A method for detecting adulteration of pseudo-ginseng or its preparation with non-medicinal rootlet

Abstract

The application discloses a detection method for detecting adulteration of non-medicinal rootlets in panax notoginseng or preparations thereof, which comprises the following steps: preparing a test sample solution by extracting the test sample with an organic solvent; performing gradient elution with 0.1% formic acid acetonitrile and 0.1% formic acid water as mobile phases; quantitatively detecting at least 37 saponin components including ginsenoside F5, notoginsenoside Fd, ginsenoside Rb3, ginsenoside F3 and pseudo-ginsenoside F11 in a multi-reaction monitoring mode; inputting quantitative data into a pre-trained extreme gradient boosting model; and determining that the sample is an adulterated sample when the predicted mass proportion of the adulterated rootlets reaches or exceeds 10%. The application breaks through the technical bottleneck that it is difficult to identify homologous non-medicinal parts by relying on only a single or a small number of components through the synergy of multi-index saponin quantification and machine learning discrimination, and has the characteristics of high sensitivity, strong specificity, accurate and reliable results, and provides an efficient and intelligent detection method for the authenticity identification and quality control of panax notoginseng powder and related traditional Chinese medicine preparations.

Description

Technical Field

[0001] This invention relates to the field of traditional Chinese medicine analysis and quality control technology, specifically to a method for detecting adulterated non-medicinal parts in Panax notoginseng and its preparations. In particular, it relates to an intelligent detection method based on a combination of liquid chromatography-triple quadrupole tandem mass spectrometry and extreme gradient boosting machine learning algorithm, for identifying illegally adulterated non-medicinal parts such as rootlets in Panax notoginseng powder or its preparations. Background Technology

[0002] Panax notoginseng (Burk.) FH Chen, belonging to the genus Panax in the family Araliaceae, is a traditional and precious Chinese medicinal herb. It possesses significant effects such as dispersing blood stasis, stopping bleeding, reducing swelling, and relieving pain, and is widely used in the prevention and treatment of cardiovascular and cerebrovascular diseases. It is a core ingredient in many well-known traditional Chinese medicine preparations. According to the current Chinese Pharmacopoeia and relevant international pharmacopoeia standards, the legally permitted medicinal parts of Panax notoginseng are limited to the main root, lateral roots, and rhizome. The fibrous roots are explicitly excluded from the officially permitted medicinal parts due to a lack of traditional medicinal records, differences in chemical composition compared to the medicinal parts, and limited related pharmacological and toxicological data.

[0003] In recent years, with the continuous growth in demand for clinical and health-promoting applications of Panax notoginseng, coupled with production capacity limitations such as its long growth cycle and continuous cropping obstacles, the contradiction between market supply and demand has become increasingly prominent. Driven by economic interests, the illegal adulteration of Panax notoginseng powder and other terminal products, as well as related compound preparations, with low-cost and easily obtainable non-medicinal parts such as fibrous roots has become increasingly serious. This type of adulteration is also known as homologous concealed adulteration because the fibrous roots and medicinal parts originate from the same plant. The content of effective active ingredients such as ginsenosides in the fibrous roots is significantly lower than that in the legally designated medicinal parts. This adulteration not only seriously dilutes the medicinal material basis of the product and harms consumer rights, but may also bring unpredictable clinical drug safety risks due to differences in composition.

[0004] To address the aforementioned issue of homologous adulteration, existing conventional detection techniques have significant limitations. Because fibrous roots share a completely identical genetic background with the main root, lateral roots, and rhizomes—all medicinal parts—their chemical composition is highly similar qualitatively, with only minor differences in the relative content and proportion of certain saponins. Traditional microscopic identification methods rely on the morphological characteristics of intact plant tissues, making it difficult to effectively identify Panax notoginseng powder that has been pulverized into a non-organic state or highly processed preparations. Existing thin-layer chromatography and conventional high-performance liquid chromatography typically use the presence or total amount of a few indicator components, such as ginsenoside Rg1, ginsenoside Rb1, and notoginsenoside R1, as the basis for judgment. This makes it difficult to effectively separate and accurately quantify homologous heteroplasms with very slight differences in content, resulting in severely insufficient sensitivity and specificity. Furthermore, although some studies have attempted to introduce multivariate statistical methods such as principal component analysis or partial least squares discriminant analysis for auxiliary discrimination, these traditional statistical methods are mainly based on the assumption that the data are normally distributed and that there is a linear correlation between variables. They cannot effectively capture and analyze the highly overlapping nonlinear metabolic characteristics between homologous parts in complex Chinese herbal medicine matrices, and are therefore prone to misjudgment results. Therefore, the industry urgently needs a detection method that combines high sensitivity, high specificity, and intelligent nonlinear judgment capabilities to effectively solve the technical problem of identifying illegally added non-medicinal rootlets in Panax notoginseng and its preparations. Summary of the Invention

[0005] This invention provides a method for detecting adulteration of non-medicinal fibrous roots in Panax notoginseng or its preparations, the method comprising the following steps:

[0006] Step (1), Sample pretreatment: Take the sample of Panax notoginseng powder or preparation containing Panax notoginseng to be tested, extract it with an organic solvent or its aqueous solution to obtain the test solution;

[0007] Step (2), Data Acquisition: The test solution was detected by liquid chromatography-triple quadrupole tandem mass spectrometry. Gradient elution was performed using 0.1% formic acid acetonitrile solution as mobile phase A and 0.1% formic acid aqueous solution as mobile phase B to obtain quantitative data of the target saponin components. The target saponin components include at least ginsenoside F5, notoginsenoside Fd, ginsenoside Rb3, ginsenoside F3 and pseudoginsenoside F11.

[0008] Step (3), intelligent determination: Input the quantitative data of the target saponin component measured in step (2) into the pre-trained Extreme Gradient Boosting (XGBoost) model. When the model predicts that the mass ratio of fibrous roots in the sample reaches or exceeds the threshold of 10%, the sample is determined to be an adulterated sample.

