Subset Selection and Docking of Human P2X7 Inhibitors

Peter   P.   Mager

Abstract

This review deals with three problems: the selection of suitable chemical descriptors from a pool of variables by a simultaneous one-regression/one-observation leaving-out resampling, the comparison of the results with a generalized-regression artificial-neural network, using an unconstrained genetic algorithm (GRNN), and the prediction of protein subdomains as potential molecular drug targets. As an example, the human P2X7 (h-P2X7) receptor subunit and a series of novel 4,5-diarylimidazoline inhibitors [Merriman et al., Bioorg. Med. Chem. Lett., 15, 435 (2005)] is used. GRNN ignores relevant and add noisy descriptors although the goodness-of-fit criterion is large. Therefore, GRNN is considered as supplementary tool which cannot replace the traditional QSAR methodology. Simultaneous one-regression/one-observation leaving-out resampling shows that the h-P2X7 inhibitory activity of 4,5- diarylimidazolines depends on electronic, steric and hydrogen-bonding properties of the substituents. Diagnostic statistic examines the validity of the results. The inhibitors are probably bound to sites that are located mainly in the subdomains 344-347 and 370-375 of h- P2X7.

Keywords: Quantitative structure-activity relationships, subset selection, one-regression/one-observation leaving-out resampling, regression analysis, diagnostic statistics, artificial neural network, 4,5-diarylimidazoline P2X7 receptor antagonists, ligand-protein docking