A temperature-constrained cascade correlation network(TCCCN), a back-propagation neural network(BP), and multiple linear regression(MLR) models were applied to quantitative structure-activity relationship(QSAR) modeling, on the basis of a set of 35 nitrobenzene derivatives and their acute toxicities. These structural quantum-chemical descriptors were obtained from the density functional theory(DFT). Stepwise multiple regression analysis was performed and the model was obtained. The value of the calibration correlation coefficient R is 0.925, and the value of cross-validation correlation coefficient R is 0.87. The standard error S=0.308 and the cross-validated(leave-one-out) standard error S_ cv =0.381. Principal component analysis(PCA) was carried out for parameter selection. RMS errors for training set via TCCCN and BP are 0.067 and 0.095, respectively, and RMS errors for testing set via TCCCN and BP are 0.090 and 0.111, respectively. The results show that TCCCN performs better than BP and MLR. A temperature-constrained cascade correlation network (TCCCN), a back-propagation neural network (BP), and multiple linear regression (MLR) models were applied to a quantitative structure-activity relationship (QSAR) modeling, on the basis of a set of 35 These structural quantum-chemical descriptors were obtained from the density functional theory (DFT). Stepwise multiple regression analysis was performed and the model was obtained. The value of the calibration correlation coefficient R is 0.925, and the value The standard error S = 0.308 and the cross-validated (leave-one-out) standard error S_cv = 0.381. The Principal component analysis (PCA) was carried out for parameter selection. RMS errors for training set via TCCCN and BP are 0.067 and 0.095, respectively, and RMS errors for testing set via TCCCN and BP are 0.090 and 0.111, respectively. The results show that TCCCN performs better than B P and MLR.