Dynamic brain network reconfiguration as a potential schizophrenia genetic risk mechanism modulated by NMDA receptor function

Significance

Converging evidence points to a role for glutamate and altered brain network dynamics in schizophrenia, but the molecular and genetic contributions are poorly understood. Here, we applied dynamic network neuroscience methods to neuroimaging working memory data to identify potential alterations in brain network flexibility related to schizophrenia genetic risk and N-methyl-d-aspartate (NMDA) receptor hypofunction. Consistent with altered network dynamics, we detected significant increases in brain network flexibility in patients with schizophrenia, healthy first-degree relatives, and healthy subjects receiving a single dose of an NMDA receptor antagonist. Our data identify a potential dynamic network intermediate phenotype related to the genetic risk for schizophrenia and point to a critical role for glutamate in the temporal coordination of neural networks and the pathophysiology of schizophrenia.

Abstract

Schizophrenia is increasingly recognized as a disorder of distributed neural dynamics, but the molecular and genetic contributions are poorly understood. Recent work highlights a role for altered N-methyl-d-aspartate (NMDA) receptor signaling and related impairments in the excitation–inhibitory balance and synchrony of large-scale neural networks. Here, we combined a pharmacological intervention with novel techniques from dynamic network neuroscience applied to functional magnetic resonance imaging (fMRI) to identify alterations in the dynamic reconfiguration of brain networks related to schizophrenia genetic risk and NMDA receptor hypofunction. We quantified “network flexibility,” a measure of the dynamic reconfiguration of the community structure of time-variant brain networks during working memory performance. Comparing 28 patients with schizophrenia, 37 unaffected first-degree relatives, and 139 healthy controls, we detected significant differences in network flexibility [F(2,196) = 6.541, P = 0.002] in a pattern consistent with the assumed genetic risk load of the groups (highest for patients, intermediate for relatives, and lowest for controls). In an observer-blinded, placebo-controlled, randomized, cross-over pharmacological challenge study in 37 healthy controls, we further detected a significant increase in network flexibility as a result of NMDA receptor antagonism with 120 mg dextromethorphan [F(1,34) = 5.291, P = 0.028]. Our results identify a potential dynamic network intermediate phenotype related to the genetic liability for schizophrenia that manifests as altered reconfiguration of brain networks during working memory. The phenotype appears to be influenced by NMDA receptor antagonism, consistent with a critical role for glutamate in the temporal coordination of neural networks and the pathophysiology of schizophrenia.

Schizophrenia is a highly heritable mental disorder for which aberrant interactions between brain regions or “dysconnectivity” have been proposed as a core neural mechanism (1). Functional magnetic resonance imaging (fMRI) studies demonstrate alterations in specific neural subcircuits in schizophrenia (2, 3) that are under genetic control (4), although recent data from network neuroscience point to more widespread disturbances in the dynamics of large-scale brain networks (5, 6). Uhlhaas (5), Uhlhaas and Singer (6), and Phillips and Silverstein (7) have proposed a plausible pathophysiological mechanism for these phenomena: They argue that alterations in the cellular excitation–inhibitory balance may lead to disturbances in the neural synchrony of large-scale cell ensembles and give rise to the dysconnectivity phenomena at the level of neural ensembles.

The neural excitation–inhibitory balance is highly dependent on glutamatergic N-methyl-d-aspartate (NMDA) receptor function (8), and alterations in NMDA receptor signaling have been associated with schizophrenia risk (9), disorder-related cognitive abnormalities (10, 11), and deficits in the temporal coordination of large-scale brain networks (5, 6). Theoretical work has related these dynamics to the generation of a “community architecture” of brain networks (12). Communities are groups of interacting brain regions that are more tightly functionally related to each other than to the rest of the brain. Defined with established network neuroscience methods, they can be interpreted as a static representation of the partitioning of the brain into functional subnetworks. Indeed, alterations in the static network community structure have been observed in schizophrenia (13, 14).

Recent developments in the evolving field of dynamic network neuroscience have extended these efforts by pioneering methods to assess the time-varying dynamics of community structure in large-scale brain networks (15–17). Unlike the “static” network approaches, these methods allow for the quantification of context-dependent changes of the network community structure over time (e.g., during learning or cognitive load), which is arguably a biologically more accurate description of normal and disordered brain network dynamics (16, 18, 19). Moreover, the temporal variability of the neural community structure predicts learning (16) and executive function (18) in normal volunteers, which are two established deficit areas in schizophrenia (20, 21). The temporal dynamic thus appears to reflect the capacity of the brain connectome to configure flexibly in support of cognitive demands, a feature that may relate to the genetic risk architecture of schizophrenia. However, the clinical, genetic, and pharmacological implications of such a potential relationship remain unexplored.

Here, we apply dynamic network neuroscience methods to study the reconfiguration of large-scale brain networks during the n-back task, a robust working memory paradigm (22, 23) and well-established task for studying the neural basis of the cognitive deficits and the genetic risk architecture of schizophrenia (2–4). We analyze brain regions (or nodes) and their interactions (or time-dependent edges) to identify densely interconnected communities (or coherent ensembles of brain regions) whose interactions change dependably during working memory performance. We quantify the dynamic changes in interactions using network flexibility, a measure of how often a brain region changes its allegiance to a community of nodes over time (Fig. 1). We tested three specific hypotheses, namely, that alterations in dynamic community structure are present in schizophrenia, may relate to the genetic liability to the disorder, and can be modulated by the NMDA-dependent challenge of the excitatory–inhibitory balance of the network.

Fig. 1.

Construction of time-dependent whole-brain connectivity matrices and computation of network flexibility indices. (A) Time courses of blood oxygen level-dependent (BOLD) signals from 270 brain regions defined by Power et al. (26) were extracted. (B) From each time window w1, w2, …, node-by-node connectivity matrices were estimated using wavelet coherence. (C) Time-dependent community structure over all time windows was computed using a multilayer modularity maximization algorithm developed by Mucha et al. (15). (D) Flexibility, F, of each node in the network reflects the average number of its community (or module) changes over time. Network flexibility was computed as the average flexibility of all nodes.

