Deciphering pathogenic mechanisms of BDCPP exposure in endometrial cancer progression via an integrated approach combining network toxicology, machine learning, and molecular docking

Abstract

Background:

Endometrial cancer (EC) is one of the most common gynecological malignancies globally. Increasing attention has been paid to the role of environmental pollutants in EC development. Epidemiological studies have demonstrated a significant association between elevated urinary concentrations of bis(1,3-dichloro-2-propyl) phosphate (BDCPP) and an increased risk of EC. However, the hub genes and underlying mechanisms of BDCPP-induced EC remain poorly understood.

Materials and methods:

Potential targets of BDCPP and EC were retrieved from multiple databases. A protein–protein interaction network was constructed based on the common targets. Enrichment analysis was performed using Gene Ontology, Kyoto Encyclopedia of Genes and Genomes, and Reactome databases. Clinical and transcriptomic data of EC patients were downloaded from The Cancer Genome Atlas Program. Subsequently, 117 machine learning models were employed to screen for hub genes. A risk score for BDCPP exposure was calculated based on the hub genes. Single-gene gene set enrichment analysis (GSEA) was conducted on the hub genes, and molecular docking was performed to predict the binding affinity of BDCPP to the hub genes.

Results:

We identified 165 potential targets implicated in BDCPP-induced EC. Machine learning pinpointed eight hub genes: PLA2G2A, PLAU, SIRT2, DRD2, GSK3A, THRB, CYP17A1, and TLR9. The BDCPP exposure risk score model offers a framework for predicting the prognosis of EC patients with moderate accuracy. Molecular docking revealed the binding potential between BDCPP and hub genes. Our findings highlight the pivotal roles of inflammatory activation, hormonal disruption, altered lipid metabolism, and epigenetic dysregulation in pathogenic mechanisms. Single-gene GSEA further emphasized the critical roles of nucleocytoplasmic transport, polycomb repressive complex, and mRNA surveillance pathway in this process.

Conclusion:

Our study investigated the hub genes and underlying mechanisms of BDCPP-induced EC. The findings not only offer novel insights into the role of environmental pollutants in EC development but also present an analytical framework for elucidating the carcinogenic mechanisms of other environmental chemicals.

HIGHLIGHTS

• 165 genes were identified as potential targets for bis(1,3-dichloro-2-propyl) phosphate (BDCPP)-induced endometrial cancer (EC).

• Eight hub genes were identified through 117 machine learning models.

• Potential mechanisms in BDCPP-induced EC were uncovered.

• The BDCPP exposure risk score can serve as a risk factor for EC.

• Molecular docking predicted the binding affinity of BDCPP to hub genes.

Introduction

Endometrial cancer (EC) ranks among the most prevalent gynecological malignancies worldwide, posing a significant public health burden. By 2020, EC emerged as the sixth leading cancer in women globally, with approximately 417 000 new cases diagnosed annually and an age-standardized incidence rate (ASR) of 8.7 per 100 000 women^[1]. Early detection of EC is associated with markedly improved clinical outcomes, whereas patients diagnosed at advanced stages exhibit substantially lower 5-year survival rates – ranging from 47% to 58% for stage III and 15% to 17% for stage IV disease – underscoring the critical urgency of prioritizing early prevention and intervention strategies^[2]. While traditional risk factors, including obesity, diabetes, and hormonal imbalances, partially elucidate its pathogenesis, the potential role of environmental pollutants in EC development has garnered increasing attention^[3]. Notably, endocrine-disrupting chemicals (EDCs), such as bisphenol A and phthalates, have been implicated in gynecological cancers through mechanisms involving estrogenic signaling disruption, oxidative stress, and epigenetic modifications^[4–6]. However, studies on organophosphate flame retardants (OPFRs), a class of emerging EDCs, in relation to gynecological cancers remain scarce, despite their widespread use, potential for bioaccumulation, and multiple identified toxic effects (including reproductive and developmental toxicity, inhibition of cell activity, and oxidative damage)^[7]. Additionally, OPFRs have been nominated by the U.S. Consumer Product Safety Commission as a priority research focus for the National Toxicology Program/National Institute of Environmental Health Sciences^[8], further underscoring the importance of conducting OPFR-related research on human health.

