目的 为实现绞股蓝总皂苷（Gynostemma pentaphyllum saponins，GPS）色谱洗脱过程实时监测，保障纯化过程绞股蓝总皂苷质量一致性。方法 采集色谱洗脱过程7批共计237个样本的拉曼光谱，将其中5批用于建模，2批用于外部测试，以总皂苷质量浓度、总固体量和人参皂苷Rb₃（Rb₃）质量浓度为指标，采用高斯过程回归（Gaussian process regression，GPR）法建立定量模型，并将GPR模型与偏最小二乘回归及支持向量机回归定量模型进行性能对比。结果 基于拉曼光谱技术结合GPR，建立了其洗脱过程的多指标定量校正模型。总皂苷质量浓度、总固体量和Rb₃质量浓度3个指标的GPR模型均具有更高的决定系数（R²），训练集R²均为1.00，验证集R²分别为0.953、0.986、0.939，以及更低的误差均方根（root mean square error，RMSE），训练集RMSE分别为70.4、224.0、31.6μg/mL，验证集RMSE分别为3.02、2.03、1.19 mg/mL。GPR模型在外部测试集的结果为总皂苷质量浓度、总固体量和Rb₃质量浓度预测R²分别达到0.947、0.954、0.837，RMSE分别为3.28、4.37、2.44 mg/mL；GPR模型能较好地反映总皂苷质量浓度和总固体量含量和变化趋势，但对Rb₃质量浓度的预测能力较弱。结论 以总皂苷质量浓度和总固体量为指标，提出的基于拉曼光谱结合GPR建模的方法可实现绞股蓝总皂苷色谱洗脱过程的实时监测。

Objective In order to realize the real-time monitoring of chromatographic process of Gynostemma pentaphyllum saponins (GPS) and ensure the quality uniformity and batch consistency. Methods The Raman spectra of 237 samples collected in seven batches during the chromatographic process, of which five batches were used as modeling sets and two batches were used as external test sets. With total saponin concentration, total solids and ginsenoside Rb₃ concentration as indexes, Gaussian process regression (GPR) method was used to establish the model, and the performance was compared with partial least squares and support vector machine regression quantitative models, and the method was applied to external test sets for validation. Results Multi-index quantitative correction models were established based on Raman spectroscopy combined with GPR. The results showed that the GPR models of the three indexes had higher coefficient of determination (R²) and lower root mean square error (RMSE). The R² of the training sets were all 1.00, and the R² of the verification sets were 0.953, 0.986, and 0.939, respectively. The RMSE of the training sets were 70.4, 224.0, 31.6 μg/mL, and the RMSE of the verification sets were 3.02, 2.03, 1.19 mg/mL, respectively. The results of external test sets showed that the prediction R² of total saponin concentration, total solid content and ginsenoside Rb₃ concentration were 0.947, 0.954 and 0.837, respectively, and RMSE were 3.28, 4.37 and 2.44 mg/mL, respectively. GPR model can predict the content and trend of total saponin and total solid well, but it is weak in predicting ginsenoside Rb₃ concentration. Conclusion With total saponins concentration and total solids as indexes, this method can realize the real-time monitoring of the chromatographic process of GPS.

R283.6

国家中医药管理局“组分中药与智能制药多学科交叉创新团队”（ZYYCXTD-D-2020002）