors its mechanism of action is distinct from chromatin compaction. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 8. PIE-1 is required for transcriptional repression in proximal oocytes. A. N2 animals were treated with either control or pie-1 RNAi. Gonads were dissected from these animals and were fixed and stained for DNA (blue) and RNAPIIpSer2 (red). Depletion of PIE-1 results in persisting RNAPIIpSer2 signal in proximal oocytes. B. Quantification of data presented in (A). RNAPIIpSer2 signal was measured using ImageJ. The average ratio of normalized raw integrated density is plotted on the y-axis. 20 samples were analyzed for each RNAi treatment over 2 independent replicates. Significance was measured using student’s t-test or Wilcoxon Rank Sum test. **p<0.01; *p<0.05. Oocytes from the pie-1 RNAi treatment had significantly increased RNAPIIpSer2 signal at the -3 and -2 positions compared to control RNAi. C. Living proximal oocytes treated with control or pie-1 RNAi were imaged for chromatin compaction using a strain harboring mCherry-tagged histone H2B. Bivalent volume in the -2 oocyte was measured using ImageJ. The average bivalent volume is plotted on the y-axis. 5 oocyte nuclei (with an average of 4 bivalents per nucleus) were analyzed for each RNAi treatment. Significance was measured using student’s t-test or Wilcoxon Rank Sum test. Exposure to pie-1 RNAi treatment did not affect bivalent volume. https://doi.org/10.1371/journal.pgen.1010831.g008 To pursue these findings, we next analyzed PIE-1 localization in oocytes, using a strain where GFP had been inserted at the endogenous pie-1 locus [41]. A typical localization pattern is shown in Fig 9A, where we see that PIE-1::GFP is mostly localized to the nucleus. Furthermore, within the -5 to -3 range of oocytes, it is clear that PIE-1::GFP accumulated within the nucleolus (Fig 9A and 9B), which can be easily observed using phase-contrast microscopy (Fig 9B). Previous work has shown that as oocytes prepare for maturation the nucleolus is lost, likely reflecting a shutdown of RNA polymerase I (RNAPI) transcription [42]. This explains why PIE-1::GFP is no longer predominantly localized to the nucleolus in -2 oocytes, as the nucleolus is undergoing dissolution (Fig 9A and 9B). Previous work in budding yeast has shown that condensin is required to remodel rDNA chromatin in preparation for cell division [43]. Given this, we wondered if TOP-2/condensin II is required for nucleolar dissolution in proximal oocytes. To address this, we used phase-contrast microscopy to image nucleoli, and we simply measured their area in control and top-2/capg-2 (RNAi) samples. As shown in Fig 9C and 9D, for control samples, we observed a significant decrease in nucleolar size in -2 oocytes, relative to -3 oocytes. This is consistent with the nucleolus undergoing dissolution in -2 oocytes. When nucleoli were assessed in samples depleted of TOP-2/CAPG-2, we saw that nucleolar size was significantly increased at both the -3 and -2 positions, relative to controls (Fig 9C and 9D). Thus, the TOP-2/condensin II axis promotes nucleolar dissolution in proximal oocytes. We also imaged PIE-1::GFP in samples depleted of TOP-2/condensin II and observed, as expected, that PIE-1 was localized to the -2 nucleoli that had resisted dissolution (Fig 9E). These data show that the TOP-2/condensin II axis controls PIE-1::GFP localization, and this likely occurs via TOP-2/condensin II’s ability to promote nucleolar dissolution. As detailed below in the Discussion, these findings suggest a model for how TOP-2/condensin II and PIE-1 work together to promote genome silencing in proximal oocytes. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 9. PIE-1 is sequestered in the nucleolus prior to silencing. A. The normal localization pattern of PIE-1::GFP in oocytes of live WM330 worms. From the -5 to -3 oocyte positions, PIE-1::GFP is sequestered within the nucleolus. Oocyte position is numbered. Scale bar represents a length of 10 μm. B. A wild-type localization pattern of PIE-1::GFP at the -3 oocyte position. PIE-1::GFP appears within the nucleolus. Nucleolus is indicated by the white circle. Scale bar represents a length of 10 μm. C. WM330 animals were treated with control or top-2/capg-2 double RNAi. Live adults were imaged at the -3 and -2 oocyte positions. Scale bar represents a length of 5 μm. D. Quantification of data presented in (C). Nucleolar size was measured using ImageJ and is plotted on the y-axis. Significance was measured using student’s t-test. **p<0.01; *p<0.05. Exposure to top-2/capg-2 treatment results in significantly larger nucleoli. E. Live adults were imaged for PIE-1::GFP after treatment with control or top-2/capg-2 double RNAi. Loss of top-2/capg-2 results in PIE-1::GFP remaining sequestered in the nucleolus at the -2 oocyte position. Scale bar represents a length of 5 μm. https://doi.org/10.1371/journal.pgen.1010831.g009 RNAPIIpSer2 is dependent on CDK-9 in proximal oocytes The goal of this study was to analyze genome silencing during oogenesis in C. elegans. To monitor transcription, we used an antibody that recognizes the active and elongating form of RNAPII. This reagent is a rabbit polyclonal antibody termed ab5095, purchased from Abcam (Waltham, MA), and it detects a phospho-epitope on the second serine within the carboxy-terminal repeat domain (CTD) of RNAPII (RNAPIIpSer2) [28]. In previous work we had validated ab5095 for use in C. elegans via the demonstration that reactivity depends on the presence of both RNAPII and phosphate, and that ab5095 accurately labels transcriptionally active nuclei in worm embryos and in the PGCs of L1 larvae [15]. For the current study, we used it to examine the transcriptional status of oocytes obtained from gonads dissected from adult hermaphrodites. Prior work that examined proximal oocytes has shown that RNAPIIpSer2 signals decreased starting with the fourth most proximal oocyte (termed -4) and became undetectable through the two most proximal oocytes (-2 and -1) [19]. We, therefore, focused our analysis on the four most proximal oocytes and found that RNAPIIpSer2 signal was detected invariably on the chromatin of oocytes at the -4 position (Fig 1B, control RNAi panel). However, as we examined the three most proximal oocytes, we observed several different patterns for the RNAPIIpSer2 signal: (1) signal exclusively on chromatin, (2) no signal at all, (3) signal present both on and off chromatin, and (4) a nucleoplasmic RNAPIIpSer2 signal that mostly excludes chromatin (Figs 1B and S1). It thus appears that while in some oocytes RNAPIIpSer2 signal completely disappeared, in others, it was merely removed from the chromatin and was present in the nucleoplasm. This form of nucleoplasmic signal has also been observed in human cells as they approach mitosis and is explained by RNAPII coming off the chromatin while maintaining CTD phosphorylations [29]. In mammals, the kinase that phosphorylates serine 2 within the RNAPII CTD is P-TEFb, which is composed of the CDK9 kinase and cyclins T1 or T2 [30]. To determine if the corresponding kinase in C. elegans, CDK-9, plays a similar role in proximal oocytes, we used RNAi to deplete the protein and we then stained for RNAPIIpSer2. As shown in Figs 1B and S1, exposure to cdk-9 RNAi caused a reduction in all three forms of RNAPIIpSer2 signal (on chromatin, off chromatin, and both on and off chromatin). This shows that the off-chromatin signals are due to CDK-9 phosphorylated RNAPII that had been displaced from chromatin but not yet dephosphorylated. For the remainder of this study, we only considered nuclei with RNAPIIpSer2 on chromatin as being transcriptionally active, and the off-chromatin signals were ignored. Interestingly, we also noticed that in more distal regions of the gonad, for example the pachytene region, RNAPIIpSer2 signal was unchanged by cdk-9 RNAi (Fig 1C). This is consistent with previous work showing that CDK-12 is the major RNAPIIpSer2 kinase in the mitotic and pachytene regions of the gonad [31]. It thus appears that there is a handoff, from CDK-12 to CDK-9, for generating RNAPIIpSer2 during oogenesis in C. elegans. The TOP-2/condensin II axis controls genome silencing during meiotic prophase Our data, together with previous findings [19], show that genome silencing initiates at the -3 position and is largely complete by the -2 position. As detailed above, there are four genome silencing systems that have been identified thus far in C. elegans: the TOP-2/condensin II axis (in starved L1 PGCs) [15], the H3K9me pathway (in starved L1 PGCs) [15]; the OMA-1/2 proteins (in one- and two-cell embryos) [11], and the PIE-1 protein (in the P2, P3, and P4 embryonic blastomeres) [12,13]. Which, if any, of these systems might also be used for genome silencing in oocytes? We know that OMA-1/2 are dispensable for oocyte silencing, as previous work had directly examined this possibility [11]. Thus, we turned our attention to the remaining three systems and focused initially on TOP-2/condensin II. To do so we examined RNAPIIpSer2 in oocytes from animals exposed to either top-2, capg-2 (which encodes a protein specific for the condensin II complex), or double top-2/capg-2 RNAi. All three of these treatments impacted the RNAPIIpSer2 pattern in a similar manner: signal now persisted in -2 oocytes, indicating that genome silencing had failed (Fig 2A and 2B). To quantify these data, we measured RNAPIIpSer2 signal intensity present on chromatin within individual oocytes. We normalized these values against signal intensities present in the pachytene region, an irrelevant tissue, in order to account for potential differences in antibody penetration across the different sample sets (see Methods for a more detailed explanation). As shown in Fig 2B, in the control