The Quagmire of Immunogenicity-Guided Antibody Design: Labels Limit Progress

An antibody format lookup beats all sequence-based learned immunogenicity predictors on a 231-therapeutic dataset, but the apparent accuracy is driven by evaluation split format composition.

• Why it matters: Immunogenicity conditioning is clinically important yet poorly specified because ADA% labels are aggregated from published trials into a single per-molecule value.
• What they did: The team extended and corrected a 217-entry dataset to 231 therapeutics with paired sequences, manually reconstructed clinical evidence for 110 from primary sources, and compared predictors against a format-based baseline.
• The result: Recomputed ADA% from trial-level measurements showed the format-based advantage disappears, while sequence predictors degrade less under cohort-weighted mean, implying label limitations and motivating trial-level, distributional evaluation.

AI summary of the abstract · may contain errors · not medical advice · read the paper.