Correct antibody-antigen prediction in AF3 and related models is partially determined by training set similarity

Summary

Correct prediction of antibody-antigen structure is partially determined by similarity to training set examples (1). This is another example of memorization in protein structure prediction, particularly among AF3-generation predictors.

Figures

Ref (1)

See also

• All-atom structure prediction of RNA is driven by memorization
• Protein-ligand co-folding methods do not generalize beyond their training set

1.

Fromm S, Ludaic M, Elofsson A. Evaluating deep learning based structure prediction methods on antibody–antigen complexes. Bioinformatics. 2026;42(4). Available from: https://doi.org/10.1093/bioinformatics/btag136