Introduction
The dysconnectivity hypothesis has been proposed as one of the core disease mechanisms of schizophrenia (SCZ), evidenced by widespread alterations in both functional and structural connectomes of SCZ patients compared to healthy controls (HCs) (Di Biase et al., Reference Di Biase, Cropley, Baune, Olver, Amminger, Phassouliotis, Bousman, McGorry, Everall, Pantelis and Zalesky2017; Hou et al., Reference Hou, Yan, Jiang, Deng, Zhang, Li, Duan, Wang, Yao, He, Rodríguez-Labrada, Yao and Luo2026; van den Heuvel & Fornito, Reference van den Heuvel and Fornito2014). These dysconnectivity patterns emerge across all disease stages (Cetin-Karayumak et al., Reference Cetin-Karayumak, Di Biase, Chunga, Reid, Somes, Lyall, Kelly, Solgun, Pasternak, Vangel, Pearlson, Tamminga, Sweeney, Clementz, Schretlen, Viher, Stegmayer, Walther, Lee and Kubicki2020), serving as early indicators of SCZ pathophysiology (Cui et al., Reference Cui, Wei, Xi, Griffa, de Lange, Kahn, Yin and van den Heuvel2019) and contributing to the development of cognitive and negative symptoms (McCutcheon, Reis Marques, & Howes, Reference McCutcheon, Reis Marques and Howes2020). Recent studies have revealed substantial inter-individual variability in SCZ dysconnectivity patterns, with patients exhibiting greater functional connectome variability in the frontoparietal and default mode networks (Santo-Angles et al., Reference Santo-Angles, Salvador, Gomar, Guerrero-Pedraza, Ramiro, Tristany, Teixidó, Ortiz-Gil, Aguirre, Bosque, López-Araquistain, Maristany, Salgado-Pineda, Sarró, McKenna, Bernardo, Pomarol-Clotet and Schwarzbach2021) and disrupted variability of sensorimotor, visual, and auditory regions (Sun et al., Reference Sun, Liu, Ma, Duan, Wang, Xu, Xu, Xu, Wang, Tang, He and Xia2021). These connectomic findings align with the known remarkable variations in genetic background (Owen et al., Reference Owen, Legge, Rees, Walters and O’Donovan2023; Wei et al., Reference Wei, de Lange, Savage, Tissink, Qi, Repple, Gruber, Kircher, Dannlowski, Posthuma and van den Heuvel2023) and clinical manifestations (Lee et al., Reference Lee, Cernvall, Borg, Plavén-Sigray, Larsson, Erhardt, Sellgren, Fatouros-Bergman and Cervenka2024; Marder & Umbricht, Reference Marder and Umbricht2023; Siskind et al., Reference Siskind, Orr, Sinha, Yu, Brijball, Warren, MacCabe, Smart and Kisely2022). Moreover, the increased inter-individual variability in SCZ was suggested to reflect a higher level of heterogeneity (Alnæs et al., Reference Alnæs, Kaufmann, van der Meer, Cordova-Palomera, Rokicki, Moberget, Bettella, Agartz, Barch, Bertolino, Brandt, Cervenka, Djurovic, Doan, Eisenacher, Fatouros-Bergman, Flyckt, Di Giorgio, Haatveit and Westlye2019), as distinct pathophysiological processes underlying the same diagnosis (Owen, Sawa, & Mortensen, Reference Owen, Sawa and Mortensen2016) would directly increase the within-group variance among SCZ patients. However, the inter-individual variability of the structural connectome in SCZ remains unclear, and delineating such structural variability is crucial for characterizing connectivity-based disease heterogeneity and elucidating SCZ’s complex etiology (Owen et al., Reference Owen, Sawa and Mortensen2016; Voineskos, Jacobs, & Ameis, Reference Voineskos, Jacobs and Ameis2020).
In light of the marked heterogeneity in SCZ, a key hypothesis proposes that multiple, partly independent disease processes ultimately converge to produce similar clinical syndromes (Owen et al., Reference Owen, Sawa and Mortensen2016). This convergence model suggests the potential for identifying disease-related dimensional patterns or subtypes that may reflect divergent biological mechanisms in SCZ. Recent studies on neuroanatomical characteristics have begun to identify such subtypes, revealing two SCZ subtypes with heterogeneous patterns of gray matter atrophy that originate separately in Broca’s area and the hippocampus (Jiang et al., Reference Jiang, Wang, Zhou, Palaniyappan, Luo, Ji, Yang, Wang, Zhang, Huang, Tsai, Chang, Xie, Zhang, Lv, Chen, Shen, Wu, Zhang and Feng2023). Additionally, researchers have characterized two SCZ subtypes exhibiting distinct trajectories of gray matter changes: one marked by cortical-predominant loss with enlarged striatum, and another by subcortical-predominant atrophy (Jiang et al., Reference Jiang, Luo, Wang, Palaniyappan, Chang, Xiang, Zhang, Duan, Huang, Gaser, Nemoto, Miura, Hashimoto, Westlye, Richard, Fernandez-Cabello, Parker, Andreassen, Kircher and Feng2024). Moreover, four distinct patterns of regional FA reduction have been identified, each associated with specific clinical features, including bizarre behavior, prominent delusions, and disorganized speech (Arnedo et al., Reference Arnedo, Mamah, Baranger, Harms, Barch, Svrakic, de Erausquin, Cloninger and Zwir2015). These subtyping studies, employing data-driven strategies and machine learning techniques, have provided important insights into the neurobiological heterogeneity underlying SCZ (Voineskos et al., Reference Voineskos, Jacobs and Ameis2020). However, it remains unknown to what extent structural dysconnectivity patterns vary across SCZ patients and whether connectivity-derived subtypes are associated with distinct biological and clinical profiles.
Here, we first identified structural circuitry that showed altered inter-individual variability in SCZ, using diffusion-weighted imaging (DWI) data from 103 SCZ patients and 97 HCs. Subsequently, the involved brain circuits were used for subtyping analysis using nonnegative matrix factorization (NMF). The resulting connectivity-derived SCZ subtypes were further compared in terms of biological and clinical features, including symptom severity, peripheral inflammatory markers, short-term treatment response, and neuroanatomical endophenotypes.
Methods
Participants
MRI data and clinical assessments of a total of 105 patients with SCZ and 97 HCs were included in the discovery dataset. Patients were recruited from the Department of Psychiatry, Xijing Hospital, and HCs were enrolled from the local community through advertisements. This study was approved by the Institutional Ethics Committee of the Xijing Hospital, The Fourth Military Medical University (number XJYYLL-2015047 and KY20183113-1). All participants provided written informed consent, followed by a full introduction of the aim and design of the study.
