Multivariate data analysis using D-optimal designs, partial least squares, and response surface modeling: A directional approach for the analysis of farnesyltransferase inhibitors

We have investigated the combined use of partial least squares (PLS) and statistical design principles in principal property space (PP-space), derived from principal component analysis (PCA), to analyze farnesyltransferase inhibitors in order to identify activity trends (an approach we call a directional approach) and quantitative structure-activity relationships (QSAR) for a congeneric series of inhibitors: the benzo[f]perhydroisoindole (BPHI) series. Trends observed in the PCA showed that the descriptors used were relevant to describe our structural data set by clearly identifying two well-defined structural subclasses of inhibitors. D-Optimal design techniques allowed us to define a training set for PLS study in PP-space. Models were derived for each biological assay under evaluation: the in vitro Ki-Ras and cellular HCT116 tests. Each of these assay-based sets was subdivided once more into two subsets according to two structural classes in this BPHI series as revealed by the PCA model. The response surface modeling (RSM) methodology was used for each subset, and the corresponding RSM plots helped us identify activity trends exploited to guide further analogue design. For more precise activity predictions more refined PLS models on constrained PP-spaces were developed for each subset. This approach was validated with predicted sets and demonstrates that useful information can be extracted from just a few very informative and representative compounds. Finally, we also showed the potential use of such a strategy at an early stage of an optimization process to extract the first activity trends that might support decision making and guide medicinal chemists in the initial design of new analogues and/or lead followup libraries.