Multi-resolution humanness scoring of antibodies via a region-isolated language model

IgScope, a 23-million-parameter antibody language model, detects back-mutated murine residues in human frameworks at 93.4% accuracy and preserves CDR accuracy when frameworks are masked.

• Why it matters: Humanness scoring is an early, repeated step in antibody development, but existing framework-dominated scores leave CDRs weakly supervised and rarely audited.
• What they did: IgScope scores humanness at chain, IMGT-region, and residue resolution using region-isolated attention and chimeric-sequence training; on 2.38 million held-out chains it masks all framework tokens.
• The result: Chain-level scoring matches the strongest published correlation with clinical anti-drug-antibody incidence (r = −0.567, n = 217) and region-level scoring shows de novo designs pass whole-chain humanness (0.966–0.988) while CDRs score far below natural (P < 1e-30).