Research On Prediction Of Drug-Target Interaction Based On Graph Neural Network

Drug-target interaction prediction often plays an important role in drug discovery and design.Using traditional biological experiments to predict the drug-target interaction is time-consuming and labor-intensive.Therefore,it is particularly important to develop suitable computational methods to speed up the progress of drug development and reduce the cost of drug development.In this thesis,drug-target interaction prediction is defined as a binary classification or link prediction problem.The following models are designed based on graph neural network for drug-target interaction prediction:1.In order to fully extract features from the expression profile of drugs and targets,graph convolution is used for the first time to extract the features of the drug and the target separately,and convolutional neural network is used to capture the latent association information of the drug-target pair.After concatenating the features,the classifier is used for prediction.2.Aiming at the problem that the drug and target similarity matrix in multi-omics data may cause information loss during the discretization process,a graph embedding strategy with weight priority-based random walk is proposed for the first time.First,it obtains multiple sets of walking sequences and the embedding representation of the nodes according to the weights,and then a classifier is used to predict the drug-target interaction.3.Aiming at the problem that it is difficult to make full use of the drug-target interaction relationships,a heterogeneous graph is constructed based on the drug and the target similarity matrices and the drug-target association matrix.Then,the graph convolution is designed on the heterogeneous graph,and the graph attention mechanism is introduced for the first time to obtain the meaningful node embedding.Finally the decoder is used to give the prediction results.According to the expression profiles and multi-omics data of the drugs and targets,three different models are proposed to solve three different problems of how to effectively extract the feature information of drug and target,how to avoid information loss in the discretization process,and how to make full use of the drug-target interactions.Different comparative experiments were done using real data with other models of the same type,and better experimental results were obtained.These works have enriched the in silico methods of drug-target interaction prediction,and also have a positive role in promoting drug development.