The gut-brain connection may link low-dose chemical exposures to early puberty

Precocious puberty (PP), defined as the onset of secondary sexual characteristics before 8 years of age in girls and 9 years in boys (1), represents an emerging pediatric health concern with increasing global incidence. This condition is associated with multiple adverse outcomes, including reduced adult height, psychological challenges, metabolic disorders, and an elevated risk of hormone-sensitive cancers later in life (2, 3). Beyond individual health, PP imposes considerable socioeconomic burdens on healthcare systems and families (4).

Endocrine-disrupting chemicals (EDCs) constitute a broad class of environmental contaminants detected in over 95% of the general population (3, 5). These include plasticizers such as bisphenol A (BPA) and phthalates (6, 7), persistent organic pollutants like polychlorinated biphenyls (PCBs) and dichlorodiphenyltrichloroethane (DDT) (4, 8), industrial compounds including polybrominated diphenyl ethers (PBDEs) and per- and polyfluoroalkyl substances (PFAS) (5, 9), as well as heavy metals such as lead (Pb) and cadmium (Cd) (10, 11). Human exposure occurs predominantly through dietary intake, water, air, and consumer products (1, 12).

Substantial epidemiological evidence supports the association between EDC exposure and altered pubertal timing. For instance, phthalate metabolites are correlated with earlier thelarche (1, 5), and prenatal exposure to PFAS and PCBs has been linked to accelerated puberty (5). Mechanistically, EDCs are known to mimic or antagonize sex steroid actions, disrupt steroidogenic enzyme activity, and alter the pulsatile release of gonadotropin-releasing hormone (GnRH) from the hypothalamus (13, 14). Nevertheless, these established pathways do not fully explain the dose-response relationships or the effects of low-level mixture exposures, indicating significant gaps in mechanistic understanding (14, 15).

Recent evidence suggests that gut microbial composition is altered in individuals with PP, exhibiting reduced bacterial diversity, decreased abundance of Lactobacillus, and an increase in Bacteroides (16). Importantly, fecal microbiota transplantation (FMT) from donors with PP into germ-free mice recapitulates early pubertal onset, supporting a potential causal role for gut microbiota in regulating puberty timing (16, 17).

The gut-brain axis (GBA) comprises multiple interactive pathways facilitating crosstalk between the gastrointestinal tract and the central nervous system. These include neural connections such as the vagus nerve (18), endocrine signals involving gut hormones and interactions with the hypothalamic-pituitary-gonadal (HPG) axis (19, 20), immune mediators including microbiota-driven cytokines (21, 22), and microbial metabolites—such as short-chain fatty acids (SCFAs), bile acids, and tryptophan derivatives—that directly or indirectly influence GnRH and kisspeptin neurons (16–18).

Although independent associations have been established between EDC exposure and PP (5, 12), between gut microbiota changes and PP (16), and between EDCs and gut dysbiosis (17, 23),the causal sequence and relative contribution of this integrated pathway constituting a sequence of events from EDC-induced gut microbial alterations to GBA signaling disruption and ultimately to PP remain largely unexplored and inadequately characterized. Critical unresolved questions include: (1) whether EDC-induced dysbiosis directly impairs GBA function to trigger PP; (2) how the contribution of the GBA pathway compares to direct endocrine disruption or oxidative stress; and (3) whether synergistic interactions exist with metabolic conditions such as obesity (14, 24). Importantly, the potential dominance of GBA-mediated mechanisms at environmentally relevant low-dose EDC mixtures (≤1 μg/kg/day) remains unvalidated, challenging the current regulatory paradigm based primarily on high-dose single-chemical toxicity.

This work is structured as a systematic review with an exploratory, hypothesis-generating component. Its primary objective is to synthesize and critically evaluate existing evidence from published studies supporting the causal pathway from perinatal EDC exposure to gut dysbiosis, GBA disruption, and PP. In light of significant heterogeneity across studies that precluded a formal meta-analysis, we performed an exploratory Random Forest analysis on integrated semi-quantitative data. This analysis was not intended as original in silico research to establish new quantitative benchmarks, but rather as a secondary, hypothesis-generating tool to identify potential key mediators and patterns within the GBA pathway, and to estimate their relative model-internal importance based on the aggregated literature. Thus, while the core of this manuscript remains a systematic evidence synthesis, the integrated exploratory modeling aims to propose a novel mechanistic framework and inform future targeted investigations. A proposed mechanistic framework illustrating the integrative pathway from EDC exposure to premature HPG axis activation via GBA disruption is presented in Figure 1. This review therefore seeks to: (1) synthesize and evaluate evidence supporting the causal pathway from EDC exposure to gut dysbiosis, GBA disruption, and PP; (2) critically assess the relative contribution of this pathway under low-dose mixture exposures compared to canonical mechanisms such as direct HPG axis interference, oxidative stress, and epigenetic modifications (11, 25, 26); and (3) identify key research gaps and propose translational strategies for prevention and intervention (17, 27).

