Botanical drugs and dosage details in the JBD study were sourced from the Chinese Pharmacopeia 2020 (National Pharmacopoeia Commission, 2020). Taxonomic validation of all botanical drugs was conducted using the Plants of the World Online (https://powo.science.kew.org/) and the Chinese Medicinal Material Images Database (https://library.hkbu.edu.hk/electronic/libdbs/mmd/) (Table 1). Information about the chemical components within these botanical drugs was gathered from various publicly accessible Traditional Chinese Medicines (TCMs) databases, such as the Traditional Chinese Medicine Information Database (TCMID, http://bidd.group/TCMID/) (Chen et al., 2006), and the Traditional Chinese Medicine and Active Ingredient Database (TCMAID, http://www.organchem.csdb.cn/scdb/main/tcm_introduce.asp). The chemical structures were prepared and converted into canonical SMILES using the OpenBabel Toolkit (O Boyle et al., 2008).

The absorption, distribution, metabolism, excretion, and toxicity (ADMET) properties and physicochemical characteristics of all compounds were predicted using the ADMETlab 2.0 web server (https://admetmesh.scbdd.com/) (Xiong et al., 2021). Under Lipinski’s rule of five (Pollastri, 2010), potential active compounds were required to fulfill the following criteria: (1) molecular weight (MW) ≤ 500; (2) number of hydrogen bond donors (nHD) ≤ 5; (3) number of hydrogen bond acceptors (nHA) ≤ 10; (4) logarithm of the n-octanol/water distribution coefficient (logP) ≤ 5; and (5) the number of rotatable bonds (nRot) ≤ 10. Furthermore, potential active compounds needed to meet the following additional conditions: (6) human oral bioavailability (OB) ≥ 30%, (7) Caco-2 permeability (Caco-2) ≥ 0.4, (8) topological polar surface area (TPSA) ≤ 60, and (9) no hepatotoxicity.

The potential targets of the potential active compounds were predicted using the following web servers: the similarity ensemble approach (SEA) search server (https://sea.bkslab.org/) (Keiser et al., 2007), HitPickV2 (http://www.hitpickv2.com/) (Hamad et al., 2019), and SwissTargetPrediction (http://www.swisstargetprediction.ch/index.php) (Daina et al., 2019). The union set of the prediction results was defined as the potential targets.

The RA pathological genes were searched and downloaded from the DisGeNet database (https://www.disgenet.org/) (Pinero et al., 2020). Genes with literature reports were defined as RA pathological genes. The intersection of targets and pathogenic genes was defined as essential common proteins (ECPs).

For the network construction methods, see our previous publication (Xu et al., 2023a). Briefly, we retrieved the interactions between targets and pathogenic genes and extracted documented protein-protein interactions from the BioGrid database (https://thebiogrid.org/) (Oughtred et al., 2019). We combined this with previous compound-target prediction results, and the compound-target-pathological gene (C-T-P) network was constructed and visualized using Cytoscape software (Kohl et al., 2011).

Transcriptome data from the macrophages of RA patients were downloaded from the National Center for Biotechnology-Gene Expression Omnibus (NCBI-GEO) database (https://www.ncbi.nlm.nih.gov/geo/) with accession identification numbers of GSE10500 and GSE97779. Data preprocessing, integration, and global normalization were conducted using our previously described method (Li et al., 2016a; Li et al., 2016b). The differential expression analysis was conducted using the empirical Bayesian algorithm in the limma package (Ritchie et al., 2015) in R statistical software (https://www.r-project.org/). Significantly differentially expressed genes were defined as absolute values of the log2-transformed fold change (logFC) > 1 and an FDR p-value <0.05.

The C-T-P network was optimized using the MOO model. The purpose of optimization was to extract target genes that were strongly associated with compounds and pathogenic genes and were dysregulated in RA patients. Details of the MOO model and network optimization can be found in a previous report (Xu et al., 2023b).

The GO terms of biological process, cellular component, molecular function, and human gene information were downloaded from the QuickGO database (https://www.ebi.ac.uk/QuickGO/) (Binns et al., 2009). The reference human genes and pathways were obtained from the Kyoto Encyclopedia of Genes and Genomes (KEGG) database (http://www.kegg.jp/) (Kanehisa and Goto, 2000). The GO terms and KEGG pathways with fewer than 10 genes were removed. The enrichment analysis was performed using the hypergeometric test. An false discovery rate (FDR)-corrected p-value ≤0.05 was considered significantly enriched.

