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

Abstract

Antibodies recognized as foreign can elicit anti-drug antibodies that compromise efficacy and safety, making humanness scoring---estimating how human-like a sequence appears---an early, repeated step in antibody development. Widely used scores reduce a variable domain to one framework-dominated number, leaving the complementarity-determining regions (CDRs)---grafted during humanization, diversified by generative models---weakly supervised and rarely audited. Here we introduce IgScope, a 23-million-parameter antibody language model trained from scratch that scores humanness at chain, IMGT-region and residue resolution; a region-isolated attention module confines attention to within-region pairs, and chimeric-sequence training supplies region-level supervision. On 2.38 million held-out chains, masking all framework tokens leaves CDR accuracy unchanged, and back-mutated murine residues in human frameworks are detected at 93.4\% accuracy. At chain level, IgScope matches the strongest published correlation with clinical anti-drug-antibody incidence (r = $-$0.567, n = 217) and outperforms a same-protocol recomputation. Region-level scoring exposes what chain-level scoring conceals: antibodies designed de novo by four generative models pass whole-chain humanness (0.966--0.988) while their CDRs score far below natural (P $<$ 1e-30) in every chain class. Quality control turns the same resolution to the mirror-image question: in humanized-mouse platforms, whose domains are human by design, IgScope scores the domain as human and, in HUGO-Nano chimeric VHHs, isolates only the FR2 signal of four camelid hallmark residues deliberately retained to stabilise the VHH. IgScope gives antibody humanization a verified foundation and generative design an audit.