GA strategy for variable selection in QSAR studies: GAPLS and D-optimal designs for predictive QSAR model

Variable selection is an important task in obtaining an interpretable and predictive model. In general, variable space has a complex landscape with many local solutions and. therefore, classical optimization techniques cannot be applied. The genetic algorithm (GA) is a recently developed optimization technique that has attracted much attention in several scientific fields. In a previous study (K. Hasegawa et al., J. Chem. Inf. Comput. Sci., 37 (1997) 306), we proposed a novel approach (GAPLS; GA-based PLS) in which variables are selected by GA and partial least squares (PLS). The purpose of this study is to examine whether GAPLS works well for a data set with a medium size of variables. The binding affinities of ligands to benzodiazepine/ GABA(A) receptor were used as a test example. The relationship between 42 physico-chemical parameters at six positions on 57 benzodiazepines (BZs) and their binding affinities has been investigated by GAPLS. The best PLS model with the selected variables was significantly more predictive than the one with all variables, and the structural requirements for the receptor binding affinity could be estimated from the selected variables. (C) 1998 Elsevier Science B.V.