A hybrid approach that combines genetic algorithms (GA) with PLS analysis i
s proposed. and a modified cross-validation correlation coefficient (MCVR2)
is used as fitness function. The method is applied to QSPR studies of PCBs
, in comparison with class regression, and many better models are obtained
for each property studied. By MCVR2, not only can the low dimensional PLS m
odels be obtained, but also the best multiple least squares regression (MLS
) models. Moreover. the results show the WHIM descriptors that contain info
rmation about the whole 3D-molecular structure are excellent descriptors fo
r molecular modeling, but by adding some common molecular parameters repres
enting total molecular properties the quality of the models is still improv
ed slightly. (C) 2001 Elsevier Science Ltd. All rights reserved.