To test these hypotheses, we first examined whole-brain network flexibility in a sample of unaffected first-degree relatives of patients with schizophrenia (2) and compared the estimates with the estimates obtained from a cohort of patients with schizophrenia and from a large sample of healthy matched controls without a prior family history of mental illness. The study of healthy first-degree relatives of patients with schizophrenia is a particularly important strategy in the search for genetically influenced biological traits, because these individuals share an enriched set of schizophrenia risk variants but are devoid of confounding factors that interfere with fMRI readouts and complicate the interpretation of patient data (e.g., medication). We hypothesized that patients will show an increase in network flexibility consistent with prior theories positing a failure in the coordination of large-scale neural ensembles and the formation of a stable network community dynamic in schizophrenia (24). Moreover, we expected to see similar alterations in network flexibility in the unaffected relatives, consistent with the idea of an intermediate phenotype reflecting the familial liability for schizophrenia.

Finally, to probe a plausible (patho)physiological mechanism of the expected altered network dynamics, we tested the effects of a single dose of the NMDA antagonist dextromethorphan (DXM) on network flexibility in an independent sample of healthy controls enrolled in an observer-blinded, placebo-controlled, randomized, cross-over pharmacological fMRI (phfMRI) experiment. We hypothesized that NMDA receptor blockade would result in a significant increase in network flexibility, consistent with a disturbance of the neural excitatory–inhibitory balance of neural networks (5, 6) and accepted theories on the role of NMDA receptor hypofunction in schizophrenia (25).

Results

Demographics, Task Performance, Head Micromovement, and Image Quality.

To examine the genetic underpinnings of alterations in the dynamic community structure, we examined the fMRI working memory data of an initial sample of 81 patients with schizophrenia, 50 unaffected first-degree relatives, and 165 healthy controls. Analyses were performed on a final sample of 28 patients with schizophrenia, 37 unaffected first-degree relatives, and 139 healthy controls after stringent head motion control and removal of individuals with more than 5% of scrubbed frames (Methods). Tables S1–S4 provide details on all variables of the final samples, including P values for the group comparisons. The relatives and control groups were well balanced for demographics, task performance, head micromovement, and image quality. The patient group was balanced for n-back task performance, head micromovement, and image quality, but differed significantly from the other two groups by a lower number of females and a lower premorbid intelligence level as quantified by the German multiple-choice vocabulary intelligence test (MWTB). Thus, we accounted for age, sex, premorbid intelligence, and site in all subsequent statistical comparisons. A more detailed clinical description of the patient sample is provided in Table S2.

Table S1.

Demographic, clinical data, and task performance of patients, relatives, and healthy controls

Characteristic	Patients (n = 28)	Relatives (n = 37)	Controls (n = 139)	F or χ2 value	P value
Demographic information
Age, y; mean ± SD	33.36 ± 9.24	29.19 ± 11.61	32.81 ± 10.15	2.00	0.138
Sex, male/female; mean ± SD	17/11	8/29	52/87	10.39	0.006
Site, Mannheim/Berlin/Bonn	28/0/0	15/14/8	43/40/56	49.32	<0.001
School years, mean ± SD	11.39 ± 1.73	11.92 ± 1.50	11.78 ± 1.55	0.976	0.382
Handedness (right/left/both)	12/2/0*	31/4/2	128/9/2	4.24	0.375
MWTB, mean ± SD	26.07 ± 6.41	29.49 ± 3.72	29.80 ± 2.80	12.22	<0.001
fMRI task performance
2-back accuracy, %; mean ± SD	62.72 ± 19.36	72.64 ± 18.92	71.01 ± 22.44	2.04	0.133
2-back reaction time, ms; mean ± SD	547.26 ± 295.10	453.19 ± 244.88	512.36 ± 320.05	0.84	0.432
Head motion parameters
Sum frame-wise displacement, mm; mean ± SD	16.66 ± 6.29	16.29 ± 5.49	16.60 ± 4.94	0.06	0.942
No. of scrubbed frames, mean ± SD	1.59 ± 1.71	1.59 ± 1.72	1.22 ± 1.40	1.10	0.337
fMRI image quality
SNR, mean ± SD	90.16 ± 10.69	93.45 ± 12.01	95.73 ± 14.52	2.09	0.127

^*No information = 14.

MWTB, multiple-choice vocabulary test (raw values).

Table S4.

Session characteristics of healthy controls in the phfMRI study

Characteristic	Placebo	DXM	T value	P value
Drug information
Intake before fMRI, minutes; mean ± SD	157.80 ± 9.62	159.00 ± 12.58	0.475	0.638
fMRI task performances
Accuracy, %; mean ± SD	88.74 ± 20.00	87.28 ± 15.53	0.802	0.428
Reaction time, ms; mean ± SD	335.08 ± 199.72	306.07 ± 166.30	1.313	0.198
Head motion parameters
Mean frame-wise displacement, mm; mean ± SD	0.71 ± 0.57	0.77 ± 63	0.036	0.971
No. of scrubbed frames, mean ± SD	1.41 ± 1.74	1.43 ± 1.83	0.063	0.950
fMRI image quality
SNR, mean ± SD	93.03 ± 11.58	90.35 ± 10.62	1.415	0.166

Table S2.

Demographic, psychological, and clinical parameters and task performance of the balanced schizophrenia sample (n = 28) and the unbalanced sample (n = 46)

Characteristic	Patients with schizophrenia (n = 28)	Patients with schizophrenia (n = 46)
Demographic information
Age, y; mean ± SD	33.36 ± 9.24	35.17 ± 9.97
Sex, male/female	17/11	36/10
Site, Mannheim/Berlin/Bonn	28/0/0	46/0/0
School years, mean ± SD	11.39 ± 1.73	11.09 ± 1.70
Psychological assessments
MWTB, mean ± SD	26.10 ± 6.41	26.48 ± 6.78
WCST–preservation score, mean ± SD	21.26 ± 16.84	23.26 ± 15.22
WCST–false concepts, mean ± SD	24.75 ± 12.36	25.24 ± 13.19
CPZE, mean ± SD	384.24 ± 166.07	363.59 ± 151.69
PANSS-positive, mean ± SD	14.00 ± 4.10	13.87 ± 3.50
PANSS-negative, mean ± SD	15.29 ± 6.03	16.39 ± 5.54
Years of illness, mean ± SD	8.55 ± 7.49	9.17 ± 7.22
Number of episodes, mean ± SD	3.81 ± 2.58	3.90 ± 2.38
Number of hospital admissions, mean ± SD	4.14 ± 3.04	3.95 ± 2.68
fMRI task performance
Accuracy, %; mean ± SD	62.72 ± 19.36	59.60 ± 20.98
Reaction time, ms, mean ± SD	547.30 ± 295.10	564.99 ± 277.57
Head motion parameters
Sum frame-wise displacement, mm; mean ± SD	16.66 ± 6.29	16.74 ± 5.82
No. of scrubbed frames, mean ± SD	1.59 ± 1.71	1.35 ± 1.48
fMRI image quality
SNR ratio	90.16 ± 10.69	84.9 ± 15.00

CPZE, chlorpromazine equivalent(s).

