This systematic review and meta-analysis followed the methodological guidelines of the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) (Page et al., 2021) to ensure methodological rigor. Moreover, the research protocol was prospectively registered with the International Prospective Register of Systematic Reviews (PROSPERO) under registration number CRD42023486389.

A comprehensive search was systematically conducted across both English and Chinese databases, covering the period from their inception to February 2025. The English databases included PubMed, Embase, Cochrane Library, EBSCO, and Web of Science, while key Chinese databases—China National Knowledge Infrastructure (CNKI), Chinese Biomedicine Literature Database (CBM), Chinese Science and Technology Journal Database (VIP), and Wanfang Database—were also thoroughly examined. The search strategy combined medical subject headings (MeSH) with free-text terms, ensuring a detailed analysis of titles, abstracts, and keywords. Specifically, the terms (“Acinetobacter baumannii” OR “A. baumannii”) AND (“Xuebijing” OR “XBJ”) were applied, with modifications tailored to the indexing systems of individual databases. To further enhance completeness, reference lists of included studies and relevant systematic reviews/meta-analyses were manually screened. Detailed search strategies are provided in Supplementary Material 1.

Inclusion criteria: Clinical trials assessing the therapeutic efficacy of XBJ in combination with other standard treatments for A. baumannii infection were eligible for inclusion, provided that the effects of XBJ could be clearly differentiated from those of other interventions. Eligible studies were required to report at least one of the following outcomes: overall effective rate, bacterial clearance rate, levels of inflammatory markers (including white blood cell (WBC) count, C-reactive protein (CRP) level, and procalcitonin (PCT)), or adverse drug events. Only randomized controlled trials (RCTs) published in English or Chinese were considered for inclusion.

Exclusion Criteria: Studies that were non-clinical, reviews, case reports, or conference abstracts were excluded. Non-randomized studies of interventions (NRSIs) were also excluded. Furthermore, studies with insufficient outcome reporting, significant methodological flaws, duplicate publications, or irrelevant data were excluded from this analysis.

Two evaluators (YHS and JQL) independently screened the literature, extracted data, and cross-verified their findings. Any discrepancies were resolved through discussion or, when necessary, consultation with a third reviewer (HW) to reach a consensus. The screening process followed a stepwise approach, beginning with an initial assessment of titles and abstracts to exclude clearly irrelevant studies, followed by a full-text review based on predefined inclusion criteria. When key data were missing or unclear, the original authors were contacted via telephone or email for clarification or additional information. The extracted data encompassed study characteristics (author, publication year, country, study design, sample size, participant demographics, and specific disease conditions), intervention details (single dose, frequency, daily dosage, treatment duration, and total dosage of XBJ), and outcome measures (overall effective rate, bacterial clearance rate, and inflammatory markers such as CRP, PCT, and WBC), along with adverse drug events. Additionally, information relevant to the risk of bias assessment, including methods of randomization, allocation concealment, and other methodological aspects, was systematically collected.

The risk of bias for all included studies was independently assessed following the guidelines of the Cochrane Handbook (version 6.4) (JPT et al., 2023). The methodological quality of RCTs was evaluated using the revised tool for risk of bias in randomized trials (RoB 2) tool, a revised framework for assessing bias in randomized trials (Sterne et al., 2019). Two researchers (HW and JMW) independently conducted the assessments, classifying studies as having low, some concerns, or high risk of bias. Any discrepancies were resolved through discussion or, if necessary, by consulting the supervisor (MYY) for consensus. The results were visualized using the R package “robvis”.

Meta-analysis was conducted using the “meta” (Balduzzi et al., 2019), “metafor” (Viechtbauer, 2010), and “dmetar” (Harrer et al., 2019) packages in R software (version 4.3.2). Risk ratios (RR) was employed for categorical variables, while standardized mean differences (SMD) was used for continuous variables, with a 95% confidence interval (95% CI) as the measure of effect size. Heterogeneity was assessed using the Cochrane Q and I² statistics, with a P-value >0.1 and I² < 50% indicating low heterogeneity, warranting the use of a fixed-effects model. Conversely, if P ≤ 0.1 and I² ≥ 50%, significant heterogeneity was assumed, necessitating the application of a random-effects model. Sensitivity analysis was performed to evaluate the influence of individual studies on overall heterogeneity. To further explore sources of heterogeneity and assess the impact of key factors on the overall effect, subgroup analysis was conducted when at least six studies were available. Publication bias was assessed using Egger’s test, funnel plots, and the trim-and-fill method.