samples, there was a significant difference in RNAPIIpSer2 signal intensity between the -3 and -2 positions, reflecting the genome silencing that occurs in -2 oocytes under normal conditions. Also shown in Fig 2B is the difference in RNAPIIpSer2 signal intensity between -2 oocytes from control samples, relative to the samples where TOP-2 and/or CAPG-2 had been depleted. Loss of TOP-2/CAPG-2 caused a significant increase in RNAPIIpSer2 signal at the -2 position. Based on these data, we conclude that loss of TOP-2 and/or CAPG-2 impacts genome silencing in -2 oocytes. Since genome silencing normally occurs in late prophase, one possibility is that loss of TOP-2/CAPG-2 somehow delays progression through diakinesis, and thus the persistence of RNAPII transcription in -2 oocytes of top-2/capg-2 depleted samples is an indirect consequence of defects in prophase timing. This is not the case, however, as when we stained for the prophase marker phospho-serine 10 of histone H3 (H3pS10) [32] we saw no difference in the timing of H3pS10 deposition (S2 Fig). We note that H3pS10 signal intensity was seemingly increased after top-2/capg-2 RNAi, however in the absence of an independent source of H3pS10 signal that can be used to normalize signal intensity across different sample sets it is impossible to know if this is a physiological effect or an artifact of differential antibody penetration. Nonetheless, it is clear that depletion of TOP-2/CAPG-2 does not delay prophase timing. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 2. Aberrant RNAPIIpSer2 signal is observed in proximal oocytes when TOP-2 and condensin II are depleted. A. N2 animals were treated with control, top-2, capg-2, or top-2/capg-2 double RNAi. Gonads were dissected from these animals and were fixed and stained for DNA (blue) and RNAPIIpSer2 (red). Each panel series (-3, -2, -1, and pachytene) represent nuclei from the same gonad. Individual depletion of TOP-2, CAPG-2, and their co-depletion results in persistent RNAPIIpSer2 signal on the chromatin of -3 and -2 oocytes. Scale bar represents a length of 2 μm. B. Quantification of data presented in (A). RNAPIIpSer2 signal was measured using ImageJ. The average ratio of normalized raw integrated density is plotted on the y-axis. Between 10 and 22 gonads were analyzed per condition and data were collected across multiple independently performed replicates. For control RNAi 22 gonads across 5 replicates were analyzed, for top-2 RNAi 15 gonads were analyzed across 2 replicates, for capg-2 RNAi it was 20 gonads across 2 replicates and for top-2/capg-2 RNAi it was 10 gonads across a single replicate. Significance was measured using student’s t-test or Wilcoxon Rank Sum test. **p<0.01; *p<0.05. In control animals, the ratio of signal density decreases significantly between -3 and -2 oocyte positions. Loss of TOP-2 or CAPG-2 results in persistent RNAPIIpSer2 signal in -2 oocytes. Co-depletion of TOP-2 and CAPG-2 produces a stronger phenotype, with persistent RNAPIIpSer2 signal in -3 and -2 oocytes. https://doi.org/10.1371/journal.pgen.1010831.g002 CDK-1 acts upstream of TOP-2/condensin II to trigger genome silencing Our data identify a role for TOP-2 and condensin II in the silencing of transcription as oocytes prepare for maturation and the meiotic divisions. Silencing initiates at the -3 position and is complete by -2, and this is also when condensin II is recruited to oocyte chromatin [26]. Previous work has also shown that condensin II activity is dependent on phosphorylation by CDK1 [33], and thus it stands to reason that CDK-1 may also play a role in genome silencing, through regulation of condensin II. To test this hypothesis, we used RNAi to reduce CDK-1 activity in proximal oocytes and we assessed the impact on transcriptional activity. As shown in Fig 3A and 3B, cdk-1 RNAi triggered a significant increase in RNAPIIpSer2 signal intensity in -3, -2, and -1 oocytes, showing that CDK-1 does indeed control genome silencing. To pursue this further, we next asked the opposite question—how does elevating CDK-1 activity impact genome silencing? For this we targeted the CDK-1 inhibitory kinase WEE-1.3 for depletion, as previous work has shown that WEE-1.3 negatively regulates CDK-1 in nematode oocytes and, further, that CDK-1 is the sole target of WEE-1.3 [34]. Treatment with wee-1.3 RNAi resulted in a decrease in RNAPIIpSer2 signal intensity across all oocytes spanning the -6 to -2 positions, showing that genome silencing was happening much earlier than in the control condition (Fig 4A and 4B). These data posed an important question: is the precocious silencing observed after wee-1.3 RNAi also dependent on TOP-2/condensin II? To address this, we co-depleted WEE-1.3 and CAPG-2, and we found that transcriptional activity was now restored to proximal oocytes (Fig 4A and 4B). Thus, co-depletion with CAPG-2 reverses the effects of WEE-1.3 depletion alone, and this shows that condensin II is required for the precocious silencing observed when CDK-1 levels are increased via depletion of WEE-1.3. Based on these data, we conclude that CDK-1 acts upstream of TOP-2/condensin II, in a positive manner, to promote genome silencing. Furthermore, our data suggest that the CDK-1 can activate condensin II for genome compaction and silencing at an activity state below that needed to trigger nuclear envelope breakdown (NEB) and entry into meiotic M-phase. This, in turn, suggests that CDK-1 activity rises gradually in the proximal gonad, as opposed to an abrupt activation in -1 oocytes. This would be consistent with previous work in HeLa cells as well as frog egg extracts showing that CDK-1-cyclin B activity rises gradually during interphase and prophase [35,36]. To gain additional evidence for a gradual acquisition of the M-phase state in proximal oocytes we stained them with MPM-2, an antibody that recognizes mitotic phosphoproteins [37]. As shown in S3A Fig, MPM-2 antigens are present at low levels in -5 oocytes and they gradually accumulate as we move proximally in the gonad, such that -1 oocytes show a high level of MPM-2 reactivity and -3 and -2 oocytes a more intermediate level. These data are consistent with the idea that CDK-1 activity gradually increases, and that levels of CDK-1 that are proficient to trigger genome compaction and silencing in -3/-2 oocytes are not yet sufficient for NEB and entry into meiotic M-phase. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 3. Aberrant RNAPIIpSer2 signal is observed in proximal oocytes when CDK-1 is depleted. A. N2 animals were treated with control or cdk-1 RNAi. Gonads were dissected from these animals and were fixed and stained for DNA (blue) and RNAPIIpSer2 (red). Loss of CDK-1 results in RNAPIIpSer2 signal persisting into the -2 and -1 oocyte positions. B. Quantification of data presented in (A). RNAPIIpSer2 signal was measured using ImageJ. The average ratio of normalized raw integrated density is plotted on the y-axis. 19 samples were analyzed for each RNAi treatment over 2 independent replicates. Significance was measured using student’s t-test or Wilcoxon Rank Sum test. **p<0.01; *p<0.05. Depletion of CDK-1 resulted in a significant increase in on-chromatin RNAPIIpSer2 signal in the three most proximal oocytes. https://doi.org/10.1371/journal.pgen.1010831.g003 Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 4. Hyperactivation of CDK by the depletion of WEE-1.3 causes unscheduled transcriptional silencing. A. N2 animals were treated with either control, wee-1.3, or wee-1.3/capg-2 double RNAi. Gonads were dissected from these animals and were fixed and stained for DNA (blue) and RNAPIIpSer2 (red). Depletion of WEE-1.3 results in the loss of on-chromatin RNAPIIpSer2 signals in the more distal (-6 to -4 position) oocytes. Co-depletion using wee-1.3/capg-2 RNAi reverses the loss of RNAPIIpSer2 signal. Scale bars represent a length of 2 μm. B. Quantification of data presented in (A). RNAPIIpSer2 signal was measured using ImageJ. The average ratio of normalized raw integrated density is plotted on the y-axis. 20 samples were analyzed for each RNAi treatment over 2 independently performed replicates. Significance was measured using student’s t-test or Wilcoxon Rank Sum test. **p<0.01; *p<0.05. Depletion of WEE-1.3 resulted in a decrease in RNAPIIpSer2 signal from -6 to -2 oocyte positions. Co-depletion of WEE-1.3 and CAPG-2 resulted in restored RNAPIIpSer2 signal in -6 to -3 oocyte positions. https://doi.org/10.1371/journal.pgen.1010831.g004 The presence of RNAPIIpSer2 signal at -1 oocytes after cdk-1 RNAi but not after top-2/capg-2 RNAi also raises the idea that there is an additional transcriptional repression mechanism in play when CDK-1 activity levels are high enough to trigger meiotic M-phase entry and NEB. To test this idea, we performed phase-contrast microscopy to assess nuclear envelope integrity of -1 oocytes in cdk-1 RNAi treated samples. We saw that more cdk-1 RNAi samples had intact nuclear envelopes than control RNAi treated samples (S3B Fig) establishing a correlation between silencing at -1 oocytes and NEB. When we performed a similar analysis on wee-1.3 RNAi oocytes (that precociously silence their genome in a top-2/condensin II dependent manner) we did not see a change in the number of samples with intact nuclear membrane relative to control RNAi animals (S3C Fig). These data show that the silencing at -1 is tied to NEB progression while the TOP-2/condensin II dependent silencing is not and suggest the existence of an additional silencing mechanism at -1 oocytes that is distinct from the silencing mechanism discussed so far in this manuscript. H3K9me3 marks accumulate on oocyte chromatin during genome silencing We next examined a role for the H3K9me pathway in oocyte genome silencing. Our previous work showed that