SCZ diagnoses were established based on the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5) (Stein et al., Reference Stein, Phillips, Bolton, Fulford, Sadler and Kendler2010) and consensus by two senior psychiatrists. Positive symptoms, negative symptoms, general psychopathological symptoms, supplementary scores, and total scores were derived from the Positive and Negative Syndrome Scale (PANSS) (Kay, Fiszbein, & Opler, Reference Kay, Fiszbein and Opler1987). Baseline intravenous blood samples were collected from the upper forearm of the 33 SCZ patients into an EDTA vacuum tube under fasting conditions to obtain inflammatory measurements, such as leukocyte count and percentage (monocytes, lymphocytes, neutrophils, eosinophils, and basophils, through a Sysmex XS Automated Hematology Analyzer). Baseline clinical assessments, blood samples, and brain scans were collected on the same day. After quality control of MRI data, two patients were additionally excluded due to large head movements, with the remaining 103 patients with SCZ and 97 HCs used for the following analyses. Short-term clinical assessment data at discharge, including PANSS assessments, were collected from 69 patients with SCZ (mean duration: 15.2 ± 8.7 days). Long-term MRI follow-up data were collected from 46 participants, including 31 patients with SCZ and 15 HCs (mean duration: 146.7 ± 49.5 days). Detailed information regarding longitudinal data collection is provided in the Supplementary Methods and Supplementary Figure S1. Demographic and clinical details of participants in the discovery dataset are provided in Supplementary Table S1.
Additionally, an independent dataset of 43 SCZ patients was used as the validation dataset. MRI data were collected at the Second Affiliated Hospital of Xi’an Jiaotong University from November 2024 to June 2025. Details of the validation dataset are provided in the Supplementary Methods and Supplementary Table S8.
Image acquisition and preprocessing
High-resolution T1-weighted MRI and DWI data were collected using a GE Discovery MR750 3.0 T scanner at the Department of Radiology, Xijing Hospital. Detailed scanning parameters are described in Supplementary Table S2. T1-weighted MRI data were used for brain tissue segmentation, using the FreeSurfer software (version 6.0.0) (Fischl et al., Reference Fischl, van der Kouwe, Destrieux, Halgren, Segonne, Salat, Busa, Seidman, Goldstein, Kennedy, Caviness, Makris, Rosen and Dale2004). The reconstructed cortical mantle was parcellated into 219 distinct cortical regions, including 111 regions in the left hemisphere and 108 regions in the right hemisphere, according to a subdivision of the Desikan–Killiany atlas (Cammoun et al., Reference Cammoun, Gigandet, Meskaldji, Thiran, Sporns, Do, Maeder, Meuli and Hagmann2012; Desikan et al., Reference Desikan, Ségonne, Fischl, Quinn, Dickerson, Blacker, Buckner, Dale, Maguire, Hyman, Albert and Killiany2006). The original DK atlas (68 regions) and the 114-region DK subdivisions were used for validation purposes in order to examine the influence of parcellation and network size (Zalesky, Fornito, Harding, et al., Reference Zalesky, Fornito, Harding, Cocchi, Yucel, Pantelis and Bullmore2010). Cortical thickness of each brain region was automatically derived for each subject through the FreeSurfer pipeline.
DWI data preprocessing was performed using CATO (v.3.1.2) (de Lange, Helwegen, & van den Heuvel, Reference de Lange, Helwegen and van den Heuvel2023) and involved three main steps: (1) realigning the 64 diffusion-weighted images (b = 1,000 s/mm2) with the reference volume (b = 0) and correcting for minor head movements and gradient-induced distortions using the FSL eddy preprocessing pipeline (Jenkinson et al., Reference Jenkinson, Beckmann, Behrens, Woolrich and Smith2012); (2) reconstructing voxel-wise diffusion orientation distribution functions using constrained spherical deconvolution (Morez, Sijbers, Vanhevel, & Jeurissen, Reference Morez, Sijbers, Vanhevel and Jeurissen2021) within the brain mask; and (3) reconstructing white matter pathways via deterministic streamline tractography with fiber assignment by continuous tracking (Mori, Crain, Chacko, & van Zijl, Reference Mori, Crain, Chacko and van Zijl1999). For each brain voxel, eight streamline seeds were generated. Fiber tracking was stopped when a streamline entered a voxel with low diffusion anisotropy (FA < 0.1), exited the defined brain region, and/or exhibited an abrupt angular deviation exceeding 45°.
Network reconstruction
Structural connectivity matrices were reconstructed based on preprocessed DWI data using CATO. Each region of the 219-region subdivision of the DK atlas was defined as a network node. A network edge was assigned if two regions were linked by reconstructed streamlines. Streamline volume density (SVD), defined as the number of streamlines (NOS) between brain regions divided by the average cortical volume, was taken as the weight of the edge (Cui et al., Reference Cui, Wei, Xi, Griffa, de Lange, Kahn, Yin and van den Heuvel2019; van den Heuvel & Sporns, Reference van den Heuvel and Sporns2011).
Statistical analysis
Double generalized linear models (DGLM) were employed to simultaneously estimate the mean and dispersion of each brain connectivity between SCZ and HC. DGLM first fits a generalized linear model (GLM) to estimate the mean parameter (mean model), with the dispersion parameter modeled using the deviances (residuals) from the mean model as responses for the second GLM (dispersion model) (Alnæs et al., Reference Alnæs, Kaufmann, van der Meer, Cordova-Palomera, Rokicki, Moberget, Bettella, Agartz, Barch, Bertolino, Brandt, Cervenka, Djurovic, Doan, Eisenacher, Fatouros-Bergman, Flyckt, Di Giorgio, Haatveit and Westlye2019; Haatveit et al., Reference Haatveit, Westlye, Vaskinn, Flaaten, Mohn, Bjella, Sæther, Sundet, Melle, Andreassen, Alnæs and Ueland2023). The process iterates between the mean and dispersion sub-models until overall convergence is achieved. Age and gender were included as covariates. DGLM analyses were performed with R (version 4.3.3) and the R package ‘dglm’. Each brain connectivity was assigned two t-scores derived separately from the mean model and dispersion model of the DGLM, representing differences in the mean and variance between patients and controls. Subsequently, network-based statistical analysis (NBS) (Zalesky, Fornito, & Bullmore, Reference Zalesky, Fornito and Bullmore2010) was applied to the DGLM-derived edge-wise t-scores to identify subnetworks showing group differences between patients with SCZ and HCs. The initial threshold value of t = 1.972, corresponding to the two-sided critical t-value at P = 0.05 with df = 198, was used as the primary edge-wise threshold in NBS to define suprathreshold edges. Connected components of positive effects (t > 1.972) and negative effects (t < −1.972) were examined separately. In NBS, the component size was defined using component extent, calculated as the number of suprathreshold edges within each connected component. A p-value was computed based on the proportion of the null distribution, generated from 5,000 permutations, in which the maximum component extent was larger than or equal to the observed component extent.
