The standards of the compounds used for quantification and method validation are shown in Supplementary Table S1. Methanol, acetonitrile, and formic acid (LC-MS grade) were obtained from Honeywell (United States). Ultrapure water with a resistivity of 18.2 MΩ cm was produced using the Milli-Q IQ 7000 water cleansing system (Merck Millipore, United States). All additional analytical-grade reagents and chemicals were purchased from Sinopharm Group Chemical Reagents (China). Stock solutions of the standard compounds were prepared in methanol at a concentration of 1 g/L and then stored at −20 °C. Calibration curves were subsequently constructed using these solutions.

William’s Medium E (WME), acetaminophen (APAP), and dimethyl sulfoxide (DMSO) were obtained from Sigma-Aldrich (United States). Fetal bovine serum (FBS), phosphate-buffered saline (PBS), 0.05% trypsin–EDTA solution, penicillin (10,000 μg/mL), and streptomycin (10,000 μg/mL) were purchased from HyClone (United States). The Cell Counting Kit-8 (CCK-8) was obtained from Dojindo Molecular Technologies (China). THSG and E.G., were dissolved in DMSO at concentrations of 700 mg/mL and 200 mg/mL, respectively, and then diluted in the medium. The final concentration of DMSO in the medium was maintained below 0.1% (v/v) (Yu et al., 2017; Lee et al., 2017).

According to the 2020 edition of the Pharmacopoeia of the People’s Republic of China (ChP), qualified PM decoction pieces are characterized by a red-brown or dark brown outer surface, a yellowish-brown or pale reddish-brown fractured surface, and a slightly astringent taste (Committee, 2020). They also exhibit 4–11 roundish abnormal vascular bundles in the bark, arranged in a ring and forming cloud-like patterns. In total, 61 batches of commercial PM samples were randomly collected from five regions in China: Yunnan (YN, n = 11), Anhui (AH, n = 10), Guangdong (GD, n = 10), Guizhou (GZ, n = 20), and Sichuan (SC, n = 10), as detailed in Supplementary Table S2. These samples displayed morphological variations in color, size, and decoction-piece specifications (Figure 1). All samples consisted of dried roots of Polygonum multiflorum Thunb., and their identity was confirmed by the China National Institutes for Food and Drug Control.

The UPLC-ESI-QqQ-MS/MS test solution was prepared as follows: the 61 batches of samples were ground into powder, sieved using a Chinese National Standard Sieve No. 4 (250 µm mesh), and weighed (1 g per sample). Each sample was transferred into a 100 mL stoppered conical flask, to which 50 mL of 70% ethanol (v/v) was added. The flask was weighed, and ultrasonic extraction was performed in a 30 °C water bath for 30 min. After the extract cooled to room temperature, the lost weight was replaced with 70% ethanol. The extract was filtered, and 25 mL of the filtrate was evaporated in a water bath. The residue was dissolved in 5 mL of 70% ethanol and filtered through a 0.22 µm microporous membrane to obtain the UPLC-ESI-QqQ-MS/MS stock solution. A 200 µL aliquot of the stock solution was diluted to 10 mL (i.e., 50-fold dilution) to prepare the working solution. The ultrasonic extraction method using 70% ethanol at 30 °C was selected based on our previous systematic optimization and validation (Lv et al., 2015; Yang et al., 2018; Zhang et al., 2020). This method consistently achieved efficiencies greater than 85% across all 26 target compounds, which included anthraquinones, stilbene glycosides, flavonoids and phenols. Furthermore, the analytes demonstrated stability under the applied conditions, with no significant degradation observed throughout the analysis (Li et al., 2022a).

The preparation of PM solutions for HepaRG cell assays followed the same procedure, involving ultrasonication in 50 mL of 70% ethanol and subsequent evaporation of 25 mL of the filtrate. The residue was dissolved in 0.5 mL of DMSO to obtain the test solution for cell-based studies. For administration in 96-well plates, the DMSO concentration was adjusted to ≤0.1%.

