The human cortex possesses a reconfigurable dynamic network architecture that is disrupted in psychosis

Jenna Reinen; Oliver Chén; R. Hutchison; B. Yeo; Kevin Anderson; Mert Sabuncu; Dost Öngür; Joshua Roffman; Jordan Smoller; Justin Baker; Avram Holmes

doi:10.1038/s41467-018-03462-y

The human cortex possesses a reconfigurable dynamic network architecture that is disrupted in psychosis

Reinen, Jenna; Chén, Oliver; Hutchison, R.; Yeo, B.; Anderson, Kevin; Sabuncu, Mert; Öngür, Dost; Roffman, Joshua; Smoller, Jordan; Baker, Justin; Holmes, Avram 2018-03-20 00:00:00 ARTICLE DOI: 10.1038/s41467-018-03462-y OPEN The human cortex possesses a reconﬁgurable dynamic network architecture that is disrupted in psychosis 1 1 2 3,4 1 Jenna M. Reinen , Oliver Y. Chén , R. Matthew Hutchison , B.T.Thomas Yeo , Kevin M. Anderson , 4,5 6 7,8 7 6 Mert R. Sabuncu , Dost Öngür , Joshua L. Roffman , Jordan W. Smoller , Justin T. Baker & 1,4,7,9 Avram J. Holmes Higher-order cognition emerges through the ﬂexible interactions of large-scale brain net- works, an aspect of temporal coordination that may be impaired in psychosis. Here, we map the dynamic functional architecture of the cerebral cortex in healthy young adults, leveraging this atlas of transient network conﬁgurations (states), to identify state- and network-speciﬁc disruptions in patients with schizophrenia and psychotic bipolar disorder. We demonstrate that dynamic connectivity proﬁles are reliable within participants, and can act as a ﬁngerprint, identifying speciﬁc individuals within a larger group. Patients with psychotic illness exhibit intermittent disruptions within cortical networks previously associated with the disease, and the individual connectivity proﬁles within speciﬁc brain states predict the presence of active psychotic symptoms. Taken together, these results provide evidence for a reconﬁgurable dynamic architecture in the general population and suggest that prior reports of network disruptions in psychosis may reﬂect symptom-relevant transient abnormalities, rather than a time-invariant global deﬁcit. 1 2 Department of Psychology, Yale University, New Haven, CT 06520, USA. Department of Psychology, Harvard University, Cambridge, MA 02138, USA. Department of Electrical & Computer Engineering, Clinical Imaging Research Centre, Singapore Institute for Neurotechnology & Memory Network Programme, National University of Singapore, Singapore 117583, Singapore. Athinoula A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital Harvard Medical School Charlestown, MA 02129, USA. School of Electrical and Computer Engineering and Meinig School of Biomedical Engineering, Cornell University, Ithaca, NY 14853, USA. Department of Psychiatry, Psychotic Disorders Division, McLean Hospital, Belmont, MA 02478, 7 8 USA. Department of Psychiatry, Massachusetts General Hospital, Harvard Medical School, Boston, MA 02114, USA. Psychiatric Neuroimaging Research Division, Massachusetts General Hospital, Charlestown, MA 02129, USA. Department of Psychiatry, Yale University, New Haven, CT 06511, USA. Correspondence and requests for materials should be addressed to A.J.H. (email: [email protected]) NATURE COMMUNICATIONS (2018) 9:1157 DOI: 10.1038/s41467-018-03462-y www.nature.com/naturecommunications 1 | | | 1234567890():,; ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-03462-y he human cortex is organized into large-scale networks with schizophrenia exhibit reduced dynamism, spending longer peri- 1–3 complex patterns of functional coupling .While sub- ods of time within single brain states and demonstrating a less- Tstantial progress has been made delineating aspects of this variable repertoire of dynamic network conﬁgurations . Relative intricate architecture, research in this domain has traditionally to healthy populations, patients with schizophrenia dwell more in relied on static analytic approaches that assume stable patterns of network conﬁgurations typiﬁed by reduced large-scale con- connectivity across time . However, the brain is not a static organ, nectivity, while also showing muted cross-network negative cor- and time-varying proﬁles of network connectivity are evident across relations, for instance, between default and other networks . 5,6 a broad range of task states and during periods of unconstrained Dynamic analyses of network function may distinguish clinical 7–9 10,11 rest ,(butsee ). Variability in the expression of dynamic groups, providing information that is inaccessible through static 12–15 17 brain states links to cognition , learning, and the presence of connectivity analysis . The incorporation of time-resolved ana- 16–18 psychiatric illnesses like schizophrenia that are characterized lyses of network function could provide a more sensitive or by a breakdown in cortical information processing. Despite the speciﬁc marker of the disease than static approaches, potentially importance of understanding the relations that link temporal associating with the illness course and/or the presence of distinct descriptions of brain function with behavior, the core features of symptom proﬁles. As impairments in attention, learning, and dynamic network organization, the stability of individually speciﬁc executive functioning are common across neuropsychiatric dis- signatures of time-resolved connectivity, and their associated rele- orders , research in this domain could provide novel insights vance to the disease, remain unclear. into the biology of illness as we work to predict the onset, track Patterns of spontaneous brain activity and connectivity have disease states, and optimize treatment response. been a focused topic of study in electrophysiological recordings at Here, we applied a sliding-window approach to characterize the level of cells, local ﬁelds, and surface electroencephalograms the reconﬁgurable architecture of large-scale brain networks in a (EEGs) . While formal biophysical models linking the activity of sample of healthy young adults and individuals with psychotic neuronal populations with large-scale brain systems have yet to illness . Among healthy adults, we demonstrate that the resulting be established , features of oscillatory neural activity, such as dynamic connectivity proﬁles are reliably expressed across scans those observed at high temporal resolutions, may be reﬂected in and visits, acting as a biological signature that can identify speciﬁc hemodynamic ﬂuctuations . Transient quasi-stable patterns individuals within a larger group. Patients with psychotic illness detected in EEGs (microstates), for example, spatially correlate exhibited intermittent disruptions within cortical association with network patterns observed through intrinsic functional networks previously associated with the disease. Individual con- 21–23 connectivity magnetic resonance imaging . Although macro- nectivity proﬁles within speciﬁc brain states predicted the pre- scale network dynamics have been linked with changes in arou- sence of active psychotic symptoms, operationalized as meeting 13 14,15 24 sal , attention , and autonomic activity , the biological bases clinician-rated symptomatic diagnostic criteria with the presence and behavioral signiﬁcance of these spontaneous ﬂuctuations of delusions and/or hallucinations in the past month . This remain unresolved. There are at least two reasons for this lack of property of dynamic network function generalized to a held-out consensus. First, prior studies of functional network dynamics sample of patients. These collective results suggest a key role for have largely focused on single clustering solutions in isolation, functional network dynamics in human cognition, and highlight choosing a ﬁxed number of temporal states a priori. As a result, a how speciﬁc breakdowns in time-varying proﬁles of network functional atlas of transient network conﬁgurations, or brain connectivity may link with the presence of distinct symptom states, that are present throughout the population has not been proﬁles in psychiatric illnesses. fully characterized. Second, analyses of transient network func- tion have principally focused on establishing the existence of a general architecture of dynamic connectivity shared across the Results population. Static patterns of intrinsic connectivity are heri- Core dynamic network conﬁgurations. The human brain can 25,26 table and act as a trait-like ﬁngerprint that can accurately exhibit a multitude of possible transient connectivity patterns 27–29 identify speciﬁc people from a large group . There is a reason comprised of varying network conﬁgurations. As an initial step in to believe that a substantial portion of the dynamic connectome understanding temporal shifts in this dynamic architecture over may be unique to each individual. Despite the importance of time, we aimed to identify a core or canonical set of transient establishing if time-resolved network function acts as an indivi- brain states conserved across individuals. First, we coupled a dual speciﬁc signature, the extent to which dynamic connectivity population atlas of large-scale cortical networks and a sliding- 7,8,40–42 proﬁles possess intra-subject reliability and capture inter-subject window approach (11 time points; width = 33 s ) to esti- variability has yet to be determined. mate time-varying connectivity proﬁles within resting-state scans Although time-resolved analyses of network organization have from 1919 healthy young adults (Fig. 1; see Methods for infor- largely focused on the study of healthy populations, there is mation on data acquisition and preparation) . These data were preliminary evidence to suggest that network dysfunction in collapsed across participants before we applied k-means cluster- psychosis may emerge through alterations in the core dynamic ing to estimate solutions yielding from 2 to 20 brain states. Note 16–18 architecture of the brain . Psychotic illnesses (including that the brain states deﬁned through k-means clustering do not schizophrenia, schizoaffective disorder, and bipolar disorder with necessarily have sharp boundaries, cleanly separating them from psychotic features) are marked by broad disruptions across cor- other network conﬁgurations. Rather, k-means clustering identi- tical association networks, potentially contributing to widespread ﬁes sets of time-varying network conﬁgurations with common 30–34 changes in information processing . By one view, impaired features, grouping them into clusters that are more similar to each network connectivity in patient populations might be time- other than to conﬁgurations in other clusters. The associated invariant, emerging through stable deﬁcits in brain function. An groupings can vary across clustering solutions as their complexity alternate possibility is that aspects of the functional impairments increases, revealing intermediate network conﬁgurations between observed in psychosis reﬂect transient abnormalities pre- more geographically separated states. To estimate the viability of ferentially evident during the expression of particular network the resulting solutions, we analyzed the population-level con- 17,35 39,44,45 conﬁgurations . Converging evidence suggests that aberrant sistency of each clustering algorithm . Consistent with an oscillatory activity may link to core symptoms of schizophrenia, expansion in the solution space, the clustering became less stable including the presence of hallucinations . Patients with as the number of estimated brain states increased (Fig. 1d and 2 NATURE COMMUNICATIONS (2018) 9:1157 DOI: 10.1038/s41467-018-03462-y www.nature.com/naturecommunications | | | NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-03462-y ARTICLE Supplementary Figure 1). Analyses indicated relative stability in a broad survey of the solution space. In the present data, these state solutions 2–8, with 2-, 4-, 5-, and 8-state solutions displaying solutions captured signiﬁcant aspects of dynamic variation in points of increased stability. The stable nature of the observed network connectivity. However, the focus on 2–8 brain states state solutions was robust to changes in data quality (see Meth- should not be taken to imply that meaningful properties are ods, Noise constraints). As such, our analyses going forward focus absent in alternative solutions. on clustering solutions containing 2–8 dynamic states to provide Brain states exhibit hierarchical features. Previous studies of Dorsal functional network dynamics have largely focused on single clustering solutions in isolation, choosing a ﬁxed number of states AP a priori. The relations between these isolated dynamic network Lateral conﬁgurations and other possible solutions remain unclear. To D address this question, we employed a matching analysis exam- Frontal Parietal ining the preservation or fractionation of individual brain states as the complexity of our clustering solution increased. An exhaustive search was performed exploring the scenario that two Medial states in solution S + 1 were subdivisions of a state within solu- tion S. Hungarian matching (Supplementary Figure 2) was used to determine which two-state combination in S + 1 was best matched to the S-state solution by minimizing network dissim- Ventral PA ilarity. As solution complexity increased, a hierarchical structure was observed, with some network conﬁgurations preserved across levels (Fig. 2; see Methods, Hierarchy analysis) and selected hybrid states in solution S breaking into substates within solution S + 1. A single state, termed state A, was evident across state solutions 2 through 7, not splitting to produce a hybrid substate Participant until the solution complexity increased to eight states. These time courses analyses demonstrate the stable expression of canonical network conﬁgurations across a variety of state solutions (2–8). Of note, when larger numbers of states were considered, increasing solu- tion complexity was reﬂected in the hierarchical fractionation of Windowed correlation particular states into substates. matrices Static analyses of network function suggest that heteromodal association cortices are more functionally variable than the unimodal cortex across the population . These aspects of the cortex, and its associated networks, are implicated in a host of Aggregrate complex cognitive functions and overlap with regions that data across participants predict individual differences in behavioral performance .We then explored if the heteromodal association cortex also exhibits heightened dynamic variability, as reﬂected in ﬂuctuating k=2 connectivity conﬁgurations over time. Analyses examining the variance of mean network connectivity across the state solutions K-means revealed evidence for dissimilar patterns of network expression clustering across the dynamic states (ANOVA of coefﬁcient of variance with k=3 k=4 k=5 Fig. 1 Detecting multiple functional connectivity states using a sliding- window approach. a The functional network organization of the human cerebral cortex is revealed through intrinsic functional connectivity. Colors reﬂect regions estimated to be within the same network determined based on the 17-network solution from Yeo et al. . The map is displayed for multiple views of the left hemisphere in Caret PALS space . b Correlation matrices are computed across regions from windowed portions (width= 33 s) of each participant’s component time series (n = 1919), and aggregated across the full sample. c K-means clustering was applied to Stability analysis identify repeated patterns of connectivity (brain states). d Instability of the 0.6 clustering algorithm is plotted as a function of the number of estimated 0.5 states (2–20). As expected based on increasing solution space (complexity), instability was greater with increasing the number of 0.4 estimated states per solution. The local minima of the graphs observable at 0.3 the 2-, 4-, 5-, and 8-state solutions indicate the number of states that can 0.2 be stably estimated by the selected clustering algorithm with the present data. Resampling over time (sliding windows) and across space (regions of 0.1 interest) yields comparable results (see Supplementary Figure 1). In this 0.0 article, we focus on state solutions 2 through 8 to provide a broad survey of 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 the solution space Number of States NATURE COMMUNICATIONS (2018) 9:1157 DOI: 10.1038/s41467-018-03462-y www.nature.com/naturecommunications 3 | | | Instability Sal/ Limbic DorsAttn Visual Default Control VentAttn SomMot –0.8 0.8 Z (r) ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-03462-y Brain states exhibit hierarchical features Global state 2 2 A B Two states 3 3 B C Three states 4 4 4 B C D Four states 5 5 5 5 B C D E Five states 6 6 6 6 6 6 A B C D E F Six states 7 7 7 7 7 7 7 C D E F G A B Seven states 8 8 8 8 8 8 8 8 A B C D E F G H Eight states Fig. 2 Brain states exhibit hierarchical features across distributed networks. Correlation matrices for state solutions 1 through 8 are shown for each region . To quantify hierarchical relationships, an exhaustive search was performed to examine the scenario that two states of the (S + 1) state solution were subdivisions of a component of the S-state solution. Hungarian matching was used to determine which two-state combination in S + 1 was best matched to the S-state solution (see Supplementary Figure 4). Values reﬂect z-transformed Pearson correlations between every region and every other region. DorsAttn indicates dorsal attention, Sal salience, SomMot somatomoto, VentAttn ventral attention state solutions treated as repeated measures, F = 44.11, the clustering solutions, the observed brain states exhibited quasi- p ≤ 0.001). Consistent with reports of heightened time-resolved stable expression, with all states maintaining nonzero dwell times network ﬂexibility within the association cortex , the greatest (ps ≤ 0.001). Reﬂecting the transition probability analyses detailed between-state variability was found in aspects of default, above, dwell times were nonuniform across brain states. Partici- attention, and control networks (Supplementary Figure 3). pants dwelled most in attractor state A (ps ≤ 0.001) relative to all other states. Several core patterns that distinguish brain states across Evidence for an attractor state. In addition to characterizing clustering solutions were evident. For ease of interpretability, the proﬁles of network connectivity across clustering solutions, the four-state solution was selected to characterize these features brain dynamics can be studied in terms of the relative likelihood of dynamic connectivity (Fig. 4). As noted in the Hungarian of transitions occurring among locally stable states over matching analyses detailed above (Fig. 2), state A most closely 7,16,17 time . To assess this feature of temporal organization, we resembled the global state revealed through traditional static examined the probability that participants would shift from a analyses of network function. Across network conﬁgurations, given state S to a different state (transition probability), as well as state A was typiﬁed by a relatively ﬂattened proﬁle of connectivity the probability that they would remain in state S. Window-by- (Fig. 4a). In line with this muted connectivity proﬁle, the window estimates were created for each state solution by expression of state A became increasingly frequent in the second matching the participant-speciﬁc connectivity matrices to the relative to the ﬁrst half of the scan across each clustering solution group atlas of brain states. This generated a vector of 110 (2–8 states; ts ≥ −6.07, ps ≤ 0.001), potentially linking with 14,15 expressed states for each participant. As solution complexity shifts in arousal and vigilance . increased, participants were more likely to transition into (ps ≤ The remaining states varied markedly from the state A 0.001), and remain in state A (all test states 2–5 ps ≤ 0.001; states connectivity pattern with differences evident across states both 6–8 ps ≥ 0.05; Fig. 3), termed the “attractor state.” Dwell time was within and between functional networks (Fig. 4b). Relative to 4 4 4 4 calculated as the percent of the total time a given participant state A , states B ,C , and D were characterized by increased expressed state S relative to the total time in states not-S. Across expression of default A (see Fig. 1a for network topographies), 4 NATURE COMMUNICATIONS (2018) 9:1157 DOI: 10.1038/s41467-018-03462-y www.nature.com/naturecommunications | | | Visual DorsAttn Sal/ Limbic SomMot VentAttn Control Default NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-03462-y ARTICLE State S+1 3 3 3 4 4 4 4 5 5 5 5 5 6 6 6 6 6 6 7 7 7 7 7 7 7 8 8 8 8 8 8 8 8 A B C A B C D A B C D E A B C D E F A B C D E F G A B C D E F G H 3 4 5 6 7 8 A A A A A A 3 4 5 6 7 8 B B B B B B 3 4 5 6 7 8 C C C C C C 4 5 6 7 8 D D D D D 5 6 7 8 E E E E 6 7 8 Transition probability F F F 7 8 0 0.45 G G 1.00 0.92 0.84 0.76 0.68 0.60 2 2 3 3 3 4 4 4 4 5 5 5 5 5 6 6 6 6 6 6 7 7 7 7 7 7 7 8 8 8 8 8 8 8 8 A B A B C A B C D A B C D E A B C D E F A B C D E F G A B C D E F G H Two Three Four Five Six Seven Eight Number of states Fig. 3 Evidence for an attractor state. a When transitioning between states, participants display an increased probability of entering state A. The probability of transitioning to each state is shown for all nonredundant possible combinations across state solutions 3–8 (1, 2 are obligated). Each row sums to 1. b State A displays the increased probability of remaining in the same state for solutions 2–5(ps ≤ 0.001). The graph reﬂects the tendency for states in solutions 2–8 to show steady-state behavior, or the probability of remaining in a given state from time point t to time point t+1 with state B also marked by heightened within-network visual Network dynamics are a marker of individual differences. Static system and default B connectivity. State B notably differed from descriptions of brain functional organization act as a biological other network conﬁgurations in its recruitment of dorsal ﬁngerprint that can identify speciﬁc individuals within a larger 4 4 27–29 attention network A. Conversely, states C and D displayed group . We aimed to determine whether the observed increased correlations within the ventral attention network. There dynamic states were reliably expressed in a manner that would was also striking variation in the coupling of the frontoparietal allow us to characterize unique within-subject proﬁles of transient network, with state C primarily characterized by increased network organization. To this end, we ﬁrst assessed within- connectivity in the control B network. session reliability. For participants with two bold runs (n = 1341) 4 4 4 States B ,C , and D showed marked departures from the off- in the same scanning session, correlation matrices were con- diagonal coupling evident in state A and the global state (Fig. 2). catenated for each individual brain state, and a Pearson correla- The negative correlations between the default and somatomotor tion was generated for run 1 relative to run 2. T-tests were used to networks were attenuated in state B and were more evident in compare the distributions of the resulting r values within and 4 4 4 4 states C and D . While states B and C displayed heightened across individuals. For each state, we observed greater within- correlations linking default and control networks, state D participant similarity for same-day scans (rs ≥ 0.49) compared to exhibited increased connectivity between default and the salience, between participants (rs ≤ 0.23, 2–8 states; ps ≤ 0.001; see Meth- attention, and somatomotor systems. Additionally, state C , and ods, Individual identiﬁcation analyses; Supplementary Figure 4). to a lesser extent state D , displayed increased cross-network We then examined a cohort of participants with two scan visits connectivity between the attention, somatomotor, and visual collected on different days (≤6 months apart; mean = 63.35 ± networks. Together, these varying network conﬁgurations 48.10 days; n = 79). Analyses demonstrated consistent within- demonstrate nonrandom departures from connectivity patterns subject dynamic state expression across visits (Fig. 5a; ps ≤ 0.001). observed in the global state. There is a strong correspondence Suggesting relatively stable intra-subject reliability across time, between the structure of intrinsic and extrinsic (task-evoked/ the observed within-participant similarity of the expressed brain coactivation) networks of the human brain, suggesting that the states (Fig. 5a) did not vary as a function of the number of days topological characteristics of the brain at rest are closely linked to between participant visits (absolute value of rs ≤ 0.21; ps ≥ 0.09). 12,48 27 cognitive function . The transient expression of integrated Permutation tests were used to assess if dynamic connectivity network conﬁgurations may enable fast and accurate cognitive proﬁles act as a signature that can accurately identify individual task performance . Emerging evidence suggests that an indivi- participants from a larger group. Correlation matrices from visit 1 dual’s unique proﬁle of dynamic network connectivity could and visit 2 were iteratively shufﬂed across participant labels and provide novel insights into the study of behavioral variability examined relative to their correct pairing. Identiﬁcation was across both health and disease. In this regard, a crucial step is the considered correct if the true visit 1 and 2 pair were maximally characterization of intra-subject reliability and inter-subject similar to each other. Correct identiﬁcation was evident within variability of the observed proﬁles of time-varying brain participants relative to the shufﬂed list across all state solutions organization across the population. (ps ≤ 0.01; see Fig. 5b), suggesting that it is possible, with high NATURE COMMUNICATIONS (2018) 9:1157 DOI: 10.1038/s41467-018-03462-y www.nature.com/naturecommunications 5 | | | State S Stationary probability ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-03462-y Mean within network connectivity Default D Default C Visual cent 1.0 Default B Visual peri Default A SomMotor A State A State B SomMotor B Control C State C State D Dorsal attention A Control B Dorsal attention B Control A Ventral attention Limbic Salience Percent deviation from state A Default D Default C Visual cent 0.30 Default B Visual peri SomMotor A Default A 4 4 State A - State B 4 4 SomMotor B Control C State A - State C 4 4 State A - State D Dorsal attention A Control B Dorsal attention B Control A Ventral attention Limbic Salience Fig. 4 Proﬁles of within-network connectivity across the four-state solution. a The mean within-network connectivity is shown across each state of the four-state solution (n = 1919). Values reﬂect z-transformed Pearson correlations. b Percent deviation from state A . Values reﬂect the percent change in 4 4 4 mean network connectivity of states B –D , relative to state A accuracy, to identify an individual from a large group of participant. For simplicity, we only consider the four-state solu- participants solely on the basis of their dynamic connectivity tion, focusing on four states due to the stability of the associated proﬁle. Dwell time was also more similar within, relative to clustering solution (Supplementary Figure 1) and the relatively between, participants for scans on the same day and across visits high within-participant reliability (Fig. 5 and Supplementary (all ps ≤ 0.01). These ﬁndings provide the ﬁrst evidence to suggest Figure 4). Importantly, while we discuss the four-state solution, that, like static characterizations of intrinsic connectivity, an meaningful properties are likely present in other dynamic net- individual’s dynamic brain function represents a unique and work conﬁgurations. We matched each windowed participant- reliable biological signature, highlighting the potential use of speciﬁc matrix to the dynamic states established in our previous intrinsic network dynamics as a neural marker of individual analyses and calculated mean correlation matrices across both differences across both health and disease. groups. Following the correction for nuisance variables (motion, age, sex, handedness, and scanner bay), group differences were calculated for each of the four states for dwell times and network Preferential state and network disruptions in psychosis. connectivity [false-discovery rate (FDR), α ≤ 0.05]. Impairments in the integration and processing of information Altered intrinsic connectivity in patients with psychosis across large-scale distributed brain networks are thought to mark observed in traditional static analyses may reﬂect both impaired psychotic disorders (including schizophrenia, schizoaffective functions within a speciﬁc brain state and the reduced tendency 30,50,51 disorder, and psychotic bipolar disorder) . Converging to enter particular network conﬁgurations over time. Analyses of evidence suggests that these disruptions may reﬂect alterations in dwell times did not reveal group variability in state A (t = 8,16,17,35 the temporal dynamics of brain function . We next 1.12, p = 0.26; Fig. 6a). However, patients displayed increased examined the expression of time-varying network conﬁgurations dwell time in state B (t = −2.8, p ≤ 0.005) and nominally in a cohort of patients with psychosis (n = 179) and a demo- signiﬁcant increases in state D (t = −2.13, p ≤ 0.03). Notably, graphically and data-quality-matched comparison sample (n = patient dwell times were reduced in state C (Fig. 5; t = 2.98, 369). To evaluate the integrity of dynamic brain functions in these p ≤ 0.003), which is characterized by increased frontoparietal groups, we created windowed correlation matrices for each control network connectivity relative to other states (Fig. 4). 