The degree model, closeness model, and betweenness model (Yang et al., 2021) were used to optimize the constructed C-T-P network and compared with the MOO model. For details see the previous publication (Xu et al., 2023a). The optimization performance of the different models was compared from the five aspects: (1) the coverage of essential common proteins; (2) the coverage of the top 100 enriched KEGG pathways; (3) the coverage of the top 1,000 enriched GO biological processes (BP); (4) the average regulating intensity; and (5) the cumulative differential expression. Genes in the optimized network in the different models were used to perform the KEGG pathway and GOBP enrichment analyses, and the reference KEGG pathways and Gene Ontology biological processes (GOBPs) were enriched using essential common proteins (ECPs). The average regulating intensity was calculated as the mean Pearson’s correlation coefficient for the gene-gene pairings in the optimized network. The cumulative differential expression was defined as the sum of the absolute values of the logFC of ECPs in the optimized network.

We extracted the pathogenic genes correlated with each compound in the optimized network and then calculated the interaction intensity flow of each pathogenic gene. Let be the interaction intensity flow of the pathogenic gene; is the set of target genes that interact with the pathogenic gene; and is the number of target genes in. The following equation was used for the calculation: I n t e n s i t y k = S M O O k ′ , k ∈ T , k ∈ P   ∑ S M O O j ′ n k , k ∉ T , k ∈ P , j ∈ T k

Where is the interaction intensity of compound, on pathogenic genes; and is the set of pathogenic genes that correlate with compound. The equation is: I n t e n s i t y i = ∑ k = 1 C k I n t e n s i t y k , k ∈ C k

The above formula was used to calculate the interaction intensity flow of a single compound on the pathogenic genes. For the interaction intensity of multiple compounds on the pathogenic genes, we first constructed a list of combinations of the different compounds and then calculated the interaction intensity of each combination using the above equation. The combination containing the same number of compounds with the strongest interaction intensity was defined as the optimal combination. Finally, we sorted these optimal combinations in ascending order of the interaction intensities and selected the least compound combination whose cumulative interaction intensity reached 90% of the interaction intensity of all the compounds in the optimized network.

The cumulative interaction intensity flow of the pathway is defined as the sum of the interaction intensity flow of pathogenic genes in the pathway. Let be the cumulative interaction intensity flow of pathway; is the number of pathogenic genes in pathway; as the number of pathways that a pathogenic gene is involved in. The calculation equation is as follows: I n t e n s i t y p = ∑ k = 1 m I n t e n s i t y k n k p

All pathways are sorted according to the, and the pathways with larger values are considered to be the primary pathways affected by the JBD.

The RAW264.7 cell line was shared by the South China University of Technology. They were cultured in the RAW264.7 cell medium, containing Dulbecco’s Modified Eagle Medium (DMEM), 10% fetal bovine serum (FBS), and a 1% penicillin-streptomycin (P/S) solution at 37°C and 5% CO2 (Procell, Wuhan, China). A total of 1 μg/ml of lipopolysaccharide (LPS) (Biosharp, Hefei, China) was administered to the RAW264.7 cells for 24 h to construct an inflammatory cell model.

The RAW264.7 cells were cultured in 96-well plates and treated with 100, 50, 25, 12.5, 6.25, 3.125, and 1.56 μM compounds for 24 h. A total of 10 μL of the CCK8 reagent (Jiancheng Bioengineering Institute, Nanjing, China) was added to the RAW264.7 cells at 37°C for 1.5 h. The absorbance was analyzed at 450 nm using a microplate reader (Bio-Rad, Hercules, CA, USA) with wells without cell serving as blanks. The cell proliferation was expressed by the absorbance.

The RAW264.7 cells were cultured in 48 well plates and incubated for 24 h to measure the anti-inflammatory activity by nitric oxide (NO) assay. The cells were treated with 25, 10, 5, 1, and 0 μM compounds at 37°C for 24 h. Subsequently, LPS (0 or 1 μg/mL) was administered into the cells and then incubated for 24 h. NO production in the supernatant was measured using the nitrite measurement according to the manufacturer’s protocol. In brief, an equal volume of supernatant and Griess reagent were mixed and incubated for 10 min at ambient temperature in the dark, and the absorbance was measured at 570 nm using a Synergy HT microplate reader from Bio-Tek Instruments Inc. (Winooski, VT, USA).

