Research On Multimodal Drug Repositioning Method Based On Prior Knowledge From Knowledge Graph

Drug discovery is an expensive,time-consuming and high-risk process.In recent years,despite the rapid progress of technology and the growing investment in drug research and development,the success rate of drug development is still remain unchanged.Drug repositioning is a process of discovering new indications for known drugs,which can significantly accelerate the process of drug development,and reduce the cost and risk.Therefore,the drug repositioning has attracted wide attention and become an important strategy in drug development.In recent years,many computational drug repositioning methods have been proposed and achieved excellent results.But existing methods still suffer from some limitations.First,these methods use very finite types of relations data and ignore the multi-type relations among drugs,diseases and other bio-entities.Second,the existing methods only considered single features and ignored the multimodal features of drugs and diseases,which leads to their poor generalization.Finally,existing methods usually use simple addition or concatenation when fusing drug features and disease features,and they do not fully consider the specificity of features,resulting in a large amount of information loss.In summary,the existing drug repositioning methods still have great room for improvement in terms of data,features and feature fusion.To solve the above problems,this paper propose a multimodal drug repositioning method based on prior knowledge from knowledge graphs,called Graph-based Prior Knowledge(Graph PK).For the improvement in terms of data,Graph PK uses various relations among drugs,diseases and other bio-entities to construct the knowledge graphs which can provide rich prior knowledge for drug repositioning.For the improvement in terms of features,Graph PK extracts the features of drugs and diseases from the knowledge graph modal,known association modal,and biological domain knowledge modal respectively.The features from these three modals can help the model better predict drug repositioning and improve the generalization of the model.For the improvement in terms of feature fusion,Graph PK designs a multimodal neural network to fuse features from three modals and predict drug repositioning.The multimodal neural network fully considers the specificity of features from different modals,and improves the utilization rate of information.Experiments show that Graph PK can achieve better performance than other existing methods.In addition,the influence of multimodal features and feature extraction methods on model performance is also studied.Finally,the case studies prove that Graph PK has strong practical application value.