Germline dependence in protein language model retrieval of adaptive immune receptors.
- 1Laboratory of Genome Informatics, Research Institute for Microbial Diseases, The University of Osaka, 3-1 Yamadaoka, Suita 565-0871, Japan.
- 2MyImmune Corporation, 251 Little Falls Drive, Wilmington, DE 19808, United States.
Abstract
Protein language models (PLMs) are increasingly used to represent adaptive immune receptors, yet their advantage over classical sequence comparison remains unclear. We benchmarked general and receptor-specific PLMs against classical sequence-comparison methods for antigen-specific nearest-neighbor retrieval across four B- and T-cell receptor databases. Using CDR3 sequences alone, classical methods consistently outperformed general PLMs: BLOSUM62 and Levenshtein exceeded ESM2 variants by 7-12 percentage points in retrieval accuracy. Among receptor-specific models, only SCEPTR and, on one database, TCR2vec matched these baselines, although both additionally encoded V-gene information. AntiBERTa2 and IgBert did not close the performance gap. Using full variable-domain sequences, alignment and PLM performance converged. Germline-reversion, region-masking, and V-gene control analyses indicated that this convergence largely reflected conserved germline-derived context, particularly V-gene identity, with limited evidence for improved extraction of antigen-specific information from the CDRs. Germline dependence also extended to SCEPTR and TCR2vec. Heavily mutated HIV-1 antibodies provided a setting-specific exception to this pattern. Random train-test splits yielded retrieval accuracy 15-28 percentage points higher than clone-aware splits, consistent with optimistic performance estimates due to clonal leakage. These findings emphasize the need for strong sequence-comparison baselines and explicit germline controls when evaluating immune-receptor representations and support clone-aware splitting as a minimum benchmark standard.
Related
- Multiple mechanisms cooperate to control PUF protein FBF-1 and FBF-2 accumulation in Caenorhabditis elegans germline
- A common germline 3′ UTR variant controls DDIT4 translation efficiency and stress responses in cancer cells
- The splicing kinase PRPF-4 is required for somatic development and germline function in C. elegans
- Toward transgenic multicellular parasites: CRISPR-Cas12a germline editing of Schistosoma mansoni eggs as key to success
- Isoform-selective modulation of peroxisome proliferator-activated receptors by quercetin and apigenin revealed through molecular simulations
BibTeX
@article{wang2026germline,
title = {{Germline dependence in protein language model retrieval of adaptive immune receptors.}},
author = {Suyue Wang and Ayan Sengupta and Songling Li and Daron M Standley},
journal = {Briefings in bioinformatics},
year = {2026},
month = nov,
doi = {10.1093/bib/bbag614},
url = {https://europepmc.org/article/MED/42998279},
note = {CC BY},
}
RIS
TY - JOUR
TI - Germline dependence in protein language model retrieval of adaptive immune receptors.
AU - Suyue Wang
AU - Ayan Sengupta
AU - Songling Li
AU - Daron M Standley
PY - 2026
DA - 2026/11/01
JO - Briefings in bioinformatics
DO - 10.1093/bib/bbag614
UR - https://europepmc.org/article/MED/42998279
AB - Protein language models (PLMs) are increasingly used to represent adaptive immune receptors, yet their advantage over classical sequence comparison remains unclear. We benchmarked general and receptor-specific PLMs against classical sequence-comparison methods for antigen-specific nearest-neighbor retrieval across four B- and T-cell receptor databases. Using CDR3 sequences alone, classical methods consistently outperformed general PLMs: BLOSUM62 and Levenshtein exceeded ESM2 variants by 7-12 percentage points in retrieval accuracy. Among receptor-specific models, only SCEPTR and, on one database, TCR2vec matched these baselines, although both additionally encoded V-gene information. AntiBERTa2 and IgBert did not close the performance gap. Using full variable-domain sequences, alignment and PLM performance converged. Germline-reversion, region-masking, and V-gene control analyses indicated that this convergence largely reflected conserved germline-derived context, particularly V-gene identity, with limited evidence for improved extraction of antigen-specific information from the CDRs. Germline dependence also extended to SCEPTR and TCR2vec. Heavily mutated HIV-1 antibodies provided a setting-specific exception to this pattern. Random train-test splits yielded retrieval accuracy 15-28 percentage points higher than clone-aware splits, consistent with optimistic performance estimates due to clonal leakage. These findings emphasize the need for strong sequence-comparison baselines and explicit germline controls when evaluating immune-receptor representations and support clone-aware splitting as a minimum benchmark standard.
SN - 1467-5463
SN - 1477-4054
ER -