Trial sequential analysis (TSA) was performed to assess potential errors arising from limited sample sizes (Wang et al., 2023). Graphical representations depicted sample size on the x-axis and cumulative Z-scores on the y-axis, with parallel lines representing conventional significance thresholds. The required information size (RIS) was automatically estimated with a pre-specified type I error rate of 5%. Robustness was determined by the intersection of the cumulative Z-curve with both conventional and TSA boundaries, indicating whether further studies were necessary.

In this study, the quality and relevance of evidence were systematically evaluated using the GRADE (Grading of Recommendations Assessment, Development, and Evaluation) framework. The GRADEpro software was employed to assess the certainty of each outcome in accordance with established guidelines. This assessment incorporated key methodological domains, including study design, risk of bias, inconsistency across findings, indirectness of evidence, and imprecision of estimates. Based on these criteria, evidence certainty was classified into four levels: high, moderate, low, or very low. To further refine the evaluation, outcome importance was rated on a nine-point scale, stratified into three significance categories: non-critical, important, and critical. This structured approach ensured a rigorous appraisal of both the strength and applicability of the evidence, providing a robust foundation for interpreting the findings within a broader scientific and clinical context.

The bioactive ingredients of XBJ were identified through previous published mass spectrometry-based studies (Huang et al., 2011; Jiang et al., 2013; Sun et al., 2017; Zuo et al., 2017a; Zuo et al., 2017b; Zuo et al., 2019; Yu et al., 2022; Li et al., 2023), and subsequently analyzed using SwissTargetPrediction (Daina et al., 2019) to predict associated protein targets of XBJ ingredients. A. baumannii-related genes were retrieved from the GEO dataset GSE69528 (Pankla et al., 2009) and reanalyzed with the limma package in R to identify differentially expressed genes (DEGs). The intersection of predicted protein targets of XBJ ingredients and A. baumannii-related genes was visualized using a Venn diagram. To explore molecular interactions, the Search Tool for the Retrieval of Interacting Genes/Proteins (STRING) database (https://cn.string-db.org/) and Cytoscape (version 3.10.1) were employed to construct the protein-protein interaction (PPI) network. The “analyze network” function in Cytoscape was applied to score the network, with the two nodes exhibiting the highest degree selected as key targets.

Molecular docking was performed using AutoDock Vina (version 1.5.7) (Trott and Olson, 2010) to predict ligand-target interactions. Crystal structures of target proteins were retrieved from the Protein Data Bank (PDB) (https://www.rcsb.org/), while 3D ligand structures were batch downloaded from the PubChem database (https://pubchem.ncbi.nlm.nih.gov/) using PubChemPy in Python (version 3.11.7). Docking results were visualized and analyzed using PyMOL (version 2.4.0) and Discovery Studio (version 2019).

Molecular dynamics simulations were conducted using GROMACS (version 2023.3) to evaluate the stability of protein-ligand complexes. Protein and ligand structures were separately processed, with topology files and simulation boxes generated using the pdb2gmx and gmx editconf commands, respectively. The system underwent initial energy minimization via the steepest descent method, followed by 100,000 steps of isothermal-isovolumetric (NVT) and isothermal-isobaric (NPT) equilibration, with coupling constants set at 0.1 ps and durations of 100 ps. Subsequently, a free dynamics simulation was performed with a 2 fs time step, extending up to 100 ns using the gmx grompp and gmx mdrun commands. Conformational data were continuously recorded to monitor the stability of the protein-ligand complexes.

