Evaluation of triphenyl phosphate’s pathogenic potential and molecular mechanisms in glioblastoma: an integrated network toxicology investigation employing multiple machine learning approaches

Abstract

Background

Glioblastoma (GBM) is a highly aggressive brain tumor with a poor prognosis, and its etiology involves both genetic and environmental factors. Triphenyl phosphate (TPHP), an organophosphorus flame retardant widely used in various consumer products, has raised public health concerns due to its environmental prevalence and detection in human samples. However, its potential role in GBM pathogenesis remains unclear. This study combined machine learning and molecular simulation docking technology to investigate the effects of TPHP on the pathogenesis and related molecular mechanisms of glioma.

Methods

Our study made use of a variety of computational learning methods along with online databases to carry out differential transcriptional expression analysis on different datasets. The aim was to identify target genes related to glioma. Based on the expression levels of key genes, we constructed a risk prediction model. Network toxicology and molecular docking technologies were adopted to investigate how TPHP binds to target proteins.

Results

Fourteen genes in total were determined to be potential target genes in relation to TPHP-induced glioma. Further machine learning analysis identified three core target genes (GRM3, IDH2, and DDR2) as key biomarkers for TPHP-induced glioma, and pathway analysis determined the relevant signaling pathways of TPHP. Molecular docking showed that there was a specific binding effect between TPHP and target proteins.

Conclusions

It is possible that TPHP influences the pathogenic mechanism of glioma through targeting particular genes and pathways. Results from molecular docking simulations demonstrated a remarkable binding specificity effect between TPHP and target proteins. This effect is highly likely to be the crucial factor contributing to the development of glioma.

Highlight box.

Key findings

• This study first identifies the pathogenic link between triphenyl phosphate (TPHP) and glioblastoma (GBM), screens 3 core biomarkers, verifies TPHP-target stable binding, establishes an effective GBM risk prediction model, and offers novel insights for GBM early prevention.

What is known and what is new?

• GBM is a lethal primary brain tumor with unclear environmental pathogenic drivers. TPHP, a widely used organophosphorus flame retardant, has documented neurotoxicity, but its role and mechanisms in GBM remain uncharacterized.

• We identified 14 cross-target genes linking TPHP exposure to GBM, and further screened 3 core biomarkers (GRM3, IDH2, DDR2) via integrated machine learning approaches. Molecular docking confirmed stable binding of TPHP to GRM3 and IDH2 proteins.

What is the implication, and what should change now?

• This study reveals the potential pathogenic link between TPHP exposure and GBM for the first time, providing novel candidate biomarkers and mechanistic insights for GBM prevention and early intervention.

Introduction

One of the frequently-seen primary malignant tumors in the central nervous system is glioma. Although certain progress has been made in diagnosis and treatment in recent years, the characteristics of high invasiveness, high recurrence rate and high mortality remain one of the major challenges in the medical field (1-4). Up to now, the mechanism underlying the pathogenesis of glioma has not been thoroughly clarified. Existing studies support that its occurrence may involve multiple factors, such as IDH1/IDH2 gene mutations being more common in low-grade gliomas (2,3); high-dose exposure to ionizing radiation is a definite risk factor for glioma; nitrate intake, viral or bacterial infections, etc., may also be involved in the occurrence of glioma; other factors such as age and gender are also associated with the onset of glioma (5). In addition, the incidence rate of men is approximately 1.5–1.6 times to that of women, and it increases with age (6). The incidence rate of glioma varies in different regions. The incidence is higher in the United States and Northern Europe, while it is relatively lower in Asia (5,7). The 5-year mortality rate of glioma is also second only to pancreatic cancer and lung cancer (8-10).