Having identified brain circuits showing significant mean and dispersion changes in SCZ, we also performed enrichment analysis to examine the associated functional networks of these identified circuits. Using hypergeometric tests, we examined whether the number of overlaps between the identified circuits and those involved in the seven functional networks (Yeo et al., Reference Yeo, Krienen, Sepulcre, Sabuncu, Lashkari, Hollinshead, Roffman, Smoller, Zollei, Polimeni, Fischl, Liu and Buckner2011), namely, the visual (VN), somatomotor (SMN), dorsal-attention (DAN), frontal–parietal (FPN), limbic (LN), ventral-attention (VAN), and default-mode networks (DMN), exceeded what could be expected by chance. We also performed enrichment analysis in terms of interregional distance by categorizing edges into short (the top 25% shortest connections), medium (the middle part of the 25th–75th percentiles of connections), and long (the top 25% longest connections) (Beul, Grant, & Hilgetag, Reference Beul, Grant and Hilgetag2015; Wei, Scholtens, Turk, & van den Heuvel, Reference Wei, Scholtens, Turk and van den Heuvel2019).
For analyses involving multiple comparisons, false discovery rate (FDR) correction was applied to control for multiple testing. FDR correction was performed within each category of related tests, and FDR-adjusted P-values are reported accordingly. Statistical significance was set at FDR-adjusted P < 0.05.
Subtyping analysis
Based on the brain connections identified from both the mean and dispersion models in the DGLM analysis, we used NMF to derive SCZ subtypes. NMF is a data-driven dual clustering method that identifies both feature clusters and individual clusters by decomposing the feature matrix into a basis matrix and a coefficient matrix (Brunet, Tamayo, Golub, & Mesirov, Reference Brunet, Tamayo, Golub and Mesirov2004). Cluster numbers from 2 to 10 were evaluated, with 200 runs performed for each cluster number. The optimal number of clusters of patients with SCZ was determined based on the cophenetic correlation coefficient and silhouette coefficient, as described in previous studies (Brunet et al., Reference Brunet, Tamayo, Golub and Mesirov2004; Luo et al., Reference Luo, Pi, Hu, Wang, Xiao, Li, Sweeney, Bishop, Gong, Xie and Lui2021) (see Supplementary Methods for details). Connections identified through other NBS thresholds were further used to assess the sensitivity of subtype assignments to the NBS threshold (Supplementary Methods). To illustrate the representative structural connectivity patterns of each cluster at varying levels, the top 10%, 20%, 30%, 40%, and 50% weighted structural connections in each column of the basis matrix were then characterized separately. Patients with SCZ were grouped into subtypes according to the index of the maximum weights in each column of the coefficient matrix. To further compare the subtype approach with dimensional approaches, NMF component loadings, as well as the difference between component loadings, were used and correlated to PANSS scores. NMF analyses were performed with R (version 4.3.3) and the R package ‘NMF’ (Gaujoux & Seoighe, Reference Gaujoux and Seoighe2010).
We also performed several sensitivity and reproducibility analyses, including analyses on sensitivity to NMF initialization, resampling, alternative clustering methods, diffusion reconstruction approaches, and parcellation schemes (see Supplementary Methods for details). To validate the corresponding subtypes in the independent dataset, we used the fitted NMF model from the discovery dataset to factorize the structural connectivity data of the independent dataset. Similarly, subtype assignment was determined by the index of the maximum estimated weight in each column of the coefficient matrix (see Supplementary Methods).
Differences in demographic data, including age, gender, years of education, and baseline medication status, were assessed between SCZ subtypes using R and are summarized in Supplementary Table S3. Clinical symptoms (i.e. PANSS subscales and total scores) and peripheral inflammation markers were also compared between SCZ subtypes using linear regression models, with age, gender, and baseline medication status included as covariates. Correlation analysis was performed within each subtype to examine the subtype-specific associations between peripheral inflammation markers and PANSS scores (see details in Supplementary Methods). We additionally examined whether the identified SCZ subtypes showed different patterns of cortical thickness abnormalities, using linear regression models with age, sex, and global measurements as covariates (Cui et al., Reference Cui, Wang, Fu, Liu, Wei, Zhao, Gu, Fan, Wu, Gong, Lin, Yin, Guan and Chang2023). The distinct patterns of cortical thickness for each subtype were further linked to cross-disorder cortical thickness patterns (Larivière et al., Reference Larivière, Paquola, Park, Royer, Wang, Benkarim, Vos de Wael, Valk, Thomopoulos, Kirschner, Lewis, Evans, Sisodiya, McDonald, Thompson and Bernhardt2021), neurotransmitter receptors/transporters density maps (Markello et al., Reference Markello, Hansen, Liu, Bazinet, Shafiei, Suárez, Blostein, Seidlitz, Baillet, Satterthwaite, Chakravarty, Raznahan and Misic2022), and network-based atrophy models (disease epicenters) (Larivière et al., Reference Larivière, Rodríguez-Cruces, Royer, Caligiuri, Gambardella, Concha, Keller, Cendes, Yasuda, Bonilha, Gleichgerrcht, Focke, Domin, von Podewills, Langner, Rummel, Wiest, Martin, Kotikalapudi and Bernhardt2020) to identify subtype-specific biological endophenotypes (see Supplementary Methods for details).
Longitudinal analysis
Given the limited amount of short-term assessment data (N = 69, collected after approximately 2–3 weeks of treatment), we applied a linear mixed-effects model to assess changes in PANSS scores in patients of different subtypes (interaction effects: subtype × time), to explore whether there were differences in treatment response between subtypes. We then assessed the medications used for short-term treatments during hospitalization. Patients receiving risperidone, paliperidone, or olanzapine monotherapy were categorized into the corresponding medication groups, whereas patients receiving other medications or combination treatments were categorized as ‘others’. Doses of current antipsychotic drugs were converted using the defined daily dose (DDD) method (Leucht, Samara, Heres, & Davis, Reference Leucht, Samara, Heres and Davis2016). More detailed information is provided in the Supplementary Material and Supplementary Table S4. We employed a linear mixed-effects model to compare changes in PANSS scores between the identified subtypes within each monotherapy group (interaction effects: subgroup × time), allowing for the detection of subtype-specific medication sensitivities. A subject ID (SubID) for each patient was taken as the random factor, and age, sex, baseline medication status, and DDD were included as covariates to control for potential confounding factors. For the long-term MRI follow-up data, longitudinal changes in regional cortical thickness between patients (N = 31) in different subtypes and HCs (N = 15) during long-term follow-up were also compared using a linear mixed-effects model (see details in the Supplementary Methods).
