Graph Convolutional Autoencoder and Fully-Connected Autoencoder with Attention Mechanism Based Method for Predicting Drug-Disease Associations

Predicting novel uses for approved drugs helps in reducing the costs of drug development and facilitates the development process. Most of previous methods focused on the multi-source data related to drugs and diseases to predict the candidate associations between drugs and diseases. There are multip...

Celý popis

Uložené v:
Podrobná bibliografia
Vydané v:IEEE journal of biomedical and health informatics Ročník 25; číslo 5; s. 1793 - 1804
Hlavní autori: Xuan, Ping, Gao, Ling, Sheng, Nan, Zhang, Tiangang, Nakaguchi, Toshiya
Médium: Journal Article
Jazyk:English
Vydavateľské údaje: United States IEEE 01.05.2021
The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
Predmet:
Ablation
Artificial neural networks
attribute-level attention
Chemicals
Disease
Diseases
Drug development
Drugs
fully-connected autoencoder
graph convolutional autoencoder
Heterogeneous networks
Modules
multiple kinds of drug similarities
Network topologies
Network topology
Neural networks
Nodes
Prediction of drug-related diseases
Predictions
Proteins
Representations
Similarity
Topology
ISSN:2168-2194, 2168-2208, 2168-2208
On-line prístup:Získať plný text
Tagy: Pridať tag
Žiadne tagy, Buďte prvý, kto otaguje tento záznam!
Popis
Shrnutí:Predicting novel uses for approved drugs helps in reducing the costs of drug development and facilitates the development process. Most of previous methods focused on the multi-source data related to drugs and diseases to predict the candidate associations between drugs and diseases. There are multiple kinds of similarities between drugs, and these similarities reflect how similar two drugs are from the different views, whereas most of the previous methods failed to deeply integrate these similarities. In addition, the topology structures of the multiple drug-disease heterogeneous networks constructed by using the different kinds of drug similarities are not fully exploited. We therefore propose GFPred, a method based on a graph convolutional autoencoder and a fully-connected autoencoder with an attention mechanism, to predict drug-related diseases. GFPred integrates drug-disease associations, disease similarities, three kinds of drug similarities and attributes of the drug nodes. Three drug-disease heterogeneous networks are constructed based on the different kinds of drug similarities. We construct a graph convolutional autoencoder module, and integrate the attributes of the drug and disease nodes in each network to learn the topology representations of each drug node and disease node. As the different kinds of drug attributes contribute differently to the prediction of drug-disease associations, we construct an attribute-level attention mechanism. A fully-connected autoencoder module is established to learn the attribute representations of the drug and disease nodes. Finally, the original features of the drug-disease node pairs are also important auxiliary information for their association prediction. A combined strategy based on a convolutional neural network is proposed to fully integrate the topology representations, the attribute representations, and the original features of the drug-disease pairs. The ablation studies showed the contributions of data related to three types of drug attributes. Comparison with other methods confirmed that GFPred achieved better performance than several state-of-the-art prediction methods. In particular, case studies confirmed that GFPred is able to retrieve more actual drug-disease associations in the top k part of the prediction results. It is helpful for biologists to discover real associations by wet-lab experiments.
Bibliografia:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
content type line 23
ISSN:2168-2194
2168-2208
2168-2208
DOI:10.1109/JBHI.2020.3039502