Tris(1,3-dichloropropyl) phosphate (TDCPP), one of the most prevalent OPFRs, is extensively incorporated into automotive components, residential furniture, and polyurethane foam-containing products to comply with flammability regulations^[9]. During routine use, TDCPP is readily released into indoor environments, accumulating in dust^[10]. Dust samples collected from offices, residential settings, and vehicles exhibit TDCPP detection rates exceeding 96%, with concentrations spanning <0.03 to 326 μg/g^[11]. Animal studies demonstrate that TDCPP is efficiently absorbed via both dermal and gastrointestinal routes^[12]. Upon human exposure to TDCPP, bis(1,3-dichloro-2-propyl) phosphate (BDCPP), its predominant metabolite, exhibits ubiquitous internal exposure, as evidenced by consistent detection in human urine through biomonitoring studies^[13]. A study on municipal wastewater treatment plants across China revealed that the mean concentration of BDCPP in influent was 22 ng/L (detection rate: 40%), while that of TDCPP was 31 ng/L (detection rate: 100%)^[14]. Their concentrations in effluent were 25 ng/L and 26 ng/L, respectively. Another study based on National Health and Nutrition Examination Survey reported a median BDCPP concentration of 1.11 ng/mL in urine^[15]. Given the bioaccumulation potential of OPFRs, it is reasonable that BDCPP concentrations in organisms are several dozen times higher than environmental levels, underscoring the importance of investigating the potential health risks posed by this persistent pollutant. In a case–control study, researchers measured the urinary levels of 10 OPFRs in EC patients and healthy individuals, revealing significantly higher concentrations of BDCPP, di-(2-butoxyethyl) phosphate, and tris(2-butoxyethyl) phosphate in EC patients^[16]. Multivariate logistic regression analysis indicated that among the 10 OPFRs, only BDCPP was an independent factor associated with EC, yet the mechanistic pathways underlying its involvement in EC progression remain incompletely elucidated.

Traditional toxicological studies often rely on single-model approaches, such as in vitro assays or animal experiments, which are limited by scalability, species-specific biases, and an inability to capture systemic molecular interactions. In contrast, the integration of network toxicology, machine learning, and molecular docking represents a novel methodological advancement. Network toxicology enables the systematic mapping of chemical-target-disease interactions, which would be intractable with conventional experiments^[17]. Machine learning algorithms are intelligent computational approaches that construct data-driven models for pattern recognition, prediction, and decision-making tasks^[18]. When trained on clinical and transcriptomic data, machine learning can identify hub genes involved in tumorigenesis from the shared target genes between chemicals and diseases. Moreover, molecular docking serves as a predictive method for assessing interactions between chemicals and hub genes. Recent studies in medicine and environmental health have demonstrated the effectiveness of similar integrated approaches. For instance, Li et al employed this integrated method to explore hub genes of prostate cancer induced by air pollution exposure and constructed a prognostic model^[19].

Building on the methodological advancements outlined previously, this study utilized an integrated approach combining network toxicology, machine learning, and molecular docking to elucidate the molecular mechanisms underlying BDCPP-induced EC (Fig. 1). Briefly, by constructing a BDCPP–target–EC interaction network, we systematically identified potential oncogenic hub genes and pathways. Subsequently, clinical and transcriptomic datasets were integrated to pinpoint hub genes linking BDCPP exposure to EC progression. Molecular docking simulations further predicted interactions between BDCPP and hub genes. The findings highlight the requirement for health risk assessment of BDCPP and offer a scientific basis for cancer prevention strategies and targeted therapeutic development. Furthermore, the analytical framework developed herein provides a methodological reference for deciphering carcinogenic mechanisms of other environmental chemicals, advancing environmental health research toward precision and systematic integration. In compliance with the TITAN Guidelines 2025, this research was carried out without the utilization of artificial intelligence tools^[20].

Figure 1.

Workflow outlining the network toxicology, machine learning, and molecular docking analyses conducted to investigate BDCPP-induced EC. BDCPP, bis(1,3-dichloro-2-propyl) phosphate; EC, endometrial cancer; GSEA, gene set enrichment analysis.

Materials and methods

Collection of BDCPP and EC related targets

Human target genes potentially affected by BDCPP were retrieved from the CTD database (https://ctdbase.org/), SuperPred database (https://prediction.charite.de/), and TargetNet database (http://targetnet.scbdd.com/). EC-related genes were collected from the GeneCards database (https://www.genecards.org/), DisGeNET database (https://disgenet.com/), and CTD database (https://ctdbase.org/). To minimize false-positive targets, we selected appropriate thresholds based on the characteristics of each database during target screening. For BDCPP targets, the filtering criteria were set as follows: literature evidence count ≥2 in the CTD database^[17], Probability ≥ median in the SuperPred database^[21], and Probability ≥0.6 in TargetNet^[22]. For EC-related targets, the median was used as the filtering threshold in both GeneCards and DisGeNET^[23], while an inference score ≥30 was applied in the CTD database^[24]. To ensure the biological relevance and comprehensiveness of our findings, all identified targets were cross-referenced and standardized using the UniProt database (https://www.uniprot.org/). After integrating and removing duplicates from the standardized predictions across multiple databases, a definitive target library was established, comprising 265 BDCPP-related targets and 7432 EC-related targets.