Table S3.

Demographics of healthy controls in the phfMRI study

Characteristic	Participants of phfMRI study
Sample size	n = 37
Age, years; mean ± SD	25.30 ± 4.20
Sex, male/females	30/7
Education, y; mean ± SD	12.79 ± 0.77

Dynamic Network Community Structure During Working Memory.

In line with our expectations, the unaffected first-degree relatives, patients with schizophrenia, and healthy controls were significantly different in network flexibility [F(2,196) = 6.541, P = 0.002]. Post hoc tests confirmed that both the relatives and patients with schizophrenia showed a significant increase in network flexibility relative to the controls [controls vs. relatives: P_corrected (P_corr) = 0.017, controls vs. patients: P_corr = 0.023]. Although the arrangement of group means was consistent with the assumed genetic risk load of the groups (Fig. 2A), the difference between the relatives and the patient group was not significant (P_corr > 0.8).

Fig. 2.

Increased brain-wide network flexibility in patients with schizophrenia, unaffected first-degree relatives, and healthy controls. (A) Significant increases in the mean dynamic reconfiguration of modular brain networks in unaffected first-grade relatives (gray bar, REL) and patients with schizophrenia (black bar, SZ) in comparison to matched healthy controls (white bar, HC) [F(2,196) = 6.541, P = 0.002, corrected for sex, age, site, and intelligence]. Bars indicate mean values, and whiskers represent SEMs. (B) Mean flexibility of each subject for the functional parcellation as introduced by Power et al. (26). Red diamonds indicate the means. Note that the uncorrected flexibility estimates are shown. AUD, auditory; CO, cinguloopercular task control; DA, dorsal attention; DMN, default-mode network; FP, frontoparietal task control; SAL, salience; SM, sensory/somatomotor; SUB, subcortical; VA, ventral attention; VIS, visual. (C) Illustration of the location of the 10, 30, and 50 most important nodes in the brain, indicating a nonlocalized effect.

Local Versus Global Changes.

To examine if the observed average changes in network flexibility are driven by a biologically coherent subset of brain regions or are instead a general feature of dynamic network reconfiguration, we conducted two supplemental analyses. First, testing the average flexibility of 10 predefined functional systems described by Power et al. (26), we could not detect a significant main effect of systems or a significant interaction of group × systems [repeated measures ANOVA with system as a within factor; group as a between-subject factor; and age, sex, site, and MWTB as covariates: main effect of system: F(9,1,764) = 1.37, P = 0.199; interaction system × group: F(18,1,764) = 1.34, P = 0.152], arguing for a brain-wide and not a system-specific effect (Fig. 2B). Second, to support this notion further, we used random forest machine learning algorithms to identify optimally predictive combinations of node-wise flexibility measures. Flexibility measures were ranked based on their importance for accurate classification, and we used group-stratified cross-validation to explore classification accuracy for different numbers of the most highly ranked predictors. We observed that the classification accuracy remained similar when using 10, 30, or 50 nodes (accuracy of 61%, 64%, and 65%, respectively). The ranked variable importance did not indicate the presence of a well-localized and biologically meaningful set of predictive nodes (Fig. 2C), with even the 10 most important nodes distributed among seven of 10 functional systems (visual system: four nodes; cinguloopercular system: two nodes; default mode and subcortical, frontoparietal, and sensory somatomotor systems: one node each), supporting our interpretation of a more generalized, system-level effect rather than a focal, subnetwork-dependent effect. (For robustness of results to definition of parcellation and the edge-wise results, please see Figs. S1, S2, and S3.)

Fig. S1.

Location of the 50 most important edges for the discrimination between patients with schizophrenia and healthy controls. The illustration depicts the location of the 10, 30, and 50 most important edges in the brain as indicated by the machine learning analysis, suggesting a rather nonlocalized effect.

Fig. S2.

(A) Mean flexibility of each subject for the different anatomical parcellations. (B) Flexibility of each subject for the 10 most discriminative brain regions as indicated by the machine learning approach. Light gray indicates healthy controls, dark gray indicates healthy first-grade relatives, and black indicates patients with schizophrenia. The red diamonds indicate the means. Note that the uncorrected flexibility estimates are shown.

Fig. S3.

Mean node-wise brain network flexibility measures for healthy controls (A), unaffected first-degree relatives (B), and patients with schizophrenia (C). For visualization purposes, the size of the nodes and the colors are scaled with the mean flexibility.

Relationship to Clinical State and Antipsychotic Medication.

We did not detect a significant association between the network flexibility measure in the patient group and either chlorpromazine equivalents (27) (Spearman’s rho = −0.145, P = 0.480), positive and negative syndrome scale scores (positive scale: Spearman’s rho = 0.068, P = 0.730; negative scale: Spearman’s rho = −0.008, P = 0.966), illness duration (Spearman’s rho = −0.311, P = 0.159), or Wisconsin Card Sorting Test (WCST) performance (perseveration score: Spearman’s rho = 0.244, P = 0.210). However, because our stringent matching procedure likely biased our patient sample toward a disproportionally higher rate of well-performing patients (thereby restricting clinically meaningful variance in the cognitive measures), we explored the association of network flexibility with cognition in the original, unmatched patient sample. Although we could not detect any significant association with the typically ceiling n-back task performance parameters (reaction time and task accuracy), we found a significant positive correlation with the perseveration score of the WCST (Spearman's rho = 0.314, P = 0.033) (28), a well-established clinical measure of cognitive flexibility. The correlation of the two measures is illustrated in Fig. S4C.

Fig. S4.

(A) Illustration of individual mean flexibility (x axis) and task accuracy (y axis) (in percentage of correct 2-back answers) measures for each of the three groups (Spearman correlations: healthy controls: rho = −0.118, P = 0.167; relatives: rho = 0.125, P = 0.460; patients with schizophrenia: rho = −0.190, P = 0.334). (B) Illustration of individual mean flexibility (x axis) and reaction time (y axis) (in milliseconds) for each of the three groups (Spearman correlations: healthy controls: rho = 0.058, P = 0.497; relatives: rho = 0.148, P = 0.382; patients with schizophrenia: rho = 0.319, P = 0.099). (C) Illustration of individual mean flexibility (x axis) and perseveration performance scores (y axis) in the WCST in the unbalanced patient group containing 46 subjects (rho = 0.314, P = 0.033).