A total of 603 relevant articles were imported into EndNote X9. Using the “find duplicates” feature, 33 duplicate records were automatically removed, with an additional 14 eliminated through manual screening. Among the remaining 556 articles, a review of titles and abstracts resulted in the exclusion of 522 studies, including 416 that did not meet the experimental requirements, interventions, or methods, 92 review articles, and 14 conference abstracts. A full-text review of the remaining 34 articles led to further exclusions: 6 for non-compliance with control principles, 1 for not meeting outcome criteria, 3 for non-compliance with diagnostic standards, 2 for not being RCTs, and 11 for being animal studies (Figure 1). Ultimately, 11 studies were included in the meta-analysis (Bai et al., 2017; Zhang et al., 2017; Ma et al., 2018; Dong, 2019; He et al., 2019; Liu et al., 2019; Liu, 2020; Ma et al., 2020; Pang et al., 2021; Niu and Zhu, 2022; Ma et al., 2024). Detailed flowchart of literature search and study selection is shown in Figure 1.

This analysis included 11 RCTs conducted in Mainland China between 2017 and 2024, enrolling a total of 1,035 participants, of whom 521 received XBJ as adjuvant treatment and 514 received conventional symptomatic therapy, with individual study sample sizes ranging from 66 to 140. The mean participant age varied from 47.67 to 71.03 years, with a generally balanced gender distribution. All studies had evaluated XBJ as an adjuvant therapy alongside conventional symptomatic treatment, either alone or in combination with antibiotics, for A. baumannii infections. XBJ was administered at doses ranging from 50 to 100 mL, with a frequency of one to three times daily, over treatment durations varying from 7 to 30 days. Nine studies assessed the overall effective rate, five evaluated bacterial clearance rate, and seven reported changes in inflammatory markers such as CRP and PCT. Additionally, five studies analyzed white blood cell count, while only three examined adverse drug events. The general characteristics of the included studies were summarized in Table 1.

Although all included studies reported randomization, only eight trials (Ma et al., 2018; Dong, 2019; He et al., 2019; Liu et al., 2019; Liu, 2020; Ma et al., 2020; Pang et al., 2021; Ma et al., 2024) explicitly described the randomization method. In contrast, the remaining studies lacked clarity in random sequence generation, which raised concerns regarding allocation concealment (Bai et al., 2017; Zhang et al., 2017; Niu and Zhu, 2022). Regarding deviations from intended interventions, one study (Zhang et al., 2017) was rated as having some concerns due to insufficient methodological detail. Overall, six studies were classified as low risk of bias (Ma et al., 2018; He et al., 2019; Liu et al., 2019; Ma et al., 2020; Pang et al., 2021; Ma et al., 2024), three as having some concerns (Bai et al., 2017; Dong, 2019; Liu, 2020), and two as high risk (Zhang et al., 2017; Niu and Zhu, 2022). A summary of the risk of bias assessment was provided in Figure 2.

Egger’s test for funnel plot asymmetry and the trim-and-fill method were employed to assess publication bias for the primary outcomes, including overall effective rate and bacterial clearance rate. For overall effective rate, Egger’s test indicated no significant publication bias (t = −0.26, df = 7, P = 0.8048), with a bias estimate of −0.2458 (SE = 0.9575), suggesting minimal asymmetry. For bacterial clearance rate, some evidence of asymmetry was observed (t = 2.87, df = 3, P = 0.0643), with a bias estimate of 1.7239 (SE = 0.6016), indicating a potential bias toward positive results. However, the further trim-and-fill method did not identify any missing studies, suggesting no need for adjustments in the funnel plots for overall effective rate (Figure 5A) or bacterial clearance rate (Figure 5B).

In our analysis, TSA was conducted to evaluate cumulative evidence and assess whether the available data provided sufficient certainty to draw reliable conclusions regarding the intervention’s efficacy. In the TSA for overall effective rate (Figure 5C) and bacterial clearance rate (Figure 5D), the cumulative Z-score line crossed the benefit boundary and met the required sample size for both outcomes, as indicated in the 90% (red) and 99% (green) analyses.

