Research On Drug Relocation Methods Based On Network Representation Learning And Deep Learning

The development of a new drug is a very long and expensive process.Determining new disease indications for known drugs,i.e.,drug repositioning,helps to reduce the cost of drug development.At present,most methods for predicting drug-related diseases are to comprehensively use data related to drugs and diseases.However,although these methods focus on integrating the features of multiple drugs,they fail to take into account the diversity of various features.In addition,although these methods use various data related to drugs and diseases,they are all proposed based on shallow models,and it is difficult to dig out the complex relationship between drugs and diseases.These deficiencies will affect the performance of the prediction method to a certain extent.We propose three drug-disease association prediction methods,The first one is the association prediction method based on traditional network representation learning(non-negative matrix factorization),The second and third are prediction methods based on deep learning.(1)A prediction method based on the non-negative matrix factorization(Dive Pred)We propose a method based on traditional network representation learning(nonnegative matrix factorization),Dive Pred,to predict drug-related diseases.Dive Pred integrates disease similarity,drug-disease association,and a variety of drug features,including drug chemical structure features,drug target domain features,drug target annotation features,and drug-related disease features.The original drug feature has sparseness and high dimensionality,Dive Pred projects the drug feature to a lowdimensional feature space based on non-negative matrix factorization to obtain a dense feature representation of the drug.Since different drug features reflect the essence of the drug from different angles,Dive Pred uses an optimization item to enhance diversity and reduce the redundancy of multiple drug features.In addition,the algorithm uses Laplacian to combine neighbor information to improve the performance of the algorithm.The fivefold crossover experiment proved that Dive Pred is superior to several other advanced drug-disease association prediction methods.(2)A prediction method of drug-disease association based on multi-channel convolutional neural network(CAPred)The original information features are decomposed by non-negative matrix to achieve feature dimensionality reduction,and feature information is obtained through iterative fitting.This traditional characterization is difficult to capture a variety of complex and non-linear connections between drugs and diseases.Therefore,in this part we propose deep network representation learning and drug-disease association prediction methods(CAPred).First,four drug-disease heterogeneous networks were constructed based on four drugs similarity combining with disease similarity and drug-disease related information.From a biological perspective,four drug-disease node pair embedding matrices are constructed.A framework based on CNN is designed to capture node pair deep feature representations from four different drug-disease embedding matrices.In addition,since features from different sources have different contributions to the performance of prediction method,we propose a feature-level attention mechanism to distinguish the contributions of different features.The comparison results show that the prediction performance of CAPred is better than several other prediction methods.(3)A prediction method of drug-disease association based on fully connected autoencoder and convolutional neural network(ANPred)ANPred learns and integrates node pair attribute information and neighbor topology information from similar and related data of drugs and diseases.A learning framework based on multi-layer convolutional neural networks is designed to learn a pair of drug and disease node attribute representations from the drug and disease related data.In order to capture the neighbor topology of a node,a random walk strategy is established to form a sequence of neighbor nodes.The neighbor topology representation of the node is extracted,based on the joint framework of fully connected autoencoder and skip-gram.The cross-validation and the case of five drugs show that ANPred is not only superior to several new methods,but also more capable of discovering potential candidate diseases.