6 NATURE COMMUNICATIONS (2018) 9:1157 DOI: 10.1038/s41467-018-03462-y www.nature.com/naturecommunications | | | NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-03462-y ARTICLE 0.65 0.52 Group 0.39 Within Between 0.26 0.13 0.00 2 2 3 3 3 4 4 4 4 5 5 5 5 5 6 6 6 6 6 6 7 7 7 7 7 7 7 8 8 8 8 8 8 8 8 A B A B C A B C D A B C D E A B C D E F A B C D E F G A B C D E F G H Two Three Four Five Six Seven Eight Number of states 1.00 0.80 Match rank st 0.60 nd rd 0.40 th th 5 + 0.20 0.00 2 2 3 3 3 4 4 4 4 5 5 5 5 5 6 6 6 6 6 6 7 7 7 7 7 7 7 8 8 8 8 8 8 8 8 A B A B C A B C D A B C D E A B C D E F A B C D E F G A B C D E F G H Two Three Four Five Six Seven Eight Number of states Fig. 5 Brain states are reliably expressed and can be used to identify individual participants. a Mean rho values within and between subjects across visits for all stable dynamic states. Correlation matrices from participants who had at least two scans within 6 months of each other (n = 79) were concatenated, and a Pearson correlation was computed for scan 1 relative to scan 2. b Permutation and correlation analyses revealed that dynamic states are consistent within individuals over time. Identiﬁcation rank accuracy is shown for each state in a subgroup of participants scanned on two different days within 6 months (n = 79). Rank denotes the order of similarity of a participant’s visit 1 to their visit 2, with 1 being a perfect identify match Analyses of region-to-region correlation strength across the Select network conﬁgurations mark active psychotic symptoms. four-state solution yielded sparse group differences in states B Together, the analyses above suggest that time-varying network and D . Conversely, widespread group differences were observed conﬁgurations capture stable aspects of inter-subject variability 4 4 4 in states A and C (Fig. 6). Analyses of state A revealed reduced (ﬁngerprints), while also serving to mark the presence of psy- within-network correlations across the cortex, including in chiatric illness. These ﬁndings suggest that dynamic approaches default and frontoparietal networks. The largest-magnitude may provide critical information that can help predict the pre- differences in state C were localized to the frontoparietal control sence of distinct clinical proﬁles within individuals, revealing 18 4 16,17 network (Figs. 6 and 7). Replicating prior reports , states A and symptom-relevant features of the disease . To determine C were associated with less negative, or muted, correlations whether individual differences in time-varying proﬁles of con- linking the aspects of the association cortex to other networks in nectivity are relevant to clinical symptomatology, we investigated psychotic illness. the extent to which speciﬁc brain states can identify patients who 4 4 Follow-up analyses of state A and C revealed evidence for experienced active psychotic symptoms at the point of initial both state-general and state-preferential impairments in network assessment. To this end, we examined clinician reports of the connectivity (Fig. 7 and Supplementary Table 1). Within the current psychotic symptoms in our patient sample, assessed frontoparietal control network, patients exhibited distributed through the DSM-IV (SCID) clinician-rated presence of delusions 4 4 38 deﬁcits across states A (state A control B and C: ts ≥ 4.12, and/or hallucinations in the past month , and selected partici- 4 4 ps ≤ 0.001) and C (state C control B: t = 4.85, p ≤ 0.001; pants who expressed all four brain states (n = 130). Using a Fig. 8a–b). All other subnetworks failed to pass multiple- machine-learning-based framework, elastic net logistic regression, comparison correction (Bonferroni p ≤ 0.05; ps ≥ 0.01). Conver- we demonstrated that the presence of active psychotic symptoms sely, reduced default network connectivity in psychosis was can be predicted based solely on the dynamic connectivity proﬁle 4 4 observed in default B in state A (state A default B: ts = 6.08, of previously unseen individuals (Methods, Prediction of active 52,53 p ≤ 0.001; default A, C, and D: ts ≤ 1.69, ps ≥ 0.09; Fig. 8c–d). psychotic symptoms) . First, training a model in 91 partici- No group differences in the default network survived correction pants (70% of the available patient sample), we used the edge 4 4 4 for multiple comparison when considering state C (state C strength from the dynamic network conﬁgurations of states A , 4 4 4 default B: ts = 2.91, p ≥ 0.01; default A, C, and D: ts ≤ 1.31, B ,C , and D as candidate features upon which we conducted 535 535 ps ≥ 0.19). These analyses suggest that prior observations of variable selection. The extracted features were then used to pre- default and control network disruptions in psychosis may reﬂect dict the presence of active psychotic symptoms in individual temporally speciﬁc impairments, preferentially manifesting dur- participants, using leave-one-subject-out cross-validation. The ing the expression of transient conﬁgurations that recruit the ﬁtted model was prospectively applied to edge strength of the function of these networks. dynamic network conﬁguration from a held-out set of NATURE COMMUNICATIONS (2018) 9:1157 DOI: 10.1038/s41467-018-03462-y www.nature.com/naturecommunications 7 | | | Percent matching Mean correlation ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-03462-y 0.45 Healthy comparison Psychotic Illness 0.38 0.31 p ≤ 0.05 p ≤ 0.005 0.24 p ≤ 0.01 0.17 0.10 4 4 4 4 State A State B State C State D 4 4 4 4 A B C D Healthy comparison –0.80 0.80 Z(r) 4 4 4 4 A B C D Psychotic illness 4 4 4 4 A B C D Functional connectivity differences –0.15 0.15 Z(r) –Z(r) comparison patients Fig. 6 Psychotic illness links to state-speciﬁc reductions in connectivity. a Bar graphs display the percent of time healthy comparison participants and individuals with psychotic illness dwell in each state of the four-state solution. b–c The 2D grids (b–c) display the complete coupling architecture of the cerebral cortex measured at rest for each group of participants. d Differences were obtained by an analysis of variance of z-transformed Pearson correlation values after linear regression of the effects of age, sex, race, ethnicity, and handedness [r(z) – r(Z) ]. The bottom panel Comparison Patient group shows signiﬁcant group differences at false-discovery rate q ≤ 0.05 participants, and yielded a scalar probability measure for each D were 0.39 (p = 0.74), 0.50 (p = 0.43), and 0.56 (p = 0.26), individual, which constituted the predicted likelihood of active respectively. psychotic symptoms. Next, we demonstrated that predictive symptom models Consistent with an increased tendency to enter state B in derived from dynamic network conﬁgurations can generalize to patients, relative to other network conﬁgurations (Fig. 6a), state data from novel individuals, applying the predictive network B speciﬁcally predicted the presence of active psychotic models in the left-out patient sample (30% of the available patient symptoms in the training set (positive network predictive sample). Here, without further ﬁtting or modiﬁcation of the network: AUC = 0.80, p ≤ 0.001, permutations = 5000; initial models, we tested the neurological signatures identiﬁed in Fig. 9a). As a comparison, the AUC values using a positive the training set for the predication of active psychotic symptoms 4 4 4 network of states A ,C , and D were 0.60 (p = 0.16), 0.68 in a held-out group previously unseen subjects (n = 39). The (p = 0.05), and 0.52 (p = 0.37), respectively. A similar, negative-network models did not generalize to the held-out although subtler, predictive proﬁle was observed for the negative sample (AUCs < 0.66, ps > 0.13). The previously identiﬁed 4 4 networks in the training set where state B predicted the presence positive-network model for state B served as a generalizable of active psychotic symptoms (AUC = 0.724, p ≤ 0.05). The predictor of active psychotic symptoms in an independent sample 4 4 remaining negative network AUC values for states A ,C , and (AUC = 0.74, p ≤ 0.05, permutation n = 5000; Fig. 9b). As a 8 NATURE COMMUNICATIONS (2018) 9:1157 DOI: 10.1038/s41467-018-03462-y www.nature.com/naturecommunications | | | Percent dwell time NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-03462-y ARTICLE positive-network model for active psychosis in state B served as a State A generalizable predictor of PANSS positive-scale symptom severity Default D in the independent sample (AUC = 0.72, p ≤ 0.05). The observed Visual cent Default C effect was preferential to the state B network model. The 4 4 0.24 associated AUC values derived across dynamic states A ,C , and Visual peri Default B D for positive out-sample predication were 0.51 (p = 0.46), 0.66 (p = 0.10), and 0.33 (p= 0.90). As above, the negative-network SomMotor A Default A models did not generalize to the PANSS positive-scale scores in the held-out sample (AUCs < 0.67, ps > 0.09). Suggesting a degree of symptom speciﬁcity, the PANSS negative (positive network: SomMotor B Control C AUC = 0.41, p = 0.83; negative network: AUC = 0.52, p = 0.42) and general psychopathology scale (positive network: AUC = 0.62, p = 0.21; negative network: AUC = 0.41, p = 0.74) scores Dorsal attention A Control B were not predicted by the state B model, or the models resulting from the other states (positive network: AUCs < 0.51, ps > 0.48; Dorsal attention B Control A negative network: AUCs < 0.59, ps > 0.18). These results provide preliminary evidence that the presence of clinical symptoms, in Ventral attention Limbic this case, active psychosis, may be associated with predictable Salience patterns within an individual’s time-varying network proﬁle, suggesting that dynamic approaches have potential utility for State C predicting a wide range of clinical symptoms and cognitive Default D abilities. Visual cent Default C 0.24 Visual peri Default B Discussion The functional coupling of cortical regions varies in response to SomMotor A Default A explicit task demands and in conjunction with shifts in arou- 13 14,15 24 sal , attention , and markers of autonomic activity . Here, using a sliding-window approach on resting-state imaging data, SomMotor B Control C we demonstrated that cortical brain networks possess a time- varying organizational structure with hierarchical properties. Select network conﬁgurations fractionated into substates in an Dorsal attention A Control B ordered manner, and a global attractor state (termed state A) was evident across clustering solutions (2–8 states). Suggesting that Dorsal attention B Control A proﬁles of dynamic network connectivity may link to behavioral differences in health and disease, the observed brain states Ventral attention Limbic reﬂected individually speciﬁc signatures, or ﬁngerprints, of time- Salience resolved connectivity that were unique and reliable within par- Fig. 7 State-speciﬁc and state-general proﬁles of network dysregulation in ticipants across both scans and independent visits separated by up psychosis. a, b Polar plots show percent difference in mean network to 6 months. Extending upon prior static analyses of network 27–29 connectivity between healthy comparison participants and individuals with function , we found that it is possible, with high accuracy, to 4 4 psychotic illness for (a) state A and (b) state C (range: –6 to 24%). The identify speciﬁc individuals from a large group of participants black hexadecagon reﬂects the mean network correlations for the healthy solely on the basis of their proﬁles of dynamic connectivity. comparison sample, set to zero. Values outside the hexadecagon reﬂect Patients with schizophrenia and psychotic bipolar disorder decreased correlation strength for patients, relative to the healthy exhibited state-speciﬁc, intermittent disruptions within cortical comparison sample association networks believed to mark the presence of psychotic illness. Finally, our analyses revealed that individual variability in 4 4 comparison, the AUC values derived across dynamic states A , select network conﬁgurations (state B ) can be used to predict the 4 4 C , and D for out-sample predication were 0.56 (p = 0.32), 0.63 presence of active psychotic symptoms in novel participants. (p = 0.18), and 0.29 (p = 0.92). Highlighting that dynamic Together, these results highlight the potential to discover indi- analyses may reveal information hidden in traditional static vidualized dynamic network proﬁles that are predictive of cog- analyses of network function, models derived from the static or nitive abilities and clinical symptoms across health and disease. global state were not predictive of active psychosis in either the Spontaneous brain activity is constrained, but not fully deter- 55–59 training (positive network: AUC = 0.62, p = 0.12; negative net- mined, by structural connectivity . This raises the possibility work: AUC = 0.64, p = 0.09) or test samples (positive network: that a quasi-stable functional architecture may anchor on ana- AUC = 0.49, p = 0.53; negative network: AUC = 0.24, p = 0.96). tomic connectivity, with transient network conﬁgurations Patients expressing active psychotic symptoms, relative to reﬂecting the inﬂuence of momentary cognitive processes, those without, presented with increased positive (present: 19.90 ± environmental demands, or other biological information. Sup- 5.67, absent: 8.88 ± 2.20; t = 7.91, p ≤ 0.001), negative (present: porting this conjecture, while functional coupling occurs in the 13.98 ± 7.76, absent: 10.18 ± 3.93; t = 1.98, p ≤ 0.05), and absence of cognition , under general anesthesia, brain activity general psychopathology (present: 31.46 ± 7.81, absent: 26.53 ± exhibits a reduction in spontaneous transitions across network 7.74; t = 2.43, p ≤ 0.05) symptoms (Supplemental Fig. 7)as conﬁgurations , settling into a restricted dynamical repertoire assessed through the positive and negative syndrome scale that closely resembles a ﬁxed network deﬁned by structural 54 56 (PANSS) . To establish the speciﬁcity of the predictive model connectivity . The present analyses indicate the expression of for the presence of psychotic features, we examined PANSS scale core, or canonical, time-varying network conﬁgurations that scores in the left-out patient sample. The previously identiﬁed separate in an ordered manner across increasingly complex NATURE COMMUNICATIONS (2018) 9:1157 DOI: 10.1038/s41467-018-03462-y www.nature.com/naturecommunications 9 | | | ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-03462-y a c Default D Control C Default C Control B Default B Control A Default A b 4 4 d 4 4 State A State C State A State C Default D Control C Default C Control B Default B Control A Default A Healthy Healthy comparison comparison Psychotic Psychotic illness illness –0.8 0.8 –0.8 0.8 Z(r) Z(r) Functional Functional connectivity connectivity differences differences –0.15 0.15 –0.15 0.15 Z(r) –Z(r) Z(r) –Z(r) patients patients comparison comparison 4 4 Fig. 8 Psychosis is associated with reduced network connectivity in states A and C . a The colored aspects of the cortex reﬂect regions estimated to be within the A, B, and C aspects of the frontoparietal control network . b Functional connectivity matrices for the 50 left and right hemisphere regions of the 4 4 frontoparietal control network shown for the healthy comparison and patient groups in states A and C . Functional connectivity difference matrices were obtained by an analysis of variance of z-transformed Pearson correlation values after linear regression of the effects of age, sex, handedness, and scanner bay. Group differences signiﬁcant at false-discovery rate q ≤ 0.05 are shown in each panel to the lower right of the unthresholded matrix. c The colored aspects of the cortex reﬂect regions estimated to be within the A, B, C, and D aspects of the default network. d Corresponding connectivity matrices are 4 4 shown in 52 left and right hemisphere regions of the default regions for states A and C . Reduced default network connectivity in psychosis is preferential to state A clustering solutions. Throughout the hierarchy, a multistable shifting . While we are unable to make direct claims regarding dynamical system was evident, ﬂuctuating around a global the association between time-varying proﬁles of network function attractor state (state A) that possesses a relatively muted con- and cognition, converging evidence suggests that broad properties nectivity proﬁle. Decreased within-network coupling is evident of network connectivity may be preserved across experimental 12,48 when spontaneous neuronal activity adheres to ﬁxed correlation contexts (e.g., intrinsic or task-evoked) and analysis strate- 61,62 48 conﬁgurations deﬁned by structural connectivity . Although gies . An important area of future work will be to establish the speculative, our results suggest the presence of an attractor state extent to which the cortical network structure adjusts as a func- that could preferentially link to the large-scale anatomical struc- tion of task and/or environment. One speculative possibility is ture of the human cerebral cortex. Future cross-modal analyses that transient periods of strong within-network coupling may will be necessary to explore these hypotheses and test the extent correspond to epochs of high efﬁciency, a property of brain to which patterns of spontaneous brain activity might reconﬁgure organization theorized to optimize communication across distinct around an underlying anatomical skeleton. functional domains . Regions within the association cortex display increased func- Human cognition is a ﬂuctuating process, and there is growing tional ﬂexibility, potentially serving to integrate information evidence that functional connectivity patterns exhibit complex 63 58 across more specialized aspects of the cortex . This proﬁle of spatiotemporal dynamics at multiple time scales . Time-varying malleability is reﬂected in static analyses of intrinsic network brain states, such as those identiﬁed in the present analyses, have function where heightened population-level variability has been been hypothesized to reﬂect changes in the ongoing cognitive 46 7 observed in association relative to unimodal cortices . Consistent processes during rest . However, the existence, putative origins, with prior reports of marked heterogeneity in the dynamic ﬂex- and cognitive correlates of dynamics in resting-state fMRI remain 13,64 10,11,19,67 ibility of neural regions , we observed the greatest cross-state a topic of empirical debate . While the relations linking variability within aspects of default, attention, and control net- intrinsic time-varying network proﬁles with cognition remain works. The trade-off between control and attention systems is speculative, there are two commonly held views. First, that by thought to be a central feature of many cognitive functions, applying clustering algorithms, the brain’s dynamic architecture including adaptive goal pursuit, working memory, and set can be carved at the joints, revealing discrete brain states. Second, 10 NATURE COMMUNICATIONS (2018) 9:1157 DOI: 10.1038/s41467-018-03462-y www.nature.com/naturecommunications | | | NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-03462-y ARTICLE a b 1.00 1.00 State A State B 0.80 0.80 State C State D 0.60 0.60 0.40 0.40 0.20 0.20 0.00 0.00 0.00 0.20 0.40 0.60 0.80 1.00 0.00 0.20 0.40 0.60 0.80 1.00 False positive rate False positive rate Fig. 9 Speciﬁc brain states are predictive of distinct clinical symptoms in novel individuals. a State B uniquely predicted the presence of active psychosis in patients with psychotic illness. Results from a leave-one-subject-out cross-validation elastic net logistic regression analysis comparing predicted and observed psychotic symptoms (n = 91). The displayed area under the receiver-operating characteristic curve (AUC) reﬂects the scalar probability measure for each individual, predicting the likelihood of active psychotic symptoms. The blue semitransparent lines reﬂect 1000 ROC curves from 1000 permutation experiments. The black dashed line indicates the mean of these 1000 curves, approximating a null curve. The purple, red, green, and orange curves 4 4 4 4 correspond to states A ,B ,C , and D , respectively. b Connectivity models deﬁned on training data predict the presence of active psychotic symptoms in an independent group of participants (n = 31). The displayed AUC graph reﬂects the scalar probability measure, predicting the likelihood of active psychosis, deﬁned using positive edges from the initial training set without further ﬁtting or modiﬁcation, in a held-out sample of patients 8,16 that the resulting macro-level brain states index, or enable, dis- organization of the brain is impaired in psychotic illness . The tinct biological and cognitive processes. In the current analyses, present analyses reveal the presence of both state and network we apply k-means clustering to identify dissociable network preferential reductions in functional connectivity. Patients with conﬁgurations in solutions from 2 to 8 brain states. Critically, we schizophrenia spend more time than healthy individuals in net- do not claim that these conﬁgurations reﬂect wholly distinct brain work conﬁgurations typiﬁed by reduced large-scale connectivity, states separated by sharp transitions, and we are not in the while also showing muted negative correlations between default position to ascribe associated cognitive functions. Rather, the data and other networks . In line with this literature, we observed a are consistent with the possibility that large-scale network cou- general proﬁle of decreased within- and increased between- pling gradually ﬂuctuates around a core functional architecture. network correlations. Our analyses were also consistent with prior However, the extent to which neurobiological and cognitive work, demonstrating executive functioning and cognitive control mechanisms may drive the observed transient connectivity pat- abnormalities in patients with psychotic illness , revealing pre- terns remains an open question. ferential disruptions in frontoparietal control network con- 4 4 nectivity in the attractor state (state A ) and state C , a network Debates regarding the existence of discrete brain states should not be taken to imply that temporal ﬂuctuations in network conﬁguration marked by increased within-network frontoparietal connectivity lack biological information. Notably, the present connectivity in healthy young adults. This proﬁle of ﬂuctuating analyses indicate that an individual’s proﬁle of dynamic func- abnormalities in network connectivity was also evident in a tional connectivity is unique and stable over the course of days default network, which exhibited state-preferential impairments and months. Patterns of connectivity deﬁned through traditional in state A . Critically, these analyses do not demonstrate selective 25,26 static analyses are heritable and function as a trait-like sig- abnormalities in discrete network conﬁgurations that are speciﬁc nature that can accurately identify participants from a large to psychotic illness. Rather, consistent with evidence for altera- 27–29 16–18 group . While prior demonstrations of cross-session identi- tions in dynamic brain architecture in patient populations , ﬁcation were established when participant visits were separated they suggest that broad disruptions across cortical association by a single day , these data suggest that single measures of time- networks in psychosis may emerge through transient abnormal- averaged connectivity may provide meaningful trait-like infor- ities preferentially evident during the expression of particular 17,35 mation about individuals. Here, we observed robust participant- network conﬁgurations . Together with growing evidence 12– speciﬁc matching when visits were separated by up to 6 months linking behavior to temporally derived network conﬁgurations (mean = 63.35 ± 48.10 days apart, range = 2–151 days). Despite , the presence of both individual speciﬁcity (ﬁngerprints) in some variability, identiﬁcation accuracy was not limited to any brain dynamics across the population and evidence for ﬂuctuat- one state or clustering solution, indicating that participants may ing, network-speciﬁc impairments in patients with severe psy- be identiﬁed with relatively thin slices of transient brain activity. chopathology have strong implications for clinical practice. For These discoveries highlight the potential to identify associations instance, these data suggest a potential avenue to identify distinct linking the unique dynamic functional architecture of an indivi- symptom proﬁles, track time-varying disease states, responses to dual’s brain to the integrity of large-scale corticocortical path- environmental perturbations, or individually speciﬁc treatment ways . The continued development of time-varying data analytic responses. approaches with high sensitivity to individual variability could Consistent with the aim of delineating disease-relevant markers facilitate the discovery of meaningful biomarkers for both cog- of brain biology, the current analyses suggest that models based nitive ability and disease states. on transient proﬁles of network function may serve as powerful, Schizophrenia and psychotic bipolar disorder are marked by generalizable predictors of clinical symptomatology. In a group of altered intrinsic network connectivity, potentially contributing to patients with psychotic illness, we identiﬁed a speciﬁc time- 30–34 widespread changes in information processing . A key ques- varying network proﬁle whose strength predicted the presence of tion facing the ﬁeld is the extent to which the temporal active psychotic symptoms at the point of clinical assessment. NATURE COMMUNICATIONS (2018) 9:1157 DOI: 10.1038/s41467-018-03462-y www.nature.com/naturecommunications 11 | | | True positive rate True positive rate ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-03462-y Patients with psychotic illness were recruited from clinical services at McLean This whole-brain network model provides preliminary evidence Hospital (n = 170; age: 32.08 ± 11.64; female: 66.47%; right handed: 84.71%), for meaningful, clinically relevant signals in patterns of dynamic including 41 patients diagnosed with schizoaffective disorder, 56 with intrinsic connectivity. Suggesting that the clinically relevant fea- schizophrenia, and 73 with psychotic bipolar disorder. Study procedures are tures of intrinsic brain dynamics are robust and generalizable, detailed in Baker et al. 2014 . Brieﬂy, exclusion criteria included neurological illness, positive pregnancy test, electroconvulsive therapy in the last 3 months, and networks deﬁned within an initial training set successfully history of head trauma. Reﬂecting the severity of the present sample, the majority predicted active psychosis in a completely independent sample. of patients (82%) reported experiencing active psychotic symptoms at the time of As reﬂected in our analyses, a randomly selected, previously their clinical assessment, as assessed through DSM-IV (SCID) clinician-rated unseen, patient with active psychosis could be distinguished symptomatic diagnostic criteria, indicating the presence of delusions and/or hallucinations in the past month . All patients were assessed for active symptoms from another patient without psychosis at ~74.0% accuracy based within 24 h of scan using the Positive and Negative Syndrome Scale (PANSS; on their dynamic expression of selected network conﬁgurations positive scale: 18.46 ± 6.51; negative scale: 13.48 ± 7.47; general psychopathology (state B ), demonstrating a relatively high level of precision. scale: 30.82 ± 7.95) . A demographically matched healthy comparison sample was Critically, these relations were not evident through traditional recruited from the surrounding Boston communities (n = 369; age: 37.16 ± 14.65; female: 62.06%; right handed: 91.33%). The McLean Hospital Institutional Review static analysis. Readers should note that these analyses leverage Board approved the study, and all participants provided written informed consent. cross-sectional data, and are technically postdictive. Additional The control group was signiﬁcantly older than the patient group (t = –3.98, p ≤ research should assess if this cross-sectional/retrospective 0.001). No signiﬁcant group differences were identiﬁed in sex, handedness, or approach generalizes to the prospective prediction of symptoms, education (ps ≥ 0.07). The comparison sample was explicitly selected to match on prior to clinical assessment. the basis data quality. The BOLD runs for the patient and comparison groups did not differ in terms of slice-based temporal signal-to-noise ratio (comparison: Suggesting a degree of speciﬁcity for the prediction of psychotic 161.17 ± 46.05; patient: 158.35 ± 71.53) or the number of relative translations in 3D symptom severity, the active psychosis model served as a unique space ≥0.1 mm (comparison: 31.05 ± 29.15; patient: 32.29 ± 29.66; ps ≥ 0.58). The predictor of PANSS positive-scale scores (binarized as greater or slice-based signal-to-noise ratio was calculated as the weighted mean of each slice’s less than one standard deviation below the sample mean) in the mean intensity over time (weighted by the size of the slice). All imaging data were collected on 3-T Tim Trio scanners (Siemens) with a 12- left-out patient sample, with poor prediction observed for both channel phased-array head coil at Harvard University, Massachusetts General negative and general psychopathology symptoms. Caution is Hospital, or McLean Hospital. Structural data included a high-resolution multi- warranted given the limited sample size; however, these data echo T1-weighted magnetization-prepared gradient-echo image (TR = 2200 ms, provide evidence to suggest that dynamic network models trained TI = 1100 ms, TE = 1.54 ms for image 1–7.01 ms for image 4, FA = 7°, 1.2 × 1.2 × 1.2 mm, and FOV = 230). Functional data were acquired using a gradient-echo on different, yet associated, symptoms have the potential to echoplanar imaging sequence sensitive to blood oxygenation level-dependent generalize across clinical measures. These data expand our cur- contrast with the following parameters: 124 time points; repetition time = 3000 ms; rent knowledge regarding abnormal large-scale network function echo time = 30 ms; ﬂip angle = 85°; 3 × 3 × 3-mm voxels; FOV = 216; and 47 axial in patients with schizophrenia and psychotic bipolar disorder, sections collected with interleaved acquisition and no gap. Participants were instructed to remain still, stay awake, and keep their eyes open. Although no and highlight the use of dynamic analytic approaches when ﬁxation image was used, participants with psychotic illness were monitored via eye- examining intrinsic connectivity across heath and disease. Taken tracking video to ensure compliance during functional scans. One to two runs were together, the present analyses provide a preliminary proof of acquired for each participant (70.89% of the main sample, 69.41% of patient concept to suggest that the altered connectivity of speciﬁc tran- participants, and 46.88% of the matched comparison received a second run). sient network conﬁgurations may link to the expression of dis- Software upgrades (VB13, VB15, and VB17) occurred during data collection. Reported results are after partialing out variance associated with scanner and crete symptom proﬁles. Future work should focus on the software upgrade. identiﬁcation of relations linking functional network dynamics to the expression of psychological and behavioral aspects of illness. Data preprocessing. Data were processed with a series of steps common to In conclusion, we demonstrated the presence of a ﬂuctuating 69–71 intrinsic connectivity analyses . Preprocessing included discarding the ﬁrst four and reconﬁgurable hierarchy across the functional connectome. volumes of each run to allow for T1-equilibration effects, compensating for slice The observed dynamic network proﬁles were unique and reliable acquisition-dependent time shifts per volume, and correcting for head motion within individuals over the course of months and impaired in using rigid body translation and rotation. Additional steps included the removal of constant offset and linear trends over each run, and the application of a temporal patients with psychotic illness. Our analyses suggest that temporal ﬁlter to retain frequencies below 0.08 Hz. Sources of spurious variance, along with patterns of connectivity between cortical regions link to the broad their temporal derivatives, were removed through linear regression. These included functional capacities of individual human brains, enabling the six parameters obtained by correction for rigid body head motion, the signal prediction of speciﬁc symptom proﬁles within patient popula- averaged over the whole brain, the signal averaged over the ventricles, and the tions. These data have important implications for the study of signal averaged over the deep cerebral white matter. Structural data and functional 39 72 data were aligned as described in Yeo et al. and Buckner et al. using the behaviors and features of psychiatric illnesses that possess time- FreeSurfer software package. This method yields a surface mesh representation of varying patterns of expression. each participant’s cortex, which is then registered to a common spherical coordi- nate system. Images were aligned with boundary-based registration from the FsFast software. Functional and structural images were then aligned to the com- Methods mon coordinate system by sampling from the middle of the cortical ribbon in a Data acquisition. Native English-speaking young adults (aged 18–35) with normal single interpolation step to reduce blurring of the functional signal across sulci and or corrected-to-normal vision were recruited from Harvard University, Massa- gyri. A 6-mm smoothing kernel was applied to the functional data in the surface chusetts General Hospital, and the surrounding Boston communities through an space, and data were downsampled to a 4-mm mesh. Additional details on the ongoing large-scale study of brain imaging and genetics (n = 1919; age: 21.35 ± 43 39 preprocessing procedures are detailed in Holmes et al. and Yeo et al. . 