Prediction Of Protein Secondary Structure Based On Temporal Convolutional Neural Network

Proteins are essential to life,and their structure determines their function.The protein secondary structure is formed by the folding of the protein primary structure and the protein tertiary structure is formed by the bending and folding of the secondary structure.Therefore,the study of protein secondary structure is very helpful to the overall study of protein structure.With the development of machine learning and deep learning,the accuracy of protein secondary structure prediction has been continuously improved,and some achievements have been made in the field of protein structure prediction,but these achievements are still not enough to meet our large demand for protein information.Because of this,we need to strengthen the exploration of protein structure field,and strive to stimulate the development of protein structure field.In this thesis,Temporal convolutional neural networks and aggregation of other models are mainly used to improve the efficiency of protein secondary structure prediction.The main work of this thesis includes the following two aspects:（1）Protein secondary structure prediction based on convolution-bidirectional temporal convolutional networks.The method was named C-BITCN by modifying TCN to bidirectional structure and feature fusion of bidirectional TCN and CNN.Because the model combines the advantages of convolutional neural network and bidirectional temporal convolutional network,it can not only capture the local correlation in amino acid sequence,but also analyze the distant interaction in amino acid sequence.In C-BITCN,PSSM is entered into CNN after windowing processing as input data 1,and the data entered into BITCN after combining features extracted from CNN and PSSM sequence as input data 2.The accuracy of Q₃ of C-BITCN in CASP10,CASP11,CASP12,CASP13,CASP14 and CB513 test sets were 83.21%,80.56%,79.53%,79.52%,78.89%and 85.34%,respectively.The accuracy of Q₈ was 70.62%,70.53%,69.40%,68.50%,67.00%and 75.03%,respectively.Compared with CNN alone,they increased by 2.64%,2.45%,2.25%,1.52%,1.83%,3.27%and 1.64%,1.45%,1.25%,1.52%,0.83%and 2.19%,respectively.The experimental results show that the combination of convolutional neural network and bidirectional temporal convolutional network can effectively predict the secondary structure of proteins.（2）Protein secondary structure prediction based on temporal convolutional neural network（TCN）and graph convolutional neural network（GCN）.In this thesis,the graph convolutional neural network and TCN were combined for the first time to predict the secondary structure of proteins.In this model,the PSSM information processed into sliding window is taken as the input data of the model,and features are extracted through the strong extraction ability of the graph neural network.The obtained features of the specified dimension are fused with the PSSM sequence and input into the bidirectional TCN for prediction.At the same time,the excellent processing ability of graph convolutional neural network for higher-level structured data and the powerful processing ability of TCN for sequences are utilized.The accuracy of Q₃ in CASP10,CASP11,CASP12,CASP13,CASP14 and CB513 test sets is 84.27%,81.35%,80.21%,79.39%and 85.90%,respectively.The accuracy of Q₈ was 71.23%,70.98%,70.26%,69.19%,67.64%and 75.81%,respectively.Compared with the simple use of two-way sequential convolution,the results are 3.05%,2.78%,2.02%,1.28%,1.80%,1.81%and1.97%,1.21%,1.66%,1.85%,1.01%,1.94%,respectively.