Seven databases (PubMed, Embase, Web of Science, China National Knowledge Infrastructure, Chinese Biological Medical Disc, Chongqing VIP, and Wanfang databases) were searched from inception to May 2025, with the last search on 5 May 2025. The search terms and their extensions, including “tanshinone ⅡA,” “Alzheimer* disease,” “Alzheimer type dementia,” “senile dementia,” “animal experiment*,” and “in vivo,” were connected using Boolean operators “AND” or “OR” to retrieve compliant publications as comprehensively and accurately as possible. The search strategy was customized to suit the specific requirements of databases. The search strategies in seven databases are provided in Supplementary Material S1 and S2. A manual screening of reference lists from included studies was conducted to identify potentially eligible publications. In addition, searches of grey literature databases (OpenGrey, bioRxiv, and Google Scholar) and identification of industry-sponsored studies were also conducted. All records were deduplicated using EndNote X9, followed by Microsoft Excel.

The eligibility criteria were established under the guidance of the PICOS framework. P (population): rat or mouse models with AD; I (intervention): Tan ⅡA; C (control): vehicle-treated animals after modeling or transgenic animal model of AD undergoing no treatment at all; O (outcomes): cognitive function assessed through the Morris water maze (MWM) test and pathological changes in the brain, neuroinflammation, oxidative stress, apoptosis, and neural and synaptic plasticity; and S (study design): preclinical controlled studies.

Inclusion criteria were as follows:

• preclinical experimental studies in rats or mice, with no restriction on strain, species, sex, and model;

Exclusion criteria were as follows:

• combined with other therapies in addition to Tan ⅡA;

A two-step screening process was used for study selection. In the first step, two research workers (YR and YD) browsed titles and abstracts back-to-back to select initially qualified studies. The studies that were not included were classified according to the exclusion criteria. The second step was a full-text review by the same two research workers independently against predetermined inclusion criteria, followed by a double-checking. Any disagreements were resolved through consultation with a third reviewer (JZ). If two articles with identical content were published in different languages, only one was retained. If two articles stemmed from the same study and utilized the same batch of animals but reported different outcomes, the one containing more information was reserved and absorbed outcomes of another.

Before data extraction, a data extraction form was designed and subsequently refined after being piloted on three studies. Two research workers (YR and WW) independently extracted data for relevant outcomes in accordance with the updated form. Complete data were obtained by emailing to the author in case where there were unreported or unclear data. The key information extracted encompassed author identification (author ID and date), animal characteristics (species, strain, age, weight, model, and number of each group), intervention (administration route, dose, and treatment duration), and outcomes. The inter-rater reliability was assessed using Cohen’s kappa. For outcome indicators measured for several consecutive days, only the value of the last day was taken. Original values with actual units were recorded truthfully and then consistently converted across studies for comparability. For continuous outcomes, the mean value and standard deviation (SD) were extracted, and data were converted using the formula SD = √ n × SEM, when standard error of the mean (SEM) was given (where n represents the number of animals). WebPlotDigitizer 4.5 was used when data were presented graphically. In the presence of multiple-dose intervention subgroups, the mean values and SDs were combined into a single group using the following formula according to the Cochrane Handbook for Systematic Reviews of Interventions (CHSRI):

Sample size: N₁ + N₂ Mean : N 1 M 1 + N 2 M 2 / N 2 M 2 , SD : N 1 − 1 S D 1 2 + N 2 − 1 S D 2 2 + N 1 N 2 N 1 + N 2 M 1 2 + M 2 2 − 2 M 1 M 2 N 1 + N 2 − 1 .

