Drug Repositioning Based On Interpretable Knowledge Graphs

Drug repositioning is an important research direction in the field of biomedicine,with the goal of unlocking the potential of existing drugs to treat new indications,which can address the problems of the high cost and low efficiency in the development of new drugs.However,the traditional methods of drug repositioning are used to convert side effects of drugs,which were difficult to meet the needs for the market and researchers for drug prediction accuracy.With the advent of big data and machine learning technologies,several emerging technologies have been applied to the field of drug repositioning,and knowledge graph(KG)stands out with its excellent data processing capabilities.However,most of the existing methods of drug repositioning based on knowledge graph display a limited ability to predict and cannot provide the interpretation of the predicted results,while the interpretability of the results plays a significant role in guiding safely a risky task of drug repositioning.To address the above issues,based on biomedical KG,this paper investigates the accuracy of drug repositioning prediction tasks and the interpretability of prediction results by using graph convolutional neural networks,integrated learning and knowledge graph embedding(KGE)technology,etc.The main research content and the modest contribution of this paper are as follows.First of all,this paper uses six publicly available datasets and constructs a largescale comprehensive biomedical KG with 13 entity types and 107 relationship types using knowledge extraction and knowledge fusion techniques.Secondly,in this paper,an ensemble-based knowledge graph convolution algorithm named EKGCN is proposed for drug repositioning prediction.To begin with,this algorithm integrates the relations between entities into the representation of nodes,so as to obtain the embeddings of entities and relations with rich semantic information.Then,EKGCN utilizes the technology of model erosion to generate multiple models and uses the ensemble technique to process these models so that the integrated embeddings not only retains the inherent features of the models,but also reduces the feature noise in different models,which improves the prediction accuracy of the models.Finally,this paper compares the EKGCN algorithm with nine related methods.The experimental results show that the EKGCN algorithm improves MRR(M e a n R e c i p r o c a l R a t e)by 2.7% compared to the state-of-the-art methods,which fully proves the effectiveness of the EKGCN algorithm.Last but not the least,in this paper,an interpretable path analysis algorithm based on subgraph perturbation named KGExplainer is proposed to provide explanations of the prediction results.The algorithm contains a greedy search strategy that looks for the critical path and a subgraph-based embedding retraining strategy.The KGExplainer algorithm narrows the critical path search scope,and its computational complexity is independent of the size of the knowledge graph.Therefore,it can be applied to largescale knowledge graphs.Moreover,this paper compares the KGExplainer algorithm with four interpretable algorithms on two datasets,and its improves by 9.7% on ACC(Accuracy)compared to the state-of-the-art methods,which proves the effectiveness of this paper’s algorithm.