The abstract is read at the publisher; the summary is JClub's.
AntiSite predicts antibody paratopes with shared sequence-and-structure modeling, achieving the highest mean PR-AUC across five antibody–antigen benchmarks and enabling nanobody zero-shot generalization.
- Why it matters: Paratope prediction is constrained by predictors that only work with either sequence or structure, limiting antibody discovery and engineering when structural coordinates are unavailable.
- What they did: AntiSite uses a shared sequence encoder, fusion block, and prediction head for both settings, adding ParaSurf molecular-surface features during structure-aware inference via frozen structural encoding and using structural-modality dropout.
- The result: Structure-aware AntiSite attains the top mean PR-AUC across evaluated methods, sequence-only PR AUC is generally lower than the strongest dedicated antigen-agnostic sequence baseline, and more accurate predicted structures yield performance near experimental-structure results.
The findingWhy it mattersWhat they didThe result
AI summary of the abstract · may contain errors · not medical advice · read the paper.