Prediction Of Protein Secondary Structure Based On Scattering Convolutional Neural Network

In the field of bioinformatics,protein secondary structure prediction is of great significance.It is necessary to fully understand the function and structure of proteins.Scientists have never stopped studying protein structure.This paper mainly uses deep learning models to further improve the prediction accuracy of protein secondary structure.The main work of this paper includes the following aspects:(1)The method is based on an optimized convolutional neural network.Firstly,the protein data is processed,and CASP11 is used as the verification set to establish a convolutional neural network model.Then the four hyper-parameters of the convolutional neural network,learning rate,gradient impulse,and regularization coefficient are combined with the training set and the verification set to construct the Bayesian optimization algorithm.Secondly,the Bayesian optimization algorithm is obtained through the training network and Bayesian optimization.Based on the network structure and parameters,the Q₃prediction accuracy rates of 81.36%,80.83%,and 84.29%were obtained on the test set CASP10,CASP11,and CB513 data sets.(2)The classification method is based on optimized convolution and the BiLSTM network.The model combines the optimized convolutional neural network and the BiLSTM neural network and uses the protein feature matrix to predict the secondary structure of the protein.The optimized convolutional neural network can extract local features between complex amino acid residues in protein sequences.In addition,the BiLSTM neural network can further extract complex long-range interactions between amino acids.The model proposed in this paper can achieve better results.(3)The prediction method is based on an adversarial network and convolutional neural network.This paper combines the generative confrontation network and the convolutional neural network and proposes a method based on the combination of the generative confrontation network and the convolutional neural network to predict the protein secondary structure.Generating adversarial networks can extract the features of amino acid residues between protein sequences,and use convolutional neural networks to predict the secondary structure of proteins after fusing the extracted features with the original protein features.On the CASP9,CASP10,CASP11,CASP12,CB513 and 25PDB datasets,Q₃prediction accuracy rates of 87.06%,87.24%,87.31%,87.39%,88.13%and 88.93%were obtained respectively.The accuracy of Q3 prediction has been significantly improved.(4)The classification method is based on wavelet scattering and a convolutional neural network.Use wavelet scattering to extract the features of protein data.Since the wavelet scattering network has different scales,different scales are set for analysis.The different protein features extracted by the wavelet scattering network are predicted by the convolutional neural network.Therefore,this chapter combines a wavelet scattering network and a convolutional neural network to perform feature secondary processing to achieve a better classification effect.Experiments show that in the method used in this paper,the public test set is used to test the model used in this paper,through Bayesian optimization of the convolutional network,combined with the BiLSTM network model,the use of generative confrontation and convolutional neural networks,and the use of wavelet scattering networks and convolutional neural networks,the methods used in this article are more accurate than the Q3 obtained by using convolutional neural networks alone.This proves that the model proposed in this paper is feasible and improves the accuracy of Q3 prediction.