In recent years, the significance of molecular pathological research in the classification of central nervous system tumors has gradually received more attention. Taking glioma as an example, some key molecular markers, such as IDH1/IDH2 gene mutations and O-6-methylguanine-DNA methyltransferase (MGMT) promoter methylation, have been proven to have extremely important values for the diagnosis, prognosis assessment, and formulation of personalized treatment plans of the disease (2,9,11). Currently, the treatment of glioma still mainly relies on surgical operations combined with adjuvant radiotherapy and chemotherapy (12-14). For glioblastoma (GBM), postoperative radiotherapy combined with temozolomide concurrent and adjuvant chemotherapy has become the standard treatment mode for adult newly diagnosed GBM (15). The overall survival prognosis of glioma patients is influenced by multiple factors, including the patient’s age, underlying health status, tumor grade and location, extent of resection, molecular variations, response to treatment, and environmental exposure, etc. (5,16-18). Further in-depth research on the mechanism of glioma occurrence and development, exploration of more potential molecular markers and therapeutic targets, are particularly important for the diagnosis and treatment of glioma.

Triphenyl phosphate (TPHP) is a widely used organic phosphorus flame retardant and plasticizer, commonly found in photographic films, nail polish, polyvinyl chloride (PVC), polycarbonate/acrylonitrile butadiene styrene (ABS) alloys, polyurethane foam, and hydraulicoil, among other products. As polybrominated diphenyl ethers (PBDEs) are gradually phased out, the usage of TPHP as a substitute has significantly increased. Nonetheless, TPHP has a high volatility and is prone to release into the environment from products, thereby entering the human body through air, dust, food, and other pathways and affecting people’s health. Relevant studies have shown that the main metabolite of TPHP in the body is diphenyl phosphate (DPHP), and it is associated with health problems such as thyroid dysfunction (19); TPHP exposure is also linked to an increased risk of metabolic diseases such as obesity and metabolic syndrome (20). Although the metabolic process and potential toxicity of TPHP in the body have been studied in recent years, its role in glioma pathogenesis remains unclear (20-22). In view of this, this study first uses network toxicology and molecular docking techniques to attempt to explore the underlying mechanism by which TPHP acts on joint tissues and leads to illness, in order to provide new perspectives and strategies for the prevention and treatment of glioma in exposed populations. We present this article in accordance with the TRIPOD reporting checklist (available at https://tcr.amegroups.com/article/view/10.21037/tcr-2025-1-2668/rc).

Methods

Data retrieval and download from Gene Expression Omnibus (GEO) database

The gene expression patterns relevant to glioma in this study were sourced from the sequencing data of three different datasets (GSE29796, normal: 6, disease: 28; GSE35158, normal: 9, disease: 14; GSE42656, normal: 10, disease: 20) in the GEO (https://www.ncbi.nlm.nih.gov/geo/) database. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Investigation of toxicity and procurement of target genes

The online tool ProTox 3.0 Chemical Substance Toxicity Prediction (https://tox.char ite.de/protox3/) was used to analyze TPHP; ChEMBL (https://www.ebi.ac.uk/chembl/), STITCH (http://stitch.embl.de/), and SwissTargetPrediction (http://www.swisstarget-pre diction.ch/) were respectively employed to explore the related target genes of the compounds.

Disease-related gene acquisition

The GeneCards (https://www.genecards.org/) database collates various genes and proteins related to various diseases. In this analysis, the GeneCards database was utilized to explore the genes associated with glioma.

Analysis based on R packages

Using “sva” to perform batch correction on the included glioma dataset. Differential expression analysis was conducted using the “limma” package in R software, and heat maps and volcano plots were generated. Multiple R packages, such as “openxlsx”, “seqinr”, “plyr”, “randomForestSRC”, “glmnet”, “plsRglm”, “gbm”, “caret”, “mboost”, “BART”, etc., were utilized to apply 113 machine learning methods to screen for diagnostic models; the “e1071/kernlab” package, “randomForest” package, and “glmnet” were used for machine learning to screen for characteristic genes. The diagnostic efficacy of characteristic genes was analyzed using the “pROC” package, and the diagnostic receiver operating characteristic (ROC) curve was drawn; risk models were constructed and validated using the “rms” and “rmda” packages.

