It is unfortunate that this diligent effort in applying machine learning to the contents of DrugBank has been substantially confounded by an apparent misunderstanding of what the chemical-to-protein links captured in this key resource actually represent. The utility of drug-target networks for pharmacologically-relevant modeling and prediction is predicated on being able to filter data sources down to direct binding interactions that have a measurable specific affinity in vitro, biochemical consequences in vivo, and be coincident w.r.t. to tissue or cell compartment. The authors have not taken into account that this relationship stringency only applies to a subset of DruBank records. This is made clear in the most recent DrugBank publication (PMID 21059682) where they explain that such mechanistically plausibly links are now tagged with “known pharmacology”. The lower limits should correspond to, as reported in reference 22, 324 primary targets for all classes of approved therapeutic drugs, although a more recent analysis of the Drugbank data set expands this to 435 human proteins (PMID: 21804595)

Even assuming a generous upper-limit estimate of ~1000 proteins with data-supported single target, multiple targets and/or off-target interactions with DrugBank chemical structures the “target” set of 3,988 used by the authors exceeds this by at least four-fold. Authentic binding interactions are thus diluted by the inclusion of largely indirect mechanistic links identified by co-citation (e.g. secondary effects or up-regulation). This results in, for example, the inclusion in Table 4 of thrombin as a target for NMDA where PubMed searches reveal this to be an indirect relationship mediated via PAR1 activation. Compounding the impression of an incomplete understanding of data sources is the inclusion of a series of central energy metabolism cofactors clustered with their many cognate binding enzymes in figure 7. Presenting these (e.g. NADH) as drug-protein hubs is mystifying when they are clearly labeled as nutracueticals in the database. Intersecting DrugBank proteins with data sets from ChEMBL, Therapeutic Target Database or published lists (e.g. PMID 21569515) would have introduced a more appropriate level of target interaction specificity for this study.