the Z2/Z3 PGCs accumulate H3K9me marks to a significant extent, relative to neighboring somatic cells, as they prepare for silencing [15]. To see if this also occurs on oocyte chromatin we used an antibody, termed ab176916 and purchased from Abcam (Waltham, MA), that recognizes the tri-methylated form of lysine 9 on histone H3. We have previously validated this antibody with the demonstration that reactivity is lost in Z2/Z3 of L1 larvae in a strain lacking SET-25, the major H3K9me3 methyltransferase in C. elegans [15]. We limited the current analysis to H3K9me3 because available antibodies against H3K9me2 do not recognize their target when the neighboring serine 10 is phosphorylated [38], and we have seen in S2 Fig that H3pS10 is prominent in proximal oocytes. As shown in Fig 5A, H3K9me3 first appears in -5 oocytes, and then gradually accumulates such that by the -2 and -1 positions the H3K9me3 and DNA signals overlap extensively. To examine the reproducibility of this pattern we quantified H3K9me3 signal intensity across 11 gonads and then compared the values obtained for the -5 to -2 positions to the value obtained for -1 position within a given gonad. As shown in Fig 5B, the pattern of gradual accumulation of H3K9me3 marks as oocytes moved from distal to proximal was indeed highly reproducible. In C. elegans, H3K9me3 is produced primarily by the SET-25 methyltransferase [39]. Indeed, when H3K9me3 was examined in set-25 mutants and compared to wild type samples, we observed that the H3K9me3 signals were attenuated (Fig 5A). Fig 5C shows additional examples of -1 and -2 oocytes for wild type and set-25 mutants at higher magnification, highlighting the differences in H3K9me3 signal intensity. No such attenuation was observed when met-2 mutants were compared to wild type (Fig 5A), consistent with the MET-2 methyltransferase primarily responsible for producing H3K9me1 and me2 [39]. Our findings are also consistent with those of Bessler et al, 2010, who observed that H3K9me3 levels in the pachytene region of the gonad do not change in met-2 mutants, relative to wild type [40]. We conclude that H3K9me3 marks accumulate dramatically on chromatin at the time that oocytes are silencing their genomes. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 5. H3K9me3 signals significantly increase as oocytes become more proximal. A. Wild-type, met-2 mutant, and set-25 mutant gonads were dissected, fixed, and stained for DNA (blue) and H3K9me3 (red). In N2s, an increase in H3K9me3 signal is observed at the -3 to -1 positions in comparison to more distal oocytes and pachytene nuclei from the same animal. Loss of MET-2 does not affect the H3K9me3 accumulation pattern. Depletion of SET-25 results in loss of H3K9me3 signal in proximal oocytes. Scale bar represents a length of 2 μm. B. Quantification of data for wild type samples in (A). H3K9me3 signals for each oocyte in -5 to -2 positions relative to the most proximal oocyte are presented. 20 samples were analyzed over 2 independent replicates. H3K9me3 signal accumulates on chromatin in the more proximal oocyte positions. C. Additional examples of -2 and -1 oocytes stained as in part (A). Either wild type (N2) or set-25 mutant oocytes are shown. Note that the set-25 samples have only trace amounts of H3K9me3 on the chromatin, relative to the wild-type sample. https://doi.org/10.1371/journal.pgen.1010831.g005 The accumulation of H3K9me3 in the oocytes was reminiscent of the H3K9me spreading we saw during L1 starvation which is dependent on TOP-2 and condensin II [15]. Given this, we wished to know if TOP-2 and condensin II also facilitated the spreading of H3K9me3 in maturing oocytes by staining gonads from top-2/capg-2 RNAi treated samples for H3K9me3. We saw that there was no difference in H3K9me3 signal between control and top-2/capg-2 RNAi treated samples (S4A and S4B Fig). Therefore, unlike in starved L1s, the accumulation of H3K9me3 in maturing oocytes does not require TOP-2 and condensin II. Both the SET-25 and MET-2 methyltransferases are required for genome silencing in oocytes Having observed SET-25 dependent accumulation of H3K9me3 marks in proximal oocytes we next asked if SET-25 plays a role in genome silencing. Staining of set-25 mutants for RNAPIIpSer2 revealed that -2 oocytes contained significantly more RNAPIIpSer2 signal than did the control samples (Fig 6A and 6B), showing that silencing was attenuated. Interestingly, we also observed a silencing defect in met-2 mutants (Fig 6A and 6B). Thus, not only do H3K9me marks accumulate during silencing, but the enzymes responsible for catalyzing these modifications are also important for genome silencing. We conclude that the H3K9me pathway globally silences transcription in developing oocytes, as it does in the PGCs of starved L1s. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 6. The H3K9 methyltransferases MET-2 and SET-25 are required for transcriptional repression in proximal oocytes. A. Gonads from wild-type, met-2 mutant, and set-25 mutant animals were dissected, fixed, and stained for DNA (blue) and RNAPIIpSer2 (red). Mutations of either met-2 or set-25 results in persistent RNAPIIpSer2 signal on chromatin, compared to wild-type animals. Scale bar represents a length of 2 μm. B. Quantification of data presented in (A). RNAPIIpSer2 signal was measured using ImageJ. The average ratio of normalized raw integrated density is plotted on the y-axis. 10 to 20 samples were analyzed for each condition over 2 independent replicates. 15 wild-type samples were analyzed while 10 met-2 mutant and 20 set-25 mutants were analyzed. Significance was measured using student’s t-test or Wilcoxon Rank Sum test. **p<0.01; *p<0.05. Loss of MET-2 resulted in a significant persistence of RNAPIIpSer2 signal on chromatin in -2 and -1 oocytes. Loss of SET-25 led to a significant persistence of RNAPIIpSer2 signal on chromatin in -2 oocytes. https://doi.org/10.1371/journal.pgen.1010831.g006 Genome silencing is coupled to chromatin compaction in oocytes Previous work has shown that condensin II loads on to oocyte chromosomes at the -3 position and is required for the intense chromatin compaction that occurs as bivalents are formed [26]. In starved L1s, genome silencing is coupled to chromatin compaction [15], and thus it was important to monitor compaction in the oocyte system. For this we turned to a previously utilized strain that carries a transgene encoding mCherry-tagged histone H2B, which marks chromatin [15,17]. Living hermaphrodites were immobilized and oocyte chromatin was imaged using confocal microscopy. We compared control samples to those that been exposed to top-2/capg-2 RNAi, and we looked at -2 oocytes, which is where genome silencing is occurring. As described in the Methods and S5 Fig, we measured the volume of the chromatin masses and found that depletion of TOP-2/CAPG-2 caused a significant increase in volume, consistent with a defect in compaction, and similar observations were made after set-25 or met-2 RNAi (Figs 7 and S5). Thus, as is the case in the PGCs of starved L1s, chromatin compaction in oocytes is driven by actions of the TOP-2/condensin II axis and components of the H3K9me pathway. We note that the effects observed here for bivalent compaction after top-2/capg-2 RNAi are less extreme than those reported by Chan and colleagues, and this is likely due to differences in how condensin II was inactivated [26]. In our experiments, we use a feeding RNAi treatment that targets capg-2. By contrast, Chan and colleagues combined RNAi with a temperature-sensitive allele, both targeting the hcp-6 subunit of condensin II, thereby using two forms of condensin II inactivation in the same experiment. As detailed by Chan and colleagues, the effect on bivalent compaction of the temperature-sensitive hcp-6 (hcp-6ts) allele alone, without RNAi, is rather modest. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 7. TOP-2, condensin II and the H3K9 methyltransferases MET-2 and SET-25 are all required for proper bivalent compaction. Living proximal oocytes, treated with control, top-2/capg-2, met-2, or set-25 RNAi, were imaged for chromatin compaction using a strain harboring mCherry-tagged histone H2B. Bivalent volume in the -2 oocyte was measured using ImageJ. The average bivalent volume is plotted on the y-axis. 5 oocyte nuclei (with an average of 4 bivalents per nucleus) were analyzed for each RNAi treatment. Significance was measured using student’s t-test or Wilcoxon Rank Sum test. **p<0.01; *p<0.05. Exposure to top-2/capg-2, met-2 or set-25 RNAi treatments results in significantly larger bivalents. https://doi.org/10.1371/journal.pgen.1010831.g007 PIE-1 is required for genome silencing during meiotic prophase in oocytes and localizes to the nucleolus prior to silencing In a final set of experiments, we examined a requirement for PIE-1 in oocyte silencing, as recent work has shown that PIE-1 is present in the adult gonad [41]. Fig 8A shows that RNAi against pie-1 causes a persistence of transcription in the -2 position, and quantification shows that RNAPIIpSer2 signal intensity is significantly higher in both -3 and -2 oocytes after pie-1 RNAi, relative to the control samples (Fig 8B). Interestingly, PIE-1 depletion had no effect on bivalent compaction (Fig 8C). Thus, like TOP-2/condensin II and the H3K9me pathway, PIE-1 is required to repress transcription in -3/-2 oocytes, but unlike these factors its mechanism of action is distinct from chromatin compaction. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 8. PIE-1 is required for transcriptional repression in proximal oocytes. A. N2 animals were treated with either control or pie-1 RNAi. Gonads were dissected from these animals and were fixed and stained for DNA (blue) and RNAPIIpSer2 (red). Depletion of PIE-1 results in persisting RNAPIIpSer2 signal in proximal oocytes. B. Quantification of data presented in (A). RNAPIIpSer2 signal was measured using ImageJ. The average ratio of normalized raw integrated density is plotted on the y-axis. 20 samples were analyzed for each RNAi treatment over 2 independent replicates. Significance was measured using student’s t-test or Wilcoxon Rank Sum test. **p<0.01; *p<0.05. Oocytes from the pie-1 RNAi treatment had significantly increased RNAPIIpSer2 signal at the -3 and -2 positions compared to control RNAi. C. Living proximal oocytes treated with control or pie-1 RNAi were imaged for chromatin compaction using a strain harboring mCherry-tagged histone H2B. Bivalent volume in the -2 oocyte was measured using ImageJ. The average bivalent volume is plotted on the y-axis. 5 oocyte nuclei (with an average of 4 bivalents per nucleus) were analyzed for each RNAi treatment. Significance was measured using student’s t-test or Wilcoxon Rank Sum test. Exposure to pie-1 RNAi treatment did not affect bivalent volume. https://doi.org/10.1371/journal.pgen.1010831.g008 To pursue these findings, we next analyzed PIE-1 localization in oocytes, using a strain where GFP had been inserted at the endogenous pie-1 locus [41]. A typical localization pattern is shown in Fig 9A, where we see that PIE-1::GFP is mostly localized to the nucleus. Furthermore, within the -5 to -3 range of oocytes, it is clear that PIE-1::GFP accumulated within the nucleolus (Fig 9A and 9B), which can be easily observed using phase-contrast microscopy (Fig 9B). Previous work has shown that as oocytes prepare for maturation the nucleolus is lost, likely reflecting a shutdown of RNA polymerase I (RNAPI) transcription [42]. This explains why PIE-1::GFP is no longer predominantly localized to the nucleolus in -2 oocytes, as the nucleolus is undergoing dissolution (Fig 9A and 9B). Previous work in budding yeast has shown that condensin is required to remodel rDNA chromatin in preparation for cell division [43]. Given this, we wondered if TOP-2/condensin II is required for nucleolar dissolution in proximal oocytes. To address this, we used phase-contrast microscopy to image nucleoli, and we simply measured their area in control and top-2/capg-2 (RNAi) samples. As shown in Fig 9C and 9D, for control samples, we observed a significant decrease in nucleolar size in -2 oocytes, relative to -3 oocytes. This is consistent with the nucleolus undergoing dissolution in -2 oocytes. When nucleoli were assessed in samples depleted of TOP-2/CAPG-2, we saw that nucleolar size was significantly increased at both the -3 and -2 positions, relative to controls (Fig 9C and 9D). Thus, the TOP-2/condensin II axis promotes nucleolar dissolution in proximal oocytes. We also imaged PIE-1::GFP in samples depleted of TOP-2/condensin II and observed, as expected, that PIE-1 was localized to the -2 nucleoli that had resisted dissolution (Fig 9E). These data show that the TOP-2/condensin II axis controls PIE-1::GFP localization, and this likely occurs via TOP-2/condensin II’s ability to promote nucleolar dissolution. As detailed below in the Discussion, these findings suggest a model for how TOP-2/condensin II and PIE-1 work together to promote genome silencing in proximal oocytes. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 9. PIE-1 is sequestered in the nucleolus prior to silencing. A. The normal localization pattern of PIE-1::GFP in oocytes of live WM330 worms. From the -5 to -3 oocyte positions, PIE-1::GFP is sequestered within the nucleolus. Oocyte position is numbered. Scale bar represents a length of 10 μm. B. A wild-type localization pattern of PIE-1::GFP at the -3 oocyte position. PIE-1::GFP appears within the nucleolus. Nucleolus is indicated by the white circle. Scale bar represents a length of 10 μm. C. WM330 animals were treated with control or top-2/capg-2 double RNAi. Live adults were imaged at the -3 and -2 oocyte positions. Scale bar represents a length of 5 μm. D. Quantification of data presented in (C). Nucleolar size was measured using ImageJ and is plotted on the y-axis. Significance was measured using student’s t-test. **p<0.01; *p<0.05. Exposure to top-2/capg-2 treatment results in significantly larger nucleoli. E. Live adults were imaged for PIE-1::GFP after treatment with control or top-2/capg-2 double RNAi. Loss of top-2/capg-2 results in PIE-1::GFP remaining sequestered in the nucleolus at the -2 oocyte position. Scale bar represents a length of 5 μm. https://doi.org/10.1371/journal.pgen.1010831.g009 Discussion The goal of this study was to define the molecular components for genome silencing in C. elegans oocytes. To monitor transcription, we relied on RNAPIIpSer2 staining, as we and others have done extensively in the past [11,15,19,44,45]. One concern with this antibody-based approach is that chromatin compaction in -2 oocytes may prevent access of the antibody to its target on chromatin, rendering false-negative data. This is clearly not the case, however, as when we depleted PIE-1 we observed strong RNAPIIpSer2 signals on chromatin in -2 oocytes, even though compaction occurs normally under this condition (Fig 8). Thus, we consider RNAPIIpSer2 staining to be an accurate and legitimate method to assess the transcriptional status of oocytes in the worm. Under normal conditions, we found that RNAPIIpSer2 signal intensity drops significantly in -2 oocytes relative to the -3 position, and that signals are often undetectable at -2 (Fig 2). This is consistent with previous work [19], and thus we conclude that genome silencing likely initiates at -3 and is largely complete by -2. At the -2 position, all oocytes have intact nuclear envelopes, and thus these cells have yet to enter meiotic M-phase. This is important as previous work has shown that transcription is repressed during M-phase [46–48], although more recent work has shown that the block is not absolute as a low-level of transcription can be detected in mitotic cells [49]. Nonetheless, it is clear that M-phase is incompatible with active transcription, and thus it is important to distinguish the genome silencing we observe in -2 oocytes from the repression of transcription that likely occurs during meiotic M-phase in -1 oocytes. Our data support this distinction in multiple ways. First, depletion of the genome silencing factors TOP-2, CAPG-2, MET-2, SET-25, and PIE-1 all yield a common phenotype—a persistence of RNAPIIpSer2 signal in -2 oocytes, and loss of the signal in -1 oocytes (Figs 2, 6, and 8). Thus, if loss of active transcription in -2 oocytes occurs mechanistically the same as in -1 oocytes, then we should see a persistence of signal in the -1 position also, and we clearly do not. Second, depletion of CDK-1 results in persistence of RNAPIIpSer2 signals in -2 and -1 oocytes (Fig 3). After exposure to cdk-1 RNAi, -1 oocytes largely retain the nuclear envelope (S3B Fig), consistent with a failure to enter meiotic M-phase. This shows that it is not simply occupancy of the -1 position that represses transcription; rather, it is M-phase entry that does so. The finding that transcription is largely repressed during M-phase was made in the early 1960s however the mechanistic basis is still unknown. A previous study examined hsp70 gene expression during mitosis and found that three of the four transcription factors (TFs) needed for hsp70 activation were physically displaced from chromatin at mitosis [50], however how this displacement occurs was not described. It may be that the TFs are physically displaced by chromosome compaction, or it may be that modification of the TFs prevents them from binding DNA. Our data shed light on this as we show that M-phase repression of RNAPIIpSer2 signals is independent of chromosome compaction. We found that loss of PIE-1 allows compaction, but not RNAPIIpSer2 repression, in -2 oocytes (Fig 8). In these same samples, at the -1 position, PIE-1 is no longer needed to suppress RNAPIIpSer2 as an additional mechanism gets activated. Because there is no further compaction happening at the -1 position [26], this additional mechanism is independent of compaction. Indeed, previous work on virus-infected human cells has shown that M-phase transcriptional repression occurs on nucleosome-free and uncompacted DNA [51], which is consistent with our data that the M-phase mechanism is distinct from compaction. In summary, we believe that the PIE-1 data demonstrate two independent mechanisms for repression at -2 relative to -1. At -2, chromatin compaction is necessary, but not sufficient, for repression. At -1, repression occurs on previously compacted chromatin, and thus is mechanistically distinct from what is happening at -2. Taken together, these data show that genome silencing precedes entry into meiotic M-phase. Why has the worm evolved a system to block transcription at -2 when it is going to happen anyway at -1? We propose that active transcription may hamper the chromosomal remodeling that occurs at -2, and thus that genome silencing at -2 has evolved to allow proper bivalent formation. Taking a candidate approach, we found that multiple genome silencing pathways are operational in -2 oocytes, as loss of the TOP-2/condensin II axis, the H3K9me pathway, and PIE-1 all impact silencing. We also found that silencing is under cell cycle control, as reducing CDK-1 activity prevents silencing in -2 oocytes and increasing CDK-1 activity promotes precocious silencing in oocytes distal to -2 (Figs 3 and 4). Furthermore, we obtained evidence that oocytes gradually approach the M-phase state, as MPM-2 antigens accumulate gradually in the proximal gonad, and not abruptly at the -1 position as one might expect if CDK-1 is abruptly activated in -1 oocytes (S3A Fig). Taken together, these data paint a picture where CDK-1 activity increases gradually as a function of oocyte position in the proximal gonad, and that it is the -2 position where CDK-1 activity crosses a threshold sufficient to trigger silencing, but not yet sufficient to trigger entry into meiotic M-phase. Thus, we propose that the timing of genome silencing is controlled by a CDK-1 activity gradient spanning the proximal gonad (Fig 10A). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 10. Models for how genome silencing occurs in proximal oocytes. A. A CDK-1 activity gradient allows for genome silencing at the -2 position to occur prior to entry into meiotic M-phase at the -1 position. B. A proposed pathway whereby CDK-1 promotes TOP-2/condensin II mediated compaction of rDNA, and this in turn promotes nucleolar dissolution and relocalization of PIE-1 to the nucleoplasm where it can block transcription. https://doi.org/10.1371/journal.pgen.1010831.g010 Our previous work had identified a linear pathway, termed GCC, that silences germline transcription during L1 starvation [15]. For GCC, TOP-2/condensin II acts upstream of the MET-2 and SET-25 methyltransferases to promote H3K9me2 and -me3 deposition on chromatin [15]. Interestingly, while TOP-2/condensin II and SET-25/MET-2 are all required for silencing in oocytes, they are not organized into a linear pathway as we found that H3K9me3 deposition in oocytes does not require TOP-2/condensin II (S4A and S4B Fig). H3K9me3 deposition is increased as oocytes approach the -1 position (Fig 5), and this is similar to what happens in late embryos, where deposition is dramatically increased in Z2/Z3, relative to their somatic counterparts [15]. How this increase is regulated, and whether it occurs via spreading of preexisting marks or via de novo deposition are fascinating questions for future research. As mentioned above, we were unable to track H3K9me2 marks, but it is interesting to note that loss of met-2, which is responsible for H3K9me2 deposition, has a stronger effect on silencing than does loss of set-25, which is responsible for H3K9me3 (Fig 6). It thus appears that both H3K9me2 and -me3 play a role in genome silencing, and another important route for future research will be to determine which of the H3K9me readers is involved and how they are acting mechanistically to promote silencing. We note that in mice, it has been appreciated for some time that oocytes undergo chromatin compaction concurrent with transcriptional repression [8,52–54]. Interestingly, very recent work has shown that a histone H3.3 chaperone complex, comprised of the Hira and Cabin1 proteins, promotes H3K9me3 deposition in chromatin, chromatin compaction, and genome silencing in mouse oocytes [55]. These data are consistent with our findings and suggest that a conserved feature of genome silencing is the H3K9me-mediated compaction of chromatin on a global scale. It will be of interest to determine if TOP-2 and/or condensin II are also required for genome silencing in murine oocytes. Our work also describes a novel function for PIE-1 in oocyte genome silencing. PIE-1 has been well studied in its role of blocking transcription in the P-lineage of early embryos, however how it does so is still a mystery. Early work suggested a model where PIE-1 binds to and sequesters cyclin T, a subunit of the CDK9 kinase that phosphorylates RNAPII on serine 2 within the CTD [56]. More recent work, however, has shown that PIE-1 mutants that fail to bind cyclin T still block transcription in the P-lineage [57] and, furthermore, that the RNAPIIpSer2 mark itself is dispensable for embryogenesis [45]. Thus, it may be that, in the embryo, PIE-1 acts by interfering with RNAPII serine 5 phosphorylation [57], via an unknown mechanism. If so, this begs the question of why has PIE-1 evolved the ability to specifically interact with cyclin T? One plausible answer is that PIE-1 targets cyclin T to block transcription in oocytes. Interestingly, we have shown here that CDK-9, presumably acting with cyclin T, is the relevant serine 2 kinase in proximal oocytes, unlike the remainder of the gonad where CDK-12 is the relevant kinase [31]. This finding makes it intriguing to speculate that the switch from CDK-12 to CDK-9 occurs so that RNAPIIpSer2 can be regulated by PIE-1 in proximal oocytes. Sorting out how PIE-1 is blocking transcription in oocytes is another important avenue for future research. Lastly, another important research question raised by our study is how is PIE-1 regulated in the proximal gonad? We have shown that PIE-1-GFP is present in the nuclei of -5, -4, and -3 oocytes, yet these cells are transcriptionally active (Figs 1B and 9A). Importantly, in oocytes distal to -2, PIE-1-GFP is localized predominantly in nucleoli (Fig 9). This might explain why these nuclei are competent for transcription despite the presence of PIE-1, if nucleolar residency prevents PIE-1 from accessing its target(s) for transcriptional repression. Previous work has shown that the nucleolus dissolves at the -2 position [42]. We have observed this as well and, furthermore, we have shown that dissolution requires TOP-2/condensin II (Fig 9). How might the various components required for genome silencing in oocytes fit together mechanistically? Our data support the model shown in Fig 10B. We propose that once CDK-1 activity passes a threshold then TOP-2/condensin II is activated and recruited to chromatin, and this has two consequences. One, chromatin compaction commences and this represses transcription, likely via occlusion of RNAPII and various transcription factors from promoters on the compacted chromatin. Two, as the rDNA is compacted, RNAPI synthesis is blocked and the nucleolus dissolves, thereby liberating PIE-1 to block transcription via an unknown mechanism. Lastly, independent of TOP-2/condensin II, the methyltransferases targeting H3K9 are stimulated and H3K9me deposition is hyper-activated, leading to chromatin compaction and genome silencing. While this model is consistent with our data, there is clearly much more work needed to establish its accuracy. Future experiments will address the role of PIE-1 nucleolar residency in its regulation as well as the mechanism by which the SET-25 and MET-2 methyltransferases are activated and how H3K9me marks accumulate so dramatically on oocyte chromatin. Materials and methods C. elegans strains N2 (wild-type), WMM1 ([pie-1::gfp::pgl-1 + unc-119(+)]; [(pAA64)pie-1p::mCherry::his-58 + unc-119(+)] IV), MT13293(met-2(n4256) III), MT17463 (set-25(n5021) III) and WM330 (pie-1(ne4301[pie-1::GFP]) III) strains were used in this study. Worms were maintained on 60mm plates containing nematode growth media (NGM) seeded with the E. coli strain OP50 or HT115. Worms were grown at 20°C and propagated through egg preparation (bleaching) every 72 hours. Bacterial strains OP50 bacteria served as the primary food source. It was grown in LB media containing 100 μg/ml streptomycin by shaking at 37°C overnight. 500 μl of the culture was seeded on Petri-dishes containing NGM + streptomycin. HT115 bacteria grown in LB media containing 100 μg/ml carbenicillin and 12.5 μg/ml tetracycline and seeded on NGM + carbenicillin + tetracycline plates were also used as a source of food. Our RNAi strains were obtained from the Ahringer library and verified by Sanger sequencing. Bacteria containing dsRNA were streaked on LB-agar plates containing 100 μg/ml carbenicillin and 12.5 μg/ml tetracycline and incubated at 37°C overnight. Single colonies were then picked and grown in 25 ml LB cultures with 100 μg/ml carbenicillin and 12.5 μg/ml tetracycline. 500 μl of this culture was seeded on 60-mm Petri dishes containing 5mM IPTG. Egg preparation Bleach solution containing 3.675 ml H2O, 1.2 NaOCl, and 0.125 ml 10N NaOH was prepared. Adult worms were washed from plates with 5 ml of M9 minimal medium (22mM KH2PO4, 22mM Na2HPO4, 85mM NaCl, and 2mM MgSO4). Worms were centrifuged at 1.9 KRPM for 1 minute and the excess medium was removed, then the bleach solution was added. Eggs were extracted by vortexing for 30 seconds and shaking for 1 minute. This was done a total of 3 times and worms were vortexed one last time. Then the eggs were spun down at 1900 rpm for 1 minute and excess bleach solution was removed, and the eggs were washed 3 times with M9 minimal medium. RNAi treatment RNAi containing NGM plates were prepared as described in the “Bacterial strains” section. For double RNAi treatments, RNAi cultures were mixed at a 1:1 ratio by volume. HT115 cells transformed with an empty pL4440 vector was used as a negative control. RNAi conditions used in this study and tests for their efficacy is described below: cdk-9 RNAi. L1 worms were plated on HT115 food plates for the first 48 hours and were then moved to plates containing cdk-9 RNAi for the remaining 24 hours. Embryonic lethality in the range of 80%—85% was observed. top-2 RNAi. L1 worms were plated on HT115 food plates for the first 24 hours and were then moved to plates seeded with top-2 RNAi for the remaining 48 hours. Embryonic lethality was observed at >90%. capg-2 RNAi. Worms were grown on HT115 food plates for the first 24 hours and were moved to plates containing capg-2 RNAi for the remaining 48 hours. An embryonic lethality of 80%-100% was seen with this RNAi treatment. top-2/capg-2 double RNAi. Worms were grown on HT115 food plates for the first 24 hours and were transferred to top-2/capg-2 double RNAi plates for the next 48 hours. Embryonic lethality ranged from 90%-100% for this RNAi treatment. cdk-1 RNAi. Worms were grown on HT115 food plates for the first 24 hours and were transferred to cdk-1 RNAi plates for the next 48 hours. Embryonic lethality ranged from 97%-100% for this RNAi treatment. wee-1.3 RNAi. Worms were grown on plates containing wee-1.3 RNAi for the entirety of their life