[0009] As a further definition of the preparation method of the test solution in step (1) above, the specific preparation method of the test solution is as follows: take Panax notoginseng powder or preparation sample, crush it through a No. 4 sieve, accurately weigh 0.6g of powder, and convert the preparation according to the Panax notoginseng content, and place it in a stoppered conical flask; accurately add 50ml of methanol; perform reflux extraction in a water bath at 80℃, cool, weigh again, make up the lost weight with methanol, shake well and filter through a 0.22μm microporous membrane, and take the filtrate to obtain the test solution.

[0010] As a further definition of the target saponin components in step (2) above, there are a total of 37 target saponin components, specifically including: Panax notoginseng saponin R3, Panax notoginseng saponin FP1, Panax notoginseng saponin R1, Ginsenoside Rg1, Ginsenoside Re, Ginsenoside Rf, Pseudoginsenoside F11, Panax notoginseng saponin R4, Panax notoginseng saponin Fa, Ginsenoside F5, Ginsenoside Rg2, Panax notoginseng saponin R2, Ginsenoside F3, Ginsenoside Ra3, Ginsenoside Rb1, Ginsenoside Rh 1. Ginsenoside Rc, Ginsenoside Ro, Panax notoginseng saponin Fc, Ginsenoside Rb2, Ginsenoside F1, Ginsenoside Rb3, Bambusaginin IVa, Ginsenoside Rd, Panax notoginseng saponin Fe, Panax notoginseng saponin Fd, Ginsenoside Rg6, Ginsenoside Rk3, Ginsenoside F4, Ginsenoside Rh4, Ginsenoside F2, Ginsenoside Rg3, Ginsenoside CK, Ginsenoside Rk1, Ginsenoside Rg5, Ginsenoside Rh2 and Ginsenoside Rk2.

[0011] As a further limitation on the liquid chromatography conditions in step (2) above, the liquid chromatography conditions are as follows: the chromatographic column is an octadecylsilane-bonded silica reversed-phase column with a specification of 2.1×150mm and 1.7μm; mobile phase A is 0.1% formic acid aqueous solution, and mobile phase B is 0.1% formic acid acetonitrile solution; the gradient elution program is as follows: 0-1 min, 22%B; 1-7 min, 22%-25%B; 7-9 min, 25%-30%B; 9-17 min, 30%-33%B; 17-28 min, 33%-47%B; 28-32 min, 47%-55%B; 32-34 min, 55%-60%B; 34-37 min, 60%-95%B.

[0012] As a preferred embodiment of the above-mentioned chromatographic column, the chromatographic column is a Waters BEH C18 column.

[0013] As a further limitation on the mass spectrometry conditions in step (2) above, the detection conditions of the triple quadrupole tandem mass spectrometer are as follows: an electrospray ionization source is used, negative ion scanning mode is used, and multiple reaction monitoring mode is used; the ion source parameters are: drying gas temperature 300℃, drying gas flow rate 5L / min, nebulizing gas pressure 40psi, sheath gas temperature 350℃, sheath gas flow rate 11L / min, capillary voltage 3.5KV, and nozzle voltage 500V.

[0014] As a further specification of the multiple reaction monitoring (MRM) mass spectrometry parameters for the five key quality markers in step (2) above, the MRM parameters for the five key quality markers are as follows: Panax notoginseng saponin Fd: parent ion m / z 961.5, adduct ion [M+HCOO]⁻, daughter ions m / z 915.5 and 783.5, Fragmentor voltage 205V, collision energies 25V and 37V respectively; Ginsenoside F5: parent ion m / z 815.5, adduct ion [M+HCOO]⁻, daughter ion m / z 149.0, Fragmentor voltage 130V, collision energy 34V; Ginsenoside Rb3: parent ion... Ginsenoside F3: m / z 1077.6, adduct ion [M+HCOO]⁻, daughter ions m / z 783.5 and 149.0, Fragmentor voltage 190V, collision energies 54V and 58V respectively; Ginsenoside F11: parent ion m / z 815.5, adduct ion [M+HCOO]⁻, daughter ion m / z 149.0, Fragmentor voltage 140V, collision energy 34V; Ginsenoside F11: parent ion m / z 799.5, adduct ion [MH]⁻, daughter ions m / z 653.5 and 101.0, Fragmentor voltage 190V, collision energies 34V and 50V respectively.

[0015] As a further definition of the XGBoost model training process in step (3) above, the XGBoost model training process includes: obtaining pure Panax notoginseng samples of known classification and adulterated samples with added rootlets as training sets; extracting and obtaining quantitative data of target saponin components in each sample for preprocessing; using the ten-fold cross-validation method to perform hyperparameter tuning on the XGBoost model and training to establish an adulteration identification model; and confirming through SHAP analysis that the five key quality markers have core feature contributions to adulteration determination.

[0016] As an implementation of the XGBoost model in step (3) above, the XGBoost model is a global adulteration identification model trained based on the quantitative data of the 37 target saponin components as input features.

[0017] As another implementation of the XGBoost model in step (3) above, the XGBoost model is an optimized adulteration identification model trained solely on the quantitative data of the five key quality markers as input features.

[0018] As a further definition of the above five key quality markers, among the five key quality markers, notoginsenoside Fd exhibits high specificity and is almost exclusively associated with adulteration categories, making it the preferred key marker for identifying adulterated samples.

[0019] As a further definition of the sample type in step (1) above, the preparation containing Panax notoginseng is a tablet, capsule or granule.

[0020] As a further limitation on the above-mentioned preparation, the preparation is red pill tablets or compound Danshen tablets.

