Omid Mokhtari
Inria, Loria, CNRS, Universite de Lorraine, Nancy
Proteins are dynamic entities, and their conformational flexibility, particularly in intrinsically disordered regions (IDRs), plays a crucial role in their function. While significant progress has been made in predicting protein-protein interfaces, existing methods often overlook conformational heterogeneity. We employ state-of-the-art geometric deep learning architectures, integrating dynamic features into protein binding site prediction. Our model leverages cooperative graph neural networks (GNNs) to optimize message passing between core and surface residues, incorporating conformational ensembles from molecular dynamics (MD) simulations, NMR, and AlphaFlow. Results demonstrate that integrating dynamic features significantly improves binding site prediction accuracy and ensures consistency across diverse benchmarks.
