Continual Learning for Emerging Epitope Landscapes in TCR peptide Binding Prediction

The TCR-peptide binding landscape evolves continuously: novel pathogens (SARS-CoV-2, emerging influenza variants) and newly characterized tumor neoantigens introduce epitope families with no training precedent. Deploying a static model trained on historical data leads to degraded performance on emerging epitopes, while naive fine-tuning on new data causes catastrophic forgetting--erasing performance on previously learned epitopes. We introduce ContinualTCR, a continual learning framework that combines reservoir replay with Elastic Weight Consolidation (EWC) regularization to balance stability (retaining old-epitope performance) and plasticity (adapting to new epitopes). Evaluated on a temporally partitioned VDJdb- IEDB benchmark across four sequential epitope arrival tasks, ContinualTCR achieves new-epitope AUROC 0.812 and old-epitope AUROC 0.781 simultaneously--reducing catastrophic forgetting by 62.9% relative to naive fine-tuning. A streaming evaluation protocol with per-task backward transfer (BWT) reporting reveals that replay alone resolves 30.6% of forgetting, EWC alone resolves 57.3%, and their combination achieves synergistic complementarity. These results establish continual learning as a necessary component of production TCR specificity systems that must adapt to evolving pathogen and neoantigen landscapes without requiring full retraining.