Fold-switching proteins push the boundaries of conformational ensemble prediction.

A protein's function depends critically on its conformational ensemble, a collection of energy weighted structures whose balance depends on temperature and environment. Though recent deep learning (DL) methods have substantially advanced predictions of single protein structures, computationally modeling conformational ensembles remains a challenge. Here, we focus on modeling fold-switching proteins, which remodel their secondary and/or tertiary structures and change their functions in response to cellular stimuli. These underrepresented members of the protein universe serve as test cases for a method's generalizability. They reveal that DL models often predict conformational ensembles by association with training-set structures, limiting generalizability. These observations suggest use cases for when DL methods will likely succeed or fail. Developing computational methods that successfully identify new fold-switching proteins from large pools of candidates may advance modeling conformational ensembles more broadly.