Intrinsic Connectivity Patterns of Task-Defined Brain Networks Allow Individual Prediction of Cognitive Symptom Dimension of Schizophrenia and Are Linked to Molecular Architecture

BACKGROUND: Despite the marked interindividual variability in the clinical presentation of schizophrenia, the extent to which individual dimensions of psychopathology relate to the functional variability in brain networks among patients remains unclear. Here, we address this question using network-based predictive modeling of individual psychopathology along 4 data-driven symptom dimensions. Follow-up analyses assess the molecular underpinnings of predictive networks by relating them to neurotransmitter-receptor distribution patterns. METHODS: We investigated resting-state functional magnetic resonance imaging data from 147 patients with schizophrenia recruited at 7 sites. Individual expression along negative, positive, affective, and cognitive symptom dimensions was predicted using a relevance vector machine based on functional connectivity within 17 meta-analytic task networks following repeated 10-fold cross-validation and leave-one-site-out analyses. Results were validated in an independent sample. Networks robustly predicting individual symptom dimensions were spatially correlated with density maps of 9 receptors/transporters from prior molecular imaging in healthy populations. RESULTS: Tenfold and leave-one-site-out analyses revealed 5 predictive network-symptom associations. Connectivity within theory of mind, cognitive reappraisal, and mirror neuron networks predicted negative, positive, and affective symptom dimensions, respectively. Cognitive dimension was predicted by theory of mind and socioaffective default networks. Importantly, these predictions generalized to the independent sample. Intriguingly, these two networks were positively associated with D 1 receptor and serotonin reuptake transporter densities as well as dopamine synthesis capacity. CONCLUSIONS: We revealed a robust association between intrinsic functional connectivity within networks for socioaffective processes and the cognitive dimension of psychopathology. By investigating the molecular architecture, this work links dopaminergic and serotonergic systems with the functional topography of brain networks underlying cognitive symptoms in schizophrenia.

Precise conceptualization of schizophrenia symptoms in terms of their underlying dimensional structure and associated neurobiology remains a challenge (1). Work focused on the subscales (positive, negative, and general symptoms) of the Positive and Negative Syndrome Scale (PANSS) (2) has not provided a clear understanding of underlying brain circuitry (3)(4)(5). We recently introduced a novel 4-dimensional conceptualization of schizophrenia symptomatology that is stable and generalizable across populations and settings (5). Because each symptom dimension captures a different clinical facet of schizophrenia (6)(7)(8), we would expect these to show differential relationships with functional brain networks. Identification of robust symptom-brain relationships (e.g., connectivity patterns and molecular substrates) in turn is key for developing reliable biomarkers and targeted treatments.
Previous studies proposed that abnormal brain connectivity might be a precipitating factor for schizophrenia (9,10), questioning region-based analyses but resonating with the dysconnection hypothesis (9)(10)(11)(12). Although resting-state functional magnetic resonance imaging (fMRI) reveals broad patterns of aberrations that may underlie the pathophysiology of schizophrenia (12)(13)(14)(15), the links between targeted symptom dimensions and associated connectivity patterns within distinct functional systems remain largely unknown. Pioneering work has explored symptom-brain associations based on regional activity and intrinsic connectivity networks using univariate group-level correlative approaches, but the results have been largely inconsistent (16)(17)(18)(19). The clinical complexity of schizophrenia together with the differences in patient populations, scanners, and study protocols across sites may have led to divergent results, posing a major challenge for establishing generalizable network-symptom relationships. Application of multivariable machine learning and cross-validation strategies to multisite data and validation of the resulting models on independent datasets is thus needed to derive robust network-symptom associations (20).
It needs to be cautioned, though, that intrinsic connectivity networks cannot readily be interpreted relative to cognitive and mental processes, owing to their unconstrained and taskindependent nature (21). In contrast, meta-analytic functional networks are derived from task-activation data, that is, the identified networks consist of brain areas robustly engaged in specific tasks and therefore mental processes (22,23). Metaanalytic networks thus provide a promising avenue to characterize association between functionally meaningful systems and specific symptom dimensions. Considering these advantages, we performed multivariable machine learning on metaanalytically defined networks to explore predictive relationships between network-specific connectivity and dimensional psychopathology.
In order to facilitate a link to treatment, we furthermore explored whether robustly symptom-related functional networks would mirror the spatial topography of underlying molecular features, given that functional brain systems relate to molecular architecture (24)(25)(26)(27). Specifically, connectivityneurotransmitter coupling has been proposed and observed in healthy populations (28,29). Similarly, network dysconnectivity in schizophrenia has been associated with altered neurotransmission (30,31) involving dopaminergic, serotonergic, GABAergic (gamma-aminobutyric acidergic), and glutamatergic pathways (32)(33)(34)(35). It is interesting to note here that current antipsychotic drugs mainly targeting the dopamine system are primarily effective against positive rather than negative or cognitive symptoms (36,37). Understanding the molecular substrate of specific dimensions of psychopathology may therefore provide leads on new treatment strategies.
We therefore assessed a broad range of meta-analytic networks relating to social, affective, executive, memory, language, and sensory-motor functions regarding their predictive power for individual positive, negative, affective, and cognitive symptom dimensions (6) in schizophrenia. Machine learning approaches with a stringent validation sequence of 10-fold cross-validation, leave-one-site-out analyses, and generalization to an independent sample was implemented to identify robust predictions. Subsequently, whole-brain density maps of 9 receptors/transporters from prior in vivo molecular imaging studies were used to investigate the molecular architecture spatially coupled to the identified predictive networks.