The GRADE evaluation revealed varying levels of certainty across key outcomes (Table 2). The overall effective rate (nine RCTs) was rated high certainty, indicating a significant improvement with XBJ compared to conventional treatment (RR 1.20, 95% CI: 1.13–1.27), despite concerns over potential bias. Similarly, the bacterial clearance rate (five RCTs) was supported by high-certainty evidence (RR 1.44, 95% CI: 1.23–1.68), reflecting a robust and consistent therapeutic benefit. In contrast, the certainty of evidence for inflammatory markers was considerably lower. CRP (seven RCTs) and WBC count (four RCTs) were rated low certainty, while PCT (five RCTs) was rated as very low-certainty evidence, primarily due to serious risk of bias, inconsistency, and suspected publication bias. For adverse drug events (three RCTs), XBJ significantly reduced incidence (RR 0.30, 95% CI: 0.13–0.72), though the certainty of evidence remained moderate to high, constrained by a small sample size.

A comprehensive analysis was conducted on 145 active ingredients of XBJ, utilizing SwissTargetPrediction to predict their target proteins. A stringent filtering process, which applied a probability threshold of >0.1, identified 758 potential target proteins of XBJ ingredients. Differentially expressed gene (DEG) analysis of the GSE69528 dataset revealed 533 genes upregulated in the A. baumannii infection group compared to the control group (Figures 6A,B). The intersection of these datasets yielded 35 common target proteins (Figure 6C).

Functional enrichment analyses, including Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analyses, revealed associations with key inflammation-related pathways such as Tumor Necrosis Factor-alpha (TNF-α) signaling pathway, Nuclear Factor Kappa B (NF-κB) signaling pathway, and Hypoxia-Inducible Factor 1 (HIF-1) signaling pathway, as well as pathways involved in infectious diseases, including Leishmaniasis, Hepatitis B, and Kaposi sarcoma-associated herpesvirus infection (Figure 6D). To further explore molecular interactions, a protein-protein interaction (PPI) network was constructed using the STRING database and visualized in Cytoscape. Network analysis ranked Matrix Metalloproteinase 9 (MMP9) and TLR4 as the top two key targets based on degree centrality (Figure 6E).

AutoDock Vina was employed for batch molecular docking between the 145 active ingredients of XBJ and the active pocket regions of MMP9 and TLR4. The results identified three compounds—scutellarin (labeled in blue), salvianolic acid C (labeled in red), and isosalvianolic acid C (labeled in green)—as potential candidates for simultaneous binding to the active pockets of both MMP9 (labeled in gray) and TLR4 (labeled in cyan). Each of these compounds formed multiple stable hydrogen bonds with MMP9 and TLR4 (Figure 7, with hydrogen bonds shown in yellow in the 3D view and green in the 2D view). Notably, their binding affinities to both proteins were relatively low, suggesting that, despite their multitargeting properties, they may exhibit efficient interaction capabilities.

GROMACS was employed to perform molecular dynamics simulations, and to investigate the binding interactions of three small-molecule ligands—scutellarin (labeled in blue), salvianolic acid C (labeled in red), and isosalvianolic acid C (labeled in green)—with MMP9 (labeled in gray) and TLR4 (labeled in cyan) proteins individually. The simulation results demonstrated that these ligand-protein complexes maintained stable binding conformations within the protein binding pockets. Key parameters, including Root Mean Square Deviation (RMSD), Root Mean Square Fluctuation (RMSF), Radius of Gyration (GYRATE), and hydrogen bond analysis, were assessed to evaluate stability and interaction dynamics. RMSD analysis indicated that the protein-ligand complexes reached equilibrium and remained stable throughout the simulation. RMSF results showed moderate constraints on protein flexibility, contributing to changes in overall conformational compactness, which was further supported by GYRATE data reflecting structural alterations. Additionally, hydrogen bond analysis confirmed stable interactions between the ligands and proteins, reinforced by the presence of stable interaction energy minima (Figure 8). These findings provided valuable insights into the binding mechanisms of these compounds, and offered a foundation for further exploration of their biological functions.