Disentangling coevolutionary constraints for modeling protein conformational heterogeneity.

Accurate characterization of multi-state protein conformations is crucial for understanding their functional mechanisms and advancing targeted therapies. Extracting coevolutionary constraints from homologous sequences helps reveal protein structure and function, which can be automatically captured by MSA Transformer leveraging attention mechanisms. Making use of the multi-conformational coevolutionary signals captured by MSA Transformer, we introduce in this study EvoSplit to disentangle coevolutionary signals associated with distinct conformations to guide protein structure predictions. EvoSplit outperforms AF-Cluster on 85 fold-switching proteins and successfully models the conformations of proteins beyond AlphaFold2's training set. We then identify 54 candidates with potential conformational diversity for cancer-related human proteins. Notably, for five GTPases, EvoSplit consistently predicts two conformations, one of which has not been previously reported. As an important example, the protein-protein interaction analysis provides new insights into novel HRAS function-associated conformations. Furthermore, the validity of these newly identified conformations is examined by evolutionary analysis and extensive molecular dynamics simulations.