神经网络用于三元不对称有机磷酸酯杀虫剂的QSAR研究

Neural network analysis in the study of the QSAR of ternary dissymmetric organic phosphate insecticide

在０乙基Ｎ烷基（取代硫脲基）硫代磷酰胺酯类杀虫剂的定量构效关系（ＱＳＡＲ）研究中，用Ｒ型聚类分析提取特征结构参数，结合多元线性回归和神经网络方法研究构效关系。回归方法为ＱＳＡＲ研究提供变量的物理解释，改进的神经网络方法———广义误差反传神经网络（ＧＢＰ）建立了更加精确的构效关系模型。研究表明神经网络在ＱＳＡＲ研究中具有良好的预测和非线性处理功能。

By use of the improved neural network method Generalized error Back Propagation (GBP) combined with the multiple linear regression(MLR),a quantitative structure activity relationship(QSAR) of a set of 22 O ethyl N alkyl(substituted thioureido) phosphoamidethioates insecticides was studied. The significant parameters were determined by R type cluster.Although the regression method gave physical explanations,it couldn′t analize the non linear QSAR.The non linear relationships between the structure and the activity were studied by GBP method with 4-4-1 network.The study rate was 0.5,modified factor was 0.05.The cross validation was used by leave one out method to avoid the overfitting phenomena.The relative errors of the prediction was 0.027.The result of study show that the neural network can play an important role in non linear QSAR and give out a more accurate model for QSAR.