Prediction Of Protein Secondary Structure Based On Deep Learning

The study of protein secondary structure prediction is a very critical sub-problem in the study of template-free three-dimensional conformation.The research content of secondary structure prediction is to find a function that can map the amino acid sequence to be tested into the secondary structure according to the characteristic matrix of the given amino acid sequence and the corresponding secondary structure.Driven by algorithms such as machine learning and evolutionary information,prediction accuracy has been slowly and steadily improved in recent years.To explore the application of deep learning in secondary structure prediction and further improve the accuracy of secondary structure prediction,this paper constructs a prediction model of a full one-dimensional convolution structure,a prediction model of a multi-scale convolution structure,and conditions.Generate a predictive model of the adversarial structure.(1)To improve the prediction accuracy of the secondary structure,this paper first constructed a prediction model of a full one-dimensional convolution structure.In this model,the protein sequence information and the specificity score matrix are spliced together to serve as the input tensor of the model.There are nine hidden layers in total,and one-dimensional convolution operation is used as the basic operation unit.(2)To further improve the model structure and the prediction accuracy,this paper introduces and improves the multi-scale convolution and channel attention mechanism on the basis of the full one-dimensional convolution,constructing a multi-scale convolution structure model.The model can be divided into two parts:feature extraction module and classification module:the feature extraction module is used to extract the features of the original amino acid sequence matrix;the classification module is used to analyze the extracted features and perform eight-state or three-state classification.The introduction and improvement of multi-scale convolution and channel attention mechanisms have greatly improved the feature extraction and function fitting capabilities of the model.(3)Based on the first two models,this paper introduces the idea of adversarial game into the prediction of secondary structure,and constructs a predictive model of conditional generation adversarial structure.Among them,the structure of the generator consists of a multi-scale convolution module and a one-dimensional convolution module for the generation of secondary structure;the discriminator uses the idea of sequence residue discrimination to determine the authenticity of the secondary structure.The three deep learning models proposed in this article have achieved excellent protein secondary structure prediction results.Among them,the prediction model of the multi-scale convolution structure has the best test results on the four benchmark data sets(CullPDB,CB513,CASP10 and CASP11).Outstanding,achieved 74.2%,70.6%,74.9%,71.5%of the eight-state prediction accuracy and 86.3%,84.7%,87.7%,84.8%of the three-state prediction accuracy.The experimental results show that compared with other methods,the three deep learning models proposed in this paper have greater competitive advantages.