Summary
Protein folding neural networks do not extrapolate to new ligand binding sites ((1), (2)). AlphaFold2 could not predict a new calcium-binding site in an engineered protein, and instead introduced a fictitious salt bridge (1). Separately, neither AlphaFold3 nor RosettaFold-AA could predict that a binding site prefers cobalt over copper (2).
Figures
Ref (1)
1.
d’Oelsnitz S, Diaz DJ, Kim W, Acosta DJ, Dangerfield TL, Schechter MW, et al. Biosensor and machine learning-aided engineering of an amaryllidaceae enzyme. Nature Communications. 2024;15(1). Available from: https://doi.org/10.1038/s41467-024-46356-y
2.
Dürr SL, Rothlisberger U. Predicting metal-protein interactions using cofolding methods: Status quo. openRxiv; 2024. Available from: https://doi.org/10.1101/2024.05.28.596236