3.20; female: 56.53%; right handed: 92.40%) . History of psychiatric illness and medication usage was assessed through a structured phone screen. On the day of MRI data collection, participants completed additional questionnaires concerning Dynamic connectivity sliding-window analysis. Cortical functional coupling their physical health, past and present history of psychiatric illness, and medication matrices were computed for each participant, across all available parcels within the usage. Exclusion criteria included a history of head trauma, current or past Axis I 17-network functional atlas of Yeo et al. .We deﬁned 114 regions (57 per pathology, neurological disorders, current or past psychotropic medication use, hemisphere) that surveyed all 17 networks. Correlation matrices were constructed current physical illness, and current or past loss of consciousness. Participants to include all region pairs arranged by network membership. provided written informed consent in accordance with guidelines set by the Connectivity across time was analyzed using a sliding-window approach 7,8 Partners Health Care Institutional Review Board and the Harvard University (width = 33 s) . Prior work suggests that a sliding-window range of 30–60 s is Committee on the Use of Human Subjects in Research. For the present study, we appropriate for dynamic connectivity analyses . Pilot analyses (available upon assessed the extent to which the dynamic network architecture of the cortex is request) revealed consistent state solution stability across varying sliding-window reliable within and across visits. To accomplish this, an additional data set (n = 79; sizes of 33–63 s. Thirty-three-second windows were chosen in order to maximize age at the ﬁrst scan: 20.99 ± 2.93; female: 45.56%; right handed: 89.87%) was signal estimates, while still capturing properties of transient functional 8,40 acquired over the course of the primary collection effort. Data were collected on 2 connectivity . A time series for each participant was extracted for the 57 regions independent days (mean = 63.35 ± 48.10 days apart; min = 2; max = 151). in each hemisphere. Time-course correlations across 110 windows per bold run for 12 NATURE COMMUNICATIONS (2018) 9:1157 DOI: 10.1038/s41467-018-03462-y www.nature.com/naturecommunications | | | NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-03462-y ARTICLE each participant (220 windows if the participant had two runs) were calculated for ﬁrst established a comparison distribution from chance. For each state, the each 114 × 114 region pair. To limit the redundancy across matrices and to reduce participant vector from visit 1 was compared to a randomly permuted list of computational load, clustering was applied to a subsample of available windowed participants from visit 2, a Pearson correlation was calculated, a matching rank was covariance matrices (1/10 windows). Results were consistent with the alternate assigned, and the identiﬁcation rate was calculated. A “correct” identiﬁcation was approach of subsampling along the temporal dimension to identify windowed deﬁned in cases where the highest-ranked rho value was within a participant (visit covariance matrices with local maxima in functional connectivity variance. The 1–visit 2) relative to other participants (for ranking results, see Fig. 4, main text). resulting correlation matrices were then aggregated and z-transformed prior to This was repeated 1000 times, and a t-test was performed comparing the number of running the clustering analyses. The clustering analysis was iteratively applied to identiﬁcations of the actual order participants relative to the distribution of the deﬁne distinct state solutions for 2 through 20 brain states. Additional details on number of identiﬁcations in the permuted order of the participant list. the selected clustering approach are provided in Yeo et al. As we did not have a priori hypotheses regarding the number of functional Noise constraints. To assess the extent to which data quality might inﬂuence on connectivity states, we assessed the stability of clustering solutions for all states our ﬁndings, we conducted a series of analyses aimed at detecting the differences in 2–20. To do so, we examined the stability of the clustering analyses by iteratively motion across state solutions. To obtain motion estimates by a participant for each and randomly splitting our data on two dimensions (sliding windows and pairwise window, we extracted the mean of the root mean square of relative motion for each connectivity) and rerunning the clustering solution 30 times . The results were participant across each window. We used the Hungarian matched state vector to 75,76 then compared using a Hungarian matching algorithm , as described in the classify relative motion within each TR to a state-speciﬁc window. We averaged next section. Greater instability was quantiﬁed as a greater summation of deviation motion values within states for each state solution. Although the relations linking between the two cluster solutions (Supplementary Figure 1). state expression and motion were limited in size (η s ≤ 0.006), state A associated with the most motion for state solutions 2, 3, and 4 (all ps ≤ 0.05; Supplementary Figure 5). State A showed a muted increase in motion relative to other states in Hierarchy analysis. To examine the relations linking each state in solution S with 75,76 5 5 the states in solution S + 1, we used a Hungarian matching technique . Every solutions 5 through 7 (ﬁve-state solution: ps ≤ 0.001 for state A relative to C and 5 5 5 6 E , p ≤ 0.05; p ≥ 0.59 for B and D ; six-state solution: ps ≤ 0.01 for state A relative possible combination of states in state solution S + 1 was compared to each state in 6 6 6 6 6 to E and F , ps ≥ 0.38 for B ,C , and D ; seven-state solution: ps ≤ 0.001 for state solution S. To illustrate this point, take the comparison of states in the three- and 7 7 7 7 7 7 A relative to E and F ; p = 0.07 for C , p = 0.06 for G , and p ≥ 0.61 for B and four-state solutions. Each possible combination in the four-state solution is 4 4 4 4 4 4 4 4 4 4 4 4 7 8 D ). In state solution 8, we found that state H , a state that our hierarchy analyses established (A B ,A C ,A D ,B C ,B D , and C D ) by calculating the mean of each cell in the 114 × 114 × n (here, n = 4) connectivity matrix across the combined identiﬁed as architecturally related to state A in solutions 2–7, was associated with more motion than all other states (ps ≤ 0.001). states to create hybrid states. Next, each hybrid state is grouped along with the other states in the four-state solution and compared to the three-state solution (A Next, we examined the consistency of state stability in participants falling 4 4 4 4 4 4 4 4 4 4 4 within the ﬁrst and fourth quartiles of the distribution for motion (low motion: + B /C /D , then A + C /B /D , then A + D /B /C , etc.). Hungarian matching is used to determine which of the hybrid combinations in the four-state solution n = 424; high motion: n = 398). No group differences were identiﬁed in age, sex, or handedness (ps ≥ 0.59). To determine the stability of our state-clustering approach, most closely approximates each state in the three-state solution (so, in the ﬁrst 4 4 4 4 example, the hybrid four-state solution comprised of A + B /C /D is matched to we applied iterative k-means clustering to each group for each population-level 3 3 3 stable state solution (2, 4, 5, and 8 states). To estimate the viability of the resulting states A ,B , and C in the three-state solution). This comparison is repeated until the match with the minimal cost is identiﬁed (Supplementary Figure 2). solutions, data were resampled across sliding windows as described previously (Fig. 1, main text). A t-test was performed to assess the differences in the consistency sampling from the high and low motion groups. No differences in Identifying variability across deﬁned state solutions. To determine the extent consistency were observed across groups (Supplementary Figure 6; all ps ≥ 0.43). of network variability across brain states, we examined the variance of mean network connectivity in state solutions 2–8 (see Supplementary Figure 3 for var- Prediction of active psychotic symptoms. Here, we demonstrate that the iance with the four-state solution). We used ANOVA to assess between-state strength of functional brain networks within speciﬁc brain states predicts the variability (coefﬁcient of variance) with state solutions treated as repeated mea- presence of active psychosis in previously unseen individuals. Elastic net logistic sures. The results conﬁrmed that variance differed across the networks (F = regression analyses were conducted with custom R code (elasticnet by Hui Zou, 44.11, p ≤ 0.001). Post hoc tests revealed increased cross-state variability within Trevor Hastie, and Robert Tibshirani). Elastic net regularization is a cross-validated default A and B relative to the default D, control A and C, limbic, somatomotor A, regularized log-linear regression procedure that combines LASSO (least and visual B networks (Bonferroni-corrected ps ≤ 0.05, all other ps ≥ 0.5). Control B absolute shrinkage and selection operator) regularization and Tikhonov (ridge) demonstrated greater variance relative to default C and D, control A and C, limbic, 77,78 regularization The resulting log-linear regression weights were applied to somatomotor, and visual networks (ps ≤ 0.05; all other ps ≥ 0.5). The salience/ the edges (ROI to ROI correlations) of each network conﬁguration within the four- ventral attention and dorsal attention A networks exhibited increased variance state solution and the associated covariates. All results were cross validated. relative to default C and D, control A and C, limbic, somatomotor, and visual Model development and validation consisted of four steps. First, model features networks (ps ≤ 0.05, all other ps ≥ 0.5). th 4 4 4 were selected. Pearson correlation between each edge of the k (e.g., k = A ,B ,C , and D ) dynamic brain state and clinical status was performed in the training set. Individual identiﬁcation analyses. To examine the extent to which the observed Note that regarding dichotomous outcomes, a mass-univariate t-test would provide dynamic connectivity proﬁles are reliably expressed across scans and visits, we ﬁrst similar results as the Pearson correlation test concerning feature selection. Here, we selected participants with two bold runs (n = 1361). Next, we examined the 79 27,52 used the Pearson correlation approach to be consistent with previous literature . participants, set aside from the original cohort, who had two separate study visits The resulting edges were separated into positive and negative groups, and within 6 months of each other. To test the relations within each individual’s proﬁle thresholded on the basis of the statistical signiﬁcance (p ≤ 0.05) and signs of of network dynamics for the two bold runs within the same visit, and then for the correlation. Second, in the model development procedure, we ﬁrst aggregated the individuals with more than one visit, we implemented the following analysis: First, k values of edges in each feature set as a summary statistics, S ,or “network we obtained each participant’s state expression across time. To accomplish this, we 27 th strength” of the k brain state, for k ¼ 1; 2; 3; 4. The network strength and used a Hungarian matching algorithm, assigning each time point in a participant’s covariates (e.g., sex, age) were then entered into the model, yielding a scalar value, windowed time-course data to individual states in the desired population-level state the predicated conditional probability of active psychotic symptoms. Formally, for solution. For instance, in the four-state solution, we found the best ﬁt for each subject i, i 2f1; 2; ; ng,we deﬁne 4 4 4 4 window in the participant’s data and classiﬁed it as state A ,B ,C ,orD . This k k k ^ ^ k β þβ s þð^γ Þci yielded a vector of states for each participant, representing the participant’s state 0 1 i k k k ^p :¼ P Y ¼ 1jS ¼ s ; C ¼ c ¼ i i k k k expression over the course of the scan. Following this, we took the average cor- i i ^ ^ k β þβ s þð^γ Þci 0 1 i 1 þ e relation matrix for each window within a state. So, for participant 1, we collapsed 4 4 across all of the state A windows, and generated a mean for state A , creating a k k k where p :¼ P Y ¼ 1jS ¼ s ; C ¼ c denotes the predicted probability that i i i i 114 × 114 × S average state matrix for each participant. Participant matrices were subject i has active psychosis, given their observed network strength s during brain vectorized, and Pearson correlations were run across every participant in two state k, and covariates c , a vector consisting of all observed covariates for subject i. analyses for (1) bold 1 and bold 2; and every participant for (2) visit 1 and visit 2. k k k ^ ^ β ; β ; and ^γ are estimated weights for brain state k from the elastic net logistic 0 1 Analyses indicated a signiﬁcant level of consistency in within-individual con- k regression, where ^γ is a vector consisting of the estimates for each covariate in c . nectivity proﬁles across bold runs and visits (Supplementary Figure 4). For each Alternatively, we could model the predicted probability that subject i does not have state solution, we ran a t-test for within-participant rho values and between- k k 1 psychosis as P Y ¼ 0jS ¼ s ; C ¼ c ¼ : Probability estimation was i i k k k k c i ^ ^ β þβ s þð^γ Þ participant rho values. All tests revealed increased rho values within, rather than 0 1 i i 1þe between, participants (all ps ≤ 0.001). iteratively performed using leave-one-subject-out cross-validation procedure. Permutation tests were performed in a manner consistent with prior studies During each iteration, the weights were estimated using data from (n–1, n = 91) examining functional connectome ﬁngerprinting . For these analyses, we participants and were used to predict the probability of the remaining participant considered participants with multiple study visits (n = 79) and utilized the state- having active psychotic symptoms. Each individual was left out once; hence, the and participant-speciﬁc connectivity vectors described above. A Pearson procedure yielded n-predicted probability scores. To evaluate the estimation correlation was calculated across an ordered list for every participant’s connectivity performance, we measured the area under the receiver-operating characteristic vector from visit 1 compared to every other participant’s vector from visit 2. We (ROC) curve (AUC), estimated directly by conducting numerical integration of the NATURE COMMUNICATIONS (2018) 9:1157 DOI: 10.1038/s41467-018-03462-y www.nature.com/naturecommunications 13 | | | ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-03462-y ROC under all thresholds that yielded unique sensitivity/speciﬁcity values, wherein 16. Calhoun, V. D., Miller, R., Pearlson, G. & Adali, T. The chronnectome: time- 0.5 indicates chance, and 1 is perfect discrimination. varying connectivity networks as the next frontier in fMRI data discovery. The model development and parameter estimation were conducted only using Neuron 84, 262–274 (2014). the training data. To evaluate the reproducibility, we applied the model obtained 17. Rashid, B., Damaraju, E., Pearlson, G. D. & Calhoun, V. D. Dynamic from the training data, without further ﬁtting or modiﬁcation, to 39 previously connectivity states estimated from resting fMRI identify differences among unseen participants. To access the speciﬁcity of the model in detecting positive schizophrenia, bipolar disorder, and healthy control subjects. Front. Hum. symptoms, we examined the positive, negative, and general psychopathology Neurosci. 8, 897 (2014). subscales of the PANSS in the held-out sample. PANSS scores were binarized as 18. Damaraju, E. et al. Dynamic functional connectivity analysis reveals transient low (less than one standard deviation below the mean) or high (greater than one states of dysconnectivity in schizophrenia. Neuroimage Clin. 5, 298–308 standard deviation below the mean) for each subscale (positive scale: 18.46 ± 6.51, (2014). cutoff = 11.95; negative scale: 13.48 ± 7.47, cutoff = 6.01; general psychopathology 19. Breakspear, M. Dynamic models of large-scale brain activity. Nat. Neurosci. scale: 30.82 ± 7.95, cutoff = 22.86). Due to the large variance in PANSS negative 20, 340–352 (2017). scores, to thoroughly explore potential relations with this subscale, we additionally 20. Schölvinck, M. L., Leopold, D. A., Brookes, M. J. & Khader, P. H. The examined alternative thresholds of <15 and of <12 selecting the optimal result contribution of electrophysiology to functional connectivity mapping. (<12) for comparison with the PANSS positive scale. Neuroimage 80, 297–306 (2013). To assess the statistical signiﬁcance of the sensitivity and speciﬁcity analyses, we 21. Van de Ville, D., Britz, J. & Michel, C. M. EEG microstate sequences in healthy performed nonparametric permutation testing. During each test, we ﬁrst randomly humans at rest reveal scale-free dynamics. Proc. Natl Acad. Sci. USA 107, permuted the active psychotic symptoms status or PANSS labels. Next, model 18179–18184 (2010). ﬁtting was carried out using the correct dynamic signatures and covariates and 22. Britz, J., Van De Ville, D. & Michel, C. M. BOLD correlates of EEG permuted (incorrect) labels. For each permutation, an AUC value was calculated. topography reveal rapid resting-state network dynamics. Neuroimage 52, The permutation tests were performed 5000 times for the training and testing data, 1162–1170 (2010). respectively. The null AUC values should be symmetrically distributed around 0.5 23. Musso, F., Brinkmeyer, J., Mobascher, A., Warbrick, T. & Winterer, G. if the procedure is unbiased. Spontaneous brain activity and EEG microstates. A novel EEG/fMRI analysis approach to explore resting-state networks. Neuroimage 52, 1149–1161 (2010). Data availability. The data summary statistics that support the ﬁndings of this 24. Chang, C. et al. Association between heart rate variability and ﬂuctuations in study are available from the corresponding author on request. As detailed in resting-state functional connectivity. Neuroimage 68,93–104 (2013). Holmes et al. , the Brain Genomics Superstruct Project (GSP) initial release data 25. Glahn, D. C. et al. Genetic control over the resting brain. Proc. Natl Acad. Sci. set of structural, functional, and behavioral measures is available for download USA 107, 1223–1228 (2010). (http://neuroinformatics.harvard.edu/gsp/). Step-by-step instructions detailing how 26. Ge, T., Holmes, A. J., Buckner, R. L., Smoller, J. W. & Sabuncu, M. R. to access the release data set are available online in the “Request Access” page Heritability analysis with repeat measurements and its application to resting- (http://neuroinformatics.harvard.edu/gsp/get). The patient data are not publicly state functional connectivity. Proc. Natl Acad. Sci. USA 114, 5521–5526 available due to information that could compromise research participant privacy/ (2017). consent. 27. Finn, E. S. et al. Functional connectome ﬁngerprinting: identifying individuals using patterns of brain connectivity. Nat. Neurosci. 18, 1664–1671 (2015). Received: 19 July 2017 Accepted: 15 February 2018 28. Miranda-Dominguez, O. et al. Connectotyping: model based ﬁngerprinting of the functional connectome. PLoS ONE 9, e111048 (2014). 29. Wang, D. et al. Parcellating cortical functional networks in individuals. Nat. Neurosci. 18, 1853–1860 (2015). 30. Baker, J. T. et al. Disruption of cortical association networks in schizophrenia and psychotic bipolar disorder. JAMA Psychiatry 71, 109–118 (2014). References 31. Buckner, R. L., Andrews-Hanna, J. R. & Schacter, D. L. The brain’s default 1. Buckner, R. L., Krienen, F. M. & Yeo, B. T. T. Opportunities and limitations of network: anatomy, function, and relevance to disease. Ann. NY Acad. Sci. intrinsic functional connectivity MRI. Nat. Neurosci. 16, 832–837 (2013). 1124,1–38 (2008). 2. Craddock, R. C. et al. Imaging human connectomes at the macroscale. Nat. 32. Cole, M. W., Repovs, G. & Anticevic, A. The frontoparietal control system: a Methods 10, 524–539 (2013). central role in mental health. Neuroscientist 20, 652–664 (2014). 3. Van Essen, D. C. & Ugurbil, K. The future of the human connectome. 33. Fornito, A. & Bullmore, E. T. Reconciling abnormalities of brain network Neuroimage 62, 1299–1310 (2012). structure and function in schizophrenia. Curr. Opin. Neurobiol. 30,44–50 4. Bullmore, E. & Sporns, O. Complex brain networks: graph theoretical analysis (2015). of structural and functional systems. Nat. Rev. Neurosci. 10, 186–198 (2009). 34. Fornito, A., Zalesky, A., Pantelis, C. & Bullmore, E. T. Schizophrenia, 5. Krienen, F. M., Yeo, B. T. T. & Buckner, R. L. Reconﬁgurable task-dependent neuroimaging and connectomics. Neuroimage 62, 2296–2314 (2012). functional coupling modes cluster around a core functional architecture. 35. Uhlhaas, P. J. & Singer, W. Abnormal neural oscillations and synchrony in Philos. Trans. R. Soc. Lond. B Biol. Sci. 369, 20130526 (2014). schizophrenia. Nat. Rev. Neurosci. 11, 100–113 (2010). 6. Spreng, R. N., Stevens, W. D., Chamberlain, J. P., Gilmore, A. W. & Schacter, 36. Miller, R. L., Yaesoubi, M., Calhoun, V. D. & Gopal, S. Higher dimensional D. L. Default network activity, coupled with the frontoparietal control fMRI connectivity dynamics show reduced dynamism in schizophrenia network, supports goal-directed cognition. Neuroimage 53, 303–317 (2010). patients. in 2014 International Workshop on Pattern Recognition in 7. Allen, E. A. et al. Tracking whole-brain connectivity dynamics in the resting Neuroimaging 3837–3840 (IEEE, 2014). state. Cereb. Cortex. 24, 663–676 (2014). 37. Sakoğlu, Ü. et al. A method for evaluating dynamic functional network 8. Hutchison, R. M. et al. Dynamic functional connectivity: promise, issues, and connectivity and task-modulation: application to schizophrenia. Magn. Reson. interpretations. Neuroimage 80, 360–378 (2013). Mater. Phys. 23, 351–366 (2010). 9. Zalesky, A., Fornito, A., Cocchi, L., Gollo, L. L. & Breakspear, M. Time- 38. American Psychiatric Association. Diagnostic and Statistical Manual of resolved resting-state brain networks. Proc. Natl Acad. Sci. USA 111, Mental Disorders (DSM-IV) (APA Press, Washington, DC, 1994). 10341–10346 (2014). 39. Yeo, B. T. T. et al. The organization of the human cerebral cortex estimated by 10. Laumann, T. O. et al. On the stability of BOLD fMRI correlations. Cereb. intrinsic functional connectivity. J. Neurophysiol. 106, 1125–1165 (2011). Cortex 27, 4719–4732 (2017). 40. Chang, C. & Glover, G. H. Time-frequency dynamics of resting-state brain 11. Liegeois, R., Laumann, T. O., Snyder, A. Z., Zhou, J. & Yeo, B. T. T. connectivity measured with fMRI. Neuroimage 50,81–98 (2010). Interpreting temporal ﬂuctuations in resting-state functional connectivity 41. Hutchison, R. M., Womelsdorf, T., Gati, J. S., Everling, S. & Menon, R. S. MRI. Neuroimage 163, 437-455 (2017). Resting-state networks show dynamic functional connectivity in awake 12. Cole, M. W., Bassett, D. S., Power, J. D., Braver, T. S. & Petersen, S. E. Intrinsic humans and anesthetized macaques. Hum. Brain Mapp. 34, 2154–2177 and task-evoked network architectures of the human brain. Neuron 83, (2013). 238–251 (2014). 42. Handwerker, D. A., Roopchansingh, V., Gonzalez-Castillo, J. & Bandettini, P. 13. Shine, J. M., Koyejo, O. & Poldrack, R. A. Temporal metastates are associated A. Periodic changes in fMRI connectivity. Neuroimage 63, 1712–1719 (2012). with differential patterns of time-resolved connectivity, network topology, and 43. Holmes, A. J. et al. Brain genomics superstruct project initial data release with attention. Proc. Natl Acad. Sci. USA 113, 9888–9891 (2016). structural, functional, and behavioral measures. Sci. Data 2, 150031 (2015). 14. Wang, C., Ong, J. L., Patanaik, A., Zhou, J. & Chee, M. W. L. Spontaneous 44. Ben-Hur, A., Elisseeff, A. & Guyon, I. A stability based method for discovering eyelid closures link vigilance ﬂuctuation with fMRI dynamic connectivity structure in clustered data. Pac. Symp. Biocomput. 7,6–17 (2002). states. Proc. Natl Acad. Sci. USA 113, 9653–9658 (2016). 45. Lange, T., Roth, V., Braun, M. L. & Buhmann, J. M. Stability-based validation 15. Kucyi, A., Hove, M. J., Esterman, M., Hutchison, R. M. & Valera, E. M. of clustering solutions. Neural Comput. 16, 1299–1323 (2004). Dynamic brain network correlates of spontaneous ﬂuctuations in attention. 46. Mueller, S. et al. Individual variability in functional connectivity architecture Cereb. Cortex 27, 1831–1840 (2016). of the human brain. Neuron 77, 586–595 (2013). 14 NATURE COMMUNICATIONS (2018) 9:1157 DOI: 10.1038/s41467-018-03462-y www.nature.com/naturecommunications | | | NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-03462-y ARTICLE 47. Goldman-Rakic, P. S. Topography of cognition: parallel distributed networks 73. Greve, D. N. & Fischl, B. Accurate and robust brain image alignment using in primate association cortex. Annu. Rev. Neurosci. 11, 137–156 (1988). boundary-based registration. Neuroimage 48,63–72 (2009). 48. Crossley, N. A. et al. Cognitive relevance of the community structure of the 74. Shirer, W. R., Ryali, S., Rykhlevskaia, E., Menon, V. & Greicius, M. D. human brain functional coactivation network. Proc. Natl Acad. Sci. USA 110, Decoding subject-driven cognitive states with whole-brain connectivity 11583–11588 (2013). patterns. Cereb. Cortex 22, 158–165 (2012). 49. Shine, J. M., et al. The dynamics of functional brain networks: integrated 75. Kuhn, H. W. The Hungarian method for the assignment problem. Nav. Res. network states during cognitive task performance. Neuron 92, 544–554 (2016). Logist. Q. 2,83–97 (1955). 50. Lynall, M. E. et al. Functional connectivity and brain networks in 76. Munkres, J. Algorithms for the assignment and transportation problems. J. schizophrenia. J. Neurosci. 30, 9477–9487 (2010). Soc. Ind. Appl. Math. 5,32–38 (1957). 51. Repovs, G., Csernansky, J. G. & Barch, D. M. Brain network connectivity in 77. Hastie, T., Tibshirani, R. & Friedman, J. H. The Elements of Statistical individuals with schizophrenia and their siblings. Biol. Psychiatry 69, 967–973 Learning- Data Mining, Inference, and Predictions (Springer, New York, (2011). 2001). 52. Rosenberg, M. D. et al. A neuromarker of sustained attention from whole- 78. Zou, H. & Hastie, T. Regularization and variable selection via the elastic net. J. brain functional connectivity. Nat. Neurosci. 19, 165–171 (2016). R. Stat. Soc. B 67, 301–320 (2005). 53. Shen, X. et al. Using connectome-based predictive modeling to predict 79. Van Essen, D. C. A population-average, landmark- and surface-based (PALS) individual behavior from brain connectivity. Nat. Protoc. 12, 506–518 (2017). atlas of human cerebral cortex. Neuroimage 28, 635–662 (2005). 54. Kay, S. R., Fiszbein, A. & Opfer, L. A. The positive and negative syndrome scale (PANSS) for schizophrenia. Schiz Bull. 13, 261–276 (1987). Acknowledgements 55. Price, C. J. & Friston, K. J. Functional ontologies for cognition: the systematic This work was supported by the Brain & Behavior Research Foundation (A.J.H.), the deﬁnition of structure and function. Cogn. Neuropsychol. 22, 262–275 (2005). Taplin Family Foundation (D.Ö.), and the National Institute of Mental Health (Grants 56. Barttfeld, P. et al. Signature of consciousness in the dynamics of resting-state R01-MH079799 and K24MH094614 to J.W.S., 23MH079982-01A1 and 5R01MH094594 brain activity. Proc. Natl Acad. Sci. USA 112, 887–892 (2015). to D.Ö, and Grant K01MH099232 to A.J.H.). Analyses were made possible by the 57. Engel, A. K., Gerloff, C., Hilgetag, C. C. & Nolte, G. Intrinsic coupling modes: resources provided through Shared Instrumentation Grants 1S10RR023043 and multiscale interactions in ongoing brain activity. Neuron 80, 867–886 (2013). 1S10RR023401. We thank Monica Rosenberg for her feedback on early versions of this 58. Honey, C. J., Kötter, R., Breakspear, M. & Sporns, O. Network structure of manuscript. Data were provided in part by the Brain Genomics Superstruct Project of cerebral cortex shapes functional connectivity on multiple time scales. Proc. Harvard University and Massachusetts General Hospital (principal investigators: Randy Natl Acad. Sci. USA 104, 10240–10245 (2007). L. Buckner, J.L.R., and J.W.S.) with support from the Center for Brain Science Neu- 59. Shen, K., Hutchison, R. M., Bezgin, G., Everling, S. & McIntosh, A. R. Network roinformatics Research Group, the Athinoula A. Martinos Center for Biomedical Ima- structure shapes spontaneous functional connectivity dynamics. J. Neurosci. ging, and the Center for Genomic Medicine. Both the Simons Foundation and the 35, 5579–5588 (2015). Howard Hughes Medical Institute supported Dr. Buckner’s work on the GSP, and we are 60. Hutchison, R. M., Hutchison, M., Manning, K. Y., Menon, R. S. & Everling, S. immensely grateful for his invaluable contributions on that project. Isoﬂurane induces dose-dependent alterations in the cortical connectivity proﬁles and dynamic properties of the brain’s functional architecture. Hum. Brain Mapp. 35, 5754–5775 (2014). Author contributions 61. Deco, G., Jirsa, V. K. & McIntosh, A. R. Resting brains never rest: J.M.R. and A.J.H. conceived the study; J.M.R., O.Y.C., K.M.A., and A.J.H. performed the computational insights into potential cognitive architectures. Trends Neurosci. analyses with technical support and guidance with data interpretation from R.M.H., M.R. 36, 268–274 (2013). S., D.Ö., J.L.R., J.W.S., J.T.B., and B.T.Y.; and J.M.R., O.Y.C., and A.J.H wrote the 62. Deco, G. et al. Identiﬁcation of optimal structural connectivity using manuscript. All authors commented on the manuscript. functional connectivity and neural modeling. J. Neurosci. 34, 7910–7916 (2014). 63. Yeo, B. T. T. et al. Functional specialization and ﬂexibility in human Additional information association cortex. Cereb. Cortex 25, 3654–3672 (2015). Supplementary Information accompanies this paper at https://doi.org/10.1038/s41467- 018-03462-y. 64. Ciric, R., Nomi, J. S., Uddin, L. Q. & Satpute, A. B. Contextual connectivity: a framework for understanding the intrinsic dynamic architecture of large-scale functional brain networks. Sci. Rep. 7, 6537 (2017). Competing interests: The authors declare no competing interests. 65. Corbetta, M. & Shulman, G. L. Control of goal-directed and stimulus-driven attention in the brain. Nat. Rev. Neurosci. 3, 201–215 (2002). Reprints and permission information is available online at http://npg.nature.com/ 66. de Pasquale, F., Della Penna, S., Sporns, O., Romani, G. L. & Corbetta, M. A reprintsandpermissions/ dynamic core network and global efﬁciency in the resting human brain. Cereb. Publisher's note: Springer Nature remains neutral with regard to jurisdictional claims in Cortex 26, 4015–4033 (2016). 67. Leonardi, N., Shirer, W. R., Greicius, M. D. & Van De Ville, D. Disentangling published maps and institutional afﬁliations. dynamic networks: Separated and joint expressions of functional connectivity patterns in time. Hum. Brain Mapp. 35, 5984–5995 (2014). 68. Perlstein, W. M., Carter, C. S., Noll, D. C. & Cohen, J. D. Relation of prefrontal cortex dysfunction to working memory and symptoms in schizophrenia. Am. Open Access This article is licensed under a Creative Commons J. Psychiatry 158, 1105–1113 (2001). Attribution 4.0 International License, which permits use, sharing, 69. Biswal, B., Yetkin, F. Z., Haughton, V. M. & Hyde, J. S. Functional adaptation, distribution and reproduction in any medium or format, as long as you give connectivity in the motor cortex of resting human brain using echo-planar appropriate credit to the original author(s) and the source, provide a link to the Creative MRI. Magn. Reson. Med. 34, 537–541 (1995). Commons license, and indicate if changes were made. The images or other third party 70. Fox, M. D. et al. The human brain is intrinsically organized into dynamic, material in this article are included in the article’s Creative Commons license, unless anticorrelated functional networks. Proc. Natl Acad. Sci. USA 102, 9673–9678 indicated otherwise in a credit line to the material. If material is not included in the (2005). article’s Creative Commons license and your intended use is not permitted by statutory 71. Van Dijk, K. R. A. et al. Intrinsic functional connectivity as a tool for human regulation or exceeds the permitted use, you will need to obtain permission directly from connectomics: theory, properties, and optimization. J. Neurophysiol. 103, the copyright holder. To view a copy of this license, visit http://creativecommons.org/ 297–321 (2010). licenses/by/4.0/. 72. Buckner, R. L., Krienen, F. M., Castellanos, A., Diaz, J. C. & Yeo, B. T. T. The organization of the human cerebellum estimated by intrinsic functional connectivity. J. Neurophysiol. 106, 2322–2345 (2011). © The Author(s) 2018 NATURE COMMUNICATIONS (2018) 9:1157 DOI: 10.1038/s41467-018-03462-y www.nature.com/naturecommunications 15 | | | http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Nature Communications Springer Journals http://www.deepdyve.com/lp/springer-journals/the-human-cortex-possesses-a-reconfigurable-dynamic-network-UYNDLBN9Ax

