We have introduced multilinear PLS in 3D QSAR and applied it to GRID d
escriptors from a set of benzamides with affinity to the dopamine D-3
receptor subtype, synthesized as potential drugs against schizophrenia
. The key issue in 3D QSAR modelling is to obtain a predictive model t
hat is easy to interpret, Each component in the multilinear PLS model
explains clearly defined details, e.g. substituent positions, while th
e bilinear PLS solution is general and more difficult to interpret. Th
e best models were obtained after four components with multilinear PLS
(Q(2) = 51%) and after only one component with bilinear PLS (Q(2) = 5
0%). The external test set was predicted better with multilinear PLS (
Q(2) = 31%) than with bilinear PLS (Q(2) = 25%). With multilinear PLS
one loses in fit and gains in stability and simplicity owing to the fe
wer parameters that need to be estimated as compared with bilinear PLS
. Finally, multilinear PLS is also less influenced by insignificant va
riation in the descriptor block, which is an advantage in 3D QSAR mode
lling. (C) 1997 John Wiley & Sons, Ltd.