Summary
Accounting for substitution likelihood (e.g., via the codon table) improves variant effect prediction relative to just using pure log-likelihood (1). This was done in DASM (2) using neutral evolution rates inferred from untranscribed portions of the genome.
Figures
Ref (1)
1.
Lu SZ, Vermani A, Sanno K, Lu J, Matsen FA, Jagota M, et al. Conditionally Site-Independent Neural Evolution of Antibody Sequences. 2026; Available from: https://arxiv.org/abs/2602.18982
2.
Matsen FA, Dumm W, Sung K, Johnson MM, Rich DH, Starr TN, et al. Separating selection from mutation in antibody language models. eLife. 2026;15. Available from: https://doi.org/10.7554/elife.109644.3