Molecular docking

To explore the effect of TPHP on the core genes, this analysis explored molecular docking to search for possible binding associations. The structure data format (SDF) format ligand documents and tridimensional structure models of the central target protein were obtained from PubChem (https://pubchem.ncbi.nlm.nih.gov/) and the Research Collaboratory for Structural Bioinformatics (RCSB) Protein Data Bank (PDB) (https://www.rcsb.or g/) (23). Protein-ligand blind docking was performed using CB Dock2 (https://cadd.labshare.cn/cb-dock2/index.php).

Statistical analysis

All statistical analyses were performed using R software (version 4.2.0). Differential gene expression was analyzed using the “limma” package [|log₂fold change (FC)| >0.585, false discovery rate (FDR) <0.05]. Weighted gene co-expression network analysis (WGCNA) was used to identify gene modules. A total of 113 machine learning models were constructed and evaluated by 10-fold cross-validation. ROC curves and nomogram were generated for model validation. All tests were two-sided, and P<0.05 was considered significant.

Results

Differential expression analysis

First, we downloaded three mRNA Seq datasets of brain gliomas from the GEO database: GSE29796, GSE35158, and GSE42656. We removed the batch effects from the included datasets, and Figure 1A,1B presents the box plots of the samples both before and after eliminating the batch effects. To verify the batch effect removal effect of the datasets, we conducted principal component analysis (PCA) on the samples incorporated in the study, and the findings are depicted in Figure 1C,1D. Subsequently, using the “limma” package (|log₂FC| >0.585), we conducted differential expression analysis, obtaining 529 differentially expressed genes (DEGs), and some heatmaps are shown in Figure 1E. Then, using WGCNA, we analyzed the included mRNA-Seq data. The optimal soft threshold with scale independence and average connectivity at six was the best (Figure 2A,2B). The module heatmap, at this juncture, demonstrated that the cyan module presented the most significant disease correlation (Figure 2C, R=0.26, P=1e−4). Figure 2D shows how the gene members of the cyan module are related to the disease (correlation =0.54, P=7.5−e161), and 2,121 related genes were obtained. This study also used GeneCards (https://www.genecards.org/) to explore genes related to brain glioma. A total of 1,055 genes related to brain glioma were obtained (relevance score >5, available at https://cdn.amegroups.cn/static/public/TCR-2025-1-2668-1.csv).

Figure 1.

Differential expression analysis. (A) Box plot of data before removing batch effects; (B) box plot of data after removing batch effects; (C) PCA of data distribution before removing batch effects; (D) PCA of data distribution after removing batch effects; (E) heatmap of the results of sample differential expression analysis. PCA, principal component analysis.

Figure 2.

Results of WGCNA analysis. (A) Scale independence: a scatter plot showing the fitting index of the scale-free topological model varying with the soft threshold; (B) the average connectivity in the right figure: a scatter plot showing the average connectivity varying with the soft threshold; (C) heatmap of the module-feature relationship; (D) the relationship between the module members in the turquoise module and the significance of genes; (E) the intersection of DEGs, genes related to the WGCNA cyan module obtained from GeneCards, and relevant genes. DEG, differentially expressed gene; ME, module member; WGCNA, weighted gene co-expression network analysis.

For the 529 DEGs, 2,121 WGCNA cyan module-related genes and 1,055 related genes obtained from GeneCards, we took the intersection. The genes that appeared in either of the two screening methods were identified as brain glioma-related genes. A total of 431 brain glioma-related genes were obtained (Figure 2E).

Acquisition of cross-target genes and their enrichment analysis

The toxicity and mutagenic ability of TPHP and its pathogenicity were analyzed using the online tool ProTox 3.0 Prediction of Toxicity of Chemicals (https://tox.charite.de/protox3/). The results showed that TPHP had significant mutagenic and pathogenic effects. Subsequently, the relevanttargets of TPHP were explored using online tools ChEMBL (https://www.ebi.ac.uk/chembl/), Stitch (http://stitch.embl.de/), and SwissTarget Prediction (http://www.swisstargetprediction.ch/). Through the integrated analysis of the obtained targets, 310 targets related to TPHP were finally determined (available at https://cdn.amegroups.cn/static/public/TCR-2025-1-2668-2.xlsx). These TPHP-related targets were further compared with the brain glioma-related targets, and 14 common cross-targets were selected (Figure 3A). Based on these cross-targets, the corresponding gene network diagrams were constructed, as shown in Figure 3B.