Construction of PPI network

Protein–protein interaction (PPI) analysis was performed on the overlapping genes between BDCPP and EC targets using the STRING database (https://cn.string-db.org/), with parameters specifically configured for Homo sapiens. Interactions with a “Medium Confidence” threshold (interaction score >0.4) were considered^[17,25,26]. The STRING-derived PPI network data were subsequently imported into Cytoscape (version 3.10.1) to assess the topological properties of the network.

Functional enrichment analysis for potential targets

Functional enrichment assessments of potential target genes were carried out utilizing the clusterProfiler R package to investigate their annotation within the Gene Ontology (GO) framework, encompassing biological processes (BPs), molecular functions (MFs), and cellular components (CCs), as well as their involvement in Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways^[27]. Furthermore, pathway analysis based on the Reactome database was performed using the ReactomePA package^[28]. For enrichment analysis, the hypergeometric test was employed. The background gene set was defined as all genes annotated in the organism-specific database. We utilized the default parameters, setting minGSSize to 10 and maxGSSize to 500. Terms or pathways were considered statistically significantly enriched when the adjusted P-value, corrected for multiple testing using the Benjamini–Hochberg method, was less than 0.05.

Screening hub genes by machine learning

To identify hub genes for the diagnosis of EC induced by BDCPP, we downloaded transcriptomic and clinical data of EC patients from The Cancer Genome Atlas Program (https://portal.gdc.cancer.gov/). Only samples categorized as tumor were selected for subsequent analysis. Subsequently, cases lacking overall survival (OS) information were excluded, resulting in a dataset of 529 EC patients (Supplemental Digital Content Table S1, available at, http://links.lww.com/JS9/F193). The data were then partitioned into training and testing sets at a 7:3 ratio, ensuring consistent survival-to-death ratios in both sets. In this study, the construction of 117 models was achieved by integrating combinations of classical machine learning algorithms (e.g., random survival forest (RSF), elastic net (Enet), CoxBoost) implemented in the Mime package with the function ML.Dev.Prog.Sig (mode set as “all,” others as default) (refer to Liu et al for a detailed introduction^[29]). Briefly, we systematically explored various parameter settings and algorithm combinations, employing K-fold cross-validation to optimize model hyperparameters, rather than relying solely on simple grid or random search. This approach enabled us to evaluate the impact of different hyperparameter combinations on model performance during training, thereby identifying the optimal model configuration. Additionally, for certain algorithms (e.g., Lasso, StepCox), we incorporated specific variable selection filters to further refine the models.

Construction of a BDCPP exposure risk model for EC patients

Initially, we performed multivariate Cox regression analysis on the identified hub genes to quantify their prognostic contributions. A risk score for each patient was subsequently calculated by integrating the regression coefficients and expression levels of these hub genes. Using the survminer package, we determined the optimal cutoff value to stratify patients into high- and low-risk groups^[30]. Kaplan–Meier (KM) survival curves and log-rank tests were applied to compare OS between the two risk groups, thereby assessing the prognostic value of the risk score. A nomogram was constructed based on multivariate Cox regression analysis using the R packages survival and rms. Model calibration was evaluated through calibration curves to assess agreement between predicted and observed survival probabilities. Additionally, multivariate Cox regression analysis was conducted to test the independence of the risk score as a prognostic factor for EC patients. Furthermore, the predictive performance of the risk model was evaluated using receiver operating characteristic (ROC) curve analysis, with the area under the curve (AUC) serving as a quantitative measure of its discriminatory ability.

Single-gene GSEA for hub genes

To elucidate the pathways associated with the identified hub genes, we performed single-gene gene set enrichment analysis (GSEA), a widely adopted method for evaluating the functional enrichment of individual genes within a dataset. Initially, gene sets were generated for each hub gene by ranking genes based on their Spearman correlation coefficients with the respective hub gene. Subsequently, KEGG GSEA was conducted on these gene sets using the clusterProfiler package. Pathways with adjusted P-values <0.05 (following correction by the Benjamini–Hochberg method) were considered significantly enriched.

