Informational and methodological differences in regional structure-function coupling in modeling approaches

Structural connectivity in the human brain constrains functional connectivity, which in turn reshapes structural connectivity, resulting in complex coupling relationships. Numerous approaches exist for quantifying structure-function coupling, and different approaches exhibit systematic differences in their estimations, but the sources and specific manifestations of these differences remain unclear. In this study, starting from the perspective of indirect connection information and using the correlational approach as a baseline, we systematically compared the differences among four modeling approaches (multivariate linear regression, multilayer perceptron, predictive graph neural network and self-supervised graph neural network) in the calculation of regional structure-function coupling by defining the informational and methodological differences. Our results show that indirect connections have a small effect on the regression approach and the multilayer perceptron, while a larger effect on the graph neural networks. Furthermore, we observed that, in predicting functional connectivity, the prediction of direct connections of brain regions with higher T1w/T2w myelination indices is more significantly influenced by indirect connections. At the level of structure-function coupling, the right-hemispheric dorsal attention network is least affected by indirect connections, while the orbito-affective network is most significantly affected. These findings reveal the mechanisms of information usage behind different approaches in the calculation of structure-function coupling, providing a basis for future research to select appropriate approaches based on the structural and functional network characteristics of the brain.