Loading next page...

References (148)

A Fornito (2012)
10.1016/j.neuroimage.2011.12.090
Neuroimage, 62
RL Buckner, FM Krienen, A Castellanos, JC Diaz, BTT Yeo (2011)
The organization of the human cerebellum estimated by intrinsic functional connectivity
J. Neurophysiol., 106
F Musso (2010)
10.1016/j.neuroimage.2010.01.093
Neuroimage, 52
E Damaraju (2014)
10.1016/j.nicl.2014.07.003
Neuroimage Clin., 5
G Deco (2014)
10.1523/JNEUROSCI.4423-13.2014
J. Neurosci., 34
D Van de Ville (2010)
10.1073/pnas.1007841107
Proc. Natl Acad. Sci. USA, 107
DN Greve, B Fischl (2009)
Accurate and robust brain image alignment using boundary-based registration
Neuroimage, 48
DA Handwerker, V Roopchansingh, J Gonzalez-Castillo, PA Bandettini (2012)
Periodic changes in fMRI connectivity
Neuroimage, 63
A Kucyi, MJ Hove, M Esterman, RM Hutchison, EM Valera (2016)
Dynamic brain network correlates of spontaneous fluctuations in attention
Cereb. Cortex, 27
RM Hutchison (2013)
Dynamic functional connectivity: promise, issues, and interpretations
Neuroimage, 80
C Chang (2013)
Association between heart rate variability and fluctuations in resting-state functional connectivity
Neuroimage, 68
P Barttfeld (2015)
Signature of consciousness in the dynamics of resting-state brain activity
Proc. Natl Acad. Sci. USA, 112
R Liegeois (2017)
10.1016/j.neuroimage.2017.09.012
Neuroimage, 163
DC Essen (2005)
A population-average, landmark- and surface-based (PALS) atlas of human cerebral cortex
Neuroimage, 28
RN Spreng (2010)
10.1016/j.neuroimage.2010.06.016
Neuroimage, 53
D Wang (2015)
10.1038/nn.4164
Nat. Neurosci., 18
C Chang (2010)
10.1016/j.neuroimage.2009.12.011
Neuroimage, 50
G Repovs, JG Csernansky, DM Barch (2011)
Brain network connectivity in individuals with schizophrenia and their siblings
Biol. Psychiatry, 69
T Lange (2004)
10.1162/089976604773717621
Neural Comput., 16
ME Lynall (2010)
10.1523/JNEUROSCI.0333-10.2010
J. Neurosci., 30
WR Shirer (2012)
10.1093/cercor/bhr099
Cereb. Cortex, 22
R Liegeois, TO Laumann, AZ Snyder, BTT Yeo (2017)
Interpreting temporal fluctuations in resting-state functional connectivity MRI
Neuroimage, 163
JM Shine, O Koyejo, RA Poldrack (2016)
Temporal metastates are associated with differential patterns of time-resolved connectivity, network topology, and attention
Proc. Natl Acad. Sci. USA, 113
WM Perlstein (2001)
10.1176/appi.ajp.158.7.1105
Am. J. Psychiatry, 158
SR Kay (1987)
10.1093/schbul/13.2.261
Schiz Bull., 13
O Miranda-Dominguez (2014)
10.1371/journal.pone.0111048
PLoS ONE, 9
CJ Price, KJ Friston (2005)
Functional ontologies for cognition: the systematic definition of structure and function
Cogn. Neuropsychol., 22
B Biswal (1995)
10.1002/mrm.1910340409
Magn. Reson. Med., 34
DA Handwerker (2012)
10.1016/j.neuroimage.2012.06.078
Neuroimage, 63
SR Kay, A Fiszbein, LA Opfer (1987)
The positive and negative syndrome scale (PANSS) for schizophrenia
Schiz Bull., 13
RL Buckner (2013)
10.1038/nn.3423
Nat. Neurosci., 16
DC Van Essen (2012)
10.1016/j.neuroimage.2012.01.032
Neuroimage, 62
10.1098/rstb.2013.0526
RM Hutchison, M Hutchison, KY Manning, RS Menon, S Everling (2014)
Isoflurane induces dose-dependent alterations in the cortical connectivity profiles and dynamic properties of the brain’s functional architecture
Hum. Brain Mapp., 35
MW Cole, G Repovs, A Anticevic (2014)
The frontoparietal control system: a central role in mental health
Neuroscientist, 20
AJ Holmes (2015)
10.1038/sdata.2015.31
Sci. Data, 2
10.1016/j.neuron.2016.09.018
P Barttfeld (2015)
10.1073/pnas.1418031112
Proc. Natl Acad. Sci. USA, 112
T Ge (2017)
10.1073/pnas.1700765114
Proc. Natl Acad. Sci. USA, 114
AJ Holmes (2015)
Brain genomics superstruct project initial data release with structural, functional, and behavioral measures
Sci. Data, 2
H Zou, T Hastie (2005)
Regularization and variable selection via the elastic net
J. R. Stat. Soc. B, 67
MW Cole (2014)
10.1016/j.neuron.2014.05.014
Neuron, 83
RL Buckner, FM Krienen, BTT Yeo (2013)
Opportunities and limitations of intrinsic functional connectivity MRI
Nat. Neurosci., 16
BTT Yeo (2015)
Functional specialization and flexibility in human association cortex
Cereb. Cortex, 25
WM Perlstein, CS Carter, DC Noll, JD Cohen (2001)
Relation of prefrontal cortex dysfunction to working memory and symptoms in schizophrenia
Am. J. Psychiatry, 158
CJ Honey (2007)
10.1073/pnas.0701519104
Proc. Natl Acad. Sci. USA, 104
T Ge, AJ Holmes, RL Buckner, JW Smoller, MR Sabuncu (2017)
Heritability analysis with repeat measurements and its application to resting-state functional connectivity
Proc. Natl Acad. Sci. USA, 114
KRA Van Dijk (2010)
10.1152/jn.00783.2009
J. Neurophysiol., 103
RL Buckner (2008)
10.1196/annals.1440.011
Ann. NY Acad. Sci., 1124
MW Cole, DS Bassett, JD Power, TS Braver, SE Petersen (2014)
Intrinsic and task-evoked network architectures of the human brain
Neuron, 83
PS Goldman-Rakic (1988)
Topography of cognition: parallel distributed networks in primate association cortex
Annu. Rev. Neurosci., 11
TO Laumann (2017)
On the stability of BOLD fMRI correlations
Cereb. Cortex, 27
D Ville, J Britz, CM Michel (2010)
EEG microstate sequences in healthy humans at rest reveal scale-free dynamics
Proc. Natl Acad. Sci. USA, 107
E Bullmore, O Sporns (2009)
Complex brain networks: graph theoretical analysis of structural and functional systems
Nat. Rev. Neurosci., 10
Uuml Sakoğlu (2010)
A method for evaluating dynamic functional network connectivity and task-modulation: application to schizophrenia
Magn. Reson. Mater. Phys., 23
DC Glahn (2010)
Genetic control over the resting brain
Proc. Natl Acad. Sci. USA, 107
G Repovs (2011)
10.1016/j.biopsych.2010.11.009
Biol. Psychiatry, 69
NA Crossley (2013)
10.1073/pnas.1220826110
Proc. Natl Acad. Sci. USA, 110
AK Engel (2013)
10.1016/j.neuron.2013.09.038
Neuron, 80
HW Kuhn (1955)
10.1002/nav.3800020109
Nav. Res. Logist. Q., 2
KRA Dijk (2010)
Intrinsic functional connectivity as a tool for human connectomics: theory, properties, and optimization
J. Neurophysiol., 103
X Shen (2017)
10.1038/nprot.2016.178
Nat. Protoc., 12
ML Schölvinck, DA Leopold, MJ Brookes, PH Khader (2013)
The contribution of electrophysiology to functional connectivity mapping
Neuroimage, 80
ES Finn (2015)
Functional connectome fingerprinting: identifying individuals using patterns of brain connectivity
Nat. Neurosci., 18
F de Pasquale (2016)
10.1093/cercor/bhv185
Cereb. Cortex, 26
JT Baker (2014)
10.1001/jamapsychiatry.2013.3469
JAMA Psychiatry, 71
Uuml Sakoğlu (2010)
10.1007/s10334-010-0197-8
Magn. Reson. Mater. Phys., 23
G Deco (2014)
Identification of optimal structural connectivity using functional connectivity and neural modeling
J. Neurosci., 34
X Shen (2017)
Using connectome-based predictive modeling to predict individual behavior from brain connectivity
Nat. Protoc., 12
A Zalesky (2014)
10.1073/pnas.1400181111
Proc. Natl Acad. Sci. USA, 111
DC Van Essen (2005)
10.1016/j.neuroimage.2005.06.058
Neuroimage, 28
MD Fox (2005)
10.1073/pnas.0504136102
Proc. Natl Acad. Sci. USA, 102
J Munkres (1957)
Algorithms for the assignment and transportation problems
J. Soc. Ind. Appl. Math., 5
F Musso, J Brinkmeyer, A Mobascher, T Warbrick, G Winterer (2010)
Spontaneous brain activity and EEG microstates. A novel EEG/fMRI analysis approach to explore resting-state networks
Neuroimage, 52
D Wang (2015)
Parcellating cortical functional networks in individuals
Nat. Neurosci., 18
HW Kuhn (1955)
The Hungarian method for the assignment problem
Nav. Res. Logist. Q., 2
AK Engel, C Gerloff, CC Hilgetag, G Nolte (2013)
Intrinsic coupling modes: multiscale interactions in ongoing brain activity
Neuron, 80
BTT Yeo (2011)
10.1152/jn.00338.2011
J. Neurophysiol., 106
RC Craddock (2013)
Imaging human connectomes at the macroscale
Nat. Methods, 10
M Breakspear (2017)
10.1038/nn.4497
Nat. Neurosci., 20
DN Greve (2009)
10.1016/j.neuroimage.2009.06.060
Neuroimage, 48
CJ Price (2005)
10.1080/02643290442000095
Cogn. Neuropsychol., 22
RN Spreng, WD Stevens, JP Chamberlain, AW Gilmore, DL Schacter (2010)
Default network activity, coupled with the frontoparietal control network, supports goal-directed cognition
Neuroimage, 53
PS Goldman-Rakic (1988)
10.1146/annurev.ne.11.030188.001033
Annu. Rev. Neurosci., 11
C Chang, GH Glover (2010)
Time-frequency dynamics of resting-state brain connectivity measured with fMRI
Neuroimage, 50
EA Allen (2014)
10.1093/cercor/bhs352
Cereb. Cortex., 24
EA Allen (2014)
Tracking whole-brain connectivity dynamics in the resting state
Cereb. Cortex., 24
RL Buckner (2011)
10.1152/jn.00339.2011
J. Neurophysiol., 106
BTT Yeo (2011)
The organization of the human cerebral cortex estimated by intrinsic functional connectivity
J. Neurophysiol., 106
G Deco (2013)
10.1016/j.tins.2013.03.001
Trends Neurosci., 36
A Fornito, ET Bullmore (2015)
Reconciling abnormalities of brain network structure and function in schizophrenia
Curr. Opin. Neurobiol., 30
N Leonardi, WR Shirer, MD Greicius, D Ville (2014)
Disentangling dynamic networks: Separated and joint expressions of functional connectivity patterns in time
Hum. Brain Mapp., 35
J Munkres (1957)
10.1137/0105003
J. Soc. Ind. Appl. Math., 5
M Corbetta, GL Shulman (2002)
Control of goal-directed and stimulus-driven attention in the brain
Nat. Rev. Neurosci., 3
J Britz, D Ville, CM Michel (2010)
BOLD correlates of EEG topography reveal rapid resting-state network dynamics
Neuroimage, 52
JT Baker (2014)
Disruption of cortical association networks in schizophrenia and psychotic bipolar disorder
JAMA Psychiatry, 71
DC Glahn (2010)
10.1073/pnas.0909969107
Proc. Natl Acad. Sci. USA, 107
10.1109/PRNI.2014.6858534
MD Fox (2005)
The human brain is intrinsically organized into dynamic, anticorrelated functional networks
Proc. Natl Acad. Sci. USA, 102
K Shen (2015)
10.1523/JNEUROSCI.4903-14.2015
J. Neurosci., 35
B Rashid (2014)
10.3389/fnhum.2014.00897
Front. Hum. Neurosci., 8
G Deco, VK Jirsa, AR McIntosh (2013)
Resting brains never rest: computational insights into potential cognitive architectures
Trends Neurosci., 36
R Ciric, JS Nomi, LQ Uddin, AB Satpute (2017)
Contextual connectivity: a framework for understanding the intrinsic dynamic architecture of large-scale functional brain networks
Sci. Rep., 7
F Pasquale, S Penna, O Sporns, GL Romani, M Corbetta (2016)
A dynamic core network and global efficiency in the resting human brain
Cereb. Cortex, 26
S Mueller (2013)
10.1016/j.neuron.2012.12.028
Neuron, 77
DC Essen, K Ugurbil (2012)
The future of the human connectome
Neuroimage, 62
H Zou (2005)
10.1111/j.1467-9868.2005.00503.x
J. R. Stat. Soc. B, 67
A Fornito, A Zalesky, C Pantelis, ET Bullmore (2012)
Schizophrenia, neuroimaging and connectomics
Neuroimage, 62
B Biswal, FZ Yetkin, VM Haughton, JS Hyde (1995)
Functional connectivity in the motor cortex of resting human brain using echo-planar MRI
Magn. Reson. Med., 34
PJ Uhlhaas (2010)
10.1038/nrn2774
Nat. Rev. Neurosci., 11
BTT Yeo (2015)
10.1093/cercor/bhu217
Cereb. Cortex, 25
CJ Honey, R Kötter, M Breakspear, O Sporns (2007)
Network structure of cerebral cortex shapes functional connectivity on multiple time scales
Proc. Natl Acad. Sci. USA, 104
B Rashid, E Damaraju, GD Pearlson, VD Calhoun (2014)
Dynamic connectivity states estimated from resting fMRI identify differences among schizophrenia, bipolar disorder, and healthy control subjects
Front. Hum. Neurosci., 8
C Wang (2016)
10.1073/pnas.1523980113
Proc. Natl Acad. Sci. USA, 113
J Britz (2010)
10.1016/j.neuroimage.2010.02.052
Neuroimage, 52
ES Finn (2015)
10.1038/nn.4135
Nat. Neurosci., 18
RM Hutchison (2014)
10.1002/hbm.22583
Hum. Brain Mapp., 35
NA Crossley (2013)
Cognitive relevance of the community structure of the human brain functional coactivation network
Proc. Natl Acad. Sci. USA, 110
R Ciric (2017)
10.1038/s41598-017-06866-w
Sci. Rep., 7
ML Schölvinck (2013)
10.1016/j.neuroimage.2013.04.010
Neuroimage, 80
MD Rosenberg (2016)
10.1038/nn.4179
Nat. Neurosci., 19
E Damaraju (2014)
Dynamic functional connectivity analysis reveals transient states of dysconnectivity in schizophrenia
Neuroimage Clin., 5
E Bullmore (2009)
10.1038/nrn2575
Nat. Rev. Neurosci., 10
C Wang, JL Ong, A Patanaik, J Zhou, MWL Chee (2016)
Spontaneous eyelid closures link vigilance fluctuation with fMRI dynamic connectivity states
Proc. Natl Acad. Sci. USA, 113
RM Hutchison, T Womelsdorf, JS Gati, S Everling, RS Menon (2013)
Resting-state networks show dynamic functional connectivity in awake humans and anesthetized macaques
Hum. Brain Mapp., 34
ME Lynall (2010)
Functional connectivity and brain networks in schizophrenia
J. Neurosci., 30
RM Hutchison (2013)
10.1002/hbm.22058
Hum. Brain Mapp., 34
RM Hutchison (2013)
10.1016/j.neuroimage.2013.05.079
Neuroimage, 80
RC Craddock (2013)
10.1038/nmeth.2482
Nat. Methods, 10
JM Shine (2016)
10.1073/pnas.1604898113
Proc. Natl Acad. Sci. USA, 113
M Corbetta (2002)
10.1038/nrn755
Nat. Rev. Neurosci., 3
PJ Uhlhaas, W Singer (2010)
Abnormal neural oscillations and synchrony in schizophrenia
Nat. Rev. Neurosci., 11
MD Rosenberg (2016)
A neuromarker of sustained attention from whole-brain functional connectivity
Nat. Neurosci., 19
S Mueller (2013)
Individual variability in functional connectivity architecture of the human brain
Neuron, 77
K Shen, RM Hutchison, G Bezgin, S Everling, AR McIntosh (2015)
Network structure shapes spontaneous functional connectivity dynamics
J. Neurosci., 35
VD Calhoun (2014)
10.1016/j.neuron.2014.10.015
Neuron, 84
C Chang (2013)
10.1016/j.neuroimage.2012.11.038
Neuroimage, 68
RL Buckner, JR Andrews-Hanna, DL Schacter (2008)
The brain’s default network: anatomy, function, and relevance to disease
Ann. NY Acad. Sci., 1124
10.1007/978-0-387-21606-5
T Lange, V Roth, ML Braun, JM Buhmann (2004)
Stability-based validation of clustering solutions
Neural Comput., 16
A Zalesky, A Fornito, L Cocchi, LL Gollo, M Breakspear (2014)
Time-resolved resting-state brain networks
Proc. Natl Acad. Sci. USA, 111
O Miranda-Dominguez (2014)
Connectotyping: model based fingerprinting of the functional connectome
PLoS ONE, 9
MW Cole (2014)
10.1177/1073858414525995
Neuroscientist, 20
N Leonardi (2014)
10.1002/hbm.22599
Hum. Brain Mapp., 35
A Fornito (2015)
10.1016/j.conb.2014.08.006
Curr. Opin. Neurobiol., 30
VD Calhoun, R Miller, G Pearlson, T Adali (2014)
The chronnectome: time-varying connectivity networks as the next frontier in fMRI data discovery
Neuron, 84
WR Shirer, S Ryali, E Rykhlevskaia, V Menon, MD Greicius (2012)
Decoding subject-driven cognitive states with whole-brain connectivity patterns
Cereb. Cortex, 22
M Breakspear (2017)
Dynamic models of large-scale brain activity
Nat. Neurosci., 20

