To address the dependence of GRIND-based 3D-QSAR on data set flexibility, we investigate a series of oxidosqualene cyclase (OSC) inhibitors. The results indicate that statistical models are determined independently of the data set but that despite identification of the same outliers and the acceptable test set prediction, not all models show good predictive correlation coefficient (q2). Moreover, the best model was obtained using a data set of the lowest energy conformers generated by a conformational analysis.