摘要
采用取代基片段值P和原子类型电拓扑状态指数Em有效表征了135个多氯咔唑化合物(PCCZs)的分子结构,通过选择变量与神经网络(BP)算法建立定量相关(QSPR)模型,以预测多氯咔唑化合物热力学性质.将选择的P,Em结构参数作为神经网络的输入层变量,热力学性质作为输出层变量,方程均采用5∶13∶1的网络结构,利用BP算法获得了3个令人满意的QSPR模型,它们的总相关系数分别为0.998 6,0.991 1和0.979 5,标准误差分别为2.123,3.237和3.952,利用这3个神经网络模型计算得到的预测值与文献值的相对平均误差分别为0.30%,1.85%和1.14%,表明模型具有良好的稳定性和预测能力.该神经网络模型所得结果优于多元回归方法所得结果,可用于对多氯咔唑化合物性质进行理论分析和预测.
We described the structures of 135 polychlorinated carbazoles （PCCZ） by the substituent fragments P and atom electrotopological state index Era. Using the variable selection and neural network back-propagation （BP） al- gorithm, we were able to predict the thermodynamic properties of polychlorinated carbazoles. Using the structural pa- rameters P and Em as the input neurons of the artificial neural network while using the thermodynamic properties as the output neurons, we constructed three satisfying QSPR models by back-propagation algorithm, whose network structures were 5 ： 13 ： 1. The total correlation coefficients R for the three models were 0. 998 6, 0. 991 1 and 0. 979 5 respectively, and the standard errors S were 2. 123, 3. 237 and 3. 952 respectively. The relative mean deviation be- tween the results in literature and the predicted values were 0.30%, 1.85% and 1.14% respectively, which showed that the models had good stability and predictive ability. Compared with the results by multiple regression methods, the results by neural network methods were better. The QSPR model could be applied in the theoretical analysis and prediction of the properties of polychlorinated carbazoles.
出处
《武汉大学学报（理学版）》
CAS
CSCD
北大核心
2014年第4期324-330,共7页
Journal of Wuhan University:Natural Science Edition
基金
江苏省自然科学基金项目(09KJD150012)
徐州市绿色技术重点实验室项目(SYS2012009)资助
关键词
多氯咔唑
神经网络
电性拓扑状态指数
片段
热力学性质
定量结构-性质相关性
polychlorinated carbazoles （PCCZ）
neural networks
electrotopological state index ragment
thermo-dynamic properties
quantitative structure-property relationship（QSPR）
