MACHINE LEARNING FOR ANALYSIS OF MD SIMULATIONS AND FOR GENERATING NEW ONES

Protein structure prediction of single stable states has advanced rapidly since the emergence of AlphaFold. However, current models still struggle to capture the conformational heterogeneity that underlies ligand binding, allostery, and molecular recognition. I will describe recent efforts to evaluate and extend AI approaches for modeling protein dynamics. We find that leading structure predictors can recover some open, closed, apo-like, and reduced-fit conformations, but that ligand-sensitive states are not always sampled without additional physical or binding-partner context. To address these limitations, we developed DeepPath, a physics-guided generative framework that uses learned conformational fields as oracles to guide pathway generation between protein states. Unlike conventional supervised approaches, DeepPath combines AI-driven inference with physical models, allowing predicted motions to be tested and refined through simulation. I will also discuss ongoing efforts to generate and organize protein-dynamics data at scale, so that simulations can better guide the next iteration of trained models. This work highlights the potential of physics-AI integration for modeling protein dynamics and improving structure-based drug discovery.