This study was conducted in strict accordance with the recommendations of the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. All methods were conducted in accordance with the Ethics Committee Institute of the School of Medicine of Foshan University. All experimental protocols were approved by Foshan University. All methods are reported in accordance with the Animal Research: Reporting of In Vivo Experiments (ARRIVE) guidelines. All surgeries were performed under chloroform anesthesia, and all efforts were made to minimize suffering.

The exact number of experimental units allocated to each group was six mice. The total number in each experiment was 30 mice. The total number of animals used was 90 mice. DBA/1J mice (males, 8 weeks old, free of murine-specific pathogens) were obtained from the Experimental Animal Center, Xiangya School of Medicine, Central South University. Random numbers were generated using the standard = RAND () function in Microsoft Excel. Randomization was conducted as follows. Animals were assigned a group designation and weighed. A total number of 30 animals were divided into five different weight groups (six animals per group). Each animal was assigned a temporary random number within the weight range group. Cages were given a numerical designation based on their position on the rack. For each group, a cage was selected randomly from the pool of all cages. One or two animals were removed from each weight range group and given their permanent numerical designation in the cages.

Mice were kept under illumination conditions, with the room temperature maintained at 22°C ± 4°C, and they were granted free access to food and water. The mice were adaptively fed for 7 days, and then procedures were conducted in the animal room of Foshan University. Each mouse was injected with 100 μL bovine type II collagen (2 mg/mL in 0.05 M acetic acid; Chondrex, Inc., Redmond, WA, USA) emulsified in complete Freund’s adjuvant (Sigma-Aldrich, St. Louis, MO, USA) on the day 7. On day 28, mice were administered 100 μL bovine type II collagen emulsified in incomplete Freund’s adjuvant by enhanced intradermal injection. Beginning on day 28, mice were administered methotrexate (MTX) (1 mg/kg) or JBD (20 mg/kg or 40 mg/kg) every 3 days. Complete adjuvants produce antibodies in mice, including IgG2a antibodies, which play an important role in the activation of the complement and subsequent development of arthritis, depending on the amount of tuberculosis bacteria in the complete adjuvant. Type II collagen and full adjuvant emulsification significantly induced widespread joint inflammation in CIA-susceptible DBA/1 mice. Mice were randomly divided into five groups, including control, CIA, CIA + MTX, CIA + BAC-L, and CIA + BAC-H. The mice were anesthetized with isoflurane and euthanized by means of cervical dislocation on day 40. Blood, knee, elbow, wrist, and ankle joints were collected from the mice. The following parameters were assessed: bone surface/volume ratio, H&E staining, safranin O-fast green staining, Western blot, and RT-qPCR.

The arthritis index (AI) was graded as following criteria: 0, normal joints; 1, slight swelling or a digital red spot; 2, red skin and slight swelling in the ankles and feet; 3, moderate swelling and erythema; 4, severe swelling and erythema involving the entire posterior claw or precipitation (Kollias et al., 2011).

A Sample of mice knee joints was selected for micro-computed tomography (CT) scanning (ZKKS-MCT-Sharp, Guangzhou, China). The voltage was 70KV, the power was 7 W, four superimposed frames were used, the angle gain was 0.72°, the exposure time was 100 ms, and the rotation was 1 week to complete the scanning for the micro-CT. A three-dimensional image screenshot and bone parameter analysis were performed after scanning and reconstruction.

The mice were anesthetized and sacrificed. The wrist, knee, ankle, and elbow joints were collected and fixed with 4% paraformaldehyde for 24 h and decalcified with 10% ethylenediaminetetraacetic acid (EDTA). After processing using conventional histological procedures, such as dehydration, transparency, wax immersion, and embedding, the paraffin-embedded tissue sections were stained by hematoxylin and eosin (H&E) staining (Sigma). The H&E staining score evaluation criteria were as follows: it could be divided into five degrees of zero to four points according to the edema, structural damage, and inflammatory cell infiltration of lesions, from light to heavy. The score was measured (Liu et al., 2022).

