Abstract
PLS (partial least square regression) has been widely used in QSAR (quantitative structure-activity relationships) studies due to its robust character against the collinear descriptors and small rations of molecules/descriptors. Although PLS is useful, its major restriction is that only linear relation can be extracted from data. Since many structure-activity data are inherently nonlinear in nature, it is desirable to have a flexible method that can model any nonlinear relations. There has been a considerable interest in neural networks for nonlinear modeling. Among them, CNN (counter propagation neural networks) is considered to be a potential candidate because of small learning time and high reproduction compared to other methods. In this study, CNN was used to analyze the nonlinear QSAR data. In order to choose the optimal descriptors from the huge possible combinations, GA (genetic algorithms) was combined with CNN routine. The predictive explained variance for the prediction set was used as the fitness function of GA. From GA and CNN modeling, the totally 15 descriptors could be reduced to 9 descriptors and the prediction ability of the CNN model could be highly improved. Furthermore, external validation was carried out by use of the best CNN model and external data set. The results really indicated that the predicted values are closer to the observed ones compared to the original CNN model.
