Research On T-cell Epitope Prediction Based On Deep Learning

With the development of bioinformatics research,T-cell epitope prediction has become a research hotspot.Antigens invade the human body to stimulate the immune response,stimulate T immune cells to differentiate and lyse target cells and produce lymphocytes to promote the generation of antibodies.During the immune process,the T-cell receptor(TCR)is a protein receptor on the surface of T-cell,and the epitope of T-cell is a segment that can be recognized by TCR and stimulate the differentiation of T-cell.T-cell epitope prediction is mainly to predict which part of the sequence of antigenic peptide binds to TCR.The prediction of T-cell epitope by traditional methods is costly and time consuming.This requires the use of computer algorithms to explore the properties of antigens and TCR sequences.In recent years,in the field of deep learning,Convolution Neural Network(CNN),Long Short-Term Memory(LSTM),Autoencoder and other models is widely used in T-cell epitope prediction and obtained excellent results.However,the T-cell epitope prediction algorithm is often encode TCR and epitope by inside sequence amino acid position or chemical and physical properties.The similarity feature between TCR sequences is ignored,and based on the existing algorithms,more complete sequence information can be extracted.CNN can extract the remarkable features in the sequence,but it is not comprehensive to extract the associated features of the positions forward and backward of the amino acids in the sequence.LSTM extracts the forward correlation information of amino acids at each position and starting from the inside of the sequence,which can solve the problems existing in CNN.However,LSTM network can only capture the features of unidirectional amino acid sequences,while ignoring the features of bidirectional amino acid sequences.At the same time,it was found that different amino acids in different positions of the TCR sequence had different effects on the TCR-epitope binding results.Therefore,this paper proposes two models to effectively solve the above problems:The main innovations and contributions of this paper are as follows:(1)The similarity of TCR topological space was found,and a new model of T-cell epitope prediction based on topological structure was proposed.The potential correlation features of TCR sequences are obtained by constructing TCR topology.This model provides a new idea for the prediction of T-cell epitope.(2)It is found that amino acids in different positions of TCR sequence have different effects on the prediction results of epitope.A new model of bidirectional long short-term memory network prediction based on attention network is proposed This model constructs by three networks,namely Bi-directional Long Short-Term Memory(Bi LSTM),Attention network and CNN,which extract sequence features from different perspectives respectively to achieve further optimization of TCR and epitope feature extraction.To sum up,the first model in this paper modeled and predicted single T-cell epitope,while the second model modeled and predicted multiple epitopes.The two models were extensively tested on the public datasets VDJDB,IEDB and Mc PAS-TCR,and evaluated by ROC curve,PR curve and other indicators.The experiments show the effectiveness of the T-cell epitope prediction model proposed in this paper based on deep learning method,which lays a foundation for the research of T-cell epitope prediction algorithm.