Publisher: Springer Journals
Copyright: Copyright © 2018 by The Author(s)
Subject: Science, Humanities and Social Sciences, multidisciplinary; Science, Humanities and Social Sciences, multidisciplinary; Science, multidisciplinary
eISSN: 2041-1723
DOI: 10.1038/s41467-018-03462-y
Publisher site: See Article on Publisher Site

Abstract

ARTICLE DOI: 10.1038/s41467-018-03462-y OPEN The human cortex possesses a reconﬁgurable dynamic network architecture that is disrupted in psychosis 1 1 2 3,4 1 Jenna M. Reinen , Oliver Y. Chén , R. Matthew Hutchison , B.T.Thomas Yeo , Kevin M. Anderson , 4,5 6 7,8 7 6 Mert R. Sabuncu , Dost Öngür , Joshua L. Roffman , Jordan W. Smoller , Justin T. Baker & 1,4,7,9 Avram J. Holmes Higher-order cognition emerges through the ﬂexible interactions of large-scale brain net- works, an aspect of temporal coordination that may be impaired in psychosis. Here, we map the dynamic functional architecture of the cerebral cortex in healthy young adults, leveraging this atlas of transient network conﬁgurations (states), to identify state- and network-speciﬁc disruptions in patients with schizophrenia and psychotic bipolar disorder. We demonstrate that dynamic connectivity proﬁles are reliable within participants, and can act as a ﬁngerprint, identifying speciﬁc individuals within a larger group. Patients with psychotic illness exhibit intermittent disruptions within cortical networks previously associated with the disease, and the individual connectivity proﬁles within speciﬁc brain states predict the presence of active psychotic symptoms. Taken together, these results provide evidence for a reconﬁgurable dynamic architecture in the general population and suggest that prior reports of network disruptions in psychosis may reﬂect symptom-relevant transient abnormalities, rather than a time-invariant global deﬁcit. 1 2 Department of Psychology, Yale University, New Haven, CT 06520, USA. Department of Psychology, Harvard University, Cambridge, MA 02138, USA. Department of Electrical & Computer Engineering, Clinical Imaging Research Centre, Singapore Institute for Neurotechnology & Memory Network Programme, National University of Singapore, Singapore 117583, Singapore. Athinoula A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital Harvard Medical School Charlestown, MA 02129, USA. School of Electrical and Computer Engineering and Meinig School of Biomedical Engineering, Cornell University, Ithaca, NY 14853, USA. Department of Psychiatry, Psychotic Disorders Division, McLean Hospital, Belmont, MA 02478, 7 8 USA. Department of Psychiatry, Massachusetts General Hospital, Harvard Medical School, Boston, MA 02114, USA. Psychiatric Neuroimaging Research Division, Massachusetts General Hospital, Charlestown, MA 02129, USA. Department of Psychiatry, Yale University, New Haven, CT 06511, USA. Correspondence and requests for materials should be addressed to A.J.H. (email: [email protected]) NATURE COMMUNICATIONS (2018) 9:1157 DOI: 10.1038/s41467-018-03462-y www.nature.com/naturecommunications 1 | | | 1234567890():,; ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-03462-y he human cortex is organized into large-scale networks with schizophrenia exhibit reduced dynamism, spending longer peri- 1–3 complex patterns of functional coupling .While sub- ods of time within single brain states and demonstrating a less- Tstantial progress has been made delineating aspects of this variable repertoire of dynamic network conﬁgurations . Relative intricate architecture, research in this domain has traditionally to healthy populations, patients with schizophrenia dwell more in relied on static analytic approaches that assume stable patterns of network conﬁgurations typiﬁed by reduced large-scale con- connectivity across time . However, the brain is not a static organ, nectivity, while also showing muted cross-network negative cor- and time-varying proﬁles of network connectivity are evident across relations, for instance, between default and other networks . 5,6 a broad range of task states and during periods of unconstrained Dynamic analyses of network function may distinguish clinical 7–9 10,11 rest ,(butsee ). Variability in the expression of dynamic groups, providing information that is inaccessible through static 12–15 17 brain states links to cognition , learning, and the presence of connectivity analysis . The incorporation of time-resolved ana- 16–18 psychiatric illnesses like schizophrenia that are characterized lyses of network function could provide a more sensitive or by a breakdown in cortical information processing. Despite the speciﬁc marker of the disease than static approaches, potentially importance of understanding the relations that link temporal associating with the illness course and/or the presence of distinct descriptions of brain function with behavior, the core features of symptom proﬁles. As impairments in attention, learning, and dynamic network organization, the stability of individually speciﬁc executive functioning are common across neuropsychiatric dis- signatures of time-resolved connectivity, and their associated rele- orders , research in this domain could provide novel insights vance to the disease, remain unclear. into the biology of illness as we work to predict the onset, track Patterns of spontaneous brain activity and connectivity have disease states, and optimize treatment response. been a focused topic of study in electrophysiological recordings at Here, we applied a sliding-window approach to characterize the level of cells, local ﬁelds, and surface electroencephalograms the reconﬁgurable architecture of large-scale brain networks in a (EEGs) . While formal biophysical models linking the activity of sample of healthy young adults and individuals with psychotic neuronal populations with large-scale brain systems have yet to illness . Among healthy adults, we demonstrate that the resulting be established , features of oscillatory neural activity, such as dynamic connectivity proﬁles are reliably expressed across scans those observed at high temporal resolutions, may be reﬂected in and visits, acting as a biological signature that can identify speciﬁc hemodynamic ﬂuctuations . Transient quasi-stable patterns individuals within a larger group. Patients with psychotic illness detected in EEGs (microstates), for example, spatially correlate exhibited intermittent disruptions within cortical association with network patterns observed through intrinsic functional networks previously associated with the disease. Individual con- 21–23 connectivity magnetic resonance imaging . Although macro- nectivity proﬁles within speciﬁc brain states predicted the pre- scale network dynamics have been linked with changes in arou- sence of active psychotic symptoms, operationalized as meeting 13 14,15 24 sal , attention , and autonomic activity , the biological bases clinician-rated symptomatic diagnostic criteria with the presence and behavioral signiﬁcance of these spontaneous ﬂuctuations of delusions and/or hallucinations in the past month . This remain unresolved. There are at least two reasons for this lack of property of dynamic network function generalized to a held-out consensus. First, prior studies of functional network dynamics sample of patients. These collective results suggest a key role for have largely focused on single clustering solutions in isolation, functional network dynamics in human cognition, and highlight choosing a ﬁxed number of temporal states a priori. As a result, a how speciﬁc breakdowns in time-varying proﬁles of network functional atlas of transient network conﬁgurations, or brain connectivity may link with the presence of distinct symptom states, that are present throughout the population has not been proﬁles in psychiatric illnesses. fully characterized. Second, analyses of transient network func- tion have principally focused on establishing the existence of a general architecture of dynamic connectivity shared across the Results population. Static patterns of intrinsic connectivity are heri- Core dynamic network conﬁgurations. The human brain can 25,26 table and act as a trait-like ﬁngerprint that can accurately exhibit a multitude of possible transient connectivity patterns 27–29 identify speciﬁc people from a large group . There is a reason comprised of varying network conﬁgurations. As an initial step in to believe that a substantial portion of the dynamic connectome understanding temporal shifts in this dynamic architecture over may be unique to each individual. Despite the importance of time, we aimed to identify a core or canonical set of transient establishing if time-resolved network function acts as an indivi- brain states conserved across individuals. First, we coupled a dual speciﬁc signature, the extent to which dynamic connectivity population atlas of large-scale cortical networks and a sliding- 7,8,40–42 proﬁles possess intra-subject reliability and capture inter-subject window approach (11 time points; width = 33 s ) to esti- variability has yet to be determined. mate time-varying connectivity proﬁles within resting-state scans Although time-resolved analyses of network organization have from 1919 healthy young adults (Fig. 1; see Methods for infor- largely focused on the study of healthy populations, there is mation on data acquisition and preparation) . These data were preliminary evidence to suggest that network dysfunction in collapsed across participants before we applied k-means cluster- psychosis may emerge through alterations in the core dynamic ing to estimate solutions yielding from 2 to 20 brain states. Note 16–18 architecture of the brain . Psychotic illnesses (including that the brain states deﬁned through k-means clustering do not schizophrenia, schizoaffective disorder, and bipolar disorder with necessarily have sharp boundaries, cleanly separating them from psychotic features) are marked by broad disruptions across cor- other network conﬁgurations. Rather, k-means clustering identi- tical association networks, potentially contributing to widespread ﬁes sets of time-varying network conﬁgurations with common 30–34 changes in information processing . By one view, impaired features, grouping them into clusters that are more similar to each network connectivity in patient populations might be time- other than to conﬁgurations in other clusters. The associated invariant, emerging through stable deﬁcits in brain function. An groupings can vary across clustering solutions as their complexity alternate possibility is that aspects of the functional impairments increases, revealing intermediate network conﬁgurations between observed in psychosis reﬂect transient abnormalities pre- more geographically separated states. To estimate the viability of ferentially evident during the expression of particular network the resulting solutions, we analyzed the population-level con- 17,35 39,44,45 conﬁgurations . Converging evidence suggests that aberrant sistency of each clustering algorithm . Consistent with an oscillatory activity may link to core symptoms of schizophrenia, expansion in the solution space, the clustering became less stable including the presence of hallucinations . Patients with as the number of estimated brain states increased (Fig. 1d and 2 NATURE COMMUNICATIONS (2018) 9:1157 DOI: 10.1038/s41467-018-03462-y www.nature.com/naturecommunications | | | NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-03462-y ARTICLE Supplementary Figure 1). Analyses indicated relative stability in a broad survey of the solution space. In the present data, these state solutions 2–8, with 2-, 4-, 5-, and 8-state solutions displaying solutions captured signiﬁcant aspects of dynamic variation in points of increased stability. The stable nature of the observed network connectivity. However, the focus on 2–8 brain states state solutions was robust to changes in data quality (see Meth- should not be taken to imply that meaningful properties are ods, Noise constraints). As such, our analyses going forward focus absent in alternative solutions. on clustering solutions containing 2–8 dynamic states to provide Brain states exhibit hierarchical features. Previous studies of Dorsal functional network dynamics have largely focused on single clustering solutions in isolation, choosing a ﬁxed number of states AP a priori. The relations between these isolated dynamic network Lateral conﬁgurations and other possible solutions remain unclear. To D address this question, we employed a matching analysis exam- Frontal Parietal ining the preservation or fractionation of individual brain states as the complexity of our clustering solution increased. An exhaustive search was performed exploring the scenario that two Medial states in solution S + 1 were subdivisions of a state within solu- tion S. Hungarian matching (Supplementary Figure 2) was used to determine which two-state combination in S + 1 was best matched to the S-state solution by minimizing network dissim- Ventral PA ilarity. As solution complexity increased, a hierarchical structure was observed, with some network conﬁgurations preserved across levels (Fig. 2; see Methods, Hierarchy analysis) and selected hybrid states in solution S breaking into substates within solution S + 1. A single state, termed state A, was evident across state solutions 2 through 7, not splitting to produce a hybrid substate Participant until the solution complexity increased to eight states. These time courses analyses demonstrate the stable expression of canonical network conﬁgurations across a variety of state solutions (2–8). Of note, when larger numbers of states were considered, increasing solu- tion complexity was reﬂected in the hierarchical fractionation of Windowed correlation particular states into substates. matrices Static analyses of network function suggest that heteromodal association cortices are more functionally variable than the unimodal cortex across the population . These aspects of the cortex, and its associated networks, are implicated in a host of Aggregrate complex cognitive functions and overlap with regions that data across participants predict individual differences in behavioral performance .We then explored if the heteromodal association cortex also exhibits heightened dynamic variability, as reﬂected in ﬂuctuating k=2 connectivity conﬁgurations over time. Analyses examining the variance of mean network connectivity across the state solutions K-means revealed evidence for dissimilar patterns of network expression clustering across the dynamic states (ANOVA of coefﬁcient of variance with k=3 k=4 k=5 Fig. 1 Detecting multiple functional connectivity states using a sliding- window approach. a The functional network organization of the human cerebral cortex is revealed through intrinsic functional connectivity. Colors reﬂect regions estimated to be within the same network determined based on the 17-network solution from Yeo et al. . The map is displayed for multiple views of the left hemisphere in Caret PALS space . b Correlation matrices are computed across regions from windowed portions (width= 33 s) of each participant’s component time series (n = 1919), and aggregated across the full sample. c K-means clustering was applied to Stability analysis identify repeated patterns of connectivity (brain states). d Instability of the 0.6 clustering algorithm is plotted as a function of the number of estimated 0.5 states (2–20). As expected based on increasing solution space (complexity), instability was greater with increasing the number of 0.4 estimated states per solution. The local minima of the graphs observable at 0.3 the 2-, 4-, 5-, and 8-state solutions indicate the number of states that can 0.2 be stably estimated by the selected clustering algorithm with the present data. Resampling over time (sliding windows) and across space (regions of 0.1 interest) yields comparable results (see Supplementary Figure 1). In this 0.0 article, we focus on state solutions 2 through 8 to provide a broad survey of 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 the solution space Number of States NATURE COMMUNICATIONS (2018) 9:1157 DOI: 10.1038/s41467-018-03462-y www.nature.com/naturecommunications 3 | | | Instability Sal/ Limbic DorsAttn Visual Default Control VentAttn SomMot –0.8 0.8 Z (r) ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-03462-y Brain states exhibit hierarchical features Global state 2 2 A B Two states 3 3 B C Three states 4 4 4 B C D Four states 5 5 5 5 B C D E Five states 6 6 6 6 6 6 A B C D E F Six states 7 7 7 7 7 7 7 C D E F G A B Seven states 8 8 8 8 8 8 8 8 A B C D E F G H Eight states Fig. 2 Brain states exhibit hierarchical features across distributed networks. Correlation matrices for state solutions 1 through 8 are shown for each region . To quantify hierarchical relationships, an exhaustive search was performed to examine the scenario that two states of the (S + 1) state solution were subdivisions of a component of the S-state solution. Hungarian matching was used to determine which two-state combination in S + 1 was best matched to the S-state solution (see Supplementary Figure 4). Values reﬂect z-transformed Pearson correlations between every region and every other region. DorsAttn indicates dorsal attention, Sal salience, SomMot somatomoto, VentAttn ventral attention state solutions treated as repeated measures, F = 44.11, the clustering solutions, the observed brain states exhibited quasi- p ≤ 0.001). Consistent with reports of heightened time-resolved stable expression, with all states maintaining nonzero dwell times network ﬂexibility within the association cortex , the greatest (ps ≤ 0.001). Reﬂecting the transition probability analyses detailed between-state variability was found in aspects of default, above, dwell times were nonuniform across brain states. Partici- attention, and control networks (Supplementary Figure 3). pants dwelled most in attractor state A (ps ≤ 0.001) relative to all other states. Several core patterns that distinguish brain states across Evidence for an attractor state. In addition to characterizing clustering solutions were evident. For ease of interpretability, the proﬁles of network connectivity across clustering solutions, the four-state solution was selected to characterize these features brain dynamics can be studied in terms of the relative likelihood of dynamic connectivity (Fig. 4). As noted in the Hungarian of transitions occurring among locally stable states over matching analyses detailed above (Fig. 2), state A most closely 7,16,17 time . To assess this feature of temporal organization, we resembled the global state revealed through traditional static examined the probability that participants would shift from a analyses of network function. Across network conﬁgurations, given state S to a different state (transition probability), as well as state A was typiﬁed by a relatively ﬂattened proﬁle of connectivity the probability that they would remain in state S. Window-by- (Fig. 4a). In line with this muted connectivity proﬁle, the window estimates were created for each state solution by expression of state A became increasingly frequent in the second matching the participant-speciﬁc connectivity matrices to the relative to the ﬁrst half of the scan across each clustering solution group atlas of brain states. This generated a vector of 110 (2–8 states; ts ≥ −6.07, ps ≤ 0.001), potentially linking with 14,15 expressed states for each participant. As solution complexity shifts in arousal and vigilance . increased, participants were more likely to transition into (ps ≤ The remaining states varied markedly from the state A 0.001), and remain in state A (all test states 2–5 ps ≤ 0.001; states connectivity pattern with differences evident across states both 6–8 ps ≥ 0.05; Fig. 3), termed the “attractor state.” Dwell time was within and between functional networks (Fig. 4b). Relative to 4 4 4 4 calculated as the percent of the total time a given participant state A , states B ,C , and D were characterized by increased expressed state S relative to the total time in states not-S. Across expression of default A (see Fig. 1a for network topographies), 4 NATURE COMMUNICATIONS (2018) 9:1157 DOI: 10.1038/s41467-018-03462-y www.nature.com/naturecommunications | | | Visual DorsAttn Sal/ Limbic SomMot VentAttn Control Default NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-03462-y ARTICLE State S+1 3 3 3 4 4 4 4 5 5 5 5 5 6 6 6 6 6 6 7 7 7 7 7 7 7 8 8 8 8 8 8 8 8 A B C A B C D A B C D E A B C D E F A B C D E F G A B C D E F G H 3 4 5 6 7 8 A A A A A A 3 4 5 6 7 8 B B B B B B 3 4 5 6 7 8 C C C C C C 4 5 6 7 8 D D D D D 5 6 7 8 E E E E 6 7 8 Transition probability F F F 7 8 0 0.45 G G 1.00 0.92 0.84 0.76 0.68 0.60 2 2 3 3 3 4 4 4 4 5 5 5 5 5 6 6 6 6 6 6 7 7 7 7 7 7 7 8 8 8 8 8 8 8 8 A B A B C A B C D A B C D E A B C D E F A B C D E F G A B C D E F G H Two Three Four Five Six Seven Eight Number of states Fig. 3 Evidence for an attractor state. a When transitioning between states, participants display an increased probability of entering state A. The probability of transitioning to each state is shown for all nonredundant possible combinations across state solutions 3–8 (1, 2 are obligated). Each row sums to 1. b State A displays the increased probability of remaining in the same state for solutions 2–5(ps ≤ 0.001). The graph reﬂects the tendency for states in solutions 2–8 to show steady-state behavior, or the probability of remaining in a given state from time point t to time point t+1 with state B also marked by heightened within-network visual Network dynamics are a marker of individual differences. Static system and default B connectivity. State B notably differed from descriptions of brain functional organization act as a biological other network conﬁgurations in its recruitment of dorsal ﬁngerprint that can identify speciﬁc individuals within a larger 4 4 27–29 attention network A. Conversely, states C and D displayed group . We aimed to determine whether the observed increased correlations within the ventral attention network. There dynamic states were reliably expressed in a manner that would was also striking variation in the coupling of the frontoparietal allow us to characterize unique within-subject proﬁles of transient network, with state C primarily characterized by increased network organization. To this end, we ﬁrst assessed within- connectivity in the control B network. session reliability. For participants with two bold runs (n = 1341) 4 4 4 States B ,C , and D showed marked departures from the off- in the same scanning session, correlation matrices were con- diagonal coupling evident in state A and the global state (Fig. 2). catenated for each individual brain state, and a Pearson correla- The negative correlations between the default and somatomotor tion was generated for run 1 relative to run 2. T-tests were used to networks were attenuated in state B and were more evident in compare the distributions of the resulting r values within and 4 4 4 4 states C and D . While states B and C displayed heightened across individuals. For each state, we observed greater within- correlations linking default and control networks, state D participant similarity for same-day scans (rs ≥ 0.49) compared to exhibited increased connectivity between default and the salience, between participants (rs ≤ 0.23, 2–8 states; ps ≤ 0.001; see Meth- attention, and somatomotor systems. Additionally, state C , and ods, Individual identiﬁcation analyses; Supplementary Figure 4). to a lesser extent state D , displayed increased cross-network We then examined a cohort of participants with two scan visits connectivity between the attention, somatomotor, and visual collected on different days (≤6 months apart; mean = 63.35 ± networks. Together, these varying network conﬁgurations 48.10 days; n = 79). Analyses demonstrated consistent within- demonstrate nonrandom departures from connectivity patterns subject dynamic state expression across visits (Fig. 5a; ps ≤ 0.001). observed in the global state. There is a strong correspondence Suggesting relatively stable intra-subject reliability across time, between the structure of intrinsic and extrinsic (task-evoked/ the observed within-participant similarity of the expressed brain coactivation) networks of the human brain, suggesting that the states (Fig. 5a) did not vary as a function of the number of days topological characteristics of the brain at rest are closely linked to between participant visits (absolute value of rs ≤ 0.21; ps ≥ 0.09). 