BioWalk-MDA: a novel approach for large-scale predicting metabolite-drug associations based on multi layered biomedical knowledge graphs

Metabolism is fundamental to sustaining human life, with changes in metabolite levels closely related to the occurrence and progression of diseases. The interaction between metabolites and drugs is intricate, encompassing drugs can modulate metabolite concentrations, as well as the metabolites generated through drug metabolism can influence pharmacological toxicity and drug interactions. Currently, a substantial proportion of metabolite-drug associations remains to be fully elucidated, and the data from individual sources are often characterized by incompleteness and noise. Here, we present BioWalk-MDA, a computational framework for large-scale predicting novel interactions between 8354 metabolites and 11 570 drugs. The framework constructs multilayered biomedical knowledge graphs (Multi-BiomedKGs) by integrating biological information across proteins, microbes, and diseases, and incorporated five types of graphs and seven types of associations. It employed random walk and heterogeneous Skip-gram model to extract feature vectors of metabolite-drug pairs and utilized a fully connected neural network (FCNN) to infer novel metabolite-drug associations. The framework demonstrated exceptional performance with an average accuracy of 0.971, an area under the receiver operating characteristic curve (AUROC) value of 0.995, and an area under the precision-recall curve (AUPRC) value of 0.994 in 5-fold cross-validation, surpassing other similar methods. Case studies on three metabolites detectable in blood and three cardiovascular drugs further demonstrated the reliability and efficiency of BioWalk-MDA, and it is anticipated to serve as a valuable tool for exploring metabolite-drug interactions and aiding in drug development and combination strategies.