With the artificial neural network（ANN） method combined with the multiple linear regression（MLR）,based on a series of quantum chemical descriptors and molecular connectivity indexes,quantitative structure-activity relationship（QSAR） models to predict the acute toxicity（-lgEC50） of substituted aromatic compounds to Photobacterium phosphoreum were established.Four molecular descriptors that appear in the MLR model,namely,the second order valence molecular connectivity index（2XV）,the energy of the highest occupied molecular orbital（EHOMO）,the logarithm of n-octyl alcohol/water partition coefficient（logKow） and the Connolly molecular area（MA）,were inputs of the ANN model.The root-mean-square error（RMSE） of the training and validation sets of the ANN model are 0.1359 and 0.2523,and the correlation coefficient（R） is 0.9810 and 0.8681,respectively.The leave-one-out（LOO） cross validated correlation coefficient（Q L2OO） of the MLR and ANN models is 0.6954 and 0.6708,respectively.The result showed that the two methods are complementary in the calculations.The regression method gave support to the neural network with physical explanation,and the neural network method gave a more accurate model for QSAR.In addition,some insights into the structural factors affecting the acute toxicity and toxicity mechanism of substituted aromatic compounds were discussed.
用量子化学MOPAC-AM1法计算21种多环芳烃(PAHs)的SEDs(steric and electronic descriptors)参数,然后用多元线性回归法建立预测多环芳烃的沸点(BP)和辛醇/水分配系数(logK_(ow))等的QSPR模型,预测BP的模型含3个变量[前线轨道能量差(E_(lumo)-E_(homo))、分子总电子能(EE)和分子总连接性(TCon)],预测logK_(ow)的模型含3个变量[偶极矩(D)、分子总能量(TE)和分子总连接性(TCon)]。所建2个模型的相关系数的平方(R^2)分别为0.997 6和0.9861,交叉验证系数(R_(LOO)~2)分别为0.9820和0.9575,说明模型均具有很好的预测能力和较强的稳健性,同时也证明SEDs参数适用于多环芳烃类化合物的QSPR研究。
Based on the quantum chemical descriptors,quantitative structure-property relationship（QSPR） models have been developed to estimate and predict the photodegradation rate constant（logK） of polycyclic aromatic hydrocarbons（PAHs） by use of linear method（multiple linear regression,MLR） and non-linear method（back propagation artificial neural network,BP-ANN）.A BP-ANN with 3-3-1 architecture was generated by using three quantum chemical descriptors appearing in the MLR model.The standard heat of formation（HOF）,the gap of frontier molecular orbital energies（ΔELH） and total energy（TE） were inputs and its output was logK.Leave-One-Out（LOO） Cross-Validated correlation coefficient（R^2CV） of the established MLR and BP-ANN models were 0.6383 and 0.7843,respectively.The nonlinear BP-ANN model has better predictive ability compared to the linear MLR model with the root mean square error（RMSE） for training and validation sets to be 0.1071,0.1514 and the squared correlation coefficient（R^2） of 0.9791,0.9897,respectively.In addition,some insights into the molecular structural features affecting the photodegradation of PAHs were also discussed.
