Chimera: Optimizing Multi-Modal, Multi-Task GNN Architectures for Precision Neuroscience

Abstract

Abstract. Deep learning offers great potential for analyzing Magnetic Resonance Imaging (MRI) data to advance our understanding of neurological disorders and guide intervention strategies. Recent work has incorporated graph neural networks to capture connectivity from structural and functional MRI; however, we identify three unresolved challenges: (1) multi-modal integration, (2) multi-task classification, and (3) underutilization of the rich architectural design space of graph operations. To address these gaps, we introduce Chimera - a hybrid N-branch GNN framework that combines classical encoder modules with graph operations for multi-modal, multi-task prediction. In our proof-of-concept, we deploy a three-leg, two-head architecture that merges resting-state functional MRI, MRI–derived morphology, and graph metrics into a unified latent representation. After cross-modal message-passing, the learned representations diverge into an age-regression and sex-classification head. When evaluated on 33,460 UK Biobank participants, Chimera achieves 92.2% accuracy in sex classification and a mean absolute error (MAE) of 4.21 years in age prediction. While some approaches with finer parcellations or temporal graphs achieve higher performance, Chimera offers a simpler, more versatile foundation for future development.