Two research workers (YR and WS) independently used a 10-item scoring checklist adapted from the study by Liu et al. (2024a) to assess the reporting quality of included studies. For an individual study, each parameter was rated as “Yes,” “No,” or “Partly.” Studies awarded “Yes” in more than 70% of parameters were regarded as high quality, whereas those awarded “Yes” in less than 50% were considered as low quality, and those awarded “Yes” in 50%–70% were considered as medium quality. The risk of bias was evaluated by two independent research workers (YR and WS) using the SYRCLE’s RoB tool (Hooijmans et al., 2014), which consists of 10 items covering the following six aspects: selection bias, performance bias, detection bias, attrition bias, reporting bias, and other biases. The categories were evaluated as “Low,” “High,” and “Unclear,” which represented low risk, high risk, and unclear risk of bias, respectively. The inter-rater reliability was assessed using Cohen’s kappa. Any discrepancies between the research workers were resolved by consulting a third research worker (QL).

Review manager 5.3 and Stata 15.0 were used for data analysis. Given that all the outcomes were predefined as continuous variables, mean differences (MDs) or standardized mean differences (SMDs) and their corresponding 95% confidence intervals (95% CIs) were calculated as effect sizes, with the cutoff for statistical significance set at P < 0.05. On the one hand, when natural units were used to measure the outcome indicator, MD should be used. On the other hand, when the method or unit of measurement showed cleavage among studies, SMD was a preferred option. Random-effects models were employed in this meta-analysis and pooled analyses were visualized with forest plots. The chi-square statistic and the Higgins index (I²) were used to quantify heterogeneity among studies, and results with I² >50% were considered substantial, which implied the existence of considerable heterogeneity. If meta-analysis was not feasible, a qualitative synthesis was adopted to report the data.

To investigate the factors influencing the efficacy of Tan ⅡA in AD animal models and explore the sources of heterogeneity, subgroup analyses were conducted based on animal characteristics (species and model) and intervention characteristics (administration route, dose, and treatment duration). For studies with multiple dose groups, the average dose was used for classification. Furthermore, meta-regression was conducted using Stata 15.0 software to assess the potential influence of covariates on heterogeneity when an adequate number of studies were available.

The funnel plot and Egger’s test complemented each other in assessing the publication bias of the included studies. When there was a sufficient quantity of studies related to the outcome, funnel plots were constructed to visually assess the potential presence of publication bias. A uniform distribution of scatter points around the combined effect size indicated the absence of publication bias. Egger’s test provided a quantitative measure of publication bias, with P < 0.05 indicating the presence of such bias.

Sensitivity analysis using the “leave-one-out” method was conducted to evaluate the robustness of the meta-analysis results. In the circumstance where there was minimal influence on the overall outcome after removal, it was reasonable to believe that the results of the meta-analysis were reliable.

A total of 435 relevant studies were identified from seven typical databases, including 76 in English and 359 in Chinese. After removing duplicates using EndNote software and Microsoft Excel, 221 records remained; two reviewers then screened the titles and abstracts and excluded 190 irrelevant studies. Full texts of 31 studies were further assessed for eligibility, and 11 studies were excluded. Out of them, four reported irrelevant outcomes, two were content duplications in different languages, four lacked raw complete data or were failed to be extracted data from figures, and one used animal model of AD combined with another diseases. Given that two studies used the same batch of animals and experimental methods but reported different outcomes, likely arising from the same project, we integrated their outcomes into one study. Two records were retrieved from citations and grey literature databases, but both were duplicates already included in the previously searched typical databases. Ultimately, a total of 19 studies were included for qualitative or quantitative synthesis (Figure 2).

Nineteen studies were included involving 771 animals (581 AD and 190 normal controls). Among the AD animals, 190 were AD controls, whereas 391 were treated with Tan ⅡA. Positive-drug control groups were excluded from data analysis. Except for one study (Jiang et al., 2010) with an equal ratio of male and female rodents, all other studies used male rodents, subordinating to 333 mice (57%) and 248 rats (43%). In these 19 studies, Sprague–Dawley (SD) rats were the most common species (240 in 8 studies, 41%), followed by B6C3-Tg mice (195 in 7 studies, 34%). Kunming (40 in 1 study, 7%), C57BL/6J (18 in 1 study, 3%), ICR (48 in 1 study, 8%), and Swiss albino (40 in 1 study, 7%) mice were also used. There were four approaches to developing AD models: injection of Aβ (eight studies), transgenic AD (APP/PS1) (seven studies), streptozotocin (STZ) (three studies), and lipopolysaccharide (LPS) (one study).

