Learning Brain Structure-Function Interactions via Modularity-Guided Attention-based Graph Neural Networks

Jing Xia, Yi Hao Chan, Deepank Girish, Conghao Wang, Jagath C. Rajapakse

Abstract: Incorporating brain structure-function interactions has been shown to enhance brain disorder diagnosis and cognitive prediction. However, current studies primarily focus on one-to-one interactions between corresponding regions of interest (ROIs), overlooking more complex interactions across different ROIs. Since ROIs within the same module exhibit more similar neural mechanisms of structure-function coupling compared to those in different modules, modular-level homogeneity provides stronger biological guidance for learning complex interactions. Though leveraging this guidance can further improve prediction performance, such studies remain scarce. To address this gap, we propose a modularity-guided attention-based graph neural network that extracts and emphasizes key structure-function interactions across ROIs relevant to specific tasks. Specifically, a modularity-guided interactive network is designed to define modularity-specific convolution operations learning interactions across structural and functional ROIs based on modular homogeneity. A modular interactive edge attention model is designed to highlight the most salient interactions by using a cross-attention mechanism. These identified interactions serve as edges connecting structural and functional ROIs, forming a unified graph. A graph transformer-based approach leverages this graph for prediction. Experimental results on three datasets demonstrate our model significantly outperforms nine state-of-the-art multi-modal methods in predicting fluid cognition, as well as in classifying Parkinson’s disease and schizophrenia. Attention maps show that the visual function engaging with default mode and frontoparietal structures is key to predicting fluid cognition, while diverse functional modules interacting with the subcortical structures are linked to Parkinson’s disease. Subcortical and visual-related structures interacting with the functional modules related to sensory and executive function are associated with schizophrenia.