[0021] This invention addresses the industry challenge of adulteration of non-medicinal parts of Panax notoginseng and its preparations by creatively proposing a detection method integrating targeted quantification and intelligent identification. The core concept of this method lies in the fact that since the medicinal and non-medicinal parts are highly similar in macroscopic chemical composition and cannot be distinguished by conventional indicator component detection, it is necessary to start from the differences in microscopic, multidimensional saponin metabolic profiles and employ algorithms capable of resolving complex nonlinear relationships to capture these differences. The technical solution of this invention revolves around this concept, and its essence is a fusion of "targeted metabolomics features + interpretable machine learning".

[0022] The LC-TQ-MS / MS targeted quantitative analysis method established in this invention exhibits extremely excellent linearity for target saponin components (correlation coefficient R² reaches 0.9905–0.9995); extremely high detection sensitivity, with a limit of detection (LOD) as low as 0.01–1.15 ng / mL and a limit of quantitation (LOQ) as low as 0.02–3.85 ng / mL. Simultaneously, the method demonstrates excellent stability, with intra-day precision (RSD) of 0.07%–14.95% and inter-day precision of 0.22%–8.50%; intra-day accuracy (expressed as relative error, RE) of 81.07%–117.79% and inter-day accuracy of 81.09%–119.59%; and good sample stability over 48 hours (RSD range 0.75%–6.83%). This provides a solid technical guarantee for the accurate quantification of trace differential markers.

[0023] The technical features of the present invention will be described in detail below.

[0024] In step (1), the sample pretreatment stage, the purpose of this invention is to efficiently and stably extract the target saponin components from Panax notoginseng powder or its complex formulation matrix. Saponins in Panax notoginseng are the main secondary metabolites and the chemical basis for distinguishing different tissue parts. This invention uses organic solvents or their aqueous solutions for extraction, preferably methanol, ethanol, or acetonitrile, with methanol being the most preferred. The extraction method can be reflux extraction or ultrasonic extraction, preferably water bath reflux extraction to obtain stable extraction efficiency. The specific preparation method of the test solution has been carefully optimized in this invention. The Panax notoginseng powder or formulation sample is crushed through a No. 4 sieve to ensure the uniformity of sample particle size, increase the extraction contact area, and ensure complete extraction. The No. 4 sieve corresponds to the pharmacopoeia standard, and its pore size is about 250 μm, which can meet the accuracy requirements of subsequent quantitative analysis. Accurately weighing 0.6 g of powder is a balance between the amount required for routine detection and the sensitivity of the instrument; for formulations containing Panax notoginseng, the amount of sample is converted according to the Panax notoginseng content indicated in the formulation specification, which ensures the comparability of the test results of different dosage forms. Extraction was performed by adding 50 ml of methanol and refluxing in an 80°C water bath. Methanol has good solubility and extraction efficiency for ginsenosides, and reflux in an 80°C water bath accelerates the extraction kinetics. After cooling, the weight loss was replenished to eliminate volume errors caused by solvent expansion and contraction and evaporation, ensuring quantitative accuracy. Filtration through a 0.22 μm microporous membrane is a common requirement for high-performance liquid chromatography and mass spectrometry, effectively removing particulate impurities that may clog the column or contaminate the ion source.

[0025] In step (2), the data acquisition stage, the purpose of this invention is to perform simultaneous quantitative analysis of saponin components covering multiple structural types in the test sample with high sensitivity and high selectivity. Liquid chromatography-triple quadrupole tandem mass spectrometry (LC-MS / MS) is chosen because it combines the high separation capability of LC with the excellent quantitative selectivity and sensitivity brought by the multiple reaction monitoring mode of triple quadrupole mass spectrometry. Facing the challenge of separating numerous isomer saponins in Panax notoginseng, this invention systematically optimizes the chromatographic conditions. An octadecylsilane-bonded silica reversed-phase column is selected as the stationary phase, whose broad hydrophobic selectivity is suitable for the separation of saponin compounds. The preferred column size is 2.1 × 150 mm, 1.7 μm, more preferably a Waters BEHC18 column. This sub-2-micron particle size column can provide a higher theoretical plate number, achieving high-resolution separation of complex components. The mobile phase system is selected as 0.1% formic acid aqueous solution and 0.1% formic acid acetonitrile solution. Adding 0.1% formic acid as a modifier to the mobile phase effectively inhibits the ionization of carboxyl or hydroxyl groups in saponin molecules, improves peak shape, and promotes the formation of characteristic formate adduct ions [M+HCOO]⁻ in negative ion electrospray mode, thereby improving the mass spectrometry response intensity and stability of saponin components. The designed refined gradient elution program, which progressively increases the proportion of organic phase B from 22% to 95% in stages over a period of 0 to 37 minutes, was optimized through extensive experiments. This program enables baseline or near-baseline separation of 37 saponins within 37 minutes, particularly for several pairs of difficult-to-separate isomers such as ginsenoside Rg6 and ginsenoside F4, and ginsenoside Rk1 and ginsenoside Rg5, laying a solid foundation for subsequent accurate quantification.

[0026] For mass spectrometry detection, an electrospray ionization source is employed, operating in negative ion scanning mode. This is because ginsenoside molecules contain multiple hydroxyl groups, readily losing protons or forming formate adducts, resulting in higher ionization efficiency and detection sensitivity in negative ion mode. Multiple reaction monitoring (MRM) is used, selecting specific precursor-daughter ion pairs for each target saponin. This is the core mode for achieving high selectivity and high sensitivity quantification in triple quadrupole mass spectrometry, minimizing the influence of other interfering substances in the matrix. The set ion source parameters, including a drying gas temperature of 300℃, a drying gas flow rate of 5 L / min, a nebulizing gas pressure of 40 psi, a sheath gas temperature of 350℃, a sheath gas flow rate of 11 L / min, a capillary voltage of 3.5 kV, and a nozzle voltage of 500 V, are a comprehensive parameter combination optimized for the characteristics of saponin compounds, aiming to stabilize and optimize the electrospray process, ensuring ionization efficiency and long-term signal stability.