Tan ⅡA was either commercially sourced from biotechnology companies or extracted in laboratory by research workers, with administration routes including both intraperitoneal injection and oral gavage. The dosages used in these 19 studies ranged from 1 mg/kg to 100 mg/kg. Thirteen studies conducted dose–control experiments on Tan ⅡA, with two or three dose groups set up. The remaining six studies set one dose group. Cohen’s kappa between YR and WW was 0.92. Further details are presented in Table 1.

A Cohen’s kappa of 0.72 reflected good agreement between reviewers YR and WS for the risk of bias evaluation using the SYRCLE’s RoB tool. For sequence generation, one study (Lin et al., 2019) used a random number table for randomization, which was considered low risk. The remaining 18 studies claimed randomization without specifying the methods, resulting in an unclear risk rating. As none of the included studies conducted baseline assessments of learning and memory performance following group allocation, the comparability of baseline characteristics across groups could not be determined, leading to an unclear risk rating in this domain. With respect to allocation concealment, the methodological details provided in all studies were insufficient to ascertain whether adequate concealment procedures were implemented, leading to all studies being rated as an unclear risk. Random housing was performed in all studies, which was therefore assessed as low risk. Caregiver blinding was assessed as unclear risk in all included studies. Regarding random outcome assessment, one study (Jiang et al., 2014) was classified as low risk, whereas the other 18 studies were deemed to have unclear risk for the obscure principles governing animal selection. Outcome assessment was processed by observers who were blinded to the experiment design in two studies (Ding et al., 2020; Peng et al., 2022), which had a low risk of bias. The remaining studies were rated as unclear risk. Six studies were rated as unclear risk for incomplete outcome data. Although the study protocols were unavailable, all prespecified outcomes were explicitly reported; thus, all studies were assessed as low risk of reporting bias. For other sources of bias, five studies failed to provide explicit declarations about potential conflicts of interest among co-authors and were assessed as unclear risk. Two studies were assessed as high risk (Wen et al., 2014; Lu et al., 2016) as they exhibited discrepancies between figures and their corresponding legends. In summary, although some studies demonstrated low risk of bias in certain domains, it should be noted that a proportion of studies exhibited unclear or even high risk of bias in specific aspects (Table 2; Figure 3).

A Cohen’s kappa of 0.83 reflected an excellent agreement on reporting quality between reviewers YR and WS. Overall, the reporting quality of the included studies was satisfactory as all studies were of high quality with a “Yes” response in more than 70% of parameters. All studies clearly reported timing of intervention, administration route, number of animals per group and defined modeling methods, outcome measures, pharmaceutical preparation, and dosage. However, none of the studies reported model validation as measures of learning and memory were recommended as the ultimate readout (Puzzo et al., 2014). Animal characteristic reporting was incomplete in most studies, typically lacking either age or weight specifications. Furthermore, six studies failed to provide essential information regarding animal dropout (Supplementary Figure S1).

A total of 14 studies conducted MWM involving 489 animals, with 340 in the experimental group and 149 in the control group. The escape latency and time spent in the target quadrant were used to evaluate cognitive function.

Aβ plaques and subsequent neuronal loss are recognized as two hallmark pathological features of AD. The hippocampal Aβ plaque burden and neuronal damage were used to assess the neuropathological effects of Tan ⅡA on brain tissue.

Inflammation markers in brain tissue were reported as TNF-α, IL-1β, and IL-6.

Oxidative stress biomarkers in brain tissue were evaluated using both antioxidant stress indicators (SOD and GSH-Px) and pro-oxidative stress indicators (MDA and ROS).

Western blot results of Caspase-3 and the Bcl-2/Bax ratio were analyzed to evaluate the antiapoptotic effect of Tan ⅡA on AD animal models.