Figure 3.

Acquisition of cross-target genes and their enrichment analysis. (A) Intersection of disease-related genes and TPHP-related targets; (B) network diagram of 14 cross-target genes; (C) GO functional enrichment analysis; (D) visualization of the GO functional enrichment network of target genes; (E) KEGG enrichment analysis results. GO, Gene Ontology; KEGG, Kyoto Encyclopedia of Genes and Genomes; PID, Pathway Interaction Database; TCA, tricarboxylic acid; TPHP, triphenyl phosphate.

Subsequently, Gene Ontology (GO) functional and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway exploration was conducted for the 14 core intersection target genes selected. The GO enrichment analysis results indicated that they mainly participated in responses to heterogenic stimuli at the biological process (BP) level, cancer pathways, Pathway Interaction Database (PID) ATF2 pathways, and other functions (Figure 3C,3D). The results of KEGG analysis indicated that these cross—target genes were predominantly focused on key pathways (see Figure 3E). Of particular note is the enrichment in the “Neuroactive ligand-receptor interaction” pathway, as TPHP is a neuroactive environmental compound, and dysregulation of synaptic signaling has been implicated in glioma growth (24). Enrichment in the “citrate cycle [tricarboxylic acid (TCA) cycle]” is also critical, as metabolic reprogramming is a hallmark of cancer, and IDH2 is a central enzyme in this pathway (25). Furthermore, the “ advanced glycation end-products (AGE)-receptor for advanced glycation end-products (RAGE) signaling pathway in diabetic complications” suggests a potential link to inflammatory and oxidative stress responses, which are mechanisms shared by many environmental toxicants. These pathway associations provide mechanistic hypotheses for how TPHP exposure might perturb cellular processes relevant to gliomagenesis.

Construction of a diagnostic model based on the combination of 113 machine learning methods

Furthermore, a diagnostic model was developed by integrating 113 machine-learning methods. The combined data from the three datasets served as the training set, while each of the three individual datasets was utilized as a validation set. The results are shown in Figure 4A. Among them, the “least absolute shrinkage and selection operator (LASSO) + GBM” combination showed the most favorable diagnostic performance among the tested algorithms. The visualization of the model diagnostic efficacy of the training set and validation set is shown in Figure 4B-4E. The confusion matrix was employed to verify the diagnostic efficacy of the model. The results, too, indicated a high level of diagnostic efficacy (Figure 4F-4I). Finally, the “LASSO + GBM” model was selected as the best development plan, and 9 key cross-target genes were determined: IDH2, GRM3, CDK4, HTR2A, PLAU, IDH1, GRIN2A, DDR2, and AR.

Figure 4.

Construction and validation of the diagnostic model. (A) Heatmap of the diagnostic model constructed based on 113 machine learning methods (yellow bars label model performance values; image shows top performers); (B-E) diagnostic ROC curves of the training set and 3 validation sets; (F-I) confusion matrices of the training cohort and 3 validation cohorts in sequence. AUC, area under the curve; CI, confidence interval; GBM, glioblastoma; LASSO, least absolute shrinkage and selection operator; LDA, linear discriminant analysis; ROC, receiver operating characteristic; SVM, support vector machine.