Effects of NMDA Receptor Challenge.

Application of DXM to 37 controls resulted in a significant increase in network flexibility compared with the placebo condition [repeated measures ANOVA with sex and age as covariates: F(1,34) = 5.291, P = 0.028; Fig. 3A]. We detected no differences between the placebo and the drug conditions in task performance [accuracy: t(36) = −0.802, P = 0.428; reaction time: t(36) = 1.313, P = 0.198], head micromovement parameters [mean frame-wise displacement: t(36) = 0.036, P = 0.971; number of motion-corrupted frames of the time series: t(36) = −0.063, P = 0.950], or fMRI signal quality [signal-to-noise-ratio: t(36) = 1.415, P = 0.166]. This observation makes the confounding influence of these variables on the effects of DXM on network flexibility unlikely.

Fig. 3.

Increased brain-wide network flexibility in healthy controls after pharmacological NMDA receptor challenge. (A) Significant increases in the mean dynamic reconfiguration of modular brain networks in healthy controls after application of dextrometorphan (DXM) [dark gray bars; repeated measures ANOVA placebo (PLA) versus DXM: F(1,34) = 5.291, P = 0.028, corrected for sex and age] relative to PLA (light gray bars). Bars indicate mean values, and whiskers represent SEMs. (B) Mean flexibility of each subject for the functional parcellation as introduced by Power et al. (26). The red diamonds indicate the means. Note that the uncorrected flexibility estimates are shown.

To support our notion that the drug effect is also rather a brain-wide effect, we repeated our analysis, testing the average flexibility of 10 predefined functional systems. We detected no significant drug × system interaction [repeated measures ANOVA with system and drug as within factors and with age and sex as covariates; interaction system × drug: F(6.62,225.12) = 2.00, P = 0.06], again suggesting an effect that was not driven by a single focal system (Fig. 3B).

Discussion

The architecture of the human brain is intrinsically organized into time-varying assemblies of functionally interacting nodes (16, 18), a dynamic neural community structure allowing for high adaptability of the system to changing environmental demands (29). Schizophrenia is increasingly recognized as a heritable brain disorder with altered neural network dynamics, yet the presence of alterations in the time-varying configuration of the functional connectome and their genetic and pharmacological implications require clarification. To address this gap in knowledge, we conducted neuroimaging studies in which we applied innovative methods from dynamic network neuroscience to fMRI data acquired during the performance of an established working memory task. We report significant increases in brain network flexibility during working memory in patients with schizophrenia, their unaffected first-degree relatives, and healthy individuals after pharmacological challenge with an NMDA receptor antagonist.

Our data extend the existing knowledge in network neuroscience in several notable directions. First, compared with healthy controls, the dynamic neural community structure of patients with schizophrenia showed an excess in flexibility in the working memory challenge. This observation extends prior network neuroscience data by demonstrating that in addition to the static network community structure (13, 14), the dynamic configuration of neural assemblies is altered in schizophrenia. More specifically, consistent with prior accounts (30), our data may indicate a temporally less stable, underconstrained, and possibly disintegrated (rather than overly rigid) network dynamic during working memory in schizophrenia. Importantly, the patient group was comparable to the controls for several potential confounds, including n-back task performance, head micromovement, and data quality measures, making these factors unlikely explanations for the outcome.

Second, we explored the neurogenetic basis of the dynamic neural network phenotype by studying healthy first-degree relatives of patients with schizophrenia relative to healthy controls without a family history of mental illness and to patients with schizophrenia. Compared with the controls, the relatives showed significant differences in dynamic network reconfiguration during working memory in a direction and magnitude that conform to the direction and magnitude of the patients. Healthy first-degree relatives are at increased risk for schizophrenia and share an enriched set of susceptibility variants with their affected family members (31). However, the relatives are devoid of confounding influences that impact brain physiology and can complicate the interpretation of patient neuroimaging data, particularly the effects of medication, substance use, or illness chronicity. Importantly, our relatives group was comparable to the controls for several other potential confounders, including demographics, n-back task performance, head micromovement, and fMRI data quality. These results suggest that these factors are not likely explanations for the observed excess in neural network flexibility in the genetically and familial at-risk individuals. Instead, our data support the notion that this dynamic network alteration is familial and may relate to the genetic liability for schizophrenia. Although the relatives’ data are the stronger evidence, the lack of correlations with antipsychotic drug dose and other clinical state measures in the patient group supports, at least indirectly, the idea that the dynamic network phenotype reflects a neurogenetic trait feature or “intermediate phenotype.” This notion is further supported by the detected association between network flexibility and WCST perseveration score in the patients. The WCST taps into a cognitive domain (cognitive set shifting or flexibility) that may plausibly relate to the quantified dynamic network flexibility measure and has previously been related to the genetic predisposition to schizophrenia (32).

Prior empirical data and theoretical accounts converge on the idea that altered oscillatory activity, temporal coordination, and excitatory–inhibitory balance of neural networks have a central role in the pathophysiology of schizophrenia (5–7). Specifically, building on the widespread topological network alteration observed in schizophrenia (33), it is assumed that the alterations in the temporal organization of the brain functional connectome represent a widely distributed, rather than a regionally circumscribed, impairment (6, 24). Although methodological differences between electrophysiology and fMRI-based network approaches call for a cautious interpretation, our results are in line with previously described global effects. The temporal coordination of node allegiances to neural systems is at the core of the quantified network flexibility measure, and our node-wise supplemental analyses highlighted a system-wide (rather than a regionally confined) increase in network flexibility in the patients and relatives. However, these findings do not argue against a differential or preferential involvement of certain brain areas at different hierarchical levels (34). For the proposed excitatory–inhibitory imbalance of neural networks in schizophrenia, altered NMDA receptor-depending input to fast-spiking γ-aminobutyric acid (GABA) interneurons has been repeatedly highlighted as a key molecular candidate mechanism for schizophrenia (5, 25, 35, 36). Indeed, prefrontal executive functions are critically regulated by NMDA receptor signaling (37, 38), and a hypofunction of NMDA receptors has been implicated in the cognitive impairments (39, 40) and the underlying genetic risk architecture (35, 36) of schizophrenia.

