Graph Neural Networks and Multimodal DTI Features for Schizophrenia Classification: Insights from Brain Network Analysis and Gene Expression.

Graph Neural Networks and Multimodal DTI Features for Schizophrenia Classification: Insights from Brain Network Analysis and Gene Expression - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Search Log in Primary site navigation Search Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Neurosci Bull . 2025 Mar 18;41(6):933–950. doi: 10.1007/s12264-025-01385-5 Graph Neural Networks and Multimodal DTI Features for Schizophrenia Classification: Insights from Brain Network Analysis and Gene Expression Jingjing Gao 1 , Heping Tang 1 , Zhengning Wang 1 , Yanling Li 2 , Na Luo 3, 4 , Ming Song 3, 4 , Sangma Xie 5 , Weiyang Shi 3, 4 , Hao Yan 6 , Lin Lu 6, 7 , Jun Yan 6, 7 , Peng Li 6, 7 , Yuqing Song 6, 7 , Jun Chen 8 , Yunchun Chen 9 , Huaning Wang 9 , Wenming Liu 9 , Zhigang Li 10 , Hua Guo 10 , Ping Wan 10 , Luxian Lv 11, 12 , Yongfeng Yang 11, 12 , Huiling Wang 13 , Hongxing Zhang 11, 12, 14 , Huawang Wu 15 , Yuping Ning 15 , Dai Zhang 6, 7, 16 , Tianzi Jiang 3, 4, 17, ✉ Author information Article notes Copyright and License information 1 School of Information and Communication Engineering, University of Electronic Science and Technology of China, Chengdu, 611731 China 2 School of Electrical Engineering and Electronic Information, Xihua University, Chengdu, 610039 China 3 Beijing Key Laboratory of Brainnetome and Brain-Computer Interface, Institute of Automation, Chinese Academy of Sciences, Beijing, 100190 China 4 Brainnetome Center, Institute of Automation, Chinese Academy of Sciences, Beijing, 100190 China 5 Institute of Biomedical Engineering and Instrumentation, School of Automation, Hangzhou Dianzi University, Hangzhou, 310018 China 6 Institute of Mental Health, Peking University Sixth Hospital, Beijing, 100191 China 7 Key Laboratory of Mental Health, Ministry of Health, and National Clinical Research Center for Mental Disorders (Peking University Sixth Hospital), Beijing, 100191 China 8 Department of Radiology, Renmin Hospital of Wuhan University, Wuhan, 430060 China 9 Department of Psychiatry, Xijing Hospital, The Fourth Military Medical University, Xi’an, 710032 China 10 Zhumadian Psychiatric Hospital, Zhumadian, 463000 China 11 Department of Psychiatry, Henan Mental Hospital, The Second Affiliated Hospital of Xinxiang Medical University, Xinxiang, 453002 China 12 International Joint Research Laboratory for Psychiatry and Neuroscience of Henan, Henan Key Lab of Biological Psychiatry of Xinxiang Medical University, Xinxiang, 453003 China 13 Department of Psychiatry, Renmin Hospital of Wuhan University, Wuhan, 430060 China 14 Department of Psychology, Xinxiang Medical University, Xinxiang, 453003 China 15 The Affiliated Brain Hospital of Guangzhou Medical University (Guangzhou Huiai Hospital), Guangzhou, 510370 China 16 Center for Life Sciences/PKU-IDG/McGovern Institute for Brain Research, Peking University, Beijing, 100871 China 17 Xiaoxiang Institute for Brain Health and Yongzhou Central Hospital, Yongzhou, 425000 China ✉ Corresponding author. Received 2024 Jul 14; Accepted 2024 Dec 5; Collection date 2025 Jun. © Center for Excellence in Brain Science and Intelligence Technology, Chinese Academy of Sciences 2025. Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law. PMC Copyright notice PMCID: PMC12158903  PMID: 40100543 Abstract Schizophrenia (SZ) stands as a severe psychiatric disorder. This study applied diffusion tensor imaging (DTI) data in conjunction with graph neural networks to distinguish SZ patients from normal controls (NCs) and showcases the superior performance of a graph neural network integrating combined fractional anisotropy and fiber number brain network features, achieving an accuracy of 73.79% in distinguishing SZ patients from NCs. Beyond mere discrimination, our study delved deeper into the advantages of utilizing white matter brain network features for identifying SZ patients through interpretable model analysis and gene expression analysis. These analyses uncovered intricate interrelationships between brain imaging markers and genetic biomarkers, providing novel insights into the neuropathological basis of SZ. In summary, our findings underscore the potential of graph neural networks applied to multimodal DTI data for enhancing SZ detection through an integrated analysis of neuroimaging and genetic features. Supplementary Information The online version contains supplementary material available at 10.1007/s12264-025-01385-5. Keywords: Schizophrenia, Magnetic resonance imaging, Classification, Deep learning, Graph neural network Introduction Schizophrenia (SZ) is a severe psychiatric disorder impacting over 20 million individuals globally. Patients suffer from debilitating symptoms such as hallucinations, delusions, and cognitive deficits throughout their lives, severely impairing social and occupational functioning [ 1 – 4 ]. Estimates from the Global Burden of Disease Study 2019 (GBD 2019) indicate that SZ continues to affect ~0.32% of the global population, with minimal change in prevalence over the past decade [ 5 ]. However, significant regional disparities exist, particularly influenced by variations in healthcare access, urbanization, and social and environmental stressors [ 6 ]. Recent studies have underscored the increasing importance of environmental risk factors, including urban living, migration, and exposure to adversity in early life [ 7 ]. These factors further complicate the clinical presentation and management of SZ, thereby making timely identification crucial. However, the considerable heterogeneity in symptom presentation poses challenges for diagnosis based solely on clinical assessments. Advances in neuroimaging and genetics have increasingly illuminated the underlying neurobiological mechanisms of the disorder, enabling more targeted approaches to diagnosis and treatment. As a result, there is growing emphasis on leveraging objective, neuroimaging-based diagnostic aids to enhance SZ detection [ 8 , 9 ]. In this regard, neuroimaging techniques such as MRI have emerged as valuable tools for elucidating the intricate neuroanatomical abnormalities underlying SZ. Moreover, in recent years, applying machine learning to neuroimaging data has shown promise in uncovering the distributed structural and functional patterns that distinguish SZ patients from healthy controls [ 10 ]. Certain studies focused on machine learning algorithms for single neuroimaging modalities such as structural MRI [ 11 ]. For instance, the utilization of structural MRI has achieved classification accuracies of ~80% [ 12 ]. In contrast, others have explored the potential of resting-state functional MRI in training multivariate classifiers, assessing functional connectivity attributes, and cross-sample generalizability [ 13 ]. Notably, utilizing functional connectivity features within the superior temporal cortex facilitated the successful identification of SZ patients with a 78.6% accuracy [ 14 ]. Although the majority of studies focused on gray matter and functional anomalies, mounting evidence indicates that white matter disturbances play a pivotal role in SZ pathology [ 15 ]. Moreover, recent evidence highlights the critical role of white matter integrity in SZ pathology, which is best captured using diffusion tensor imaging (DTI) metrics such as fractional anisotropy (FA) and fiber number (FN) [ 16 ]. FA measures the degree of directional water diffusion, providing an indirect assessment of white matter microstructural integrity. At the same time, FN quantifies the number of fiber tracts, reflecting the connective strength between brain regions. The motivation behind focusing on FA and FN features lies in the need to explore white matter integrity in SZ. It offers complementary perspectives on the underlying disruptions of brain networks in SZ. By analyzing both features, we aim to provide a more comprehensive understanding of SZ’s neuropathology. Focusing on these white matter features allows for a deeper investigation into the disruptions of brain networks that are characteristic of SZ. Though some incorporated multimodal data with white matter features, these were typically supplementary classification variables. Few studies have separately examined white matter networks for classification purposes, often utilizing small, single-site, single-modality samples, thereby limiting generalizability across diverse populations [ 17 – 20 ]. For example, when exclusively applying DTI data, the classification accuracy for SZ declined to ~70% and 60% in two datasets from the previous 75.05% [ 15 ]. Furthermore, the complexities and diversities associated with handling large-scale datasets further exacerbate the inadequacies of these research methodologies [ 21 , 22 ]. The graph neural networks in deep learning are capable of capturing high-dimensional relationships from multimodal neural imaging data and exhibit superior performance; they have the ability to overcome these challenges [ 23 – 25 ]. While deep learning techniques tailored for SZ detection via white matter networks are scarce, their effectiveness is often limited by significant variability in data collected across different sites, impacting the generalizability and robustness of models. In recent years, significant advances in transcriptomics and genomics have provided deeper insights into the genetic underpinnings of SZ. Large-scale genomic studies, such as those by the Psychiatric Genomics Consortium, have identified numerous genetic loci linked to SZ, emphasizing the polygenic nature of the disorder [ 26 ]. Transcriptomics research has further revealed altered gene expression in brain regions associated with neurotransmission and synaptic plasticity, which may contribute to structural changes in neuroimaging studies [ 27 ]. By integrating genomic and transcriptomic data with neuroimaging, recent studies have begun to identify gene networks affecting white matter integrity, providing a more comprehensive view of SZ’s underlying pathology [ 28 ]. These advances underscore the importance of combining genetic data with imaging markers to improve our understanding of SZ and enhance diagnostic precision. Therefore, this study aimed to develop a reliable deep learning framework leveraging graph neural networks and diffusion MRI to uncover SZ-associated