Results
Brain circuits of VAN, FPN, and DMN showed large individual differences in SCZ
As revealed by the dispersion model of DGLM, one subnetwork of 124 connections (i.e. dispersion-model subnetwork) was observed, showing significantly larger inter-individual variability among patients with SCZ compared to HCs (P NBS < 0.001; Figure 1a). This subnetwork covered widespread brain regions, including the bilateral caudal anterior cingulate gyri, superior and middle temporal regions, orbital, superior, medial, and inferior frontal regions, insula, and so forth (Figure 1a). Validation analyses using different parcellations (DK-68 and DK-114) showed similar patterns (Supplementary Figure S2). Enrichment analysis further showed the identified connections to be enriched in the set of within-network connections of the FPN and sets of between-network connections between the VAN and DMN, the FPN and DMN, and the LN and SMN (P FDR = 0.045, <0.001, 0.002, and 0.008, respectively; corrected across 28 within- and between-networks; hypergeometric tests; Figure 1b). These results suggested that structural connectivity across higher-order cognitive networks, namely, VAN, FPN, and DMN, exhibited larger concentrated inter-individual variability in patients with SCZ. In line with connection-wise analysis, we also performed region-wise enrichment analysis and showed that brain regions involved in the dispersion-model subnetwork were predominantly enriched in the VAN and DMN (both P FDR = 0.037; corrected across seven functional networks; hypergeometric tests; Figure 1c). Also, regions involved in VAN, FPN, and DMN in general showed significantly higher levels of dispersion effects than the VN and the SMN (Wilcoxon rank-sum test, Z = 1.983, P = 0.047; Figure 1d), suggesting that increased inter-individual variability in SCZ was observed in those higher-order cognitive networks. In addition, functional annotation analysis using neuromaps showed associations between brain variability patterns and spatial patterns of 5-HT1a, 5-HT6, D2, MOR, α4β2, CB1, H3, and KOR (Supplementary Table S6).
The dispersion model of structural connectivity in schizophrenia. (a) Connections with significantly higher heterogeneity of SVD values in patients with SCZ compared to healthy controls. Connected components of positive effects (t > 1.972) and negative effects (t < −1.972) were examined separately. The summed t-scores for regions linked by connections associated with significant effects in the DGLM were plotted to represent the contribution of each brain region within the dispersion model. (b) Enrichment of connections in the dispersion model for seven functional networks at the connectivity level. The connection color represents enrichment, with red indicating significant enrichment and gray indicating nonsignificance. (c) Enrichment of brain regions in the dispersion model for seven functional networks at the brain regional level. Asterisks indicate P FDR < 0.05. (d) Comparison of t-scores in brain regions of higher-order cognitive networks (DMN, FPN, and VAN) versus primary cognitive networks (SMN and VN). Asterisk indicates P < 0.05. (e) Enrichment of connections in the dispersion model across three distance classes: short-, medium-, and long-distance connections. Asterisk indicates P < 0.05. SVD, streamline volume density; SCZ, schizophrenia; DGLM, double generalized linear models; VN, visual network; SMN, somatomotor network; DAN, dorsal-attention network; FPN, frontal–parietal network; LN, limbic network; VAN, ventral-attention network; DMN, default-mode network; FDR, false discovery rate.
Figure 1. Long description
Panel a shows a glass brain with a network of nodes and red edges, alongside four cortical surface maps colored by t value from 0 to 16.35, indicating regional contributions to the dispersion model. Panel b is a circular chord diagram showing connectivity enrichment within and between seven functional networks: V N, S M N, D A N, V A N, L N, F P N, and D M N. Thick orange bands indicate significant enrichment (P F D R less than 0.05) within F P N and between V A N and D M N, F P N and D M N, and L N and S M N. Panel c is a stacked bar chart of brain region numbers across the same seven networks. D M N and V A N show significant enrichment marked with asterisks. Panel d is a vertical dot plot comparing t values between Primary and Higher-order cognitive networks, with Higher-order showing a significantly higher distribution marked by an asterisk. Panel e is a bar chart showing the negative log 10 P-value for connection distances. The medium-distance category significantly exceeds the significance threshold (dashed line) and is marked with an asterisk, while short and long distances do not.
In parallel with connectivity showing larger individual variability in SCZ, we also identified 61 connections, namely the mean-model subnetwork, showing significant mean differences between patients and controls according to the mean-model of DGLM (P NBS < 0.001 for decreased mean weights; P NBS = 0.019 for increased mean weights; detailed in the Supplementary Results and Supplementary Figure S3A). Connections observed in the DGLM mean-model were enriched in within-network connectivity of the VAN and LN, as well as in between-network connectivity between the VAN and DMN, and between the SMN and DMN (P FDR = 0.007, 0.038, <0.001, and 0.002, respectively; corrected across 28 within- and between-networks; hypergeometric tests; Supplementary Figure S3B). This result suggests the key role of connections between VAN and DMN, where SCZ patients exhibited convergent disruptions involving both dispersion and mean models. No significant regional enrichment across the seven functional networks was observed for the mean-model subnetwork (Supplementary Figure S3C and S3D).
Given that several regions of DMN and VAN, such as the insula, superior frontal, and superior temporal, were reported as hub regions of the connectome (Supplementary Figure S4), we additionally performed enrichment analysis in the context of hub regions. We found that both the dispersion-model and mean-model subnetworks showed enrichment for edges of hub regions (Supplementary Figure S3E). Additionally, in the context of connectivity distance, the dispersion-model subnetwork showed a significant association with medium-distance connections (P FDR = 0.012, corrected across six distance-related analyses; Figure 1e). No significant enrichment in connection-level distance classes was observed for the mean-model subnetwork (Supplementary Figure S3F).
Two SCZ subtypes were identified based on connectivity
Based on the brain circuits revealed by DGLM, SCZ patients were grouped into two subtypes with distinct dysconnectivity patterns (Supplementary Figure S5A), with 57 patients (55.3%) classified as subtype-I and 46 patients (44.7%) as subtype-II. Sensitivity analysis using other NBS thresholds showed similar subtype assignments among SCZ patients (Supplementary Results). The two subtypes showed no difference regarding age, gender, education, and baseline medication status (Supplementary Table S3). The top 20% connections contributing to each subtype are shown in Figure 2a, with results from other thresholds presented in Supplementary Figure S5B. Brain circuits driving subtype-I were specifically enriched in the between-network connectivity between the DAN and LN (P FDR = 0.031, top 20% weighted connections; corrected across 28 within- and between-networks; hypergeometric tests; Figure 2b), whereas those driving subtype-II were specifically enriched in between-network connectivity between the VAN and DMN (P FDR = 0.026, top 20% weighted connections; corrected across 28 within- and between-networks; hypergeometric tests; Figure 2b). Results from additional thresholds of top connections are provided in Supplementary Figure S5C.
