Explicit germline encoding improves prediction of non-germline CDRH3 residues

Summary

Encoding germline and non-germline amino acids as separate token types improves antibody-language-model prediction of non-germline residues in CDRH3 without sacrificing strong framework prediction. PRISM uses a factorized 53-token vocabulary and separates germline from mutated residues in representation space (1).

See also

• PLMs are biased toward germline antibodies
• Germline

1.

Kim J, Blalock N, Kulkarni A, Nakamura K, Romero PA. Explicit representation of germline and non-germline residues improves antibody language modeling. openRxiv; 2026. Available from: https://doi.org/10.64898/2026.05.06.723387