论文部分内容阅读
Two kinds of radial basis networks, radial basis function neural networks(RBFN) and general regression neural networks (GRNN), have been used forthe quantitative structure & activity relationship (QSAR) study ofpolychlorinated dibenzo-p-dioxins (PCDDs) and the simultaneous determinationof polycyclic aromatic hydrocarbons (PAHs). A GRNN was used for the first time in the QSAR study of organicpollutants to predict n-octanol/water partition coefficients of PCDDs from theirtopological molecular descriptors. In total, 42 PCDDs and dibenzo-p-dioxinswere available for this study – 42 PCDDs and dibenzo-p-dioxins in the trainingdata set and 41 PCDDs in the test data set. Partial least squares (PLS), backpropagation neural network (BPN) and general regression neural network(GRNN) models were trained using the training data set, and the accuracy of themodels obtained were examined by the use of leave-one-out cross-validation.For predicting the n-octanol/water partition coefficients, the GRNN modelperformed best. With the test data set, the correlation coefficient and root meansquare error for the GRNN model are 0.9276 and 0.22 respectively. The optimalGRNN model was used to predict the n-octanol/water partition coefficients of33 PCDDs outside the training data set. For describing the structure of PCDDs,the topological molecular descriptors (nine Cl-substitution descriptors, numberof Cl-substitution and molecular weight) outperformed the mobile order anddisorder thermodynamic method. A RBFN was used to correlate the Ah-receptor binding affinity ofpolychlorinated dibenzo-p-dioxins (PCDDs) and polybrominateddibenzo-p-dioxins (PBDDs) with six quantum chemical descripors-electronicenergy(EE), polarizability(α), first hyperpolarizability(β), the most negativeelectrostatic potential charge on atom, total energy(TE) and Elumo-Ehomo . PLS,BPN and RBFN models were trained with 14 PCDDs and 10 PBDDs as trainingdata set, and the accuracy of the models obtained was examined by the use of I<WP=5>leave-one-out cross-validation. For predicting the Ah-receptor binding affinity,the RBFN model performed best. With the training data set, the correlationcoefficient and root mean square error for the optimal RBFN model are 0.9498and 0.439. The optimal RBFN model was used to predict the Ah-receptorbinding affinity of 61 PCDDs and dibenzo-p-dioxin outside the training data set. RBFN, BPN and PLS were compared in order to establish the bestmultivariate model for the analysis of polycyclic aromatic hydrocarbons (PAHs)mixtures containing pyrene, benzo[a]anthracene and chrysene. The synchronousfluorescence spectra (recorded at wavelength increments of 100,115 and 150 nm)of 29 standards (20 standards as calibration set and 9 standards as validation set)were used for this purpose. For predicting the validation set, the RBFN and BPNmodels performed best with lower root mean square errors of 1.15 and 0.99respectively in contrast to 2.23 for the PLS model. With comparable predictingaccuracy with the BPN model, the RBFN model needs only 4 epochs of trainingin contrast to 104 epochs of training for the BPN model.