[0027] The quantitative target saponin components in this invention are one of the core elements of the technical solution. At the essential element level, the target saponin components include at least five key quality markers: ginsenoside F5, notoginsenoside Fd, ginsenoside Rb3, ginsenoside F3, and ginsenoside F11. These five components are not randomly selected, but scientifically screened from a large library of 37 saponins through subsequent SHAP interpretability analysis. They are key variables that contribute significantly to distinguishing adulterated samples. Among them, notoginsenoside Fd exhibits extremely high specificity. Its content in samples adulterated with fibrous roots differs significantly from that in samples of pure medicinal parts, and is almost solely associated with the adulteration category, making it the preferred core marker for identifying adulterated samples. In a more comprehensive approach, this invention performs targeted quantification of up to 37 saponins. This list of 37 saponins covers major components such as ginsenosides Rb1, Rg1, Re, Rd, and Rc, as well as a series of trace isomers and metabolic intermediates that vary in proportion across different tissues and sites. Examples include notoginsenosides R1, R4, Fa, Fc, Fe, and Fd; ginsenosides F1, F2, F3, F4, and F5; ginsenoside F11; ginsenoside IVa; and rare saponins such as Rg2, Rg3, Rg5, Rg6, Rh1, Rh2, Rh4, Rk1, Rk2, Rk3, and CK. Comprehensive quantitative information on these 37 saponins allows for the construction of a high-dimensional dataset that fully reflects the metabolic fingerprint characteristics of different tissues and sites within Panax notoginseng. For each target saponin, this invention provides its preferred multiple reaction monitoring mass spectrometry parameters, including the parent ion mass-to-charge ratio, the type of adductor ion, the daughter ion mass-to-charge ratio, the fragmentor voltage, and the collision energy. These are the optimal values ​​obtained directly on the instrument using standards, which can ensure the highest sensitivity and specificity of the detection method and are an important guarantee for the standardization and reproducibility of the method of this invention.

[0028] In step (3), the intelligent judgment stage, the purpose of this invention is to use a machine learning model to perform pattern recognition on multidimensional saponin quantitative data, thereby achieving objective and accurate adulteration judgment. The extreme gradient boosting model is chosen as the core classifier because of its excellent performance in processing high-dimensional, nonlinear, and interactive small sample data. Compared with traditional linear analysis models such as partial least squares discriminant analysis, XGBoost can automatically capture the complex nonlinear synergistic changes and threshold effects between saponin components by integrating a large number of decision trees and continuously fitting the residuals of the previous tree. This nonlinearity is precisely the key to distinguishing metabolic differences between different tissue sites of the same origin. Compared with other ensemble learning or deep learning algorithms such as random forest, support vector machine, LightGBM, and CatBoost, the comprehensive experimental verification results show that XGBoost performs best in multiple indicators such as classification accuracy, precision, recall, and F1 score.

[0029] Model training is a rigorous process. The training set consists of samples with known classifications: pure Panax notoginseng powder (identified by botanical experts) and adulterated samples artificially mixed with fibrous roots according to a known gradient ratio. After extracting quantitative data of each saponin from the samples and performing preprocessing such as data standardization, a ten-fold cross-validation method is used. The training set is randomly divided into ten parts, and nine parts are used to train the model, while one part is used to validate the model. Through iterative optimization, the optimal set of hyperparameters for the XGBoost model is determined, including parameters such as maximum tree depth, learning rate, and subsampling ratio. This process effectively prevents overfitting and improves the model's generalization ability.

[0030] After training, this invention creatively introduces SHAP analysis to interpret the model. Based on game theory, SHAP analysis can calculate the contribution of each input feature to each prediction result. Through SHAP global feature importance analysis, not only were the aforementioned five key quality markers successfully screened from 37 saponin features, but the correlation direction and strength between each marker and different classification results were also clearly revealed. For example, it was confirmed that N-Fd almost exclusively drives the "adulteration" judgment, G-F3 is highly correlated with the "cut-out" category, and G-Rb3 and PseG-F11 drive the "branch root" category prediction. This interpretability is crucial for the establishment of drug regulatory standards because it transforms the "black box" decision-making process of the algorithm into a traceable and verifiable chemical basis. In the final judgment, the target saponin data of the sample to be tested is input into the trained model, and the model outputs a probability prediction value for the presence of admixtured rootlets. This invention has established that when the proportion of admixtured rootlets corresponding to this predicted probability reaches or exceeds 10%, the sample is judged as adulterated. The 10% threshold setting strikes a reasonable balance between detection sensitivity and the practical need to combat counterfeiting. It is sufficient to identify adulteration with real economic benefits while avoiding false positives due to excessive sensitivity.

[0031] This invention provides two application paths for the model, forming a flexible dual-track judgment scheme. Path one utilizes a global adulteration identification model trained with data from 37 target saponins for prediction. This model leverages the most comprehensive chemical information, boasts the highest identification accuracy, and is suitable for scenarios requiring the highest detection precision or confirmatory testing of suspicious samples. Path two utilizes an optimized adulteration identification model trained with data from only 5 key quality markers for prediction. This model significantly reduces the number of saponins to be detected, greatly reducing detection costs and analysis time, improving the method's scalability and efficiency for routine high-throughput detection, while still maintaining an extremely high identification rate for adulterated samples. The two paths can be flexibly selected or mutually verified according to actual needs.

[0032] The method of this invention is applicable to various test samples containing Panax notoginseng. The forms include not only Panax notoginseng raw powder, but also traditional Chinese medicine preparations made from Panax notoginseng through various formulation processes. The dosage forms of the preparations include, but are not limited to, tablets, capsules, and granules. Specific preparation varieties include red pill tablets, compound Danshen tablets, etc.