Molecular docking

To elucidate the molecular mechanisms underlying the interactions between BDCPP and the hub genes identified in this study, we employed molecular docking methodology. The three-dimensional structures of the proteins of hub genes were predicted using AlphaFold 3^[31]. Molecular docking was performed using the CB-Dock2 web server (https://cadd.labshare.cn/cb-dock2/php/index.php) to assess the binding affinity between BDCPP (ligand) and hub proteins (receptors) with default parameters (specifically, no manual adjustments were made to the grid center position or grid size; CB-Dock2 algorithm automatically identified potential pockets on the protein surface and generated docking boxes). Key protein–ligand interactions were identified through docking analysis and visualized in two dimensions using Discovery Studio.

Results

Identification and PPI analysis of potential targets

We compiled a total of 265 BDCPP-related targets from the CTD, SuperPred, and TargetNet databases. Additionally, 7432 EC-associated targets were retrieved from the GeneCards, DisGeNET, and CTD databases. By intersecting these two datasets, we identified 165 potential targets implicated in BDCPP-induced EC (Fig. 2A). Subsequently, a PPI network was constructed for these 165 shared targets using the STRING database. Following network analysis and visualization with Cytoscape, proteins without any interactions were excluded, revealing a network comprising 162 nodes and 1311 edges (Fig. 2B). In this PPI network, the average number of neighbors is 16.185, the clustering coefficient is 0.457, the network density is 0.101, and the network centralization is 0.351.

Figure 2.

(A) Venn diagram illustrating the intersection between BDCPP targets and EC targets. (B) PPI network of the intersecting targets. Different colors represent the magnitude of degree, with a gradient from blue to red indicating an increase in degree, while the size of nodes also increases accordingly. Edges between nodes are depicted with gray lines. In the overall layout, nodes located in the inner circles indicate higher degrees. BDCPP, bis(1,3-dichloro-2-propyl) phosphate; EC, endometrial cancer; PPI, protein–protein interaction.

Functional enrichment analysis of all potential targets

Subsequently, we performed enrichment analysis on all potential targets using the KEGG, Reactome, and GO databases, with results sorted in ascending order by adjusted p-value (Supplemental Digital Content Tables S2–S4, available at, http://links.lww.com/JS9/F193). KEGG enrichment analysis revealed that all potential targets were involved in multiple biological categories, including “Metabolism,” “Environmental Information Processing,” “Organismal Systems,” and “Cellular Processes.” The top three pathways ranked by adjusted P-value were “Apoptosis,” “Neutrophil extracellular trap formation,” and “Arachidonic acid metabolism” (Fig. 3A, Supplemental Digital Content Table S2, available at, http://links.lww.com/JS9/F193). For Reactome enrichment analysis, the top three pathways were “Biosynthesis of DHA-derived SPMs (specialized proresolving mediators),” “Biosynthesis of specialized proresolving mediators (SPMs),” and “Diseases of signal transduction by growth factor receptors and second messengers” (Fig. 3B, Supplemental Digital Content Table S3, available at, http://links.lww.com/JS9/F193). GO enrichment analysis demonstrated that in the GO BP category, the top three pathways were “response to molecule of bacterial origin,” “response to lipopolysaccharide,” and “response to xenobiotic stimulus” (Fig. 3C, Supplemental Digital Content Table S4, available at, http://links.lww.com/JS9/F193). In the GO MF category, they were “histone deacetylase activity,” “protein lysine deacetylase activity,” and “heme binding.” In the GO CC category, the top three terms were “ficolin-1-rich granule,” “membrane raft,” and “membrane microdomain.”

Figure 3.

(A, B) KEGG/Reactome enrichment analysis (top 15 results) of potential targets. The x-axis represents the fold enrichment, while the y-axis displays the KEGG/Reactome descriptions with respective categories. The numbers within the circles indicate the gene count corresponding to each description. The text symbols on the right side represent the adjusted P-values, where asterisks denote statistical significance levels as follows: *: 0.01–0.05; **: 0.001–0.01; ***: <0.001. (C) GO enrichment analysis (top 10 results for each ontology). The x-axis indicates the fold enrichment, and the y-axis lists the GO descriptions. Different colors represent various ontologies, with the intensity of the color reflecting the magnitude of the adjusted P-values. The size of each dot corresponds to the gene count associated with the respective GO description. BP, biological process; CC, cellular component; GO, Gene Ontology; HIF-1, Hypoxia-inducible factor 1; KEGG, Kyoto Encyclopedia of Genes and Genomes; MF, molecular function; SPMs, specialized proresolving mediators; VEGF, vascular endothelial growth factor.