12,48 27 cognitive function . The transient expression of integrated Permutation tests were used to assess if dynamic connectivity network conﬁgurations may enable fast and accurate cognitive proﬁles act as a signature that can accurately identify individual task performance . Emerging evidence suggests that an indivi- participants from a larger group. Correlation matrices from visit 1 dual’s unique proﬁle of dynamic network connectivity could and visit 2 were iteratively shufﬂed across participant labels and provide novel insights into the study of behavioral variability examined relative to their correct pairing. Identiﬁcation was across both health and disease. In this regard, a crucial step is the considered correct if the true visit 1 and 2 pair were maximally characterization of intra-subject reliability and inter-subject similar to each other. Correct identiﬁcation was evident within variability of the observed proﬁles of time-varying brain participants relative to the shufﬂed list across all state solutions organization across the population. (ps ≤ 0.01; see Fig. 5b), suggesting that it is possible, with high NATURE COMMUNICATIONS (2018) 9:1157 DOI: 10.1038/s41467-018-03462-y www.nature.com/naturecommunications 5 | | | State S Stationary probability ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-03462-y Mean within network connectivity Default D Default C Visual cent 1.0 Default B Visual peri Default A SomMotor A State A State B SomMotor B Control C State C State D Dorsal attention A Control B Dorsal attention B Control A Ventral attention Limbic Salience Percent deviation from state A Default D Default C Visual cent 0.30 Default B Visual peri SomMotor A Default A 4 4 State A - State B 4 4 SomMotor B Control C State A - State C 4 4 State A - State D Dorsal attention A Control B Dorsal attention B Control A Ventral attention Limbic Salience Fig. 4 Proﬁles of within-network connectivity across the four-state solution. a The mean within-network connectivity is shown across each state of the four-state solution (n = 1919). Values reﬂect z-transformed Pearson correlations. b Percent deviation from state A . Values reﬂect the percent change in 4 4 4 mean network connectivity of states B –D , relative to state A accuracy, to identify an individual from a large group of participant. For simplicity, we only consider the four-state solu- participants solely on the basis of their dynamic connectivity tion, focusing on four states due to the stability of the associated proﬁle. Dwell time was also more similar within, relative to clustering solution (Supplementary Figure 1) and the relatively between, participants for scans on the same day and across visits high within-participant reliability (Fig. 5 and Supplementary (all ps ≤ 0.01). These ﬁndings provide the ﬁrst evidence to suggest Figure 4). Importantly, while we discuss the four-state solution, that, like static characterizations of intrinsic connectivity, an meaningful properties are likely present in other dynamic net- individual’s dynamic brain function represents a unique and work conﬁgurations. We matched each windowed participant- reliable biological signature, highlighting the potential use of speciﬁc matrix to the dynamic states established in our previous intrinsic network dynamics as a neural marker of individual analyses and calculated mean correlation matrices across both differences across both health and disease. groups. Following the correction for nuisance variables (motion, age, sex, handedness, and scanner bay), group differences were calculated for each of the four states for dwell times and network Preferential state and network disruptions in psychosis. connectivity [false-discovery rate (FDR), α ≤ 0.05]. Impairments in the integration and processing of information Altered intrinsic connectivity in patients with psychosis across large-scale distributed brain networks are thought to mark observed in traditional static analyses may reﬂect both impaired psychotic disorders (including schizophrenia, schizoaffective functions within a speciﬁc brain state and the reduced tendency 30,50,51 disorder, and psychotic bipolar disorder) . Converging to enter particular network conﬁgurations over time. Analyses of evidence suggests that these disruptions may reﬂect alterations in dwell times did not reveal group variability in state A (t = 8,16,17,35 the temporal dynamics of brain function . We next 1.12, p = 0.26; Fig. 6a). However, patients displayed increased examined the expression of time-varying network conﬁgurations dwell time in state B (t = −2.8, p ≤ 0.005) and nominally in a cohort of patients with psychosis (n = 179) and a demo- signiﬁcant increases in state D (t = −2.13, p ≤ 0.03). Notably, graphically and data-quality-matched comparison sample (n = patient dwell times were reduced in state C (Fig. 5; t = 2.98, 369). To evaluate the integrity of dynamic brain functions in these p ≤ 0.003), which is characterized by increased frontoparietal groups, we created windowed correlation matrices for each control network connectivity relative to other states (Fig. 4). 6 NATURE COMMUNICATIONS (2018) 9:1157 DOI: 10.1038/s41467-018-03462-y www.nature.com/naturecommunications | | | NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-03462-y ARTICLE 0.65 0.52 Group 0.39 Within Between 0.26 0.13 0.00 2 2 3 3 3 4 4 4 4 5 5 5 5 5 6 6 6 6 6 6 7 7 7 7 7 7 7 8 8 8 8 8 8 8 8 A B A B C A B C D A B C D E A B C D E F A B C D E F G A B C D E F G H Two Three Four Five Six Seven Eight Number of states 1.00 0.80 Match rank st 0.60 nd rd 0.40 th th 5 + 0.20 0.00 2 2 3 3 3 4 4 4 4 5 5 5 5 5 6 6 6 6 6 6 7 7 7 7 7 7 7 8 8 8 8 8 8 8 8 A B A B C A B C D A B C D E A B C D E F A B C D E F G A B C D E F G H Two Three Four Five Six Seven Eight Number of states Fig. 5 Brain states are reliably expressed and can be used to identify individual participants. a Mean rho values within and between subjects across visits for all stable dynamic states. Correlation matrices from participants who had at least two scans within 6 months of each other (n = 79) were concatenated, and a Pearson correlation was computed for scan 1 relative to scan 2. b Permutation and correlation analyses revealed that dynamic states are consistent within individuals over time. Identiﬁcation rank accuracy is shown for each state in a subgroup of participants scanned on two different days within 6 months (n = 79). Rank denotes the order of similarity of a participant’s visit 1 to their visit 2, with 1 being a perfect identify match Analyses of region-to-region correlation strength across the Select network conﬁgurations mark active psychotic symptoms. four-state solution yielded sparse group differences in states B Together, the analyses above suggest that time-varying network and D . Conversely, widespread group differences were observed conﬁgurations capture stable aspects of inter-subject variability 4 4 4 in states A and C (Fig. 6). Analyses of state A revealed reduced (ﬁngerprints), while also serving to mark the presence of psy- within-network correlations across the cortex, including in chiatric illness. These ﬁndings suggest that dynamic approaches default and frontoparietal networks. The largest-magnitude may provide critical information that can help predict the pre- differences in state C were localized to the frontoparietal control sence of distinct clinical proﬁles within individuals, revealing 18 4 16,17 network (Figs. 6 and 7). Replicating prior reports , states A and symptom-relevant features of the disease . To determine C were associated with less negative, or muted, correlations whether individual differences in time-varying proﬁles of con- linking the aspects of the association cortex to other networks in nectivity are relevant to clinical symptomatology, we investigated psychotic illness. the extent to which speciﬁc brain states can identify patients who 4 4 Follow-up analyses of state A and C revealed evidence for experienced active psychotic symptoms at the point of initial both state-general and state-preferential impairments in network assessment. To this end, we examined clinician reports of the connectivity (Fig. 7 and Supplementary Table 1). Within the current psychotic symptoms in our patient sample, assessed frontoparietal control network, patients exhibited distributed through the DSM-IV (SCID) clinician-rated presence of delusions 4 4 38 deﬁcits across states A (state A control B and C: ts ≥ 4.12, and/or hallucinations in the past month , and selected partici- 4 4 ps ≤ 0.001) and C (state C control B: t = 4.85, p ≤ 0.001; pants who expressed all four brain states (n = 130). Using a Fig. 8a–b). All other subnetworks failed to pass multiple- machine-learning-based framework, elastic net logistic regression, comparison correction (Bonferroni p ≤ 0.05; ps ≥ 0.01). Conver- we demonstrated that the presence of active psychotic symptoms sely, reduced default network connectivity in psychosis was can be predicted based solely on the dynamic connectivity proﬁle 4 4 observed in default B in state A (state A default B: ts = 6.08, of previously unseen individuals (Methods, Prediction of active 52,53 p ≤ 0.001; default A, C, and D: ts ≤ 1.69, ps ≥ 0.09; Fig. 8c–d). psychotic symptoms) . First, training a model in 91 partici- No group differences in the default network survived correction pants (70% of the available patient sample), we used the edge 4 4 4 for multiple comparison when considering state C (state C strength from the dynamic network conﬁgurations of states A , 4 4 4 default B: ts = 2.91, p ≥ 0.01; default A, C, and D: ts ≤ 1.31, B ,C , and D as candidate features upon which we conducted 535 535 ps ≥ 0.19). These analyses suggest that prior observations of variable selection. The extracted features were then used to pre- default and control network disruptions in psychosis may reﬂect dict the presence of active psychotic symptoms in individual temporally speciﬁc impairments, preferentially manifesting dur- participants, using leave-one-subject-out cross-validation. The ing the expression of transient conﬁgurations that recruit the ﬁtted model was prospectively applied to edge strength of the function of these networks. dynamic network conﬁguration from a held-out set of NATURE COMMUNICATIONS (2018) 9:1157 DOI: 10.1038/s41467-018-03462-y www.nature.com/naturecommunications 7 | | | Percent matching Mean correlation ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-03462-y 0.45 Healthy comparison Psychotic Illness 0.38 0.31 p ≤ 0.05 p ≤ 0.005 0.24 p ≤ 0.01 0.17 0.10 4 4 4 4 State A State B State C State D 4 4 4 4 A B C D Healthy comparison –0.80 0.80 Z(r) 4 4 4 4 A B C D Psychotic illness 4 4 4 4 A B C D Functional connectivity differences –0.15 0.15 Z(r) –Z(r) comparison patients Fig. 6 Psychotic illness links to state-speciﬁc reductions in connectivity. a Bar graphs display the percent of time healthy comparison participants and individuals with psychotic illness dwell in each state of the four-state solution. b–c The 2D grids (b–c) display the complete coupling architecture of the cerebral cortex measured at rest for each group of participants. d Differences were obtained by an analysis of variance of z-transformed Pearson correlation values after linear regression of the effects of age, sex, race, ethnicity, and handedness [r(z) – r(Z) ]. The bottom panel Comparison Patient group shows signiﬁcant group differences at false-discovery rate q ≤ 0.05 participants, and yielded a scalar probability measure for each D were 0.39 (p = 0.74), 0.50 (p = 0.43), and 0.56 (p = 0.26), individual, which constituted the predicted likelihood of active respectively. psychotic symptoms. Next, we demonstrated that predictive symptom models Consistent with an increased tendency to enter state B in derived from dynamic network conﬁgurations can generalize to patients, relative to other network conﬁgurations (Fig. 6a), state data from novel individuals, applying the predictive network B speciﬁcally predicted the presence of active psychotic models in the left-out patient sample (30% of the available patient symptoms in the training set (positive network predictive sample). Here, without further ﬁtting or modiﬁcation of the network: AUC = 0.80, p ≤ 0.001, permutations = 5000; initial models, we tested the neurological signatures identiﬁed in Fig. 9a). As a comparison, the AUC values using a positive the training set for the predication of active psychotic symptoms 4 4 4 network of states A ,C , and D were 0.60 (p = 0.16), 0.68 in a held-out group previously unseen subjects (n = 39). The (p = 0.05), and 0.52 (p = 0.37), respectively. A similar, negative-network models did not generalize to the held-out although subtler, predictive proﬁle was observed for the negative sample (AUCs < 0.66, ps > 0.13). The previously identiﬁed 4 4 networks in the training set where state B predicted the presence positive-network model for state B served as a generalizable of active psychotic symptoms (AUC = 0.724, p ≤ 0.05). The predictor of active psychotic symptoms in an independent sample 4 4 remaining negative network AUC values for states A ,C , and (AUC = 0.74, p ≤ 0.05, permutation n = 5000; Fig. 9b). As a 8 NATURE COMMUNICATIONS (2018) 9:1157 DOI: 10.1038/s41467-018-03462-y www.nature.com/naturecommunications | | | Percent dwell time NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-03462-y ARTICLE positive-network model for active psychosis in state B served as a State A generalizable predictor of PANSS positive-scale symptom severity Default D in the independent sample (AUC = 0.72, p ≤ 0.05). The observed Visual cent Default C effect was preferential to the state B network model. The 4 4 0.24 associated AUC values derived across dynamic states A ,C , and Visual peri Default B D for positive out-sample predication were 0.51 (p = 0.46), 0.66 (p = 0.10), and 0.33 (p= 0.90). As above, the negative-network SomMotor A Default A models did not generalize to the PANSS positive-scale scores in the held-out sample (AUCs < 0.67, ps > 0.09). Suggesting a degree of symptom speciﬁcity, the PANSS negative (positive network: SomMotor B Control C AUC = 0.41, p = 0.83; negative network: AUC = 0.52, p = 0.42) and general psychopathology scale (positive network: AUC = 0.62, p = 0.21; negative network: AUC = 0.41, p = 0.74) scores Dorsal attention A Control B were not predicted by the state B model, or the models resulting from the other states (positive network: AUCs < 0.51, ps > 0.48; Dorsal attention B Control A negative network: AUCs < 0.59, ps > 0.18). These results provide preliminary evidence that the presence of clinical symptoms, in Ventral attention Limbic this case, active psychosis, may be associated with predictable Salience patterns within an individual’s time-varying network proﬁle, suggesting that dynamic approaches have potential utility for State C predicting a wide range of clinical symptoms and cognitive Default D abilities. Visual cent Default C 0.24 Visual peri Default B Discussion The functional coupling of cortical regions varies in response to SomMotor A Default A explicit task demands and in conjunction with shifts in arou- 13 14,15 24 sal , attention , and markers of autonomic activity . Here, using a sliding-window approach on resting-state imaging data, SomMotor B Control C we demonstrated that cortical brain networks possess a time- varying organizational structure with hierarchical properties. Select network conﬁgurations fractionated into substates in an Dorsal attention A Control B ordered manner, and a global attractor state (termed state A) was evident across clustering solutions (2–8 states). Suggesting that Dorsal attention B Control A proﬁles of dynamic network connectivity may link to behavioral differences in health and disease, the observed brain states Ventral attention Limbic reﬂected individually speciﬁc signatures, or ﬁngerprints, of time- Salience resolved connectivity that were unique and reliable within par- Fig. 7 State-speciﬁc and state-general proﬁles of network dysregulation in ticipants across both scans and independent visits separated by up psychosis. a, b Polar plots show percent difference in mean network to 6 months. Extending upon prior static analyses of network 27–29 connectivity between healthy comparison participants and individuals with function , we found that it is possible, with high accuracy, to 4 4 psychotic illness for (a) state A and (b) state C (range: –6 to 24%). The identify speciﬁc individuals from a large group of participants black hexadecagon reﬂects the mean network correlations for the healthy solely on the basis of their proﬁles of dynamic connectivity. comparison sample, set to zero. Values outside the hexadecagon reﬂect Patients with schizophrenia and psychotic bipolar disorder decreased correlation strength for patients, relative to the healthy exhibited state-speciﬁc, intermittent disruptions within cortical comparison sample association networks believed to mark the presence of psychotic illness. Finally, our analyses revealed that individual variability in 4 4 comparison, the AUC values derived across dynamic states A , select network conﬁgurations (state B ) can be used to predict the 4 4 C , and D for out-sample predication were 0.56 (p = 0.32), 0.63 presence of active psychotic symptoms in novel participants. (p = 0.18), and 0.29 (p = 0.92). Highlighting that dynamic Together, these results highlight the potential to discover indi- analyses may reveal information hidden in traditional static vidualized dynamic network proﬁles that are predictive of cog- analyses of network function, models derived from the static or nitive abilities and clinical symptoms across health and disease. global state were not predictive of active psychosis in either the Spontaneous brain activity is constrained, but not fully deter- 55–59 training (positive network: AUC = 0.62, p = 0.12; negative net- mined, by structural connectivity . This raises the possibility work: AUC = 0.64, p = 0.09) or test samples (positive network: that a quasi-stable functional architecture may anchor on ana- AUC = 0.49, p = 0.53; negative network: AUC = 0.24, p = 0.96). tomic connectivity, with transient network conﬁgurations Patients expressing active psychotic symptoms, relative to reﬂecting the inﬂuence of momentary cognitive processes, those without, presented with increased positive (present: 19.90 ± environmental demands, or other biological information. Sup- 5.67, absent: 8.88 ± 2.20; t = 7.91, p ≤ 0.001), negative (present: porting this conjecture, while functional coupling occurs in the 13.98 ± 7.76, absent: 10.18 ± 3.93; t = 1.98, p ≤ 0.05), and absence of cognition , under general anesthesia, brain activity general psychopathology (present: 31.46 ± 7.81, absent: 26.53 ± exhibits a reduction in spontaneous transitions across network 7.74; t = 2.43, p ≤ 0.05) symptoms (Supplemental Fig. 7)as conﬁgurations , settling into a restricted dynamical repertoire assessed through the positive and negative syndrome scale that closely resembles a ﬁxed network deﬁned by structural 54 56 (PANSS) . To establish the speciﬁcity of the predictive model connectivity . The present analyses indicate the expression of for the presence of psychotic features, we examined PANSS scale core, or canonical, time-varying network conﬁgurations that scores in the left-out patient sample. The previously identiﬁed separate in an ordered manner across increasingly complex NATURE COMMUNICATIONS (2018) 9:1157 DOI: 10.1038/s41467-018-03462-y www.nature.com/naturecommunications 9 | | | ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-03462-y a c Default D Control C Default C Control B Default B Control A Default A b 4 4 d 4 4 State A State C State A State C Default D Control C Default C Control B Default B Control A Default A Healthy Healthy comparison comparison Psychotic Psychotic illness illness –0.8 0.8 –0.8 0.8 Z(r) Z(r) Functional Functional connectivity connectivity differences differences –0.15 0.15 –0.15 0.15 Z(r) –Z(r) Z(r) –Z(r) patients patients comparison comparison 4 4 Fig. 8 Psychosis is associated with reduced network connectivity in states A and C . a The colored aspects of the cortex reﬂect regions estimated to be within the A, B, and C aspects of the frontoparietal control network . b Functional connectivity matrices for the 50 left and right hemisphere regions of the 4 4 frontoparietal control network shown for the healthy comparison and patient groups in states A and C . Functional connectivity difference matrices were obtained by an analysis of variance of z-transformed Pearson correlation values after linear regression of the effects of age, sex, handedness, and scanner bay. Group differences signiﬁcant at false-discovery rate q ≤ 0.05 are shown in each panel to the lower right of the unthresholded matrix. c The colored aspects of the cortex reﬂect regions estimated to be within the A, B, C, and D aspects of the default network. d Corresponding connectivity matrices are 4 4 shown in 52 left and right hemisphere regions of the default regions for states A and C . Reduced default network connectivity in psychosis is preferential to state A clustering solutions. Throughout the hierarchy, a multistable shifting . While we are unable to make direct claims regarding dynamical system was evident, ﬂuctuating around a global the association between time-varying proﬁles of network function attractor state (state A) that possesses a relatively muted con- and cognition, converging evidence suggests that broad properties nectivity proﬁle. Decreased within-network coupling is evident of network connectivity may be preserved across experimental 12,48 when spontaneous neuronal activity adheres to ﬁxed correlation contexts (e.g., intrinsic or task-evoked) and analysis strate- 61,62 48 conﬁgurations deﬁned by structural connectivity . Although gies . An important area of future work will be to establish the speculative, our results suggest the presence of an attractor state extent to which the cortical network structure adjusts as a func- that could preferentially link to the large-scale anatomical struc- tion of task and/or environment. One speculative possibility is ture of the human cerebral cortex. Future cross-modal analyses that transient periods of strong within-network coupling may will be necessary to explore these hypotheses and test the extent correspond to epochs of high efﬁciency, a property of brain to which patterns of spontaneous brain activity might reconﬁgure organization theorized to optimize communication across distinct around an underlying anatomical skeleton. functional domains . Regions within the association cortex display increased func- Human cognition is a ﬂuctuating process, and there is growing tional ﬂexibility, potentially serving to integrate information evidence that functional connectivity patterns exhibit complex 63 58 across more specialized aspects of the cortex . This proﬁle of spatiotemporal dynamics at multiple time scales . Time-varying malleability is reﬂected in static analyses of intrinsic network brain states, such as those identiﬁed in the present analyses, have function where heightened population-level variability has been been hypothesized to reﬂect changes in the ongoing cognitive 46 7 observed in association relative to unimodal cortices . Consistent processes during rest . However, the existence, putative origins, with prior reports of marked heterogeneity in the dynamic ﬂex- and cognitive correlates of dynamics in resting-state fMRI remain 13,64 10,11,19,67 ibility of neural regions , we observed the greatest cross-state a topic of empirical debate . While the relations linking variability within aspects of default, attention, and control net- intrinsic time-varying network proﬁles with cognition remain works. The trade-off between control and attention systems is speculative, there are two commonly held views. First, that by thought to be a central feature of many cognitive functions, applying clustering algorithms, the brain’s dynamic architecture including adaptive goal pursuit, working memory, and set can be carved at the joints, revealing discrete brain states. Second, 10 NATURE COMMUNICATIONS (2018) 9:1157 DOI: 10.1038/s41467-018-03462-y www.nature.com/naturecommunications | | | NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-03462-y ARTICLE a b 1.00 1.00 State A State B 0.80 0.80 State C State D 0.60 0.60 0.40 0.40 0.20 0.20 0.00 0.00 0.00 