Drug-disease Association Prediction Research For Multi-source Data Integration

Predicting new uses for approved drugs can help reduce drug development costs and facilitate the development process.Most previous approaches focus on drug-related and disease-related multi-source data to predict candidate associations between drugs and diseases.Drugs have a variety of informative attribute information that is beneficial for drug-disease pair association prediction,which are mostly ignored by previous methods.According to different attributes,multiple similarities of drugs be calculated.Most of the previous methods failed to deeply integrate these similarities.In addition,the topology of multiple drug-disease heterogeneous networks constructed with different drug similarities is also underutilized.Drug subnets and a disease subnet included in heterogeneous networks respectively cover the common topology information between drug and disease nodes,the specific information between drug nodes and the specific information between disease nodes.This project proposes three prediction methods for predicting drug-disease associations:（1）Prediction method GAPred based on convolutional autoencodersA prediction method based on convolutional autoencoders,GAPred,for association prediction.GAPred integrates a disease similarity,a drug-disease association,and three drug similarities（including drug chemical substructure attributes,drug target domain attributes,and drug target annotation attributes）.According to the proposed feature construction strategy,GAPred constructs three attribute embeddings of drug-disease pairs.Each attribute embeddings are separately learned by convolutional autoencoders to obtain information-dense attribute representations of node pairs in a low-dimensional feature space.GAPred integrates the three attribute representations of node pairs through an attribute-level attention mechanism,and finally makes predictions.Experimental results and case studies confirm the improvement of GAPred performance.（2）Prediction method GFPred based on graph convolutional autoencoders and fully connected autoencodersThe GFPred method is constructed to learn and integrate topological representations and node attribute representations of multiple heterogeneous networks.First,GFPred builds drug-disease heterogeneous networks with multiple drug similarities that reflect the degree of similarity between drugs at different levels.GFPred learns the complex topology of drug and disease nodes in three networks through a graph convolutional autoencoder framework with an attention mechanism,and finally obtains the lowdimensional topological representation of the nodes.In addition,GFPred learns multiple attribute information of drugs and diseases through a fully connected autoencoder framework with an attribute-level attention mechanism to mine the attribute representations of nodes.Finally,GFPred deeply fuses topological representations,attribute representations,and original features of drug-disease node pairs through a convolutional neural network-based integration strategy.（3）Prediction method CTST based on graph convolutional autoencoders with parameter sharingCTST model is designed for learning and integrating common and specific topological representations in heterogeneous networks and subnetworks.First,CTST builds heterogeneous networks containing multiple drug subnetworks and one disease subnetwork,which cover the similarities and associations between drug and disease nodes.CTST proposes a graph convolutional autoencoder with parameter sharing to facilitate the extraction of common representations of drug and disease nodes in heterogeneous networks.A common-representation-level attention mechanism（C-Attention）is proposed to distinguish the contributions of nodes’ common representations to association prediction and fuse them adaptively.In order to learn the topology and attributes of nodes in subnets,CTST proposes a specific graph convolutional autoencoder with specificrepresentation-level attention mechanism to capture specific representations of nodes.CTST finally passes through an integrated module with a representational feature-level attention mechanism to achieve the final association prediction.