AntiSite: Unified paratope prediction for antibodies and nanobodies with or without structure

Accurate paratope prediction provides insights into antibody--antigen recognition while supporting antibody discovery and engineering. Existing predictors are fixed to a single input setting: sequence-based methods cannot use an antibody structure, structure-based methods require coordinates, and methods covering both combine separately trained predictors. We present AntiSite, an antigen-agnostic predictor whose sequence encoder, fusion block and prediction head are shared across both settings; structure-aware inference adds molecular-surface features from our previously developed ParaSurf model, used here as a frozen, separately trained structural encoder. Structural-modality dropout enables both inference settings without retraining. Across five antibody--antigen benchmarks, structure-aware AntiSite achieved the highest mean PR-AUC among all evaluated methods, while sequence-only PR AUC was generally lower than that of the strongest dedicated antigen-agnostic sequence baseline. AntiSite also demonstrated zero-shot generalization to nanobodies without additional training. We further evaluated structures from recent antibody-specific and general protein folding methods, including Ibex, Protenix-v1, Boltz-2 and Chai-1, in antibody-only or antigen-aware co-folding settings. Their benefit was associated with structural accuracy, with the most accurate inputs yielding performance close to that obtained with experimental structures. AntiSite thus provides a shared predictor for antibody and nanobody binding sites, usable from sequence alone and refined when reliable structural information is available. Code and trained checkpoints are openly available.