Application Of Graph Neural Network Based On Graph Structure Learning To Drug Target Prediction

Drug development is a laborious,time-consuming,and costly task.Despite the significant amount of funding that has been invested in drug development over the past few decades,it still cannot meet the current demand for new drugs.Drug target prediction,as a very important method of drug development,can greatly shorten the drug development cycle and reduce development costs by predicting the potential targets of a compound.In addition,drug target prediction also helps to study the mechanism of drug action and understand the relationship between drugs and diseases.This information is very important for selecting appropriate drugs to treat a certain disease,designing more effective drug molecules,and improving the success rate of drug development.With the development of artificial intelligence,researchers have also started to combine deep learning algorithm models to predict drug-target relationships.As research deepens,people have a deeper understanding of the mechanism of drug action and find that there is a great correlation between drugs and targets,but this data structure is different from traditional Euclidean space data,but is a graph-structured data.Traditional neural networks perform well on Euclidean space data problems,but are severely restricted in processing graph-structured data features.Therefore,graph neural networks have been proposed as the optimal framework for data analysis of graph-structured data,which can effectively solve the problems in drug target prediction.However,traditional homogeneous graph structure learning methods mainly model and predict isomorphic graphs,that is,the connection relationships between nodes are the same,and in this case,the features and connection relationships of nodes can be represented by a fixed vector.But in the field of drug target prediction,many graphs are not isomorphic graphs,and the relationships and attributes between nodes may be heterogeneous,and different connection relationships may have different meanings.Traditional homogeneous graph structure learning methods ignore the differences in heterogeneity,and therefore cannot adapt well to heterogeneous graph structures.In addition,traditional homogeneous graph structure learning methods often only focus on the connection relationships between nodes and ignore the attribute information of nodes,which may lead to information loss or errors in modeling and prediction,resulting in a decrease in the prediction accuracy of the model.To address these issues,this paper proposes a new graph neural network drug target prediction model based on graph structure learning(HGDTA).In this model,the original relationship subgraph,node features,and semantic embeddings are first used as inputs to generate relationship subgraphs separately.Then,by learning the feature graph and semantic graph,and fusing them with the original graph,the learned relationship subgraphs are obtained,and finally,these relationship subgraphs are fused into a heterogeneous graph.To better represent node features,the model introduces the Poincaré ball model in hyperbolic space and uses node embeddings based on meta-paths to construct a semantic embedding matrix.Finally,by jointly training the heterogeneous graph structure and GNN parameters,the model can avoid most of the assumptions of the neighbor based on the original heterogeneous structure in HGNN,and improve the accuracy and portability of the model prediction.Experiments on drug target-related datasets show that HGDTA can effectively predict the relationship between drugs and targets,with high accuracy and portability.At the same time,this paper also sets up ablation experiments and finally verifies the effectiveness of the feature graph generator and semantic graph generator.