white matter biomarkers. Specifically, a large multi-site dataset with >1500 Chinese participants was utilized, which enabled us to derive a robust and generalizable model. Moreover, through graph neural network analysis and model interpretation, we endeavored to identify reliable structural alterations linked to SZ. To achieve this, we introduce a novel dual-branch graph neural network (GNN) model that leverages DTI data, specifically FA and FN features, in order to enhance the classification of SZ. This approach not only integrates critical white matter integrity and connectivity metrics but also effectively mitigates the influence of multi-site data variability through advanced harmonization techniques. Furthermore, by integrating genetic analysis, we aimed to elucidate the intricate relationships between imaging markers and genetic biomarkers. Taken together, it is hoped that this work will contribute to enhancing clinical-assisted diagnosis and enable more accurate detection of this psychiatric disorder. Materials and Methods Overall Flowchart The methodology, depicted in Fig. 1 , comprised three principal components: SZ classification, identification of brain imaging biomarkers, and genetic biomarker analysis. The specific procedures undertaken were as follows: SZ Classification: Brain network structural data obtained from DTI, specifically FA and FN measures, were used to construct graph structural representations at the individual level. The SZ was classified by applying both single-branch and dual-branch GNN architectures. Identification of brain imaging biomarkers: For each individual-level graph structure, node weight values were computed, and subsequently, the average weights were determined. This comprehensive evaluation enabled the identification of brain imaging biomarkers that exhibited significant associations with SZ. Analysis of genetic biomarkers: Partial least squares (PLS) regression was applied to elucidate the relationships between the weight matrix and the gene expression matrix. Through this analytical approach, genetic biomarkers that demonstrated linkages to the identified brain imaging biomarkers were explored. Fig. 1. Open in a new tab The overall flowchart. A SZ Classification. Brain network structure data are obtained from diffusion tensor imaging (DTI) and individual graph structure data are constructed. Classification using single-branch and double-branch GNN architectures. B Identification of brain imaging biomarkers, and genetic biomarker analysis. Analysis of the brain imaging biomarkers significantly associated with SZ based on the corresponding average brain area scores of the FA branch in the dual branches. PLS analysis of the relationship between the importance weight matrix and gene expression matrix, and exploration of the genetic biomarkers associated with identified brain imaging biomarkers. Participants The study utilized a multi-site dataset comprising 1507 participants from across China. Of these, 838 were diagnosed with SZ, while 669 constituted a healthy control cohort (see Supplementary Table S1 ). SZ patients were recruited from seven renowned hospitals: Peking University Sixth Hospital, Beijing Hui Long Guan Hospital, Anhui Medical University, Xinxiang Medical College Second Affiliated Hospital, Guangzhou Psychiatric Hospital, Wuhan University People’s Hospital, and Zhumadian Psychiatric Hospital. Symptom severity was assessed using the Positive and Negative Syndrome Scale (see Supplementary Table S2 ), administered to patients within one week of their MRI scans. Image Acquisition Dataset 1 was obtained at Peking University Sixth Hospital using a Siemens TrioTim scanner. Dataset 2 was procured at Beijing Huilongguan Hospital using a Siemens TrioTim scanner. Dataset 3 was sourced from Anhui Medical University using a GE Medical Systems Signa HDxt scanner. Dataset 4 was obtained from Xinxiang Medical College Second Affiliated Hospital using two devices. Some diffusion-weighted imaging (DWI) data were acquired by the Siemens Verio scanner, some DWI data were acquired by the GE Medical Systems Signa HDxt scanner. Dataset 5 was collected at Guangzhou Psychiatric Hospital using a Philips scanner. Dataset 6 was secured at Wuhan University People’s Hospital using a GE Medical Systems Signa HDxt scanner. Dataset 6 was secured at Wuhan University People’s Hospital using a GE Medical Systems Signa HDxt scanner. Dataset 7 was obtained at Zhumadian Psychiatric Hospital using a GE Medical Systems Signa HDxt scanner. All DWI data were acquired using a single-shot spin-echo planar imaging sequence with the parameters listed in Supplementary Table S1 . Data Preprocessing The preprocessing of DTI datasets was meticulously conducted using Pipeline for Analyzing Brain DTI (PANDA) [ 29 ], specialized software for diffusion MRI analysis. It encompasses essential processing steps, including skull stripping, motion correction, eddy current artifact correction, and the calculation of diffusion tensor parameters. Consistent with prior recommendations [ 30 ], default parameters were applied for processing the DTI images of each subject in this study. The preprocessing pipeline included the following sequential steps: (1) Conversion of images from DICOM to NIFTI format. (2) Removal of non-brain tissues based on a whole-brain mask generated from the T1-weighted image. (3) Mitigation of head motion and correction of eddy current artifacts to reduce their impact on image deformations. (4) Computation and fitting of diffusion tensor metrics for each voxel using a linear least square fitting method. (5) Co-registration of individual T1-weighted structural images with the corresponding non-diffusion-weighted (b = 0) image in the native diffusion space. (6) Alignment of the transformed structural image with the standardized template in the Montreal Neurological Institute space using the Functional Magnetic Resonance Imaging of the Brain Software Library. Among these steps, eddy currents were acknowledged as a significant factor contributing to image deformations. An affine transformation was applied to align DTI images with the T1-weighted image, effectively mitigating the adverse effects of head motion and eddy currents. After excluding data that failed to meet quality standards, 1507 viable samples were retained. To facilitate the localization and analysis of brain regions, we utilized the Brainnetome Atlas [ 31 ], comprising 246 subregions, including 210 cortical and 36 subcortical subregions, providing a comprehensive framework for brain mapping. Following previous research recommendations [ 32 ], the Fiber Assignment by Continuous Tracking algorithm, incorporated within the Diffusion Toolkit, was applied for fiber tractography to reconstruct whole-brain connectomes [ 33 ]. This algorithm involves seed-based bi-directional tracking along the principal eigenvector of the diffusion tensor, with fiber propagation halted when the deflection angle threshold of the propagation line exceeds 45° [ 34 ]. Systematic identification of white matter connections between each pair of Brainnetome Atlas regions allowed for the extraction of fibers corresponding to these connections, establishing two brain network connectivity matrices: the mean FA matrix and the FN matrix, the dimensions of these two matrices are 246 × 246. Population Stratification and Site Effect Adjustment Given that the dataset utilized in this study consists of 1507 participants collected from seven different sites, there was a potential risk of significant population stratification due to demographic and scanner-related differences. To mitigate these effects, we applied the ComBat harmonization technique, which is widely used for multi-site neuroimaging data to reduce site-specific variability while preserving biological differences. The ComBat method was applied after adjusting for covariates such as age and gender, thereby reducing the influence of non-biological variability introduced by different scanners or acquisition protocols. This adjustment ensured that the subsequent analysis focused on the inherent neurobiological characteristics of SZ, rather than artifacts introduced by site differences. In our study, the ComBat harmonization technique was applied to the reconstructed FA and FN data, rather than raw data, to mitigate site-specific variability while preserving biological variance. We used site variables as batch identifiers and adjusted for covariates such as age and gender to more accurately eliminate batch effects while retaining meaningful neurobiological differences. The harmonization process was applied following the preprocessing of the DTI data, which included corrections for eddy currents and motion artifacts, ensuring consistency in the data prior to applying the ComBat adjustment. The Construction of Graph Data Upon completion of the image preprocessing phase, two brain connectivity networks were generated for each participant, each with a size of 246×246. These networks were designed to represent interconnections among various regions as delineated in the Brainnetome Atlas. Specifically, the networks consisted of weighted values, denoting either the average FA or the FN between connected nodes. Since both the FA and FN networks resemble graph-structured data, we independently constructed graph data for each. These graph data sets were subsequently utilized as inputs in our model. In our model, each sample corresponds to two sets of graph-structured data, namely the FA graph-structured data and the FN graph-structured data. Within each of these data structures, the nodes correspond to distinct brain regions, totaling 246 nodes in all. When constructing the node features, we used each row of the FA network as the node features for the corresponding node. This approach effectively encapsulated the connectivity between the specific node and the remaining 246 nodes. Simultaneously, we treated the network as an adjacency matrix, the value of which is used to determine the edge connection between the corresponding nodes. We set the threshold to 0. Those greater than 0 are considered connected, and those less than or equal to 0 are considered not connected. This methodology culminated in the formation of the corresponding graph-structured data, as represented in the equation: G FA = V , E , X 1 in which G represents the graph-structured data, V represents the set of