Chinese databases of China National Knowledge Infrastructure (CNKI), Wanfang, VIP, and English databases of PubMed, Embase, and Cochrane Library were retrieved from inception to 14 August 2023. The basic search framework was: (“total glucosides of paeony”) and (“rheumatoid arthritis” or “experimental arthritis”). Moreover, retrieve references cited by the included studies and include them if they met the inclusion criteria. The previous step was repeated for that included reference until no references met the inclusion criteria. This is to minimize the rate of missed detections. Detailed search strategies can be found in Supplementary Material S1.

1) Study Types: In vivo animal experiments.

(1) The same study with duplicate publications (retaining the one with more complete data); (2) Reviews; (3) Clinical studies and case reports; (4) Conference literature, scientific and technological achievements, and guidelines; (5) Cell experiments; (6) Interventions using paeoniflorin or other components of TGP; (7) Studies that do not report data.

Two researchers (Mengdi He and Zhipeng Hu) independently screened the literature by reading the title, abstract, and full text to obtain the literature that met the inclusion criteria. Disagreements were resolved by consulting a third researcher (Maoyi Yang). Two researchers (Mengdi He and Zhipeng Hu) independently extracted data in Excel, including literature author, publication year, country, sex ratio, animal species and strains, age, body weight, equivalent dose, type of animal model, sample size, co-intervention, interventions in experimental versus control groups, route of administration, intervention time, manufacturer of TGP, paeoniflorin content, measurement specimens, outcomes and their detection methods. Then cross-check to reduce the error rate. Outcome data were recorded as mean and standard deviation. If the literature provided standard error, the data were transformed. If the literature presented data only in the form of images, data were extracted three times using WebPlotDigitizer and averaged. If the literature lacked relevant data, Mengdi He attempted to contact the authors and excluded the literature when no response resulted in incomplete key data.

Two investigators (Mengdi He and Zhipeng Hu) used the SYRCLE tool to assess the risk of bias in animal experiments (Zeng et al., 2015), which included the following 10 perspectives: (1) sequence generation; (2) baseline characteristics; (3) allocation concealment; (4) random housing; (5) blinding of animal keepers and investigators; (6) assessment of randomization outcomes; (7) blinding of outcome evaluators; (8) reporting of incomplete data; (9) selective reporting of outcomes; (10) other sources of bias. Each included study was classified as Yes (low risk of bias), No (high risk of bias), or Uncertain (uncertain risk of bias). Disagreements were resolved by consulting a third researcher (Maoyi Yang).

Meta-analysis was performed using RevMan 5.4 provided by the Cochrane Collaboration and Stata 12.0 software. When multiple intervention groups were included in a study, the common control group was evenly divided into several small sample groups (Control group sample size divided by the number of doses in the intervention group), and then all pairwise comparison combinations were included in the meta-analysis to avoid artificially expanding the sample size (Higgins et al., 2023). For example, in a study with three doses of TGP intervention groups (n = 10) and a control group (n = 10), the sample size of the control group was split into 3, 3, and 4, which were combined with the intervention groups into the 3–10, 3–10, 4–10 comparison combinations.

All outcomes were continuous variables. If the effect size units and measurement methods were the same among studies, effect sizes were represented by weighted mean difference (WMD) and 95% confidence interval (CI). Conversely, standardized mean difference (SMD) and 95% CI were used. Combined effect sizes were considered statistically significant at p < 0.05. I ² and Q-test were used to assess the heterogeneity of the studies. If I ² ≤ 50% or p ≥ 0.1, the heterogeneity among the studies was small, and the fixed-effects model was used. If I ² > 50% or p < 0.1, the heterogeneity among the studies was large, and leave-one-out sensitivity analysis was performed first to explore the source of heterogeneity and judge the stability of the results. If the source could not be found, the random-effects model was used and subgroup analysis was performed.

Subgroup analysis was performed from five perspectives: dose and manufacturer of TGP, model type (CIA or AIA), animal species (such as rat or mouse), and animal strain [such as Wistar or Sprague Dawley (SD) rat, DBA/1 mouse]. Subgroup analysis for each perspective was carried out when the number of subgroups was less than or equal to 3. The dose subgroup analysis was based on 1/2 of the equivalent dose, the equivalent dose, and two times the equivalent dose as the division points (Tang, 2020). Specifically, in rats (200 g), 0 < low dose (LD) group ≤81.5 mg·kg-1·day-1, 81.5 mg·kg-1·day-1< medium dose (MD) group ≤163 mg·kg-1·day-1, 163 mg·kg-1·day-1 < high dose (HD) group ≤326 mg·kg-1·day-1; in mice (20 g), 0< LD ≤ 117 mg·kg-1·day-1, 117 mg·kg-1·day-1 < MD ≤ 234 mg·kg-1·day-1, 234 mg·kg-1·day-1 < HD ≤ 468 mg·kg-1·day-1. Although the equivalent doses for 200 g rats and 20 g mice obtained by conversion from the standard dose for a 70 kg human are acceptable within a certain body weight range, large conversion errors occur when the actual body weights of the rats or mice are outside that range (Xia et al., 2021). Therefore, in the dose subgroup analysis, we convert the equivalent dose and adjust the low, medium, and high dose ranges one by one according to the mean of the actual body weight of the animals in each study.

