A comparison of methods for modeling quantitative structure-activity relationships

A large number of methods are available for modeling quantitative structure-activity relationships (QSAR). We examine the predictive accuracy of several methods applied to data sets of inhibitors for angiotensin converting enzyme, acetylcholinesterase, benzodiazepine receptor, cyclooxygenase-2, dihydrofolate reductase, glycogen phosphorylase b, thermolysin, and thrombin. Descriptors calculated with CoMFA, CoMSIA, EVA, HQSAR, and traditional 2D and 2.5D descriptors were used for developing models with partial least squares (PLS). In addition, the genetic function approximation algorithm, genetic PLS, and back-propagation neural networks were used for deriving models from 2.5D descriptors (i.e., 2D descriptors and 3D descriptors calculated from CORINA structures and Gasteiger-Marsili charges). Predictive accuracy was assessed using designed test sets. It was found that HQSAR generally performs as well as CoMFA and CoMSIA; other descriptor sets performed less well. When 2.5D descriptors were used, only neural network ensembles were found to be similarly or more predictive than PLS models. In addition, we show that many cross-validation procedures yield similar estimates of the interpolative accuracy of methods. However, the lack of correspondence between cross-validated and test set predictive accuracy for four sets underscores the benefit of using designed test sets.