An Evaluation of Biomolecular Energetics Learned by AlphaFold

Abstract

Deep learning has revolutionized protein structure prediction, with function prediction on the horizon1,2. Biomolecular properties, including structure and all aspects of function, emerge from atomic-level interactions and the probabilities of their formation3,4. Learning the physical rules that govern these probabilities allows models to deliver accurate structure predictions and may enable extrapolation beyond the training data--a capability needed to predict the many biologically important functional properties where comprehensive data are not readily attainable5. Current structure-based models follow a training logic primarily focused on matching atomic coordinates, rather than atomic interactions and their probabilities. It remains unknown whether the models have learned the physical rules that underlie atomic interactions, the extent of their knowledge, and the prediction errors that arise from limits to this knowledge. We found that state-of-the-art structure prediction models, AlphaFold2, AlphaFold3, and ESMFold, capture basic energetic principles but show pervasive biases in the conformational preferences of molecular interactions. These biases manifest as widespread prediction errors, including the misassignment of a large fraction of side-chain non-covalent interactions--[~]30% for the AlphaFold models and [~]60% for ESMFold--and an inability to reproduce experimentally derived conformational ensembles. More than half the errors made by AlphaFold2 and AlphaFold3 are in common, suggesting limitations not overcome by using different model architectures. Overall, our multifaceted, physics-grounded evaluation identified previously unknown, system-wide deficiencies in current structure prediction models. This framework is applicable to and needed for all biomolecular structure and function prediction models that deliver atomic-level structural information. The insights derived from these evaluations will allow researchers to judiciously apply current models and will guide the development of next-generation models to achieve accurate prediction of biomolecular function.