nodes, E represents the set of edges, and X represents the set of node features. In parallel to the construction of the FA network, a similar procedure is applied to construct graph data for the FN network, facilitating its incorporation into the graph neural network for classification purposes. The representation of the FN graph-structured data can be represented as follows: G FN = V , E , X 2 Gene Data and Preprocessing For this study, the Allen Human Brain Atlas (AHBA) dataset, renowned for its comprehensive high-resolution coverage of the entire human brain, was used as a foundational resource. It comprises microarray and RNA sequencing data derived from 3702 distinct spatial samples extracted from six adult brain donors [ 35 ]. To enhance the accuracy of correlating whole-genome and brain imaging data, only the RNA sequencing data was utilized. The genes were meticulously filtered and relabeled leveraging the tools provided by Abagen ( https://github.com/rmarkello/abagen ). Furthermore, the MRI images of the AHBA donors were mapped to the 246 delineated brain regions in the Brainnetome Atlas. Consequently, a gene expression matrix with dimensions of 246 ×13561, with values ranging from 0 to 1, represents the expression levels of these 13561 genes in 246 brain regions, was successfully derived. Graph Neural Network and Analysis The Graph Attention Network (GAT), a prominent graph neural network architecture incorporating an attention mechanism, has garnered significant attention due to its remarkable performance across diverse domains [ 36 – 38 ]. Notably, its effectiveness in MRI data classification tasks has been validated by previous investigations [ 39 ]. Given the relevance of FA and FN data, which effectively capture the brain’s connectivity patterns, we chose to construct graph-structured representations from these modalities as input to the GAT for our SZ classification task. Expanding upon this foundation, we introduce a dual-branch GAT network meticulously designed to integrate both FA and FN data for the classification task. Concurrently, each branch retained the capacity to independently perform classification using the respective FA or FN data, as illustrated in Fig. 2 . This dual-branch architecture allows leveraging complementary information from the two modalities while evaluating their individual discriminative power. Fig. 2. Open in a new tab Graph neural network. A The single branch graph neural network. Construction of the same network structure for FA and FN data: input, (GAT+Layer norm+Relu+Topkpooling) × 3, Global pooling, mlp, and output. B The dual branch graph neural network. The same network structure as the single branch for FA and FN data is applied, and features are fused using concat after global pooling. The GAT architecture plays a crucial role in capturing the complex, nonlinear relationships between different brain regions, thanks to its multi-layered design that allows features to be dynamically learned based on node importance. Specifically, each layer in the GNN is tasked with processing connectivity data and learning unique node representations by iteratively aggregating and transforming neighborhood information. The first GNN layer captures the most local connections between nodes, focusing on the immediate relationships within individual brain regions. Subsequent layers then build upon these relationships, allowing for the extraction of higher-order features that represent increasingly global and abstract interactions across the entire network. This multi-layered structure of GAT helps the model learn nonlinear interactions by weighting node features differently depending on their connectivity, enhancing the ability to capture complex functional relationships between regions of interest (ROIs). In our proposed network architecture, we implemented a single-branch classification approach for both FA and FN graph-structured data, utilizing an identical network structure. The overall network design encompasses three layers of GAT, each layer adeptly extracting features from the graph-structured data. Notably, GAT distinguishes itself from traditional Graph Convolutional Networks (GCN) by considering not only information from neighboring nodes but also the weights between nodes during the node feature update process, thereby adeptly capturing significant inter-node relationships [ 40 ]. The following formula can be used to represent the specific procedure. The input to GAT is a set of node features: h = h → 1 , h → 2 , … h → N , h → i ∈ R F 3 Using the learnable weight matrix W ∈ R F ′ × F , which is applied to every node. A shared attention mechanism: a : R ′ × R ′ → R , which is performed self-attention on the nodes. The importance of the characteristics of node j to node i is obtained: e ij = a W h → i , W h → j 4 They are normalized across all choices of j using the softmax function: a ij = s o f t m a x j e ij = e x p e ij ∑ K ∈ N i e x p e ik 5 a nonlinearity σ is applied to get the output: h → i ′ = σ ∑ j ∈ N i a ij W h → j 6 h = h → 1 , h → 2 , … h → N , h → i ∈ R F 7 where N is the number of nodes, F is the number of features in each node, and the output has potentially different cardinality F’. Following each graph convolutional layer, we apply the Rectified Linear Unit activation function and layer normalization techniques. This effectively addresses concerns related to gradient explosion or vanishing, significantly reducing the training time of the feed-forward neural network and lending further stability to the classification process [ 41 ]. Finally, a multi-layer perceptron (MLP) is applied to classify the globally pooled features. By utilizing a dual-branch network, FA and FN data are treated as two related sub-tasks and processed independently in parallel branches. For each branch, we adopt the same graph neural network structure and data input as a single branch. Specifically, FA and FN are encoded separately using three layers of GAT, allowing the model to learn specific features unique to each modality. After feature extraction, the learned features from each branch are subjected to a global pooling layer, which reduces the data’s dimensionality while retaining the most informative features from each modality. The features are then fused by concatenation to integrate the complementary information provided by both FA and FN. This concatenated representation captures both modalities’ unique and shared aspects, enhancing the network’s ability to model complex relationships. The fused features are subsequently passed to an MLP for the final classification, providing a robust mechanism to integrate and utilize the unique characteristics of each modality effectively. This strategic design optimally leverages the inherent information across the two datasets, potentially improving classification accuracy by combining microstructural insights from FA with connectivity measures from FN. Furthermore, our network architecture boasts scalability, as we proficiently use the same graph neural network structure to extract features from FA and FN graph-structured data separately. The globally pooled features obtained from each branch are effectively merged through a concatenation layer. Finally, a two-layer fully connected network undertakes classification based on the merged features, discerning between individuals with SZ and NCs with precision. Interpretable Model Given the promising capabilities of GAT in analyzing complex neural connections, it becomes essential to explore how specific features of white matter, such as FA and FN, can be integrated into this framework to enhance our understanding of SZ. The motivation behind studying the influence of FA and FN features on DTI in the context of SZ lies in the understanding of how white matter abnormalities contribute to the disorder’s pathology. Given that SZ is characterized by widespread structural and functional disruptions in various brain regions, investigating the role of FA and FN provides insights into the specific connectivity patterns associated with these changes. Abnormal FA values can indicate compromised white matter integrity, while FN features can reveal complex network interactions that are crucial for understanding the neurobiological basis of SZ. Focusing on these features may enhance the understanding of the relationship between white matter connectivity and SZ, which may lead to improved diagnostic and therapeutic approaches. In our study, we achieved favorable results in SZ classification through the implementation of the GAT. However, the inherent non-linearity of the GAT model poses challenges in comprehending its internal dynamics and decision-making processes, thereby rendering the model a “black box” that is difficult to explain and comprehend [ 42 ]. To address this issue and reduce its complexity, the TopKpooling method is used to improve interpretability by selecting the most important nodes based on learned attention scores. The specific procedure involves several steps: Firstly, we create a 246 * 1 trainable parameter and multiply the node features of 246 * 246 with the trainable parameter of 246 * 1 to obtain a more manageable form of 246 * 1. Next, by applying an activation function, we derive weight coefficients for each node, thereby forming a weight coefficient vector of size 246 x 1 [ 43 , 44 ]. Subsequently, we retain the top 50% of nodes with the highest scores, which allows us to focus on brain regions that contribute significantly to the classification outcome. By examining these selected nodes (brain regions), we can infer that the connectivity and integrity of these regions are particularly relevant for distinguishing between SZ patients and healthy controls. This selective pruning not only reduces model complexity but also highlights specific brain regions that serve as potential biomarkers, thereby enhancing the clinical interpretability of our findings. The specific procedure can be represented by the following formulae: y = σ x p p 8 i = top k y 9 x ′ = x · tan y i 10 V ′ = V i 11 where x represents the node feature of 246 * 246, p represents the trainable vector of size 246 * 1, and x is multiplied by p to obtain x p of 246 * 1, y represents the scores of each node, K represents the number of retained nodes, x’ represents the updated features, and V’ represents the updated corresponding node sets. By applying this approach, we