Predicting Human Intrinsic Functional Connectivity From Structural Connectivity: An Artificial Neural Network Approach

How structural connectivity (SC) constrains and shapes functional connectivity (FC) in the human brain to support rich cognitive functions has long been a core issue in neuroscience. Although evidence accumulate to suggest that FC strength is correlated with multiple aspects of SC, few studies has analyzed the SC-to-FC relationship in a multivariate manner. This paper proposed a novel usage of the feedforward neural network to predict FC strength as a nonlinear combination of 115 features that described the geometric and topological aspects of SC. The resulting model outperformed four state-of-the-art models in both terms of predictive power and generalizability. Model interpretation analyses found that the geometric features were generally more predictive than the topological ones, providing novel evidence for the significant impact of geometric relationships on FC generation. Comparison of feature contributions to predicting FC with different structural properties further revealed the crucial role of indirect structural paths for inducing FC, particularly between disconnected and/or distanced regions. Together, our results suggested that the flexible FC is significantly but unevenly influenced by the combination of geometric and topological characteristics of the structural network. The proposed framework would also be used for link prediction on top of an underlying topology.

Automated summary

The study proposes a novel approach to predict functional connectivity strength in the human brain based on geometric and topological features of structural connectivity, showing that geometric features are more predictive than topological features. Indirect structural paths play a crucial role in inducing functional connectivity, especially between disconnected and/or distanced regions.