A total of 100 years ago, Paul Ehrlich, a Nobel Laureate in physiology and medicine, proposed the concept of 'magic bullets' -that is, chemical compounds precisely interacting with a specific receptor. For decades, the 'one drugone target -one disease' paradigm dictated much of the drug development process. However, in the past 10 years, tremendous advances in transcriptomics and genomics research have shifted this simplistic view of mechanism of action (MoA) to a more complex systems pharmacology paradigm where a compound can bind to several targets.Experimental elucidation of drug MoA constitutes a critical element in the early preclinical drug development process. This step is labor intensive, costly and very complex since many drug-target interactions (DTI) remain elusive [1,2]. Biochemical assays and genetic interaction approaches were traditionally used to understand MoA of drug candidates [3]. However, these experimental approaches are time consuming. The increasing availability of computational resources and public biochemical assay and pharmacogenomic data has enabled the development of computational drug-target identification strategies over their more traditional experimental counterparts. Computational approaches are categorized into two main categories: chemoinformatics, which includes structure-based virtual screening or molecular docking and ligand-based structural similarity approach; and systems biology, which focuses more on biological and chemogenomic data [4].Chemoinformatics mainly relies on structure-based protein ligand docking which requires a three-dimensional structure of a protein and a library of small molecules and uses sophisticated computational tools in order to model the docking of a ligand into a protein binding pocket [5]. However, many protein structures are unavailable. Ligand-based similarity approaches assume that two similar protein sequences bind similar drugs, so for a given small drug with a known protein partner, a highly similar protein would be predicted as a putative target for the tested drug [6]. The limitation of this method is that it is only applicable to drugs with established binding information toward corresponding targets which is not the case for chemical compounds in early screening phases.In contrast, systems-based approaches leverage a diverse set of high-dimensional pharmacogenomic data. Some of these methods [7] infer new DTI from comprehensive pharmacological and structural data regarding known DTI from online resources, such as DrugBank [8] and PubChem [9]. Cheng et al. developed a network-based inference method that predicts a new DTI using a two-step diffusion process on a bipartite graph of drugs and targets, and their known interactions as edges [10]. Other methods have extended this seminal work with sophisticated graph traversal methods to infer new DTI [11,12]. These methods have shown improved accuracy and provided better scalability than structure-and ligand-based approaches. However, they usually rely on high-level annotations of drugs and...