strive to enhance the transparency and interpretability of our model’s outcomes while simultaneously facilitating the identification of crucial brain regions relevant to SZ classification. Gene Expression Analysis In this study, we applied the Partial Least Squares (PLS) method to investigate the relationship between genetic biomarkers and the weights associated with SZ brain images. The overarching objective was to discover the regression relationship between gene expression matrices and explanatory weights. The PLS regression is a multivariate statistical method that offers a valuable approach to simultaneously analyze multiple dependent variables. It proves particularly advantageous in handling limited sample sizes and addressing multicollinearity issues with independent variables. As a result, PLS regression has been widely applied in this field [ 45 – 47 ]. In our PLS analysis, gene expression data served as the independent (predictor) variables, while the brain region weights derived from the graph neural network were treated as the dependent (response) variables. This approach was used to analyze the potential relationships between gene expressions and brain ROIs, thereby identifying genetic markers strongly linked to specific brain regions involved in SZ. Our analysis revealed that PLS1 had the strongest association with the predictor variable, with a cumulative explanatory variance of 83.385% in node weights. To assess whether this association was greater than expected by chance, we conducted 1,000 bootstrap resampling iterations ( P <0.05) [ 48 ], recalculating the PLS model for each resample. The bootstrap analysis revealed that <1% of the resampled models showed a variance explained by PLS1 equal to or greater than that of the original model ( P <0.01), suggesting that the association is statistically significant and unlikely to have occurred by chance. PLS1, the first latent variable, explained 83.385% of the variance in the node weights, which is substantially higher than would be expected by chance, as indicated by a P -value of <0.01 from bootstrap resampling. Consequently, genes with substantial weights on PLS1 have the potential for association with SZ brain imaging markers. Although we did not conduct spatial permutation tests due to computational limitations and data availability, we used bootstrap resampling with 1,000 iterations as an alternative. This approach provided a robust estimate of variance significance. However, future work should consider incorporating spatial permutation testing to directly address spatial autocorrelation effects more comprehensively. Results Classification Performance The principal aim of this investigation was to evaluate the potential of utilizing DTI data in conjunction with deep learning methodologies to classify and investigate structural brain alterations in individuals diagnosed with SZ. Our experimental endeavors focused on evaluating the effectiveness of FA and FN data within single-branch and dual-branch network classification. To improve the generalization of the model, we merged data from 7 distinct sites. Subsequently, we partitioned this combined data into 10 non-overlapping segments to conduct experiments utilizing a 10-fold cross-validation approach. In each experiment, 9 partitions were used for training the model, while one partition was held out for validation. Model training was executed using the Adam optimization algorithm, with a learning rate of 0.001 and a batch size of 16. The cross-entropy loss function served as the foundation for the classification process. In addition, to obtain reliable experimental results, we applied multiple different random seeds, fully shuffled the data distribution of the training and testing sets, and repeated multiple 10-fold cross-validation experiments. The average accuracy error of multiple 10-fold cross-validation was within 0.01, indicating that the data distribution had no impact on this experiment. The accuracy differences across multiple 10-fold cross-validation runs showed slight variations, generally remaining within a reasonable range. This indicates that the data distribution had a limited impact on the stability of the experimental results. In the evaluation of our classification models, we applied a range of metrics, including accuracy (ACC), specificity (SPE), sensitivity (SEN), and the F1 score. Ultimately, our findings encompassed the results obtained from both FA and FN data in the context of single-branch network classification and dual-branch network classification. Detailed data (see Supplementary Tables S3 , S4 and S5 ) have been included for a comprehensive reference. Results of Classification in a Single Branch According to the results in Supplementary Table S3 , our proposed single-branch GNN exhibited superior classification performance on the FA dataset. It achieved an average ACC of 72.59%, a SEN of 76.40%, a SPE of 67.83%, and an F1 score of 75.50%. These outcomes highlight the proficiency of the single-branch GNN in accurately identifying the majority of samples within the FA dataset and effectively avoiding misclassifications. In contrast, the results in Supplementary Table S4 indicate that our single-branch GNN had a slightly inferior classification performance on the FN dataset compared to the FA dataset. It achieved an average ACC of 70.86%, a SEN of 73.19%, a SPE of 67.65%, and an F1 score of 73.73%. This suggests that our single-branch GNN may encounter limitations when handling samples in the FN dataset, indicating a need for further optimization and improvement. Results of Classification in Dual Branch As detailed in Supplementary Table S5 , our proposed dual-branch GNN had a significant performance, with an average ACC of 73.79%, a SEN of 77.59%, a SPE of 70.11%, and an F1 score of 75.96% across the entire dataset. These results highlight a noteworthy improvement in classification performance when using the dual-branch structure compared to the single-branch GNN, indicating the effectiveness of this approach. The high SEN implies that the model is particularly effective at identifying true positive cases of SZ, which is crucial for ensuring that individuals receive timely and appropriate treatment. Meanwhile, the SPE values indicate that the model effectively minimizes false positive rates, thereby reducing the potential for misdiagnosis of healthy individuals. Together, these metrics signify a promising advancement in the reliability of imaging-based diagnosis of SZ, suggesting the potential clinical value of this approach. Such improvements in classification performance underscore the significance of integrating white matter brain network features (i.e., FA and FN) with GNN-based analysis. This integration has substantial potential for clinical application, providing a non-invasive and objective tool to aid clinicians in diagnosing SZ with enhanced accuracy. By combining these complementary features, our dual-branch GNN not only boosts the overall accuracy of the classification but also enhances interpretability by emphasizing important features linked to SZ pathology. Based on these findings, our dual-branch GNN demonstrated optimal classification performance on datasets integrating FA and FN, which clearly indicates the effective utilization of unique features from these two datasets by the dual-branch network. Furthermore, having established the robustness of our dual-branch GNN model for SZ classification, we next aimed to explore the specific brain imaging markers that contribute to the classification outcomes. By identifying and analyzing these markers, we sought to provide a biological interpretation of the structural alterations in individuals with SZ, thereby enhancing our understanding of the neurobiological mechanisms underlying the disorder. In light of the excellent performance in integrating the FA and FN datasets, as well as the importance of FA as an indicator of corresponding brain regions in DTI images, we also plan to explore potential brain imaging markers and genetic biomarkers based on the FA network in future studies. Ultimately, this exploration aims to deepen our understanding of the neurobiological mechanisms underlying SZ, with a focus on identifying new biomarkers that could further enhance the diagnostic process. Results of Classification Compared with Other Methods We compared our model with five other competing methods: Random Forest [ 49 ], Support Vector Machine (SVM) [ 50 ], XGBoost [ 51 ], ResNet [ 52 ], and GCN [ 53 ]. GCN uses the same network structure as this experimental method, with a penalty parameter of 1.0 in the SVM, 100 decision trees in the Random Forest, 100 boosting rounds in XGBoost, and a maximum depth of 6 trees. ResNet uses the classic ResNet18 but changes the input size and uses binary classification. All other parameters in the above methods are defaults. We evaluated their performance across diverse modal data. We utilized the F-score algorithm for feature selection for the FA and FN matrix, specifically selecting the top 1000 features as input for the machine learning methods. In the case of ResNet, we considered the matrix as a 2D image and utilized it as input for the network. Although the FA and FN matrices are represented differently (i.e., feature vectors, 2D images, graph structures), the core information contained in the data remains consistent across all models. We applied the 10-fold cross-validation method for the comparison method, and for the deep learning method, we applied the same training strategy as in this experiment. As illustrated in Table 1 and Fig. 3 , our model exhibited superior performance in terms of ACC metrics compared to the other classification methods. Furthermore, concerning ACC, SPE, SEN, F1_score, and the area under the curve metric, our model also demonstrated superior performance, especially when applied to the FA + FN dataset, surpassing the performance of alternative classification methods. These outcomes demonstrate the exceptional classification abilities of our model, effectively distinguishing abnormal samples and minimizing false positive rates. Table 1. Comparison of our method and other baseline methods. Method Model ACC (%) SEN (%) SPE (%) F1_score(%) Our method FA+FN 73.79 77.59 70.11 75.96 FA 72.59 76.40 67.92 75.50 FN 70.86 73.19 67.65 73.73 GCN FA+FN 71.47 74.12 67.72 73.85 FA 70.61 72.09 69.63 74.37 FN 69.27 