Hierarchical multi-atlas fusion network for robust and interpretable cross-site analysis of functional connectomes in major depressive disorder.

Major depressive disorder (MDD) is associated with distributed disruptions in functional brain networks, yet current connectome-based biomarkers often rely on a single parcellation and show limited cross-site robustness and biological interpretability. We propose a Hierarchical Multi-Atlas Fusion Network (HMAF-Net) for robust and interpretable cross-site analysis of functional connectomes in MDD. HMAF-Net reconciles heterogeneous anatomical and functional atlases by learning a small set of shared fusion nodes that serve as latent anchors for cross-atlas semantic alignment. For each atlas, it extracts node-wise descriptors and uses Transformer-based sparse soft assignment with a structure-aware contrastive objective to align atlas nodes in a common semantic space while retaining atlas-specific information. The resulting fusion-node embeddings support subject-level classification and spatial mapping for biomarker localization and imaging-transcriptomic association. We evaluated HMAF-Net on the multi-site REST-meta-MDD cohort (N = 1,611). Under 10-fold cross-validation, HMAF-Net achieved 72.28% accuracy and 71.49% AUC, showing the highest mean accuracy among representative atlas-fusion and graph-based baselines under identical splits. Under leave-one-site-out evaluation, it reached 77.48% mean accuracy, with site-wise analyses indicating residual variability associated with cohort size and head motion. External validation on two independent MDD datasets further supported cross-dataset generalizability. Fusion-node involvement shifts revealed reproducible fronto-limbic and default-mode-salience alterations. These spatial profiles were associated with Allen Human Brain Atlas expression gradients and enriched for synaptic and neurotransmission pathways, providing a molecularly grounded axis for interpretation. By integrating cross-atlas alignment, cross-site evaluation, and transcriptomic interpretation, HMAF-Net yields compact connectomic embeddings for reproducible and mechanistically informed MDD biomarker development.