Identification of hub genes and construction of BDCPP exposure risk model

Through systematic evaluation of 117 machine learning models, we identified the RSF + GBM model as the top performer, achieving a C-index of 0.75 on the test set (the highest among all models) (Fig. 4A). This model comprised eight genes (PLA2G2A, PLAU, SIRT2, DRD2, GSK3A, THRB, CYP17A1, TLR9), which were also designated as hub genes. For the four models demonstrating higher C-index values (>0.7) in the test set, we calculated feature importance and visualized the top 15 genes ranked by importance for each model (Supplemental Digital Content Figure S1, available at, http://links.lww.com/JS9/F193). The results indicated that the importance of the eight hub genes remained relatively stable across these models. Using the risk score calculated from these eight genes, EC patients were stratified into low-risk and high-risk groups (Fig. 4B). KM survival analysis demonstrated significantly poorer OS in the high-risk group compared to the low-risk group P < 0.0001) (Fig. 4C). A nomogram was constructed to predict 1-, 3-, and 5-year OS based on these eight genes (Fig. 4D). Among the calibration curves for 1-, 3-, and 5-year survival, the 1-year curve closely approximates the ideal diagonal line, indicating relatively high predictive accuracy; however, the 3- and 5-year curves demonstrate a risk of overestimating probabilities (Fig. 4E). The variance inflation factors for age, clinical stage, and risk group were 1.03, 1.027, and 1.08, respectively, all closely approaching 1 and indicating no multicollinearity. The results of proportional hazards (PH) assumption checks are presented in Supplemental Digital Content Figure S2, available at, http://links.lww.com/JS9/F193, demonstrating that both the overall model and individual variables satisfy the PH assumption. Multivariate Cox regression analysis, adjusted for age, stage, and risk group, demonstrated the potential of the risk score as a prognostic factor for EC patients (Fig. 4F). Time-dependent ROC curves for the risk score model yielded AUC values of 0.655, 0.673, and 0.745 for 1-year, 3-year, and 5-year survival predictions, respectively, suggesting moderate predictive accuracy (Fig. 4G). In summary, the BDCPP exposure risk score model, constructed using eight hub genes identified by machine learning, provides a framework for predicting the prognosis of EC patients with moderate accuracy.

Figure 4.

(A) The C-index values derived from 117 machine learning models were visualized using a heatmap, with model names displayed along the y-axis. Models were ordered based on their C-index performance in the testing set. (B) Distribution plot of BDCPP exposure risk scores and survival time. (C) Kaplan–Meier survival curves with risk table for EC patients stratified into high-risk and low-risk groups based on the risk score. (D) Nomogram based on identified hub genes. (E) Calibration curves for 1-, 3- and 5-year survival. (F) Multivariate Cox regression analyses of age, stage, and risk group. (G) Time-dependent ROC curves with AUC values indicating the predictive accuracy of the risk score model for 1-year, 3-year, and 5-year survival. BDCPP, bis(1,3-dichloro-2-propyl) phosphate; EC, endometrial cancer; AUC, area under the curve; ROC, receiver operating characteristic.

Single-gene GSEA for hub genes

To further elucidate the potential regulatory mechanisms of the eight hub genes in BDCPP-induced EC, we conducted single-gene GSEA (Supplemental Digital Content Table S5, available at, http://links.lww.com/JS9/F193). For each hub gene, we visualized the top three pathways ranked by GSEA results (Fig. 5A–H). Notably, “Nucleocytoplasmic transport” was enriched by three hub genes, while “Polycomb repressive complex” and the “mRNA surveillance pathway” were each enriched by two hub genes (Fig. 5I).

Figure 5.

(A–H) Single-gene GSEA enrichment results (top 3) of each hub gene. (I) Table of the top 3 enrichment results and corresponding genes. The terms were ordered bases on the number of hub genes associated. GSEA, gene set enrichment analysis.

Molecular docking

To validate the binding affinity between BDCPP and the eight identified hub genes, we performed molecular docking using CB-Dock2 (Fig. 6A) and visualized the interactions with Discovery Studio (Fig. 6B–I). Binding energy served as a crucial metric for assessing interaction stability: values <0 kcal/mol indicated that the binding process was thermodynamically spontaneous, ≤−5.0 kcal/mol suggested good binding potential, and ≤−7.2 kcal/mol denoted strong binding potential^[17]. Among the eight hub genes examined, binding energies ranged from −4.6 to −5.4 kcal/mol, with THRB, TLR9, PLAU, and SIRT2 displaying binding energies below −5 kcal/mol, indicating their good potential to bind with BDCPP. The identified hub genes represent candidates for developing exposure biomarkers and designing therapeutic strategies (e.g., inhibitors or competitive binders) to mitigate adverse effect.

