Germline dependence in protein language model retrieval of adaptive immune receptors.

For antigen-specific nearest-neighbor retrieval using CDR3 sequences, classical sequence comparison (BLOSUM62, Levenshtein) beat general ESM2 protein language models by 7–12 percentage points across four B- and T-cell receptor databases.

• Why it matters: It is unclear whether protein language models truly add value beyond classical sequence similarity for adaptive immune receptor retrieval, and which sequence features drive any gains.
• What they did: The study benchmarked general and receptor-specific PLMs versus classical methods for nearest-neighbor retrieval across four receptor databases, using CDR3-only and full variable-domain inputs plus germline-reversion, region-masking, and V-gene controls.
• The result: When using full variable-domain sequences, alignment and PLM performance converged, largely explained by conserved germline-derived context (especially V-gene identity), with limited evidence of antigen-specific extraction from CDRs.

AI summary of the abstract · may contain errors · not medical advice · read the paper.