AlphaFold3 beats Boltz-2 and Chai-1 at predicting how nanobodies bind viral antigens, with practical rules for trusting the output
- Compares three co-folding models (AlphaFold3, Boltz-2, Chai-1) on viral antigen-nanobody complexes deposited after AlphaFold3's training cutoff.
- AlphaFold3 performed best, and sampling many more seeds per target steadily raised the share of correct epitope predictions.
- Confidence-score thresholds combining interface and CDR3 loop metrics filter out most wrong models, and antigen multimeric state and glycosylation matter for some targets.
- Applied retrospectively to rank neutralizing nanobodies and map a broadly neutralizing epitope across alphaviruses; no new experiments.
💡 Points Based on the abstract
Post-cutoff structures are the right test, but the thresholds were tuned and applied on the same retrospective sets, so expect them to loosen on your target. It is also worth asking whether more seeds make wrong poses more consistent too; a wet-lab epitope binning or mutagenesis check on two or three nanobodies would show if the ranking is real.
Benchmarking Machine Learning Models for Viral Nanobody Epitope Prediction ↗