Heavy-chain immune repertoire sequencing enables language-model prediction of antigen-specific antibodies

Rapidly identifying antigen-specific antibodies within complex B cell repertoires is important for therapeutic antibody discovery, vaccine development, disease surveillance, and immune condition monitoring, especially for emerging pandemics. RNA deep sequencing can rapidly provide antibody sequence candidates, but predicting their binding-specificity has remained difficult. Here we show that antigen-specific antibodies can be predicted directly from mRNA-derived heavy-chain V(D)J deep sequencing of unselected immune repertoires by parameter-efficient fine-tuning of the protein language model ESM-2. We have achieved high antigen recognition accuracies across antibodies specific for SARS-CoV-2 spike protein, influenza hemagglutinin, and HIV glycoprotein gp120 antigens, respectively. Our fine-tuned language model, Antigen Specificity Predictor, when applied to unsorted peripheral blood repertoires from immunized mice by single-cell deep sequencing, could predict specific B cell receptors at high frequency, which were then experimentally validated. A significant overlap was obtained with the predicted receptors when benchmarked on previously unseen human B cell receptor sequences identified by barcoding-enabled affinity selection. In bulk mRNA-sequences of human immune repertoires, the predicted antigen-specific B cells exhibited characteristics reminiscent of biology-aware learning. Our model’s performance cannot be explained by sequence memorization. We establish that unselected heavy chain antibody sequences alone carry sufficient signal for repertoire-scale computational antibody discovery and immune profiling, thus motivating potential extension to autoantibody identification in cancer and autoimmune diseases. Teaser Language models decode antibody-antigen specificity from sequence alone, enabling large-scale immune profiling without experimental selection.