A comprehensive literature search was performed across four electronic databases: PubMed, Web of Science, Scopus, and Embase. The search encompassed studies published from January 1, 2010, to March 1, 2025, with a final update conducted on May 1, 2025. The search strategy was constructed by combining terms from three conceptual domains using Boolean operators:

1) EDCs: Key terms included “Endocrine Disruptors”[Mesh], “Environmental Pollutants”[Mesh], “Bisphenol A”, “phthalate”, “Di-(2-ethylhexyl) phthalate (DEHP)”, “PCB”, “polybrominated diphenyl ethers”, “PBDE”, “perfluoroalkyl substances”, “PFAS”, “heavy metals”, “lead”, and “cadmium”.

2) Gut Microbiota & GBA: Key terms included “Gastrointestinal Microbiome”[Mesh], “Dysbiosis”[Mesh], “gut microbio”, “Brain-Gut Axis”[Mesh], “vagus nerve”, “SCFAs”, “short chain fatty acid”, “bile acid”, “tryptophan”, “inflammation”, “intestinal barrier”, and “tight junction”.

3) Precocious Puberty & Puberty Timing: Key terms included “Puberty, Precocious”[Mesh], “precocious puber”, “early puber”, “GnRH”, “HPG axis”, and “kisspeptin”.

The complete, reproducible search syntax for all databases is provided in. 1

Study selection was guided by the PECOS (Population, Exposure, Comparator, Outcome, Study Design) framework: 1) Population: Human children/adolescents or experimental mammals; 2) Exposure: EDCs (individual compounds or mixtures), measured in environmental or biological matrices; 3) Comparator: Unexposed control groups or baseline microbiota profiles; 4) Outcome: Studies were required to report on at least two sequential components of the hypothesized EDC–microbiota–GBA–PP pathway; 5) Study Design: Observational studies, randomized controlled trials, animal experiments, and relevant in vitro mechanistic studies were eligible. Reviews, conference abstracts, and non-English publications were excluded.

Studies were also excluded if they were of low quality (Newcastle-Ottawa Scale [NOS] score <6 for human studies; SYRCLE risk of bias score <5 for animal studies) or lacked primary data.

The study selection process followed the PRISMA 2020 flow diagram (). After duplicate removal, titles and abstracts of the retrieved records were screened for relevance. To enhance efficiency, an artificial intelligence (AI)-assisted screening tool (ASReview version 1.0) was employed during the initial title/abstract screening phase to prioritize potentially relevant records. However, all final inclusion and exclusion decisions were independently verified by two human reviewers (H.W. and G.W.), with any discrepancies resolved through discussion or consultation with a third reviewer (C.H.). This approach ensured that AI-derived suggestions did not solely determine eligibility, thereby mitigating potential algorithmic bias in study selection. Subsequently, the full text of potentially eligible articles was assessed against the predefined eligibility criteria. To facilitate cross-species comparison, animal exposure doses were converted to Human Equivalent Doses (HEDs) using physiologically based pharmacokinetic (PBPK) modeling, following established EPA guidelines (EPA/600/R-23/238) and incorporating species-specific parameters detailed in. 1 1

Data were extracted using a standardized template, which captured the following information: (1) study metadata (design, population, exposure timing); (2) EDC details (compounds, doses, matrices); (3) microbiota metrics (α-diversity, taxonomic shifts); (4) GBA markers (SCFAs, cytokines, intestinal permeability); (5) PP outcomes (pubertal onset timing, Kiss1/GnRH expression); and (6) key covariates (e.g., BMI, diet).

The methodological quality of included studies was critically appraised using validated tools: the Newcastle-Ottawa Scale (NOS) for human studies, the SYRCLE risk of bias tool for animal studies, and a modified CREDIBILITY checklist for in vitro studies.

