Research On Drug Target Prediction Based On Knowledge Grap

Drug-target interactions are an important step in drug development,but the number of drug-target pairs passed through clinical trial studies only remains small,with shortcomings such as small experimental scale,high workload,and labor-intensive.Therefore,there is an urgent need for researchers to perform drug-target interaction prediction in an innovative way to solve the problems of time-consuming and inefficient drug development.With the continuous advancement of computer and other technologies,various drug databases are emerging,and drug-target prediction methods based on knowledge mapping can provide a priori knowledge for drug development and predict new drug-target interactions based on existing drug information,which greatly accelerates the drug development process.However,their reliance on a priori knowledge makes it extremely difficult to predict new drug or new target interactions,and most of the current models that can predict new drugs do not combine existing drug and existing target information with unknown drug and target information,ignoring a large amount of other potential information.Meanwhile,the drugtarget prediction models based on knowledge graphs rely on complete biomedical knowledge graphs,which are mostly lagging and slow to update.Therefore,this paper focuses on the dynamic update of knowledge graph and drug-target interaction prediction based on knowledge graph,and the main contents include:(1)To address the problem of lagging biomedical knowledge map update in existing knowledge map prediction models,a dynamic knowledge map update method based on the Unified Medical Language System(UMLS)and large-scale network embedding algorithm is proposed on the basis of existing knowledge maps.The method utilizes the UMLS knowledge base to add biomedical entities to the triad through a hyper-syntactic lexicon,and uses expert dictionaries to bring together different types of entities such as nouns,verbs,and adjectives of each vocabulary to supplement the entity part of the triad.Secondly,semantic relations among biomedical entities are extracted from text sentences by Sem Rep,and semantic predicates are used as relations to supplement the triad.Finally,based on the existing knowledge graph,a large-scale network embedding method is used to calculate the connection probability between the nodes of unrelated entities and other nodes for the entities in the database from which relationships cannot be extracted from the text,so as to supplement the relationships between entities.The experimental results show that the method reduces 8% redundant nodes on Yamanishi＿08 dataset,5% redundant nodes on KEGG dataset,and 7% less redundant nodes on Drug Bank dataset,and the feasibility of the method is verified in the example.(2)To address the problem that existing drug-target interaction models are difficult to predict the interaction relationship between under-studied drugs and targets and do not sufficiently consider prior knowledge,a drug-target prediction model based on knowledge mapping and a dual network logistic matrix decomposition method is proposed,in which all input data sets become in the form of a triad,the Rotat E model is trained,and the embedding is optimized using a negative sampling loss function.Finally,the interaction adjacency matrix is generated by the scoring function so that the prior knowledge can be integrated into the prediction model.Finally,a prediction model based on the dual network logistic matrix decomposition is designed to predict the interaction relationships of understudied new drugs or new targets,and the potential variable matrix and the final diffusion matrix are integrated into the logistic function to obtain the interaction probabilities between new drugs and targets.The experimental results show that the method converges faster than the baseline model as the number of samples increases,and the AUC value of the method exceeds 0.96 when the training samples are increased to 1000.in the best case,the AUPR is 0.912 and the AUC is 0.986.and the optimal combination of the method is verified by experiments with multiple combinations of knowledge graph models and new drug prediction models.