For safranin O/fast green staining, the paraffin-embedded slices were deparaffinized, sequentially immersed in xylene I for 20 min, xylene II for 20 min, absolute ethanol I for 5 min, absolute ethanol II for 5 min, 75% alcohol for 5 min and washed with tap water. The slices were stained with fast green staining for 5–10 min, dried to discard the excess dyeing solution until the cartilage was colorless, soaked in a differentiation solution, and washed in tap water. The slices were stained with the safranin staining solution for 15–30 s and dehydrated quickly with absolute ethanol. The slices were made transparent with xylene for 5 min, mounted with neutral gum, observed, and photographed under the microscope (Zhang et al., 2021a).

The RAW264.7 cells or synovial tissues were homogenized and lysed in radioimmunoprecipitation assay (RIPA) buffer (Sigma-Aldrich) that contained a protease inhibitor (Roche, Mannheim, Germany). A total of 50 μg protein isolated from the RAW264.7 cells or synovial tissues was separated using 10% sodium dodecyl sulfate-polyacrylamide gel electrophoresis (SDS-PAGE) and transferred to a polyvinylidene difluoride (PVDF) membrane (Millipore, Billerica, MA, USA). The membranes were incubated with the following primary antibodies: CYP17A1 (ab125022, Abcam), HSD17B3 (ab126228, Abcam), TLR4 (ab13556, Abcam), p-MAPK (AF4001, Affinity), MAPK (BF8015, Affinity), p-p65 (#11011, SAB) and p65 (#48676, SAB). Subsequently, a horseradish peroxidase (HRP)-conjugated secondary antibody was added for the 1-h incubation. The blots were developed using the ECL Western blot Kit (LabLead, E1050) on the gel imaging system (ChampgelTM 5,000 plus, SINSAGE, China). The exposure time for blots ranged from 100 ms to 3 s. Film density was measured using ImageJ densitometry software and was normalized against glyceraldehyde 3-phosphate dehydrogenase (GAPDH) (#60004-1-Ig, Proteintech) (Liu et al., 2022).

The total RNA was extracted using Trizol (Invitrogen) according to the manufacturer’s protocol. One milligram of RNA was reverse transcribed into cDNA using avian myeloblastosis virus (AMV) reverse transcriptase (Qiagen, Vento, Netherlands) with an RNase inhibitor and oligo d(T) primer at 40°C for 50 min followed by heating at 90°C for 5 min. Then, 1 μL of the reverse-transcript was added to a 30 μL PCR mixture for 40 cycles. Each cycle included 93°C for 30 s and 54°C for 60 s and used Taq polymerase. Negative controls consisted of an equal volume of water substituted for the volume of RNA in the reverse transcriptase (RT) reaction. The normalization of mRNA expression data for sample-to-sample variability in RNA input, RNA quality, and reverse transcription efficiency was achieved by a the comparison between the copy numbers of the target gene and the housekeeping gene, and GAPDH. The relative gene expression was measured using the 2^−ΔΔCT method. Primer sequences are shown in Table 2.

All experiments were run in a minimum of triplicates, and PRISM (GraphPad, La Jolla, CA, USA) statistical software was used for the statistical analysis. Statistical differences between two groups were determined using a Mann-Whitney test. The Kruskal–Wallis test determined differences among multiple groups. The effects of different concentrations of bornyl acetate on the CIA mouse model were assessed using a Mann-Whitney test. The Kruskal–Wallis test determined differences among multiple groups. All data are presented as mean values ±standard deviation (SD). p < 0.05 was considered statistically significant.

This study involved the collection of a total of 1,252 compounds derived from eight different herbs (Bai Shao, Dang Gui, Fang Feng, Gan Cao, Huang Qi, Jiang Huang, Qiang Huo, and Sheng Jiang). The pharmacological and ADMET properties of these compounds were assessed and predicted (Supplementary Table S1). Following our screening criteria, we identified and retained 274 potentially active compounds. Among these, the numbers of screened potential active compounds found in Bai Shao, Dang Gui, Fang Feng, Gan Cao, Huang Qi, Jiang Huang, Qiang Huo, and Sheng Jiang were 12, 33, 46, 27, 1, 16, 39, and 100, respectively (Figure 1; Supplementary Table S2). In addition, it is worth noting that there were 48 compounds that were common to two or more of these herbs.

