Research On The Association Prediction Between Drugs And Diseases Based On The Integration Of Multiple Heterogeneous Network Informatio

Identifying new uses of approved drugs(drug repositioning)is an effective way to reduce drug development time and cost.Predicting potential drug-disease associations allows biologists to screen more reliable candidates,thereby shortening the drug development cycle.So how to effectively predict reliable candidates is a topic worthy of further study in the field of biology.Recent drug-disease association prediction methods have focused on integrating multi-modal data for drugs and diseases.In this paper,we use a deep learning approach to merge different levels of information about drugs and diseases,including multi-perspective property information of drugs and diseases,multi-scale neighbor topology information in drug-disease heterogeneous networks,and rich semantic information.In this paper,combining multi-modal data of drugs and diseases,three models are proposed to predict drug-disease associations.A drug-disease association prediction model(MVpred)with multi-receptive field convolutional neural networks(CNN)is proposed based on the rich associations and connections embedded in drug and disease node pairs.First,a node pair embedding strategy is proposed to fuse many similarities and connections between drugs and diseases under different biological premises.Then,different sizes of receptive fields are used to capture the attribute information of different perspectives in the node pairs.Finally,the attribute information from different perspectives is fused by convolutional neural networks and association prediction is performed.The ability of MVpred to predict potential candidates is further demonstrated by comparative experiments and a case study of five drugs.In view of the rich semantic information in drug-disease heterogeneous networks,we propose a fully connected autoencoder association prediction model(NSpred)with node-level attention and semantic-level attention mechanisms.Firstly,three heterogeneous networks were constructed based on similarity of chemical substructure,similarity of target domain and similarity of target annotation.The neighbor set of drug(disease)node is constructed based on meta path.Since the significance of different neighbouring nodes varies,we propose node-level attention to distinguish the significance of different neighbours.The contribution of different meta-paths to correlation prediction is different.We propose that semantic-level attention should be assigned different weights to different meta-paths.Then the learned node attribute representation containing rich semantic information is fed into the fully connected automatic encoder to obtain the low-dimensional semantic representation.Finally,a fully connected neural network is used to obtain the final predicted score.It was demonstrated that both AUROC and AUPRC of NSpred significantly outperformed the other six drug-disease association prediction methods.Aiming at the multi-scale neighbor topology information and multi-dimensional similarity information of drugs and diseases in heterogeneous networks,we proposed a prediction model MGpred based on convolutional autoencoder and graph convolutional autoencoder.First,we constructed three heterogeneous networks based on multiple kinds of drug similarities.Each network comprises drug and disease nodes and edges created based on node-wise similarities and associations that reflect specific topological structures.We also propose an embedding mechanism to formulate topologies that cover different ranges of neighbors.To encode the embeddings and derive multi-scale neighboring topology representations of drug and disease nodes,we propose a module based on graph convolutional autoencoders with shared parameters for each heterogeneous network.We also propose scale-level attention to obtain an adaptive fusion of informative topological representations at different scales.To integrate the multi-dimensional similarity of drugs,we propose a similarity-level attention mechanism that assigns different weights to different levels of similarity.The properties of drugs and diseases are then captured by convolutional autoencoders.Comprehensive experiment results demonstrate that MGpred outperforms other state-of-art methods in comparison to drug-related disease prediction,and the recall rates for the top-ranked candidates and case studies on five drugs further demonstrate the ability of MGpred to retrieve potential drug-disease associations.