Figure 6.

(A) Binding energy of BDCPP with the proteins corresponding to the hub genes. (B–I) Molecular docking result of BDCPP with the corresponding proteins. The bond lengths between the ligand and receptor are labeled adjacent to the corresponding interaction. The relevant amino acid residues in the receptor protein are displayed using their abbreviated names along with the residue numbers. BDCPP, bis(1,3-dichloro-2-propyl) phosphate.

Discussion

BDCPP, a predominant metabolite of the widely used OPFR TDCPP, has emerged as an increasingly concerning environmental pollutant linked to multiple health adversities, including EC. In this study, we conducted a comprehensive investigation into the potential mechanisms underlying BDCPP-induced EC. Specifically, we identified 165 targets implicated in BDCPP-related EC and constructed a PPI network (Fig. 2). Subsequently, leveraging 117 machine learning models, we screened and identified eight hub genes: PLA2G2A, PLAU, SIRT2, DRD2, GSK3A, THRB, CYP17A1, and TLR9 (Fig. 4). Molecular docking analysis confirmed the spontaneous binding of BDCPP to the proteins encoded by these hub genes (Fig. 6).

Based on KEGG, GO, and Reactome pathway enrichment results, we categorized BDCPP’s pathogenic mechanisms into four axes: inflammatory activation, hormonal disruption, lipid metabolism, and epigenetic dysregulation. The related hub genes and enriched pathways are summarized in Supplemental Digital Content Table S1, available at, http://links.lww.com/JS9/F193.

Within the inflammatory axis, hub genes PLA2G2A, PLAU, and TLR9 mechanistically drive endometrial carcinogenesis by orchestrating a pro-tumorigenic microenvironment. PLA2G2A (phospholipase A2 group IIA) releases arachidonic acid that is converted to prostaglandin E2 via cyclooxygenase-2, directly stimulating endometrial cell proliferation and angiogenesis while suppressing apoptosis^[32,33]. This lipid mediator also amplifies “Neutrophil extracellular trap formation” (KEGG) by activating neutrophils, thereby fostering a pro-tumor microenvironment^[34]. PLAU (urokinase plasminogen activator) cleaves plasminogen to plasmin, activating matrix metalloproteinases that degrade basement membranes, a prerequisite for tumor invasion and metastasis^[35]. Its overexpression facilitates the “Cell recruitment” pathway (Reactome) by processing chemokines that attract tumor-associated macrophages^[36]. TLR9 recognizes hypomethylated CpG DNA from necrotic tumor cells, triggering Nuclear Factor kappa-light-chain-enhancer of activated B cells (NF-κB)-mediated transcription of IL-6 and TNF-α (“Signaling by Interleukins,” Reactome)^[37]. This establishes a chronic inflammatory loop that induces epithelial-to-mesenchymal transition (EMT) in endometrial cells, fostering an immunosuppressive niche permissive for EC progression^[38]. Collectively, this axis promotes persistent inflammation, enhances tissue invasion, and creates an immunosuppressive niche, facilitating EC initiation and progression.

The hormonal disruption axis features CYP17A1, THRB, and DRD2 as key endocrine disruptors. CYP17A1 catalyzes androgen synthesis from progesterone, providing substrates for intratumoral estrogen production via aromatase – a critical driver of estrogen-dependent EC pathogenesis^[39]. Its induction by BDCPP may explain pollutant-associated EC risk in postmenopausal women where ovarian estrogen is absent^[40]. THRB (thyroid hormone receptor β) acts as a tumor suppressor by regulating Phosphatidylinositol 3-Kinase (PI3K)/Also known as PKB Protein Kinase B (AKT) signaling^[41]. Its downregulation in EC permits uncontrolled cell cycle progression through cyclin D1 overexpression^[42]. The disrupted “Thyroid hormone signaling” (KEGG) further impairs mitochondrial metabolism, shifting cells toward glycolytic phenotypes^[43]. DRD2 (dopamine receptor D2) exhibits paradoxical roles – while canonically inhibiting prolactin secretion, its ectopic expression in endometrial cells activates Src/FAK pathways, enhancing cell motility and invasion^[44]. Acting through a non-canonical β-arrestin2-mediated transcriptional program, DRD2 can upregulate oncogenes such as MYC, providing an endocrine-independent carcinogenic mechanism^[45]. In conclusion, this axis contributes to EC progression by disrupting endocrine homeostasis, promoting estrogen-driven proliferation, and altering cellular responses to hormonal cues.