Structural connectivity phenotypes and clinical characteristics differ between schizophrenia subtypes. (a) Representative structural networks (top 20% weighted connections) and composite brain regions for two SCZ subtypes. (b) Enrichment analysis of the top 20% representative connections for two SCZ subtypes across the seven functional networks. Connection colors indicate enrichment, with colored connections representing significant enrichment and gray representing nonsignificant results after correction for multiple comparisons. (c) Different levels of peripheral inflammation markers, including neutrophil percentage, lymphocyte percentage, NLR, and SIRI, between the two subtypes. (d) Comparisons of PANSS positive symptoms between the two subtypes at baseline. Asterisk indicates P < 0.05. (e) Changes in PANSS positive symptoms between the two subtypes (subtype by time interaction effect) at baseline and during short-term follow-up. (f) Changes in PANSS total scores between subtypes during short-term follow-up in the risperidone and olanzapine groups (subtype × time interaction effect). Asterisk indicates P < 0.05. SCZ, schizophrenia; NLR, neutrophil-to-lymphocyte ratio; SIRI, systemic inflammation response index; PANSS, Positive and Negative Syndrome Scale.
Figure 2. Long description
Panel a displays brain network maps for Subtype-I and Subtype-II. Subtype-I shows denser mean connectivity in blue and dispersion connectivity in orange. Lateral and medial brain views show regional enrichment on a color scale from 0 to 0.075. Panel b uses two chord diagrams to show functional network enrichment. Subtype-I has significant blue connections between S M N and L N, S M N and D M N, and D A N and L N, and within V A N. Subtype-II shows significant enrichment in connections between V A N and D M N, S M N and L N, and D M N and S M N, as well as within V A N. Panel c contains four box plots comparing Neutrophil percentage, Lymphocyte percentage, N L R, and S I R I. Subtype-I has a significantly higher Neutrophil percentage and trend-level higher N L R, and S I R I levels, while Subtype-II has a trend-level higher Lymphocyte percentage. Panel d is a raincloud plot showing P A N S S Positive scores at baseline, with Subtype-II having a slightly higher median than Subtype-I. Panel e is a line graph showing changes in P A N S S Positive scores from baseline to follow-up, with Subtype-II showing trend-level greater improvement than Subtype-I. Panel f shows P A N S S Total score changes for Risperidone and Olanzapine groups. In the Olanzapine group, Subtype-II shows trend-level greater symptom improvement than Subtype-I.
Random seed analysis showed that the two-subtype solution was consistently identified as optimal across 500 random seed initializations. Patient-level subsampling analysis also showed stable subtype assignments and consistent connectivity patterns between the subsampling and full-sample analyses. Validation using alternative clustering methods, including NMF with the Brunet algorithm and k-means clustering, further supported the division of the two subtypes (Supplementary Results and Supplementary Figures S6–S9).
We next examined whether the two subtypes differed in clinical symptoms. Patients in subtype-II showed trend-level higher PANSS positive scores than those in subtype-I (t = 2.272, P = 0.025, not corrected; Figure 2d and Supplementary Table S3). No apparent difference was observed between the two subtypes across negative symptoms, general psychopathological symptoms, supplementary scores, or total scores (P ≥ 0.05; Supplementary Table S3). Consistent with the discovery dataset, the assigned subtypes in the validation dataset showed a difference in PANSS positive scores (t = 2.921, P = 0.006), with subtype-II showing higher positive symptom severity than subtype-I (Supplementary Table S8).
In addition to subtyping, we further examined dimensional characteristics of the NMF-derived structural connectivity patterns. We assessed associations between NMF component loadings and clinical symptoms, showing that component 1 was negatively associated with PANSS positive scores (r = −0.210, P = 0.038; Supplementary Figure S10A), whereas component 2 showed a positive but nonsignificant association (r = 0.184, P = 0.070; Supplementary Figure S10B). Moreover, the difference between the two component loadings was associated with PANSS positive scores (r = −0.202, P = 0.047; Supplementary Results; Supplementary Figure S10D), suggesting that the NMF-derived structural connectivity patterns may also provide complementary dimensional information related to positive symptom severity.
Subtype-I was associated with peripheral inflammatory markers
We observed a significant difference in the neutrophil percentage (t = −3.207, P = 0.004, FDR-corrected) between the two subtypes (Supplementary Table S3). We also observed trend-level, nonsignificant differences in other inflammatory markers, including neutrophil/lymphocyte ratio (NLR; t = −2.383, P = 0.025, not corrected) and systemic inflammation response index (SIRI; t = −2.067, P = 0.049, not corrected) in subtype-I, and higher lymphocyte percentage (t = 2.201, P = 0.037, not corrected) in subtype-II (Figure 2c and Supplementary Table S3). Only trend-level, nonsignificant positive correlations between peripheral inflammation markers and clinical symptoms were observed within subtype-I (Supplementary Table S9). No significant association was found between peripheral inflammation markers and clinical symptoms in patients from subtype-II. These findings suggest that the two connectivity-derived subtypes may show marginally different peripheral inflammatory profiles and possible subtype-specific inflammation–symptom associations.
Subtype-II experienced a trend-level higher degree of symptom relief, particularly within the olanzapine subgroup
Using longitudinal data, a trend-level subtype-by-time interaction effect was observed for PANSS positive scores, showing that patients from subtype-II tended to experience a higher degree of PANSS positive symptom relief compared with those in subtype-I (t = 2.366, P = 0.021, not corrected; Figure 2e). Further, we stratified patients with SCZ by medication group (risperidone, paliperidone, and olanzapine) and evaluated the subtype-by-time interaction effect on PANSS scores within each medication group. In the olanzapine group, trend-level subtype-by-time interaction effects were observed for negative symptoms (t = 2.915, P = 0.015, not corrected), general psychopathological symptoms (t = 2.485, P = 0.032, not corrected; Supplementary Figure S5D), and total scores (t = 3.509, P = 0.006, not corrected; Figure 2f), with patients in subtype-II showing a nonsignificant, trend-level higher degree of short-term symptom relief compared with those in subtype-I. In contrast, no apparent subtype-by-time interaction effects were observed for changes in PANSS scores in the risperidone and paliperidone groups (Figure 2f and Supplementary Figure S5D). More detailed information is provided in Supplementary Table S5. These results suggest a potential subtype-related difference in short-term symptom improvement within the olanzapine subgroup.
Subtype-I showed similar neuroanatomical vulnerability patterns with bipolar disorder
Given that the normative connectome organization plays a key role in anchoring anatomical changes in SCZ (Georgiadis et al., Reference Georgiadis, Larivière, Glahn, Hong, Kochunov, Mowry, Loughland, Pantelis, Henskens, Green, Cairns, Michie, Rasser, Catts, Tooney, Scott, Schall, Carr, Quidé and Kirschner2024), we next examined whether the two connectome-derived subtypes are associated with distinct neuroanatomical alteration patterns. Compared with HCs, patients from subtype-I exhibited significantly reduced cortical thickness in the left inferior temporal and precentral regions (P FDR < 0.05; corrected across 219 brain regions; Figure 3a). In contrast, patients from subtype-II demonstrated significantly lower cortical thickness in the left postcentral and superior frontal regions and higher thickness in the right superior parietal region (P FDR < 0.05; corrected across 219 brain regions; Figure 3a). No significant differences in cortical thickness were observed between the two subtypes (P FDR ≥ 0.05; corrected across 219 brain regions).
