Virtual screening using PLS discriminant analysis and ROC curve approach: an application study on PDE4 inhibitors.

Abstract

Virtual screening (VS) represents an important tool for the drug discovery process, in particular for the hit generation phase. Classifiers are often inserted as filters at the beginning of a VS path, and in the present paper the performances of several PLS-DA classifiers (QikProp, Dragon, EVA descriptors) are evaluated in the effort to distinguish PDE4 inhibitors from other druglike molecules. As benchmark also docking scores and the fitness to pharmacophore hypotheses were used to perform the same task, checking in this way if docking or 3D search can be anticipated in the VS process. The visual analysis of the Receiver Operating Characteristic (ROC) curve was useful to have an overall picture of the classification and to select the right threshold that marks the boundary between active and inactive classes. The best classification was obtained by a model based on the Dragon descriptors that are calculated from the molecular 2D structure. Its performance was good for the training set in terms of recall, enrichment factor, and area under the ROC curve and was confirmed in the prediction of the test set.