改进的BP网络应用于部分硝基芳烃的QSAR研究

Improved BP Networks As a QSAR Model for Nitrobenzene Derivatives

应用改进的ＢＰ网络法进行部分硝基芳烃的ＱＳＡＲ研究，用所建立的模型进行毒性预报，并与传统的ＢＰ网络和多元线性回归模型比较，表明改进的ＢＰ网络模拟和预报能力均优于传统的ＢＰ网络和多元线性回归模型。在学习集中，改进的ＢＰ网络，传统的ＢＰ网络，多元线性回归三者的均方差分别为：０．０２４０，０．０１０７，０．０４４７，决定系数为：０．９３６８，０．９７１８，０．８８２４；预报集中，三者的均方差分别为：０．０４４０，０．０５５４，０．０７７２，决定系数为：０．８２１６，０．７７５３，０．６８７０。文中还讨论了网络改进的理由。

Three kinds of quantitative structure activity relationship (QSAR) models for nitrobenzene derivatives were constructed by using the improved BP network and multivariate linear regression analysis and used to predict toxicities of the nitrobenzene derivatives not included in the training set. By comparing the calculated values and the experimental results, it was concluded that BP network approach for QSAR study of environmental pollutants was better than traditional BP network and multivariate linear regression analysis. For above three models, in learning set, MSE was 0 024 0, 0 010 7, 0 044 7, R 2 0 936 8, 0 971 8, 0 882 4 respectively; in predicted set, MSE 0 044 0, 0 055 4, 0 0772 ,R 2 0 821 6, 0 775 3, 0 687 0 respectively. The reasons for improving the BP network and how to select some parameters in BP network were also discussed.