[0033] The technical solution provided by this invention effectively addresses the deficiencies and technical problems in existing technologies. Existing microscopic identification methods are completely unsuitable for powders and preparations; traditional thin-layer chromatography and high-performance liquid chromatography rely on a few indicator components, and because the content ranges of these indicator components in fibrous roots and medicinal parts often overlap, effective differentiation is impossible; linear statistical methods such as principal component analysis and partial least squares discriminant analysis are incapable of analyzing the highly overlapping nonlinear metabolic characteristics between homologous tissues, often resulting in adulterated and non-adulterated samples appearing mixed up on the score map. The core solution of this invention lies in: firstly, utilizing the high resolution and sensitivity of liquid chromatography-triple quadrupole tandem mass spectrometry, it achieves precise quantification of 37 saponin profiles, including multiple trace isomers, reflecting subtle metabolic differences in different tissue parts, maximizing the capture and presentation of the intrinsic differences in samples from a chemical analysis perspective. Next, the XGBoost nonlinear machine learning algorithm is introduced to model high-dimensional saponin quantitative data. This algorithm can efficiently learn and utilize the complex, nonlinear synergistic change patterns and discrimination thresholds among various saponin components, thus achieving a breakthrough in the traditionally intractable problem of homologous tissue classification. Furthermore, through SHAP analysis, five key quality markers with chemical entity significance directly related to adulteration are extracted from the complex model, such as the highly specific Panax notoginseng saponin Fd. This not only provides a more efficient and concise detection scheme, but also provides clear material basis and "white-box" interpretation of the detection results in accordance with the requirements of drug regulatory science, completely solving the pain point of traditional biochemical identification and ordinary machine learning models lacking molecular-level interpretability. It is this deep organic integration of "cutting-edge analytical technology to obtain deep differential data" and "intelligent and interpretable algorithm to analyze nonlinear patterns" that enables this invention to accurately and reliably identify the fibrous roots of Panax notoginseng and its preparations with an adulteration rate as low as 10%, with extremely high sensitivity, specificity, and resistance to matrix interference. Attached Figure Description

[0034] Figure 1 This is the total ion chromatogram.

[0035] Figure 2 Ion chromatograms of 37 ginsenosides extracted.

[0036] Figure 3 This is a confusion matrix diagram of 37 ginsenosides based on the XGBoost model.

[0037] Figure 4 This is a multi-class ROC curve of 37 ginsenosides based on the XGBoost model.

[0038] Figure 5 This is a global feature importance diagram of 37 ginsenosides based on the mean absolute SHAP value.

[0039] Figure 6 This is a confusion matrix diagram of five quality markers based on the XGBoost model.

[0040] Figure 7 This is a multi-class ROC curve for five quality markers based on the XGBoost model.

[0041] Figure 8 A t-SNE visualization of the entire dataset based on five quality indicators.

[0042] Figure 9 This is a global feature importance map of five quality markers based on the mean absolute SHAP value.

[0043] Figure 10 PLS-DA score plot based on quantitative data of 37 ginsenosides.

[0044] Figure 11 Radar charts showing the performance of different machine learning models.

[0045] Figure 12 Heatmaps showing the performance of different machine learning models. Detailed Implementation

[0046] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0047] like Figures 1 to 12 As shown, in one specific embodiment of the present invention, a method for detecting adulterated non-medicinal rootlets in Panax notoginseng or its preparations is provided. This method integrates optimized pretreatment process, high-sensitivity liquid chromatography-triple quadrupole tandem mass spectrometry targeted quantitative analysis, and intelligent judgment based on interpretable machine learning. The complete operation process is as follows:

[0048] First, sample pretreatment and preparation of the test solution are performed. Take the powder of Panax notoginseng to be tested, or traditional Chinese medicine preparations such as tablets, capsules, or granules made from Panax notoginseng, such as Hongyao tablets or Compound Danshen tablets, and pulverize them and pass them through a No. 4 sieve to ensure uniform particle size and complete extraction. Accurately weigh 0.6 grams of the sieved powder. For preparations containing Panax notoginseng, convert the amount according to the Panax notoginseng content indicated on the specification before sampling. Place the sample in a stoppered conical flask. Accurately add 50 ml of methanol, weigh it, and let it stand overnight or directly perform water bath reflux extraction.

[0049] In this embodiment, reflux extraction at 80°C for 2 hours is preferred. Under these conditions, methanol exhibits good solubility and extraction efficiency for various ginsenosides in Panax notoginseng, and reflux accelerates the extraction kinetics, resulting in high extraction efficiency and good reproducibility. After cooling, the weight is recalculated, and methanol is used to compensate for the weight loss caused by evaporation or thermal expansion and contraction during heating and cooling to eliminate solvent volume errors and ensure the accuracy of subsequent quantitative analysis. After shaking well, the extract is filtered, and the filtrate is then filtered through a 0.22-micron microporous membrane. The resulting filtrate is used as the test solution. The preparation method for the reference solution is as follows: accurately weigh appropriate amounts of the target saponin component reference standards, dissolve and dilute them in methanol to prepare a single reference standard stock solution with a concentration of 1 mg / mL. Then, mix the stock solutions and dilute them with methanol to prepare a series of mixed standard working solutions of different concentrations, which are stored at 4°C for later use.

[0050] The above pretreatment steps are simple to operate, and the validation results show that the method recovery rate can be stably maintained between 81.83% and 118.83%, and the matrix effect test results are strictly controlled between 79.46% and 118.63%, which fully demonstrates that complex matrices do not significantly interfere with the quantification of the analyte.

[0051] Next, liquid chromatography-triple quadrupole tandem mass spectrometry data acquisition was performed.

