Brain Topogr. 2026 Jul 23;39(5):83. doi: 10.1007/s10548-026-01237-z.
ABSTRACT
Understanding how the brains structural architecture supports task-evoked functional activity is a fundamental goal of neuroscience. However, most predictive models rely on single-scale structural features, overlooking the brains intrinsically hierarchical organization. Here, we introduce the Collaborative Graph Attention Multi-Task Network (CoGA-MTN), a deep learning framework that integrates multi-scale structural features to jointly predict task-based functional connectivity (FC) and identify potential disease-related network alterations. CoGA-MTN employs a dual-branch graph attention network to extract complementary global statistical and local topological features from structural MRI, and a cross-modal task-coordinated learning mechanism that enables task-conditioned FC prediction alongside multi-disease classification. Validated on the Consortium for Neuropsychiatric Phenomics dataset (152 participants, three tasks, four diagnostic groups), CoGA-MTN outperforms single-scale baselines in task-conditioned FC prediction (PCC = 0.657 ± 0.02) and achieves macro F1 scores of 0.68-0.75 across three psychiatric disorders. Crucially, the model reconstructs a stable whole-brain connectivity architecture that is conserved across tasks, while simultaneously revealing diagnosis-related discrepancies that are consistent with established pathophysiological models. By modeling psychiatric conditions within a normative structure-function framework, this work provides a unified approach for characterizing the relationships between multi-scale brain structure, dynamic function, and psychopathology.
PMID:42489954 | DOI:10.1007/s10548-026-01237-z
