Cruciani G, Watson K A
Department of Chemistry, University of Perugia, Italy.
J Med Chem. 1994 Aug 5;37(16):2589-601. doi: 10.1021/jm00042a012.
A primary goal in any drug design strategy is to predict the activity of new compounds. Comparative molecular field analysis (CoMFA) has been used in drug design and three-dimensional quantitative structure/activity relationship (3D-QSAR) methods. The CoMFA approach permits analysis of a large number of quantitative descriptors and uses chemometric methods such as partial least squares (PLS) to correlate changes in biological activity with changes in chemical structure. One of the characteristics of the 3D-QSAR method is the large number of variables which are generated in order to describe the nonbonded interaction energies between one or more probes and each drug molecule. Since it is difficult to know a priori which variables affect the biological activity of the compounds, much effort has been devoted to developing methods that optimize the selection of only those variables of importance. This work focuses on some of the aspects involved in the selection of such variables, applied to a series of glucose analogue inhibitors of glycogen phosphorylase b, using the program GRID to describe the molecular structures and using a method of generating optimal partial least squares estimations (program GOLPE) as the chemometric tool. This data set, consisting of over 30 compounds in which the three-dimensional ligand-enzyme bound structures are known, is well suited to study the effect of different data pretreatment procedures on the final model used for the prediction of new drug molecules. By relying on our knowledge of the real physical problem (i.e., using the combined crystallographic and kinetic results), it has been shown that suitable data pretreatment and variable selection have been found that does not result in a significant loss of relevant information. Moreover, by using an appropriate scaling procedure, GOLPE variable selection minimizes the risk of overfitting and overpredicting.
