AlphaFold2 generalizes to unseen secondary structures and topologies

Summary

In training OpenFold, (1) found that excluding all sheet proteins or all helical proteins from the training set did not prevent it from learning to sample those topologies correctly. Additionally, it could predict the structures of CATH domains removed from training, although that had a more dramatic effect on predictive performance than random subsampling of the training set.

Figures

Ref (1)

See also

• DL structure prediction methods cannot predict CDR conformations unseen during training

1.

Ahdritz G, Bouatta N, Floristean C, Kadyan S, Xia Q, Gerecke W, et al. OpenFold: retraining AlphaFold2 yields new insights into its learning mechanisms and capacity for generalization. Nature Methods. 2024;21(8):1514–24. Available from: https://doi.org/10.1038/s41592-024-02272-z