Research On Drug-Target Interaction Prediction Based On Matrix Completion

Drug research and development is an important mean to promote the steady development of social health.The identification of drugs and target proteins is the key to the development of modern new drugs.Over the past few decades,many biological experiments have been conducted to determine the interactions between drugs and targets,but they have been overcome by three major obstacles: long development cycles,high cost,low effectiveness,and large financial and material requirements.With the rapid development of big data processing technology,intelligent computing emerges at a historic moment.Researchers can solve the shortcomings of traditional methods such as long research and development time,large capital and labor costs,and low success rate of drug research and development by computer simulation,calculation and budget of the relationship between drugs and target proteins,and integrate and extract the characteristic data of drugs and target proteins.It is of great significance to explore the interaction between drug molecules and target proteins by computer simulation for the development of new drugs and the improvement of human medical treatment.In this thesis,methods based on matrix completion and the deep learning correlation model were used to classify and predict drug-target interactions.Drug-target interaction prediction was regarded as a matrix completion task to complete the drug target interaction matrix,and drug compound molecules and target protein sequences were numerically characterized respectively.The powerful graph neural network model is used to conduct correlation prediction for drug targets,and the missing value in the original interaction matrix is finally completed.The proposed method is as follows:(1)A drug-target interaction prediction method GCMCDTI based on graph convolution matrix completion was proposed.The structural framework of the graph encoder was used to comprehensively consider the structural information between the drug and the target as well as the respective edge information of the drug and the target.The heterogeneous nodes in the drug-target bipartite graph were encoded by the graph encoder framework to obtain the drug and target characteristics after embedding.A bilinear decoder was used to reconstruct the drug target interaction matrix and complete the missing values,to predict the potential drug target interaction.The results showed that the AUC values of GCMCDTI model were 95.78%,95.31%,93.90% and 91.77%in four data sets of the enzyme,GPCR,ion channel and nuclear receptor,and the AUPR values were 94.50%,93.26%,92.02% and 95.68% respectively.(2)A drug-target interaction prediction method RMCDTI based on a relational graph neural network completion matrix was proposed.Based on the prediction method of drug-target interaction based on graph convolution matrix completion,a method different from the direct inference method is adopted,which does not need to take the entire interaction matrix as input,but to infer whether a single drug-target pair is associated.In this method,closed subgraphs about target drug-target pairs were extracted from the drug-target interaction matrix in an inductive way,and the relational graph convolutional network was trained to map the subgraphs to the predicted results.The results showed that the AUC values of RMCDTI model in four data sets of the enzyme,GPCR,ion channel and nuclear receptor were 96.97%,91.27%,95.41% and86.68%,and the AUPR values were 97.37%,92.18%,95.75% and 88.36%,respectively.