采用偏最小二乘法和径向基神经网络结合近红外光谱技术建立蝙蝠蛾拟青霉发酵菌丝体中虫草酸、 多糖和腺苷含量的定量分析模型, 模型泛化能力强且预测精度高, 能够满足原料药及相关产品实际检测中的应用。 通过化学诱变和液体深层发酵获得214个蝙蝠蛾拟青霉菌丝体样品, 扫描获得近红外光谱, 采用常规方法测定样品中虫草酸、 多糖和腺苷的含量。 在应用蒙特卡罗偏最小二乘法识别异常样品、 确定校正集样品数量的基础上, 以逼近度(Da)为评价指标, 采用可移动窗口偏最小二乘法和径向基神经网络筛选特征波长变量, 最佳光谱预处理方法及建模重要参数。 通过比较分析, 最终确定蝙蝠蛾拟青霉菌丝体中虫草酸、 多糖和腺苷含量定量分析模型分别为RBFNN, RBFNN和PLS模型, 其校正集和预测集样品实验测定值与预测值间相关系数(R2p和R2c)分别为0.941 7和0.966 3, 0.980 3和0.985 0, 0.976 1和0.972 8, 表明模型具有很好的拟合度和预测性能。

Partial least squares (PLS) and radial basis function neural network (RBFNN) combined with near infrared spectroscopy (NIR) were applied to develop models for cordycepic acid, polysaccharide and adenosine analysis in Paecilomyces hepialid fermentation mycelium. The developed models possess well generalization and predictive ability which can be applied for crude drugs and related productions determination. During the experiment, 214 Paecilomyces hepialid mycelium samples were obtained via chemical mutagenesis combined with submerged fermentation. The contents of cordycepic acid, polysaccharide and adenosine were determined via traditional methods and the near infrared spectroscopy data were collected. The outliers were removed and the numbers of calibration set were confirmed via Monte Carlo partial least square (MCPLS) method. Based on the values of degree of approach (Da), both moving window partial least squares (MWPLS) and moving window radial basis function neural network (MWRBFNN) were applied to optimize characteristic wavelength variables, optimum preprocessing methods and other important variables in the models. After comparison, the RBFNN, RBFNN and PLS models were developed successfully for cordycepic acid, polysaccharide and adenosine detection, and the correlation between reference values and predictive values in both calibration set (R2c) and validation set (R2p) of optimum models was 0.941 7 and 0.966 3, 0.980 3 and 0.985 0, and 0.976 1 and 0.972 8, respectively. All the data suggest that these models possess well fitness and predictive ability.