Neural and synaptic plasticity was assessed through the quantification of PSD-95 and BDNF expressions.

Meta-regression was performed for escape latency, time spent in the target quadrant, SOD, and MDA to examine the impact of study quality (risk of bias assessment scores), population characteristics (animal species and model type), and intervention features (administration route, dose, and treatment duration) as covariates on the overall effect, thereby exploring heterogeneity. The meta-regression results for escape latency showed that only treatment duration significantly influenced the effect size (P < 0.05), indicating it as a potential factor contributing to heterogeneity. This finding was inconsistent with the results of the subgroup analysis, possibly due to residual confounding factors that prevented complete elimination of heterogeneity within subgroups (Supplementary Table S9). Meta-regression of time spent in the target quadrant, SOD, and MDA failed to detect any significant association, indicating that the factors examined may not account for the heterogeneity among studies (Supplementary Tables S10–S12).

Publication bias was evaluated for escape latency, time spent in the target quadrant, SOD, and MDA. Except SOD, visual inspection of the funnel plots revealed evident asymmetry in the distribution of data points. Notably, some points fell outside the 95% confidence intervals, further suggesting the presence of potential publication bias (Figure 9). This observation was statistically confirmed through Egger’s test for escape latency, target quadrant occupancy, and MDA (all P < 0.05); however, SOD showed borderline significance (P = 0.054). The trim-and-fill method was used to correct funnel plot asymmetry caused by publication bias, followed by repeated meta-analysis using fixed- and random-effects models. Notably, no statistically significant differences were observed in pooled effect sizes before and after trimming, which indicated minimal impact of publication bias and robust results (Supplementary Figure S6; Table 7).

Sensitivity analyses were performed using Stata 15.0 software for outcomes including escape latency, time spent in the target quadrant, SOD, and MDA levels to explore the impact of individual studies on the overall effect (Figure 10). The results revealed that the exclusion of any single study had negligible effects on the pooled effect sizes, supporting the consistency and reliability of the meta-analysis outcomes.

To our knowledge, no previous SR/MA has comprehensively evaluated the preclinical efficacy of Tan ⅡA in AD models. Herein, we conducted this work. A total of 19 in vivo studies involving 581 AD animals were included, all of which were conducted in China between 2010 and 2024, with the majority published in English. The overall reporting quality was comprehensive and standardized, whereas methodological assessment revealed potential risks of bias. Results of the present study showed that Tan ⅡA significantly improved cognitive performance, as indicated by reduced escape latency and increased target quadrant retention time in the MWM test. Furthermore, Tan ⅡA treatment attenuated AD-related neuropathological changes, evidenced by reduced hippocampal Aβ plaque burden and neuronal damage in brain tissue. At the molecular level, Tan ⅡA significantly downregulated pro-inflammatory cytokines (TNF-α, IL-1β, and IL-6). The compounds also exhibited potent antioxidant properties, as demonstrated by decreased MDA and ROS levels and increased SOD and GSH-Px levels. Meanwhile, Tan ⅡA displayed notable antiapoptotic effects through Caspase-3 suppression and Bcl-2/Bax ratio elevation. Additionally, treatment upregulated neural and synaptic plasticity markers (PSD-95 and BDNF), suggesting the potential of Tan ⅡA for neural and synaptic restoration in AD pathology.

The observed outcomes in this study corroborate earlier findings that Tan ⅡA exhibits neuroprotective effects in AD treatment (Subedi and Gaire, 2021; Sherawat and Mehan, 2023), primarily through modulation of inflammation, oxidative stress, apoptosis, synaptic plasticity, Aβ aggregation, and Tau hyperphosphorylation (Figure 11).

