The landscape of machine learning approaches for modeling protein conformational ensembles

The conformational ensemble of a protein and its corresponding probabilities and dynamics are crucial determinants of its function, but are difficult to access with traditional experimental and computational technologies. This review examines the landscape of machine learning for modeling protein conformational ensembles. We categorize computational methods into three classes: AlphaFold-based approaches that modify the input multiple sequence alignment, score-based generative models that use diffusion or flow-matching algorithms, and protein language models that link sequence evolution with sequential dynamics. We discuss the data available for training and benchmarking, including molecular dynamics simulations and experimental repositories. We highlight current limitations in the field, including the lack of standardization in benchmarking and the high variability of mechanisms and environmental conditions that challenge current methods. Drawing lessons from the success of AlphaFold, we identify key opportunities for further improvement, including accurate modeling of kinetics and thermodynamics, and linking model uncertainty with targeted collection of new data.