Third, specifically to probe the question of whether alterations in NMDA receptor signaling impact the proposed dynamic network intermediate phenotype, we conducted a pharmacological challenge experiment consisting of a single dose of an NMDA receptor antagonist before fMRI. Under DXM, we observed a significant reorganization of the dynamic configuration of the functional connectome in the working memory challenge. Specifically, the data pointed to a drug-induced “network hyperflexibility” in the hypoglutamatergic state, a direction of change that is consistent with patients with schizophrenia and the healthy at-risk individuals. Notably, the phfMRI protocol followed an observer-blinded, placebo-controlled, randomized, cross-over design, and other critical variables, such as n-back task performance, head micromovement, and fMRI data quality, were comparable between the sessions. These circumstances makes the influence of secondary confounders on the observed drug effect unlikely explanations. Instead, the data conform to popular theories assuming a critical role for NMDA receptor function in the excitatory–inhibitory balance and temporal coordination of neural networks (6, 24, 25) and related connectivity alterations in schizophrenia (1, 6, 24, 25), and might indicate a plausible molecular mechanism for the putative genetic risk-associated dynamic network phenotype we identified in this work.

Fourth, in the context of prior empirical data and theoretical accounts, the combined results of our experiments might offer a conceptual means of linking network properties to underlying molecular and cellular pathways: The transient communications patterns between distant regions (which we measure by network flexibility) are established through synchronous firing of mostly long-range pyramidal glutamatergic neurons. The activity of these neurons is regulated via a feedback loop involving fast-spiking inhibitory and parvalbumin (PV)-containing interneurons, which regulate the firing rate of the long-range pyramidal glutamatergic neurons through GABAergic input (8, 41). This feedback loop is dependent on NMDA receptor functioning, because input from the long-range pyramidal glutamatergic neurons into the PV interneurons is transduced via NMDA receptors. As outlined above, an altered NMDA receptor-depending input to those fast-spiking GABAergic interneurons has been repeatedly highlighted as a key candidate mechanism in schizophrenia pathophysiology. Uhlhaas and Singer (6) propose that a hampered feedback loop leads to asynchronous firing of long-range pyramidal neurons and, consequently, to impaired synchrony between two distant brain regions (5, 6). The impaired synchrony is believed to translate to more time-variant connectivity patterns between brain regions, which may plausibly result in a less stable community structure of the network in schizophrenia and after NMDA receptor blockade.

Our study has several limitations worth noting. First, although our sample sizes were relatively large and the observed effects were consistent across a range of analyses, we cannot fully rule out that our findings in the first-degree relatives may reflect familial rather than genetic influences. Second, we cannot rule out that in addition to NMDA receptor blockade, the mechanism of action of DXM involved partial binding to other receptor types (e.g., σ1, serotonin transporter). However, DXM binds to the same receptors as ketamine (42), and the applied dose results in brain concentrations similar to the brain concentrations evoking an NMDA receptor block in vitro (43–45). Third, although we demonstrate similar alterations in network reconfiguration in patients with schizophrenia, healthy first-degree relatives, and healthy subjects receiving a single dose of an NMDA receptor antagonist, this observation does not definitively prove that both phenomena are provoked by the same underlying biological mechanism or that NMDA receptors have a causal role in generating the schizophrenia-related network alterations. Here, further research is warranted, including pharmacological experiments in schizophrenia patients and their healthy first-degree relatives. Fourth, although the idea of an excitatory–inhibitory imbalance is a popular concept in current psychiatric neuroscience (34, 46), the biological mechanisms require further clarification. Specifically, although the ubiquitous presence of glutamatergic neurons and NMDA receptors across the brain makes distributed genetic risk effects at the neural network level plausible, further translational research is needed to dissolve some of the ambiguities of this pathophysiological concept. Finally, although the applied dynamic network neuroscience methods are innovative and have gained increasing impetus in the neuroscience community recently (16, 18, 47, 48), more work is needed to define the relative similarities and differences in the reconfiguration of dynamic neural networks in healthy individuals and clinical populations under different empirical contexts with different behavioral demands.

In summary, we provide evidence for a potential dynamic network intermediate phenotype for schizophrenia manifesting as a “hyperflexible” configuration of the brain functional connectome during working memory function. We provide evidence that the phenotype is a possible indicator of the genetic liability to schizophrenia, and is modulated by NMDA receptor antagonism in a way that conforms to existing concepts of a critical role for NMDA receptor input for the temporal organization of neural networks and related impairments in the pathophysiology of schizophrenia. Finally, our data may provide a valuable basis for the identification of the genetic contributions to the more proximal glutamatergic mechanisms of altered neural network dynamics and suggest a useful system-level target for the study of the effects of novel and established antipsychotic treatments.

Methods

Participants, Data Acquisition, and Preprocessing.

All participants provided written informed consent for protocols approved by the Institutional Review Boards of the Medical Faculty Mannheim of the University of Heidelberg, the Medical Faculty of the University of Bonn, and the Charité–University Medicine Berlin. A total of 337 subjects (165 healthy controls, 50 first-degree relatives, 81 patients with schizophrenia, and an additional sample of 41 healthy controls for the phfMRI experiment) were included in this study (details and the supporting information for a detailed description of the recruitment and inclusion criteria are provided in Tables S1–S4). Further processing, quality control, and balancing (methods details are provided below) left a total of 204 subjects for the genetic risk analysis (139 healthy controls, 37 first-degree relatives, and 28 patients with schizophrenia) and 37 healthy controls for the phfMRI experiment.

Blood oxygen level-dependent fMRI was acquired for all subjects while performing an established n-back working memory task (49). Details on the task, acquisition parameters, and quality assurance measures can be found in SI Methods. Functional image preprocessing of the working memory task was performed in SPM8 using previously described (4, 22, 23) standard procedures. We then extracted the mean time series from 5-mm spheres that were centered at 270 published peak coordinates and corrected it for signals from white matter and cerebrospinal fluid, as well as the six head motion parameters from the realignment step by regression as previously described (23, 50). To minimize the potential for artificial inflation of coherence estimates, we removed the mean effects of the task conditions by regression (4, 18, 23). In addition, because in-scanner head motion poses a severe problem for dynamic functional connectivity analysis, all frames of the time series with a frame-wise displacement of >0.5 mm were scrubbed and interpolated using spline interpolation. Further, all subjects with more than 5% of scrubbed frames after correction were excluded from the analysis.

Estimation of Network Connectivity and Identification of Putative Functional Communities.