METHODS AND MATERIALS Sample
A total of 147 patients with schizophrenia from 7 centers located in Europe and the United States represented the main sample (Table S1; Supplement). These sites differed significantly in illness duration (p , .001) (Table S1). An independent sample with 117 patients with schizophrenia (Table S2; Supplement) retrieved from the Bipolar-Schizophrenia Network on Intermediate Phenotypes (B-SNIP) database (38) was used for independent model validation. For both samples, diagnosis of schizophrenia was established based on DSM-IV, DSM-IV-TR, or ICD-10 criteria [see Supplement and (6)]. These international datasets cover a broad range of clinical states, settings, and medical systems, facilitating identification of robust network-symptom associations. Current drug dosages of antipsychotic medication were olanzapine-equivalent transformed (39). For each site, subjects gave written informed consent, and study approval was given by the respective ethics committees or institutional review boards. Approval for pooled reanalysis was provided by the ethics committee of the University of Düsseldorf, Germany.

Calculation of Dimensional Symptom Scores
Severity of psychopathology was assessed using the PANSS (2). The 30 PANSS items were compressed into 4 (negative, positive, affective, and cognitive) symptom dimensions ( Figure S1A) identified in our prior factorization analysis on two large multisite schizophrenia samples as stable and wellgeneralizable across populations, settings, and medical systems (6). Each dimensional score for an individual was calculated as the dot product between the coefficients loaded on the particular dimension within the dimensional structure ( Figure S1) and the 30 PANSS item scores of that individual (as implemented in DCTS [a dimensions and clustering tool for assessing schizophrenia symptomatology; http://webtools. inm7.de/sczDCTS/]). Higher scores on a dimension indicate more-severe symptoms ( Figure S1B).