Using a previously validated UPLC-ESI-QqQ-MS/MS method, we quantified 26 compounds in PM, including anthraquinones, stilbene glycosides, flavonoids and phenols (Li et al., 2022a). This method involved comprehensive assessments of extraction recovery for all analytes as well as their stability throughout the analytical process, ensuring the reliability of the quantitative results. Concentrated and diluted PM solutions were analyzed using an LC-20AD system (Shimadzu, Japan), with an injection volume of 1 µL and a column temperature of 27 °C. The column was an ACQUITY UPLC™ HSS C₁₈ column (100 mm × 2.1 mm, 1.8 μm), mobile phase A was 100% acetonitrile, and mobile phase B was a 0.2% formic acid solution. Gradient elution was performed as follows: 0–3.2 min, 10%–15% A; 3.2–6.5 min, 15% A; 6.5–9.7 min, 15%–26% A; 9.7–12 min, 26%–32% A; 12–15 min, 32%–36% A; 15–16.5 min, 36% A; 16.5–31 min, 36%–80% A; 31–35.9 min, 80% A; 35.9–36 min, 80%–10% A; and 36–41 min, 10% A. The flow rate was set at 0.4 mL/min.

Detection was conducted on an LCMS-8050-QqQ-mass-spectrometer (Shimadzu, Japan) operated in both positive and negative ionization modes, with the mode providing the highest compound clarity selected for quantification. The electrospray ionization (ESI) source parameters were set at 300 °C for temperature and 3,000 V for ion spray voltage. For each analyte, the two most abundant ions were selected as precursor ions, and the product ions with the strongest responses were identified through product-ion scanning. For quantification, the two strongest transitions (a quantifier and a qualifier) were selected, and their collision energies were optimized. The method was validated for trueness, repeatability, specificity, linearity, and sensitivity. Further details of the methodological validation are provided in our previous study (Li et al., 2022a).

Quantitative analysis of the chemical composition of each PM batch was performed using the LabSolutions browser (Shimadzu, version 5). Compound concentrations were determined by comparing analyte peak areas with those of corresponding standards, and the concentrations of the 26 compounds were expressed as mean values from three replicates. For chemometric analysis, the preprocessed data were exported to OriginPro 2021 for normalization, and the normalized model was validated for interpretability using variable importance in projection (VIP) and Corr.Coeffs values. Then, heatmaps combined with hierarchical cluster analysis (HCA) were generated using Tbtools to visualize sample distribution patterns. PCA with seven-fold cross-validation was performed using SIMCA-P 14.1 to verify classification reliability and to identify discriminant metabolites associated with grouping (Yin et al., 2024; Ji et al., 2024). To assess the influence of geographic origin on PM chemical composition, one-way analysis of variance (ANOVA) was conducted using OriginPro 2021. The total concentration of the 26 compounds in each sample was used as the dependent variable, and the five geographic origins (Yunnan, Anhui, Guangdong, Guizhou, and Sichuan) were set as the categorical independent variable.

The HepaRG cell line was obtained from Thermo Fisher Scientific (United States), and cells from the 12th to 13th generations were used in the experiments. Cells were cultured in WME supplemented with 10% fetal bovine serum (FBS, v/v) and 1% penicillin–streptomycin (v/v), and maintained at 37 °C in a humidified atmosphere containing 5% CO₂. Cells in the logarithmic growth phase were used for subsequent experiments. To evaluate the potential hepatotoxicity of compliant PM samples, 29 batches were systematically selected from three groups of qualified samples (n = 9 in Group B, n = 11 in Group C, and n = 9 in Group D). In addition, two samples from Group A (S1 and S3), which did not meet the pharmacopeial standard for THSG content (<1%), were included in the cytotoxicity assays to provide a more comprehensive understanding of the relationship between compositional deviation and toxicity. The selection of samples across groups considered both geographical origin and morphological traits, covering five provinces (Yunnan, Anhui, Guangdong, Guizhou, and Sichuan) and three chromaticity categories (white, yellow, and brown), thereby ensuring representativeness, typicality, and randomness.

Cell viability was assessed using the CCK-8 assay (Wang T. K. et al., 2023). Cells in the logarithmic growth phase were seeded into 96-well plates at a density of 8 × 10³ cells/well and incubated for 24 h. To minimize edge effects, the outer wells of each plate were filled with PBS. Cells were then exposed to serum-free medium containing different concentrations of PM extracts (1,000, 100, 10, and 1 μg/mL) for 72 h. Following exposure, 100 µL of CCK8 solution in serum-free medium was added to each well, and cells were incubated at 37 °C for 2 h. Absorbance was measured at 450 nm using a Biotek Epoch microplate reader (Agilent, United States). Each treatment group included four replicates. Cell viability was calculated as (dosing − blank)/(control − blank). APAP was used as a positive control for hepatotoxicity. It was dissolved in DMSO and diluted in WME to final concentrations of 1–20 mM, with the DMSO concentration maintained below 0.1% (v/v) (Yokoyama et al., 2018; Van den Eede et al., 2015). Cells were treated with APAP for 72 h under the same conditions as PM samples. The half-maximal inhibitory concentration (IC₅₀) values of the 29 p.m. batches were determined, and significant differences were analyzed using one-way ANOVA with Duncan’s post hoc test. Statistical significance was set at p < 0.05.