Given the significant heterogeneity in experimental designs, outcomes, and reporting formats across the included studies, a formal meta-analysis was deemed unfeasible. To enable an integrative analysis and identify key mediators within the GBA pathway, we instead performed a semi-quantitative synthesis and an exploratory Random Forest analysis.

The systematic literature search identified 4,582 records from databases and 32 records from manual searches. After duplicate removal, 3,129 unique records were screened by title and abstract. Of these, 412 full-text articles were assessed for eligibility, resulting in 87 studies that met all inclusion criteria for the final synthesis (Supplementary Figure 1). The included studies comprised human epidemiological cohorts (n=45), animal experiments (n=32), and in vitro studies (n=10). Key characteristics of representative studies are summarized in Table 1.

The synthesis of evidence from human and animal studies revealed a consistent pattern of gut microbial disruption following EDC exposure. A consistent reduction in microbial α-diversity was observed following perinatal EDC exposure, with a pooled estimate for Shannon index change of Δ = -1.8 (95% CI: -2.2 to -1.4; I² = 65%), derived from n = 28 studies (20 human, 8 animal) that reported diversity metrics (5, 17, 23). Meta-analysis of taxon-specific changes revealed a weighted mean decrease in Lactobacillus abundance of -42% (95% CI: -48% to -36%) in human studies (n=16) (5, 16) and -58% (95% CI: -65% to -51%) in rodent models (n=12), with moderate heterogeneity (I² = 55% and 60%, respectively) (23, 29), and a concurrent enrichment of Bacteroides (odds ratio [OR] = 1.68) (5).

Functionally, exposure to EDCs such as BPA, PFAS, and PCBs was linked to profound metabolic alterations. A marked suppression of butyrate production (30-50% reduction) and a decrease in secondary bile acid metabolism, including a 40% reduction in taurochenodeoxycholic acid (TCDCA; a secondary bile acid), were commonly observed (11, 16–18). Furthermore, a shift in tryptophan metabolism towards the kynurenine pathway was indicated by a 1.8-fold increase in the kynurenine-to-tryptophan ratio (11, 16–18). These taxonomic and functional shifts are visually summarized in Figure 2, which illustrates the conserved dysbiosis pattern across species.

Across 7 case-control studies, children with PP consistently showed a pooled reduction in Lactobacillus of -38% (95% CI: -45% to -31%; p < 0.001) and a 2.3-fold increase in Bacteroides (95% CI: 1.8–2.9; p < 0.001), with low between-study heterogeneity (I²= 30%) (5, 16). Animal studies provided compelling causal evidence. Fecal microbiota transplantation (FMT) from PP donors into germ-free mice resulted in accelerated pubertal onset, concomitant with increased hypothalamic Kiss1 mRNA expression (p < 0.001) (16). Similarly, antibiotic-induced dysbiosis advanced vaginal opening by 3.2 days (p < 0.01) in murine models (30). These findings suggest that the association between dysbiosis and PP is likely mediated by functional impairments, such as reduced SCFA production (16, 18), rather than changes in single microbial species alone.

The synthesis of mechanistic data revealed a multi-step pathway linking gut dysbiosis to neuroendocrine disruption:

1) Metabolite-Mediated Mechanisms: The reduction in butyrate levels compromised its function as an endogenous histone deacetylase (HDAC) inhibitor, particularly targeting Class I HDACs (e.g., HDAC1, HD2). This led to decreased histone acetylation at the promoter region of the Kiss1 gene. Consequently, the repressive chromatin state was alleviated, promoting Kiss1 transcription in kisspeptin neurons of the hypothalamic arcuate nucleus (18). Concurrently, the deficit in the secondary bile acid taurochenodeoxycholic acid (TCDCA) reduced agonism of the membrane bile acid receptor TGR5 (GPBAR1). This attenuated TGR5-mediated activation of the AMP-activated protein kinase (AMPK) signaling cascade in GnRH neurons. Since AMPK acts as an energy sensor and inhibits neuronal excitability, its dampened activity likely contributed to increased pulsatile firing of GnRH neurons, thereby facilitating premature GnRH release (16).

