1992 年 40 巻 11 号 p. 2962-2969
The quantitative structure-activity relationships (QSARs) of 37 carboquone derivatives with antileukemic activity and 51 benzodiazepine derivatives with anti-pentylentetrazole activity were studied using two neural computing methods-the functional-link net (FUNCLINK) and the generalized delta rule net with the back propagation of error(GDR). Both methods showed good fitting of the activity values in the two data sets. A great difference appeared, however, in the prediction of the activity values : GDR's predictive ability is much lower than FUNCLINK's. To elucidate the difference of the predictive ability, we examined the contribution of parameters to activity using the QSAR models of carboquone derivatives by plotting the contribution curves. Well-regulated and similar contribution curves resulted for all parameters in both the FUNCLINK and GDR models for the entire data. On the other hand, the contribution curves for the leave-one-out models derived by eliminating one compound from the data set showed that much greater deviations occurred in GDR than in FUNCLINK. The QSAR models of GDR seemed to depend greatly upon each individual compound.
