Research On Binding Prediction Of T-cell Receptor And Antigen Based On Attention Mechanism

The immune response in which the body recognizes and binds to specific pathogens during contact with pathogenic microorganisms,and ultimately removes them from the body,is known as specific immunity.In specific immunity,T lymphocytes specifically bind to antigenic epitopes through surface antigen receptors,leading to lymphocyte activation,proliferation and thus immunity.Therefore,predicting whether the T cell receptor(TCR)binds to an antigenic epitope plays a significant role in the study of the immune process and the immunogenicity of antigens.The traditional experimental approach to verify whether an antigen can bind to a TCR has been labor-intensive and costly.With the recent increase in protein sequence data and the development of bioinformatics,methods have been developed to use computer techniques to analyze whether an antigen can bind to a given TCR.In particular,with the recent development of deep learning,deep learning models such as Convolutional Neural Networks(CNN)and Recurrent Neural Networks(RNN)have been proposed to predict the binding of T cell receptors to antigens with good results.However,current deep learning-based prediction methods focus more on the relationship between a per-amino acid and adjacent amino acids,ignoring the importance of the amino acid relative to the whole sequence.In addition,the relationship between TCR and epitope is unclear,and the potential relationship that may exist between the two is ignored.At the same time,the prediction of unknown epitopes has been a challenge in this field.Current prediction algorithms for unknown epitopes mostly start from the model itself,using the model to learn sequence features autonomously.There is redundant information in the extracted features,and the current prediction algorithms for unknown epitopes ignore the key information of epitope homology.Based on this,two models are proposed in this thesis to solve the above problems.(1)A prediction model based on convolutional multi-head attention and bidirectional attention is proposed to solve the problem that CNN-based methods ignore global features and the information between sequences is not fully utilized.The attention mechanism is taken as the main feature extraction method,and the convolution combined with attention mechanism is used to extract global features.The bidirectional attention mechanism is adopted to learn the potential interaction feature information between TCR and antigenic epitopes,and further extract features between TCR and epitopes.(2)A prediction model based on integrated attention distillation network is proposed to solve the problem that current deep learning methods ignore the homology between epitopes and have great difficulty in identifying unknown epitopes for TCRs.The attention mechanism is taken as the main feature extraction method and the model improves the prediction ability of unknown epitopes by defining the distance between different epitopes to characterize the similarity and potential homology between epitopes.A similar effect is achieved by building a ’teacher-student’network to compress the model parameters,training the teacher network and transferring the’knowledge’ from the teacher network to the student network,and reducing information redundancy by increasing the interval attention and adjusting the number of attentions.In summary,the first model in this thesis models single T-cell epitopes and multiple T-cell epitopes,and the second model predicts T-cell epitopes that do not appear.Both models were analyzed and evaluated on the VDJdb dataset,the Mc PAS-TCR dataset and the IEDB dataset.The validity of each module and each method proposed was demonstrated by ablation experiments.The results showed that the proposed model increases the depth of feature extraction,improves the accuracy of prediction through experimental evaluation and comparison tests.And the results validate that the alpha chain of TCR receptors for this prediction task with a good facilitation effect.Meanwhile,the generalization ability of the model was verified by the IEDB dataset.The experiments demonstrate the effectiveness and robustness of the T cell receptor-antigen binding prediction model based on the attention mechanism proposed in this thesis.