Definition of Functional Brain Networks
Seventeen functional networks, which cover a broad range of domains reflecting cognitive, socioaffective, and sensorymotor functions that have been implicated in schizophrenia, were used (Table S3). These networks were based on coordinate-based meta-analyses (21,22) and represent regions demonstrating convergent activations associated with specific functional domains across many prior task-fMRI studies. They thus provide the best a priori estimate of the location of specific functional networks and were therefore here assessed by resting-state fMRI in new subjects. For convenience, we grouped these 17 networks into 6 broad functional domains ( Figure 1; Table S3), though it must be stressed that each network was analyzed separately.

fMRI Data Processing
All resting-state fMRI scans (Tables S4 and S5) were preprocessed using SPM12 (http://www.fil.ion.ucl.ac.uk/spm) [see (6) and Supplement]. After excluding subjects with excessive head motion (40) or poor image quality (see Supplement), 126 and 100 patients with schizophrenia were retained in the main and B-SNIP validation samples, respectively (Table 1). Head motion differed significantly between the sites but did not correlate with the residuals of any symptom dimensions after adjusting for age, gender, and site. Still, we adjusted head motion effects in predictive modeling as a conservative approach to rule out any possible predictability based on movement.
White matter and cerebrospinal fluid signals as well as 24 head motion parameters (41,42) were regressed out from the overall fMRI time series (42), but the global mean signals were not removed given ongoing controversies (43,44). Voxels within a 6-mm sphere around each meta-analytically derived peak-activation coordinate formed a node. The first eigenvariate of the time series from all the voxels within each node was calculated as the regional characteristic time series containing better signal-to-noise ratio than the peak voxel alone.
For each network, Pearson's correlation was calculated between the characteristic time series for each pair of the nodes, resulting in an N 3 N matrix of resting-state functional connectivity (rsFC). The upper triangle of the resulting connectivity matrix was extracted, linearized, and Fisher's Z transformed to provide the features for our predictive modeling.

Prediction of Symptom Dimensions Using Network rsFC
Multivariable regression via a relevance vector machine (RVM) (45) was implemented and evaluated using Figure 1. Overview of the 17 functional networks derived from prior coordinate-based meta-analytical studies. Networks were assigned to 6 broad domains according to their main functional roles, as implicated in the tasks included in the source publications of these meta-analytical networks, in multiple neurocognitive and socioaffective processes. We also note that networks such as the task-deactivation default mode network (22) and the extended socioaffective default network which is derived from default mode network regions (71) can be engaged during multiple processes (69,70). Details are provided in Table S3. Yellow nodes represent the spheres created from the coordinates with a 6-mm radius, and blue lines denote the pairwise connections between the nodes. Connections within each network were used separately in our predictive modeling to investigate network-specific relationships with individual dimensions of psychopathology.

310
Biological Psychiatry February 1, 2021; 89:308-319 www.sobp.org/journal Biological Psychiatry 5003 repeated 10-fold cross-validation, individually for each combination of functional network and symptom dimension in the main sample. Using each network individually in the model was intended to ensure the functional specificity of our results. Importantly, RVM is a sparse learning method, that is, only a few of the learned feature weights are nonzero, lending interpretability as to which features (connections) are predictive. As recommended in (46), the symptom dimensional scores and rsFC features were adjusted for confounding effects of age, gender, site, and head motion. To avoid data leakage within crossvalidation (47), confound regression models were learned only on the training set and then applied to both training and test data (48,49). The RVM model trained on the (confoundadjusted) training data and was then applied to the (confound-adjusted) held-out data. Folds were stratified to accommodate different sample sizes across sites. Prediction performance was evaluated using Pearson's correlation between the (adjusted) scores and their predictions. Significant predictions were further validated for their generalizability across sites using leave-one-site-out crossvalidation following the same schematic but training on all sites but one and testing on the left-out site (Supplement). Statistical significance of the cross-validation-based correlations was determined through 1000 permutations by shuffling the symptom dimensional scores (lowest p = .001, right-tailed) (Supplement) (47,50).
Critically, we validated predictive associations confirmed by leave-one-site-out cross-validation in the independent B-SNIP sample. For this, we trained RVM models on the significantly predictive networks in the entire main sample and then, without further fitting or modification, applied them to held-out B-SNIP data. Robust associations passing this strict 3-step validation procedure were then further assessed as described below.
In addition, connections within the networks identified with robust associations with symptoms were combined to test the possible improvement in prediction performance. Control analyses were performed by repeating all validation procedures including illness duration or olanzapineequivalent dosage as confounds. For comparison, 1) we predicted the 3 PANSS subscales (positive, negative, and general psychopathology) using the same 3-step validation procedure and each network individually as in our main analyses, and 2) we predicted the 4 symptom dimensions using the whole connectome (39,060 connections) that comprised pairwise connectivity between all the nodes pooled from the 17 meta-analytic networks.