Elevated inflammatory cytokines have been observed in the brain of patients with early-stage AD, suggesting a significant contribution of neuroinflammation to AD pathology (Heneka et al., 2015). Aβ peptide accumulation induces chronic microglia activation, triggering the inflammatory cascade by eliciting expressions of pro-inflammatory cytokines, including TNF-α, IL-1β, and IL-6. These cytokines subsequently contribute to neuronal dysfunction, apoptosis, and synaptic loss (Leng and Edison, 2021; Thakur et al., 2023). Tan ⅡA can suppress glial activation and subsequent neuroinflammatory responses by downregulating Iba-1 and GFAP expressions, an effect partially attributed to its inhibition of the RAGE/NF-κB signaling cascade (Ding et al., 2020). Additionally, Tan ⅡA exerts anti-inflammatory effects via upregulation of PI3K/AKT phosphorylation (Fang et al., 2021).

Compared with other organs, the brain is particularly susceptible to oxidative stress damage. The imbalance between oxidation and antioxidation generates an excess of ROS, leading to damage to lipids, proteins, and DNA, which may underlie the cognitive impairment in AD (Butterfield et al., 2001). MDA, which is the end product of lipid peroxidation, exerts severe cytotoxic effects. SOD, the first line of defense against oxidative damage, and GSH-Px, a key antioxidant enzyme in the body, work together to protect the brain from oxidative stress. Tan ⅡA can suppress oxidative stress-induced brain tissue damage and promote cognitive function recovery. Mechanistically, these effects are associated with the activation of the CREB-BDNF-TrkB pathway (Xiang et al., 2024a), stimulation of the PI3K/Akt/GSK3β pathway (Peng et al., 2022; Liu et al., 2024b), and inhibition of the p38 MAPK pathway (Liu et al., 2016), thereby enhancing the activity of endogenous antioxidants (SOD and GSH), reducing MDA levels, and scavenging ROS.

Neuronal apoptosis compromises brain structure and function. Postmortem studies have identified distinct apoptotic features in AD brains, including altered expression profiles of proapoptotic and antiapoptotic regulators in brain extracts (Kitamura et al., 1998). Endoplasmic reticulum stress (ERS) represents a critical apoptotic pathway, triggering the unfolded protein response (UPR) which either orchestrates adaptive programs to restore homeostasis or triggers apoptosis of irreversibly damaged cells (Gerakis and Hetz, 2018). Tan ⅡA exerts antiapoptotic effects through multiple mechanisms: a. by downregulating GRP78, a core ERS chaperone protein, and suppressing PERK/eIF2α, IRE1α, ATF6, CHOP, and JNK signaling pathways to mitigate ERS-induced apoptosis (He et al., 2020a; He et al., 2020b); b. through PI3K/Akt-mediated regulation of apoptotic proteins, suppressing Bax while enhancing Bcl-2 expression (Liu et al., 2024b); and c. by inhibiting caspase cascade activation, particularly preventing Caspase-3 activation, thereby attenuating apoptotic progression (Wen et al., 2014; He et al., 2020a; He et al., 2020b).

Neuronal death or a reduction in neuronal density is observed in brain regions critically associated with memory function in AD patients (Miller et al., 2022). Neuronal loss in AD initiates during the preclinical stage and progressively advances through the prodromal phase, including mild cognitive impairment (MCI), ultimately culminating in dementia. Synaptic degeneration, serving as an early harbinger of neuronal degeneration, was discovered to typically precede neuronal loss, with both pathological mechanisms synergistically contributing to AD-associated cognitive dysfunction (Abdi et al., 2022; Koch and Spampinato, 2022). Neuronal and synaptic functions are regulated by a complex interplay of neurotransmitters, neurotrophic factors, and associated proteins. Tan ⅡA can promote synaptogenesis and enhance neuronal plasticity by upregulating synaptic proteins SYN and PSD-95 and activating BDNF (Ding et al., 2020; Liu et al., 2024b). Moreover, it demonstrates neuroprotective properties by rescuing long-term potentiation defects, which is associated with the activation of the CREB-BDNF-TrkB pathway (Xiang et al., 2024a).