72.46 68.50 74.67 ResNet18 FA+FN 66.87 74.36 50.08 70.59 FA 64.49 70.13 41.84 61.44 FN 63.82 72.09 42.73 71.16 XGBoost FA+FN 69.20 76.75 59.82 73.31 FA 66.62 74.25 57.04 71.07 FN 69.99 77.86 60.03 74.14 Random forest FA+FN 68.81 83.87 50.01 74.84 FA 64.69 82.26 42.75 72.02 FN 69.13 81.99 53.05 74.52 SVM FA+FN 62.18 73.08 48.33 67.99 FA 69.73 82.36 54.13 79.37 FN 58.66 68.10 46.83 64.37 Open in a new tab Fig. 3. Open in a new tab The area under the curve of this experimental method and the comparative experimental method; x-axis, false positive rate, and y-axis, true positive rate. Eliminating the Impact of Age, Gender, and Site Differences To ensure the integrity of our findings, we rigorously screened our samples to address potential influences from factors such as age and gender disparities across groups. Initially, we assessed the mean differences in age between the NC and SZ groups, and to minimize age-related confounding effects, we removed samples from the 31 youngest participants in the NC group and the 30 oldest participants in the SZ group. This exclusion of 61 samples ensured that our analysis focused on a core age group where brain structural changes related to SZ are most pronounced and comparable. Including participants at age extremes could introduce variability associated with normal developmental or degenerative changes, which are not specific to SZ and could confound our analysis. By excluding these outliers, we aimed to better isolate SZ-specific brain alterations. Subsequently, we conducted statistical tests to ensure the comparability of demographic variables between the NC and SZ groups. Our analyses, including t -tests for age and χ 2 tests for gender, indicated no significant differences between the groups ( P >0.05), reaffirming the comparability of these groups after screening. Following this, we proceeded with 10-fold cross-validation using the refined samples to validate our results, ensuring consistency in our experimental setup. The summarized outcomes are presented in Table 2 . Table 2. The 10-fold cross-validation results of screened samples. Method Model ACC (%) SEN (%) SPE (%) F1_score (%) Our method FA+FN 73.65 (0.02) 76.28 (0.03) 69.91 (0.06) 76.31 (0.02) FA 72.07 (0.02) 77.82 (0.09) 64.26 (0.10) 75.54 (0.43) FN 69.70 (0.03) 76.99 (0.04) 60.47 (0.06) 73.88 (0.03) Open in a new tab Means and standard deviations are reported. To further address potential biases arising from variations across data collection sites, we applied the ComBat harmonization method to the screened samples, as outlined by [ 54 ]. This method standardizes data across different sites by identifying batch effects through site variables while adjusting for covariates such as age and gender to more accurately eliminate these effects while retaining meaningful biological differences. The impact of this normalization process on our experimental findings is detailed in Table 3 . In addition, for our baseline comparison method, we applied the same preprocessing steps as described above; the results after processing are presented in Supplementary Table S6 . Table 3. The 10-fold cross-validation result after combat. Method Model ACC (%) SEN (%) SPE (%) F1_score (%) Our Method FA+FN 73.59 (0.02) 76.04 (0.03) 70.00 (0.08) 76.31 (0.01) FA 70.21 (0.03) 76.64 (0.05) 61.78 (0.08) 73.99 (0.03) FN 68.02 (0.02) 72.84 (0.05) 62.43 (0.07) 71.60 (0.03) Open in a new tab Means and standard deviations are reported. Upon scrutinizing the filtered data, we found no discernible inter-group disparities in age and gender, reaffirming the robustness of our screening process. Notably, the experimental outcomes remained consistent with those obtained before data screening, with marginal changes in the FA, FN, and FA+FN dual branches. Furthermore, the application of ComBat to address site differences yielded comparable results, underscoring the effectiveness of this approach in maintaining experimental integrity. Although slight variations were noted in the performance metrics of the FA, FN, and dual-branch models, these changes remained within an acceptable range, indicating the reliability of our findings. Based on the performance of our best-performing fusion dual-branch model, we conclude that the influence of demographic variables such as age and gender, as well as site variations, has been effectively mitigated. Moreover, using the 10-fold cross-validation method with merged samples from 7 sites has bolstered the generalizability and robustness of our model. Potential Brain Image Markers In our model, we implemented the TopKpooling method to calculate scores for individual brain nodes, aiming to enhance classification performance and identify potential brain imaging markers. Through evaluating the FA branch GNN in the dual branch, we organized the average scores from 10-fold cross-validation for each node in descending order. Notably, a distinct clustering of the top 50 brain regions emerged, prominently including the superior frontal gyrus, cingulate gyrus, middle frontal gyrus, thalamus, orbital gyrus, and anterior cuneiform lobe. These regions, depicted in Fig. 4 , represent key areas of interest in SZ research due to their involvement in cognitive and emotional processing. For instance, the superior frontal gyrus has been associated with working memory and executive function deficits in SZ, consistent with findings by Ehrlich et al. (2012) and Mubarik et al. (2016) [ 55 , 56 ]. The cingulate gyrus, particularly its anterior part, has been implicated in emotional regulation and social cognition, aligning with studies by Thielen et al. (2022) that highlighted its role in SZ-related cognitive impairments [ 57 ]. Furthermore, the thalamus has been shown to play a central role in sensory integration and cognitive processing, with disrupted thalamic connectivity being a hallmark feature in SZ pathology, as reported by Wang et al. [ 58 ]. The current findings underscore the convergence between identified imaging markers and well-established SZ neuropathology, providing strong evidence for the relevance of these regions as potential biomarkers for SZ. The bilateral distribution of regions such as the superior frontal gyrus and cingulate gyrus underscores their significance in understanding the neural mechanisms underlying SZ symptoms. Fig. 4. Open in a new tab Visualization of potential imaging markers associated with FA. Separate labeling of brain regions on the left and right sides that may be potential biomarkers. Potential Gene Biomarkers and Enrichment Pathways In Fig. 5 , Part A illustrates the visualization of ontology terms associated with the PLS1 gene (with pFDR <0.05). The size of each circle in the visualization represents the prevalence of the associated terms. Part B displays the enrichment network, revealing similarities within and between clusters of enriched terms. Circular nodes represent enriched terms, with node size indicating the number of encompassed genes. In addition, a color-coded scheme highlights cluster identity, where nodes of the same color belong to the same cluster [ 59 ]. Fig. 5. Open in a new tab Functional enrichment of gene transcripts. A Enrichment dot bubble. Bubble color, the converted p-value; bubble size, the number of genes; x-axis, the enrichment value; y-axis, the GO term. B Network of enriched terms. Each term is represented by a circle node, where its size is proportional to the number of input genes that fall under that term, and its color represents its cluster identity (i.e., nodes of the same color belong to the same cluster). Terms with a similarity score >0.3 are linked by an edge (the thickness of the edge represents the similarity score). In our genetic analysis, we sorted the PLS1 weights in descending order and identified 526 top- and 526 bottom-ranked genes (the top 10 and bottom 10 genes are shown in Supplementary Table S7 ). We conducted a comprehensive analysis of these genes using Metascape [ 60 ], focusing on Gene Ontology (GO) categories such as Biological Processes, Cellular Components (CC), and Molecular Functions (MF). This analysis allowed us to explore the enriched terms across these GO categories, providing insights into the biological roles of the identified genes. For example, the identified GO terms under CC include structures such as synaptic membrane and axonal components, whereas those under MF cover neurotransmitter binding and ion channel activity. By categorizing these enriched terms, we offer a detailed overview of the biological functions, highlighting their potential contributions to SZ pathology. In particular, following the enrichment analysis and adjustment for enrichment conditions (pFDR <0.05), along with the exclusion of isolated clusters, we identified the top 11 gene ontologies associated with FA. Notable terms included “postsynapse,” “axon,” and “presynapse.” These findings provide valuable insights into the potential roles of these gene ontologies in SZ pathogenesis. Discussion In this study, we introduce a novel classification model based on a GNN, specifically designed to identify individuals with SZ by analyzing DTI data. A key aspect of our research is addressing the challenges associated with verifying the performance of the classification model across multi-site datasets within the field of clinical research. These challenges include mitigating potential confounding factors such as variations in scanning equipment, diverse imaging protocols, and limited sample sizes. To effectively tackle these issues, we utilized a dataset containing 1507 samples from 7 different research sites. To further reduce potential biases introduced by differences between research sites, we applied the ComBat method to eliminate site-specific differences within the data. To enhance the robustness of our findings, we conducted 10-fold cross-validation experiments, which allowed us to validate our model’s performance across different subsets of the data. Moreover, our model leverages graph-structured data, which demonstrates its ability to capture complex non-linear relationships between different brain regions and to preserve critical topological information through the application of a GAT. Compared to traditional methods relying on one-dimensional features or two-dimensional matrix operations using Convolutional Neural Networks (CNNs), our classification model exhibited superior performance, highlighting the strength of