Identification of Reliably Predictive Connections and Subnetworks
The intrinsic feature selection in RVM through its sparse modeling was leveraged to identify reliably predictive connections and the potential subnetworks formed by them. A connection was identified as reliably predictive when it had nonzero weights in 1) at least 80% of the 10-fold crossvalidation repetitions, 2) at least 6 out of the 7 (i.e., .80%) leave-one-site-out analyses, and 3) the models trained on the entire main sample for validation in B-SNIP. The cutoff of 80% is suggested as a conservative threshold in both real and simulated data (51). To assess the predictive capacity of the subnetworks, RVM models were trained using the rsFC of the subnetworks on the main sample and tested in B-SNIP.

Spatial Correlation With Receptor/Transporter Densities
Finally, we evaluated the topographical relationship between network node location and the distribution of several receptor/ transporter systems, assessing for highly expressed receptors/ transporters in the identified networks relative to the entire brain. This was tested by comparing the average receptor/ transporter density across all nodes within a given network against a null distribution based on 1000 random network configurations generated by redistributing the nodes throughout the gray matter while preserving the between-node distances (66 mm tolerance). Seven dopamine and serotonin receptors (dopaminergic: D 1  . For comparability, these maps, in MNI152 space, were all resampled to an isotropic 2-mm spatial resolution as in our fMRI data and linearly rescaled to 0-100. Furthermore, the significantly higher expressed receptors/ transporters were entered into a spatial correlation analysis (52,53) calculated as rank correlation between the node importance scores and receptor/transporter densities calculated for these nodes. The node importance score was calculated by summing the selection frequency of the connections of each node derived from the repeated 10-fold cross-validation. Bootstrap analysis was conducted to ensure robustness. To establish the statistical significance of a spatial correlation against chance, a spatial permutation test was used where the null distribution was estimated based on the correlations between the node importance scores for a given network and the nodal receptor/ transporter densities extracted from 1000 simulated (random) networks (Supplement).

Network-Based Prediction of Specific Symptom Dimensions
Predictive modeling with stringent 3-step validation revealed several predictive network-psychopathology relationships. In the first step (10-fold cross-validation), the affective, negative, and positive symptom dimensions could be significantly predicted using rsFC within the mirror neuron system, theory of mind (ToM), and cognitive emotion regulation networks, respectively, while the cognitive dimension was predicted by 3 networks: ToM, empathy, and extended socioaffective default (eSAD) (Figure 2A, B). Except for the empathy-cognitive prediction, all these predictive network-symptom associations ( Figure 2B) were confirmed in the second validation step, that is, the leave-one-site-out analysis ( Figure 2C).
As both the 10-fold cross-validation and leave-one-site-out analyses were performed in the main sample, these analyses may still be overly optimistic regarding generalization to new patient populations. Therefore, we added a third validation step using a completely independent sample. Three of the 5 leave-one-site-out validated associations (ToM-negative, cognitive emotion regulation-positive, mirror neuron systemaffective) were not confirmed in this step. However, training RVM models on the entire main sample and testing it in the B-SNIP dataset revealed that the ToM and eSAD networks were significantly predictive of cognitive symptoms ( Figure 2D). Comparatively, combing the connections in these two networks yielded only a marginal improvement in the prediction of cognitive dimension in B-SNIP (r = .24, p = .018).
In complementary analyses, no significant effects of illness duration or olanzapine-equivalent dosage on the 4 symptomdimension scores were found (all p . .05 in the fitted general linear models). Conversely, controlling for illness duration or olanzapine-equivalent dosage did not alter the overall predictive patterns. Highlighting the utility of our 4-dimensional conceptualization of schizophrenia symptomatology (6), predicting the traditional 3 PANSS subscales, 10-fold cross-validation and leave-one-site-out experiments on the main sample revealed only 2 predictive patterns (ToM-negative, eSAD-general psychopathology), which did not, however, generalize to B-SNIP ( Figure S2). The whole-connectome comprising nodes from all networks was not predictive of any of the 4 symptom dimensions (all r , .11, p . .13; 5003 repeated 10-fold cross-validation), supporting our use of individual networks as an a priori biologically meaningful dimensionality reduction.