When there were ten or more comparison groups for the same outcome, publication bias was assessed by plotting a funnel plot or performing an Egger’s test (with p < 0.05 as the presence of publication bias).

A total of 1478 studies were retrieved from the databases (CNKI, 545; Wanfang, 462; VIP, 295; PubMed, 86; Embase, 77; Cochrane Library, 13). First, 743 studies were excluded due to duplication, and then 693 studies were excluded by reading the title and abstract. 20 studies were excluded by reading the full text and one of them was categorized as unclassified because the measurement specimen was unknown (Fu et al., 2022), and then 22 studies were included. By searching the references of the included studies, two additional studies were included (Chang et al., 2016; Asenso et al., 2019), and finally, 24 studies were included (Asenso et al., 2019; Chang et al., 2016; Jia, 2015; Li et al., 2019; Li, 2012; Li et al., 2011; Li et al., 2010; Liu et al., 2010; Liu et al., 2012; Peng et al., 2012; Qin, 2008; Qu et al., 2022; Shen et al., 2019; Wang et al., 2013a; Wang et al., 2013b; Wang, 2019; Xie, 2019; Yan et al., 2021; Yao, 2007; Yuan, 2013; Zhou et al., 2019; Zhou et al., 2018; Zhou et al., 2012; Zhou et al., 2009). It should be noted that during the screening process, regarding characteristics and outcome data, study (Chang et al., 2012a) is similar to study (Li et al., 2010), and study (Chang et al., 2012b) is similar to study (Li et al., 2011), and there was a great possibility of duplicated reporting. So Mengdi He tried to contact the authors for more detailed information, but the authors did not respond, and out of an abundance of caution, we excluded two studies by Jingzhi Chang (Chang et al., 2012a; Chang et al., 2012b). The specific literature screening flow chart is shown in Figure 1.

The meta-analysis included a total of 41 independent comparison groups, 581 experimental and control animals, all of which were rodents, including 510 rats (392 SD rats, 98 Wistar rats, 20 Lewis rats) and 71 mice (31 DBA/1 mice, 40 C57BL/6J mice). Fifteen studies used SD rats, five studies used Wistar rats, one study used Lewis rats, two studies used DBA/1 mice, and one study used C57BL/6J mice. In terms of model types, 12 studies used the AIA model, and 12 studies used the CIA model.

The dose range of TGP was 25–370 mg·kg-1·day-1, with six studies involving three doses, one study involving two doses, and 17 studies involving one dose. The intervention time of TGP ranged from 1 to 8 weeks, with 23 studies for 1–4 weeks and one study for 8 weeks. Regarding the reported time points of outcomes, only two studies reported multiple time points. TGP originated from six different manufacturers: (1) Institute of Clinical Pharmacology, Anhui Medical University. (2) Anhui Bozhou Chinese herbal medicine market (self-made purification). (3) Ningbo Lihua Pharmaceutical Co., Ltd. (4) Shanxi Weikangtang Chinese Herbal Pieces Co., Ltd. (5) Sanjiu Pharmaceutical Co., Ltd. (6) Ningbo Yicuijian Biological Technology Co., Ltd. Among them, the first two are self-made purification, and the last four are proprietary Chinese medicine. The specific basic characteristics are shown in Table 1.

Only one study described the specific method of random sequence generation, one study had an incorrect method of sequence generation, and the rest mentioned randomization. The baseline characteristic of only one study was rated as low risk because it described serum cytokine levels after modeling. This phenomenon may be caused by the fact that existing studies were still accustomed to using outcomes such as paw swelling and arthritis index to measure baseline characteristics, and the topic of some studies was not that type of outcome, so it was not paid attention to before intervention of TGP. None of the studies described allocation concealment or blinding of animal keepers and investigators. Seven studies were rated as unclear risk in random outcome assessment. One study stated blinding of outcome evaluators. 13 studies mentioned randomization of animal placement. One study had incomplete data reporting. Two studies were rated as high risk for selective reporting of outcomes, one of which had multiple time points but reported only one, the other described measurement of serum IL-10 levels in the experimental method, but the results were not reported. All other sources of bias were low risk. The risk of bias summary and risk of bias graph are shown in Figure 2 and Figure 3.

