Protein-ligand co-folding methods do not generalize beyond their training set

Summary

Protein-ligand co-folding methods such as AlphaFold3 struggle to generalize beyond their training sets (Škrinjar et al 2025 (1,2)). They continue to dock ligands into the training set poses even when active site residues are heavily mutated (2).

Figures

Ref Škrinjar et al 2025 (1)

Ref (2)

See also

• Protein folding neural networks do not extrapolate to new ligand binding sites
• All-atom structure prediction of RNA is driven by memorization

1.

Škrinjar P, Eberhardt J, Tauriello G, Schwede T, Durairaj J. Have protein-ligand cofolding methods moved beyond memorisation?. openRxiv; 2025. Available from: https://doi.org/10.1101/2025.02.03.636309

2.

Masters MR, Mahmoud AH, Lill MA. Investigating whether deep learning models for co-folding learn the physics of protein-ligand interactions. Nature Communications. 2025;16(1). Available from: https://doi.org/10.1038/s41467-025-63947-5