Abstract
Understanding protein structure and dynamics is crucial for basic biology and drug design. Conventional methods often provide static conformations that inadequately capture protein flexibility. We present PackDock, a framework that integrates deep learning and physics-based modeling to represent protein-ligand interactions. PackDock's core, PackPocket, uses diffusion models to sample diverse binding pocket conformations and predict ligand-induced changes. We validate PackDock through side-chain packing, redocking, and cross-docking experiments, demonstrating its ability to address protein flexibility challenges. In a real-world application, PackDock identified nanomolar affinity compounds with unreported scaffolds for the protein of interest. Additionally, it revealed key amino acid conformational changes, offering insights into protein-ligand interactions. By accurately predicting complex conformations in various scenarios, PackDock enhances our understanding of protein dynamics and provides perspectives for both basic biological research and drug discovery efforts.