Heavy-chain templates and self-distillation improve nanobody structure prediction

Summary

Augmenting nanobody structure predictors with conventional heavy-chain structures and filtered self-distilled VHH structures improves prediction accuracy. Nanobody-specific filtering of synthetic structures reduced H-CDR3 RMSD and structural violations at a fixed training-compute budget (1).

See also

• Self-distillation exposes more data to Evoformer
• Deep learning methods produce different CDRH3 conformers

1.

Ali M, Greenig M, Jaskolowski M, Crnogaj M, Smorodina E, Zhao H, et al. Improving nanobody structure prediction with self-distillation. openRxiv; 2025. Available from: https://doi.org/10.64898/2025.12.01.691162