Evaluating predictive patterns of antigen-specific B cells by single-cell transcriptome and antibody repertoire sequencing.

The field of antibody discovery typically involves extensive experimental screening of B cells from immunized animals. Machine learning (ML)-guided prediction of antigen-specific B cells could accelerate this process but requires sufficient training data with antigen-specificity labeling. Here, we introduce a dataset of single-cell transcriptome and antibody repertoire sequencing of B cells from immunized mice, which are labeled as antigen specific or non-specific through experimental selections. We identify gene expression patterns associated with antigen specificity by differential gene expression analysis and assess their antibody sequence diversity. Subsequently, we benchmark various ML models, both linear and non-linear, trained on different combinations of gene expression and antibody repertoire features. Additionally, we assess transfer learning using features from general and antibody-specific protein language models (PLMs). Our findings show that gene expression-based models outperform sequence-based models for antigen-specificity predictions, highlighting a promising avenue for computationally guided antibody discovery.