冬虫夏草作为著名的传统中药材, 由于其良好的药用价值而备受青睐。 目前多数工作集中研究其活性成分含量以及药理药效。 而对其不同部位的识别研究较为匮乏。 基于红外光谱数据, 结合化学计量学对多维度复杂体系的解析优势对冬虫夏草不同部位进行分类识别。 首先对野生冬虫夏草五个不同部位包括子座头、 子座中、 头部、 虫体中段、 虫体尾段总共808个光谱数据使用标准正态变换(SNV)、 多元散射矫正(MSC)进行数据预处理。 而后用竞争自适应再权重取样(CARS)、 变量组合种群分析(VCPA)挑选具有代表意义的特征变量。 最后使用偏最小二乘判别分析(PLS-DA)、 线性判别分析(LDA)进行建模预测分析。 模型对训练集使用十倍交叉验证, 以准确率(Acc)作为评价指标。 结果表明, 在该数据上PLS-DA模型在10倍交叉验证和独立测试集上的预测准确率分别是90.1%和92.0%, 而使用LDA模型时, 预测准确率分别降低到86.7%和85.8%。 采用CARS和VCPA特征挑选方法可有效将特征从3 601维分别降到699和420维, 同时保持预测准确率与全部特征的预测准确率相当。 而挑选的特征波数630, 625, 1 024, 1 028, 1 084和1 089 cm^-1与虫草的甘露醇相关, 879和874 cm^-1与虫草的多糖相关。 通过对挑选的波数进行Wilcoxon rank-sum检验进一步表明虫草五个部位之间存在显著差异。 研究表明化学计量学方法结合红外光谱能够有效识别冬虫夏草不同部位, 有助于在分子层面上加深对冬虫夏草形成的认识, 为针对虫草不同部位高效利用提供参考。

Cordyceps Sinensis, a famous Chinese medicinal material, is favored due to its good medicinal value. Recently, investigations have focused on the study of its active ingredient content and pharmacological effects. However, scarce studies were reported on the identification of different parts of wild Cordyceps. This study is based on infrared spectroscopy data, combined with the analytical preponderance of chemometrics in multi-dimensional complex systems to classify and identify different parts of Cordyceps Sinensis. First, preprocessing methods, standard normal variation (SNV) and multiplicative scatter correction (MSC) were used on a total of 808 spectral data of five different parts of wild Cordyceps, including head of stroma(HS), middle of stroma(MS), head(HD), the middle larva body(ML) and the end larva body(EL). Then, competitive adaptive reweighted sampling (CARS) and variable combination population analysis (VCPA) were hired to select characteristic variables with representative significance. Ultimately, partial least squares discriminant analysis (PLS-DA) and linear discriminant analysis (LDA) were engaged for modeling and predictive analysis. Ten-fold cross-validation was used on the training set, and accuracy (Acc) was employedas the evaluation index. The results showed that the prediction accuracies of the PLS-DA model on the 10-fold cross-validation and independent test set on this data were 90.1% and 92.0%, respectively, while using the LDA model, the prediction accuracies reduced to 86.7% and 85.8%, respectively. In addition, the dimensions of the features can be effectively reduced from 3 601 to 669 and 420, respectively, when using CARS and VCPA feature selection methods, but keeping the prediction accuracies equivalent to that of all features. The selected wavenumbers 630, 625, 1 024, 1 028, 1 084, and 1 089 cm^-1were related to mannitol in cordyceps, and 879 and 874 cm^-1 were related polysaccharides in cordyceps. The Wilcoxon rank-sum test on the selected wavenumbers further showed significant differences between the five parts of Cordyceps. This study showed that chemometric methods combined with infrared spectroscopy could effectively identify different parts of Cordyceps Sinensis, thereby deepening the understanding of the formation of Cordyceps at the molecular level and providing a reference for the efficient use of different parts of Cordyceps.