Previous studies have shown that the color characteristics of herbal medicines provide insight into their internal composition (Zhang et al., 2023; Liu et al., 2018). On this basis, 15 batches of PM samples (n = 5 per group) that passed the content test were selected from the three groups to study the correlation between the chromaticity and toxicity of decoction pieces. A stereoscopic fluorescence microscope was used to capture images of four anatomical regions (the cork, phloem, central xylem, and heteromorphic vascular bundles) from both the front and back surfaces of PM decoction pieces. Compared to colorimeters and spectrophotometers, this microscope provides enhanced magnification of morphological characteristics. To ensure reproducibility and account for sample heterogeneity, five different sites were measured for each anatomical region. The images were imported into Photoshop CS6, where chromaticity indices (R, G, B, L*, a*, and b*) were extracted. For each anatomical region, the mean chromaticity value was calculated from the five measured sites, resulting in four averaged chromaticity values for each surface. Subsequently, mean values from all 40 measurements per sample (4 regions × 5 sites × 2 surfaces) were used to calculate the overall color parameter according to the formula E*ab = [L*2+a*2 + b*2]^1/2. HCA was then performed using OriginPro 2021 to explore the correlations between the chemical-composition-based toxicity classifications of PM and the chromaticity values of different anatomical regions of the decoction pieces. To verify the above correlation, the RGB values were normalized using the following equations: Normalized R = r/(r + g + b), Normalized G = g/(r + g + b), Normalized B = b/(r + g + b). The normalized values were subsequently subjected to regression analysis against the IC₅₀ concentrations obtained from the hepatocellular cytotoxicity assays. This approach follows the normalization strategy described by Guo et al. (2020), who applied RGB values to evaluate the vegetation index of maize leaves and demonstrated a strong linear relationship between them.

The preceding analyses demonstrated that hepatocytotoxicity data were correlated with the quantitative profiles of PM components. To further investigate this relationship, an OPLS-DA model was constructed using SIMCA-P 14.1 software to maximize intergroup separation and identify the key compounds responsible for these differences (Hou et al., 2024; Zhou et al., 2024). The validity of the chemometric model was evaluated using the R ² and Q ² parameters, with values greater than 0.5 indicating good model quality, and values approaching 1 reflecting excellent model performance. Compounds were ranked by their VIP values, and those with VIP ≥1 were considered to contribute significantly to group classification. To minimize overfitting, the OPLS-DA models were rigorously validated through a 200× permutation test (Uttl et al., 2019). In this test, the order of the Y-variable (sample class) was randomly permuted, and a new model was generated for each permutation. A model was considered valid if the R² and Q² values of the permuted models were significantly lower than those of the original model, and if the regression line of Q² against the correlation coefficient between the original and permuted Y-variables displayed a negative intercept. For further chemometric verification, the quantitative data from different PM groups were exported into Excel for one-sample t-tests. To mitigate the risk of false positives in marker selection, p-values from the t-tests were adjusted using the False Discovery Rate (FDR) correction according to the Benjamini–Hochberg method. Compounds with VIP >1 and an FDR-adjusted q-values <0.05 were considered statistically significant differential constituents.

Stock solutions of potentially toxic compounds were prepared in DMSO, and hepatocellular cytotoxicity assays were conducted at physiologically relevant concentrations (30–700 μg/mL for THSG and 1–80 μg/mL for E.G.,), as described in Section 2.5. These ranges were selected based on plasma and tissue concentrations reported in previous in vivo pharmacokinetic studies following PM extract administration, particularly those associated with hepatotoxic outcomes (Yu et al., 2017, Li D. et al., 2022). The “One-belt, One-line” method, a novel drug combination assessment approach analogous to the Chou–Talalay method, was applied to evaluate compound interactions. In this approach, the expected additive effect of a multi-drug combination is displayed as a curved band on a two-dimensional coordinate plot, while the actual dose–response relationship of the tested combination is represented as a curve within the same space. The resulting alignment is referred to as the “One-belt, One-line” phenomenon (Yuan and Chen, 2019). Using Prism 5.02 and OriginPro 2021, dose–response curves for individual compounds and their combinations were generated, and CI values were calculated to evaluate interactions between potential toxic compounds. A CI value of 1 indicates an additive effect, a CI value <1 suggests synergy, and a CI value >1 indicates antagonism. Although initially developed for efficacy studies, the “One-belt, One-line” model can be extended to toxicological studies, as it is based on the same isobolographic principles as the Chou–Talalay method, which has been widely used in both efficacy and toxicity studies (Sriramulu et al., 2024; Wu et al., 2024).