cycle. An embryonic lethality of approximately 40% was observed. Additionally, a significant reduction in brood size, and the coalescence of bivalents into one chromatin mass in proximal oocytes of some samples, were observed for wee-1.3 RNAi worms, as previously reported [34]. wee-1.3/capg-2 double RNAi. Worms were grown on plates containing wee-1.3 RNAi for 24 hours then were moved to plates containing wee-1.3/capg-2 RNAi where they remained for the rest of their life cycle. An embryonic lethality of approximately80%, and the coalescence of bivalents in proximal oocytes of some samples were observed. met-2 RNAi. Worms were grown on plates containing met-2 RNAi for the entirety of their life cycle. Some of the adult worms were bleached and an L1 chromatin compaction assay was performed on the resulting larvae to test RNAi efficacy. See Belew et al., 2021, for details on the L1 compaction assay. set-25 RNAi. Worms were grown on plates containing set-25 RNAi for the entirety of their life cycle. RNAi efficacy was tested via the same method as for met-2 RNAi. pie-1 RNAi. Worms were grown on pie-1 RNAi plates for the entirety of their life cycle. An embryonic lethality of 100% was observed for this RNAi. Antibodies and dilutions RNAPIIpSer2: Rabbit antibody from Abcam (ab5095, Waltham, Massachusetts) was used at a dilution of 1:100. H3pSer10: Rabbit antibody from Rockland Immunochemicals (600-401-I74, Pottstown, Pennsylvania) was used at a dilution of 1:500. H3K9me3: Rabbit antibody from Abcam (ab176916, Waltham, Massachusetts) was used at a dilution of 1:1000. MPM-2: Mouse antibody (isotype—IgG1) from Sigma-Aldrich (05–368, St. Louis, Missouri) was used at a dilution of 1:500. Secondary antibodies: Alexa Fluor conjugated secondary antibodies from Invitrogen (Thermo Fisher Scientific, Waltham, Massachusetts) were used at a dilution of 1:200. Immunofluorescence staining Adult worms were first washed off plates with 10 ml of M9 minimal medium and rinsed 3 more times. Then, they were centrifuged at 1.9 KRPM and the excess medium was removed. 20 μl of media containing about 50 worms were spotted on a coverslip and 3 μl of anesthetic (20mM Sodium Azide and 0.8M Tetramisole hydrochloride) was added to immobilize them. Worms were dissected using 25Gx5/8 needles (Sigma Aldrich, St. Louis, Missouri). To release gonads, adult worms were cut twice, once right below their pharyngeal bulb and once near the tail. The coverslip was then mounted onto poly-L-lysine covered slides and let rest for 5 minutes. Slides were put on dry ice for 30 minutes. Samples were then freeze-cracked by flicking the coverslips off for permeabilization. For RNAPIIpSer2, H3pSer10 and MPM-2 antibody staining experiments, once samples were permeabilized, slides were put in cold 100% methanol (-20°C) for 2 minutes and then fixing solution (0.08M HEPES pH 6.9, 1.6mM MgSO4, 0.8mM EGTA, 3.7% formaldehyde, 1X phosphate-buffered saline) for another 30 minutes. After fixing, slides were washed three times with TBS-T (TBS with 0.1% Tween-20) and were blocked for 30 minutes with TNB (containing 100mM Tris-HCl, 200 mM NaCl, and 1% BSA). Primary antibodies were then applied at the dilutions described above in TNB and slides were incubated at 4°C overnight. For H3K9me3 staining experiments, permeabilized samples were put in cold 100% methanol (-20°C) for 10 seconds and then fixing solution (0.08M HEPES pH 6.9, 1.6mM MgSO4, 0.8mM EGTA, 3.7% formaldehyde, 1X phosphate-buffered saline) for 10 minutes. After fixing, slides were washed three times with TBS-T (TBS with 0.1% Tween-20) and were blocked for 2 hours with TNB (containing 100mM Tris-HCl, 200 mM NaCl, and 1% BSA) supplemented with 10% normal goat serum. Primary antibodies were then applied at the dilutions described above in TNB supplemented with 10% goat serum and slides were incubated at 4°C overnight. On the next day, the slides were washed 3 times with TBS and slides were incubated with secondary antibodies and Hoechst 33342 dye for 2 hours at room temperature. Slides were washed 3 times with TBS, mounting medium (50% glycerol in PBS), and coverslips were applied and sealed with Cytoseal XYL (Thermo Fisher Scientific, Waltham Massachusetts). Live animal imaging Adult N2, WMM1 and EGW83 worms were collected off plates and were washed 3 times with 10 ml M9 minimal medium. After the last wash, worms were spun down at 1.9 KRPM and excess medium was removed. 0.3% agarose pads were made on slides, and a 10 μl aliquot of adult worms was mounted. 4 μl of anesthetic (20mM Sodium Azide and 0.8M Tetramisole hydrochloride) was added to stop the worms from moving. A coverslip was gently applied, and the slides were imaged. Immunofluorescent imaging All slides were imaged using an Olympus Fluoview FV1000 confocal microscope using Fluoview Viewer software at a magnification of 600x (60x objective and 10x eyepiece magnifications). Laser intensity was controlled for experiments to achieve consistency among samples. Quantification of data RNAPIIpSer2 signal quantification. For each oocyte nucleus, two images were taken: one of the Hoechst-stained DNA and one of the RNAPIIpSer2 signal. Images were analyzed using ImageJ. An outline was drawn using the polygon selection tool around the Hoechst-stained area to mark the space occupied by DNA. The region of interest was copied and pasted to the RNAPIIpSer2 signal image and the raw integrated density (the sum of the values of the pixels in the selection) was measured. The raw integrated density was then normalized by the area to get a measure we called “signal density”. To account for possible variability in signal intensity due to different degrees of antibody penetration from sample to sample, each oocyte’s signal density was normalized to an average signal density from 5 pachytene nuclei found in the same image. The final normalized signal densities for the oocytes are presented. For each condition, were collected from two independently performed experimental replicates and the data were then pooled for statistical analysis and presentation. H3K9me3 signal quantification. Signal densities for each oocyte were calculated similarly to what is described above in the RNAPIIpSer2 signal quantification. For data on Fig 5, H3K9me3 signal of each oocyte in the -2 to -5 positions was normalized to the most proximal (-1) oocyte and the ratios are presented as percentages. For data presented in S4 Fig signal densities of control and top-2/capg-2 RNAi samples were shown side by side for comparison. H3pSer10 signal quantification. Signal densities for each oocyte were calculated similarly to what is described above in the RNAPIIpSer2 signal quantification. Signal densities were not normalized to pachytene nuclei since pachytene nuclei either did not harbor any signal. (for H3pSer10 staining) or the signals were affected by our RNAi treatments (for H3K9me3 staining) which precluded us from using them for unbiased normalization. Bivalent volume quantification. Z-stacks of oocyte nuclei were acquired. For each condition, 5 nuclei were analyzed and every bivalent within those nuclei that were visually distinguishable from one another was included in our analysis. On average 4 bivalents per nucleus were analyzed using ImageJ. The scale was set to 69nm per pixel. To measure the volume of a bivalent, a polygon was tightly drawn around it on each stack the bivalent appears in. The areas of these polygons were measured and summed up. Finally, the sum was multiplied by the distance between each stack to calculate an approximation of the bivalent’s volume. Averages of the volumes of all bivalents analyzed were presented. Nucleolar size measurement. Images were captured of focal planes corresponding to the maximum diameter of the nucleolus in living oocytes for each oocyte position. Diameters were measured using ImageJ and then converted to area. Statistical analysis Prior to performing any statistical test, data was tested whether it was parametric or not. To do so, the Shapiro-Wilk test was used to test for normal distribution and F-test was used to test for variance homogeneity of the datasets we were comparing. Data were then analyzed using a student’s t-test or Wilcoxon Rank Sum test depending on whether the datasets fulfill the requirements for a parametric test or not. Differences between any two datasets were considered statistically significant if a P-value of <0.05 was obtained. C. elegans strains N2 (wild-type), WMM1 ([pie-1::gfp::pgl-1 + unc-119(+)]; [(pAA64)pie-1p::mCherry::his-58 + unc-119(+)] IV), MT13293(met-2(n4256) III), MT17463 (set-25(n5021) III) and WM330 (pie-1(ne4301[pie-1::GFP]) III) strains were used in this study. Worms were maintained on 60mm plates containing nematode growth media (NGM) seeded with the E. coli strain OP50 or HT115. Worms were grown at 20°C and propagated through egg preparation (bleaching) every 72 hours. Bacterial strains OP50 bacteria served as the primary food source. It was grown in LB media containing 100 μg/ml streptomycin by shaking at 37°C overnight. 500 μl of the culture was seeded on Petri-dishes containing NGM + streptomycin. HT115 bacteria grown in LB media containing 100 μg/ml carbenicillin and 12.5 μg/ml tetracycline and seeded on NGM + carbenicillin + tetracycline plates were also used as a source of food. Our RNAi strains were obtained from the Ahringer library and verified by Sanger sequencing. Bacteria containing dsRNA were streaked on LB-agar plates containing 100 μg/ml carbenicillin and 12.5 μg/ml tetracycline and incubated at 37°C overnight. Single colonies were then picked and grown in 25 ml LB cultures with 100 μg/ml carbenicillin and 12.5 μg/ml tetracycline. 