Research On Drug Repositioning Method Based On Multi-level Information Integration

The development of new drugs is an extremely time-consuming and costly process with a high failure rate.Therefore,identifying new indications for existing drugs(drug repositioning)can effectively facilitate drug development efficiency and reduce the cost of experimental development.However,how to predict the association between drugs and diseases more accurately and efficiently is the key element of our research.Currently,available drug repositioning methods focus on multiple sources of data about drugs and diseases and do not consider the connection between different levels of information.In this paper,we learn and construct models by targeting different levels of information about drugs and diseases,which include node attribute information at the drug and disease node pair level,global topology information at the heterogeneous network level,and integration of neighbor topology information at the heterogeneous network level and attribute information at the node pair level.We propose three deep learning models for drug-disease association prediction based on different levels of information,respectively.Comparison experiments with previous methods show that all three of our proposed methods achieve the best prediction performance.Our main tasks are as follows:(1)Drug-disease association prediction method based on convolutional autoencoderIn this part of the study,a convolutional autoencoder-based prediction method(PAPred)is proposed for drug and disease node pair level node attribute information to extract and integrate multiple similarities and associations and encode the embedded representation to predict drug-related candidate diseases.First,three drug networks and a disease network were constructed using drug-related chemical substructures,structural domain and protein gene ontology data of proteins,and disease signature data,respectively.The attribute embedding strategy for node pairs was obtained by extracting node features from multiple drug networks,disease networks and drugdisease association information based on attention at the similarity level and embedding them based on biological prerequisites related to drugs and diseases.Similarity-level attention can assign corresponding weights to feature vectors of nodes in multiple drug networks adaptively to effectively fuse multiple similarity features of drug nodes.Finally,the attribute encoding of the node pairs is obtained by convolutional autoencoder learning and a classifier is used to obtain scores for drug-disease association prediction.Experimental results on a public dataset show that PAPred performs better than several other newer drug-disease association prediction models.In addition,a five-drug case study further validated PAPred's ability to identify promising drug-related disease candidates.(2)A method for predicting drug-related diseases based on multiplexed graph convolutional neural networksIn this section,global topological information at the level of heterogeneous nets of drug and disease nodes is investigated and a drug-related disease prediction model based on a multiplexed graph convolutional neural network,GTPred,is proposed.First,three drug-disease heterogeneous networks were constructed using drug networks and drug-disease association information built from three different perspectives together with disease networks,respectively.Then,multiplexed graph convolutional neural network models are built while extracting the global topological representation matrices of different drug and disease nodes from different drug-disease heterogeneous networks,respectively.In addition,based on the different contributions of global topological representations from different sources to drug-disease association prediction,this paper also proposes an attention mechanism at the global topological level to learn the augmentation matrix of global topological representations of drug and disease nodes.Finally,a joint training approach was utilized to keep the learning of the embedding representation of the model consistent,and the classifier predicted potential drugrelated candidate diseases.The results of the comparison experiments show that GTPred has superior performance in terms of AUC and AUPR than several other comparative methods and will be more accurate in predicting drug-related diseases.In addition,case studies in five drugs further demonstrate GTPred's ability to identify promising drug-related disease candidates.(3)Integration of neighbor topology based on meta-paths and node attributes for drug-disease association prediction methodsIn this section,this paper addresses the integration of neighbor topology information at the heterogeneous network level and attribute information at the nodepair level,leading to the proposal of a dual-attention mechanism based on meta-paths and a two-branch model prediction method consisting of a model based on convolutional neural networks in conjunction with a model based on convolutional autoencoders,NAPred.First,three drug-disease heterogeneous networks were constructed using the same strategy as GTPred,using similarities and associations extracted from different perspectives related to drugs and diseases.We then extracted two levels of representations between pairs of drugs and diseases,including a metapath level neighbor topology representation that learns the topological structure relationships between multiple neighbors,and a node-pair attribute representation that reveals the potential relationships between common drug and disease-related data sources.Neighbor topology representations were obtained from multiple drug-disease heterogeneous networks via multiscale meta-paths,fully connected neural networks,and dual attention mechanisms,respectively,and were encoded using convolutional neural networks to represent the implied neighbor topology relationships between pairs of drug-disease.New attention mechanisms at the neighbor scale level and at the neighbor topology level to distinguish the different contributions of neighbor features and topologies at different scales to drug-disease prediction,respectively.The nodepair attribute representations are obtained from multiple drug-disease heterogeneous networks using an embedding strategy,and then the attribute information of the drugdisease pairs is encoded by convolutional autoencoder modeling.Finally,the final drugdisease association prediction scores were obtained by weighted summation of the prediction scores from the two branches.The experimental results showed that NAPred achieved the best performance across multiple assessment criteria and case study analyses,outperforming six state-of-the-art drug-disease prediction models.