Based on the literature review by Li et al. (2022a), 26 compounds were selected for quantitative analysis of PM samples. These compounds were required to be present in PM and have been reported in literature as being associated with toxic effects. Under identical quality control conditions, 12 anthraquinones, 5 stilbene glycosides, 4 flavonoids, and 5 phenols in test solutions from 61 p.m. batches were analyzed (Figure 2; Supplementary Tables S3–S5). The total concentrations of anthraquinones in PM samples ranged from 0.22 to 7.92 mg/g. Among them, emodin, physcion, physcion-8-glucoside, trans-emodin dianthrones, EG, and emodin-8-O-(6′-methylmalonyl)-glucopyranoside were the main anthraquinones, accounting for 94.27% of the total anthraquinone content (Supplementary Table S3). In contrast, rhein, emodin-6-glucoside, 1-methyl emodin, emodin-1-glucoside, cis-emodin dianthrones, and physcion-8-O-(6′-methylmalonyl)-glucopyranoside were the minor anthraquinones (Supplementary Table S3). Stilbene glycosides represented the most abundant compound class, with concentrations ranging from 4.35 to 59.43 mg/g. THSG was the dominant constituent, accounting for 96.24% of the total stilbene glycoside content (Supplementary Table S4). Meanwhile, 2,3,5,4′-tetrahydroxystilbene-2-O-(2″-O-feruloyl)-β-D-glucopyranoside, cis-2,3,5,4′-tetrahydroxystilbene-2-O-β-D-glucoyranoside, resveratrol, and polydatin were minor constituents, with average contributions of 1.74%, 0.50%, 0.26%, and 1.25%, respectively (Supplementary Table S4). Among the flavonoid and phenolic compounds, catechin, epicatechin, and torachrysone-8-O-glucoside exhibited average concentrations above 0.1 mg/g (Supplementary Table S5). A graphical summary of the average total contents of the four major compound classes (anthraquinones, stilbene glycosides, flavonoids and phenols) across all 61 batches is presented in Figure 3. This visualization clearly demonstrates that stilbene glycosides are the most abundant compound class in PM, followed by anthraquinones, consistent with the quantitative results described above.

According to the 2020 edition of the Chinese Pharmacopeia, the THSG content in PM should not be less than 1.0% (Committee, 2020). Based on this criterion, Group A samples (S1–S5) were classified as unqualified and excluded from subsequent hepatocellular cytotoxicity assays.

To explore similarities among PM samples from different batches, an HCA model was constructed using all 61 samples. The samples clustered into four distinct categories (Figure 4): Group A (S1–S5. unqualified THSG content), Group B (S11–S19, S22–S26, and S28–S30), Group C (S27 and S31–S50), and Group D (S6–S10, S20–S21, and S51–S61). There was no strong correlation between the chemical-composition-based classifications of PM samples and their geographic origins, suggesting that variations in morphological traits of PM are greater than those associated with geographical origin. One-way ANOVA further confirmed that the total content of the four major compound classes (anthraquinones, stilbene glycosides, flavonoids and phenols) in PM samples showed no significant difference compared to geographical origins. For anthraquinones (Figure 5A), pairwise comparisons between geographical areas showed no significant differences among the 10 pairs: AH vs. YN (p = 0.9893), GD vs. YN (p = 0.7472), GD vs. AH (p = 0.4768), GZ vs. YN (p = 0.5021), GZ vs. AH (p = 0.8462), SC vs. YN (p = 0.5419), GZ vs. GD (p = 0.0393), SC vs. AH (p = 0.8357), SC vs. GD (p = 0.069), SC vs. GZ (p = 0.9997). For stilbene glycosides (Figure 5B), there was a significant difference between GZ and YN (p < 0.01), GZ and AH (p < 0.01), GZ and GD (p < 0.01), SC and GZ (p < 0.01), while the other 6 pairs showed no significant differences: AH vs. YN (p = 0.9989), GD vs. YN (p = 0.8838), GD vs. AH (p = 0.9649), SC vs. YN (p = 0.9937), SC vs. AH (p = 0.9999), SC vs. GD (p = 0.9865). For flavonoids (Figure 5C), there was no significant difference between three pairs of regions, including AH vs. YN (p = 0.0163), GD vs. YN (p = 0.8619), GD vs. AH (p = 0.1911), SC vs. AH (p = 0.2933). Similarly, for phenols (Figure 5D), except for significant difference between GZ vs. YN (p = 0.0041), there were no significant differences in the other 9 pairs: AH vs. YN (p = 0.7092), GD vs. YN (p = 0.8511), GD vs. AH (p = 0.9989), GZ vs. AH (p = 0.2064), GZ vs. GD (p = 0.1127), SC vs. YN (p = 0.2085), SC vs. AH (p = 0.9061), SC vs. GD (p = 0.789), SC vs. GZ (p = 0.7721). Overall, there was no significant difference between the geographical origins and compound content of the vast majority of samples, further indicating that the chemical grouping of PM was not related to its geographical origin.

