Revealing bias in antibody language models through systematic training data processing with OAS-explore

Antibody language models (LMs) trained on immune receptor sequences have been applied to diverse immunological tasks such as humanization and prediction of antigen specificity. While promising, these models are often trained on datasets with limited donor diversity, raising concerns that biases in the training data may hinder their generalizability. To quantify the impact of biased training data, we introduce an open-source processing pipeline for the 2.4 billion unpaired antibody sequences in the Observed Antibody Space (OAS) database, enabling customizable filtering and balanced sampling by donor, species, chain type and other metadata. Analysis of OAS revealed that 13 individuals contribute over 70% of human antibody sequences. Using our pipeline, we trained 17 RoBERTa antibody LMs on datasets of different compositions. Models failed to generalize across chain types and showed limited transfer between human and mouse repertoires. Both individual- and batch-specific effects influenced model performance, and expanding donor diversity did not improve generalization to unseen individuals from unseen publications.