Edge-Aware Graph Attention Networks for Interpreting Biophysical Mechanisms from Molecular Dynamics Simulations

Graph attention networks (GATs) are emerging as powerful aritficial intelligence (AI) tools for learning biomolecular dynamics, yet extracting mechanistic insight from learned attention remains challenging. Here, we present an interpretable AI approach that combines molecular dynamics prediction with mechanistic interpretation through attention-derived communication networks. We develop three edge-aware GAT models - Edge-Conditioned, Edge-Injected, and Edge-Gated - that differ in how edge information is incorporated during message passing. Relative to a standard GAT, our edge-aware models improve coordinate prediction while recovering complementary aspects of residue communication. Application to Chignolin and HIV-1 protease demonstrates their ability to characterize communication networks underlying protein folding, allostery, and mutation-induced functional remodeling. Analysis of the communication networks using graph-based descriptors of communication throughput, intensity, and relay revealed that the Edge-Conditioned model preferentially emphasizes communication hubs characterized by high communication throughput and intensity, the Edge-Injected model preferentially highlights high-intensity communication within functionally important regions, and the Edge-Gated model most clearly resolves long-range communication relay. Overall, this approach provides an interpretable AI strategy for uncovering the communication mechanisms that underlie folding, allostery, and long-range signal propagation in molecular machines, while also supporting AI-guided modulation of biomolecular function and rational protein engineering.