500 μl of this culture was seeded on 60-mm Petri dishes containing 5mM IPTG. Egg preparation Bleach solution containing 3.675 ml H2O, 1.2 NaOCl, and 0.125 ml 10N NaOH was prepared. Adult worms were washed from plates with 5 ml of M9 minimal medium (22mM KH2PO4, 22mM Na2HPO4, 85mM NaCl, and 2mM MgSO4). Worms were centrifuged at 1.9 KRPM for 1 minute and the excess medium was removed, then the bleach solution was added. Eggs were extracted by vortexing for 30 seconds and shaking for 1 minute. This was done a total of 3 times and worms were vortexed one last time. Then the eggs were spun down at 1900 rpm for 1 minute and excess bleach solution was removed, and the eggs were washed 3 times with M9 minimal medium. RNAi treatment RNAi containing NGM plates were prepared as described in the “Bacterial strains” section. For double RNAi treatments, RNAi cultures were mixed at a 1:1 ratio by volume. HT115 cells transformed with an empty pL4440 vector was used as a negative control. RNAi conditions used in this study and tests for their efficacy is described below: cdk-9 RNAi. L1 worms were plated on HT115 food plates for the first 48 hours and were then moved to plates containing cdk-9 RNAi for the remaining 24 hours. Embryonic lethality in the range of 80%—85% was observed. top-2 RNAi. L1 worms were plated on HT115 food plates for the first 24 hours and were then moved to plates seeded with top-2 RNAi for the remaining 48 hours. Embryonic lethality was observed at >90%. capg-2 RNAi. Worms were grown on HT115 food plates for the first 24 hours and were moved to plates containing capg-2 RNAi for the remaining 48 hours. An embryonic lethality of 80%-100% was seen with this RNAi treatment. top-2/capg-2 double RNAi. Worms were grown on HT115 food plates for the first 24 hours and were transferred to top-2/capg-2 double RNAi plates for the next 48 hours. Embryonic lethality ranged from 90%-100% for this RNAi treatment. cdk-1 RNAi. Worms were grown on HT115 food plates for the first 24 hours and were transferred to cdk-1 RNAi plates for the next 48 hours. Embryonic lethality ranged from 97%-100% for this RNAi treatment. wee-1.3 RNAi. Worms were grown on plates containing wee-1.3 RNAi for the entirety of their life cycle. An embryonic lethality of approximately 40% was observed. Additionally, a significant reduction in brood size, and the coalescence of bivalents into one chromatin mass in proximal oocytes of some samples, were observed for wee-1.3 RNAi worms, as previously reported [34]. wee-1.3/capg-2 double RNAi. Worms were grown on plates containing wee-1.3 RNAi for 24 hours then were moved to plates containing wee-1.3/capg-2 RNAi where they remained for the rest of their life cycle. An embryonic lethality of approximately80%, and the coalescence of bivalents in proximal oocytes of some samples were observed. met-2 RNAi. Worms were grown on plates containing met-2 RNAi for the entirety of their life cycle. Some of the adult worms were bleached and an L1 chromatin compaction assay was performed on the resulting larvae to test RNAi efficacy. See Belew et al., 2021, for details on the L1 compaction assay. set-25 RNAi. Worms were grown on plates containing set-25 RNAi for the entirety of their life cycle. RNAi efficacy was tested via the same method as for met-2 RNAi. pie-1 RNAi. Worms were grown on pie-1 RNAi plates for the entirety of their life cycle. An embryonic lethality of 100% was observed for this RNAi. cdk-9 RNAi. L1 worms were plated on HT115 food plates for the first 48 hours and were then moved to plates containing cdk-9 RNAi for the remaining 24 hours. Embryonic lethality in the range of 80%—85% was observed. top-2 RNAi. L1 worms were plated on HT115 food plates for the first 24 hours and were then moved to plates seeded with top-2 RNAi for the remaining 48 hours. Embryonic lethality was observed at >90%. capg-2 RNAi. Worms were grown on HT115 food plates for the first 24 hours and were moved to plates containing capg-2 RNAi for the remaining 48 hours. An embryonic lethality of 80%-100% was seen with this RNAi treatment. top-2/capg-2 double RNAi. Worms were grown on HT115 food plates for the first 24 hours and were transferred to top-2/capg-2 double RNAi plates for the next 48 hours. Embryonic lethality ranged from 90%-100% for this RNAi treatment. cdk-1 RNAi. Worms were grown on HT115 food plates for the first 24 hours and were transferred to cdk-1 RNAi plates for the next 48 hours. Embryonic lethality ranged from 97%-100% for this RNAi treatment. wee-1.3 RNAi. Worms were grown on plates containing wee-1.3 RNAi for the entirety of their life cycle. An embryonic lethality of approximately 40% was observed. Additionally, a significant reduction in brood size, and the coalescence of bivalents into one chromatin mass in proximal oocytes of some samples, were observed for wee-1.3 RNAi worms, as previously reported [34]. wee-1.3/capg-2 double RNAi. Worms were grown on plates containing wee-1.3 RNAi for 24 hours then were moved to plates containing wee-1.3/capg-2 RNAi where they remained for the rest of their life cycle. An embryonic lethality of approximately80%, and the coalescence of bivalents in proximal oocytes of some samples were observed. met-2 RNAi. Worms were grown on plates containing met-2 RNAi for the entirety of their life cycle. Some of the adult worms were bleached and an L1 chromatin compaction assay was performed on the resulting larvae to test RNAi efficacy. See Belew et al., 2021, for details on the L1 compaction assay. set-25 RNAi. Worms were grown on plates containing set-25 RNAi for the entirety of their life cycle. RNAi efficacy was tested via the same method as for met-2 RNAi. pie-1 RNAi. Worms were grown on pie-1 RNAi plates for the entirety of their life cycle. An embryonic lethality of 100% was observed for this RNAi. Antibodies and dilutions RNAPIIpSer2: Rabbit antibody from Abcam (ab5095, Waltham, Massachusetts) was used at a dilution of 1:100. H3pSer10: Rabbit antibody from Rockland Immunochemicals (600-401-I74, Pottstown, Pennsylvania) was used at a dilution of 1:500. H3K9me3: Rabbit antibody from Abcam (ab176916, Waltham, Massachusetts) was used at a dilution of 1:1000. MPM-2: Mouse antibody (isotype—IgG1) from Sigma-Aldrich (05–368, St. Louis, Missouri) was used at a dilution of 1:500. Secondary antibodies: Alexa Fluor conjugated secondary antibodies from Invitrogen (Thermo Fisher Scientific, Waltham, Massachusetts) were used at a dilution of 1:200. Immunofluorescence staining Adult worms were first washed off plates with 10 ml of M9 minimal medium and rinsed 3 more times. Then, they were centrifuged at 1.9 KRPM and the excess medium was removed. 20 μl of media containing about 50 worms were spotted on a coverslip and 3 μl of anesthetic (20mM Sodium Azide and 0.8M Tetramisole hydrochloride) was added to immobilize them. Worms were dissected using 25Gx5/8 needles (Sigma Aldrich, St. Louis, Missouri). To release gonads, adult worms were cut twice, once right below their pharyngeal bulb and once near the tail. The coverslip was then mounted onto poly-L-lysine covered slides and let rest for 5 minutes. Slides were put on dry ice for 30 minutes. Samples were then freeze-cracked by flicking the coverslips off for permeabilization. For RNAPIIpSer2, H3pSer10 and MPM-2 antibody staining experiments, once samples were permeabilized, slides were put in cold 100% methanol (-20°C) for 2 minutes and then fixing solution (0.08M HEPES pH 6.9, 1.6mM MgSO4, 0.8mM EGTA, 3.7% formaldehyde, 1X phosphate-buffered saline) for another 30 minutes. After fixing, slides were washed three times with TBS-T (TBS with 0.1% Tween-20) and were blocked for 30 minutes with TNB (containing 100mM Tris-HCl, 200 mM NaCl, and 1% BSA). Primary antibodies were then applied at the dilutions described above in TNB and slides were incubated at 4°C overnight. For H3K9me3 staining experiments, permeabilized samples were put in cold 100% methanol (-20°C) for 10 seconds and then fixing solution (0.08M HEPES pH 6.9, 1.6mM MgSO4, 0.8mM EGTA, 3.7% formaldehyde, 1X phosphate-buffered saline) for 10 minutes. After fixing, slides were washed three times with TBS-T (TBS with 0.1% Tween-20) and were blocked for 2 hours with TNB (containing 100mM Tris-HCl, 200 mM NaCl, and 1% BSA) supplemented with 10% normal goat serum. Primary antibodies were then applied at the dilutions described above in TNB supplemented with 10% goat serum and slides were incubated at 4°C overnight. On the next day, the slides were washed 3 times with TBS and slides were incubated with secondary antibodies and Hoechst 33342 dye for 2 hours at room temperature. Slides were washed 3 times with TBS, mounting medium (50% glycerol in PBS), and coverslips were applied and sealed with Cytoseal XYL (Thermo Fisher Scientific, Waltham Massachusetts). Live animal imaging Adult N2, WMM1 and EGW83 worms were collected off plates and were washed 3 times with 10 ml M9 minimal medium. After the last wash, worms were spun down at 1.9 KRPM and excess medium was removed. 0.3% agarose pads were made on slides, and a 10 μl aliquot of adult worms was mounted. 