our GNN-based approach. Our results also underscore that SZ is associated with complex, nonlinear interactions between different brain regions. The GAT model effectively captured these relationships by using attention mechanisms that assigned variable weights to connections based on their importance. These nonlinear relationships reflect how certain brain regions contribute disproportionately to SZ pathology, potentially providing insights into previously unidentified pathways of disease progression. However, due to the inherent complexity of neural networks, these relationships remain challenging to interpret fully. We plan to address these challenges in future research by integrating explainable AI techniques to better understand the model’s decision-making process. Our proposed model’s classification accuracy for distinguishing SZ patients from HCs is reported at 73.79%. Although this accuracy exceeds chance and reflects the balanced nature of the study groups, it appears somewhat lower than the results from other studies in the literature. For instance, [ 15 ] achieved an accuracy of approximately 75.05% using a Gradient Boosting Decision Tree algorithm, while varying accuracies ranging from 70% to 80% have been documented in other DTI research. However, it is essential to recognize the methodological and dataset differences that may account for these variations in performance. It may influence the classification results due to the unique characteristics of the sample. Furthermore, the feature selection process differs significantly; Liang et al. extracted features from multiple neuroimaging metrics [ 15 ], whereas our research emphasizes features derived specifically from FA and FN data. This targeted approach enables a deeper investigation into white matter connectivity, potentially affecting our model’s performance. Variations in classification accuracy can also be attributed to different evaluation methodologies, such as cross-validation strategies and the algorithms employed. Finally, the clinical context in which these studies were conducted can impact the accurate interpretation of results. Therefore, while our model performs competitively, it is vital to consider these contextual factors when comparing with the literature. Extensive research consistently indicates markedly lower FA values in specific brain regions of individuals with SZ compared to HCs. Earlier studies examining the white matter microstructure of individuals with early-onset SZ reveal notable reductions in FA scores among early-stage patients compared to their healthy counterparts [ 61 ]. These alterations in FA extend to fiber tracts that connect the cortex, cerebellum, thalamus, and cortex in individuals with psychotic disorders [ 62 ]. It is noteworthy that the pathway from the left ventral posterior cingulate cortex to the middle temporal gyrus demonstrates a substantial decrease in FA among individuals with psychiatric disorders [ 63 ]. Concurrently, studies have identified noteworthy alterations in brain connectivity in individuals with SZ, including changes in the visual network, sensorimotor network, frontoparietal network, and parietal network [ 64 – 66 ]. Our classification model capitalizes on this compelling evidence, building upon foundational data supported by distinctive FA brain networks and FN brain networks in individuals with SZ. Applying model interpretation techniques enables the identification of pivotal brain regions essential to the classification task. Our findings reveal structural alterations in brain regions closely linked to SZ, including the cingulate gyrus, frontal lobe, and parietal lobe. Particularly, the cingulate cortex plays a pivotal role in processes such as emotion formation, processing, learning, and memory. The anterior cingulum serves as the central hub of the salience network, and changes in these areas may be linked to attention deficits in SZ [ 67 ]. Earlier studies have shown decreased FA values in the posterior cingulate gyrus and right temporal middle gyrus in SZ patients compared to HCs [ 68 ]. In addition, reduced FA has been reported in the left cingulum and thalamus [ 69 ], aligning with the disruption of the cingulum bundle integrity found in patients with SZ [ 70 ]. Subsequent investigations suggest that deficits in working memory, a prominent feature of SZ pathophysiology, may be associated with abnormalities in the frontal lobe, temporal lobe, and parietal lobe [ 71 ]. These regions are implicated in cognitive functions, language processing, and working memory. Irregularities in the structure and function of these networks within these regions may influence selective attention, language processing, and working memory execution, potentially leading to impaired cognitive function [ 72 ]. Research has demonstrated that individuals with SZ display white matter abnormalities, with a specific focus on the temporal lobe (inferior longitudinal fasciculus) and cingulate regions [ 73 ]. In addition, patients with SZ display notable reductions in FA in various brain regions such as the left external capsule, right thalamic-prefrontal bundle, and thalamic-parietal bundle [ 74 ], corroborating the structural changes found in these areas, consistent with our findings in the frontal lobe and parietal lobe. Moreover, patients with SZ display markedly lower FA in bilateral frontal lobes, the right superior occipital gyrus, and the left anterior limb of the internal capsule [ 75 ], corroborating the structural changes in these regions, aligning with our findings in the frontal lobe and parietal lobe. Irregularities in the structure and function of these networks within these regions may contribute to impaired cognitive function, influencing selective attention, language processing, and working memory. These research findings enhance our comprehension of structural alterations in the brains of individuals with SZ, with implications for diagnosis, treatment, and prevention. Ultimately, through enrichment analysis, we pinpointed critical gene ontologies, including “postsynapse,” “axon,” and “presynapse.” These results are consistent with previous studies emphasizing the importance of synaptic pathology in SZ [ 76 ]. For example, patients with SZ exhibit dendritic spine density and glutamatergic postsynaptic abnormalities [ 77 ]. The changes in synaptic density are believed to be associated with the pathological changes in brain connectivity underlying SZ [ 78 ]. The degeneration and overgrowth of monoamine axons are associated with negative, cognitive, and positive symptoms of the disease, respectively [ 79 ]. Synaptic pathology and abnormalities in the neurotransmitter system may play important roles in the etiology of this disease. Therefore, understanding the potential mechanisms of synaptic pathology in SZ is crucial for designing new treatment strategies and interventions. The dataset used in this study did not collect genetic data from participants, there was no genetic data or MRI data available for the study of SZ. However, studies have already used genetic and MRI data for the classification of SZ, confirming the possibility of using genetic data for classification [ 80 ]. Integrating genetic data into our classification framework could significantly enhance accuracy and interpretability. Future iterations of our research could explore multi-modal approaches that leverage both genetic markers and imaging features, enriching our understanding of SZ’s neurobiological underpinnings. Conclusion and Limitations To explore the feasibility of using DTI data to classify patients with SZ and HC groups, we developed a graph attention mechanism-based model. Our research demonstrated that utilizing a dual-branch graph neural network model and extracting multimodal features from DTI data is superior to single-branch graph neural network models and other comparative experimental methods. Based on this dual-branch graph attention mechanism, we further localized key brain regions relevant to classification, such as the superior frontal gyrus, cingulate gyrus, middle frontal gyrus, thalamus, orbital gyrus, and anterior cuneiform lobe. We also identified key elements such as “postsynapse,” “axon,” and “presynapse.” These findings are consistent with the literature and contribute to the understanding of the neurobiological basis of SZ. The framework developed in this study, which integrates FA and FN brain networks alongside Graph neural network, enriches the landscape of SZ research and reveals new avenues for therapeutic interventions. By highlighting critical brain regions and genetic markers, our work deepens the understanding of the intricate neural mechanisms underlying SZ and offers potential implications for its diagnosis, treatment, and prevention. While significant strides have been made in identifying SZ, our study still faces certain challenges that warrant attention. Firstly, despite utilizing a relatively large sample size from the psychiatric cohort, our study is constrained by the number of deep-learning samples available. Moreover, conducting experiments across multiple locations with diverse scanning protocols may impact the model’s classification performance. Therefore, we recommend that future studies adopt standardized scanning protocols in larger cohorts, encompassing participants from various locations to enhance the robustness and generalizability of the classification models. The classification accuracy of our model may appear lower compared to some reported values, which can be attributed to the diverse nature of the dataset used in this study—data collected from multiple sites with varying protocols inherently introduces greater variability. To address this, future research could benefit from standardized data collection and preprocessing approaches, helping to mitigate the noise arising from data heterogeneity. In addition, incorporating other imaging modalities and optimizing the model’s architecture could further improve classification outcomes. Although the current study prioritized the FA network for gene expression analysis due to its superior performance and established role in assessing white matter integrity, we recognize that incorporating FN data could provide complementary insights. Future studies could explore the combined use of FA and FN to gain a deeper understanding of the genetic underpinnings of SZ, thus advancing