Patients in different subtypes demonstrated distinct biological endophenotypes. (a) Regional cortical thickness differences between each subtype and HCs. The t-scores derived from linear regression models were plotted for each subtype. Annotated brain regions indicate significant differences (P FDR < 0.05). Blue color indicates decreased cortical thickness in patients with SCZ compared to HCs, whereas red color indicates an increased value. (b) Correlations between the ENIGMA-derived cortical thickness maps and regional cortical thickness alterations at each subtype and group level. Asterisks indicate P spin -FDR < 0.05. (c) Mapping connectivity-based epicenters in each subtype. The correlation coefficients were plotted to represent the spatial similarity between subtype-specific cortical thickness changes and seed-based functional (left) and structural (right) connectivity measures across all cortical regions. Regions with significant associations were highlighted (P spin -FDR < 0.05) and identified as epicenters. (d) Percentage of cortical thickness changes during long-term follow-up in patients with SCZ from each subtype and HCs. The ratio of cortical thickness changes after long-term follow-up was defined as (follow-up − baseline)/(baseline). A decrease in cortical thickness at follow-up compared to baseline is indicated by blue, while an increase is represented in red. Asterisks indicate brain regions with significant group-by-time interactions (P FDR < 0.05) in patients from subtype-II compared to HCs. SCZ, schizophrenia; HC, healthy control.
Figure 3. Long description
Panel a, Baseline cortical thickness differences. Displays lateral and medial views of brain hemispheres for Subtype-I and Subtype-II. A color scale for t values ranges from negative 4.4 in blue to positive 4.4 in red. Subtype-I shows significant blue areas in the left precentral and inferior temporal regions. Subtype-II shows significant blue areas in the left postcentral and superior frontal regions and a red area in the right superior parietal region.
Panel b, Between disease similarity. A horizontal heatmap comparing Subtype-I, Subtype-II, and All groups against eight conditions: 22q11.2, Epilepsy, A D H D, A S D, B D, M D D, O C D, and S C Z. The r value scale ranges from negative 0.6 in blue to positive 0.6 in red. White asterisks indicate significant positive correlations of Subtype-I and the all group with B D and S C Z .
Panel c, Functional and structural epicenters. Two sets of brain maps showing r values from 0 to negative 0.62 in blue. Functional epicenters are concentrated in frontal and temporal regions in Subtype-I and frontal, parietal, and cingulate regions in Subtype-II. Structural epicenters are limited to the left pars opercularis in both subtypes, with the left caudal middle frontal region additionally identified in Subtype-II.
Panel d, Longitudinal changes in cortical thickness. Brain maps for Subtype-I, Subtype-II, and Control groups. The change ratio scale ranges from negative 9 percent in blue to positive 9 percent in red. Black asterisks mark regionswith significant group-by-time interactions between Subtype-II and the H C group.
We then examined whether the subtype-specific patterns of neuroanatomical vulnerabilities are associated with other mental disorders by means of spatial correlations between cortical thickness alteration patterns of each subtype and brain vulnerability maps from eight psychiatric disorders. We found that the cortical thickness alteration pattern for subtype-I was significantly correlated with the corresponding patterns observed in bipolar disorder (r = 0.527, P spin -FDR = 0.004) and SCZ (r = 0.574, P spin - FDR = 0.002; corrected across eight disorders; Figure 3b) as reported by ENIGMA (Hibar et al., Reference Hibar, Westlye, Doan, Jahanshad, Cheung, Ching, Versace, Bilderbeck, Uhlmann, Mwangi, Kramer, Overs, Hartberg, Abe, Dima, Grotegerd, Sprooten, Boen, Jimenez and Andreassen2018; van Erp et al., Reference van Erp, Walton, Hibar, Schmaal, Jiang, Glahn, Pearlson, Yao, Fukunaga, Hashimoto, Okada, Yamamori, Bustillo, Clark, Agartz, Mueller, Cahn, de Zwarte, Hulshoff Pol and Turner2018). In contrast, cortical thickness changes in subtype-II were found to be spatially correlated with 22q11.2 (r = −0.292, Pspin = 0.027) and autism (r = −0.389, P spin = 0.049, Figure 3b), highlighting the distinct neurobiological features and uniqueness of subtype-II. These two subtype-specific patterns of neuroanatomical changes also showed divergent correlations with the spatial patterns of various receptor and transporter densities, including D2, α4β2, KOR, 5-HT2a, 5-HT6, MOR, CB1, VAChT, GABAa, 5-HT1b, 5-HT4, M1, 5-HTT, NET, and SV2A (see details in Supplementary Table S7).
Additionally, we examined the subtype-specific epicenters by correlating whole-brain patterns of cortical thickness change for each subtype with the functional and structural connectivity profiles of each cortical region. In subtype-I, we identified functional epicenters predominantly in the bilateral pars orbitalis, pars triangularis, banks of the superior temporal sulcus, and lateral orbitofrontal regions, left superior frontal, caudal middle frontal, and middle temporal gyri, left pars opercularis, entorhinal, supramarginal, and temporal pole regions (P spin -FDR < 0.05; corrected across 68 brain regions; Figure 3c). In subtype-II, functional epicenters were primarily located in the bilateral rostral middle frontal regions, the left pars opercularis, posterior cingulate, and supramarginal regions, and the right superior frontal gyrus (P spin -FDR < 0.05; corrected across 68 brain regions; Figure 3c), further suggesting subtype-specific epicenters in SCZ.
We additionally investigated whether longitudinal changes in cortical thickness varied across subtypes. Significant group-by-time interactions in cortical thickness were observed in the bilateral superior parietal regions, left caudal anterior cingulate, pars opercularis, postcentral, and superior frontal regions, as well as in right inferior parietal and supramarginal regions in patients from the subtype-II compared to HCs (P FDR < 0.05; corrected across 219 brain regions; Figure 3d). We did not observe any significant group-by-time interaction effect in patients from the subtype-I compared to HCs, nor between patients in the two subtypes.
Discussion
In this study, we systematically assessed the inter-individual variability of the structural connectome in SCZ and performed connectivity-based subtyping analysis. Our results revealed that structural brain circuits involved in higher-order cognitive networks were preferentially vulnerable in SCZ patients. Based on the identified features, patients were classified into two connectivity-derived subtypes that differed in clinical and biological profiles. Specifically, one subtype (i.e. subtype-II) showed trend-level effects of more severe positive symptoms and lower NLR levels, while exhibiting distinct neuroanatomical features that diverged from the averaged pattern of SCZ. Notably, our exploratory analyses suggested greater short-term symptom improvement in subtype-II than in subtype-I within the olanzapine monotherapy subgroup, suggesting a potential subtype-related difference in short-term treatment response. Collectively, these findings suggest that altered structural dysconnectivity patterns may help characterize connectivity-defined variation among patients and contribute to understanding neurobiological and clinical heterogeneity in SCZ.
