Predicting human microbe-drug associations via graph attention network with multiple kernel fusion

Microbial dysregulation may lead to the occurrence of diseases, and using microbe and drug data to infer the microbe-drug association has attracted extensive attention. There have been many studies to build association prediction models, but most of them are through biological experiments, which are time-consuming and expensive. Therefore, it is necessary to develop a computational method focusing on microbe-drug spatial information to predict the microbe-drug associations. In this work, we use the biological information to construct heterogeneous networks of drugs and microbes. We propose a new method based on Multiple Kernel fusion on Graph Attention Network (GAT) to predict human microbe-drug associations, called GATMDA. Our method extracts multi-layer features based on GAT which can learn the embedding of microbes and drugs on each layer and achieve the purpose of extracting multiple information. We further fuse multiple kernel matrices based on average weighting method. Finally, combined kernel in the microbe space and drug space are used to infer microbe-drug associations. Compared with eight state-of-the-art methods, our method receives the highest AUC and AUPR on MDAD and aBiofilm dataset. Case studies for Human immunodeficiency virus 1 (HIV-1) and adenovirus further confirm the effectiveness of GATMDA in identifying potential microbe-drug associations.