The bile was filtered through gauze and then prefrozen at −80°C for 12 h. All samples were lyophilized in a vacuum freeze-dryer (EYELA, TOKYO RIKAKIKAI CO., LT D, FDU-2110) for 48 h. Approximately 5 mg of bile powder was weighed and extracted by ultrasonication for 45 min with a mixed solvent (water containing 1% formic acid mixed with methanol in a volume ratio of 5:95). The solution was fixed in a 10 mL volumetric flask, mixed well, and filtered through a microporous membrane (0.22 μM) before use (Li et al., 2022). All bile samples were mixed to obtain a quality control sample (QC).

The separation was performed on an ultrahigh-performance liquid chromatography (UPLC) system (I-Class, Waters Corp., Manchester, United Kingdom) combined with a quadrupole time-of-flight mass spectrometer (Q-TOF/MS, Xevo G2-S, Waters Corp., Manchester, United Kingdom). The mobile phase consisted of a 0.1% formic acid aqueous solution (A) and a 0.1% formic acid acetonitrile solution (B). The gradient elution procedure used was as follows: 0–0.5 min, 5% B; 0.5–1 min, 5% B→20% B; 1–3 min, 20% B→40% B; 3–12 min, 40% B→60% B; 12–16 min, 60% B→90% B; 16–17 min, 90% B→100% B; 17–18 min, 100% B; 18–19 min, 100% B→5% B; and 19–20 min, 5% B. The flow rate was 0.25 mL/min, and the injection volume was 5 μL. The column and autosampler were maintained at 40°C. An electrospray ion source (ESI) was used to perform a full scan in the negative ion mode. Data collection ranged from 50 to 1000 Da. The operating parameters of the ESI source were as follows: capillary voltage, 2.0 kV; cone voltage, 30 V; cone gas flow rate, 50 L/h; ion source temperature, 100°C; and dissolvent gas temperature, 450°C.

1) Instrumental precision: The same QC sample was injected for 6 times, 10 peaks were randomly selected and the RSD values of peak area and retention time were calculated for each of the 10 peaks.

To compare and distinguish the bile of each species, UPLC-Q-TOF/MS was used for sample analysis in this study. First, raw data were exported to MassLynx version 4.1 software (Waters Corp., Milford, MA), and the data provided the m/z value and retention time as well as the detected peak with a certain intensity. Then, the data were normalized and imported into SIMCA-P14.1 (Umetrics AB, Sweden) for multivariate data analysis. The aggregation degree of bile can be determined from the score plot of principal component analysis (PCA). To clarify the differences between them, chemical markers of the bile from three animals were screened. Then, an OPLS-DA model of the samples was performed, and the variable importance projection (VIP) of each chemical component was calculated. A variable whose VIP was greater than 1.0 was usually considered to be a potential marker (Li et al., 2020). Therefore, the components with VIP>1 were screened and subjected to statistical analysis by SPSS 20.0, and chemical markers with statistically significant differences were obtained (p < 0.05).

Before the luciferase experiments, cell viability under different concentrations of key markers was first investigated. 293T cells at the logarithmic growth stage were inoculated in 96-well cell culture plates at a density of 2 × 10⁵ cells/mL. Then, 100 μL of 0.01, 0.1, 1, 10, and 100 μM CA, GCA, GCDCA, and TDCA were added to each well in 4 parallels. After 24 h of treatment, the supernatant was aspirated, and CCK-8 solution was added to each well and then incubated for 30 min at 37°C. The absorbance OD value was measured at 450 nm by an enzyme marker.

The NF-κB luciferase reporter plasmid (pGL4.32) and PGMLR-TK Renilla Fluorescein Reporter Plasmid (pGL4.75) were simultaneously transfected into 293T cells with PEI (1 mg/mL) transfection reagent. Cells transferred into plasmid were cultured for 24 h, and then dexamethasone (10 μM) and the key marker with a safe concentration were separately added. Control and Model groups were set up and incubated for another 6 h. The cells were then lysed and assayed with the Dual-Luciferase assay system. Six parallel wells were set up and replicated three times. Finally, the relative luciferase activity was obtained by comparing firefly luciferase activity with Renilla reniformis luciferase activity.

First, cell viability assays were performed on RAW264.7 cells. RAW264.7 cells at the logarithmic growth stage were collected and added to 0.01, 0.1, 1, 10, 100 μM CA, GCA, GCDCA, and TDCA solutions. Four parallel wells were set for each concentration, and the control group and DMSO group without drugs were set at the same time. After 24 h of dose addition, 100 μL CCK-8 solution was added to each well and incubated at 37°C for 30 min. The absorbance OD value was measured at 450 nm with a microplate reader, and the cell viability was calculated.

RAW264.7 cells were treated with LPS at a concentration of 1 μg/mL and CA, GCA, GCDCA, and TDCA at a final concentration of 0.1 μM. After 24 h of dose addition, the medium supernatant was collected from microporous plates. Then, the solution was mixed with the same amount of Griess reagent, and the absorbance value of the solution was measured at 540 nm. The concentration of NO in the samples was calculated according to the standard curve.