Our findings revealed greater inter-individual variability of structural connectivity in SCZ compared with HCs. Connections with significantly higher inter-individual variability in SCZ were primarily localized to the temporal, frontal, and insular regions, as well as bilateral caudal anterior cingulate gyri. Notably, the temporal, frontal, and insular regions also exhibited greater variability in cortical thickness in patients with SCZ compared to HCs (Alnæs et al., Reference Alnæs, Kaufmann, van der Meer, Cordova-Palomera, Rokicki, Moberget, Bettella, Agartz, Barch, Bertolino, Brandt, Cervenka, Djurovic, Doan, Eisenacher, Fatouros-Bergman, Flyckt, Di Giorgio, Haatveit and Westlye2019; Omlor et al., Reference Omlor, Rabe, Fuchs, Surbeck, Cecere, Huang, Homan, Kallen, Georgiadis, Spiller, Seifritz, Weickert, Bruggemann, Weickert, Potkin, Hashimoto, Sim, Rootes-Murdy, Quide and Homan2025). Gray matter and resting-state functional abnormalities in these regions have frequently been reported in previous neuroimaging studies and are extensively associated with the core pathophysiology of SCZ (Li et al., Reference Li, Zhang, Wang, Yang, Yang, Gao and Du2025; Snelleksz et al., Reference Snelleksz, Rossell, Gibbons, Nithianantharajah and Dean2022; Tian et al., Reference Tian, Zalesky, Bousman, Everall and Pantelis2019). Our reported higher inter-individual variability in the structural brain connectome extends these previous studies, offering further support for increased variability in the dysconnectivity patterns of SCZ at the macroscale level. Although inter-individual variability and pathophysiological heterogeneity are conceptually distinct, increased variability was likely associated with divergent pathophysiological processes that result in greater within-group variance among SCZ patients.
Our observations of the inter-individual variability in SCZ further pointed to significant alterations in the higher-order cognitive networks. Notably, higher-order cognitive networks serve as central components of the human brain network, playing important roles in supporting global neural integration and cognitive function (Baggio et al., Reference Baggio, Segura, Junque, de Reus, Sala-Llonch and van den Heuvel2015; Buehler, Zhang, & Roiser, Reference Buehler, Zhang and Roiser2025; Dong et al., Reference Dong, Zhang, Labache, Zhang, Ooi, Yeo, Margulies, Holmes and Zuo2024; Smallwood et al., Reference Smallwood, Bernhardt, Leech, Bzdok, Jefferies and Margulies2021; van den Heuvel, Kahn, Goñi, & Sporns, Reference van den Heuvel, Kahn, Goñi and Sporns2012). Disruptions in these central connections of the brain network have emerged as prominent features in both first-episode and chronic SCZ (Cui et al., Reference Cui, Wei, Xi, Griffa, de Lange, Kahn, Yin and van den Heuvel2019; Dong et al., Reference Dong, Wang, Chang, Luo and Yao2018; Hua et al., Reference Hua, Cummings, Roach, Fryer, Loewy, Stuart, Ford, Vinogradov and Mathalon2023; Menon, Palaniyappan, & Supekar, Reference Menon, Palaniyappan and Supekar2023), contributing to the core clinical phenotype of the disorder, such as maladaptive behaviors and cognitive deficits (Fan et al., Reference Fan, Yu, Pinkham, Zhu, Tang, Wang, Zhang, Ma, Zhang, Zhang and Dai2022; Zhao et al., Reference Zhao, Tian, Yan, Yue, Yan and Zhang2017). Our observations of significant enrichment in connections with higher inter-individual variability in central networks suggest varying degrees of dysconnectivity effects in key brain functional networks, reflecting the variability of connectome disturbance across patients with SCZ.
Two SCZ subtypes were identified based on the dysconnectivity patterns, showing different levels of peripheral inflammatory markers. Contributions of immune factors to the onset and prognosis of SCZ have been established previously (Buckley, Reference Buckley2019). Specifically, increased NLR, a biomarker of systemic inflammation, has consistently been reported in patients with SCZ (Bhikram & Sandor, Reference Bhikram and Sandor2022; Zheng et al., Reference Zheng, Zhou, Chen, Fu, Zhang and Zhu2024), with positive associations observed between NLR and symptom severity (Šagud et al., Reference Šagud, Madžarac, Nedic Erjavec, Šimunović Filipčić, Mikulić, Rogić, Bradaš, Bajs Janović and Pivac2023), cognitive functions (Liang et al., Reference Liang, Guan, Sun, Hao and Xiu2024), and treatment response (Labonté, Zhand, Park, & Harvey, Reference Labonté, Zhand, Park and Harvey2022). Our results are thus consistent with recent reports demonstrating significant relationships between NLR and PANSS total scores (Wang, Chen, Guan, & Li, Reference Wang, Chen, Guan and Li2024). Meanwhile, patients from subtype-II tended to show lower NLR values than those from subtype-I, with no association between inflammatory markers and symptoms, suggesting a preliminary, potential subtype-related difference in peripheral inflammatory profiles. It should be noted that our peripheral inflammatory data were available only in a subset of patients, which may limit the statistical power and generalizability. Moreover, these markers were derived from peripheral blood cell profiles and may not directly reflect inflammatory processes occurring within the central nervous system. Further validation in larger cohorts with more comprehensive immune profiling would thus be important to clarify how immune-related alterations are linked to dysconnectivity patterns in SCZ.
SCZ patients from subtype-II tended to show marginally greater symptom improvement during short-term treatment, particularly within the olanzapine monotherapy subgroup. Previous studies have reported variations in treatment response to antipsychotic medications among patients with SCZ. For instance, Case et al. identified distinct response trajectories (classes) in PANSS total scores over 12 weeks of treatment with risperidone or olanzapine in patients with SCZ or schizoaffective disorder (Case et al., Reference Case, Stauffer, Ascher-Svanum, Conley, Kapur, Kane, Kollack-Walker, Jacob and Kinon2011). Another study, based on changes in the reduction rate of the PANSS total scores over 6 weeks, classified patients with SCZ into high-trajectory and low-trajectory groups, corresponding to better and worse antipsychotic responses, respectively (Dai et al., Reference Dai, Wu, Tang, Yue, Yan, Zhang, Tan, Deng, Chen, Yang, Lu, Wang, Yang, Zhang, Yang, Li, Lv, Tan, Zhang and Li2020). These studies have primarily focused on heterogeneity in clinical symptom changes among patients, with limited exploration of the underlying heterogeneous pathophysiology. Our exploratory findings extend these observations by suggesting that patient subgroups driven by connections with different patterns of mean and/or dispersion effects may differ in short-term treatment response. Further studies with larger medication-specific samples and controlled treatment designs are needed to determine whether connectivity-derived subtypes may inform treatment-response stratification.