1) Barrier-Immune Axis Activation: Exposure to DEHP was quantified as an 80% increase in intestinal permeability, as measured by FITC-dextran translocation (17). Systemic lipopolysaccharide (LPS), upon reaching the brain, binds to Toll-like receptor 4 (TLR4) on microglia and potentially on GnRH neurons themselves. This binding initiates an intracellular signaling cascade involving MyD88, leading to the phosphorylation and degradation of IκBα, thereby releasing NF-κB p65 for nuclear translocation. Nuclear NF-κB subsequently transactivates pro-inflammatory genes (e.g., Il6, Tnf), creating a local inflammatory milieu that can directly potentiate the secretion of GnRH (22).

2) Neural Signaling: Vagotomy experiments confirmed the critical role of the vagus nerve, as it blocked the effects of microbial metabolites on GnRH neuronal activity (p < 0.01) (18).

A temporal sequence was established in animal models, where EDC exposure initiated microbiota shifts within 1–2 weeks, followed by metabolic and inflammatory alterations at 2–4 weeks, culminating in HPG axis activation and pubertal onset after 4–8 weeks (17).

The exploratory Random Forest analysis, employed to identify key mediators within the GBA pathway, identified butyrate depletion (Mean Decrease Gini MDGini, MDGini = 12.3) and IL-6 elevation (MDGini = 9.7) as the most prominent features associated with PP risk in the model (Figure 3A). The model demonstrated robust predictive performance, with 10-fold cross-validation yielding R² = 0.73 and root mean square error (RMSE) = 0.18 (Figure 3B), indicating good internal consistency within the integrated dataset. To explore potential key drivers, we applied an exploratory Random Forest model. This analysis, constrained by our integrated dataset, identified that mediators within the GBA pathway collectively exhibited the highest relative importance, accounting for approximately 68% of the model’s internal variance explanation for PP risk at the examined low-dose range (≤1 μg/kg/day) (17). This suggests the GBA pathway may be a predominant feature in the model’s prediction under these conditions. It is crucial to note, however, that this estimate is derived from studies that, despite being categorized as “low-dose” in our analysis, often utilized doses higher than typical human environmental exposures (e.g., BPA at 5-50 μg/kg/day in rodents vs. human urinary levels around 3.2 μg/L). This model-derived estimate suggests the GBA pathway may be highly influential at low doses, but it does not establish a definitive causal proportion, and its primacy at truly human-relevant environmental doses requires direct experimental confirmation. Conversely, the model pattern was consistent with a shift towards direct HPG axis interference as the dominant mechanism in high-dose scenarios (>5 mg/kg/day) within the studied experiments (17). Furthermore, the correlation matrix among the top GBA mediators (Figure 3C) revealed significant interrelationships within the integrated dataset. Notably, a strong negative correlation was observed between butyrate levels and intestinal permeability (r=−0.55), while butyrate exhibited a strong positive correlation with total SCFAs (r=0.85). These associations underscore the interconnected nature of microbial metabolic output, gut barrier integrity, and systemic inflammatory tone within the GBA pathway, supporting the hypothesis that coordinated multi-factor disruptions contribute to EDC-induced precocious puberty. The exploratory Random Forest analysis suggested a dose-dependent shift in mechanistic dominance. At low doses (≤1 μg/kg/day), GBA-related mediators collectively accounted for approximately 68% of the model’s internal variance explanation for PP risk, whereas direct HPG axis interference became dominant at higher doses (>5 mg/kg/day). These dose-dependent contributions are illustrated in Figure 4. Key experimental findings supporting this analysis are summarized in Table 2.

The evidence synthesis supports an integrated network wherein EDC-induced GBA disruption interacts synergistically with canonical pathways, including direct endocrine disruption, oxidative stress, epigenetic reprogramming, and metabolic dysregulation. Key crosstalk mechanisms include an inflammation-ROS cycle that compromises gut barrier integrity, a microbiota-epigenetics-metabolism axis linking SCFA depletion to leptin dysregulation, and transgenerational epigenetic feedforward loops exacerbated by microbial metabolite deficits (16, 18, 21, 22, 27).

In addition to GBA disruption, evidence from the included studies supports several complementary mechanisms through which EDCs may influence pubertal timing.

