A Metagraph-Based Model for Predicting Drug-Target Interaction on Heterogeneous Network

Determining drug-target interactions (DTIs) is an important task in drug discovery and drug relocalization. Currently, different models have been proposed to predict the potential interactions between drugs and targets. However, how to make full use of the information of drugs and targets to improve the prediction performance is still a great challenge. We define the problem of DTI prediction as a link prediction problem in a heterogeneous network and propose a new method, named MGDTI. The heterogeneous network includes known drug-target interactions and drug-drug and target-target similarity relationships. Firstly, we use the frequent subgraph mining algorithm to extract important metagraphs representing the network structure and semantic features without using domain knowledge and experience; then the matrix factorization method based on multiple commuting matrices is used to obtain the embedding representations of drugs and targets from multiple metagraphs; finally link prediction tasks are performed to predict the potential interactions between drugs and targets. We compare MGDTI with four classic heterogeneous network embedding methods and the experimental results show that MGDTI could achieve a better prediction performance.