0.20 0.40 0.60 0.80 1.00 0.00 0.20 0.40 0.60 0.80 1.00 False positive rate False positive rate Fig. 9 Speciﬁc brain states are predictive of distinct clinical symptoms in novel individuals. a State B uniquely predicted the presence of active psychosis in patients with psychotic illness. Results from a leave-one-subject-out cross-validation elastic net logistic regression analysis comparing predicted and observed psychotic symptoms (n = 91). The displayed area under the receiver-operating characteristic curve (AUC) reﬂects the scalar probability measure for each individual, predicting the likelihood of active psychotic symptoms. The blue semitransparent lines reﬂect 1000 ROC curves from 1000 permutation experiments. The black dashed line indicates the mean of these 1000 curves, approximating a null curve. The purple, red, green, and orange curves 4 4 4 4 correspond to states A ,B ,C , and D , respectively. b Connectivity models deﬁned on training data predict the presence of active psychotic symptoms in an independent group of participants (n = 31). The displayed AUC graph reﬂects the scalar probability measure, predicting the likelihood of active psychosis, deﬁned using positive edges from the initial training set without further ﬁtting or modiﬁcation, in a held-out sample of patients 8,16 that the resulting macro-level brain states index, or enable, dis- organization of the brain is impaired in psychotic illness . The tinct biological and cognitive processes. In the current analyses, present analyses reveal the presence of both state and network we apply k-means clustering to identify dissociable network preferential reductions in functional connectivity. Patients with conﬁgurations in solutions from 2 to 8 brain states. Critically, we schizophrenia spend more time than healthy individuals in net- do not claim that these conﬁgurations reﬂect wholly distinct brain work conﬁgurations typiﬁed by reduced large-scale connectivity, states separated by sharp transitions, and we are not in the while also showing muted negative correlations between default position to ascribe associated cognitive functions. Rather, the data and other networks . In line with this literature, we observed a are consistent with the possibility that large-scale network cou- general proﬁle of decreased within- and increased between- pling gradually ﬂuctuates around a core functional architecture. network correlations. Our analyses were also consistent with prior However, the extent to which neurobiological and cognitive work, demonstrating executive functioning and cognitive control mechanisms may drive the observed transient connectivity pat- abnormalities in patients with psychotic illness , revealing pre- terns remains an open question. ferential disruptions in frontoparietal control network con- 4 4 nectivity in the attractor state (state A ) and state C , a network Debates regarding the existence of discrete brain states should not be taken to imply that temporal ﬂuctuations in network conﬁguration marked by increased within-network frontoparietal connectivity lack biological information. Notably, the present connectivity in healthy young adults. This proﬁle of ﬂuctuating analyses indicate that an individual’s proﬁle of dynamic func- abnormalities in network connectivity was also evident in a tional connectivity is unique and stable over the course of days default network, which exhibited state-preferential impairments and months. Patterns of connectivity deﬁned through traditional in state A . Critically, these analyses do not demonstrate selective 25,26 static analyses are heritable and function as a trait-like sig- abnormalities in discrete network conﬁgurations that are speciﬁc nature that can accurately identify participants from a large to psychotic illness. Rather, consistent with evidence for altera- 27–29 16–18 group . While prior demonstrations of cross-session identi- tions in dynamic brain architecture in patient populations , ﬁcation were established when participant visits were separated they suggest that broad disruptions across cortical association by a single day , these data suggest that single measures of time- networks in psychosis may emerge through transient abnormal- averaged connectivity may provide meaningful trait-like infor- ities preferentially evident during the expression of particular 17,35 mation about individuals. Here, we observed robust participant- network conﬁgurations . Together with growing evidence 12– speciﬁc matching when visits were separated by up to 6 months linking behavior to temporally derived network conﬁgurations (mean = 63.35 ± 48.10 days apart, range = 2–151 days). Despite , the presence of both individual speciﬁcity (ﬁngerprints) in some variability, identiﬁcation accuracy was not limited to any brain dynamics across the population and evidence for ﬂuctuat- one state or clustering solution, indicating that participants may ing, network-speciﬁc impairments in patients with severe psy- be identiﬁed with relatively thin slices of transient brain activity. chopathology have strong implications for clinical practice. For These discoveries highlight the potential to identify associations instance, these data suggest a potential avenue to identify distinct linking the unique dynamic functional architecture of an indivi- symptom proﬁles, track time-varying disease states, responses to dual’s brain to the integrity of large-scale corticocortical path- environmental perturbations, or individually speciﬁc treatment ways . The continued development of time-varying data analytic responses. approaches with high sensitivity to individual variability could Consistent with the aim of delineating disease-relevant markers facilitate the discovery of meaningful biomarkers for both cog- of brain biology, the current analyses suggest that models based nitive ability and disease states. on transient proﬁles of network function may serve as powerful, Schizophrenia and psychotic bipolar disorder are marked by generalizable predictors of clinical symptomatology. In a group of altered intrinsic network connectivity, potentially contributing to patients with psychotic illness, we identiﬁed a speciﬁc time- 30–34 widespread changes in information processing . A key ques- varying network proﬁle whose strength predicted the presence of tion facing the ﬁeld is the extent to which the temporal active psychotic symptoms at the point of clinical assessment. NATURE COMMUNICATIONS (2018) 9:1157 DOI: 10.1038/s41467-018-03462-y www.nature.com/naturecommunications 11 | | | True positive rate True positive rate ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-03462-y Patients with psychotic illness were recruited from clinical services at McLean This whole-brain network model provides preliminary evidence Hospital (n = 170; age: 32.08 ± 11.64; female: 66.47%; right handed: 84.71%), for meaningful, clinically relevant signals in patterns of dynamic including 41 patients diagnosed with schizoaffective disorder, 56 with intrinsic connectivity. Suggesting that the clinically relevant fea- schizophrenia, and 73 with psychotic bipolar disorder. Study procedures are tures of intrinsic brain dynamics are robust and generalizable, detailed in Baker et al. 2014 . Brieﬂy, exclusion criteria included neurological illness, positive pregnancy test, electroconvulsive therapy in the last 3 months, and networks deﬁned within an initial training set successfully history of head trauma. Reﬂecting the severity of the present sample, the majority predicted active psychosis in a completely independent sample. of patients (82%) reported experiencing active psychotic symptoms at the time of As reﬂected in our analyses, a randomly selected, previously their clinical assessment, as assessed through DSM-IV (SCID) clinician-rated unseen, patient with active psychosis could be distinguished symptomatic diagnostic criteria, indicating the presence of delusions and/or hallucinations in the past month . All patients were assessed for active symptoms from another patient without psychosis at ~74.0% accuracy based within 24 h of scan using the Positive and Negative Syndrome Scale (PANSS; on their dynamic expression of selected network conﬁgurations positive scale: 18.46 ± 6.51; negative scale: 13.48 ± 7.47; general psychopathology (state B ), demonstrating a relatively high level of precision. scale: 30.82 ± 7.95) . A demographically matched healthy comparison sample was Critically, these relations were not evident through traditional recruited from the surrounding Boston communities (n = 369; age: 37.16 ± 14.65; female: 62.06%; right handed: 91.33%). The McLean Hospital Institutional Review static analysis. Readers should note that these analyses leverage Board approved the study, and all participants provided written informed consent. cross-sectional data, and are technically postdictive. Additional The control group was signiﬁcantly older than the patient group (t = –3.98, p ≤ research should assess if this cross-sectional/retrospective 0.001). No signiﬁcant group differences were identiﬁed in sex, handedness, or approach generalizes to the prospective prediction of symptoms, education (ps ≥ 0.07). The comparison sample was explicitly selected to match on prior to clinical assessment. the basis data quality. The BOLD runs for the patient and comparison groups did not differ in terms of slice-based temporal signal-to-noise ratio (comparison: Suggesting a degree of speciﬁcity for the prediction of psychotic 161.17 ± 46.05; patient: 158.35 ± 71.53) or the number of relative translations in 3D symptom severity, the active psychosis model served as a unique space ≥0.1 mm (comparison: 31.05 ± 29.15; patient: 32.29 ± 29.66; ps ≥ 0.58). The predictor of PANSS positive-scale scores (binarized as greater or slice-based signal-to-noise ratio was calculated as the weighted mean of each slice’s less than one standard deviation below the sample mean) in the mean intensity over time (weighted by the size of the slice). All imaging data were collected on 3-T Tim Trio scanners (Siemens) with a 12- left-out patient sample, with poor prediction observed for both channel phased-array head coil at Harvard University, Massachusetts General negative and general psychopathology symptoms. Caution is Hospital, or McLean Hospital. Structural data included a high-resolution multi- warranted given the limited sample size; however, these data echo T1-weighted magnetization-prepared gradient-echo image (TR = 2200 ms, provide evidence to suggest that dynamic network models trained TI = 1100 ms, TE = 1.54 ms for image 1–7.01 ms for image 4, FA = 7°, 1.2 × 1.2 × 1.2 mm, and FOV = 230). Functional data were acquired using a gradient-echo on different, yet associated, symptoms have the potential to echoplanar imaging sequence sensitive to blood oxygenation level-dependent generalize across clinical measures. These data expand our cur- contrast with the following parameters: 124 time points; repetition time = 3000 ms; rent knowledge regarding abnormal large-scale network function echo time = 30 ms; ﬂip angle = 85°; 3 × 3 × 3-mm voxels; FOV = 216; and 47 axial in patients with schizophrenia and psychotic bipolar disorder, sections collected with interleaved acquisition and no gap. Participants were instructed to remain still, stay awake, and keep their eyes open. Although no and highlight the use of dynamic analytic approaches when ﬁxation image was used, participants with psychotic illness were monitored via eye- examining intrinsic connectivity across heath and disease. Taken tracking video to ensure compliance during functional scans. One to two runs were together, the present analyses provide a preliminary proof of acquired for each participant (70.89% of the main sample, 69.41% of patient concept to suggest that the altered connectivity of speciﬁc tran- participants, and 46.88% of the matched comparison received a second run). sient network conﬁgurations may link to the expression of dis- Software upgrades (VB13, VB15, and VB17) occurred during data collection. Reported results are after partialing out variance associated with scanner and crete symptom proﬁles. Future work should focus on the software upgrade. identiﬁcation of relations linking functional network dynamics to the expression of psychological and behavioral aspects of illness. Data preprocessing. Data were processed with a series of steps common to In conclusion, we demonstrated the presence of a ﬂuctuating 69–71 intrinsic connectivity analyses . Preprocessing included discarding the ﬁrst four and reconﬁgurable hierarchy across the functional connectome. volumes of each run to allow for T1-equilibration effects, compensating for slice The observed dynamic network proﬁles were unique and reliable acquisition-dependent time shifts per volume, and correcting for head motion within individuals over the course of months and impaired in using rigid body translation and rotation. Additional steps included the removal of constant offset and linear trends over each run, and the application of a temporal patients with psychotic illness. Our analyses suggest that temporal ﬁlter to retain frequencies below 0.08 Hz. Sources of spurious variance, along with patterns of connectivity between cortical regions link to the broad their temporal derivatives, were removed through linear regression. These included functional capacities of individual human brains, enabling the six parameters obtained by correction for rigid body head motion, the signal prediction of speciﬁc symptom proﬁles within patient popula- averaged over the whole brain, the signal averaged over the ventricles, and the tions. These data have important implications for the study of signal averaged over the deep cerebral white matter. Structural data and functional 39 72 data were aligned as described in Yeo et al. and Buckner et al. using the behaviors and features of psychiatric illnesses that possess time- FreeSurfer software package. This method yields a surface mesh representation of varying patterns of expression. each participant’s cortex, which is then registered to a common spherical coordi- nate system. Images were aligned with boundary-based registration from the FsFast software. Functional and structural images were then aligned to the com- Methods mon coordinate system by sampling from the middle of the cortical ribbon in a Data acquisition. Native English-speaking young adults (aged 18–35) with normal single interpolation step to reduce blurring of the functional signal across sulci and or corrected-to-normal vision were recruited from Harvard University, Massa- gyri. A 6-mm smoothing kernel was applied to the functional data in the surface chusetts General Hospital, and the surrounding Boston communities through an space, and data were downsampled to a 4-mm mesh. Additional details on the ongoing large-scale study of brain imaging and genetics (n = 1919; age: 21.35 ± 43 39 preprocessing procedures are detailed in Holmes et al. and Yeo et al. . 3.20; female: 56.53%; right handed: 92.40%) . History of psychiatric illness and medication usage was assessed through a structured phone screen. On the day of MRI data collection, participants completed additional questionnaires concerning Dynamic connectivity sliding-window analysis. Cortical functional coupling their physical health, past and present history of psychiatric illness, and medication matrices were computed for each participant, across all available parcels within the usage. Exclusion criteria included a history of head trauma, current or past Axis I 17-network functional atlas of Yeo et al. .We deﬁned 114 regions (57 per pathology, neurological disorders, current or past psychotropic medication use, hemisphere) that surveyed all 17 networks. Correlation matrices were constructed current physical illness, and current or past loss of consciousness. Participants to include all region pairs arranged by network membership. provided written informed consent in accordance with guidelines set by the Connectivity across time was analyzed using a sliding-window approach 7,8 Partners Health Care Institutional Review Board and the Harvard University (width = 33 s) . Prior work suggests that a sliding-window range of 30–60 s is Committee on the Use of Human Subjects in Research. For the present study, we appropriate for dynamic connectivity analyses . Pilot analyses (available upon assessed the extent to which the dynamic network architecture of the cortex is request) revealed consistent state solution stability across varying sliding-window reliable within and across visits. To accomplish this, an additional data set (n = 79; sizes of 33–63 s. Thirty-three-second windows were chosen in order to maximize age at the ﬁrst scan: 20.99 ± 2.93; female: 45.56%; right handed: 89.87%) was signal estimates, while still capturing properties of transient functional 8,40 acquired over the course of the primary collection effort. Data were collected on 2 connectivity . A time series for each participant was extracted for the 57 regions independent days (mean = 63.35 ± 48.10 days apart; min = 2; max = 151). in each hemisphere. Time-course correlations across 110 windows per bold run for 12 NATURE COMMUNICATIONS (2018) 9:1157 DOI: 10.1038/s41467-018-03462-y www.nature.com/naturecommunications | | | NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-03462-y ARTICLE each participant (220 windows if the participant had two runs) were calculated for ﬁrst established a comparison distribution from chance. For each state, the each 114 × 114 region pair. To limit the redundancy across matrices and to reduce participant vector from visit 1 was compared to a randomly permuted list of computational load, clustering was applied to a subsample of available windowed participants from visit 2, a Pearson correlation was calculated, a matching rank was covariance matrices (1/10 windows). Results were consistent with the alternate assigned, and the identiﬁcation rate was calculated. A “correct” identiﬁcation was approach of subsampling along the temporal dimension to identify windowed deﬁned in cases where the highest-ranked rho value was within a participant (visit covariance matrices with local maxima in functional connectivity variance. The 1–visit 2) relative to other participants (for ranking results, see Fig. 4, main text). resulting correlation matrices were then aggregated and z-transformed prior to This was repeated 1000 times, and a t-test was performed comparing the number of running the clustering analyses. The clustering analysis was iteratively applied to identiﬁcations of the actual order participants relative to the distribution of the deﬁne distinct state solutions for 2 through 20 brain states. Additional details on number of identiﬁcations in the permuted order of the participant list. the selected clustering approach are provided in Yeo et al. As we did not have a priori hypotheses regarding the number of functional Noise constraints. To assess the extent to which data quality might inﬂuence on connectivity states, we assessed the stability of clustering solutions for all states our ﬁndings, we conducted a series of analyses aimed at detecting the differences in 2–20. To do so, we examined the stability of the clustering analyses by iteratively motion across state solutions. To obtain motion estimates by a participant for each and randomly splitting our data on two dimensions (sliding windows and pairwise window, we extracted the mean of the root mean square of relative motion for each connectivity) and rerunning the clustering solution 30 times . The results were participant across each window. We used the Hungarian matched state vector to 75,76 then compared using a Hungarian matching algorithm , as described in the classify relative motion within each TR to a state-speciﬁc window. We averaged next section. Greater instability was quantiﬁed as a greater summation of deviation motion values within states for each state solution. Although the relations linking between the two cluster solutions (Supplementary Figure 1). state expression and motion were limited in size (η s ≤ 0.006), state A associated with the most motion for state solutions 2, 3, and 4 (all ps ≤ 0.05; Supplementary Figure 5). State A showed a muted increase in motion relative to other states in Hierarchy analysis. To examine the relations linking each state in solution S with 75,76 5 5 the states in solution S + 1, we used a Hungarian matching technique . Every solutions 5 through 7 (ﬁve-state solution: ps ≤ 0.001 for state A relative to C and 5 5 5 6 E , p ≤ 0.05; p ≥ 0.59 for B and D ; six-state solution: ps ≤ 0.01 for state A relative possible combination of states in state solution S + 1 was compared to each state in 6 6 6 6 6 to E and F , ps ≥ 0.38 for B ,C , and D ; seven-state solution: ps ≤ 0.001 for state solution S. To illustrate this point, take the comparison of states in the three- and 7 7 7 7 7 7 A relative to E and F ; p = 0.07 for C , p = 0.06 for G , and p ≥ 0.61 for B and four-state solutions. Each possible combination in the four-state solution is 4 4 4 4 4 4 4 4 4 4 4 4 7 8 D ). In state solution 8, we found that state H , a state that our hierarchy analyses established (A B ,A C ,A D ,B C ,B D , and C D ) by calculating the mean of each cell in the 114 × 114 × n (here, n = 4) connectivity matrix across the combined identiﬁed as architecturally related to state A in solutions 2–7, was associated with more motion than all other states (ps ≤ 0.001). states to create hybrid states. Next, each hybrid state is grouped along with the other states in the four-state solution and compared to the three-state solution (A Next, we examined the consistency of state stability in participants falling 4 4 4 4 4 4 4 4 4 4 4 within the ﬁrst and fourth quartiles of the distribution for motion (low motion: + B /C /D , then A + C /B /D , then A + D /B /C , etc.). Hungarian matching is used to determine which of the hybrid combinations in the four-state solution n = 424; high motion: n = 398). No group differences were identiﬁed in age, sex, or handedness (ps ≥ 0.59). To determine the stability of our state-clustering approach, most closely approximates each state in the three-state solution (so, in the ﬁrst 4 4 4 4 example, the hybrid four-state solution comprised of A + B /C /D is matched to we applied iterative k-means clustering to each group for each population-level 3 3 3 stable state solution (2, 4, 5, and 8 states). To estimate the viability of the resulting states A ,B , and C in the three-state solution). This comparison is repeated until the match with the minimal cost is identiﬁed (Supplementary Figure 2). solutions, data were resampled across sliding windows as described previously (Fig. 1, main text). A t-test was performed to assess the differences in the consistency sampling from the high and low motion groups. No differences in Identifying variability across deﬁned state solutions. To determine the extent consistency were observed across groups (Supplementary Figure 6; all ps ≥ 0.43). of network variability across brain states, we examined the variance of mean network connectivity in state solutions 2–8 (see Supplementary Figure 3 for var- Prediction of active psychotic symptoms. Here, we demonstrate that the iance with the four-state solution). We used ANOVA to assess between-state strength of functional brain networks within speciﬁc brain states predicts the variability (coefﬁcient of variance) with state solutions treated as repeated mea- presence of active psychosis in previously unseen individuals. Elastic net logistic sures. The results conﬁrmed that variance differed across the networks (F = regression analyses were conducted with custom R code (elasticnet by Hui Zou, 44.11, p ≤ 0.001). Post hoc tests revealed increased cross-state variability within Trevor Hastie, and Robert Tibshirani). Elastic net regularization is a cross-validated default A and B relative to the default D, control A and C, limbic, somatomotor A, regularized log-linear regression procedure that combines LASSO (least and visual B networks (Bonferroni-corrected ps ≤ 0.05, all other ps ≥ 0.5). Control B absolute shrinkage and selection operator) regularization and Tikhonov (ridge) demonstrated greater variance relative to default C and D, control A and C, limbic, 77,78 regularization The resulting log-linear regression weights were applied to somatomotor, and visual networks (ps ≤ 0.05; all other ps ≥ 0.5). The salience/ the edges (ROI