Our methodology involved a series of steps. Initially, duplicate compounds were removed, after which target prediction was conducted using three different tools: SEA, HitPickV2, and SwissTargetPrediction, resulting in predictions of potential targets for the remaining 175 compounds. These predictions indicated that these compounds could potentially interact with a range of 16–178 targets (Supplementary Table S3). Furthermore, we collected a total of 2,766 distinct pathological genes associated with rheumatoid arthritis (Supplementary Table S4). These targets and pathological genes were cross-referenced with the BioGRID database to acquire protein-protein interaction data. Subsequently, we merged the compound-target network with the protein-protein interaction network to create the comprehensive C-T-P network. The original C-T-P network encompassed (Ru et al., 2014), 535 compound-target pairs and 46,838 protein-protein interactions (Figure 2A). Notably, 438 proteins were found to be both target proteins and pathogenic genes. Subsequent KEGG pathway and GOBP enrichment analyses highlighted their involvement in multiple signaling pathways linked to human diseases (Supplementary Tables S5, S6).

The original C-T-P network was refined using the MOO method. The optimized C-T-P network now comprises 10,239 compound-target pairs and 24,091 protein-protein interactions, incorporating 175 compounds, 1,085 predicted targets, and 2,081 rheumatoid arthritis pathological genes (with 438 genes shared between the target and pathological gene categories), as depicted in Figure 2B. Notably, the top 20 genes with the highest degrees in both the original and optimized networks have been labeled. Of these top genes, 13 are common to both the target genes and pathological genes (AR, APP, BRD4, CTNNB1, ESR1, ESR2, GSK3B, HDAC1, HSP90AA1, MDM2, NTRK1, PARP1, and NTRK1).

In this study, we conducted a comprehensive examination of unique genes within the optimized network generated by the MOO model, as well as networks created using the degree model, closeness model, and betweenness model. These genes were subsequently subjected to overlap assessments concerning the essential genes, the KEGG pathway enrichment, and the GOBP enrichment. To gauge the model performance, we focused on the top 100 enriched KEGG pathways and the top 1,000 enriched GOBPs.

The results revealed varying degrees of overlap in the essential genes, with 438, 272, 259, and 273 common genes identified in the MOO model, degree model, closeness model, and betweenness model, respectively. Similarly, for the top 100 enriched KEGG pathways, there were 90, 86, 85, and 83 shared pathways among these four models, respectively. Furthermore, in the case of the top 1,000 enriched GOBPs, the numbers of overlapping terms were 775, 538, 521, and 542 in the MOO model, degree model, closeness model, and betweenness model, respectively. In addition, we evaluated the average regulating intensity and cumulative differential expression values, finding that these metrics were consistently higher in the MOO model compared with the other three models (Figure 3). These findings collectively indicated that the optimization performance of the MOO model surpassed that of the other models.

The interaction intensity flow on the pathogenic genes associated with each screened compound was calculated (Supplementary Table S7). Remarkably, the cumulative impact of 15 specific compounds on these pathogenic genes accounted for 90% of all active compounds (Figure 4). The details of these compounds, which originate from Bai Shao (one compound), Dang Gui (three compounds), Fang Feng (seven compounds), Gan Cao (four compounds), Huang Qi (one compound), Qiang Huo (three compounds), and Sheng Jiang (eight compounds), are elaborated in Table 3. Notably, among these compounds, JBD00018 (stearic acid) and JBD01030 (eugenol) were FDA-approved drugs, while the remaining 13 compounds are natural ingredients derived from botanical sources. The cumulative interaction intensity flow of these screened compounds on the enriched pathways suggested that these 15 compounds primarily affect metabolic pathways, immune pathways and signaling transduction pathways (Figure 5; Supplementary Table S8). Next, we verified the effectiveness of these compounds using cell and animal experiments.