[0052] This embodiment uses a liquid chromatography-triple quadrupole tandem mass spectrometry (LC-MS) system equipped with an autosampler, such as the Agilent 1290-6475. The specific LC conditions are as follows: an octadecylsilane-bonded silica gel reversed-phase column, preferably a Waters BEH C18 column with dimensions of 2.1 × 150 mm and a packing particle size of 1.7 μm. This column, after comparative screening, is significantly superior to other conventional reversed-phase columns in terms of peak shape symmetry, isomer resolution, and reproducibility. Mobile phase A is an aqueous solution containing 0.1% formic acid, and mobile phase B is an acetonitrile solution containing 0.1% formic acid. The addition of a small amount of formic acid effectively improves the ionization efficiency of saponin components, resulting in sharper chromatographic peaks. The gradient elution program is as follows: 0 to 1 minute, 22% B; 1 to 7 minutes, 22% to 25% B; 7 to 9 minutes, 25% to 30% B; 9 to 17 minutes, 30% to 33% B; 17 to 28 minutes, 33% to 47% B; 28 to 32 minutes, 47% to 55% B; 32 to 34 minutes, 55% to 60% B; 34 to 37 minutes, 60% to 95% B. The flow rate is set at 0.35 mL / min, the column temperature is maintained at 40°C, and the injection volume is typically 2 μL.

[0053] like Figure 1 and Figure 2 As shown, the separation of saponin components in Panax notoginseng is difficult due to the large number of isomers (such as G-Rg6, G-F4, G-Rk1, G-Rg5). This invention compares four conventional reversed-phase chromatography columns: Waters ACQUITY UPLCORTECS C18 (2.1 mm × 150 mm, 1.7 μm), Waters Luna Omega C18 (2.1 mm × 150 mm, 1.6 μm), Waters ACQUITY UPLC BEH Shield RP C18 (2.1 mm × 100 mm, 1.7 μm), and Waters ACQUITY UPLC BEH C18 (2.1 mm × 150 mm, 1.7 μm). The results showed that, compared with other column types, the WatersACQUITY UPLC BEH C18 (2.1 mm × 150 mm, 1.7 μm) column performed best in terms of peak shape symmetry, resolution, and reproducibility, especially in the excellent separation of isomers. The gradient design described above, after fine optimization, achieved effective separation of 37 saponin components, particularly good baseline separation of isomers such as ginsenosides Rg6 and F4, laying the foundation for accurate subsequent quantification.

[0054] Mass spectrometry conditions were set as follows: an electrospray ionization source was used, operating in negative ion scanning mode, with multiple reaction monitoring (MRM) as the monitoring mode. The general parameters for the ion source were: drying gas temperature 300°C, drying gas flow rate 5 L / min, nebulizing gas pressure 40 psi, sheath gas temperature 350°C, sheath gas flow rate 11 L / min, capillary voltage 3.5 kV, and nozzle voltage 500 V. Dedicated MRM parameters were established for 37 saponins targeting the specified saponin components.

[0055] The target saponin components of this invention include at least five key quality markers, namely ginsenoside F5, notoginsenoside Fd, ginsenoside Rb3, ginsenoside F3, and pseudoginsenoside F11. The multiple reaction monitoring (MRM) parameters for these five key quality markers were specifically optimized as follows: For Panax notoginseng saponin Fd, the parent ion m / z was 961.5, the addendum ion form was [M+HCOO]⁻, and the daughter ions were selected at m / z 915.5 and 783.5, respectively. The fragmentator voltage was 205 V, and the collision energies were 25 V and 37 V, respectively. For Ginsenoside F5, the parent ion m / z was 815.5, the addendum ion form was [M+HCOO]⁻, and the daughter ion was m / z 149.0. The fragmentator voltage was 130 V, and the collision energy was 34 V. For Ginsenoside Rb3, the parent ion m / z was 1077.6, the addendum ion form was [M+HCOO]⁻, and the daughter ions were m / z 783.5 and 149.0, respectively. The fragmentator voltage was 190 V, and the collision energies were 54 V and 58 V, respectively. For Ginsenoside F3, the parent ion m / z was 961.5, the addendum ion form was [M+HCOO]⁻, and the daughter ions were m / z 783.5 and 149.0, respectively. The fragmentator voltage was 190 V, and the collision energies were 54 V and 58 V, respectively. The target saponin F11 has a parent ion of m / z 799.5, an adduct ion of m / z [M+HCOO]⁻, a daughter ion of m / z 149.0, a fragmentor voltage of 140 V, and a collision energy of 34 V. The target saponin F11 has a parent ion of m / z 799.5, an adduct ion of m / z [MH]⁻, daughter ions of m / z 653.5 and 101.0, a fragmentor voltage of 190 V, and collision energies of 34 V and 50 V, respectively. The mass spectrometry parameters for the remaining 32 target saponin components were also optimized and set according to their optimal adduct ion forms and characteristic daughter ions. Quantitative data of the target saponin components can be obtained using the above conditions.

[0056] In method validation, the established LC-TQ-MS / MS targeted quantitative analysis method demonstrated excellent performance. The 37 saponin components showed good linearity within their respective concentration ranges, with correlation coefficients (R²) ranging from 0.9905 to 0.9995. The detection sensitivity was extremely high, with limits of detection as low as 0.01 to 1.15 ng / mL and limits of quantitation as low as 0.02 to 3.85 ng / mL. Intra-day precision (RSD) ranged from 0.07% to 14.95%, and inter-day precision ranged from 0.22% to 8.50%. Accuracy (expressed as relative error) ranged from 81.07% to 117.79% intra-day and 81.09% to 119.59% inter-day. The test solution exhibited good stability over 48 hours, with RSDs ranging from 0.75% to 6.83%. The recovery rate of this method remained stable at 81.83% to 118.83%, confirming the consistency and extremely high reproducibility of the extraction method. The matrix effect test results ranged from 79.46% to 118.63%, clearly demonstrating that the complex matrix of Panax notoginseng powder and its preparations did not significantly interfere with the quantitative determination of the ginsenosides targeted by this invention. This fully demonstrates that the quantitative accuracy, reproducibility, and resistance to matrix interference of the method of this invention meet the requirements of trace analysis, and is particularly suitable for the detection of complex matrix samples such as Panax notoginseng powder and its compound preparations. It is precisely because of such high quantitative accuracy and sensitivity that this invention can accurately capture the minute component concentration differences between medicinal and non-medicinal parts (rootlets) that were previously overlooked by conventional HPLC, thus overcoming the deficiency of existing technologies in distinguishing homologous non-medicinal parts.