to ROI correlations) of each network conﬁguration within the four- ventral attention and dorsal attention A networks exhibited increased variance state solution and the associated covariates. All results were cross validated. relative to default C and D, control A and C, limbic, somatomotor, and visual Model development and validation consisted of four steps. First, model features networks (ps ≤ 0.05, all other ps ≥ 0.5). th 4 4 4 were selected. Pearson correlation between each edge of the k (e.g., k = A ,B ,C , and D ) dynamic brain state and clinical status was performed in the training set. Individual identiﬁcation analyses. To examine the extent to which the observed Note that regarding dichotomous outcomes, a mass-univariate t-test would provide dynamic connectivity proﬁles are reliably expressed across scans and visits, we ﬁrst similar results as the Pearson correlation test concerning feature selection. Here, we selected participants with two bold runs (n = 1361). Next, we examined the 79 27,52 used the Pearson correlation approach to be consistent with previous literature . participants, set aside from the original cohort, who had two separate study visits The resulting edges were separated into positive and negative groups, and within 6 months of each other. To test the relations within each individual’s proﬁle thresholded on the basis of the statistical signiﬁcance (p ≤ 0.05) and signs of of network dynamics for the two bold runs within the same visit, and then for the correlation. Second, in the model development procedure, we ﬁrst aggregated the individuals with more than one visit, we implemented the following analysis: First, k values of edges in each feature set as a summary statistics, S ,or “network we obtained each participant’s state expression across time. To accomplish this, we 27 th strength” of the k brain state, for k ¼ 1; 2; 3; 4. The network strength and used a Hungarian matching algorithm, assigning each time point in a participant’s covariates (e.g., sex, age) were then entered into the model, yielding a scalar value, windowed time-course data to individual states in the desired population-level state the predicated conditional probability of active psychotic symptoms. Formally, for solution. For instance, in the four-state solution, we found the best ﬁt for each subject i, i 2f1; 2; ; ng,we deﬁne 4 4 4 4 window in the participant’s data and classiﬁed it as state A ,B ,C ,orD . This k k k ^ ^ k β þβ s þð^γ Þci yielded a vector of states for each participant, representing the participant’s state 0 1 i k k k ^p :¼ P Y ¼ 1jS ¼ s ; C ¼ c ¼ i i k k k expression over the course of the scan. Following this, we took the average cor- i i ^ ^ k β þβ s þð^γ Þci 0 1 i 1 þ e relation matrix for each window within a state. So, for participant 1, we collapsed 4 4 across all of the state A windows, and generated a mean for state A , creating a k k k where p :¼ P Y ¼ 1jS ¼ s ; C ¼ c denotes the predicted probability that i i i i 114 × 114 × S average state matrix for each participant. Participant matrices were subject i has active psychosis, given their observed network strength s during brain vectorized, and Pearson correlations were run across every participant in two state k, and covariates c , a vector consisting of all observed covariates for subject i. analyses for (1) bold 1 and bold 2; and every participant for (2) visit 1 and visit 2. k k k ^ ^ β ; β ; and ^γ are estimated weights for brain state k from the elastic net logistic 0 1 Analyses indicated a signiﬁcant level of consistency in within-individual con- k regression, where ^γ is a vector consisting of the estimates for each covariate in c . nectivity proﬁles across bold runs and visits (Supplementary Figure 4). For each Alternatively, we could model the predicted probability that subject i does not have state solution, we ran a t-test for within-participant rho values and between- k k 1 psychosis as P Y ¼ 0jS ¼ s ; C ¼ c ¼ : Probability estimation was i i k k k k c i ^ ^ β þβ s þð^γ Þ participant rho values. All tests revealed increased rho values within, rather than 0 1 i i 1þe between, participants (all ps ≤ 0.001). iteratively performed using leave-one-subject-out cross-validation procedure. Permutation tests were performed in a manner consistent with prior studies During each iteration, the weights were estimated using data from (n–1, n = 91) examining functional connectome ﬁngerprinting . For these analyses, we participants and were used to predict the probability of the remaining participant considered participants with multiple study visits (n = 79) and utilized the state- having active psychotic symptoms. Each individual was left out once; hence, the and participant-speciﬁc connectivity vectors described above. A Pearson procedure yielded n-predicted probability scores. To evaluate the estimation correlation was calculated across an ordered list for every participant’s connectivity performance, we measured the area under the receiver-operating characteristic vector from visit 1 compared to every other participant’s vector from visit 2. We (ROC) curve (AUC), estimated directly by conducting numerical integration of the NATURE COMMUNICATIONS (2018) 9:1157 DOI: 10.1038/s41467-018-03462-y www.nature.com/naturecommunications 13 | | | ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-03462-y ROC under all thresholds that yielded unique sensitivity/speciﬁcity values, wherein 16. Calhoun, V. D., Miller, R., Pearlson, G. & Adali, T. The chronnectome: time- 0.5 indicates chance, and 1 is perfect discrimination. varying connectivity networks as the next frontier in fMRI data discovery. The model development and parameter estimation were conducted only using Neuron 84, 262–274 (2014). the training data. To evaluate the reproducibility, we applied the model obtained 17. Rashid, B., Damaraju, E., Pearlson, G. D. & Calhoun, V. D. Dynamic from the training data, without further ﬁtting or modiﬁcation, to 39 previously connectivity states estimated from resting fMRI identify differences among unseen participants. To access the speciﬁcity of the model in detecting positive schizophrenia, bipolar disorder, and healthy control subjects. Front. Hum. symptoms, we examined the positive, negative, and general psychopathology Neurosci. 8, 897 (2014). subscales of the PANSS in the held-out sample. PANSS scores were binarized as 18. Damaraju, E. et al. Dynamic functional connectivity analysis reveals transient low (less than one standard deviation below the mean) or high (greater than one states of dysconnectivity in schizophrenia. Neuroimage Clin. 5, 298–308 standard deviation below the mean) for each subscale (positive scale: 18.46 ± 6.51, (2014). cutoff = 11.95; negative scale: 13.48 ± 7.47, cutoff = 6.01; general psychopathology 19. Breakspear, M. Dynamic models of large-scale brain activity. Nat. Neurosci. scale: 30.82 ± 7.95, cutoff = 22.86). Due to the large variance in PANSS negative 20, 340–352 (2017). scores, to thoroughly explore potential relations with this subscale, we additionally 20. Schölvinck, M. L., Leopold, D. A., Brookes, M. J. & Khader, P. H. The examined alternative thresholds of <15 and of <12 selecting the optimal result contribution of electrophysiology to functional connectivity mapping. (<12) for comparison with the PANSS positive scale. Neuroimage 80, 297–306 (2013). To assess the statistical signiﬁcance of the sensitivity and speciﬁcity analyses, we 21. Van de Ville, D., Britz, J. & Michel, C. M. EEG microstate sequences in healthy performed nonparametric permutation testing. During each test, we ﬁrst randomly humans at rest reveal scale-free dynamics. Proc. Natl Acad. Sci. USA 107, permuted the active psychotic symptoms status or PANSS labels. Next, model 18179–18184 (2010). ﬁtting was carried out using the correct dynamic signatures and covariates and 22. Britz, J., Van De Ville, D. & Michel, C. M. BOLD correlates of EEG permuted (incorrect) labels. For each permutation, an AUC value was calculated. topography reveal rapid resting-state network dynamics. Neuroimage 52, The permutation tests were performed 5000 times for the training and testing data, 1162–1170 (2010). respectively. The null AUC values should be symmetrically distributed around 0.5 23. Musso, F., Brinkmeyer, J., Mobascher, A., Warbrick, T. & Winterer, G. if the procedure is unbiased. Spontaneous brain activity and EEG microstates. A novel EEG/fMRI analysis approach to explore resting-state networks. Neuroimage 52, 1149–1161 (2010). Data availability. The data summary statistics that support the ﬁndings of this 24. Chang, C. et al. Association between heart rate variability and ﬂuctuations in study are available from the corresponding author on request. As detailed in resting-state functional connectivity. Neuroimage 68,93–104 (2013). Holmes et al. , the Brain Genomics Superstruct Project (GSP) initial release data 25. Glahn, D. C. et al. Genetic control over the resting brain. Proc. Natl Acad. Sci. set of structural, functional, and behavioral measures is available for download USA 107, 1223–1228 (2010). (http://neuroinformatics.harvard.edu/gsp/). Step-by-step instructions detailing how 26. Ge, T., Holmes, A. J., Buckner, R. L., Smoller, J. W. & Sabuncu, M. R. to access the release data set are available online in the “Request Access” page Heritability analysis with repeat measurements and its application to resting- (http://neuroinformatics.harvard.edu/gsp/get). The patient data are not publicly state functional connectivity. Proc. Natl Acad. Sci. USA 114, 5521–5526 available due to information that could compromise research participant privacy/ (2017). consent. 27. Finn, E. S. et al. Functional connectome ﬁngerprinting: identifying individuals using patterns of brain connectivity. Nat. Neurosci. 18, 1664–1671 (2015). Received: 19 July 2017 Accepted: 15 February 2018 28. Miranda-Dominguez, O. et al. Connectotyping: model based ﬁngerprinting of the functional connectome. PLoS ONE 9, e111048 (2014). 29. Wang, D. et al. Parcellating cortical functional networks in individuals. Nat. Neurosci. 18, 1853–1860 (2015). 30. Baker, J. T. et al. Disruption of cortical association networks in schizophrenia and psychotic bipolar disorder. JAMA Psychiatry 71, 109–118 (2014). References 31. Buckner, R. L., Andrews-Hanna, J. R. & Schacter, D. L. The brain’s default 1. Buckner, R. L., Krienen, F. M. & Yeo, B. T. T. Opportunities and limitations of network: anatomy, function, and relevance to disease. Ann. NY Acad. Sci. intrinsic functional connectivity MRI. Nat. Neurosci. 16, 832–837 (2013). 1124,1–38 (2008). 2. Craddock, R. C. et al. Imaging human connectomes at the macroscale. Nat. 32. Cole, M. W., Repovs, G. & Anticevic, A. The frontoparietal control system: a Methods 10, 524–539 (2013). central role in mental health. Neuroscientist 20, 652–664 (2014). 3. Van Essen, D. C. & Ugurbil, K. The future of the human connectome. 33. Fornito, A. & Bullmore, E. T. Reconciling abnormalities of brain network Neuroimage 62, 1299–1310 (2012). structure and function in schizophrenia. Curr. Opin. Neurobiol. 30,44–50 4. Bullmore, E. & Sporns, O. Complex brain networks: graph theoretical analysis (2015). of structural and functional systems. Nat. Rev. Neurosci. 10, 186–198 (2009). 34. Fornito, A., Zalesky, A., Pantelis, C. & Bullmore, E. T. Schizophrenia, 5. Krienen, F. M., Yeo, B. T. T. & Buckner, R. L. Reconﬁgurable task-dependent neuroimaging and connectomics. Neuroimage 62, 2296–2314 (2012). functional coupling modes cluster around a core functional architecture. 35. Uhlhaas, P. J. & Singer, W. Abnormal neural oscillations and synchrony in Philos. Trans. R. Soc. Lond. B Biol. Sci. 369, 20130526 (2014). schizophrenia. Nat. Rev. Neurosci. 11, 100–113 (2010). 6. Spreng, R. N., Stevens, W. D., Chamberlain, J. P., Gilmore, A. W. & Schacter, 36. Miller, R. L., Yaesoubi, M., Calhoun, V. D. & Gopal, S. Higher dimensional D. L. Default network activity, coupled with the frontoparietal control fMRI connectivity dynamics show reduced dynamism in schizophrenia network, supports goal-directed cognition. Neuroimage 53, 303–317 (2010). patients. in 2014 International Workshop on Pattern Recognition in 7. Allen, E. A. et al. Tracking whole-brain connectivity dynamics in the resting Neuroimaging 3837–3840 (IEEE, 2014). state. Cereb. Cortex. 24, 663–676 (2014). 37. Sakoğlu, Ü. et al. A method for evaluating dynamic functional network 8. Hutchison, R. M. et al. Dynamic functional connectivity: promise, issues, and connectivity and task-modulation: application to schizophrenia. Magn. Reson. interpretations. Neuroimage 80, 360–378 (2013). Mater. Phys. 23, 351–366 (2010). 9. Zalesky, A., Fornito, A., Cocchi, L., Gollo, L. L. & Breakspear, M. Time- 38. American Psychiatric Association. Diagnostic and Statistical Manual of resolved resting-state brain networks. Proc. Natl Acad. Sci. USA 111, Mental Disorders (DSM-IV) (APA Press, Washington, DC, 1994). 10341–10346 (2014). 39. Yeo, B. T. T. et al. The organization of the human cerebral cortex estimated by 10. Laumann, T. O. et al. On the stability of BOLD fMRI correlations. Cereb. intrinsic functional connectivity. J. Neurophysiol. 106, 1125–1165 (2011). Cortex 27, 4719–4732 (2017). 40. Chang, C. & Glover, G. H. Time-frequency dynamics of resting-state brain 11. Liegeois, R., Laumann, T. O., Snyder, A. Z., Zhou, J. & Yeo, B. T. T. connectivity measured with fMRI. Neuroimage 50,81–98 (2010). Interpreting temporal ﬂuctuations in resting-state functional connectivity 41. Hutchison, R. M., Womelsdorf, T., Gati, J. S., Everling, S. & Menon, R. S. MRI. Neuroimage 163, 437-455 (2017). Resting-state networks show dynamic functional connectivity in awake 12. Cole, M. W., Bassett, D. S., Power, J. D., Braver, T. S. & Petersen, S. E. Intrinsic humans and anesthetized macaques. Hum. Brain Mapp. 34, 2154–2177 and task-evoked network architectures of the human brain. Neuron 83, (2013). 238–251 (2014). 42. Handwerker, D. A., Roopchansingh, V., Gonzalez-Castillo, J. & Bandettini, P. 13. Shine, J. M., Koyejo, O. & Poldrack, R. A. Temporal metastates are associated A. Periodic changes in fMRI connectivity. Neuroimage 63, 1712–1719 (2012). with differential patterns of time-resolved connectivity, network topology, and 43. Holmes, A. J. et al. Brain genomics superstruct project initial data release with attention. Proc. Natl Acad. Sci. USA 113, 9888–9891 (2016). structural, functional, and behavioral measures. Sci. Data 2, 150031 (2015). 14. Wang, C., Ong, J. L., Patanaik, A., Zhou, J. & Chee, M. W. L. Spontaneous 44. Ben-Hur, A., Elisseeff, A. & Guyon, I. A stability based method for discovering eyelid closures link vigilance ﬂuctuation with fMRI dynamic connectivity structure in clustered data. Pac. Symp. Biocomput. 7,6–17 (2002). states. Proc. Natl Acad. Sci. USA 113, 9653–9658 (2016). 45. Lange, T., Roth, V., Braun, M. L. & Buhmann, J. M. Stability-based validation 15. Kucyi, A., Hove, M. J., Esterman, M., Hutchison, R. M. & Valera, E. M. of clustering solutions. Neural Comput. 16, 1299–1323 (2004). Dynamic brain network correlates of spontaneous ﬂuctuations in attention. 46. Mueller, S. et al. Individual variability in functional connectivity architecture Cereb. Cortex 27, 1831–1840 (2016). of the human brain. Neuron 77, 586–595 (2013). 14 NATURE COMMUNICATIONS (2018) 9:1157 DOI: 10.1038/s41467-018-03462-y www.nature.com/naturecommunications | | | NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-03462-y ARTICLE 47. Goldman-Rakic, P. S. Topography of cognition: parallel distributed networks 73. Greve, D. N. & Fischl, B. Accurate and robust brain image alignment using in primate association cortex. Annu. Rev. Neurosci. 11, 137–156 (1988). boundary-based registration. Neuroimage 48,63–72 (2009). 48. Crossley, N. A. et al. Cognitive relevance of the community structure of the 74. Shirer, W. R., Ryali, S., Rykhlevskaia, E., Menon, V. & Greicius, M. D. human brain functional coactivation network. Proc. Natl Acad. Sci. USA 110, Decoding subject-driven cognitive states with whole-brain connectivity 11583–11588 (2013). patterns. Cereb. Cortex 22, 158–165 (2012). 49. Shine, J. M., et al. The dynamics of functional brain networks: integrated 75. Kuhn, H. W. The Hungarian method for the assignment problem. Nav. Res. network states during cognitive task performance. Neuron 92, 544–554 (2016). Logist. Q. 2,83–97 (1955). 50. Lynall, M. E. et al. Functional connectivity and brain networks in 76. Munkres, J. Algorithms for the assignment and transportation problems. J. schizophrenia. J. Neurosci. 30, 9477–9487 (2010). Soc. Ind. Appl. Math. 5,32–38 (1957). 51. Repovs, G., Csernansky, J. G. & Barch, D. M. Brain network connectivity in 77. Hastie, T., Tibshirani, R. & Friedman, J. H. The Elements of Statistical individuals with schizophrenia and their siblings. Biol. Psychiatry 69, 967–973 Learning- Data Mining, Inference, and Predictions (Springer, New York, (2011). 2001). 52. Rosenberg, M. D. et al. A neuromarker of sustained attention from whole- 78. Zou, H. & Hastie, T. Regularization and variable selection via the elastic net. J. brain functional connectivity. Nat. Neurosci. 19, 165–171 (2016). R. Stat. Soc. B 67, 301–320 (2005). 53. Shen, X. et al. Using connectome-based predictive modeling to predict 79. Van Essen, D. C. A population-average, landmark- and surface-based (PALS) individual behavior from brain connectivity. Nat. Protoc. 12, 506–518 (2017). atlas of human cerebral cortex. Neuroimage 28, 635–662 (2005). 54. Kay, S. R., Fiszbein, A. & Opfer, L. A. The positive and negative syndrome scale (PANSS) for schizophrenia. Schiz Bull. 13, 261–276 (1987). Acknowledgements 55. Price, C. J. & Friston, K. J. Functional ontologies for cognition: the systematic This work was supported by the Brain & Behavior Research Foundation (A.J.H.), the deﬁnition of structure and function. Cogn. Neuropsychol. 22, 262–275 (2005). Taplin Family Foundation (D.Ö.), and the National Institute of Mental Health (Grants 56. Barttfeld, P. et al. Signature of consciousness in the dynamics of resting-state R01-MH079799 and K24MH094614 to J.W.S., 23MH079982-01A1 and 5R01MH094594 brain activity. Proc. Natl Acad. Sci. USA 112, 887–892 (2015). to D.Ö, and Grant K01MH099232 to A.J.H.). Analyses were made possible by the 57. Engel, A. K., Gerloff, C., Hilgetag, C. C. & Nolte, G. Intrinsic coupling modes: resources provided through Shared Instrumentation Grants 1S10RR023043 and multiscale interactions in ongoing brain activity. Neuron 80, 867–886 (2013). 1S10RR023401. We thank Monica Rosenberg for her feedback on early versions of this 58. Honey, C. J., Kötter, R., Breakspear, M. & Sporns, O. Network structure of manuscript. Data were provided in part by the Brain Genomics Superstruct Project of cerebral cortex shapes functional connectivity on multiple time scales. Proc. Harvard University and Massachusetts General Hospital (principal investigators: Randy Natl Acad. Sci. USA 104, 10240–10245 (2007). L. Buckner, J.L.R., and J.W.S.) with support from the Center for Brain Science Neu- 59. Shen, K., Hutchison, R. M., Bezgin, G., Everling, S. & McIntosh, A. R. Network roinformatics Research Group, the Athinoula A. Martinos Center for Biomedical Ima- structure shapes spontaneous functional connectivity dynamics. J. Neurosci. ging, and the Center for Genomic Medicine. Both the Simons Foundation and the 35, 5579–5588 (2015). Howard Hughes Medical Institute supported Dr. Buckner’s work on the GSP, and we are 60. Hutchison, R. M., Hutchison, M., Manning, K. Y., Menon, R. S. & Everling, S. immensely grateful for his invaluable contributions on that project. Isoﬂurane induces dose-dependent alterations in the cortical connectivity proﬁles and dynamic properties of the brain’s functional architecture. Hum. Brain Mapp. 35, 5754–5775 (2014). Author contributions 61. Deco, G., Jirsa, V. K. & McIntosh, A. R. Resting brains never rest: J.M.R. and A.J.H. conceived the study; J.M.R., O.Y.C., K.M.A., and A.J.H. performed the computational insights into potential cognitive architectures. Trends Neurosci. analyses with technical support and guidance with data interpretation from R.M.H., M.R. 36, 268–274 (2013). S., D.Ö., J.L.R., J.W.S., J.T.B., and B.T.Y.; and J.M.R., O.Y.C., and A.J.H wrote the 62. Deco, G. et al. Identiﬁcation of optimal structural connectivity using manuscript. All authors commented on the manuscript. functional connectivity and neural modeling. J. Neurosci. 34, 7910–7916 (2014). 63. Yeo, B. T. T. et al. Functional specialization and ﬂexibility in human Additional information association cortex. Cereb. Cortex 25, 3654–3672 (2015). Supplementary Information accompanies this paper at https://doi.org/10.1038/s41467- 018-03462-y. 64. Ciric, R., Nomi, J. S., Uddin, L. Q. & Satpute, A. B. Contextual connectivity: a framework for understanding the intrinsic dynamic architecture of large-scale functional brain networks. Sci. Rep. 7, 6537 (2017). Competing interests: The authors declare no competing interests. 65. Corbetta, M. & Shulman, G. L. Control of goal-directed and stimulus-driven attention in the brain. Nat. Rev. Neurosci. 3, 201–215 (2002). Reprints and permission information is available online at http://npg.nature.com/ 66. de Pasquale, F., Della Penna, S., Sporns, O., Romani, G. L. & Corbetta, M. A reprintsandpermissions/ dynamic core network and global efﬁciency in the resting human brain. Cereb. Publisher's note: Springer Nature remains neutral with regard to jurisdictional claims in Cortex 26, 4015–4033 (2016). 67. Leonardi, N., Shirer, W. R., Greicius, M. D. & Van De Ville, D. Disentangling published maps and institutional afﬁliations. dynamic networks: Separated and joint expressions of functional connectivity patterns in time. Hum. Brain Mapp. 35, 5984–5995 (2014). 68. Perlstein, W. M., Carter, C. S., Noll, D. C. & Cohen, J. D. Relation of prefrontal cortex dysfunction to working memory and symptoms in schizophrenia. Am. Open Access This article is licensed under a Creative Commons J. Psychiatry 158, 1105–1113 (2001). Attribution 4.0 International License, which permits use, sharing, 69. Biswal, B., Yetkin, F. Z., Haughton, V. M. & Hyde, J. S. Functional adaptation, distribution and reproduction in any medium or format, as long as you give connectivity in the motor cortex of resting human brain using echo-planar appropriate credit to the original author(s) and the source, provide a link to the Creative MRI. Magn. Reson. Med. 34, 537–541 (1995). Commons license, and indicate if changes were made. The images or other third party 70. Fox, M. D. et al. The human brain is intrinsically organized into dynamic, material in this article are included in the article’s Creative Commons license, unless anticorrelated functional networks. Proc. Natl Acad. Sci. USA 102, 9673–9678 indicated otherwise in a credit line to the material. If material is not included in the (2005). article’s Creative Commons license and your intended use is not permitted by statutory 71. Van Dijk, K. R. A. et al. Intrinsic functional connectivity as a tool for human regulation or exceeds the permitted use, you will need to obtain permission directly from connectomics: theory, properties, and optimization. J. Neurophysiol. 103, the copyright holder. To view a copy of this license, visit http://creativecommons.org/ 297–321 (2010). licenses/by/4.0/. 72. Buckner, R. L., Krienen, F. M., Castellanos, A., Diaz, J. C. & Yeo, B. T. T. The organization of the human cerebellum estimated by intrinsic functional connectivity. J. Neurophysiol. 106, 2322–2345 (2011). © The Author(s) 2018 NATURE COMMUNICATIONS (2018) 9:1157 DOI: 10.1038/s41467-018-03462-y www.nature.com/naturecommunications 15 | | |

Journal

Nature Communications – Springer Journals

Published: Mar 20, 2018

Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 7-Day Trial for You or Your Team.

Learn More →

The human cortex possesses a reconfigurable dynamic network architecture that is disrupted in psychosis

The human cortex possesses a reconfigurable dynamic network architecture that is disrupted in psychosis

Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 7-Day Trial for You or Your Team.

Learn More →

The human cortex possesses a reconfigurable dynamic network architecture that is disrupted in psychosis

The human cortex possesses a reconfigurable dynamic network architecture that is disrupted in psychosis

References (148)

Abstract

Journal

Recommended Articles

There are no references for this article.

Our policy towards the use of cookies