Generation of protein dynamics by machine learning

Abstract

Machine learning has advanced protein structure prediction to deliver accurate but mostly static models. Capturing protein dynamics as conformational ensembles remains a significant challenge. Recent developments, especially generative models, are enabling the prediction of structural ensembles beyond traditional simulations. This review examines emerging machine learning approaches for modeling protein dynamics, in terms of generating PDB-like ensembles, accelerating molecular simulations, modeling non-globular protein ensembles, and integrating experimental data. General-purpose and system-specific models are discussed, particularly in terms of conformational coverage, transferability, and responsiveness to environmental conditions. Hybrid models, which combine experimental and simulation data, represent a promising direction. Nonetheless, key challenges remain, including generating states with correct probabilities, modeling unseen conformations, and integrating experimental constraints rigorously.