Research On Prediction Method Of Drug-Target Based On Heterogeneous Graph Attention Network

Identifying the complex interactions between drugs and targets is the key to drug discovery and drug repositioning,and it usually plays a crucial role in drug development and evaluation of drug effectiveness and safety.Traditional biological experiment methods to identify drug-target interactions are often costly and time-consuming.Therefore,the key to the development of new drugs and drug repositioning is how to accurately and quickly identify drug-target interactions.As a powerful graph representation learning technology based on deep learning,graph neural network can mine large-scale graph data and greatly improve the performance of many network-related prediction tasks,which has aroused extensive research interest.However,heterogeneous graphs containing different types of nodes and links have not been fully considered in graph neural networks.At the same time,the great potential of the attention mechanism in deep learning has been fully demonstrated in various fields.First,this paper proposes a hierarchical heterogeneous graph attention data representation learning method to learn the feature representation of drugs and target nodes in a drug-target heterogeneous network.Second,combining PU learning and inductive matrix completion to predict potential drug-target interactions.Specifically,node-level attention aims to learn the importance between nodes and their neighbors based on meta-paths.Semantic attention can learn the importance of different meta-paths.With the learned importance from both node-level and semantic-level,the importance of nodes and meta-paths can be fully considered.Then,the proposed model can generate node embeddings by aggregating features from neighbors based on meta-paths in a hierarchical manner.Considering the PU learning problem in the biological association prediction problem,this paper adopts the processing method of taking positive loss and non-labeled samples to take different loss weights in the loss function and combining the induction matrix completion method to propose learning goals and improve prediction ability.The characteristics of drugs and target nodes in the heterogeneous network learned through the model in this paper integrate various information in the drug-target heterogeneous network,and automatically learn the rich structural information and semantic information in the network.It can promote the prediction of drug-target interactions.The method proposed in this paper is compared in various experiments.Compared with some other existing methods for predicting drug-target interaction,he experimental results verify that the model is more efficient and effective.At the same time,further experiments prove that the model proposed in this paper can predict the unknown drugtarget interaction relationship.These results indicate that the calculation method model proposed in this paper can provide strong support for drug development and drug relocation.