Research On Drug Interaction Prediction Method Based On Knowledge Graph Interpretability

The combination of multiple drugs may cause drug-drug interaction(DDI).DDI is a doubleedged sword.A positive DDI will enhance the curative effect,while a negative DDI will weaken the efficacy of the drug,causing Drug-Drug Adverse reactions(DDADR)and even life-threatening.Also,as the number of approved drugs increases,the likelihood of drug-drug interactions increases.Therefore,how to effectively predict DDI has become an urgent and challenging task in clinical practice.Although in vivo and in vitro wet experiments can discover potential DDIs,they require long experimental periods and consume many research resources.At present,computing methods based on machine learning and deep learning have been widely used in DDI prediction tasks,and the continuous enrichment of drug data resources has also improved the performance of these methods.However,DDI prediction tasks still face many challenges:(1)Most of the DDI prediction methods based on machine learning and deep learning are data-driven,limited by data sparsity and cold start issues,overly rely on labeled data but ignore a large amount of unlabeled data;(2)Existing data-driven DDI prediction methods cannot integrate rich multi-modal information of the drug,such as the chemical structure information of the drug,thus making the feature representation semantics of the drug wizened and further the DDI prediction results unsatisfactory.At the same time,these methods lack the interpretability of the results,which is critical for medicine domain;(3)The knowledge-driven DDI prediction methods are limited by the traditional knowledge graph embedding method and training mode,thus making the obtained drug feature representation quality not good enough.It is necessary for the knowledge-driven DDI prediction method to obtain a more discriminative drug feature representation,thereby improving the DDI prediction effect.In response to the above problems,this thesis deeply studies the DDI prediction problem and key technologies,and proposes a series of DDI prediction methods integrated with knowledge graphs.The main contribution of this thesis are as follows:(1)Proposing a drug-drug interaction prediction method based on structure-aware and multiview consensus-constrained graph contrastive representation(SAMVC-GCL)to solve the problem of sparse data and over-reliance on labeled data.Firstly,drug-related biomedical information and drug substructure information are used to construct a drug knowledge graph,and the information flow theory is applied to disseminate information on the drug knowledge graph to generate a drugdrug interaction network;secondly,a multimodal graph convolution network(GCN)is applied to the DDI networks which is trained by contrastive learning and multi-view consensus constraint graphs to obtain discriminative drug feature representations;finally,multi-source drug features are fused with the attention mechanism,and fed to the DDI prediction module.(2)Presenting a drug-drug interaction prediction method based on a knowledge graph and multiple attention coding(KGMAC)to solve the problem that current DDI prediction methods lack of the interpretability for DDI prediction results.Firstly,we construct a chemical element knowledge graph which is initialized by knowledge graph embedding technology,and then,it is augmented into the molecular drug structure to generate a drug molecule enhanced graph with domain knowledge;after that,the medicinal chemistry is fused by a multi-attention encoder.The topological information of the structure and the domain information in the knowledge graph can generate expressive drug feature representations which are fed to the next DDI prediction task.This work is also the basis for further optimization research.(3)Proposing a drug-drug interaction prediction method based on a knowledge graph and dual-channel contrastive learning(KGDCL)to solve the problem that the knowledge-driven DDI prediction method is limited by the traditional knowledge graph embedding method and training mode,thus making the learned features representations low-quality.First,we encode the dualchannel data by the GCN encoders and the multi-attention encoders,and obtain the feature representations of the drug from the perspective of the knowledge-driven molecular drug structure diagram and the original drug molecular structure diagram;after that,it is generated by hard negative sample mining.Negative samples,compared with random addition and deletion of nodes,are more conducive to get the high-quality negative samples and contrastive training;finally,training through dual-channel contrastive learning further improves the quality of drug feature representation,thereby improving DDI prediction accuracy.This thesis conducts extensive experiments on TWOSIDES,Drug Bank,SIDER,OFFSIDES,and Pub Chem datasets and uses multiple indicators for evaluation.The experimental results show that the series of models proposed in this thesis is superior to some of the best models in the DDI prediction task in terms of accuracy and interpretability,and realize the interpretable prediction of DDI.Our methods will help to formulate more effective drug treatment solutions and improve the efficiency of new drug development.