Prediction Of Drug-drug Interactions Based On Tensor Decomposition And Deep Neural Networks

Drug-drug interaction（DDI）is critical to drug research and an important task in drug safety monitoring,providing effective and safe combination prescriptions for multi-ple drugs.The combination use of multiple drugs may improve efficacy or reduce adverse drug reactions on the one hand,but on the other hand,there are also adverse drug interac-tions and harmful side effects.Hence,predicting potential drug interactions is critical for physicians,patients,and society.Because prediction through experimental studies is of-ten complex,expensive,and time consuming,the development of computational methods to predict drug-drug interactions is urgently needed.However,most prediction methods only predict the existence of interactions between two drugs,and the type of interactions between drugs is relatively few.Therefore,a hybrid model of tensor decomposition with constraints and deep neural network is proposed in this thesis to predict multiple types of drug-drug interactions.Firstly,in this study,we calculated various drug similarities for different data sets by using different methods and information,such as Gauss kernel similarity,Minkows-ki distance similarity,based on chemical similarity,based on ligand chemical similarity,based on side effects similarity,based on anatomical,therapeutic,and Chemical（ATC）classification systems similarity,based on drug target similarity networkgenetic ontology（GO）semantic similarity,sequence similarity,protein-protein interaction（PPI）network distance.For making full use of various similarity information of drugs,the weight-ed average method was used for fusion.Then,a new tensor decomposition model with constraint conditions（TEDSEL）is constructed.Based on the traditional tensor decom-position model,the model takes the fused drug similarity information as the constraint condition of tensor decomposition,the Hessian and L_2,1regularization terms are added to prevent the model from overfitting.The model obtained Top1-precision 0.9759 under five-fold cross validation on the Drug DS3 dataset,compared with several other typical tensor decomposition methods,the performance is better.Finally,a hybrid model of tensor decomposition and deep neural network with con-straints（TDDNN）is proposed.The feature matrix decomposed by the tensor decompo-sition model is input to the deep neural network as input information to extract features from both linear and nonlinear perspectives,the Accuracy of the mixed model is 0.9957under the five-fold cross validation on the Drug Bank5.0 dataset,better performance than several other existing methods.In addition,we analyzed the related cases studies of Am-phetamine and Bromocriptine.For these two drugs,18 and 19 of the top 20 drugs with which the model predicted interaction have been confirmed in the relevant drug database,indicating that the prediction ability of the hybrid model of tensor decomposition and deep neural network is reliable.