DL structure prediction methods cannot predict CDR conformations unseen during training

Summary

The structures of antibody CDRH3 conformations absent from training sets cannot be accurately predicted by deep learning methods (1). An inability to extrapolate to new CDRH3 conformations may be why fine-tuned versions of RF-diffusion for antibody design supplement their training datasets with those of other loop-mediated PPIs (2,3), although no ablations have been done to my knowledge.

See also

• Focused protein sequence libraries are poor training sets
• Deep learning methods cannot generalize T-cell receptor binding to new epitopes

1.

Greenshields-Watson A, Abanades B, Deane CM. Investigating the ability of deep learning-based structure prediction to extrapolate and/or enrich the set of antibody CDR canonical forms. Frontiers in Immunology. 2024;15. Available from: https://doi.org/10.3389/fimmu.2024.1352703

2.

Peng Z, Han C, Wang X, Li D, Yuan F. Generative Diffusion Models for Antibody Design, Docking, and Optimization. openRxiv; 2023. Available from: https://doi.org/10.1101/2023.09.25.559190

3.

Bennett NR, Watson JL, Ragotte RJ, Borst AJ, See DL, Weidle C, et al. Atomically accurate de novo design of antibodies with RFdiffusion. Nature. 2025;649(8095):183–93. Available from: https://doi.org/10.1038/s41586-025-09721-5