Drug-Drug Interactions Prediction Based On Biomedical Knowledge Graph And Drug Molecular Structure

With the improvement of medical treatment,it is more and more common to use multiple drugs to treat complex diseases.Simultaneous administration of multiple drugs may lead to changes in the efficacy of drugs.To avoid the medical accidents caused by which,mining Drug-Drug Interactions(DDIs)has received a lot of attention.However,there are few known DDIs at present,and it is time-consuming and labor-intensive to find potential DDIs through traditional wet experiments.In recent years,machine learning has been applied to predict DDIs.Most of the existing prediction methods based on machine learning can predict DDIs accurately,but there are still several problems to be solved.For example,link prediction based methods ignore the influence of multiple biological entities in the human body on DDIs.Similarity based methods cannot effectively integrate the association information between drugs and multiple biological entities.Biomedical knowledge graph contains the association information among drugs and various biological entities,and knowledge graph embedding methods can capture the association information to improve the ability of DDIs prediction.However,the traditional knowledge graph embedding methods fail to deal with the complex association patterns among biological entities,and knowledge graph based methods generally ignore the influence of drug chemical properties on DDIs.To solve these problems,a drug-drug interactions prediction model based on biomedical Knowledge Graph and drug Molecular Structure(KGMS)is proposed in this paper.Firstly,KGMS optimizes the methods of extracting drug features from knowledge graph and learns the biomedical knowledge graph based drug features through Node-typeAware based Multi-relational Graph Neural Networks(NAMGNN).NAMGNN can mine the complex association patterns in biomedical knowledge graph to integrate association information among drugs and different types of biological entities.Secondly,considering the influence of drug chemical properties on DDIs,KGMS adopts Graph Isomorphism Networks(GIN)to learn the molecular structure based drug features from drug molecular graphs.Then,KGMS uses the self-attention mechanism to fuse the biomedical knowledge graph based features and the molecular structure based features for the information interaction between the both drug features,and the fused drug features are used to construct DDIs prediction model.KGMS can perform two DDIs prediction tasks:(1)predicting the existence of DDIs between the candidate drug pairs or not;(2)predicting the types of DDIs between the candidate drug pairs.Experimental results show that KGMS performs better than existing DDIs prediction methods in the both kinds of prediction tasks.The performance of KGMS on DDIs prediction is improved by integrating the biomedical knowledge graph based drug features and the molecular structure based drug features,and the feature fusion method based on self-attention mechanism performs better than traditional feature fusion methods.