Structure-based De Novo Nanobody Design Without Antibody Data

Nanobodies (VHHs) are compact, stable single-domain binders with broad thera- peutic and reagent utility, but structure-based de novo nanobody design is limited by the scarcity of antibody–antigen complex structures. We present Proteina- Nanocuerpo, a 160M-parameter latent flow-matching model for target-conditioned nanobody design trained without a single antibody structure or sequence. We initialize from Proteina-Complexa, a general protein-complex generative model that jointly denoises backbone Cαcoordinates and per-residue local latents from a frozen all-atom autoencoder, and fine-tune exclusively on general synthetic protein– protein dimers that are stochastically relabeled as pseudo-VHH complexes. This procedure enables antibody-specific conditioning using only non-antibody data. We extend Proteina-Complexa with a framework channel that supplies frame- work sequence and structure along the binder axis under staged dropout. At inference, this allows a germline VHH framework to remain fixed while CDR loops of specified lengths are co-designed. We also extend the target channel to provide secondary-structure and interface descriptors, together with hotspot masks, on both the sequence and pair tracks. In addition, we introduce an in silico design benchmark that enables fair comparisons among methods that have not previously been evaluated head-to-head. Our preview checkpoint achieves strong unique strict success rates without reward-based ranking, filtering, or test- time scaling. We further see that structurally successful designs perform well on open-source, sequence-based nanobody metrics, including developability scores, antibody pseudo-perplexity, and humanness scores. We emphasize that this work is in-progress and we are actively working to extend the benchmark targets and model performance. Wet-lab validation is in progress.