Online quality control of panaxatriol saponins percolation extraction using near-infrared technology

Purpose: To establish a new prediction model for online quality control of the percolation extraction of panaxatriol saponins (PTS), viz, ginsenoside Rg1, ginsenoside Re and notoginsenoside R1, from notoginseng by near-infrared (NIR) technology coupled with partial least squares (PLS) analysis. Methods: Ten batches of PTS (420 samples) were collected and the constituents were determined using HPLC. The NIR spectroscopy of samples was determined using a Fourier-Transform nearinfrared spectrometer with an optical fiber transmission PbS detector. Eight sample batches were the calibration set, and two batches were the forecast set. Calibration models were established based on min-max normalization (MMN). Results: The root mean square errors of cross-validation (RMSECV) of Rg1, Re, and R1 were 0.798, 0.095, and 0.259 mg/mL, respectively. The root mean square errors of prediction (RMSEP) were 1.110, 0.496, and 0.390 mg/mL, respectively. The correlation coefficients (R 2 ) of cross-validation were 0.9682, 0.9681, and 0.9626, respectively, while the correlation coefficients (R 2 ) of prediction were 0.9831, 0.9198, and 0.9661, respectively. Conclusion: The results indicate that NIR is a quick and effective tool for online quality control of PTS (ginsenoside Rg1, ginsenoside Re, and notoginsenoside R1) in the percolation extraction process.

In the present study, new mathematical models for PTS using near-infrared(NIR) combined with partial least squares (PLS) were established for the online quality control of the percolation extraction process, which helps to guarantee the PTS quality.

EXPERIMENTAL Chemicals and reagents
Standards of ginsenoside Rg1, notoginsenoside R1, and ginsenoside Re were purchased from the National Institute for the Control of Pharmaceutical and Biological Products (Beijing, China).Chromatographic acetonitrile was purchased from Fisher Chemicals (Fair Lawn, New Jersey, USA).The ultrapure water used was purified using the Millipore purification system (Millipore, MA, USA).All other reagents used in this study were of analytical grade.

Sample preparation
Samples were powdered and strained through a standard 24 mesh sieve (850 ± 29 μm).The powder was first immersed in 60 % ethanol for 24 h and then percolated with a flow rate of 5 -8 mL/min.The percolation process lasted for 19 and 25 mL samples were collected: 0 -5 h, collected every 1 h; 5 -15 h, collected every 30 min; 15 -19 h, collected every 15 min.A total of 420 samples were obtained (42 samples for each batch).Each sample was directly determined using a PbS detector at room temperature.Batches 1, 2, 3, 4, 6, 7, 8 and 9 were the calibration set, whereas batches 5 and 10 were the forecast set.

HPLC chromatographic conditions
HPLC was performed using an Agilent 1200 high-performance liquid chromatography (HPLC) system (Agilent Technologies, CA, USA).The samples were analyzed on an Agilent C18 column (250 mm × 4.6 mm, 5 μm, Agilent).The elution system contained acetonitrile:water (19.5:80.5)at a 1.0 mL/min flow rate and the column temperature was maintained at 30 °C.The wavelength was set at 210 nm.

The near-infrared spectroscopy conditions
The NIR spectroscopy was determined using a Bruker Matrix-I Fourier-Transform Near-Infrared (FT-NIR) spectrometer (Bruker Optik, Ettlingen, Germany).An optical fiber transmission PbS detector was used for spectra acquisition and the OPUS workstation was employed for dataprocessing.The absorbance of air was regarded as the reference standard.Each sample was scanned 64 times with the resolution of 8 cm -1 over the scan range from 12500 to 3600 cm -1 .
After the automatic baseline corrections of the FT-NIR, the samples were added in 10 mL beakers and the optical fiber transmission PbS detector was inserted.Every sample was scanned three times after the sample was stable without bubbles and the mean value was used for the final analysis.

Saponin content of the samples
The calibration set had 336 samples and the validation set had 84 samples.Table 1 shows the contents of notoginsenoside R1, ginsenoside Rg1, and ginsenoside Re during the percolation extraction process.

Primary selection of the spectrum region
In this study, PLS method was employed for analyzing the models.Although PLS can assess the whole spectrum, this increases computational cost and reduces predictive power.In Figure 1, the absorbance for the wavelength range 12000-8000 cm -1 was almost a flat line, whereas the absorbance for the wavelength range 8000-4000 cm -1 changed significantly.This suggests that the wavelength range of 8000-4000 cm -1 may contain abundant information, which could reflect the sample composition.Therefore, the wavelength range of 8000 -4000 cm -1 was selected for analysis using PLS method.

Selecting pretreatment methods
The OPUS workstation was developed by Bruker Instruments LTD (Bruker Optik, Ettlingen, Germany) and has 11 complete spectral pretreatment methods including min-max normalization (MMN), vector normalization (VN), and constant offset elimination (COE).  .As shown in Table 2 and Table 3, MMN was the most appropriate method for ginsenoside Rg1, ginsenoside Re, and notoginsenoside R1 analyses.

Determination of the dimensions (D) of PLS factors
The PLS D values greatly affected the prediction results of the calibration models.Even in the same spectral pretreatment method, the D value could directly influence the values of R 2 , RMSECV, and RMSEP.A D value that is too small could lead to insufficient spectral information, whereas a D value that is too large could cause over fitting of the model.The OPUS could successfully recommend the optimal D value.The statistical parameters versus D are presented in Figure 2. As can be seen in Figure 2, when the D value increased, the value of RMSECV decreased, but R 2 increased in the calibration sets.Moreover, the two parameters were stable when the D value achieved a certain value.These D values for ginsenoside Rg1, ginsenoside Re, and notoginsenoside R1 acquired by OPUS were 17, 18, and 24, respectively.

Predicted result of the validation set
The calibration models were established using the PLS method with the obtained optimal parameters.Then the models were used to predict the values of ginsenoside Rg1, ginsenoside Re, and notoginsenoside R1 in the 5 th and 10 th batch of samples.Figure 3 shows there is no significant difference between the values predicted by the NIR models and the values directly measured by HPLC.Furthermore, the results by NIR were obtained in 20 s, but HPLC required at least 30 min.Thus, the established models could satisfy the precision requirement of the online analysis for percolation extraction process of PTS.The parameters of the three optimal models are shown in Table 4.To find meaningful associations from NIRs, appropriate mathematical models need to be established.Recently, some chemometrics methods were used to establish NIR mathematical models, such as partial least squares (PLS), artificial neural networks (ANN), and multiple linear regression.Among these methods, PLS is the most popular because it can reveal information for the dependent variable as well as reduce the dimensions of the spectral matrix [25-27].In the present study, NIR coupled with PLS was employed, and a mathematical model was successfully established for online monitoring of PTS (ginsenoside Rg1, ginsenoside Re, and notoginsenoside R1) in the percolation extraction process.

CONCLUSION
An online monitoring model has been successfully established for the percolation extraction process of PTS extracted from notoginseng using NIR technology combined with PLS.Overall, the proposed method and models are efficient and accurate, and can be applied for rapid analysis and online quality control in PTS manufacturing.

Figure 1 :
Figure 1: The NIR spectrum of the samples From the OPUS, the predictive ability of different models was assessed for low root mean square errors of cross-validation (RMSECV) and root mean square errors of prediction (RMSEP) and high R 2

Table 4 :
Parameters of optimal models for three active ingredients