Reliably Predictive Connections and the Predictiveness of Subnetworks
Ten connections within eSAD and 8 connections within ToM were identified as consistently relevant for predicting cognitive symptoms ( Figure 3A; Table S6), that is, were selected in more than 80% of the different cross-validation runs (repeated 10fold cross-validation and leave-one-site out) and in the final models trained on the entire main sample for validation in B-SNIP. The ensuing ToM and eSAD subnetworks featured 3 spatially overlapping nodes located in the ventromedial prefrontal cortex, left middle temporal gyrus, and posterior cingulate cortex/precuneus and highlighted the ventromedial prefrontal cortex-posterior cingulate cortex/precuneus connection ( Figure 3B; Table S7). In turn, connections to subcortical regions including the bilateral amygdala/hippocampus and ventral striatum were specific to the eSAD subnetwork. A model based on the identified eSAD subnetwork showed almost identical prediction performance for the B-SNIP data compared with the one trained on the whole eSAD ( Figure 3C) (note that subnetwork definition was based only on the main sample, i.e., there is no leakage of information about the B-SNIP data). Interestingly, compared with the whole ToM network (Figure 2D), the ToM subnetwork demonstrated an improved performance ( Figure 3C). This confirms the power of sparse modeling in RVM to identify the truly relevant features.

Relationship to Molecular Architecture
Results showed that 8 receptors/transporters ( Figure 4A) are highly expressed in one or both networks (ToM, eSAD) relative   to the simulated networks with perturbed spatial configuration of nodes (histograms of node distance are shown in Figure S3). Spatial correlation between the node importance (Table S9) and the spatially corresponding density of those significant receptors/transporters (the procedure was detailed in Figure  4B) revealed a relationship to both the dopaminergic and the serotonergic systems ( Figure 4C). Specifically, the nodes of ToM that showed higher importance in predicting the cognitive dimension tracked with higher dopamine synthesis capacity (r = .54, p = .02). The prediction importance of the nodes within the eSAD network correlated positively with densities of D 1 (r = .66, p = .007) and 5-HTT (r = .53, p = .046), as well as dopamine synthesis capacity (r = .54, p = .036). The significance was corroborated by bootstrapped confidence intervals.

DISCUSSION
By using predictive modeling on multisite fMRI data with a strict out-of-sample validation procedure (repeated 10-fold crossvalidation, leave-one-site-out analysis, independent dataset), two a priori meta-analytically defined functional networks, ToM and eSAD, were identified as significantly and robustly predictive of the cognitive symptom dimension in patients with schizophrenia. In contrast, prediction using the original PANSS subscales failed to generalize to the independent sample, supporting the notion that these traditional dimensions do not correspond well with underlying neurobiology. Through the implicit feature selection of RVM, reliably predictive connections were identified that constituted subnetworks connecting nodes mainly distributed in the (medial) prefrontal cortex, posterior cingulate cortex, temporal regions, and subcortical structures. Moreover, higher densities of D 1 receptor and 5-HTT as well as elevated dopamine synthesis capacity were related to the node importance of the ToM and the eSAD networks in the prediction of cognitive symptomatology.