This study summarized potential mechanisms of TGP on RA in the included studies, which were mainly divided into four aspects: regulating inflammatory cytokines, inhibiting synoviocyte proliferation and promoting synoviocyte apoptosis, inhibiting bone destruction, and regulating immune function. The specific mechanisms are shown in Table 2.

By combining 24 studies, this article found that compared to the model control group, TGP decreased the levels of TNF-α, IL-1β, IL-6, and PGE2 and increased the levels of TGF-β1 after 1–2 weeks of intervention. After 3–4 weeks of intervention, the levels of TNF-α, IL-1β, IL-6, and so on were decreased and the levels of IL-10 and IL-4 were increased. After 8 weeks of intervention, the levels of TNF-α, IL-6, and IL-17α were decreased and the level of IL-10 was increased. There was no significant difference in the effects of TGP on the levels of IL-10, IL-17, and IFN-γ after 1–2 weeks of intervention and IL-1 and TGF-β1 after 3–4 weeks of intervention. The specific results are summarized in Table 3.

Due to the differences in animal species, strain, model type, TGP’s dose and manufacturer, intervention time, and other factors, clinical heterogeneity may be caused. Therefore, we reduced the possible clinical heterogeneity from two perspectives. First, before including the studies, we limited the dose to no more than twice the standard dose of the corresponding species. And the route of administration is limited to gavage. Second, before the results were combined, they were divided into 1–2 weeks, 2–4 weeks, and 8 weeks according to the different intervention times of the included studies.

After combining the results, this study found that the heterogeneity of most results was generally high. For this reason, the sensitivity analyses performed first proved that the results were all robust, and the subgroup analyses performed later found that the heterogeneity mainly originated from the animal strain, TGP’s manufacturer, and dose, while the animal species and model type were not the sources of heterogeneity. In terms of model types, neither AIA nor CIA can fully imitate human RA and can only reflect the nature of the disease from some aspects. Although there may be clinical heterogeneity between them, the results of this study showed that the model type was not the source of heterogeneity.

Although many studies were included in this study, most of them reported fewer outcomes, resulting in fewer comparison groups included in some outcomes. Moreover, the sources of heterogeneity are not consistent across outcomes, so sources of heterogeneity obtained through subgroup analyses need to be viewed with caution (Sun et al., 2010).

This study has five implications for future research. First of all, there may be some methodological heterogeneity in the included studies. There were deficiencies or lack of relevant instructions in random sequence generation, baseline characteristics, allocation concealment, blinding of animal keepers and investigators, and blinding of outcome evaluators, so it is hoped that future researchers will be more rigorous in designing and executing their experimental protocols, and more standardized in writing their manuscripts.

Secondly, the outcomes vary greatly among the included studies, and core outcome sets need to be studied and popularized (Kirkham et al., 2017). And then, at present, the experimental animals used in the relevant studies are mainly in the lower age stages, and there is a lack of studies in the higher age stages. It is hoped that future researchers can consider that the main incidence population of RA is in the high age stage, and supplement the research of related age stages.

Next, the intervention time of TGP was obviously discontinuous, and there was a lack of studies on corresponding intervention time (5–7 weeks or even longer). It is not difficult to find that the low, medium, and high dosages in some studies were 25, 50, and 100 mg·kg-1·day-1, respectively (Li et al., 2010; Liu et al., 2010; Li et al., 2011; Liu et al., 2012; Yan et al., 2021), which were different from the pre-designed dose subgroup analysis. This may be due to the fact that researchers are not starting from a clinical reality. Therefore, the intervention time and dose of TGP need to be reported more comprehensively, the minimum intervention time and dose that begin to take effect should reach a consensus, and the maximum intervention time and dose that stop increasing effect remain to be investigated. More high-quality animal studies are expected to emerge as a reference for clinical research. Finally, funnel plots of TNF-α (Intervention time is 3–4 weeks) and IL-1β (Intervention time is 3–4 weeks) suggest publication bias, so negative results need to be encouraged for publication.

This study is the first systematic review and meta-analysis of the effects of TGP on serum inflammatory cytokines in animal models of RA, with a complete set of outcomes included. There is a comprehensive study in this aspect. However, some studies have also reported local-level outcomes such as inflammatory factors and protein molecules in foot and claw synovial tissues, articular cartilage, and joint fluids. For example, TGP may exert anti-inflammatory effects by regulating the balance between Gαi and Gαs and decreasing the expression of β-arrestins (Wang et al., 2011; Jia et al., 2014). This aspect is not covered in this study, mainly due to the difference in measurement specimens and the lack of relevant studies. This perspective is expected to serve as an entry point for future systematic reviews and Meta-analyses to further evaluate the improvement of localized lesions in RA after drug administration, which will help to uncover and summarize the mechanism of TGP in the treatment of RA.