Lipid metabolic reprogramming pivots on the dual functionality of PLA2G2A. Beyond inflammation, its hydrolysis of membrane phospholipids generates lysophosphatidic acid (LPA) – a potent mitogen that activates LPA receptors to stimulate EC cell proliferation via Extracellular Signal-Regulated Kinase (ERK) phosphorylation^[46]. The enzyme’s role in “Regulation of lipolysis” (KEGG) liberates free fatty acids that serve as energy substrates and Peroxisome Proliferator-Activated Receptor γ (gamma) (PPARγ) ligands, promoting adipokine-mediated tumor growth^[47]. In EC, this reprogramming supports the high energy demands of proliferating tumor cells, contributes to the tumor-promoting inflammatory milieu, and may be a key link between obesity-associated metabolic dysfunction and EC pathogenesis^[48,49].

Epigenetic dysregulation features SIRT2 and GSK3A. SIRT2, an NAD⁺-dependent histone/non-histone deacetylase, targets non-histone substrates such as α-tubulin and p53, while also entering the nucleus during mitosis to deacetylate H4K16ac, thereby regulating chromatin dynamics^[50]. Additionally, SIRT2 activates the transcription factor FOXO3a through deacetylation, which indirectly represses the expression of tumor-suppressor genes such as Phosphatase and tensin homolog deleted on chromosome ten (PTEN)^[51]. On the other hand, GSK3A phosphorylates histone H3 at Thr6 (H3T6ph), preventing the recruitment of the demethylase LSD1 and thus preserving pro-proliferative methylation patterns when overexpressed^[52]. Collectively, this axis drives EC progression by establishing aberrant gene expression programs.

In single-gene GSEA, “Nucleocytoplasmic Transport,” “Polycomb Repressive Complex (PRC),” and “mRNA Surveillance Pathway” consistently ranked among the top three enriched pathways across multiple hub genes, indicating their pivotal roles in BDCPP-induced EC.

Nucleocytoplasmic transport, a core process governing the exchange of materials (e.g., transcription factors, RNAs, and regulatory proteins) between the nucleus and cytoplasm, can be disrupted by environmental toxicants^[53]. This disruption may occur through interference with nuclear pore complex function or altered expression of transport proteins, leading to aberrant activation of oncogenic signaling pathways^[54]. Specifically, SIRT2 regulates the nuclear-cytoplasmic shuttling of key proteins like FOXO3a and p53; its dysregulation by BDCPP could trap tumor suppressors in the cytoplasm or mislocalize oncoproteins^[50,51]. GSK3A phosphorylation events can influence the activity of nuclear import/export machinery components^[52]. THRB, as a nuclear hormone receptor, relies heavily on regulated nuclear import for its transcriptional function; its perturbation disrupts thyroid hormone signaling and downstream metabolic/growth control^[41,43].

PRC mediates epigenetic silencing of developmental and tumor suppressor genes via histone modifications such as H3K27 trimethylation^[55]. Our finding implicates DRD2 and TLR9 in PRC function. DRD2’s non-canonical β-arrestin2 signaling can potentially modulate chromatin remodelers or recruit PRC components to specific genomic loci, facilitating the silencing of tumor suppressors like PTEN and promoting oncogene (e.g., MYC) expression^[45]. TLR9 activation by endogenous ligands can trigger inflammatory signaling cascades (e.g., NF-κB) that intersect with epigenetic regulators, potentially influencing PRC-mediated silencing programs involved in EMT or immune evasion^[37,38]. This suggests BDCPP might exploit PRC-mediated epigenetic silencing through these hub genes, contributing to the stable repression of critical tumor suppressor pathways in endometrial cells, a hallmark of cancer progression with poor prognostic implications.

The mRNA Surveillance Pathway, particularly nonsense-mediated mRNA decay (NMD), maintains proteome stability by identifying and degrading aberrant transcripts (e.g., those containing premature termination codons)^[56]. THRB dysfunction could generate aberrant THRB transcripts susceptible to NMD, further amplifying the loss of tumor-suppressive functions and dysregulating PI3K/AKT signaling^[41,42]. TLR9 signaling induces potent inflammatory cytokines (e.g., IL-6, TNF-α); efficient NMD is crucial to prevent the accumulation of potentially harmful transcripts arising from the intense transcriptional activity in inflammation. Compromised NMD due to TLR9 pathway overactivation could lead to the persistence of pro-inflammatory and pro-tumorigenic transcripts^[37]. Dysfunctional mRNA surveillance allows the accumulation of truncated or aberrant proteins with dominant-negative or gain-of-function oncogenic potential, highlighting a novel mechanism linking environmental toxicants to post-transcriptional control failure in carcinogenesis.

