Research On Drug-Drug Interaction Prediction Based On Knowledge Graph

DDI(Drug-drug interaction)prediction aims to identify DDI effectively to reduce the adverse drug reactions when two or more drugs are taken together..In contemporary clinical treatment,multiple-drug treatment is increasingly common,which makes the prediction of drug-drug interaction critical in drug discovery and clinical application.Traditional methods for DDI prediction relaying on vitro and vivo experiments are labor-intensive and limited in experimental scale.The development of the Internet and computer technology,and the establishments of lots of drug databases makes computer-based drug-drug interaction prediction methods popular and effective.More recently,a tool for data integration and representation called knowledge graph has drawn great research attention from researchers.Deep learning models enhanced by knowledge graph can obviously promote the performance of downstream tasks.In biomedicine,several large-scale knowledge graphs related to drugs have been established and opened,which provide data basis for DDI prediction based on the mining of knowledge graphs.Existing computer-based prediction methods usually extract drug information from multiple data sources and treat the task as a binary classification problem.However,the binary classification is difficult to clearly tell the boundary between positive and negative samples owing to the incompleteness and uncertainty of derived data.Moreover,existing methods tend to focus on the properties of the drug itself,and ignore the information on the relations between the drug entity and other biomedical entities.And exiting methods lack an automated way to fuse multiple features obtained from diverse information.Finally,existing methods mainly focus on whether there are interactions between two drugs and only a few studies the prediction of finegrained interaction types.And There is still room remained for improvement in the prediction of multi-type drug-drug interactions.To solve the problems above,we carry out the following research:(1)To solve the problem that binary classification cannot clearly model the boundary between the positive and negative sample owing to the incompleteness and uncertainty of derived data,we propose a novel prediction model named3 WDDI,which introduces three-way decision and knowledge graph embedding to enhance the performance of DDI prediction.3WDDI utilizes three-way decision,a granular computing method,to model the DDI prediction as uncertain decision and treats an external drug-related knowledge graph as supplementary information.Firstly,during the boundary dividing of the three-way decision,the drug pairs are divided into positive,negative,and boundary regions by Convolutional Neural Network(CNN)according to the drug chemical structure feature.Further,delay decision is made for objects in the boundary region by integrating the knowledge graph embedding feature to promote decision-making accuracy.The empirical results show that 3WDDI outperforms several baseline models.(2)To solve the problem that existing methods only focus on the properties of the drug itself and neglect the information of the relations between the drug entity and other biomedical entities,as well as the lack of automated feature fusion methods,we propose a deep learning method called DANDDI,which applies the knowledge graph embedding and attention mechanism in deep learning to prediction multi-type drug-drug interactions.Firstly,the model extracts four kinds of drug attributes including substructure,target,enzyme and pathway to construct four local representations of drugs base on the similarity of each attribute.Then,a sub-knowledge graph is collected from a large-scale biomedical knowledge graph.To generate the global representation of drugs,the sub-knowledge graph is fed into a knowledge graph embedding model called Compl Ex to encode the information of relations between drug and other entities.Further,we design a feature fusion module base on the attention mechanism to fuse the five drug representations above.Finally,the fused representation will be used to predict the result by a deep neural network.The empirical results show that DANDDI outperforms several baseline models.