Heatmaps were then generated to visualize compositional differences among the four sample groups. This approach highlighted several major quantitative composition patterns and facilitated comparisons of differential compounds among groups (Wang Q. et al., 2023; Yin et al., 2024; Ziółkowska et al., 2016).

To validate the reliability of the classification, an unsupervised pattern recognition approach was applied (Luo et al., 2024). The PCA score plot (Figure 6) revealed that the first (PC1) and second (PC2) principal components accounted for 83.50% and 8.83% of the total variance, with a cumulative explanatory power R²X (cum) of 92.40%. Between Groups B and D was observed, suggesting shared chemical characteristics, such as comparable levels of flavonoids and phenols (Supplementary Table S5). However, significant differences in THSG and EG contents were detected between these groups. Combined with the trait identification and HCA-based grouping (Figure 4), these findings confirmed that Groups B and D were distinct. Therefore, the observed overlap did not undermine the classification but rather highlighted the complexity of the relationships among morphological traits, chemical composition, and hepatotoxicity in PM. Overall, the PCA score plot confirmed the classification trend based on compound content across the 61 samples, consistent with the heatmap analysis results.

To investigate how the chemical composition of PM influences its cytotoxicity to HepaRG cells and to evaluate potential differences in toxicity among different groups of PM samples, the CCK-8 assay was employed to measure the viability of HepaRG cells across the four sample groups. IC₅₀ values were determined after exposing HepaRG cells to 1–1000 μg/mL PM for 72 h (Supplementary Table S6). The positive control, APAP, showed an IC₅₀ of 8.2 ± 0.9 mM, consistent with previous reports on APAP-induced toxicity in HepaRG cells (Van den Eede et al., 2015; Yokoyama et al., 2018; Deng et al., 2024). These results confirmed the sensitivity of the assay to hepatotoxicants and supported the observed differences in toxicity across PM groups. The calculated IC₅₀ values aligned with the previously established PM groupings (Figure 7A). Group B samples exhibited IC₅₀ values greater than 450 μg/mL, representing the lowest toxicity; Group C samples demonstrated IC₅₀ values below 390 μg/mL, indicating the highest toxicity; and Group D samples had intermediate values between 390 and 450 μg/mL, reflecting moderate toxicity (Luo et al., 2024; Li et al., 2024). Significant differences in toxicity were observed among Groups B, C, and D (Figure 7B). Unexpectedly, Group A samples, despite their non-compliant THSG content, exhibited low cytotoxicity, with IC₅₀ values around 540 μg/mL This finding suggests THSG and related compounds as predictive markers of hepatotoxicity.

To validate the representativeness of the 31 tested batches, THSG content, monitored as a key index in the Chinese Pharmacopeia (Committee, 2020), was compared between tested and untested batches. No significant differences were observed within the same group (p > 0.05, t-test). For example, the average THSG content in tested Group B batches was 12.54 mg/g, compared to 11.98 mg/g in untested batches (p = 0.32). Overall, these results demonstrated that the cytotoxic effects on HepaRG cells were closely related to the chemical composition of PM samples, providing a basis for screening potential hepatotoxic constituents.

