Integrating multi-order spectral filtering with structural subgraph generation for predicting functional connectivity in brain networks

Understanding how brain structural connectivity (SC) shapes functional connectivity (FC) is essential for revealing the organizational principles of brain information processing. However, existing model-based approaches often struggle to characterize nonlinear SC–FC associations, while conventional graph neural networks remain limited in capturing the multiscale and heterogeneous organization of brain networks. To address these limitations, we propose Cheb-GINet, a Chebyshev graph isomorphism network for subject-level SC-to-FC prediction. First, each SC matrix is converted into an initial brain graph, from which a subgraph is then obtained through an adaptive subgraph generator using node features and anatomical connections to estimate edge-wise retention probabilities between brain regions. The connection strength of each retained edge in the subgraph is retrieved from the original subject-level SC matrix. Next, the generated subgraph is fed into Cheb-GINet, in which the selected SC edge strengths are scaled and normalized to build the spectral operators for multi-order Chebyshev graph convolutions, while the unprocessed edge values are retained to modulate GIN-based neighbor aggregation and the multi-head attention module. The resulting node-level features are then pooled into a subject-level graph embedding, which summarizes the whole-brain SC-derived representation and is then decoded to reconstruct the symmetric FC matrix. Experiments on three brain connectome datasets with increasing resolutions demonstrate that Cheb-GINet consistently outperforms state-of-the-art methods. These results suggest that our proposed architecture provides an effective and structurally grounded framework for modeling nonlinear SC–FC relationships, offering a step toward a deeper mechanistic understanding of brain network organization.