甘草中甘草苷和甘草酸红外定量模型特征变量的筛选与解析

借助变量筛选方法可以从复杂的光谱背景下选择部分变量构建定量预测模型，在一定程度上提高建模变量的解释性。然而模型解释性的提高并不意味着建模变量有确切的理化意义。本研究以甘草中红外定量预测模型为载体，解析移动窗口偏最小二乘(mwPLS)、组合间隔偏最小二乘(siPLS)和竞争自适应抽样方法(CARS)三种变量筛选方法所得变量与目标成分化学特征的相关性，比较不同变量筛选方法下所筛变量解释性的差异。结果表明，mwPLS优先筛出黄酮和皂苷两类成分红外光谱上区别明显的苯环骨架振动和皂苷母核上甲基取代基弯曲振动所对应的波段，siPLS筛出了黄酮类成分的(φ)C—O，(φ)CC， (φ)C—H伸缩振动的特征区间组合和皂苷类成分的C—O，C—H，O—H伸缩振动的特征区间组合。相对于以上两种变量筛选方法，CARS筛选得到的变量能够更好地归属于甘草苷和甘草酸在中红外1 000～4 000 cm-1特征区的特征峰，而且基于CARS筛选的变量建模，模型的预测性能得到了提高。因此，CARS筛选的变量能实现目标成分红外特征区大部分化学特征的解析，有利于增强模型的解释性。

Feature Selection and Interpretation in Infrared Quantitative Models of Liquiritin and Glycyrrhizin in Radix Glycyrrhizae

Feature selection can improve the interpretation of the modeling variables to a certain extent by selecting variables from the complex spectra backgrounds. However, the improvement of models interpretation does not mean that the modeling variables have the exact physical or chemical significance. In this paper, We explore the relation between the chemical characteristics of target components and the spectrum variables selected with 3 kinds of variables selection methods which are moving window partial least squares regression(mwPLS), synergy interval partial least squares regression(siPLS) and competitive adaptive reweighted sampling(CARS), and compare the interpretation difference of the variables selected with the above variables selection methods. The results show that the variables selected with mwPLS accord with ν(φ)CC of liquiritin and δ_CH3 or δ_CH2 of glycyrrhizin, which are the obvious spectra differences between the flavonoids and saponins in Radix Glycyrrhizae, and the variables selected with siPLS are the characteristic intervals combinations of the flavonoids or saponins in Radix Glycyrrhizae, which is the combination of ν_()CC, ν_()C—O, ν_()C—H of flavonoids or the combination of ν_C—O, ν_C—H, ν_O—H of saponins while the variables selected with CARS can better accord with most of the characteristic peaks from 1 000 to 4 000 cm-1 of liquiritin or glycyrrhizin in Radix Glycyrrhizae, and the predict performance of the infrared quantitative model established on the spectroscopic variables selected with CARS can be improved. Therefore, most of the variables selected with CARS can be interpreted by the characteristic peaks in the infrared characteristic region of the target components, which is beneficial to improve the interpretation of the quantitative model.