Assessment of computational methods in predicting TCR-epitope binding recognition

T-cell receptors (TCRs) are critical for the immune system's ability to recognize specific epitopes. Accurate prediction of TCR-epitope interactions is fundamental for understanding and enhancing immune responses. Despite significant advances in computational methods for TCR-epitope binding prediction, a thorough evaluation of these tools remains lacking. Here, we assessed 50 state-of-the-art TCR-epitope prediction models using 21 datasets covering 762 epitopes and hundreds of thousands corresponding binding TCR sequences. Our analysis revealed that while the ratio of positive to negative samples subtly influences performance, the source of negative TCR samples significantly impacts model accuracy. External negative data may introduce uncontrolled confounders, compromising evaluation reliability. A positive correlation between the number of TCRs per epitope and model performance highlights the importance of large, diverse datasets. Multi-feature models generally outperform single CDR3β models, but generalization to unseen epitopes remains a challenge across all evaluated models. Using independently sourced test sets, which offer practical predictions for real-world applications, for both seen and unseen epitopes is crucial for objective performance assessment. These insights would guide the development of more accurate and robust computational tools for TCR-epitope interaction prediction, accelerating advancements in this evolving field.