To ensure the accuracy and representativeness of the stereo fluorescence microscopy images, specific samples were selected: S11, S14, S18, S22, and S28 from Group B; S27, S31, S38, S41, and S46 from Group C; and S8, S9, S52, S58, and S60 from Group D. Images of the cork site (16x magnification), phloem (16x), heteromorphic vascular bundles (12x), and central xylem (12x) were collected under standardized conditions, with an identical imaging angle and light intensity (Light = 68) (Figure 8). Each sample was measured at four anatomical regions, with five sites per region on two sides, resulting in 40 chromaticity measurements per sample for mean calculation (Supplementary Table S7). Stereo fluorescence microscopy was employed to enable high-resolution, localized measurements of heterogeneous tissue structures, which bulk colorimetry could not resolve. All imaging parameters (light intensity, magnification, and angle) were rigorously standardized to ensure reproducibility. To evaluate whether chromaticity values could discriminate among PM groups, HCA was performed based on these chromaticity measurements. Clustering based on chromaticity values effectively distinguished the three groups (Figure 9A). RGB analysis revealed that Group B samples exhibited high R, G, and B values, appearing white. Group C and Group D samples showed elevated R and G values, appearing yellow. Combined with Lab values (L* for luminance, b* for yellow–blue, and a* for red–green), the color traits were identified as follows: Group B (low toxicity) appeared white, Group C (high toxicity) appeared light yellow, and Group D (medium toxicity) appeared yellow–brown. These findings indicate that chromaticity profiles, when analyzed by HCA, are associated with hepatocytotoxicity and may serve as predictors of PM toxicity. We used regression analysis to verify the correlation between chromaticity values and hepatotoxicity, Figure 9B. The results demonstrated a significant positive correlation between normalized B values and IC₅₀ (p < 0.0001), and a significant negative correlation between normalized R values and IC₅₀ (p < 0.0001). These findings support the reliability of using chromaticity values to predict the hepatotoxicity of PM.

It is important to note that factors such as storage conditions, drying methods, and geographic origin may influence the color stability of PM decoction pieces (Zhang et al., 2023). To minimize such variability, all samples in this study were stored under consistent conditions (room temperature, dry, and dark) prior to analysis. Furthermore, because drying methods (such as sun drying vs oven drying) and geographical origins may vary among commercially available PM, 61 batches were randomly selected to represent market diversity, thereby enhancing the representativeness and robustness of the experiment.

Survival experiments using HepaRG cells revealed significant cytotoxicity differences among the three groups of commercial PM samples, with toxicity ranked as Group B < Group D < Group C. To minimize the influence of irrelevant factors (such as intragroup variance) and to enhance intergroup discrimination, OPLS-DA analysis was employed to identify differential toxic compounds (Yin et al., 2024; Xue et al., 2022). The first OPLS-DA analysis compared the least toxic Group B and the most toxic Group C (Figure 10). The results showed R²X = 92.40%, R²Y = 91.10%, and Q² = 88.10%, indicating that the model had strong discriminative and predictive performance for distinguishing Group B from Group C. Cross-validation results from the 200× permutation test confirmed that the model was not overfitted. A second OPLS-DA analysis compared Group B with the moderately toxic Group D (Figure 11). The model achieved R²X = 87.30%, R²Y = 89.50%, and Q² = 85.30%, suggesting that the expected differentiation effect was achieved, and the permutation test again confirmed that the model did not overfit. The results indicated that between Groups B and C, THSG (VIP = 4.7704) and EG (VIP = 1.0427) were potential cytotoxic markers. Between Groups B and D, THSG (VIP = 3.8540), EG (VIP = 1.8470), emodin (VIP = 1.7722), and catechin (VIP = 1.3553) were identified as differential compounds distinguishing weakly toxic and moderately toxic samples (Table 1).

To further validate the differential markers, a t-test model was established to identify compounds primarily responsible for cytotoxicity differences. When Group B was compared with Groups C and D, compounds with p-values <0.05 were considered as differential markers of toxicity. These results, together with VIP values >1, allowed the selection of appropriate markers (Table 1). Based on the t-test and VIP results, two common toxicity-related compounds (Compounds 5 and 13) were identified. THSG, widely distributed in PM samples (VIP = 4.11, p < 0.01), has been reported to induce hepatotoxicity under certain concentrations or ratios (Qiu et al., 2013; Xue et al., 2020). Similarly, EG (VIP = 1.51, p < 0.01), the main active ingredient of PM, has been shown to exert hepatotoxic effects (Zhang et al., 2020). The significance of THSG and EG as primary hepatotoxic candidates was further confirmed as they remained statistically significant (q < 0.05) after FDR correction for multiple testing, underscoring the robustness of these findings. In addition, previous studies have shown that when stilbene glycosides are combined with anthraquinones, they induce greater metabolic pathway disruption compared to anthraquinones alone, suggesting synergistic toxic interactions between these compound classes (Zhang et al., 2020; Lin et al., 2015).