Core cross-target gene screening

Then we conducted a screening of nine key cross-target genes. LASSO regression analysis (Figure 5A), random forest (RF) algorithm (Figure 5B), and support vector machine (SVM) algorithm (Figure 5C,5D) were used to screen these nine key cross-target genes. LASSO regression analysis obtained five genes when the cross-validation error was the lowest; the SVM algorithm also obtained eight genes when the cross-validation error was the lowest; with seed set to 123 and ntree set to 500, the error was the smallest when ntree was 23, and the top five genes with the highest importance score were selected; the intersection of the three machine learning methods was taken, and finally, three core cross-target genes were obtained: GRM3, DDR2, and IDH2, as shown in Figure 5E. These genes have been These genes have been implicated in crucial cancer-related processes. GRM3 modulates glutamate signaling and has been associated with glioma progression and therapy response. IDH2 mutations are well-established drivers in gliomagenesis, affecting cellular metabolism and epigenetics. DDR2 is a collagen receptor involved in cell adhesion and migration, processes critical for cancer invasion. Their identification as core targets strengthens the biological plausibility of TPHP’s potential involvement in glioma.

Figure 5.

Screening of core cross-target genes. (A) Core target genes are screened based on LASSO regression analysis; (B) core target genes are singled out based upon the SVM algorithm; (C,D) core target genes are screened based on the RF algorithm; (E) the intersection of the screening results yielded by the three machine-learning methods is determined. LASSO, least absolute shrinkage and selection operator; RF, random forest; RFE, recursive feature elimination; SVM, support vector machine.

Risk model construction

In addition, a risk prediction model was constructed based on the three-core cross-targets, as shown in Figure 6A. From the model, it can be observed that the upregulation of DDR2 and IDH2 expression is positively correlated with the risk of glioma occurrence, while the downregulation of GRM3 expression is negatively correlated with the risk of glioma occurrence. To assess the precision of this risk prediction model, a calibration curve was plotted. The results showed that the model prediction curve was highly consistent with the ideal standard curve, indicating that the model performed well in predicting glioma (see Figure 6B). Further decision curve analysis (DCA) indicated that the model could provide higher net benefits for clinical decision-making at lower threshold probabilities (see Figure 6C). Moreover, a clinical decision curve was also plotted in this analysis to evaluate the potential influence of the prediction model in clinical practice. The results suggested that the predicted results had a good correspondence with actual events as the threshold changed, demonstrating the potential that this model holds for clinical application (Figure 6D).

Figure 6.

Risk model construction and validation. (A) Nomogram for risk prediction model was constructed based on the three-core cross-target genes; (B) calibration curve; (C) decision curve; (D) clinical impact curve.

Molecular docking

Furthermore, molecular docking was used to explore the potential binding interactions so as to study the influence of TPHP on the hub genes in glioma. The ligand file in SDF format and the three-dimensional structure model of the core target protein were down-loaded from PubChem (https://pubchem.ncbi.nlm.nih.gov/) and RCSB PDB (https://www.rcsb.org/). Protein-ligand blind docking was carried out using the online tool CB Dock2 (https://cadd.labshare.cn/cb-dock2/index.php). TPHP was docked with three core cross-target genes, respectively. It is generally believed that when the binding affinity is less than −7 kcal/mol, the binding is relatively stable. The results showed that TPHP exhibited high binding stability with GRM3 and IDH2 (Figure 7A,7B). The specific binding sites are shown in the figures. These sites indicate the specific amino acid residues of TPHP binding to the protein, and this is of great significance for comprehending how TPHP impacts the functions of these proteins and their potential biological consequences.

Figure 7.

Molecular docking results. (A) Molecular docking result of TPHP with GRM3; (B) molecular docking result of TPHP with IDH2. TPHP, triphenyl phosphate.

Discussion

The metabolic process of TPHP in the body mainly relies on the cytochrome P450 enzymes in the liver and serum, converting into various metabolites including DPHP (26). TPHP can migrate from the final product to the surrounding environmental media, such as indoor dust and soil, and has been found in human specimens (such as the urine of pregnant women and infants) (27). This indicates that TPHP and its metabolites have extensive distribution and migration capabilities in the environment and may enter the human body through multiple exposure pathways, posing a potential threat to human health. Existing studies suggest that TPHP has definite interference with lipid metabolism, endocrine disruption, and potential developmental toxicity effects. Long-term exposure and cumulative intake will cause immeasurable damage to the human body (26). Considering that glioma is a central nervous system malignant tumor that is sensitive to environmental factors, the environmental contact with TPHP might have a crucial part in the occurrence and evolution of glioma.

