Abstract
Major depressive disorder (MDD) and schizophrenia (SZ) are among the most debilitating psychiatric disorders, characterized by widespread disruptions in large-scale brain networks. However, the commonalities and distinctions in their large-scale network distributions remain unclear. The present study aimed to leverage advanced deep learning techniques to identify these common and distinct patterns, providing insights into the shared and disorder-specific neural mechanisms underlying MDD and SZ. Recent advances in graph neural networks (GNNs) offer a powerful framework for analyzing brain connectivity patterns, enabling automated learning of complex, high-dimensional network features. In this study, we applied state-of-art GNN architectures to classify MDD and SZ patients from healthy controls (HCs), respectively, using a multisite resting-state fMRI dataset. The attention-based hierarchical pooling GNN (SAGPool) model achieved the highest performance, with mean accuracies of 71.50% for MDD and 75.65% for SZ classification. Using a perturbation-based explainability method, we identified prominent functional connections driving model decisions, revealing distinct patterns of the large-scale network disruption across disorders. In MDD, alterations were dominantly observed in the default mode network (DMN), whereas SZ exhibited prominent alterations in the ventral attention network (VAN). Notably, specific functional connections identified by our model showed significant correlations with clinical symptoms, particularly positive and general symptoms measured by the positive and negative syndrome scale (PANSS) in SZ patients. Our findings demonstrate the utility of GNNs for uncovering complex connectivity patterns in psychiatric disorders and provide novel insights into the distinct neural mechanisms underlying MDD and SZ. These results highlight the potential of graph-based models as tools for both diagnostic classification and biomarker discovery in psychiatric research.