4 μl of anesthetic (20mM Sodium Azide and 0.8M Tetramisole hydrochloride) was added to stop the worms from moving. A coverslip was gently applied, and the slides were imaged. Immunofluorescent imaging All slides were imaged using an Olympus Fluoview FV1000 confocal microscope using Fluoview Viewer software at a magnification of 600x (60x objective and 10x eyepiece magnifications). Laser intensity was controlled for experiments to achieve consistency among samples. Quantification of data RNAPIIpSer2 signal quantification. For each oocyte nucleus, two images were taken: one of the Hoechst-stained DNA and one of the RNAPIIpSer2 signal. Images were analyzed using ImageJ. An outline was drawn using the polygon selection tool around the Hoechst-stained area to mark the space occupied by DNA. The region of interest was copied and pasted to the RNAPIIpSer2 signal image and the raw integrated density (the sum of the values of the pixels in the selection) was measured. The raw integrated density was then normalized by the area to get a measure we called “signal density”. To account for possible variability in signal intensity due to different degrees of antibody penetration from sample to sample, each oocyte’s signal density was normalized to an average signal density from 5 pachytene nuclei found in the same image. The final normalized signal densities for the oocytes are presented. For each condition, were collected from two independently performed experimental replicates and the data were then pooled for statistical analysis and presentation. H3K9me3 signal quantification. Signal densities for each oocyte were calculated similarly to what is described above in the RNAPIIpSer2 signal quantification. For data on Fig 5, H3K9me3 signal of each oocyte in the -2 to -5 positions was normalized to the most proximal (-1) oocyte and the ratios are presented as percentages. For data presented in S4 Fig signal densities of control and top-2/capg-2 RNAi samples were shown side by side for comparison. H3pSer10 signal quantification. Signal densities for each oocyte were calculated similarly to what is described above in the RNAPIIpSer2 signal quantification. Signal densities were not normalized to pachytene nuclei since pachytene nuclei either did not harbor any signal. (for H3pSer10 staining) or the signals were affected by our RNAi treatments (for H3K9me3 staining) which precluded us from using them for unbiased normalization. Bivalent volume quantification. Z-stacks of oocyte nuclei were acquired. For each condition, 5 nuclei were analyzed and every bivalent within those nuclei that were visually distinguishable from one another was included in our analysis. On average 4 bivalents per nucleus were analyzed using ImageJ. The scale was set to 69nm per pixel. To measure the volume of a bivalent, a polygon was tightly drawn around it on each stack the bivalent appears in. The areas of these polygons were measured and summed up. Finally, the sum was multiplied by the distance between each stack to calculate an approximation of the bivalent’s volume. Averages of the volumes of all bivalents analyzed were presented. Nucleolar size measurement. Images were captured of focal planes corresponding to the maximum diameter of the nucleolus in living oocytes for each oocyte position. Diameters were measured using ImageJ and then converted to area. RNAPIIpSer2 signal quantification. For each oocyte nucleus, two images were taken: one of the Hoechst-stained DNA and one of the RNAPIIpSer2 signal. Images were analyzed using ImageJ. An outline was drawn using the polygon selection tool around the Hoechst-stained area to mark the space occupied by DNA. The region of interest was copied and pasted to the RNAPIIpSer2 signal image and the raw integrated density (the sum of the values of the pixels in the selection) was measured. The raw integrated density was then normalized by the area to get a measure we called “signal density”. To account for possible variability in signal intensity due to different degrees of antibody penetration from sample to sample, each oocyte’s signal density was normalized to an average signal density from 5 pachytene nuclei found in the same image. The final normalized signal densities for the oocytes are presented. For each condition, were collected from two independently performed experimental replicates and the data were then pooled for statistical analysis and presentation. H3K9me3 signal quantification. Signal densities for each oocyte were calculated similarly to what is described above in the RNAPIIpSer2 signal quantification. For data on Fig 5, H3K9me3 signal of each oocyte in the -2 to -5 positions was normalized to the most proximal (-1) oocyte and the ratios are presented as percentages. For data presented in S4 Fig signal densities of control and top-2/capg-2 RNAi samples were shown side by side for comparison. H3pSer10 signal quantification. Signal densities for each oocyte were calculated similarly to what is described above in the RNAPIIpSer2 signal quantification. Signal densities were not normalized to pachytene nuclei since pachytene nuclei either did not harbor any signal. (for H3pSer10 staining) or the signals were affected by our RNAi treatments (for H3K9me3 staining) which precluded us from using them for unbiased normalization. Bivalent volume quantification. Z-stacks of oocyte nuclei were acquired. For each condition, 5 nuclei were analyzed and every bivalent within those nuclei that were visually distinguishable from one another was included in our analysis. On average 4 bivalents per nucleus were analyzed using ImageJ. The scale was set to 69nm per pixel. To measure the volume of a bivalent, a polygon was tightly drawn around it on each stack the bivalent appears in. The areas of these polygons were measured and summed up. Finally, the sum was multiplied by the distance between each stack to calculate an approximation of the bivalent’s volume. Averages of the volumes of all bivalents analyzed were presented. Nucleolar size measurement. Images were captured of focal planes corresponding to the maximum diameter of the nucleolus in living oocytes for each oocyte position. Diameters were measured using ImageJ and then converted to area. Statistical analysis Prior to performing any statistical test, data was tested whether it was parametric or not. To do so, the Shapiro-Wilk test was used to test for normal distribution and F-test was used to test for variance homogeneity of the datasets we were comparing. Data were then analyzed using a student’s t-test or Wilcoxon Rank Sum test depending on whether the datasets fulfill the requirements for a parametric test or not. Differences between any two datasets were considered statistically significant if a P-value of <0.05 was obtained. Supporting information S1 Fig. All patterns of RNAPIIpSer2 signal in C. elegans proximal oocytes are dependent on CDK-9. A. The different patterns of RNAPIIpSer2 (red) signal on and/or off DNA (blue) observed in N2s treated either control or cdk-9 RNAi. Depletion of CDK-9 results in the loss of RNAPIIpSer2 signal in proximal oocytes regardless of RNAPIIpSer2 signal localization. Scale bar represents a length of 2 μm. B. Visualization of data presented in (A). N2s treated with cdk-9 RNAi showed reduced RNAPIIpSer2 signal in all forms. The number of samples analyzed over 2 independent replicates is presented below the charts for each oocyte position. https://doi.org/10.1371/journal.pgen.1010831.s001 (TIFF) S2 Fig. TOP-2 and condensin II mediated transcriptional repression in oocytes is independent of cell-cycle timing. A. N2 animals were treated with control or top-2/capg-2 double RNAi. Gonads from young adults were dissected, fixed, and stained for DNA (blue) and H3pSer10 (green). Exposure to top-2/capg-2 RNAi does not affect the timing of H3pSer10. Scale bar represents a length of 2 μm. B. Quantification of data presented in (A). 11 control RNAi and 10 top-2/capg-2 RNAi samples were analyzed for each RNAi treatment. There was no significant difference in H3pSer10 signal after TOP-2 and CAPG-2 co-depletion. Signal density was not normalized. https://doi.org/10.1371/journal.pgen.1010831.s002 (TIFF) S3 Fig. CDK-1 levels gradually increase as oocytes become more proximal and are required for NEB at -1 oocytes. A. N2 gonads were dissected, fixed, and stained for DNA (blue) and MPM-2 (white). An increase in MPM-2 signal is observed at the proximal oocyte positions in comparison to more distal oocytes. Scale bar represents a length of 2 μm. B. N2 samples were treated with either control or cdk-1 RNAi. Nuclear membrane of -1 oocytes were evaluated using live phase-contrast microscopy. 20 samples were analyzed over two replicates and quantifications are presented below each representative image. cdk-1 RNAi resulted in a higher number of samples with intact nuclear envelope when compared to control RNAi treatment. C. N2 samples were treated with either control or wee-1.3 RNAi. Nuclear envelope integrity was evaluated like in (B) for oocytes in -1 to -6 positions. 5 gonads were analyzed for each treatment and the quantifications are presented below each representative image. Treatment with wee-1.3 RNAi did not alter the progression of NEB in proximal oocytes. https://doi.org/10.1371/journal.pgen.1010831.s003 (TIFF) S4 Fig. Depletion of TOP-2 and condensin II does not alter H3K9me3 deposition in proximal gonads. A. Gonads from N2 adults treated with either control or top-2/capg-2 RNAi were dissected, fixed, and stained for DNA (blue) and H3K9me3 (red). Treatment with top-2/capg-2 RNAi did not alter the deposition of H3K9me3 in proximal gonads. Scale bar represents a length of 2 μm. B. Quantification of signal density for the data presented in (A). 10 samples were analyzed for each treatment. H3K9me3 signal remained the same when top-2/capg-2 RNAi treated samples were compared with those treated with control RNAi. https://doi.org/10.1371/journal.pgen.1010831.s004 (TIFF) S5 Fig. Representative images for bivalent volume measurements. Z-stacks were taken from living oocytes at the -2 position in animals treated with control, top-2/capg-2, met-2, set-25, and pie-1 RNAi. Shown here are representative images of a bivalent volume measurement from each RNAi treatment. See Methods for details on bivalent volume measurement. Scale bar represents a length of 2 μm. https://doi.org/10.1371/journal.pgen.1010831.s005 (TIFF) S1 Table. Raw data that underlie all graphs presented in this study. https://doi.org/10.1371/journal.pgen.1010831.s006 (XLSX) Acknowledgments We are grateful to Erik Griffin and Craig Mello for the kind gifts of worm strains.