Research On Drug Repositioning Based On Deep Collaborative Filtering Algorithm

Drug repositioning refers to exploring new therapeutic effects of existing drugs and has become an important research field of drug design.The effective way of drug repositioning is to predict the drug-target interaction(DTI).Relying only on biological experiments to identify the interaction between drugs and targets,it is difficult to find effective drug-target pairs among a large number of drugs and targets.The collaborative filtering technology based on matrix factorization has achieved good results in the field of DTI prediction.However,there are still shortcomings: First,label information and unlabeled information cannot be effectively used in matrix factorization,so how to effectively learn the data is an urgent problem to be solved.Second,the interaction matrix between the drug and the target is very sparse,which can easily affect the accuracy of matrix factorization.In order to overcome the above difficulties and improve the prediction effect of DTI,the main work and innovations are as follows:To solve the problem that matrix factorization cannot effectively learn data information,a semi-supervised graph regularized matrix factorization(SSGRMF)algorithm is proposed.First,in order to learn the limited drug label information,the label matrix is introduced to improve the unsupervised NMF,and a semi-supervised matrix factorization model is constructed.Then,in order to learn the local geometric structure of the drugs data and the targets data,graph regularization of the drugs and graph regularization of the targets are used to make the drugs or targets that are close to each other in the original space also be close to each other in the learned manifold during the factorization process.Finally,the NMF model is extended by the graph regularization of drugs,targets and label constraint to predict the DTI.In order to evaluate the performance of the model,Experiments on the yamanishi dataset,compared with the two matrix factorization methods;SSGRMF has the highest AUC and AUPR.The results show that SSGRMF algorithm can effectively learn the feature of the data when the label data is limited,and has better performance in DTI prediction.For the sparse problem of DTI matrix,auxiliary deep Auto-Encoder is introduced into SSGRMF algorithm,and a deep collaborative filtering algorithm is proposed.Firstly,in order to utilize the original structure information of drugs and targets,a parallel input auxiliary deep Auto-Encoder is designed to simultaneously extract the Latent features of the DTI matrix and drugs or targets.Then,the auxiliary deep Auto-Encoder is integrated into the SSGRMF,Train the model to solve the optimal drug potential impact factor and target potential impact factor.Finally,matrix completion is used to predict the possible DTI pairs.In order to verify the model,the deep collaborative filtering algorithm was tested on the yamanishi data set with added feature data of drugs and targets.The results show that it has a better effect than the latest two DTI prediction methods.Prove that the algorithm in this thesis can alleviate the sparse problem of DTI matrix and improve the accuracy of DTI prediction.Furthermore,comparing the top ten recommended drug-target pairs with public databases,the results showed that 70% of the drug-target pairs were certified,proving the feasibility of this method in drug repositioning.