According to the OPLS-DA models, we initially screened eight bile acids that could distinguish three kinds of bile. However, when establishing a machine learning model, not every feature contributes significantly to the predicted target value (Kaplan et al., 2021). Removing low contributing variables and using key feature markers to construct machine learning models can improve the generalization ability of the models. NCA is an effective means of finding key features and is commonly used by researchers for feature selection when modeling against high-dimensional data (Jiménez-Grande et al., 2021). In this study, NCA was used to select key features for chemical markers between the bile of different animals, and key feature markers were modeled as feature variables. The weight values for each feature are shown in Supplementary Figure S1. The features with higher weights were the ones that were more significant to the classification value. The weights of TDCA, GCA, CA, and GCDCA were all greater than 1, indicating that these four components contributed more to differentiating the bile of different animals. All four components have been shown to have antibacterial, anti-inflammatory, antihypertensive, anti-cough, and expectorant pharmacological effects. Previous reports showed that these four feature markers had pharmacological effects, such as bacteriostasis, anti-inflammatory, antihypertensive, antitussive, and expectorant effects (Zhou et al., 2021).

To validate the accuracy and specificity of the key feature markers, a polynomial kernel function classifier (PL-SVM) model was established using these four feature markers as feature variables. The classification performance of SVM is affected by many factors, among which the following two factors are crucial: 1) the error penalty Factor C and 2) the form of the kernel function and its parameters (Maltarollo et al., 2019). The polynomial kernel function that we chose for modeling is one of the most widely used and applicable kernel functions (Ye et al., 2013). In this study, the grid search method was used to perform a global search for optimization of the parameters C and g to achieve the best discriminative performance of the PL-SVM model. When C was 0.134 and g was 3.732, the PL-SVM model had the most accurate discriminant performance.

In addition to g, the main parameters of the PL-SVM model also include the polynomial coefficient and coef 0. The optimal g was determined by the grid search method, and then an optimization search was performed for the polynomial coefficients as well as the coef 0 value. The results showed that the PL-SVM models with optimal performance were obtained when the polynomial coefficients were 3 and 4, the models were fitted with 100% CV accuracy, and the classification accuracy was up to 100% (Supplementary Table S8). However, the larger the polynomial coefficients are, the higher the dimensionality of the mapping and the consequent increase in computational effort. When the coefficients are too large, learning is too complex and tends to lead to overfitting (Ellmann et al., 2020). The factor of 3 was chosen to take into account the size of the calculation and the amount of memory used. The model performed best when the polynomial coefficient was 3 and coef 0 was 7.4, with a CV accuracy of 100% and a classification accuracy of 100%. The results showed that the established PL-SVM model was stable and reliable and could accurately differentiate bile from different sources. In addition, the four key feature markers were highly accurate and specific.

Cytokine expression, apoptosis regulation, and cell proliferation are highly dependent on the transcription factor NF-κB, which plays a key role in the regulation of the inflammatory response (Emam et al., 2021). In this study, the viability of 293T cells was assessed after treatment with different concentrations of four key feature markers (Supplementary Figure S2). GCDCA and TDCA did not show cytotoxic effects on 293T cells at the test concentrations of 10⁻²–10² μM; GCA and CA concentrations of 10⁻⁵–10^–3 μM and 10⁻⁵–10^–1 μM, respectively, showed no cytotoxic effect on 293T cells. TNF-α was used as the modelling drug and 10 μg/mL of TNF-α complete medium solution was added to the model group; positive drug and different concentrations of bile acid solutions were prepared using 10 μg/mL of TNF-α complete medium solution and the prepared solutions of each group were added to the cells of the corresponding group. Then, the cells were treated with the maximum nontoxic concentrations of each key marker and evaluated for anti-inflammatory effects. The nuclear factor NF-κB pathway has long been recognized as a classic proinflammatory signaling pathway. This was mainly based on the role of NF-κB in the expression of proinflammatory genes, including cytokines, chemokines, and adhesion molecules (Lawrence 2009). Dual luciferase reporter gene assay systems were used to determine the effect of key feature markers on the expression levels of NF-κB in cells. As shown in Supplementary Figure S3, the positive control medicine dexamethasone (10 μM) significantly inhibited TNF-α-induced NF-κB production (p < 0.01), and TDCA, GCA, CA, and GCDCA also had a significant inhibitory effect on NF-κB. This finding suggested that four bile acids were potential NF-κB inhibitors, with preliminary confirmation of their anti-inflammatory activity.

In this study, the effects of four markers at concentrations ranging from 10⁻²–10² μM on the viability of RAW264.7 macrophage cells were first examined by the CCK-8 assay, and the results obtained are shown in Supplementary Figure S4. Compared with the control group, none of the four feature markers showed cytotoxicity to RAW264.7 cells at a concentration of 10⁻²–10² μM. Therefore, 10⁻²–10² μM of four markers was used to evaluate the anti-inflammatory effects.

Macrophages can sense and respond to many stimuli and trigger inflammation by secreting NO. Excessive production of NO is associated with restriction of the immune response and remission of inflammation, and excessive NO production by macrophages has been observed in many inflammatory diseases (Lee et al., 2021). According to the results observed (Figures 4A–D), nontoxic doses of TDCA, GCA, and GCDCA significantly inhibited NO release compared to the LPS group (p < 0.01), as did 10⁻¹ μM CA (p < 0.05). Four key feature markers could exert their anti-inflammatory activity in LPS-stimulated RAW264.7 cells by inhibiting NO production.