[0057] The next step is machine learning-based intelligent judgment, which requires prior training and establishing an extreme gradient boosting model. During training, pure Panax notoginseng samples with known and clearly defined classifications and adulterated samples containing fibrous roots are obtained as training sets. Adulterated samples are prepared by adding fibrous root powder to genuine Panax notoginseng main root powder at mass ratios of 5%, 10%, 20%, 50%, 80%, and 100%. This invention defines samples with an fibrous root content of 10% or higher as adulterated samples. Quantitative data of the aforementioned target saponin components are extracted from each training set sample and preprocessed, such as by standardization to eliminate differences in the dimensionality of different saponins. The hyperparameters of the XGBoost model are fine-tuned using a 10-fold cross-validation method to train a classification model capable of distinguishing different tissue parts and adulteration situations.

[0058] After training, two models with different feature dimensions can be formed: one is a global adulteration identification model based on quantitative data of all 37 target saponin components as input features, and the other is an optimized adulteration identification model trained only on quantitative data of the five key quality markers, namely ginsenoside F5, notoginsenoside Fd, ginsenoside Rb3, ginsenoside F3 and pseudoginsenoside F11, as input features.

[0059] like Figures 3 to 9 As shown, in the model evaluation, the global model achieved 100% accuracy in identifying adulterated samples, with an area under the receiver operating characteristic (AUC) of 0.9989. The optimized model also achieved 100% accuracy in identifying adulterated samples, with an AUC of 0.9762. Furthermore, this invention employs SHAP additive interpretability analysis to analyze the model, breaking the "black box" limitation of machine learning and meeting the interpretability requirements of pharmaceutical testing. SHAP analysis results show that notoginsenoside Fd and ginsenoside F5 make significant contributions to adulteration classification. Notoginsenoside Fd exhibits high specificity, with its level changes almost exclusively correlated with the adulteration category, confirming it as the preferred key quality marker for identifying adulterated samples. Ginsenoside F3 is mainly related to the cut category, while ginsenoside Rb3 and pseudoginsenoside F11 drive the identification of the root category. This finding provides a clear chemical basis for the subsequent development of national or industry supplementary testing standards.

[0060] When judging the actual samples to be tested, the quantitative data of the target saponin components in the Panax notoginseng powder or preparation sample obtained according to the above steps are input into a pre-trained XGBoost model. When the model predicts that the mass proportion of fibrous roots in the sample reaches or exceeds the threshold of 10%, the sample is judged as an adulterated sample. In the screening of actual commercially available samples, the optimized XGBoost model based on 5 key quality markers was applied to predict the adulteration of 48 batches of commercially available Panax notoginseng powder, successfully identifying 34 batches of samples as adulterated samples, with an average confidence score of 0.894; 36 batches of red pills with Panax notoginseng as the core raw material were tested, of which 11 batches were judged to be adulterated with fibrous roots, while no adulteration was found in the 25 batches of raw materials from well-known large pharmaceutical companies. This confirms that the method of the present invention has a strong generalization ability and practical anti-counterfeiting effectiveness even in complex formulation environments.

[0061] Comparative Analysis: A Comparison of Existing Conventional Statistical Analysis with Various Machine Learning Algorithms

[0062] Currently, the industry commonly uses linear statistical methods such as partial least squares discriminant analysis (PLS-DA) or conventional support vector machines (SVM) for the identification of traditional Chinese medicine.

[0063] 1. Comparison with the traditional PLS-DA linear model: such as Figure 10 As shown, using quantitative data input into the PLS-DA model, the results reveal a significant overlap between non-medicinal parts (fibrous roots / FR) and medicinal parts (lateral roots / BR, main roots / MR) on the spatial score map. This indicates that traditional linear dimensionality reduction methods cannot effectively resolve the highly nonlinear component overlap problem between homologous tissues.

[0064] 2. Compared with other machine learning algorithms: such as Figure 11 and Figure 12 As shown, Support Vector Machine (SVM), Random Forest (RF), LightGBM, and CatBoost models were constructed simultaneously for horizontal comparative evaluation. The results show that the preferred XGBoost model of this invention performs best in accuracy, AUC, precision, recall, and F1 score. Conventional machine learning black boxes are prone to misclassification of feature weights, while the optimized XGBoost model constructed in this invention, containing five key quality markers, breaks the black box through SHAP interpretation, possessing both extremely high transparency and 100% specificity in detecting adulteration. This demonstrates that the solution of this invention has unparalleled technical advantages in the field of identifying homologous hidden adulteration.

[0065] In summary, this invention organically combines high-performance liquid chromatography-triple quadrupole tandem mass spectrometry (HPLC-TQMS), information-rich multi-component saponin characteristic spectra, and an interpretable extreme gradient boosting machine learning model. This allows for the precise capture of trace differences in the proportions of metabolic components between homologous medicinal and non-medicinal rootlets. It not only solves the technical pain point of traditional microscopic, chromatographic, and linear statistical methods failing to identify homologous hidden adulteration, but also possesses extremely high sensitivity, specificity, accuracy, and resistance to matrix interference. This provides a reliable, intelligent, and scientifically interpretable detection method for the quality control of Panax notoginseng and its preparations.

[0066] The specific embodiments described in this invention are merely illustrative of the invention and are not intended to limit it. Those skilled in the art can make modifications to these embodiments without contributing any inventive step after reading this specification, but such modifications are protected by patent law as long as they fall within the scope of the claims of this invention.