By synthesizing evidence from 87 studies, this review addresses a key knowledge gap and supports a potential role for GBA disruption in the pathogenesis of EDC-induced PP. Our findings indicate that perinatal exposure to environmentally relevant low-dose EDC mixtures induces conserved gut dysbiosis, characterized by reduced microbial diversity (Shannon Δ = −1.8), decreased Lactobacillus (−40%), and increased Bacteroides (1.5-fold). These alterations are associated with impaired production of microbial metabolites, including butyrate (↓50%) and bile acids, increased intestinal permeability (FITC-dextran ↑80%), systemic inflammation (IL-6 ↑1.8-fold), and epigenetic dysregulation of hypothalamic Kiss1/GnRH neurons (5, 16–18, 21, 22). Crucially, fecal microbiota transplantation from PP donors into germ-free mice recapitulated early pubertal onset, providing experimental support for a potential causal role of gut microbiota (16).

Complementary mechanisms identified include direct HPG axis disruption (14, 35), oxidative stress (10, 11), epigenetic reprogramming (10, 11, 38–40), and obesity-leptin dysregulation (5, 38). Variations in the extent of Lactobacillus depletion between rodents (−60%) (23) and humans (−40%) (5, 16) may reflect species-specific susceptibilities or differences in EDC potencies.

Based on the synthesized evidence, we propose an integrated model wherein EDC-induced GBA disruption may interact synergistically with canonical pathways (Figure 5). Several key crosstalk mechanisms emerge from the literature:

GBA–Inflammation–Oxidative Stress Axis: EDC-induced dysbiosis is associated with elevated intestinal permeability, which may facilitate LPS translocation and systemic inflammation (IL-6↑1.8×, TNF-α↑) (21, 22). This inflammatory milieu could, in turn, activate hypothalamic NF-κB, potentially sensitizing GnRH neurons to EDC receptor agonism (14, 22). Beyond transcriptional regulation, post-translational modifications directly impair barrier integrity. ROS-induced carbonylation of occludin disrupts its interaction with ZO-1 proteins, leading to the disassembly of tight junctions and increased paracellular permeability. This creates a self-amplifying cycle wherein barrier breach perpetuates inflammation and oxidative stress (21, 39).

Microbiota–Epigenetics–Metabolism Axis: The interplay between microbial metabolites and EDC-driven epigenetics is critical. EDCs like BPA can induce DNA hypomethylation at the Kiss1 promoter. Butyrate, through HDAC inhibition, creates a permissive chromatin state (e.g., hyperacetylated histones) that synergistically enhances the transcriptional activity of the already hypomethylated Kiss1 gene. This suggests a convergent epigenetic mechanism whereby EDCs prime the Kiss1 gene for dysbiosis-mediated activation (18, 40). Reduced SCFAs suppress GLP-1 secretion (↓40%), promoting obesity and hyperleptinemia (OR = 3.21), which may directly stimulate kisspeptin neurons (22, 27). EDC-obesity synergy appears to occur via PPARγ/RXR activation and Bacteroides-enriched dysbiosis (27, 40), consistent with the obesogenic properties of EDCs demonstrated in epidemiological studies (5, 14).

Transgenerational Feedforward Loops: Prenatal EDC exposure induces epigenetic alterations (e.g., Kiss1 hypomethylation (34)) that may be exacerbated by deficits in microbial metabolites such as butyrate (18). Microbial tryptophan metabolites (e.g., indole-3-propionic acid) modulate DNMT1 activity (16), potentially linking dysbiosis to transgenerational HPG-axis dysregulation.

Potential Synergistic Amplification: The calculated synergy indices (SI > 2.3) for some mixture exposures suggest the occurrence of supra-additive effects, possibly through the convergence of multiple pathways. Co-exposure to BPA (ERα agonist) and LPS amplified GnRH secretion by 160% compared to individual exposures (16), suggesting inflammatory potentiation of endocrine disruption. DEHP and cadmium synergistically increased hypothalamic ROS (300%) and gut permeability (↑120%) (11, 21). Epidemiological synergy indices (SI) for PP risk exceeded 2.3 under combined exposures (e.g., BPA+PFAS+phthalates), compounded by obesity (leptin OR = 3.21) and inflammation (5, 38).

Collectively, the integrated model reveals that EDCs orchestrate PP through a concerted multi-hit mechanism targeting key cellular nodes:

The crosstalk between these nodes, such as NF-κB regulating both inflammatory genes and epigenetic enzymes, creates a robust, self-reinforcing network that drives the premature activation of the HPG axis.

