Training antibody language models on normalized mutation frequencies improves zero-shot expression prediction

Summary

Training antibody language models on normalized mutation frequencies improves zero-shot expression prediction (1). This approach relies on A) normalizing amino acid mutation frequencies by their likelihood in the codon table as well as substitution rates in non-transcribed regions of DNA, and B) germline-descendant substitution pairs observed in phylogenies derived from next-generation sequencing of antibody repertoires.

Figures

Ref (1)

See also

1.

Bitbol A-F. eLife Assessment: Separating selection from mutation in antibody language models. 2026; Available from: https://doi.org/10.7554/elife.109644.3.sa0