Performance of DrugFEATURE, MAPPOD, F-pocket, SiteMap, and DLID on the drug discovery data set (see Supplementary Table S2). The data set contains 17 druggable targets and 10 difficult targets (6 targets for prodrugs/transporters and 4 undruggable ones). In terms of discriminating druggable targets from difficult ones, Cheng's method MAPPOD shows the best performance (but on data used for calibration), whereas DrugFEATURE shows better performance than F-pocket, SiteMap, and DLID. AUC, area under the curve.