For the connectivity analysis, we first applied a sliding time window with a length of 15 vol (51) [and no gap between the windows (16, 18)] to all node time series. We then used wavelet coherence to estimate the functional connectivity between each pair of brain nodes using the MATLAB (MathWorks) package for cross-wavelet and wavelet coherence analysis (52), as described previously (16, 18, 48). For each subject, this estimation yielded 114 (or 109 in the pharmacological challenge study) weighted matrices describing the functional connectivity in each time window during the n-back task performance. For each subject, the resulting coherence matrices were partitioned into time-dependent communities using a multilayer community detection algorithm based on a Louvain-like locally greedy optimization method to maximize the modularity quality function (15, 16, 18, 48). For each optimization, we calculated a time-dependent network flexibility matrix (18), F, whose binary elements, F_ij, indicate whether (1) or not (0) a node i changes its community at the transition j between two consecutive time windows. Individual whole-brain network flexibility estimates were obtained by averaging the mean network flexibility matrix over time windows and over brain regions.

Supplemental Machine Learning Analysis.

Random forest machine learning was used to identify combinations of nodes that could optimally differentiate between patients with schizophrenia and controls (53). Details are provided in SI Methods.

SI Results

Influence of Head Movement and Signal Quality.

The patient, relative, and control groups were balanced for head motion [ANOVA main effect of group for frame-wise displacement: F(2,246) = 0.72, P = 0.489] and signal-to-noise ratio (SNR) [ANOVA main effect of group for SNR: F(2,246) = 1.82, P = 0.164]. Including neither frame-wise displacement nor SNR as additional covariates of no interest into the univariate ANOVA model changed the reported group differences for brain network flexibility [ANOVA for main effect of group, corrected for age, sex, site, premorbid intelligence, and frame-wise displacement: F(2,240) = 6.74, P = 0.001; ANOVA for main effect of group, corrected for age, sex, site, premorbid intelligence, and SNR: F(2,240) = 5.47, P = 0.005].

Robustness of Results.

Omega value.

The multilayer community detection algorithm has two free parameters: gamma and omega. Whereas gamma is a structural resolution parameter and tunes the number and size of detected communities, omega affects the temporal stability of the detected communities. In our analyses, we adopted the most abundantly chosen omega value in dynamic network neuroscience by setting this parameter to 1 (18, 47). Nonetheless, to validate that our results are independent of the particular choice of omega, we repeated our analysis for a range of omega values (omega = 0.2, 0.6, 0.8, 0.9, 1, 1.1, 1.2, 1.4, 1.6). Using repeated measures ANOVA with group as a within-subject factor and sex, age, site, and premorbid intelligence (as in the main analysis) as covariates, we detected a main effect of omega [F(8,1568) = 120.64, P < 0.001], but no significant group by omega interaction effect [F(16,1568) = 1.140, P = 0.311]. Also, the reported between-subject group effect remained significant [F(2,196) = 4.14, P = 0.017], which argues against a relevant influence of the choice of omega on the reported findings.

Window size.

Our choice of using a window size of 15 repetition times (TRs; 30 s) was based on the intrinsic design of the n-back task, which is a block design with a block length of 15 TRs. Although 15 vol is a rather short window for estimating the connectivity between 270 nodes, it is in line with recent recommendations by Leonardi and van de Ville (51), who demonstrated the adequacy of a 30-s time window for the quantification of the dynamic reconfiguration of neural networks in the frequency range of 0.08–0.15 Hz. Second, the n-back task applied here uses a “repeated measurement” design by repeating each block four times. Effectively, because we report averages over the entire task, this design should give us more robust estimates of the true underlying community structure. Third, we directly examined this question by repeating our analysis with a broader window of 17, 20, and 30 TRs. As expected, we found a significant main effect of window length [repeated measures ANOVA with group as a between-subjects factor and sex, age, site, and premorbid intelligence as covariates: F(3,585) = 223.63, P < 0.001], but no significant interaction of window length with group [F(6,585) = 1.37, P = 0.224]. Also, the reported between-subject group effect remained significant [F(2,195) = 7.29, P = 0.001], which argues against a relevant influence of the choice of window size on the reported findings.

Removal of task block-related variance.

As in previous reports (4, 18, 23), we removed the mean task variance of task blocks to correct for coactivation patterns that would artificially inflate our connectivity measures. Importantly, this correction only removes the effects of the block structure of the task by removing the difference in the mean intermodal connectivity between conditions. Nonetheless, we reanalyzed our data without correction for task variance and detected a main effect of task variance removal [F(1,195) = 8.166, P = 0.005], but we did not find a significant “group-by-task removal” interaction effect [F(1,194) = 0.496, P = 0.609]. Also, the reported between-subject group effect remained significant [F(2,194) = 5.433, P = 0.005], which argues against a relevant influence of the task block-related signals on the reported findings.

Influence of scanning site.

Because diagnostic group and scanning site are confounded in our sample, we repeated our analysis with a considerably reduced sample from one site only (43 healthy controls, 15 first-degree relatives, and 28 patients with schizophrenia). Although the main effect of group remained significant [F(2,80) = 3.84, P = 0.026], as well as the post hoc comparison testing of schizophrenic patients versus healthy controls (P_corr = 0.033), the difference between healthy controls and first-degree relatives was not significant anymore (P_corr = 0.287). Because this result is probably due to the drastic reduction in group size for the first-degree relative group, we also used a regression model to test if our data are still in accordance with the assumed genetic risk load [healthy controls < first-degree relatives < patients with schizophrenia: b = 0.002, t(4,81) = 2.74, P = 0.008]. Overall, these additional analyses demonstrate that scanning site is unlikely to explain our findings.

Supplemental analysis with machine learning methods.

Random forest machine learning algorithms were used to identify optimally predictive combinations of node-wise and edge-wise flexibility measures. Flexibility measures were ranked based on their importance for accurate classification, and we used group-stratified cross-validation to explore classification accuracy for different numbers of the most highly ranked predictors. We observed that the classification accuracy remained similar when using 10, 30, or 50 nodes (accuracy of 61%, 64%, and 65%, respectively). The ranked variable importance did not indicate the presence of a well-localized set of predictive nodes (Fig. 2C and Fig. S1), supporting our interpretation of a more generalized, system-level effect rather than a focal, subnetwork-dependent effect. All node-wise flexibility measures together yielded a mean sensitivity and specificity of 49% and 87%, respectively. Independent prediction of relatives showed that a mean of 36% of these individuals were classified as patients, supporting an intermediate position of these subjects between the two diagnostic groups used for classifier training.

Analysis of edge-wise flexibility measures yielded similar results, showing no relevant changes when including 10, 30, or 50 edges (accuracy of 65%, 66%, and 64%, respectively).