research in this domain. Supplementary Information Below is the link to the electronic supplementary material. Supplementary file1 (PDF 199 KB) (198.5KB, pdf) Acknowledgments This work was supported by the National Natural Science Foundation of China (62276049, 61701078, 61872068, and 62006038), the Natural Science Foundation of Sichuan Province (2025ZNSFSC0487), the Science and Technology Innovation 2030 - Brain Science and Brain-Inspired Intelligence Project (2021ZD0200200), the National Key R&D Program of China (2023YFE0118600), Sichuan Province Science and Technology Support Program (2019YJ0193, 2021YFG0126, 2021YFG0366, and 2022YFS0180), and Medico-Engineering Cooperation Funds from the University of Electronic Science and Technology of China (ZYGX2021YGLH014). Conflict of interest The authors declare that there are no conflicts of interest. References 1. Fraguas D, Díaz-Caneja CM, Pina-Camacho L, Janssen J, Arango C. Progressive brain changes in children and adolescents with early-onset psychosis: A meta-analysis of longitudinal MRI studies. Schizophr Res 2016, 173: 132–139. [ DOI ] [ PubMed ] [ Google Scholar ] 2. Moylan S, Maes M, Wray NR, Berk M. The neuroprogressive nature of major depressive disorder: Pathways to disease evolution and resistance, and therapeutic implications. Mol Psychiatry 2013, 18: 595–606. [ DOI ] [ PubMed ] [ Google Scholar ] 3. Whiteford HA, Degenhardt L, Rehm J, Baxter AJ, Ferrari AJ, Erskine HE. Global burden of disease attributable to mental and substance use disorders: Findings from the global burden of disease study 2010. Lancet 2013, 382: 1575–1586. [ DOI ] [ PubMed ] [ Google Scholar ] 4. Ma S, Chen T, Jia W, Liu J, Ding S, Li P, et al . Enhanced Beta2-band oscillations denote auditory hallucination in schizophrenia patients and a monkey model of psychosis. Neurosci Bull 2024, 40: 325–338. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 5. Same K, Shobeiri P, Rashidi MM, Ghasemi E, Saeedi Moghaddam S, Mohammadi E, et al . A global, regional, and national burden and quality of care index for schizophrenia: Global burden of disease systematic analysis 1990–2019. Schizophr Bull 2024, 50: 1083–1093. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 6. Andleeb H, Moltrecht B, Gayer-Anderson C, Arango C, Arrojo M, D’Andrea G, et al . Age-at-migration, ethnicity and psychosis risk: Findings from the EU-GEI case-control study. PLoS Ment Health 2024, 1: e0000134. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 7. Domschke K. Prevention in psychiatry: A role for epigenetics? World Psychiatry 2021, 20: 227–228. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 8. Ecker C, Rocha-Rego V, Johnston P, Mourao-Miranda J, Marquand A, Daly EM, et al . Investigating the predictive value of whole-brain structural MR scans in autism: A pattern classification approach. Neuroimage 2010, 49: 44–56. [ DOI ] [ PubMed ] [ Google Scholar ] 9. Pettersson-Yeo W, Allen P, Benetti S, McGuire P, Mechelli A. Dysconnectivity in schizophrenia: Where are we now? Neurosci Biobehav Rev 2011, 35: 1110–1124. [ DOI ] [ PubMed ] [ Google Scholar ] 10. Talpalaru A, Bhagwat N, Devenyi GA, Lepage M, Mallar Chakravarty M. Identifying schizophrenia subgroups using clustering and supervised learning. Schizophr Res 2019, 214: 51–59. [ DOI ] [ PubMed ] [ Google Scholar ] 11. Xiao Y, Yan Z, Zhao Y, Tao B, Sun H, Li F, et al . Support vector machine-based classification of first episode drug-naïve schizophrenia patients and healthy controls using structural MRI. Schizophr Res 2019, 214: 11–17. [ DOI ] [ PubMed ] [ Google Scholar ] 12. Kambeitz J, Kambeitz-Ilankovic L, Leucht S, Wood S, Davatzikos C, Malchow B, et al . Detecting neuroimaging biomarkers for schizophrenia: A meta-analysis of multivariate pattern recognition studies. Neuropsychopharmacology 2015, 40: 1742–1751. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 13. Skåtun KC, Kaufmann T, Doan NT, Alnæs D, Córdova-Palomera A, Jönsson EG, et al . Consistent functional connectivity alterations in schizophrenia spectrum disorder: A multisite study. Schizophr Bull 2017, 43: 914–924. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 14. Cao B, Cho RY, Chen D, Xiu M, Wang L, Soares JC, et al . Treatment response prediction and individualized identification of first-episode drug-naïve schizophrenia using brain functional connectivity. Mol Psychiatry 2020, 25: 906–913. [ DOI ] [ PubMed ] [ Google Scholar ] 15. Liang S, Li Y, Zhang Z, Kong X, Wang Q, Deng W, et al . Classification of first-episode schizophrenia using multimodal brain features: A combined structural and diffusion imaging study. Schizophr Bull 2019, 45: 591–599. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 16. Cetin-Karayumak S, Di Biase MA, Chunga N, Reid B, Somes N, Lyall AE, et al . White matter abnormalities across the lifespan of schizophrenia: A harmonized multi-site diffusion MRI study. Mol Psychiatry 2020, 25: 3208–3219. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 17. Mikolas P, Hlinka J, Skoch A, Pitra Z, Frodl T, Spaniel F, et al . Machine learning classification of first-episode schizophrenia spectrum disorders and controls using whole brain white matter fractional anisotropy. BMC Psychiatry 2018, 18: 97. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 18. Ingalhalikar M, Kanterakis S, Gur R, Roberts TPL, Verma R. DTI based diagnostic prediction of a disease via pattern classification. Med Image Comput Comput Assist Interv 2010, 13: 558–565. [ DOI ] [ PubMed ] [ Google Scholar ] 19. Lei D, Pinaya WHL, Young J, van Amelsvoort T, Marcelis M, Donohoe G, et al . Integrating machining learning and multimodal neuroimaging to detect schizophrenia at the level of the individual. Hum Brain Mapp 2020, 41: 1119–1135. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 20. Wang J, Ke P, Zang J, Wu F, Wu K. Discriminative analysis of schizophrenia patients using topological properties of structural and functional brain networks: A multimodal magnetic resonance imaging study. Front Neurosci 2022, 15: 785595. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 21. Lu XB, Zhang Y, Yang DY, Yang YZ, Wu FC, Ning YP, et al . Analysis of first-episode and chronic schizophrenia using multi-modal magnetic resonance imaging. Eur Rev Med Pharmacol Sci 2018, 22: 6422–6435. [ DOI ] [ PubMed ] [ Google Scholar ] 22. Yang Y, Zhang Y, Wu F, Lu X, Ning Y, Huang B, et al . Automatic classification of first-episode, drug-naive schizophrenia with multi-modal magnetic resonance imaging. Sheng Wu Yi Xue Gong Cheng Xue Za Zhi 2017, 34: 674–680. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 23. Chen X, Zhou J, Ke P, Huang J, Xiong D, Huang Y, et al . Classification of schizophrenia patients using a graph convolutional network: A combined functional MRI and connectomics analysis. Biomed Signal Process Contr 2023, 80: 104293. [ Google Scholar ] 24. Gao J, Qian M, Wang Z, Li Y, Luo N, Xie S, et al . Exploring schizophrenia classification through multimodal MRI and deep graph neural networks: Unveiling brain region-specific weight discrepancies and their association with cell-type specific transcriptomic features. Schizophr Bull 2024, 51: 217–235. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 25. Lei D, Qin K, Pinaya WHL, Young J, Van Amelsvoort T, Marcelis M, et al . Graph convolutional networks reveal network-level functional dysconnectivity in schizophrenia. Schizophr Bull 2022, 48: 881–892. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 26. Ripke S, Walters J, O’Donovan M. Mapping genomic loci prioritises genes and implicates synaptic biology in schizophrenia. medRxiv 2016, 10.1101/2020.09.12.20192922. [ Google Scholar ] 27. Fromer M, Roussos P, Sieberts SK, Johnson JS, Kavanagh DH, Perumal TM, et al . Gene expression elucidates functional impact of polygenic risk for schizophrenia. Nat Neurosci 2016, 19: 1442–1453. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 28. Guan F, Ni T, Zhu W, Keoki Williams L, Cui LB, Li M, et al . Integrative omics of schizophrenia: From genetic determinants to clinical classification and risk prediction. Mol Psychiatry 2022, 27: 113–126. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 29. Cui Z, Zhong S, Xu P, He Y, Gong G. PANDA : A pipeline toolbox for analyzing brain diffusion images. Front Hum Neurosci 2013, 7: 42. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 30. Lin L, Jin C, Fu Z, Zhang B, Bin G, Wu S. Predicting healthy older adult’s brain age based on structural connectivity networks using artificial neural networks. Comput Methods Programs Biomed 2016, 125: 8–17. [ DOI ] [ PubMed ] [ Google Scholar ] 31. Fan L, Li H, Zhuo J, Zhang Y, Wang J, Chen L, et al . The human brainnetome atlas: A new brain atlas based on connectional architecture. Cereb Cortex 2016, 26: 3508–3526. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 32. Long Z, Duan X, Wang Y, Liu F, Zeng L, Zhao JP, et al . Disrupted structural connectivity network in treatment-naive depression. Prog Neuropsychopharmacol Biol Psychiatry 2015, 56: 18–26. [ DOI ] [ PubMed ] [ Google Scholar ] 33. Wang B, Fan Y, Lu M, Li S, Song Z, Peng X, et al . Brain anatomical networks in world class gymnasts: A DTI tractography study. Neuroimage 2013, 65: 476–487. [ DOI ] [ PubMed ] [ Google Scholar ] 34. Shu N, Liu Y, Li K, Duan Y, Wang J, Yu C, et al . Diffusion tensor tractography reveals disrupted topological efficiency in white matter structural networks in multiple sclerosis. Cereb Cortex 2011, 21: 2565–2577. [ DOI ] [ PubMed ] [ Google Scholar ] 35. Hawrylycz MJ, Lein ES, Guillozet-Bongaarts AL, Shen EH, Ng L, Miller JA, et al . An anatomically comprehensive atlas of the adult human brain transcriptome. Nature 2012, 489: 391–399. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 36. Hong Y, Luo P, Jin S, Liu X. LaGAT: Link-aware graph attention network for drug-drug interaction prediction. Bioinformatics 2022, 38: 5406–5412. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 37. Hu J, Cao L, Li T, Dong S, Li P. GAT-LI: A graph attention network based learning and interpreting method for functional brain network classification. BMC Bioinformatics 2021, 22: 379. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 38. Wang Q, Guo Y, Shen Y, Tong S, Guo H. Multi-layer graph attention network for sleep stage classification based on EEG. Sensors (Basel) 2022, 22: 9272. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 39. Safai A, Vakharia N, Prasad S, Saini J, Shah A, Lenka A, et al . Multimodal brain connectomics-based prediction of Parkinson’s disease using graph attention networks. Front Neurosci 2022, 15: 741489. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 40. Veličković P, Cucurull G, Casanova A, Romero A, Liò P, Bengio Y. Graph attention networks. 2017: 1710.10903. https://arxiv.org/abs/1710.10903v3 . 41. Ba JL, Kiros JR, Hinton GE. Layer normalization. 2016: 1607.06450. https://arxiv.org/abs/1607.06450v1 . 42. Zhang C, Bengio S, Hardt M, Recht B, Vinyals O. Understanding deep learning (still) requires rethinking generalization. Commun ACM 2021, 64: 107–115. [ Google Scholar ] 43. Cangea C, Velikovi P, Jovanovi N, Kipf T, Liò P, Towards sparse hierarchical graph classifiers. arXiv.org 2018, 10.48550/arXiv.1811.01287. 44. Knyazev B, Taylor GW, Amer MR, Understanding attention and generalization in graph neural networks. 2019: 4202–4212. 45. McColgan P, Gregory S, Seunarine KK, Razi A, Papoutsi M, Johnson E, et al . Brain regions showing white matter loss in Huntington’s disease are enriched for synaptic and metabolic genes. Biol Psychiatry 2018, 83: 456–465. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 46. Romero-Garcia R, Warrier V, Bullmore ET, Baron-Cohen S, Bethlehem RAI. Synaptic and transcriptionally downregulated genes are associated with cortical thickness differences in autism. Mol Psychiatry 2019, 24: 1053–1064. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 47. Vértes PE, Rittman T, Whitaker KJ, Romero-Garcia R, Váša F, Kitzbichler MG, et al . Gene transcription profiles associated with inter-modular hubs and connection distance in human functional magnetic resonance imaging networks. Philos Trans R Soc Lond B Biol Sci 2016, 371: 20150362. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 48. Romero-Garcia R, Seidlitz J, Whitaker KJ, Morgan SE, Fonagy P, Dolan RJ, et al . Schizotypy-related magnetization of cortex in healthy adolescence is colocated with expression of schizophrenia-related genes. Biol Psychiatry 2020, 88: 248–259. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 49. Breiman L. Random forests. Mach Learn 2001, 45: 5–32. [ Google Scholar ] 50. Cortes C, Vapnik V. Support-vector networks. Mach Learn 1995, 20: 273–297. [ Google Scholar ] 51. Chen T, Guestrin C, XGBoost: A scalable tree boosting system. In: Proc 22nd ACM SIGKDD Int Conf Knowl Discov Data Min, 2023. 52. He K, Zhang X, Ren S, Sun J, Deep residual learning for image recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). June 27-30, 2016, Las Vegas, NV, USA. IEEE, 2016: 770–778. 53. Kipf TN, Welling M. Semi-supervised classification with graph convolutional networks. 2016: 1609.02907. https://arxiv.org/abs/1609.02907v4 . 54. Fortin JP, Cullen N, Sheline YI, Taylor WD, Aselcioglu I, Cook PA, et al . Harmonization of cortical thickness measurements across scanners and sites. NeuroImage 2018, 167: 104–120. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 55. Ehrlich S, Brauns S, Yendiki A, Ho BC, Calhoun V, Charles Schulz S, et al . Associations of cortical thickness and cognition in patients with schizophrenia and healthy controls. Schizophr Bull 2012, 38: 1050–1062. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 56. Mubarik A, Tohid H. Frontal lobe alterations in schizophrenia: A review. Trends Psychiatry Psychother 2016, 38: 198–206. [ DOI ] [ PubMed ] [ Google Scholar ] 57. Thielen JW, Müller BW, Chang DI, Krug A, Mehl S, Rapp A, et al . Cortical thickness across the cingulate gyrus in schizophrenia and its association to illness duration and memory performance. Eur Arch Psychiatry Clin Neurosci 2022, 272: 1241–1251. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 58. Wang HS, Rau CL, Li YM, Chen YP, Yu R. Disrupted thalamic resting-state functional networks in schizophrenia. Front Behav Neurosci 2015, 9: 45. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 59. Li J, Seidlitz J, Suckling J, Fan F, Ji GJ, Meng Y, et al . Cortical structural differences in major depressive disorder correlate with cell type-specific transcriptional signatures. Nat Commun 2021, 12: 1647. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 60. Zhou Y, Zhou B, Pache L, Chang M, Khodabakhshi AH, Tanaseichuk O, et al . Metascape provides a biologist-oriented resource for the analysis of systems-level datasets. Nat Commun 2019, 10: 1523. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 61. Tamnes CK, Agartz I. White matter microstructure in early-onset schizophrenia: A systematic review of diffusion tensor imaging studies. J Am Acad Child Adolesc Psychiatry 2016, 55: 269–279. [ DOI ] [ PubMed ] [ Google Scholar ] 62. Ye H, Zalesky A, Lv J, Loi SM, Cetin-Karayumak S, Rathi Y, et al . Network analysis of symptom comorbidity in schizophrenia: Relationship to illness course and brain white matter microstructure. Schizophr Bull 2021, 47: 1156–1167. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 63. Joo SW, Chon MW, Rathi Y, Shenton ME, Kubicki M, Lee J. Abnormal asymmetry of white matter tracts between ventral posterior cingulate cortex and middle temporal gyrus in recent-onset schizophrenia. Schizophr Res 2018, 192: 159–166. [ DOI ] [ PubMed ] [ Google Scholar ] 64. Das TK, Kumar J, Francis S, Liddle PF, Palaniyappan L. Parietal lobe and disorganisation syndrome in schizophrenia and psychotic bipolar disorder: A bimodal connectivity study. Psychiatry Res Neuroimaging 2020, 303: 111139. [ DOI ] [ PubMed ] [ Google Scholar ] 65. de Filippis R, Carbone EA, Gaetano R, Bruni A, Pugliese V, Segura-Garcia C, et al . Machine learning techniques in a structural and functional MRI diagnostic approach in schizophrenia: A systematic review. Neuropsychiatr Dis Treat 2019, 15: 1605–1627. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 66. Wang S, Zhan Y, Zhang Y, Lyu L, Lyu H, Wang G, et al . Abnormal long- and short-range functional connectivity in adolescent-onset schizophrenia patients: A resting-state fMRI study. Prog Neuropsychopharmacol Biol Psychiatry 2018, 81: 445–451. [ DOI ] [ PubMed ] [ Google Scholar ] 67. Tamminga CA, Vogel M, Gao X, Lahti AC, Holcomb HH. The limbic cortex in schizophrenia: Focus on the anterior cingulate. Brain Res Brain Res Rev 2000, 31: 364–370. [ DOI ] [ PubMed ] [ Google Scholar ] 68. Ou J, Lyu H, Hu M, Li J, Guo W, Guo X, et al . Decreased white matter FA values in the left inferior frontal gyrus is a possible intermediate phenotype of schizophrenia: Evidences from a novel group strategy. Eur Arch Psychiatry Clin Neurosci 2018, 268: 89–98. [ DOI ] [ PubMed ] [ Google Scholar ] 69. Zeng J, Zhang W, Xiao Y, Fu G, Liu L, Tang B, et al . Pretreatment abnormalities in white matter integrity predict one-year clinical outcome in first episode schizophrenia. Schizophr Res 2021, 228: 241–248. [ DOI ] [ PubMed ] [ Google Scholar ] 70. Kubicki M, Westin CF, Nestor PG, Wible CG, Frumin M, Maier SE, et al . Cingulate Fasciculus integrity disruption in schizophrenia: A magnetic resonance diffusion tensor imaging study. Biol Psychiatry 2003, 54: 1171–1180. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 71. Goldman-Rakic PS. Working memory dysfunction in schizophrenia. J Neuropsychiatry Clin Neurosci 1994, 6: 348–357. [ DOI ] [ PubMed ] [ Google Scholar ] 72. Honey GD, Bullmore ET, Sharma T. De-coupling of cognitive performance and cerebral functional response during working memory in schizophrenia. Schizophr Res 2002, 53: 45–56. [ DOI ] [ PubMed ] [ Google Scholar ] 73. Lener MS, Wong E, Tang CY, Byne W, Goldstein KE, Blair NJ, et al . White matter abnormalities in schizophrenia and schizotypal personality disorder. Schizophr Bull 2015, 41: 300–310. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 74. Joo SW, Kim H, Jo YT, Yoon W, Kim Y, Lee J. Shared and distinct white matter abnormalities in schizophrenia and bipolar disorder. Prog Neuropsychopharmacol Biol Psychiatry 2021, 108: 110175. [ DOI ] [ PubMed ] [ Google Scholar ] 75. Grazioplene RG, Bearden CE, Subotnik KL, Ventura J, Haut K, Nuechterlein KH, et al . Connectivity-enhanced diffusion analysis reveals white matter density disruptions in first episode and chronic schizophrenia. Neuroimage Clin 2018, 18: 608–616. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 76. Chelini G, Pantazopoulos H, Durning P, Berretta S. The tetrapartite synapse: A key concept in the pathophysiology of schizophrenia. Eur Psychiatry 2018, 50: 60–69. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 77. Matas E, William DJ, Toro CT. Abnormal expression of post-synaptic proteins in prefrontal cortex of patients with schizophrenia. Neurosci Lett 2021, 745: 135629. [ DOI ] [ PubMed ] [ Google Scholar ] 78. Berdenis van Berlekom A, Muflihah CH, Snijders GJ, MacGillavry HD, Middeldorp J, Hol EM, Kahn RS, De Witte LD. Synapse pathology in schizophrenia: A meta-analysis of postsynaptic elements in postmortem brain studies. Schizophr Bull 2020, 46(2): 374–386. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 79. Nakamura S. Integrated pathophysiology of schizophrenia, major depression, and bipolar disorder as monoamine axon disorder. Front Biosci (Schol Ed) 2022, 14: 4. [ DOI ] [ PubMed ] [ Google Scholar ] 80. Shaik NS, Cherukuri TK, Calhoun VD, Ye DH. Multi-modal imaging genomics transformer: Attentive integration of imaging with genomic biomarkers for schizophrenia classification. 2024: 2407.19385. Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials Supplementary file1 (PDF 199 KB) (198.5KB, pdf) Articles from Neuroscience Bulletin are provided here courtesy of Springer ACTIONS View on publisher site PDF (1.7 MB) Cite Collections Permalink RESOURCES Similar articles Cited by other articles Links to NCBI Databases Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top

Related documents

• Exploration of schizophrenia-associated gene modules using graph theory, co-expression networks, and dimensionality reduction#13950
• Machine-learning analysis of temporal molecular dynamics stratifies autism likelihood - a multinational study#13908
• Bioinformed idiographic symptom networks to identify neurobiological predictors of affective-state transitions in bipolar disorder.#430639
• A generative explainable model for antimicrobial peptide prediction using bidirectional temporal convolutional neural network.#148394
• GRIA Gene Expression in Schizophrenia: A Participant-Level Meta-Analysis.#351632