Improving nanobody structure prediction with self-distillation

Abstract

Nanobodies are increasingly attractive therapeutic and biotechnological molecules, yet accurate structure prediction of their highly variable H-CDR3 loops remains a central challenge for machine learning models. Here, we investigate whether nanobody-specific structure prediction can be improved through curated synthetic data strategies. We systematically evaluate different data augmentation regimes, including self-distillation from unlabelled VHH sequences. To ensure structural plausibility of synthetic training samples, we develop NanoKink, the first sequence-based classifier of kinked versus extended H-CDR3 conformations, and apply stringent filtering criteria for non-canonical disulfide bond placement and confor-mational accuracy. On a curated benchmark enriched for challenging nanobody features, we show that, for a fixed training compute budget, a nanobody-specific model trained with filtered synthetic data significantly improves over baseline models and NanobodyBuilder2, achieving lower mean H-CDR3 RMSD and fewer structural violations, while remaining competitive with AlphaFold3 at approximately two orders of magnitude lower per-structure inference time. Our results highlight promising directions in synthetic data generation for nanobody structure modelling and provide a practical framework for optimisation of VHH structure prediction models.