SI Methods

Participants.

All participants provided written informed consent for protocols approved by the institutional review boards of the medical faculties in Mannheim, Bonn, and Berlin. A total of 337 subjects were included in this study (details are provided in Tables S1–S4). From a total sample of 215 subjects who were recruited from local residents’ registration offices, psychiatric hospitals, support groups, and media advertisements in and around Mannheim, Bonn, and Berlin in Germany, we finally included 139 healthy controls and 37 unaffected first-degree relatives of patients with schizophrenia after careful balancing for demographic, neuropsychological, and imaging quality parameters, including correction for sharp head movements. The same quality procedures applied to the 41 healthy participants of the phfMRI experiment recruited by flyers from the Mannheim area, leaving 37 participants for the final analysis. Exclusion criteria included the presence of a lifetime history of psychiatric, neurological, or significant general medical illness; pregnancy; a history of head trauma; and current alcohol or drug abuse. None of the healthy volunteers had a first-degree relative with a psychiatric disorder or received psychotropic pharmacological treatment. All participants were of European descent. The clinical diagnosis of the index patients of the first-degree relatives were verified by a trained psychiatrist or psychologist.

From a total sample of 82 patients with schizophrenia, we included 28 individuals with balanced n-back task performance, head micromovement, and fMRI signal quality parameters as described above. This selection was done because these confounds are known to compromise functional connectivity analyses (23, 54), and we wished to minimize their influence on the study outcome. The patients were recruited from the Department of Psychiatry and Psychotherapy at the Central Institute of Mental Health in Mannheim. General inclusion criteria were a Diagnostic and Statistical Manual of Mental Disorders, 4th Edition diagnosis of a schizophrenia spectrum disorder, monotherapy with antipsychotics, and a stable medication (less than 10% change in dose) and clinical state [total positive and negative syndrome scale (PANSS) score of <70] for at least 2 wk as verified by the Clinical Global Impression (CGI-S) and the PANSS scales.

MRI Data Acquisition and Functional Preprocessing.

Functional data acquisition and preprocessing were performed as previously described (18, 22, 23). In short, we acquired blood oxygen level-dependent fMRI on three identical scanners (3T Siemens Trio) in Mannheim, Bonn, and Berlin. Identical sequence protocols were used at all sites. Functional data were acquired using an echo-planar imaging sequence with the following scanning parameters: TR/echo time = 2,000/30 ms, α = 80°, 28 axial slices of 4-mm thickness, 1-mm gap, descending acquisition, field of view = 192 mm, and 64 × 64 matrix. Functional image preprocessing of the working memory task was performed in SPM8 using previously described (4, 22, 23) standard procedures, including realignment of the data to the first image of the time series, slice-time correction, spatial normalization to the Montreal Neurological Institute template (resulting in a resampled voxel resolution of 3 mm³), and spatial smoothing with a 9-mm full-width at half-maximum Gaussian kernel.

fMRI Quality Assurance Measures.

fMRI data quality assurance followed previously published methods (22). For quantification of head micromovements, we calculated the mean voxel-based, frame-wise displacements across the time series (54). Individual SNRs were calculated using the NYU CBI Data Quality toolbox (cbi.nyu.edu/software/dataQuality.php). We used ANOVA models for group comparison.

Working Memory Task.

Brain function during working memory was studied with an established n-back paradigm during fMRI as described in more detail previously (4, 18, 22). Briefly, the n-back task is a block-designed working memory task in which a series of numbers are displayed on a screen in a random order at set locations in a diamond-shaped box (stimulus presentation time, 500 ms; interstimulus interval, 1,500 ms). In the 2-back condition, the participants are asked to encode a currently seen number, simultaneously recall the number seen two presentations earlier, and press the button corresponding to the position of the number two presentations earlier. In the control condition, the subjects are asked to press the button corresponding to the position of the currently seen number. The task is presented in eight blocks of 28 s each, with alternating epochs of 0-back and 2-back performance (task duration: 4.1 min or 124 whole-brain scans).

NMDA Receptor Challenge.

On each visit during the phfMRI study, the subjects received an oral dose of placebo or 120 mg of DXM, a potent noncompetitive NMDA receptor antagonist that induces brain concentrations in humans similar to the brain concentrations elicited by NMDA receptor blockade in vitro at this dose (42–44). The fMRI experiment was scheduled 2.5 h after oral intake of the capsules, following a prior report on the expected time interval until the peak plasma level is reached in humans (55). Drug and placebo conditions were assigned to testing days in a randomized and counterbalanced order. Both the investigators and subjects were blinded to the treatment assignment.

Estimating Networks of Functional Connectivity.

After preprocessing, the mean time series in 5-mm spheres around coordinates defined by Power et al. (26) were extracted. Because the nodal template defined by Power et al. (26) does not provide coverage of the hippocampus, amygdala, and nucleus accumbens, we manually added bilateral coordinates for these three regions based on three metaanalysis articles (56–58).

Using a sliding time window with a length of 15 vol and no gap between windows, we use the Morlet wavelet transform to estimate the functional connectivity between each pair of brain regions, as described previously (noc.ac.uk/using-science/crosswavelet-wavelet-coherence) (18). Unlike standard correlation approaches, wavelet coherence estimates have the benefits of a high sensitivity to small signal changes in nonstationary time series with noisy backgrounds (59). We note that due to the short time intervals of only 30 s and due to our TR of 2 s, we restricted our analysis to the wavelet scales corresponding approximately to the frequency band of 0.08–0.15 Hz. For each subject, this procedure yielded 114 (or 109 in the pharmacological challenge study) weighted matrices describing the temporal evolution of functional connectivity across the task.

Identifying Putative Functional Modules.