Symptom Dimensions Were Differentially Predicted by Different Functional Networks
In contrast to the prevalent notion of global or widespread brain deficits (10)(11)(12) in schizophrenia, we here revealed specific network-psychopathology relationships based on the predictive capacity of functional systems for patient-specific symptom severity along 4 dimensions. Two networks, ToM and eSAD, both predicting the cognitive dimension, passed our strict validation steps (see the Supplement and Table S8 for an additional explorative analysis correlating cognitive scores assessed using a scale other than the PANSS with the true and the predicted cognitive dimensional scores). As ToM subserves social cognition while the eSAD network relates to socioaffective processes, these results support previous findings indicating that compromised "social brain" development in schizophrenia relates to higher-level cognitive deficits (54,55). Somewhat surprisingly, these two networks did not allow a robust prediction of negative or affective symptoms, even though other socioaffective networks [cognitive emotion regulation: reappraisal process of emotional stimuli (56); mirror neuron system: emotional aspect of social cognition (18)] predicted positive and affective dimensions, respectively, in the main sample. Thus, it stands to reason that different socioaffective networks capture individual variance of different symptom dimensions. Of note, these predictions did not generalize to B-SNIP, potentially because of between-sample differences in clinical characteristics, highlighting that building models that generalize to novel cohorts remains a challenge (57). We also noted that the cognitive dimension was not robustly predicted by any of the assessed cognitive networks, mirroring previous work in healthy subjects showing that taskbased functional connectivity yields better predictions of cognition than networks at rest (58)(59)(60).
In turn, intrinsic connectivity patterns within the ToM network are robustly predictive of the cognitive status in individual patients with schizophrenia. ToM is the cognitive ability of an individual to infer others' mental states, intentions, and beliefs (61). As these involve complex cognitive processes and considering that the cognitive dimension of schizophrenia psychopathology encompasses symptoms such as "conceptual disorganization," "lack of insight," and "disturbed abstract thinking," this predictive relationship is not unexpected. Conversely, ToM deficits, which are prevalent in schizophrenia and are a well-established feature and vulnerability marker of this disorder (62), are known to be associated with symptoms of disorganization (63). Abnormal neural activation in response to tasks targeting ToM also has been reported in schizophrenia, involving the temporoparietal junction as well as prefrontal and temporal regions where the current ToM metanetwork is distributed (64,65).
It is interesting to note that the ToM network also predicted negative symptoms in the main sample, although this did not generalize to the B-SNIP data. Given that the subnetworks of ToM predicting negative and cognitive dimensions, respectively, were largely divergent ( Figure 3B; Figure S4), this resonates with the notion that ToM encompasses multiple components (66).
Intriguingly, cognitive symptoms in schizophrenia were also linked to socioaffective processes via the prediction using the eSAD network. This corroborates recent studies consistently linking neural activity patterns in the default mode network (DMN) to cognitive abilities (67,68), and the implication of the DMN-derived eSAD network (69) in cognitive processing is therefore unsurprising. Our finding moreover resonates well with literature suggesting that socioaffective factors affect and modulate cognitive performance such as working memory and attention in patients with schizophrenia (70)(71)(72). The use of seemingly noncognitive psychosocial methods has been proposed as a potential remediation strategy for cognitive deficits in schizophrenia (70), and indeed has been successfully applied in practice (72). Consequently, we would hypothesize that the cognitive dysfunctions in schizophrenia might relate to an impaired integration of self-and other-related processes. Since the interaction of the (core) DMN with other brain regions and networks is more reflective of schizophrenia symptoms than within-DMN connectivity (18,73,74), it is not surprising that the DMN was not predictive but that the eSAD, which comprises regions going beyond the DMN, was. including mainstay antidopaminergic agents (37,75). In line with the notion that dopaminergic dysfunction alone does not account for the whole picture of schizophrenia psychopathology (76,77), here we revealed that the networks and nodes predictive of individual cognitive symptom load are related to both dopaminergic (32) and serotonergic (76) systems. These data extend previous region-of-interest analyses (78,79) to a comprehensive topographical level and are supported by findings that cognitive deficits relate to D 1 receptor elevation in schizophrenia (78). The present extension to a network-level investigation is important, as patients with schizophrenia are characterized by dysconnection between distant brain regions, and network-based analyses can nominate mechanisms of action for the development of novel pharmacological treatments (80). In line with involvement of dopamine functioning in cognitive deficits in schizophrenia (32), the cognitive dimension of psychopathology was associated with dopamine synthesis capacity. Increased dopamine synthesis capacity in schizophrenia was reported in not only the striatum (32) but also cortical areas, including the prefrontal cortex (81,82), where multiple nodes within the ToM and eSAD networks are located. Moreover, cognitive symptoms were linked to the density of 5-HTT, which plays a critical role in regulating serotonergic concentration and signaling. Serotonergic dysfunction and 5-HTT polymorphism (83) have been involved in schizophrenia pathophysiology (34,76). Previous postmortem and in vivo positron emission tomography studies have yielded inconsistent results on the alteration of 5-HTT density in schizophrenia (34), while an overexpression of 5-HTT messenger RNA in the prefrontal and temporal cortex has been demonstrated (84). We here revealed a potential role of 5-HTT in cognitive deficits via the networks involved in socioaffective processes, resonating with the proposed implication of 5-HTT in the affective domain (85,86). While atypical antipsychotics, including olanzapine and risperidone, also target the 5-HT 2A receptor (87,88), their effects on 5-HTT seem to be equivocal (89,90). Interestingly, antipsychotic drugs pimavanserin and SEP-363856 have just been introduced that preferentially target serotonin and not dopamine receptors (91,92), suggesting increased focus also on serotonergic pathways.

