Integrative quantitative and qualitative analysis for the quality evaluation and monitoring of Danshen medicines from different sources using HPLC-DAD and NIR combined with chemometrics

Abstract

The root and rhizome of Salvia miltiorrhiza (Danshen in short) is a well-known herbal medicine used to treat cardiovascular diseases in the world. In China, the roots and rhizomes of several other Salvia species (Non-Danshen in short) are also used as this medicine in traditional folk medicine by local herbalists. Differences have been reported in these medicines originating from different sources, and their quality variation needs to be clearly investigated for effective clinical application. This study presented a comprehensive quality evaluation and monitoring for Danshen from 27 sampling sites and Non-Danshen from other 5 Salvia species based on a high-performance liquid chromatography-diode array detector (HPLC-DAD) and near-infrared (NIR), with the combination of chemometric models. The results showed that cryptotanshinone, tanshinone IIA, tanshinone I, salvianolic acid B, salvianic acid A sodium, dihydrotanshinone I, and rosmarinic acid in these medicines from different sources exhibited great variations. Referring to the standards in Chinese Pharmacopoeia (CP), European Pharmacopeia (EP), and United States Pharmacopeia (USP), Non-Danshen from S. brachyloma, S. castanea, S. trijuga, S. bowleyana, and S. przewalskii were assessed as unqualified, and Danshen in the Shandong Province had the best quality due to the high qualified rate. Based on random forest (RF) and partial least-squares discriminant analysis (PLS-DA), NIR technique could successfully monitor the quality of these medicines by discriminating the species and regions with the accuracies of 100.00 and 99.60%, respectively. Additionally, modified partial least-squares regression (MPLSR) models were successfully constructed to investigate the feasibility of NIR fingerprints for the prediction of the quality indicators in these medicines. The optimized models obtained the best results for the total of tanshinone IIA, tanshinone I, and cryptotanshinone (TTC), tanshinone IIA, and salvianolic acid B, with the relative prediction deviation (RPD) of 4.08, 3.92, and 2.46, respectively. In summary, this study demonstrated that HPLC-DAD and NIR techniques can complement each other and could be simultaneously applied for evaluating and monitoring the quality of Danshen medicines.

Keywords: HPLC-DAD; NIR; Salvia medicines; chemometric models; quality evaluation and monitoring.

Figure 1

The distribution of sampling sites in 7 provinces in China. (Description: This figure shows parts of China consisting of the 27 sampling sites from 7 provinces for Danshen medicines. To obtain a better quality evaluation, we artificially divided these sites into 5 regions according to different environmental conditions).

Figure 2

The simple data analysis flow diagram of this study. (Description: This figure shows the main steps of HPLC-DAD quantitation for active compounds, NIR fingerprint for overall metabolites, and the application of chemometric models, which complement each other for a comprehensive quality analysis for Salvia medicines from different sources).

Figure 3

The representative chromatograms of medicines based on the HPLC-DAD technique (Description: (A) represents the chromatograms of mixed standard solution and Danshen medicines; (B) represents the chromatograms of roots and rhizomes of 6 Salvia species. The chemical compounds are salvianic acid A sodium, rosmarinic acid, salvianolic acid B, dihydrotanshinone I, cryptotanshinone, tanshinone I, and tanshinone IIA according to the retention time, respectively. The roots and rhizomes for Salvia species of 1–6 are S. miltiorrhiza, S. trijuga, S. castanea, S. przewalskii, S. bowleyana, and S. brachyloma, respectively.).

Figure 4

The raw NIR spectra fingerprints of Salvia medicines (Description: this figure shows the original NIR spectra of different medicines from 850 to 2,500 nm at 2 nm intervals. Several common peaks are also marked in this figure.).

Figure 5

The spectra fingerprints after the pretreatments (Description: Appropriate pretreatments can improve the quality of raw spectra. In this study, the smoothing, SNV, and 2D algorithm are applied simultaneously and some valuable peaks are also marked.).

Figure 6

The visualization distribution of multi-source samples using t-SNE (Description: (A) represents the visualization result for Danshen medicines from different regions; (B) represents the visualization result between Danshen and Non-Danshen medicines; T-SNE 1 and t-SNE 2 represent the scores on the X and Y axes of the spectral data after dimensionality reduction, respectively.).

Figure 7

The parameter selection of the RF model (Description: (A) represents the parameter selection approaches for the n-estimator; (B) represents the parameter selection approaches for max_feature. These parameter selection were performed according to OOB error, which can guarantee an excellent predicted and generalization ability for the developed RF model.).

Figure 8

The variable importance of NIR variables based on the Gini impurity index (Description: this figure displays the variable evaluation procedure after parameter selection. After the elimination of unimportant variables, the accuracy of the RF model can be further improved.).

Figure 9

The permutation and AUC plots of the developed PLS-DA model (Description: (A) represents the permutation plot; (B) represents the AUC plot. According to the visualization result, PLS-DA confidently exports a 100% accuracy for discriminating between Danshen and Non-Danshen medicines. We show the AUC and permutation plots to indicate the reliability of this model.).

Figure 10

The correlation diagrams of the NIR predicted and HPLC-DAD data regarding three quality indicators (Description: TTC represents the total of tanshinone IIA, tanshinone I, and cryptotanshinone; TA2 and SAB represents the compounds of tanshinone IIA and salvianolic acid B, respectively. As seen from this figure, these observations are evenly distributed on both sides of the correlation equation, indicating these models are effective to predict these quality indicators.).