The combined study based on network toxicology and various machine learning methods aimed to provide a new perspective for exploring the possible mechanism of TPHP acting on glioma. GO functional enrichment analysis revealed that at the BP level, processes related to TPHP covered responses to heterologous stimuli and positive regulation of the neuronal apoptosis process. These processes may participate in the occurrence and development of glioma by regulating cell metabolism and neuronal survival. For example, the response to heterologous stimuli may involve the metabolism and detoxification of TPHP and its metabolites by cells, while the positive regulation of the neuronal apoptosis process may promote abnormal death of neurons, thereby creating conditions for the formation of glioma. At the cellular component level, related processes involve the tyrosine kinase signal pathway of cell surface receptors, which may affect cell signal transduction and proliferation. The aromatic amino acid metabolism process and others may affect the progression of glioma by regulating neurotransmitter levels and cell metabolism. KEGG enrichment analysis suggested that TPHP and its metabolites may affect glioma through multiple signaling pathways, such as the AGE-RAGE signaling pathway in diabetic complications: after AGEs bind to RAGE, it activates the NADPH oxidase 4 (NOX4) in the NOX family of NADPH oxidase, leading to excessive production of reactive oxygen species (ROS). Activation of RAGE also may contrleibute to the secretion of pro-inflammatory cytokines, such as tumor necrosis factor alpha (TNF-α) and interleukin-6 (IL-6), further exacerbating the inflammatory response.

In addition, AGEs activate multiple signaling pathways, such as nuclear factor kappa B (NF-κB), mitogen-activated protein kinase (MAPK), and phosphoinositide 3-kinase (PI3K)-protein kinase B (AKT)-mammalian target of rapamycin (mTOR), through binding to RAGE, causing adverse effects such as inflammation, oxidative stress, and cell apoptosis, which are closely related to the occurrence and development of glioma (28,29). In addition, the core target, IDH2, is one of the key enzymes in the citric acid cycle (TCA cycle), responsible for catalyzing the dehydrogenation of isocitrate to generate α-KG. Mutant IDH2 leads to the accumulation of 2-HG, which directly disrupts TCA cycle homeostasis, affecting energy metabolism and the supply of biosynthetic precursors. As a structural analog, 2-HG competitively inhibits α-KG-dependent dioxygenases (such as prolyl hydroxylase and histone demethylases), resulting in metabolic dysregulation and increasing the risk of glioma development.

Among the brain glioma diagnostic models constructed by combining 113 machine learning methods, the “LASSO + GBM” combination demonstrated the best diagnostic efficacy. Further screening based on RF, LASSO regression analysis, and SVM algorithms resulted in the identification of three core cross-targets. A risk prediction model constructed based on these three-core cross-targets showed good predictive ability for brain glioma in our dataset. This conclusion was verified by calibration curves and DCA, suggesting that this model has potential application value in clinical practice. The construction and application of this model provide new tools for the early diagnosis and risk assessment of brain glioma, which may help clinicians intervene at the early stage of the disease and improve the survival rate and quality of life of patients.

The molecular docking predictions provide a structural basis for functional hypotheses. The high binding affinity predicted between TPHP and GRM3 is particularly intriguing. GRM3 is a metabotropic glutamate receptor known to modulate synaptic transmission and neuronal survival. If TPHP binding disrupts GRM3 signaling, it could potentially lead to neuronal excitotoxicity or altered glial-neuronal communication, creating a permissive environment for tumorigenesis (30,31). Similarly, the predicted binding to IDH2, a key metabolic enzyme, raises the possibility that TPHP might interfere with its normal function, potentially mimicking or exacerbating the effects of oncogenic IDH mutations that drive metabolic alterations in glioma (32,33). Thus, our findings shift the research question from ‘if’ TPHP interacts with these proteins to ‘how’ such interactions might mechanistically contribute to glioma pathogenesis, guiding future functional experiments. The molecular docking result allows us to hypothesize about the interaction mechanism between TPHP and proteins related to glioma, thereby providing molecular-level evidence for revealing its role in the occurrence and development of glioma. These findings offer new perspectives for the early prevention and later diagnosis and treatment of glioma, and may help to develop specific antagonists or degraders targeting TPHP to alleviate its promoting effect on glioma. However, it must be clarified that molecular docking is merely a computational algorithmic simulation, providing supportive evidence for the conclusions. Further detailed experimental validation is required to determine whether TPHP modulates the GRM3/DDR2/IDH2 and subsequent signaling pathways in glioma-relevant systems, including comprehensive cellular assays (specific experimental design is detailed in Appendix 1).