Pattern from Exploratory Modeling and Its Interpretation: The exploratory Random Forest analysis, applied to our integrated dataset, yielded a pattern wherein GBA-related mediators collectively showed high relative importance, explaining approximately 68% of the model’s internal variance for PP risk at the studied low-dose range (≤1 μg/kg/day in animal models). It is essential to explicitly clarify that the term “low-dose” in this context refers to doses used in experimental toxicology, which are often orders of magnitude higher than typical human environmental exposures. For example, the rodent BPA dose of 5–50 μg/kg/day considered “low” in our analysis corresponds to exposure levels substantially exceeding the 95th percentile of human urinary BPA (≈3.2 μg/L) (41). Therefore, while the model suggests that GBA disruption may be a prominent feature under these experimental conditions, this finding cannot be directly extrapolated to conclude that the GBA pathway predominates at truly human-relevant environmental exposure levels (e.g., ≤1 μg/kg/day in humans). The model-derived estimate should be interpreted strictly as a hypothesis-generating indicator that warrants validation in studies specifically designed with environmentally realistic, low-dose mixture exposures (15, 17). Key mediators included butyrate depletion (MDGini = 12.3) and IL-6 elevation (MDGini = 9.7). Notably, these effects occurred below established no-observed-adverse-effect levels (NOAELs)—for instance, dysbiosis was observed at BPA doses as low as 5 μg/kg/day (23). In contrast, direct endocrine disruption appeared dominant at higher doses (>5 mg/kg/day), accounting for only 32% of the effect (17). At sub-environmental exposures (<0.1 μg/kg/day), epigenetic mechanisms contributed approximately 41% (36–46%) of PP risk (34, 36).

Critical Knowledge Gaps: Human evidence remains largely indirect, with mechanistic scores ranging from 2–4 compared to 6 in animal studies, reflecting a lack of longitudinal data on GBA biomarkers (5, 15). Moreover, despite real-world coexposures (e.g., BPA and PFAS), only 3.4% of included studies quantified mixture effects, with synergy indices (SI) exceeding 2.3 indicating significant supra-additivity (23, 28).

A fundamental limitation of the current evidence base, which directly challenges the translational relevance of our findings, is the stark underrepresentation of studies examining realistic EDC mixtures. As noted in our synthesis, only 3.4% (3/87) of the included studies investigated chemical mixtures, despite the unequivocal consensus that human exposure occurs to complex, low-dose mixtures of EDCs, not to single chemicals in isolation (15, 28). This overwhelming focus on single-chemical paradigms means that the synergistic, additive, or antagonistic interactions between EDCs—which could critically alter dose-response relationships, mechanistic dominance, and overall risk—remain largely uncharacterized within the context of the GBA-PP pathway. For instance, while our analysis identified potential synergistic interactions (SI > 2.3), these were derived from a minuscule subset of data. The vast majority of the mechanistic insights on GBA disruption (e.g., BPA-induced dysbiosis, DEHP-increased permeability) come from single-chemical experiments, leaving a critical unanswered question: Do these pathways operate similarly, or are they potentiated, when exposed to the “chemical soup” representative of real-world environments? This gap limits our ability to accurately model human risk. It also underscores a misalignment between current toxicological research paradigms and the reality of human environmental exposure.

Dose-Translation Discrepancy and Human Relevance: A critical limitation that challenges the translational relevance of our synthesized evidence is the prevalent use of animal doses that far exceed typical human environmental exposures. For instance, rodent studies often employ BPA doses of 5–50 μg/kg/day (17), which are not environmentally relevant when compared to median human biomonitoring data (urinary BPA ≈ 1–3 μg/L) (12, 41). Similarly, DEHP doses of 5 mg/kg/day in rats translate to human equivalent doses vastly higher than background serum metabolite concentrations in children (1). Consequently, the mechanistic insights derived from these high-dose experimental studies—including the proposed prominence of the GBA pathway—cannot be directly translated to real-world human exposure scenarios. The assertion that GBA disruption is the dominant mechanism at human-relevant low doses remains a hypothesis that requires rigorous testing in models employing environmentally realistic exposure regimens (15, 17). To definitively establish the role of the GBA at human-relevant exposures, future research must prioritize studies employing environmentally relevant, low-dose EDC mixtures (ΣEDCs ≤1 μg/kg/day) in models with human physiological relevance. A particularly promising approach is the use of humanized-microbiota rodent models, where animals are colonized with human gut microbiota. This would allow for a more direct investigation of the EDC-microbiota-GBA-PP pathway under exposure scenarios that genuinely mirror the human condition, thereby bridging the current translational gap.

