HLAResolver: A statistical framework for resolving HLA ambiguity in T-cell response data

Large-scale models of T-cell recognition require training examples that specify the complete T-cell receptor (TCR)-peptide-human leukocyte antigen (HLA) complex. Multiplexed Identification of T-cell Receptor Antigen Specificity (MIRA) is a high-throughput assay that identifies TCRs associated with responses to specific peptides, but does not determine which HLA presents each peptide. Because a donor can carry up to six distinct class I HLA-A, HLA-B, and HLA-C alleles, each response may have several candidate presenting HLAs. To narrow this set of candidate presenting HLAs, we developed HLAResolver, a statistical framework that combines MIRA response patterns across HLA-typed donors with NetMHCpan predictions of peptide presentation. For each peptide, HLAResolver uses negative-binomial regression to jointly estimate associations between HLA alleles and TCR response counts, accounting for differences in assay yield. NetMHCpan eluted-ligand predictions guide sparse regularization, favoring alleles with stronger predicted presentation while allowing strong MIRA response evidence to support alternatives. To assess whether this integration provides an advantage over either approach alone, we compared HLAResolver with NetMHCpan, which predicts peptide-HLA presentation, and the Restrictor Analysis Tool for Epitopes (RATE), which infers HLA associations from T-cell response patterns. Using held-out MIRA experiments, we assessed each method's ability to predict observed responses and the number of candidate HLAs it assigned. At matched specificity for predicting MIRA responses, HLAResolver achieved higher sensitivity while selecting fewer candidate HLAs per peptide than either comparator.