Prediction Of Protein Secondary Structure By GAN Feature Extraction

Protein secondary structure is the basis of studying protein tertiary structure,drug design and application development.8-state protein secondary structure can provide more adequate protein information than 3-state protein secondary structure.Therefore,this thesis focuses on the study of 8-state of secondary structures of proteins,and mainly uses the popular deep learning model to extract protein features,so as to enhance the prediction accuracy of 8-state of secondary structures.The main work is divided into the following aspects:(1)In this thesis,generative adversarial network(GAN)and Bidirectional Long Short-Term Memory Network(Bi-LSTM)were combined to predict the secondary structure of 8-state of proteins.GAN simulates PSSM to extract features,and then combines with PSSM to enhance features.Two different classification methods are proposed for this model,namely G-BLS and G-BLS-3TO8.G-BLS directly uses the combined model to predict 8-state of structures,while G-BLS-3to8 refers to the relation between the 3 and 8-state of structures.Firstly,3 classifications were made,and then8-state of prediction were made based on them.The experimental results show that both the two methods achieve good prediction accuracy of 8-state of secondary structures.(2)In this thesis,a prediction method named WG-Res is proposed by combining Wasserstein generative adversarial network(WGAN)and residuals network(Res Net).Firstly,WGAN uses Wasserstein distance to optimize GAN,which improves its gradient instability.Res Net uses a deeper model than the traditional convolutional network to fully train,thereby improving the model’s prediction performance and,to some extent,the prediction accuracy.In the end,the prediction accuracy of 8-state secondary structure was 72.53%,71.43%,70.67%,68.83%,69.17% and 75.41% for CASP10-14 and CB513,respectively.The experimental results show that WG-Res has good performance in predicting 8-state secondary structure.(3)This thesis proposes a WG-ICRN structure that combines WGAN with Res Net and Inception for 8-state of secondary structure prediction.After extracting features and fusing PSSM,WGAN enters into Res Net with improved Inception model to get final prediction results.Inception using convolution kernels of different sizes can extract features of different sizes and fuse them together,while also reducing the number of training parameters.The proposal of WG-ICRN not only shortens the training time of data,but also further improves the prediction accuracy of 8-state of structures.Finally,WG-ICRN obtained the prediction accuracy of 73.32%,71.55%,70.81%,68.88%,69.29%,75.56% on CASP10-14 and CB513 data sets,respectively.The experimental results showed that,It is feasible and practical that WG-ICRN can effectively improve the prediction accuracy of 8-state of protein structures.