Research On Computational Models For Drug Repositioning

There are some problems in drug discovery,such as low profit to risk ratio and long cycle.Drug repositioning methods use existing researched drugs as the object to discover,confirm and apply new indications to solve problems in drug discovery.In this thesis,aiming at the effectiveness and safety of drugs in the process of drug repositioning,we conduct drug-target interaction prediction and drug side effects identification,combine with drug known disease data for drug repositioning,and build a drug repositioning data analysis platform to provide theoretical guidance for drug research and development.The work of this thesis is as follows:1.Aiming at the problem of insufficient utilization of the network topology information and node neighborhood information of the drug-target interaction network when using the matrix factorization method to predict the drug-target interaction,we propose a drug-target interaction prediction model based on collaborative matrix decomposition,which integrates the topological and attribute features of drugs and targets.In this thesis,self-paced learning is applied to avoid the model falling into bad local minima during the training process.In the case of predicting the interaction between a known drug and a known target,the value of AUPR(Area Under the Precision-Recall Curve)on the ion channel data of the Yamanishi dataset is 0.962,and the value of AUC(Area Under ROC Curve)is 0.995.2.Aiming at the problem that the drug side effect recognition calculation method uses a single feature information to express limited information,we propose a drug side effect identification method based on multiple kernel learning.Gaussian interaction profile kernel,correlation coefficient kernel,cosine similarity kernel and mutual information kernel of drugs and side effects are calculated as the basic kernel respectively.The optimal core of drugs and side effects is obtained by maximizing the cosine similarity method.Finally,the Kronecker RLS(Recursive Least Square)method is utilized to optimize the objective function to identify the side effects of drugs.The experimental results show that in the Pauwels dataset,AUPR score is 0.668 and AUC score is 0.951.3.In view of the problems of incomplete information and redundant data in largescale biomedical data,we propose a drug reposition model based on matrix projection,which decomposes the original drug-disease association matrix into feature similarity matrix and noise matrix.The objective function is optimized by the augmented Lagrange multiplier method to obtain the best feature similarity matrix,and we combine the known disease data of the drug to finally obtain the drug-disease association relationship prediction matrix,the new indication of the drug is predicted for drug relocation.The experimental results show that AUPR score of this method is 0.823,and AUC score is0.972.4.In this thesis,we design and implement a drug repositioning data analysis platform,which utilizes My SQL database and is developed based on Spring Boot open source framework and browser/server structure.The platform functions include information retrieval,similarity calculation,prediction service,user management,and data management.This platform can assist drug researchers in discovering new indications for drugs and improve the success rate of drug development,which is of great significance to drug development and disease research.