The Protein Secondary Structure Prediction Based On Convolutional Neural Network

Protein makes all life present the complex and changeable state,and it plays an important role in the construction of life system.The protein secondary structure is the important basis for protein to form stable conformation,and it is the premise of researching protein tertiary structure.The research of protein secondary structure provides the premise for understanding the function of protein and the mode of interaction between sequences.The research of secondary structure is useful to understand biological activities and various biological traits,and it helps to develop the new drugs.Therefore,it is an important task to extract useful biological information from a large number of protein sequences.In this paper,a new encoding method called radical group encoding method is proposed.The radical group encoding method is defined according to the stable molecular functional groups within amino acids,the new encoding method contains forty two functional groups.(1)the protein secondary structure prediction based on new radical group:The radical group encoding method and the quadrature encoding method use Support Vector Machine(SVM)to achieve results,and the results are compared.The datasets are CB513 and 25 pdb.Two experiments are set for SVM,the experiment one is to compare the results of radical group and the quadrature encoding method,and the radical group encoding method result is 1.08% higher than the quadrature encoding method result.The experiment two is to choose two different simplifying methods which are G,H,I?H,B,E?E,others?C and H?H,E?E,others?C,the simplifying method of H,G,I?H;B,E?E;others?C is the most difficult and the result is the lowest.(2)the secondary structure prediction based on Convolutional Neural Network :The first step of the experiment is to combine the radical group encoding and position specific scoring matrix,the second step is to achieve the accuracy based on Bayes Classifier after extracting feature through Convolutional Neural Network(CNN).The accuracies show that the result through CNN is higher 5% than Bayes Classifier.(3)the secondary structure prediction based on Auto Encoder:The first step of the experiment is to combine the radical group encoding and position specific scoring matrix(PSSM),the second step is to extract features using the single layer Auto Encoder and two layers stacked Auto Encoder.The extracted data will be predicted through Bayes Classifier.The results are divided into the Bayes Classifier accuracy,the accuracy of single layer Auto Encoder and the accuracy of stacked Auto Encoder.The stacked Auto Encoder can achieve the highest accuracy.The result of two layers is higher 2.69% than Bayes Classifier.