DynamicGT: a dynamic-aware geometric transformer model to predict protein binding interfaces in flexible and disordered regions

Abstract

Protein-protein interactions are fundamental to cellular processes, yet existing deep learning approaches for binding site prediction often rely on static structures, limiting their performance when disordered or flexible regions are involved. To address this, we introduce a novel dynamic-aware method for predicting protein-protein binding sites by integrating conformational dynamics into a cooperative graph neural network (Co-GNN) architecture with a geometric transformer (GT). Our approach uniquely encodes dynamic features at both the node (atom) and edge (interaction) levels, and consider both bound and unbound states to enhance model generalization. The dynamic regulation of message passing between core and surface residues optimizes the identification of critical interactions for efficient information transfer. We trained our model on an extensive overall 1-ms molecular dynamics simulations dataset across multiple benchmarks as the gold standard and further extended it by adding generated conformations by AlphaFlow. Comprehensive evaluation on diverse independent datasets containing disordered, transient, and unbound structures showed that incorporating dynamic features in cooperative architecture significantly boosts prediction accuracy when flexibility matters, and requires substantially less amount of data than leading static models.