Limitations and Considerations
First, the effect sizes for the correlation between the symptom scores and their predictions were moderate. However, despite the clinical complexity of the population and the heterogeneities in scanners and protocols, the effect sizes are similar to those previously reported for predicting, e.g., creativity (93), personality (46), and memory performance (94) in healthy subjects (r values mostly w.2-.35). Since olanzapineequivalent dosage did not correlate with symptom scores or alter the prediction pattern after additionally controlling for the dosage in cross-validation, medication would be largely a source of random variation in our data and thus make our results more conservative. Second, we acknowledge that taskspecific changes in network topology and related dynamics may also explain the variance in individual symptoms (95,96) and should be the scope of future studies to complement our resting-state findings. Third, although meta-analytically derived peak coordinates are the most likely location of specific functional networks, there may exist individual-specific topological differences that in turn may reflect pathophysiological mechanisms underpinning individual symptoms. Future studies assessing the contribution of individual-specific topology to the prediction of symptoms are desired. Fourth, although glutamatergic dysfunction has been increasingly implicated in schizophrenia neurocognitive deficits (77,97), there are no publicly available in vivo maps reflecting the glutamatergic system. Fifth, within-subject (longitudinal) studies assessing symptoms, rsFC, and receptor densities are needed. Finally, we acknowledge that the sample sizes are modest for some of the receptor/transporter maps that we used from prior molecular imaging studies, as acquisition of molecular imaging data in large samples remains difficult.
Based on rsFC within different meta-analytic task-activation networks covering a broad range of functional domains and predictive modeling with strict validations, intrinsic connectivity patterns of networks implicated in socioaffective processing was revealed to be robustly associated with the cognitive dimension of psychopathology. Our investigation of the molecular architecture of the identified predictive networks implied a potential involvement of 5-HTT, besides the dopaminergic system, in schizophrenia cognitive symptomatology, possibly providing hints into treatments. This article was published as a preprint on bioRxiv: https://doi.org/10.11 01/2020.07.02.185124.

ACKNOWLEDGMENTS AND DISCLOSURES
Maps of the meta-analytic networks and the code for predictive modeling can be found at https://github.com/inm7/Network-basedprediction. Density maps of neurotransmitter receptors and transporters and the MATLAB-based toolbox for the integration of positron emission tomography-and single photon emission computed tomography-derived modalities with other brain imaging data are available at https://github. com/juryxy/JuSpace.