TTCR: Accurate TCR-Epitope Binding Affinity Prediction Using Transformers

Predicting the binding affinity between a T cell receptor (TCR) and an epitope is essential in cancer immunotherapy. The existing predictor epiTCR employs Random Forest to distinguish between binding and non-binding TCR-epitopes. However, the feature representation learning capability of Random Forest is limited, resulting in unsatisfactory prediction outcomes. Therefore, there is a need to build an improved TCR-epitope predictor. In this paper, we build a predictor based on transformers, called TTCR. This model uses two different tokenizers to convert the TCR and epitope sequences into the number sequences that can be understood by the network. We embed the transformers into TTCR encoders, in order to increase the interpretability of the model. The developed models are trained and tested based on five public datasets. The experimental results demonstrate that TTCR is a powerful tool to conduct TCR-epitope binding affinity prediction.