For each subject, the resulting matrices were partitioned into time-respecting modules using a multilayer community detection algorithm introduced and extensively described by Mucha et al. (15) and applied to neuroimaging data by Bassett et al. (16) and Braun et al. (18). In short, for each subject, the 114 (109 for the pharmacological dataset) weighted adjacency matrices can be combined to form a rank 3 adjacency tensor A that can be used to represent time-dependent or multilayered networks. One can thereby define a multilayer modularity:

$$\text{Q}=\hspace{0.17em}\frac{1}{2\mathrm{\mu }}\hspace{0.17em}{\displaystyle \sum }_{\text{ijlr}}\left\{\left({\text{A}}_{\text{ijl}}-{\hspace{0.17em}\mathrm{\gamma }}_{\text{l}}{\text{P}}_{\text{ijl}}\right){\mathrm{\delta }}_{\text{lr}}+{\hspace{0.17em}\delta }_{\text{ij}}{\mathrm{\omega }}_{\text{jlr}}\right\}\delta \left({\text{g}}_{\text{il}},{\text{g}}_{\text{jr}}\right),$$ [S1]

where the adjacency matrix of layer 1 has components A_ijl, the element P_ijl gives the components of the corresponding layer l matrix for the optimization null model, γ_l is the structural resolution parameter of layer l, the quantity g_il gives the community assignment of node i in layer l, the quantity g_jr gives the community assignment of node j in layer r, the element ω_jlr gives the connection strength (i.e., “interlayer coupling parameter,” temporal resolution parameter) from node j in layer r to node j in layer l, the total edge weight in the network is $\mathrm{\mu }=\frac{1}{2}{\displaystyle \sum _{\text{jr}}}{\text{k}}_{\text{jr}}$, the strength (i.e., weighted degree) of node j in layer l is ${\text{k}}_{\text{jl}}={\hspace{0.17em}\text{k}}_{\text{jl}}+{\text{c}}_{\text{jl}}$, the intralayer strength of node j in layer l is ${\text{k}}_{\text{jl}}={\displaystyle \sum _{\text{i}}}{\text{A}}_{\text{ill}}$, and the interlayer strength of node j in layer l is ${\text{c}}_{\text{jl}}={\displaystyle \sum _{\text{r}}}{\mathrm{\omega }}_{\text{jlr}}$.

This procedure yields a community assignment for every region and every time window, which indicates the module allegiance. Maximizing this modularity quality function is nondeterministic polynomial-time hard (NP-hard), and we therefore use a Louvain-like locally greedy heuristic algorithm (15) to seek the optimal modularity index Q. Due to the heuristic nature of the algorithm and to the near-degeneracy of the optimization landscape of the multilayer modularity quality function, each independent run of the algorithm provided a slightly different partition of network nodes into communities across the 114 (109) time slices (or network layers). Therefore, we repeated the modularity optimization procedure 100 times for each subject. For each repetition, we calculated a flexibility change matrix, F∆, whose binary elements F∆_i,j indicate if node i changes its module allegiance at the transition j between two consecutive time windows. Averaging F∆ over all repetitions of the modularity optimization, we obtained the average flexibility change matrix for each subject, F∆, whose elements F∆_i,j estimate the probability that a brain region changes its allegiance to putative functional modules between any two consecutive time windows. Brain-wide or network flexibility, fnet, is then calculated as the average flexibility of all regions over all transitions: fnet = $\overline{\text{F}\mathrm{\Delta }}$.

Neuropsychological Assessment.

The WCST evaluates general executive function and requires strategic planning, the ability to use feedback, the ability to shift cognitive sets, and the ability to modulate impulsive responding. The WCST was administered to 73 of the total 81 patients with the standard instructions described by Heaton et al. (28): Patients viewed one card at a time on a computer screen. Each card contained stimuli that were characterized by three different stimulus dimensions: shape (circle, crosses, triangles, or stars), color (yellow, blue, red, or green), and quantity (one, two, three, or four stimuli). Patients were asked to sort the cards into four piles on the screen headed by a key card representing a distinct stimulus constellation. At the beginning of each trial, patients started sorting without any information on the sorting rules and had to generate hypotheses about the sorting rules from feedback (“correct” or “incorrect”) provided by the computer after each sort. To obtain correct feedback, patients focused on one stimulus dimension and ignored the other two dimensions. After 10 consecutive correct sorts, the sorting rule changed, requiring patients to perform an extradimensional shift. Each stimulus dimension was relevant twice so that, overall, patients performed six extradimensional shifts. The task ended when all shifts had been mastered or when all cards had been used. The main outcome parameter to characterize performance was the perseverative score and the number of perseverative errors, which indicate how often a subject erroneously adhered to a specific category.

Supplemental Machine Learning Analysis.

To explore further which brain regions contributed most to the observed differences in whole-brain network flexibility, we used a random forest machine learning algorithm to identify combinations of nodes that could optimally differentiate between patients with schizophrenia and controls (53). We began by ranking variables based on their importance, which we defined as the mean decrease in classification accuracy in out-of-bag samples. Variable importance estimation was repeatedly performed during cross-validation, and accuracy of the resulting classifier was determined on the fold not used for training. Due to the low number of patients with schizophrenia, we performed stratified, 28-fold cross-validation, where each fold contained data from one patient and an equal number of randomly selected controls.

Random forests were built using 1,000 trees, and classifier training was performed using group-stratified sampling. Due to the high variable number for edge-wise flexibility analyses, random forest variable selection was preceded by correlation-based feature selection, identifying the 300 edges most strongly associated with the outcome based on Spearman’s correlation. All machine learning performance estimates are mean values from 20 independent repeats of the cross-validation procedure. The means of sensitivity and specificity, averaged across cross-validation folds, were used to quantify classification accuracy.

Additionally, we explored whether a subset of anatomically or functionally meaningful connections drove the whole-brain flexibility difference. To that end, we constructed for each subject a modular allegiance matrix T whose binary elements Tij indicate if two nodes, i and j, have been assigned to the same module or not. Summing up all modular allegiance matrices for each subject, we obtained the consensus matrix Tcons, whose elements indicate how often two nodes have been assigned to the same module over a set of matrices, again indicating how often two nodes have been in the same module over the entire time of the task. These edge-wise reconfigurations were then rendered into the random forest machine learning algorithm described above to identify the set of edge-wise reconfigurations that optimally differentiated between patients with schizophrenia and controls (Fig. S1).

Random forest machine learning was performed using the software package R (https://www.r-project.org/).

Statistical Analysis.

Statistical analyses were conducted using SPSS (version 21; IBM) with an alpha level set at P < 0.05 for all tests. For differences in network flexibility between healthy controls, first-degree relatives, and patients with schizophrenia, we used an analysis of covariance model with the individual whole-brain network flexibility estimates as dependent variables; group as a factor; and age, sex, site, and premorbid intelligence level as covariates of no interest. In the case of significant group differences, we conducted pairwise post hoc t tests with Bonferroni correction for multiple comparisons. In patients, the potential relationship between network flexibility, antipsychotic drug dose (expressed in chlorpromazine equivalents), and neuropsychological (WCST) and clinical parameters (illness duration and illness severity as indexed by the CGI-S and PANSS) was explored using Spearman’s rank order correlation test.