BP网络用于香梨酒香气成分的QSRR研究

Study of QSRR on aroma compounds in Kuerle fragrant pear wine using back-propagation artificial neural network

采用误差反传前向人工神经网络建立54种香梨酒香气成分的结构与色谱保留之间的定量关系模型(ANN模型).以54种香梨酒香气成分的分子连接性指数和分子形状属性指数作为输入,色谱保留时间作为输出,采用内外双重验证的方法分析和检验所得模型的稳定性和外部预测能力,所构建网络模型的相关系数为0.998、交叉检验相关系数为0.997、标准偏差为0.289、残差绝对值≤1.12,应用于外部预测集,外部预测集相关系数为0.984;而多元线性回归(MLR)法模型的相关系数为0.951、标准偏差为1.33、残差绝对值≤3.08,外部预测集相关系数为0.953.结果表明:ANN模型获得了比MLR模型更好的拟合效果.

The systematic study of the quantitative structure-retention relationship （QSRR） of 54 aroma compounds in Kuerle fragrant pear wine was performed by the artificial neural network （ANN） based on the back-propagation algorithm. The stabilization and generalization ability of the model constructed by ANN was verified by the inner-external test when the molecular connectivity indexes and Kappa shape indexes of 54 aroma compounds in Kuerle fragrant pear wine were used as the input of the neural network and the retention times of these compounds were used as output of the neural network. For the artificial neural network method, the correlation coefficient was O. 998, the leave one out cross-validation regression coefficient was 0. 997, the standard deviation was O. 289, the correlation coefficient of the test set was 0. 984 and the absolute values of residual were less than 1.12. In order to make a contrast, the QSRR model was set up by multiple linear regressions （MLR） method. For the model built by MLR, the correlation coefficient was 0.951, the standard deviation was 1.33, the absolute values of residual were less than 3.08 and the correlation coefficient of the test set was O. 953. The results showed that the performance of neural network method was better than that of MLR method.