A simple antibody-format lookup beats sequence-based immunogenicity predictors, exposing weak ADA labels
- The authors corrected and expanded a public antibody immunogenicity dataset and manually rebuilt trial-level clinical evidence for a subset of therapeutics.
- On the original labels, a format-only lookup outperformed every learned sequence model, and apparent accuracy tracked split composition.
- Recomputing labels from trial-level data erased that advantage, so label quality limits progress; they advise reporting trial-level ADA, modelling distributions and using format baselines.
💡 Points Based on the abstract
A lookup table beating learned models suggests the models learned little about sequence. Treat any immunogenicity predictor as a format-level prior until labels are rebuilt at trial level, and don't use it to rank candidates within one format. If a vendor tool is offered, ask first how it compares with the format-only baseline.
The Quagmire of Immunogenicity-Guided Antibody Design: Labels Limit Progress ↗