Prediction of brain functions from structural connectome using graph neural network

Abstract

AO_SCPLOWBSTRACTC_SCPLOWThe relationship between brain structure and function remains elusive, amidst the tremendous advances in brain mapping techniques. In this work, we attempt to partially disentangle this relationship by connecting task-evoked functional MRI (fMRI) responses with the underlying structural connectome using graph neural network (GNN). MRI data (n = 1,063) were collected from the Human Connectome Project. We demonstrate that our GNN-based model predicts task-evoked fMRI responses with high fidelity. Using a graph attention mechanism, it is possible to infer the subsets of neighboring cortical regions whose structural connections are important for the prediction of the functional responses of individual cortical regions. Notably, for each cortical region, such subset of neighboring cortical regions is predominantly localized to the ipsilateral hemisphere and much smaller than that with direct structural connections. We found that the higher cognitive functions subserved by the cingulo-opercular, dorsal attention, frontoparietal and default mode clusters may depend on neighboring cortical regions across a wide range of functional brain clusters in the ipsilateral hemisphere, whilst the sensory functions subserved by the visual1 and auditory clusters on neighboring cortical regions across much fewer functional brain clusters.