Large-scale predictions of alternative protein conformations by AlphaFold2-based sequence association

The many successes of AlphaFold2 (AF2) have inspired methods to predict multiple protein conformations, many of which have biological significance. These methods often assume that AF2 relies on evolutionary couplings to predict alternative protein conformations, but they perform poorly on fold-switching proteins, which remodel their secondary structures and modulate their functions in response to cellular stimuli. Here we present a method designed to leverage AF2’s learning of protein structure more than evolutionary couplings. This method–called CF-random–outperforms other methods for predicting alternative conformations of not only fold switchers but also dozens of other proteins that undergo rigid body motions and local conformational rearrangements. It also enables predictions of fold-switched assemblies unpredicted by AlphaFold3. Several lines of evidence suggest that CF-random sometimes works by sequence association: relating patterns from homologous sequences to a learned structural landscape. Through a blind search of thousands of Escherichia coli proteins, CF-random suggests that up to 5% switch folds.