Aβ plaque and NFT accumulation, the two typical pathological hallmarks of AD, synergize with neuroinflammation, oxidative stress, apoptosis, and synaptic damage, ultimately resulting in cognitive decline. Tan IIA not only indirectly reduces Aβ deposition and NFT formation by ameliorating the above processes but also exerts direct effects. It enhances Aβ degradation by upregulating IDE and NEP (Liu et al., 2024b), two pivotal Aβ-degrading enzymes with dual intracellular and extracellular activities. It has been shown that Tan IIA can upregulate sAPPα expression, which is a cleavage product of APP by α-secretase, implying a reduction in Aβ production via the β-secretase pathway (Zhang et al., 2020). Additionally, it modulates the RAGE/LRP1 transport system on vascular endothelial cells to accelerate trans-endothelial clearance of Aβ (Wan et al., 2023). In the formation and aggregation of NFTs, Tan IIA suppresses Tau hyperphosphorylation by inhibiting the activation of GSK-3β and ERK (Lin et al., 2019). An in vitro experiment demonstrates that Tan IIA induces Tau polyubiquitination, leading to its proteasomal degradation. Meanwhile, Tan IIA can bind to Tau protein to inhibit the formation of Tau fibrils (Cai et al., 2020).

First, the preponderance of included studies conducted in China potentially introduces geographical bias into these meta-analysis results. We advocate for multicenter, international studies from different geographical regions. Moreover, studies should standardize experimental protocols to enhance repeatability and comparability of results. By incorporating a broader range of contexts, future research can generate more globally applicable evidence for Tan IIA’s efficacy in AD treatment.

Second, the evaluation of risk of bias and methodological quality uncovered significant flaws in research implementation, especially unclear risks regarding blinded outcome assessors and the lack of model validation in most studies. Failure to apply the blinding method is likely to lead to an overestimation of the effect size of Tan IIA. Model validation based on pathology or ethology is the foundation for simulating the clinical features of AD and evaluating the effect of Tan IIA. These limitations threaten the stability of results and the generalizability of the conclusion, warranting cautious interpretation. In future experimental studies, scientific and rigorous protocols should be formulated with reference to the guidelines, and the principles of randomization and blinding should be followed to reduce bias and strengthen experimental validity.

Finally, the limitations of this SR/MA must be acknowledged. PICO-related heterogeneity across included studies, such as varied animal models, intervention protocols, and outcome assessments, may limit the reliability of synthesized results. Despite efforts to retrieve grey literature, no additional studies were available. The absence of unpublished research, including negative, ongoing, and industry-sponsored studies, compromised the comprehensiveness of the search. Future studies should expand the search scope, implement rigorous literature screening and bias assessment, and adopt advanced statistical methods to address heterogeneity, thereby enhancing the quality and credibility of the findings.

Tan IIA may exert side effects in humans or animals. Safety outcomes were derived from 22 clinical studies using sodium tanshinone IIA sulfonate (STS) injection (Wang et al., 2014). Among 27 cases, skin and adnexa injuries (e.g., rash) accounted for 29.9%, systemic damages (e.g., anaphylactic shock) accounted for 21.5%, circulatory system injuries (e.g., chest tightness) accounted for 20.6%, and central/peripheral nervous system injuries (e.g., dizziness) accounted for 15.9%. All the same, these adverse reactions resolved or subsided after symptomatic treatment, without obvious sequelae. We noted that a zebrafish model study demonstrated developmental, cardiovascular, and neurotoxicity of Tan ⅡA at 40 μM concentration (Lai, 2020). In contrast, STS injection at clinical concentrations (1 mg/ml) showed no hemolytic or erythrocyte agglutination effects in guinea pigs, and equivalent doses (4 mg/kg) exhibited no irritant effects on rabbit auricular veins or quadriceps muscles, nor did it induce systemic allergic reactions in guinea pigs (Cao et al., 2010). Notably, none of the included studies in this SR/MA measured safety for Tan IIA in AD treatment, posing challenges to clinical translation of our findings. Given Tan IIA’s potential for AD, future research should focus on safety evaluation to determine safe dosage, concentration, and administration frequency.