A Multi-graph Deep Learning Model for Predicting Drug-Disease Associations

Computational drug repositioning is essential in drug discovery and development. The previous methods basically utilized matrix calculation. Although they had certain effects, they failed to treat drug-disease associations as a graph structure and could not find out more in-depth features of drugs and diseases. In this paper, we propose a model based on multi-graph deep learning to predict unknown drug-disease associations. More specifically, the known relationships between drugs and diseases are learned by two graph deep learning methods. Graph attention network is applied to learn the local structure information of nodes and graph embedding is exploited to learn the global structure information of nodes. Finally, Gradient Boosting Decision Tree is used to combine the two characteristics for training. The experiment results reveal that the AUC is 0.9625 under the ten-fold cross-validation. The proposed model has excellent classification and prediction ability.