Research On Drug-disease Association Prediction Based On Heterogeneous Network Informatio

Drug development suffers from high time-consuming and costly problems.Therefore,drug repositioning can effectively reduce R&D costs and uncover the potential value of approved drugs.Currently,most computational approaches focus on building drugdisease heterogeneous networks to predict candidate associations between drugs and diseases by constructing graph inference models.However,these methods do not deeply integrate the topology and unique features of the learned drug and disease nodes,nor do they perform attribute complementation for disease nodes with missing attributes based on the topological information of the heterogeneous network.In order to make full use of the heterogeneous network information between drugs and diseases,this paper proposes three deep learning-based association prediction methods.The results show that the performance of the three association prediction methods is improved the main work accomplished is as follows:(1)Prediction of drug-disease associations based on attribute complementationPrevious approaches did not perform attribute complementation of disease nodes with missing attributes based on topological information of heterogeneous networks.We propose a new prediction method,FACDDP,to encode and integrate the semantics of multiple meta-paths and learn to obtain topological embeddings of drug and disease nodes.The topological relationships between nodes are used as a guide to weighted aggregation of drug node attributes with attributes to complement disease nodes without attributes.We designed a meta-path level and a neighbor level attention mechanism to fuse semantic information from multiple meta-paths and information from node neighbors,respectively.Five-fold cross-validation was used to evaluate the performance of the prediction model.Also,five different drugs were used as case studies to validate the proposed model in identifying potential candidate diseases.(2)Predicting drug-related diseases based on multilayer convolutional neural networksPrevious convolutional neural network models have mainly focused on learning deep information,ignoring the importance of detailed features learned by shallow networks for drug-disease association prediction.To this end,in this section,we propose a new multilayer convolutional neural network-based prediction model,DRCNDDP,to predict drug-related diseases.We use drug and disease data from different sources,which include information on similarities between drugs,similarities between diseases,and associations between drugs and diseases,to build a matrix of attributes of drug-disease node pairs.For these attributes,we use a multilayer convolutional neural network to deeply integrate the detailed attribute features of the node pairs learned by the shallow network and the representative attribute features learned by the deep network to obtain a richer low-dimensional feature representation of the node pairs.The method exhibits better performance than other compared methods,specifically in terms of AUC and AUPR.This indicates that our proposed DRCNDDP method has good potential for drugdisease association prediction.In addition,the ability of DRCNDDP in mining potential drugs with related candidate diseases was confirmed by a case study.(3)Drug-disease association prediction based on topology and unique features of nodes in fused heterogeneous networksModels with multiple graph convolutional coding layers or graph neural coding layers can easily learn the transition smoothing features of drug and disease nodes.To this end,this paper proposes a new prediction method,GFDDA,for encoding features specific to drug and disease nodes in multiple heterogeneous networks,as well as nodepair properties.First,by constructing similarities for different kinds of drugs,we build three drug-disease heterogeneous networks to facilitate the integration of various drugand disease-related data.Since each heterogeneous network has its own specific topology and node features,we construct a separate graph convolutional autoencoder with feature complement for each network to learn the topological representation of the nodes.To alleviate the over-smoothing problem during node feature learning,we learn node features specific to each network and add them to each encoding layer of the graph convolutional network.Finally,we also designed attention mechanisms at the representation level and feature type level to obtain more informative representations and feature types.In this paper,we also conduct ablation experiments to explore the contribution of multiple drug attributes and feature complements to improve performance.The experimental results on a public dataset show that GFDDA achieves better performance than several other comparative prediction methods.In addition,GFDDA was found to have the ability to discover potential drug-related disease candidates based on a case study of five drugs.