Predicting\nDrug–Target Interactions Using Probabilistic\nMatrix Factorization

Abstract

Quantitative analysis of known drug–target\ninteractions\nemerged in recent years as a useful approach for drug repurposing\nand assessing side effects. In the present study, we present a method\nthat uses probabilistic matrix factorization (PMF) for this purpose,\nwhich is particularly useful for analyzing large interaction networks.\nDrugBank drugs clustered based on PMF latent variables show phenotypic\nsimilarity even in the absence of 3D shape similarity. Benchmarking\ncomputations show that the method outperforms those recently introduced\nprovided that the input data set of known interactions is sufficiently\nlargewhich is the case for enzymes and ion channels, but not\nfor G-protein coupled receptors (GPCRs) and nuclear receptors. Runs\nperformed on DrugBank after hiding 70% of known interactions show\nthat, on average, 88 of the top 100 predictions hit the hidden interactions.\nDe novo predictions permit us to identify new potential interactions.\nDrug–target pairs implicated in neurobiological disorders are\noverrepresented among de novo predictions.