Our comparative analyses show that only patients from subtype-I displayed a high degree of proximity in brain morphological features to SCZ and BD, whereas no significant relationships were identified in subtype-II. Previous neuropathological and neuroimaging studies have highlighted the common biological mechanisms underlying SCZ and BD, including high genetic correlations (Cross-Disorder Group of the Psychiatric Genomics Consortium et al., Reference Lee, Ripke, Neale, Faraone, Purcell, Perlis, Mowry, Thapar, Goddard, Witte, Absher, Agartz, Akil, Amin, Andreassen, Anjorin, Anney and Anttila2013), shared psychotic symptoms (Pearlson, Reference Pearlson2015), and overlapped disruptions in gray matter volume (Chang et al., Reference Chang, Womer, Edmiston, Bai, Zhou, Jiang, Wei, Wei, Ye, Huang, He, Xu, Tang and Wang2018; Wei et al., Reference Wei, Huang, Cui, Xu, Wei, Tang, Tang, Hu, Zhang, Liu, Wang and Wang2026). Our reported associations between global patterns of brain cortical thickness in subtype-I patients and BD emphasize the shared pathophysiological changes between the two disorders. By contrast, the distinct global cortical thickness patterns observed in patients from subtype-II, which are unrelated to SCZ, BD, or other psychiatric and neurological disorders, highlight the subtype-specific cortical thickness features.
With regard to network-related cortical changes, we found that patients from different subtypes demonstrated different functional and structural disease epicenters. Epicenters are regions where normative connectivity features most closely resemble disease-related patterns of morphological alterations, and reflect the degree to which network connections match the morphological alteration patterns (Larivière et al., Reference Larivière, Rodríguez-Cruces, Royer, Caligiuri, Gambardella, Concha, Keller, Cendes, Yasuda, Bonilha, Gleichgerrcht, Focke, Domin, von Podewills, Langner, Rummel, Wiest, Martin, Kotikalapudi and Bernhardt2020; Shafiei et al., Reference Shafiei, Markello, Makowski, Talpalaru, Kirschner, Devenyi, Guma, Hagmann, Cashman, Lepage, Chakravarty, Dagher and Mišić2020). The functional and structural epicenters we observed in patients from subtype-I are largely in line with recent reports, with the left pars opercularis and middle temporal regions identified as the most consistent disease epicenters (Georgiadis et al., Reference Georgiadis, Larivière, Glahn, Hong, Kochunov, Mowry, Loughland, Pantelis, Henskens, Green, Cairns, Michie, Rasser, Catts, Tooney, Scott, Schall, Carr, Quidé and Kirschner2024). In patients from subtype-II, epicenters were more specific relative to the group average results and distinctly differed from subtype-I. Integrating findings from the similarities in cortical thickness changes across multiple brain disorders, the observed subtype-specific neuroanatomical features in patients from subtype-II provide converging evidence for the unique neurobiological profile associated with this subtype and further underscore the heterogeneous pathophysiology underlying SCZ.
Several considerations should be addressed in the present study. First, patients with SCZ in our study were, on average, 2 years older than HCs. Given that white matter connectivity patterns can vary with age (Cetin-Karayumak et al., Reference Cetin-Karayumak, Di Biase, Chunga, Reid, Somes, Lyall, Kelly, Solgun, Pasternak, Vangel, Pearlson, Tamminga, Sweeney, Clementz, Schretlen, Viher, Stegmayer, Walther, Lee and Kubicki2020; Damoiseaux, Reference Damoiseaux2017), we controlled for age by including it as a covariate in our analysis, ruling out age as a major contributor to the variability in structural connectivity. Second, DWI data were preprocessed using the FSL eddy preprocessing pipeline, but susceptibility distortion correction was not performed because reverse phase-encoded images or field maps were not available. Third, the independent validation cohort was collected at a different hospital; however, both the discovery and validation cohorts originated from the same geographic region. In addition, because the observed difference in clinical symptoms between the identified subtypes did not survive correction for multiple testing, the clinical distinction based on this finding should be considered preliminary. Replication and validation in larger, independent, multisite, and geographically diverse cohorts are needed to confirm this finding and establish the broader generalizability and clinical relevance of these subtypes. Fourth, although baseline medication status did not differ between subtypes, some patients had a history of antipsychotic exposure, and detailed illness-duration information was not fully available; therefore, potential effects of illness chronicity and medication exposure on structural connectivity cannot be fully excluded. Fifth, some patients with SCZ in our study dropped out during short- and long-term follow-up, resulting in a reduced sample size for the longitudinal data, especially for long-term MRI follow-up. In addition, medication specificity analyses were restricted to patients receiving antipsychotic monotherapy, resulting in limited sample sizes within each medication subgroup. These analyses were also based on short-term treatment during hospitalization rather than on randomized treatment assignment. Therefore, conclusions regarding subtype-specific treatment response within specific medication subgroups should be considered preliminary, and future studies with larger samples and controlled treatment designs are warranted.
Collectively, our study reveals increased inter-individual variability in the structural connectome of SCZ, concentrated in central connections supporting higher-order cognitive functions. Based on the identified connections, patients were grouped into two connectivity-derived subtypes that showed potential differences in clinical symptoms and short-term treatment response, together with distinct endophenotypic features. SCZ may thus be conceptualized as a disorder of heterogeneous structural connectome organization, in which individual-specific dysconnectivity profiles may be associated with potentially distinct clinical trajectories and differential treatment responsiveness.
Supplementary material
The supplementary material for this article can be found at http://doi.org/10.1017/S0033291726105972.
Funding statement
This work was supported by the National Natural Science Foundation of China (Grant Nos. 82572326, U25A2083, and 82202264 to Y.W.; Grant Nos. 82572176 and 82271949 to L.-B.C.); the Beijing Municipal Natural Science Foundation (Grant No. 7232341 to Y.W.); the Natural Science Basic Research Program of Shaanxi (Grant No. 2025SYS-SYSZD-061 to L.-B.C.), the State Key Laboratory of Neurology and Oncology Drug Development (Grant No. SKLSIM-F-2024-66 to L.-B.C.), and IIT Clinical Research Fund of The Second Affiliated Hospital of Xi’an Jiaotong University (Grant No. 2025IIT-M05 to J.-N.Z.).
Competing interests
The authors declare none.
Ethical standards
The authors assert that all procedures contributing to this work comply with the ethical standards of the relevant national and institutional committees on human experimentation and with the Helsinki Declaration of 1975, as revised in 2008.
Open access
