FIGURE 1.

AttnTAP improved the prediction accuracy of TCR-peptide binding. (A) AttnTAP was a dual-input deep learning framework, which included the feature extractors for TCR and peptide sequences, Attn-BiLSTM and Attn-MLP. The corresponding feature vectors extracted by the two models were then concatenated for predicting the likelihood of TCR and peptide binding using the multilayer perceptron network. (B) The feature extractor for TCR sequences, Attn-BiLSTM, was divided into four parts: the input layer, bi-directional long short term memory (BiLSTM) layer, attention layer, and output layer. Sequences were preprocessed and encoded into embeddings in the input layer. The embeddings were then fed into the BiLSTM and attention layers, respectively. The BiLSTM layer extracted the sequences’ feature vectors, while the attention layer computed the weights of each position in the sequences. Finally, the output layer outputted the weighted feature vectors.