Assessment of computational methods in predicting TCR-epitope binding recognition

T cell receptors (TCRs) play a vital role in immune recognition by binding specific epitopes. Accurate prediction of TCR-epitope interactions is fundamental for advancing immunology research. Although numerous computational methods have been developed, a comprehensive evaluation of their performance remains lacking. Here we assessed 50 state-of-the-art TCR-epitope prediction models using 21 datasets covering 762 epitopes and hundreds of thousands binding TCRs. Our analysis revealed that the source of negative TCRs substantially impacts model accuracy, with external negatives potentially introducing uncontrolled confounders. Model performance generally improved with more TCRs per epitope, highlighting the importance of large and diverse datasets. Models incorporating multiple features typically outperformed those using only complementarity-determining region 3β information, yet all struggle to generalize to unseen epitopes. The use of independent test sets proved crucial for unbiased assessment on both seen and unseen epitopes. These insights will guide the development of more accurate and generalizable TCR-epitope prediction models for real-world applications.