An Ensemble Learning Model for Predicting Unseen TCR-Epitope Interactions

The recognition of TCR-Epitope interactions is key to the development of vaccines. In this study, we develop an ensemble learning model to predict the interactions between unseen TCR and epitope sequences. The proposed ensemble learning model consists of three individual BiLSTM models with different codings of sequences as their inputs, namely, BiLSTM with one hot vector, BiLSTM with ProtBERT embedding vector, and attention-based BiLSTM with physicochemical property vector. During the model training, the equally-long sequences are created by zero padding, which may introduce noise to the data. To mitigate such noise, a mask strategy is adopted in this study. The experiments on dataset VDJdb+Covid-19 shows that BiLSTM with one hot vector, BiLSTM with ProtBERT embedding vector, and attention-based BiLSTM with physicochemical property vector achieves AUC-ROC values of 0.78, 0.77 and 0.74, respectively, which are already comparable to or better than the existing competing methods. Furthermore, the ensemble learning model based on the predicted posterior probabilities from three individual BiLSTM models achieve the AUC-ROC value of 0.79, which is convincingly better than the existing competing methods.