Drug-Target Interaction Prediction Based On Deep Learning

Accurately identifying drug-target interactions is crucial in drug discovery,since it not only can deepen our understanding of drug action mechanisms,but also contribute to drug repositioning in pharmacology.As traditional experimental methods are limited by cost and throughput,it is important to develop effective computational methods to predict drug-target interactions.Althought variety of computational methods have been used to identify interactions between drugs and targets,their accuracies were relatively low,which in turn requires the development of more advanced algorithms to improve prediction accuracy.Deep learning has been successfully applied in many fields as an emerging class of machine learning algorithms.Therefore,based on the existing computation methods and from the perspective of systematic biology and network pharmacology,in this thesis we made full use of the massive data resources from existing public databases,integrated multiple information of drugs and targets,and adopted various types of deep learning algorithms to establish more efficient and accurate prediction models for drug-target interactions.The specific research content includes the following three parts:The first chapter outlines the background and significance of identifying drug-target interactions,summarizes the existing research methods,and analyzes the advantages and limitations of these methods.On such basis,we briefly demonstrate the deep learning algorithm.We focus on several deep learning algorithms used in our work,analyze the advantages of the algorithms and give examples of some applications in the field of drug development.The second chapter studies a new drug-target interaction prediction method based on deep learning algorithm.The DrugBank dataset was used as a benchmark data.We firstly constructed drug-target interaction features based on drug structure information and protein sequence information,and then employed three types of deep learning framework to effectively learn the representations from input vectors,including deep neural network(DNN),highway network(HN)and recurrent neural network(RNN).The results on the test set were state-of-the-art with the Area Under roc Curve(AUC)and the prediction accuracies(Acc)of the DNN,HN and RNN were 0.96and 0.90,0.94 and 0.88,0.95 and 0.86,respectively.In addition,we established two traditional machine learning models,Random Forest(RF)and Support Vector Machine(SVM).The AUC and Acc of the RF and SVM models were 0.90 and 0.84,0.92 and 0.85,respectively.It can be seen that deep learning models had a certain improvement over the performance of traditional machine learning models.In order to further illustrate the generalization ability of the models,we introduced the Experimental set as the external validation set.The prediction results showed that our methods had better generalization ability and could be used to predict new drug-target interactions.In the third chapter,we mainly research on the new methods of drug-target binding affinity prediction based on deep learning algorithm.Binding affinity reflects the strength of drug-target interactions.In this work,we hope to establish a regression model that can directly predict the binding affinity using deep learning algorithm.We used the PDBBind dataset as the benchmark data.We constructed eigenvectors using two methods,one based on drug structure information and target sequence information,the other analyzed drug-target interaction information based on BINANA algorithm.Then we used DNN to establish a drug-target binding affinity prediction model.Considering the different types of binding affinity(dissociation constant K_d and inhibition constant K_i),we established DL-K_d model,DL-K_i model,and DL-All model without considering grouping strategies.The results on the test set were superior with the Pearson correlation coefficient R and RMSE of the descriptor-based DL-K_i model,BINANA-based DL-All model were 0.85 and 1.07,0.83 and 0.87,respectively.The results suggest that our models had good predictive power.In addition,the prediction performance of the DL-All model based on the two features was better than the DL-K_i and DL-K_d models,indicating that the deep learning algorithm is more suitable for big data.Finally,by comparing with SVM and RF,it is elucidated that the deep learning models have much better prediction performance than the traditional machine learning models.In this thesis,we used deep learning algorithm to establish classification and regression models for drug-target interaction prediction,which not only qualitatively determined whether there was interaction between the drug and the target,but also directly predicted the binding affinity of drug-target interactions.This is of great significance for the repositioning of old drugs and the development of new drugs.