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

Generative sequence-based antibody design is increasingly influenced by learned property predictors, but the success of these approaches is limited by the quality of the labels guiding them. Immunogenicity is a crucial property to consider, yet it is difficult to specify: the only metric used is a single anti-drug-antibody percentage (ADA%) from published trials.

To examine this metric, the authors extended and corrected the widely used 217-entry dataset to include 231 therapeutics, covering various formats from mouse to fully human antibodies. They also manually reconstructed the clinical evidence for 110 of these therapeutics from primary sources. The study reveals that using an antibody format lookup surpasses all sequence-based learned predictors when evaluating on this expanded dataset.

The apparent accuracy is actually determined by the format composition of the evaluation split, not the sequence signal itself. Moreover, the labels underestimate their own evidence, as antibodies labeled at or below 6 ADA% can span up to 86% across different trials. By recomputing the labels from the trial-level measurements, the advantage of the format-based approach disappears, while sequence-based predictors degrade less under the cohort-weighted mean.

The authors conclude that the progress in antibody design is constrained by the labels themselves, suggesting that researchers should report ADA% at the trial level, treat the target as a distribution rather than a single point value, and evaluate against a format-based baseline.