(203i) A Flow-Matching Approach for Predicting Protein Transition Pathways

Understanding protein conformational dynamics is crucial for drug design, as static structures alone fail to capture the full range of biologically relevant motions. Traditional molecular dynamics (MD) simulations, while effective, are computationally expensive due to the vast timescales required to sample rare but functionally important transitions. Enhanced sampling methods, such as biased and adaptive sampling, attempt to accelerate these transitions but suffer from limitations like hysteresis or inefficient exploration. To overcome these challenges, we propose a generative AI-based framework that leverages flow matching on Riemannian manifolds to predict transition pathways between metastable protein states. Our method formulates protein conformations as points on a curved manifold and uses flow matching to learn optimal velocity fields that drive conformational changes along geodesic paths. This approach enables accurate, physically meaningful intermediate structure generation while significantly reducing computational cost. We demonstrated this technique on adenylate kinase, a protein with well-characterized large-scale conformational changes. Our results show that the predicted transition pathways closely align with those obtained from long-timescale MD simulations, validating the method's ability to capture realistic biomolecular motions without requiring exhaustive sampling.