Limitations of the Exploratory Machine Learning Analysis: While the Random Forest analysis served its primary goal of hypothesis generation and mediator prioritization, several key limitations must be acknowledged. First, the model was built upon a semi-quantitatively integrated dataset not originally designed for such analysis, introducing potential biases from heterogeneous study designs and reporting formats. Most critically, the model lacked external validation, meaning its performance and the specific quantitative estimates (e.g., the 68% pathway contribution) are inherently internal to our dataset and may not generalize. The good cross-validation metrics (R² = 0.73) indicate internal consistency but do not mitigate the need for external validation. Therefore, the numerical estimates should be interpreted as illustrative of a compelling pattern that merits targeted investigation, not as definitive measures of causal attribution. Future studies with standardized, pre-specified outcomes are essential for building quantitatively reliable and generalizable models. Although the Random Forest model showed reasonable internal performance (cross-validation R² = 0.73, RMSE = 0.18), these metrics should be interpreted with caution given the heterogeneous nature of the integrated dataset.

The Predominance of Single-Chemical Paradigms: A paramount translational limitation is that over 90% of the included mechanistic studies investigated single EDCs in isolation. This stands in stark contrast to the reality of human exposure, which is to complex, lifelong mixtures of chemicals (e.g., BPA, PFAS, phthalates, and metals) (15). Consequently, the field critically lacks data on how interactions within such mixtures might alter the GBA disruption pathway, potentially through synergistic or additive effects that could lower the effective dose of concern. No study to date has comprehensively assessed mixture effects on the microbiota–SCFA–inflammation axis. Furthermore, experimental doses [e.g., PFAS at 100 μg/kg/day (5)] vastly surpass median human serum levels (0.5–2 ng/mL) (5), limiting translational relevance.

Translational Bottlenecks: In vitro concentrations (e.g., DEHP at 10 μM ≈ 3,900 ng/mL (22)) often lack physiological relevance to human tissue accumulation. Heterogeneous exposure units (e.g., oral μg/kg/day vs. inhalation μg/m³) further complicate cross-study comparisons and meta-analyses (41).

Regulatory Implications: Our findings highlight the potential value of exploring microbiota-derived biomarkers and mixture interactions in future risk assessments, consistent with recent recommendations for mixture toxicity evaluation (42). Feasible GBA biomarkers include fecal butyrate (<2.5 μmol/g), the Bacteroides/Lactobacillus ratio (>1.5), and plasma IL-6 (>5 pg/mL) (5, 16, 18, 21). A tiered testing strategy—incorporating high-throughput screening (e.g., with colonic organoids), in vivo assessment in models like zebrafish or humanized-microbiota rodents, and low-dose mixture exposures—could be evaluated in future research to better assess GBA-mediated toxicity (16–18, 21, 23).

Intervention Strategies: Microbiota-targeted therapies, such as supplementation with Lactobacillus and inulin to restore butyrate levels, represent a promising area for future investigation (16, 18). Adjunctive agents such as curcumin [NF-κB inhibition (22)] or N-acetylcysteine [ROS scavenging (11)] may mitigate inflammatory and oxidative components of GBA disruption.

Research Priorities: Future studies should focus on: (1) investigating Real-World Mixtures: Priority must be given to studies that move beyond single-chemical exposures. Future experimental models, particularly humanized-microbiota rodents, should be exposed to defined, low-dose mixtures of EDCs (e.g., ΣEDCs ≤1 μg/kg/day) that reflect common human co-exposure profiles (e.g., BPA, DEHP, PFOS, and lead). This is essential to determine whether the GBA disruption pathway observed in single-chemical studies is amplified, attenuated, or qualitatively different under mixture conditions; (2) integrating longitudinal EDC measurements with multi-omics profiling in prospective cohorts; (3) evaluating Lactobacillus and butyrate-producing probiotics in high-exposure pediatric subgroups; and (4) applying machine learning to identify critical susceptibility windows and gene–environment interactions (5, 12, 14, 16, 23, 24), building on computational toxicology approaches (28, 31).