Even if TPHP can be detected in human body fluid samples, the threshold concentration for causing disease is also very important (20,27). Then, can environmental TPHP exposure realistically result in biologically relevant concentrations reaching the human brain? TPHP is rapidly metabolized to DPHP in the body. While our study focuses on TPHP, existing toxicokinetic data suggest that both TPHP and DPHP can cross the blood-brain barrier, albeit with limited efficiency (34). Estimated human exposure levels from environmental sources like dust are in the nanogram to low microgram per kilogram of body weight per day range (35). The binding affinities we predicted (below −7 kcal/mol) for TPHP with GRM3 and IDH2 indicate high stability, which could be biologically significant even at low concentrations if the compounds accumulate in the brain over time or if exposure is chronic. This provides a plausible, though not yet proven, link between environmental exposure levels and the potential for target interaction identified in our docking simulations. Therefore, while our computational model suggests a potential mechanism, the dosimetric relationship between external exposure, internal brain concentration, and biological effect remains a critical uncertainty that must be addressed in future toxicokinetic and in vivo studies.

Although this study provides a new perspective for understanding the association between TPHP and its metabolite DPHP and GBM, its limitations are also obvious. Firstly, this study is based on publicly available gene expression data and network toxicology analysis, which may not fully consider the influence of individual differences and environmental factors on the pathogenic potential of TPHP. For example, there may be significant differences in genetic background, lifestyle, and environmental exposure levels among different individuals, and these factors may interact with the effects of TPHP and jointly influence the risk of GBM occurrence. Secondly, the construction of the machine learning model relies on the selected dataset, and the sources and characteristics of different datasets may limit the generalization ability of the model. Although we improved the reliability of the model through cross-validation, it is still necessary to verify it in a larger range of independent samples to ensure its stability and accuracy in different populations and environmental conditions. Moreover, the model demonstrated extremely high predictive performance in both the training and validation sets, indicating an overfitting issue. In the future, if conditions permit, it will be necessary to use our own sequencing data for further validation. Although the results of molecular docking revealed the binding sites of TPHP to target proteins, further verification of its biological effects through experimental methods, such as in vitro cell experiments and in vivo animal model studies, is still needed to directly observe the impact of TPHP on the functions of GBM-related proteins and cellular biological behaviors. Therefore, future research should integrate multiple experimental methods, including in vitro cell experiments, in vivo animal models, population epidemiological studies, and clinical sample analysis, to more comprehensively elucidate the mechanism of action of TPHP and its metabolite DPHP in the occurrence of GBM, and provide more powerful scientific evidence for the prevention and treatment of GBM.

Despite the insights gained, this study has several limitations that should be acknowledged. First, as a purely computational analysis, our findings are hypothesis-generating and require validation in experimental models (e.g., in vitro or in vivo studies). Second, the predictive performance of our model, though promising, needs to be verified in larger, independent clinical cohorts. Finally, the precise biological mechanisms underlying the predicted interactions remain to be fully elucidated through functional studies.

Conclusions

In this study, integrating network toxicology, machine learning methods, and molecular docking experiments, initially revealed the potential association and mechanism of action between TPHP and GBM, and constructed a risk prediction model with preliminary predictive ability. However, these findings still require further experimental verification and clinical research to confirm their scientificity and clinical application value.

Supplementary

The article’s supplementary files as

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.