Collectively, these pathways may form a synergistic network in BDCPP-induced EC. Dysregulated nucleocytoplasmic transport facilitates the nuclear accumulation of oncogenic signals, activating proliferation and inflammatory programs. Epigenetic silencing mediated by PRC represses tumor suppressor genes, thereby deregulating the PI3K–AKT and cell cycle pathways. Their interaction with previously discussed mechanisms involving inflammatory activation, hormonal disruption, lipid metabolism, and epigenetic dysregulation underscores the multi-target nature of BDCPP’s carcinogenic effects.

Our findings carry significant public health implications. Given the escalating environmental exposure to BDCPP, our mechanistic elucidation of its role in EC provides critical scientific evidence for regulatory action. The identified hub genes and molecular pathways offer biomarkers for early exposure assessment in vulnerable populations. Furthermore, the developed BDCPP exposure risk score model serves as a potential clinical tool for stratifying high-risk EC patients, informing targeted screening and prevention strategies. This work directly addresses the urgent need for evidence-based chemical risk assessments to guide public health policies aimed at reducing environmental carcinogen exposure. However, several methodological constraints warrant consideration. First, the variability in target predictions across different databases, computational methods, and filtering criteria leads to inherent inconsistencies in raw data quality, highlighting the current absence of standardized procedures in this research area. Additionally, while evaluating up to 117 machine learning algorithms in this study offers advantages in exploring optimal models, it also significantly increases the risk of overfitting. More importantly, reliance solely on a single dataset for training and evaluation further amplifies the overfitting risk and limits the generalizability of the results. Future studies should assess the generalization performance of the final selected model in independent, high-quality external validation cohorts. Furthermore, given the moderate predictive power demonstrated by the current model, subsequent studies should focus on optimizing algorithms and enhancing dataset diversity to enable more precise prognostic analysis of EC induced by BDCPP. Finally, the findings of this study are entirely derived from bioinformatics and computational biology approaches, lacking functional validation through in vitro or in vivo experiments, and thus carry a degree of speculative nature. To elucidate the practical biological implications of these findings more comprehensively and enhance their translational potential, future research will prioritize experimental validation. This will involve using qPCR and Western Blot techniques to assess mRNA and protein expression changes of hub genes in relevant cell models following BDCPP exposure, as well as employing siRNA-mediated gene knockdown to specifically silence hub genes in vitro and subsequently evaluate their phenotypic effects induced by BDCPP exposure, thereby functionally confirming the core roles of these genes.

Conclusion

In conclusion, our study integrated network toxicology, machine learning, and molecular docking approaches to elucidate the mechanisms underlying BDCPP exposure-induced EC. Initially, we identified potential targets of BDCPP and EC using public databases, yielding 165 overlapping targets implicated in BDCPP-induced EC pathogenesis. We then fitted 117 machine learning models and selected the RSF + GBM model as the optimal predictor based on the C-index values in the test dataset. This model identified eight hub genes: PLA2G2A, PLAU, SIRT2, DRD2, GSK3A, THRB, CYP17A1, and TLR9. A risk score was calculated based on these hub genes and subsequently utilized to construct a prognostic analysis model. Molecular docking predicted the binding affinity of BDCPP to the proteins encoded by these hub genes. Enrichment analyses using the GO, KEGG, and Reactome databases highlighted the pivotal roles of inflammatory activation, hormonal disruption, altered lipid metabolism, and epigenetic dysregulation in pathogenic mechanisms. Additionally, single-gene GSEA analysis of the hub genes highlighted the critical roles of nucleocytoplasmic transport, polycomb repressive complex, and mRNA surveillance pathway. By elucidating hub genes and underlying mechanisms, our study contributes to the growing body of evidence linking BDCPP exposure to EC, providing novel insights into environmental health risks and cancer etiology.

Ethical approval

The data used in this study were all sourced from freely accessible public databases.

Consent

This study did not involve patients or volunteers.

Sources of funding

This study was conducted without financial support.

Author contributions

Z.W.: writing – original draft, data curation, investigation; Y.F.: writing – review and editing, investigation; Q.C.: writing – review & editing, investigation; L.Z.: Conceptualization, methodology, software, formal analysis, investigation, data curation, visualization, writing – original draft, writing – review & editing, supervision, project administration.

Conflicts of interest disclosure

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Guarantor

Linhao Zong.

Research registration unique identifying number (UIN)

This study does not involve human subjects.

Provenance and peer review

Not commissioned, externally peer-reviewed.

Data availability statement

Data will be available upon reasonable request.