Drug-Disease Association Prediction Method Based On Multi-View Information Fusion

Drug repositioning is an emerging approach in drug research to identify new therapeutic potentials of approved drugs and discover drug treatment options for untreated diseases.Advances in genomics,and the tremendous growth of biological data,have further boosted the performance of computational drug repositioning methods.In particular,deep learning-based drug-disease association prediction methods have made the field of computational drug repositioning an active field.Therefore,it is of great significance to develop deep learning-based computational methods to quickly and accurately determine the relationship between drugs and diseases.In this thesis,a drug-disease association prediction method based on multi-view information fusion is carried out,that is,a multi-view data fusion model(Multiple Views Fusion,MVF)and an association prediction scoring model based on inductive matrix completion(Inductive Matrix Completion,IMC)end-to-end.A method of end-to-end training to achieve drug-disease association prediction.MVF-IMC consists of two parts:(1)Multi-view data fusion(MVF)module.MVF first obtains the information of multiple perspectives of the drug(disease),that is,the similarity matrix of multiple perspectives of the drug(disease),and through the deep learning method,the multiple similarity matrices of the drug(disease)are fused to output the drug(disease)fusion matrix,The drug(disease)fusion matrix supplements the multi-view data of the drug(disease)before fusion,so it has more complete drug(disease)feature information.(2)Inductive matrix completion(IMC)prediction scoring module.After multi-view data fusion,the drug(disease)fusion matrix is input into the induction matrix completion method as a feature matrix.During the training process,the corresponding drug fusion completion matrix(drug parameter matrix)and disease fusion completion matrix(disease parameter matrix)are updated,and finally a drug-disease association prediction matrix is generated.MVFIMC is a model architecture that is trained end-to-end through the combination of a deep learning-based multi-view data fusion model and an association prediction scoring model based on an inductive matrix completion method.This model architecture is important for studying drug-disease association prediction problems.meaning.In this thesis,four datasets(Cdataset,LRRSL,PREDICT,SND)are compared with the existing methods.Under the same experimental scheme,8 accuracy,specificity,precision,recall,f1＿score,MCC,AUC and AUPR are used.The evaluation indicators are better than the existing methods.In addition,based on these four datasets,this thesis conducted a case analysis for two different cases(breast tumor and lung tumor),selected the top 20 ranked drug candidates for breast tumor and lung tumor,and integrated the four data The top 10 drug-disease association pairs were selected and the accuracy of their prediction results was verified.