The effects of the compounds on cell viability and inflammatory response were verified using LPS-induced RAW.264.7 cells. Because some compounds cannot be purchased or have no Chemical Abstracts Service (CAS) registry number, six selected compounds were utilized in the experiment: JBD01030 (eugenol), JBD00018 (stearic acid), JBD00198 (isoamylbenzene), JBD01230 (dihydrocapsaicin), JBD00978 (p-hydroxyphenethyl anisate), and JBD00233 (bornyl acetate). The secretion of IL-6, IL-1β, and TNF-α significantly increased in the LPS-induced RAW264.7 cells (Figure 6A). We then examined the cell viability of the RAW264.7 cells after treatment with JBD01030 (eugenol), JBD00018 (stearic acid), JBD00198 (isoamylbenzene), JBD01230 (dihydrocapsaicin), JBD00978 (p-hydroxyphenethyl anisate) and JBD00233 (bornyl acetate) (0, 1.56, 3.125, 6.25, 12.5, 25, 50 and 100 μM). There was no significant change in the cell viability of RAW264.7 cells after being treated with stearic acid, dihydrocapsaicin, or p-hydroxyphenethyl anisate (Figure 6B). Administration of isoamylbenzene and bornyl acetate significantly decreased the cell viability of RAW264.7 cells (Figure 6B). NO production significantly decreased after treatment with eugenol, stearic acid, and bornyl acetate at 10 and 25 μM in the LPS-induced RAW264.7 cells (Figure 6C). The administration of 3-methybutylbenzene at 5, 10, and 25 μM significantly decreased the NO production in LPS-induced RAW264.7 cells (Figure 6C). Moreover, treatment with dihydrocapsaicin and p-hydroxyphenethyl anisate at 25 μM decreased the NO production in LPS-induced RAW264.7 cells (Figure 6C). Intriguingly, administration of p-hydroxyphenethyl anisate at 1, 5 and 10 μM increased the NO production in LPS-induced RAW264.7 cells (Figure 6C). The above results showed that isoamylbenzene and bornyl acetate (BAC) both reduced cell viability and inhibited inflammation. Studies have shown that BAC has antioxidant and anti-inflammatory properties in different types of tissues and cells (Yang et al., 2018). Isoamylbenzene has only a benzene ring and a hydrophobic chain segment, so its water solubility is poor. It is an oily liquid, and it lacks other functional groups, such as hydroxyl, sulfhydryl, sulfonic acid and carboxyl. Oxygen in the ether bond can increase the water solubility of the drug, and aminos and amides can form hydrogen bonds with the target. In addition, the presence of benzene rings alone may present a risk of cancer. Therefore, we used BAC for the animal experiment verification.

To verify the efficacy of BAC against RA, a CIA mouse model, which is widely used to clarify the mechanisms of RA and to explore the potential therapeutic targets, was established. The mice were then treated with MTX, the first-line treatment for RA, and two dosing regimens of BAC (Figure 7A).

The arthritis index (AI) of the CIA mice was higher than that of any other group. Furthermore, the AI of the MTX-treated CIA mice was lower than BAC-treated CIA mice. Importantly, the AI in the CIA + BAC-H group was lower than in the CIA + BAC-L group (Figure 7B). The effect of MTX or BAC on the knee joints of mice was measured by using micro-CT. The bone surface/volume (BS/BV) ratio of the CIA group was obviously higher than that of the control group. Administration of MTX or BAC significantly decreased the BS/BV ratio. Intriguingly, BAC at a high concentration was more effective in decreasing the BS/BV ratio (Figures 7C,D).

To further explore the effects of MTX and BAC on the CIA mice, the wrist, knee, elbow, and ankle joints of the mice were subjected to histological analysis by H&E and Safranin O staining. Compared with the control group, the joint structure was incomplete, the cartilage surface was rough, and a large number of inflammatory cells had infiltrated the joint tissues in the CIA group. Administration of MTX or BAC alleviated synovial hyperplasia and decreased the severity of cartilage damage and inflammatory cell infiltration (Figure 7E–N). Intriguingly, BAC at a low concentration was more effective in decreasing histopathological scores and modified Mankin’s scores of wrist and knee joints. BAC at a high concentration was more effective in decreasing histopathological scores and modified Mankin’s scores of the elbow and ankle joints of mice (Figure 7E–N).

The bioinformatics analysis above revealed that the screened compounds exerted their most significant impact on the biosynthesis of steroid hormones. We selected several markers in steroid hormone biosynthesis to verify the effect of BAC on the levels of these proteins (including CYP17A1, HSD17B3, TLR4, MAPK, and p65). The expressions of CYP17A1, HSD17B3 and TLR4 increased in the CIA mice compared with the control mice. The phosphorylation of MAPK and p65 both significantly increased in the CIA mice compared with the control mice. However, administration of MTX and BAC decreased the expressions of these proteins (Figure 8). Importantly, BAC at a high concentration was more effective in decreasing the expressions of CYP17A1 and HSD17B3 and the phosphorylation levels of MAPK and p65 compared with BAC at a low concentration (Figure 8).