Predicting Drug-Disease Associations via Meta-path Representation Learning based on Heterogeneous Information Net works

Identifying new indications of drugs plays an important role in the drug research and development process. However, traditional methods are labor-intensive and financially demanding to discover new indications. Computational methods are regarded as an effective way to predict underlying drug-disease associations (DDAs). Therefore it is a great urgent to develop computational-based methods to improve the accuracy of DDAs prediction. In this paper, a novel Meta-path Representation Learning-based model called MRLDDA is proposed to predict new DDAs on a heterogeneous information network (HIN). Specifically, MRLDDA first constructs a meta-path strategy based on rich HIN, i.e., drug-protein-disease-drug, and then the network representation of drugs and diseases is obtained by a heterogeneous representation model. Finally, a typical machine learning strategy-random forest classifier is applied to solve the prediction task of DDAs. Experimental results on the two benchmark datasets show that MRLDDA has a better prediction performance for the new DDAs under ten-fold cross-validation, with AUC of 0.8427 on B-Dataset and 0.9482 on F-Dataset.