Claims

1. A method for detecting adulteration of non-medicinal fibrous roots in Panax notoginseng or its preparations, characterized in that, Includes the following steps: (1) Sample pretreatment: Take the sample of Panax notoginseng powder or preparation containing Panax notoginseng to be tested, extract it with organic solvent or its aqueous solution to obtain the test solution; (2) Data acquisition: The test solution was detected by liquid chromatography-triple quadrupole tandem mass spectrometry. Gradient elution was performed using 0.1% formic acid acetonitrile solution as mobile phase A and 0.1% formic acid aqueous solution as mobile phase B to obtain quantitative data of the target saponin components. The target saponin components include at least ginsenoside F5, notoginsenoside Fd, ginsenoside Rb3, ginsenoside F3 and pseudoginsenoside F11. (3) Intelligent judgment: The quantitative data of the target saponin component measured in step (2) is input into the pre-trained extreme gradient boosting XGBoost model. When the model predicts that the mass ratio of the fibrous roots in the sample reaches or exceeds the threshold of 10%, the sample is judged to be an adulterated sample.

2. The method according to claim 1, characterized in that, The specific preparation method of the test solution in step (1) is as follows: Take Panax notoginseng powder or preparation sample, crush it through a No. 4 sieve, accurately weigh 0.6g of powder, and convert the preparation according to the amount of Panax notoginseng, and place it in a stoppered conical flask; accurately add 50ml of methanol; perform reflux extraction in a water bath at 80℃, cool, weigh again, make up the lost weight with methanol, shake well and filter through a 0.22μm microporous membrane, and take the filtrate to obtain the test solution.

3. The method according to claim 1, characterized in that, The target saponin components mentioned in step (2) total 37, specifically including: notoginsenoside R3, notoginsenoside FP1, notoginsenoside R1, ginsenoside Rg1, ginsenoside Re, ginsenoside Rf, pseudoginsenoside F11, notoginsenoside R4, notoginsenoside Fa, ginsenoside F5, ginsenoside Rg2, notoginsenoside R2, ginsenoside F3, ginsenoside Ra3, ginsenoside Rb1, ginsenoside Rh1, ginsenoside Rc, Ginsenoside Ro, Panax notoginseng saponin Fc, Ginsenoside Rb2, Ginsenoside F1, Ginsenoside Rb3, Panax notoginseng saponin IVa, Ginsenoside Rd, Panax notoginseng saponin Fe, Panax notoginseng saponin Fd, Ginsenoside Rg6, Ginsenoside Rk3, Ginsenoside F4, Ginsenoside Rh4, Ginsenoside F2, Ginsenoside Rg3, Ginsenoside CK, Ginsenoside Rk1, Ginsenoside Rg5, Ginsenoside Rh2 and Ginsenoside Rk2.

4. The method according to claim 1, characterized in that, The chromatographic conditions for liquid chromatography described in step (2) are as follows: the chromatographic column is an octadecylsilane-bonded silica reversed-phase column with a size of 2.1 × 150 mm and a diameter of 1.7 μm; mobile phase A is 0.1% formic acid aqueous solution, and mobile phase B is 0.1% formic acid acetonitrile solution; the gradient elution program is as follows: 0-1 min, 22% B; 1-7 min, 22%-25% B; 7-9 min, 25%-30% B; 9-17 min, 30%-33% B; 17-28 min, 33%-47% B; 28-32 min, 47%-55% B; 32-34 min, 55%-60% B; 34-37 min, 60%-95% B.

5. The method according to claim 4, characterized in that, The chromatographic column is a Waters BEH C18 column.

6. The method according to claim 1, characterized in that, The detection conditions for the triple quadrupole tandem mass spectrometry described in step (2) are as follows: an electrospray ionization source, negative ion scanning mode, and multiple reaction monitoring mode are used; the ion source parameters are: drying gas temperature 300℃, drying gas flow rate 5 L / min, nebulizing gas pressure 40 psi, sheath gas temperature 350℃, sheath gas flow rate 11 L / min, capillary voltage 3.5 KV, and nozzle voltage 500 V.

7. The method according to claim 1, characterized in that, The multiple reaction monitoring mass spectrometry parameters of the five key quality markers mentioned in step (2) are as follows: N-Fd: Mother ion m / z 961.5, adduct ion [M+HCOO]⁻, daughter ions m / z 915.5 and 783.5, fragmentor voltage 205 V, collision energies 25 V and 37 V respectively; G-F5: Mother ion m / z 815.5, adduct ion [M+HCOO]⁻, daughter ion m / z 149.0, Fragmentor voltage 130V, collision energy 34V; G-Rb3: parent ion m / z 1077.6, adduct ion [M+HCOO]⁻, daughter ions m / z 783.5 and 149.0, fragmentor voltage 190 V, collision energies 54 V and 58 V respectively; G-F3: Mother ion m / z 815.5, adduct ion [M+HCOO]⁻, daughter ion m / z 149.0, Fragmentor voltage 140V, collision energy 34V; PseG-F11: Mother ion m / z 799.5, adduct ion [MH]⁻, daughter ions m / z 653.5 and 101.0, fragmentor voltage 190 V, collision energies 34 V and 50 V respectively.

8. The method according to claim 1, characterized in that, The training process of the extreme gradient boosting XGBoost model in step (3) includes: obtaining pure Panax notoginseng samples with known classification and adulterated samples with added rootlets as training sets; extracting and obtaining quantitative data of the target saponin components in each sample for preprocessing; using the ten-fold cross-validation method to perform hyperparameter tuning on the extreme gradient boosting XGBoost model and training to establish an adulteration identification model; and confirming through SHAP analysis that the five key quality markers have core feature contributions to adulteration determination.

9. The method according to claim 3, characterized in that, The extreme gradient boosting XGBoost model described in step (3) is a global adulteration identification model trained based on the quantitative data of the 37 target saponin components as input features.

10. The method according to claim 1, characterized in that, The extreme gradient boosting XGBoost model described in step (3) is an optimized adulteration identification model trained solely on quantitative data of the five key quality markers as input features.

Images