The structures of THSG and E.G., are shown in Figure 12A. Their cytotoxic effects on HepaRG hepatocytes were evaluated individually at concentrations of 30–700 μg/mL for THSG and 1–80 μg/mL for EG. THSG was non-toxic at concentrations up to 350 μg/mL; however, at 400–500 μg/mL, only 50% of the cells survived (Figure 12B). In contrast, E.G., exhibited stronger toxicity, with 50% inhibition observed at 10 μg/mL, and complete cell death was visually observed at 80 μg/mL. Mass spectrometry confirmed the presence of both compounds in all PM samples, although their relative levels varied across Groups B, C, and D, corresponding to different degrees of hepatocytotoxicity. These findings suggest that the ratio of THSG to E.G., is an important determinant of toxicity.

To further investigate the antiproliferative effects of THSG and EG, 10 concentrations of THSG (30, 120, 180, 250, 300, 350, 400, 450, 500, and 550 μg/mL) were combined with 10 concentrations of E.G., (1.25, 2.5, 5, 7.5, 10, 12.5, 15, 17.5, 20, and 22.5 μg/mL). These concentration ranges not only covered the IC₅₀ values of the two compounds but also reflected physiologically achievable concentrations in vivo. As reported by Yu et al. (2017), THSG can accumulate in the liver at concentrations up to several hundred µg/mL, which is consistent with the mid-to-high concentration range tested in this study (250–550 μg/mL). After 72 h of incubation, the IC₅₀ values for THSG and EG were 417.0 μg/mL and 12.7 μg/mL, respectively. When the two compounds were co-incubated, the IC₅₀ decreased to 247.7 + 7.5 μg/mL (Figure 12C), indicating that the toxicity of EG was significantly enhanced when THSG was co-administered at non-toxic concentrations (250–350 μg/mL). These findings demonstrated that THSG and EG exerted greater cytotoxic effects when combined. The hepatotoxic interactions between THSG and EG (additive, synergistic, or antagonistic) were evaluated using the “One-belt, One-line” model (Yuan and Chen, 2019). In this approach, the belt represents the observed dose–effect relationship of the two-drug combination (Figure 12D), while the red curve represents the expected additive effect. When the observed efficacy falls within the expected additive range, the interaction is classified as additive; when it exceeds the range, it is synergistic; and when it is lower than expected, it is antagonistic. Our results showed that combinations of 30–180 μg/mL THSG with 1.25–5 μg/mL EG exhibited antagonistic effects. In contrast, combinations of 250–350 μg/mL THSG with 10–15 μg/mL EG showed additive effects. Notably, combinations of 400–550 μg/mL THSG with 15–22.5 μg/mL EG produced synergistic effects (Figure 12D; Table 2). The CI was used to further quantify the additive, synergistic, and antagonistic interactions between THSG and EG (Table 2). Both the “One-belt, One-line” model and CI values consistently indicated that increasing the concentrations of the two compounds enhanced hepatotoxic effects, thereby narrowing the drug safety window and increasing the risk of liver injury. The consistency of results between the approaches confirmed the classification of additional, synergistic, and antagonistic interactions between THSG and EG. Quantitative analysis revealed that THSG and EG were widely distributed in PM samples (4.90–57.24 mg/g and 0.10–4.73 mg/g, respectively). Notably, Group C samples, which exhibited the strongest cytotoxicity, contained higher levels of THSG (32.58–57.24 mg/g) and E.G., (1.45–4.73 mg/g). These elevated concentrations were positively correlated with toxicity, consistent with the interaction patterns observed in the “One-belt, One-line” model. Similar trends were observed in Group D, which displayed moderate toxicity.

Overall, these findings demonstrate that the hepatotoxicity of PM is not attributable to a single compound but arises from interactions among multiple constituents. The dose-dependent synergistic toxicity observed between THSG and E.G., highlights the risks associated with high-dose consumption or specific compound combinations in PM products.