Introduction

Neonicotinoid (NEO) pesticides are the fourth generation of insecticides, which have good effects on the control of stack-inhaling mouth pests, small Lepidopteran and Coleopteran pests1. The commonly used NEOs include Imidacloprid (IMI), Acetamiprid (ACE), Thiamethoxam (THX), Clothianidin (CLO) and Dinotefuran (DIN)2. Unlike previous generations of pesticides, such as organophosphorus pesticides and pyrethroids, which inhibit acetylcholinesterase and interfere with sodium channels, NEOs kill insects by directly binding to the insect’s acetylcholine receptor and simulating acetylcholine acting on its nervous system3,4,5. Its insecticidal mechanism is more selective, systematic, and less residual, so it has been widely used worldwide6. As the utilization of neonicotinoid pesticides has increased, they have been detected in various environmental media, such as domestic soil, water resources, and residential areas7,8,9. Residues have also been found in foods, such as fruit and other products consumed by humans, which has prompted concerns about potential exposure risks10. Moreover, traces of these pesticides have been detected in biological samples like blood and urine, further emphasizing the necessity for discourse surrounding their safety11,12. Studies have demonstrated that NEOs can have adverse effects on human health, including contributing to chronic diseases such as neurological disorders and cancer13,14. Recent clinical research suggested that exposure to neonicotinoids and their metabolites may increase the risk of liver cancer15.

Hepatocellular carcinoma (HCC) is the most common type of primary liver cancer and ranks as the seventh most common frequently occurring malignant tumor and is the fourth leading cause of cancer-related mortality in the world16,17. Researches indicate that the development of HCC is associated with chronic liver diseases and pesticide exposure18. Nevertheless, few studies have undertaken a comparative analysis of the hepatic toxicity and carcinogenic potential of commonly used neonicotinoid insecticides. Furthermore, research on the mechanisms by which these pesticides may induce HCC is also limited.

To address the aforementioned, we initially utilized the ADMETlab3.0 and Protox3.0 databases to predict the toxicity of five commonly used NEOs and found the pesticide with the highest potential for carcinogenicity and liver damage. Subsequently, network toxicology, enrichment analysis, GEO database and four machine learning algorithms combined with molecular docking were used to investigate the mechanism of the pesticide inducing HCC. The above methods are valuable due to their capacity to predict and detect possible toxic effects by analyzing large datasets and biological data, thereby providing an in-depth understanding of the risks related to pesticide exposure at the molecular level. Moreover, assessing the potential carcinogenic and hepatotoxic effects of this pesticide is crucial for evaluating its safety in both agricultural and medical contexts, as well as for generating scientific evidence that can guide the establishment of safety protocols and protective guidelines.

Methods

Toxicity analysis of common neonicotinoid pesticides and pesticide selection

The chemical information and Simplified Molecular Input Line Entry System (SMILES) identifiers for five NEOs were obtained from PubChem (https://pubchem.ncbi.nlm.nih.gov/) and PubMed (https://pubmed.nc- bi.nlm.nih.gov/). The toxicity scores for these pesticides were assessed using the ProTox3.0 (https://tox.charite.de/protox3/index.php?site=home) and ADMETlab3.0 (https://admetlab3.scb- dd.com/server/screening) databases. The pesticide with the highest potential for carcinogenicity and hepatotoxicity was selected for further analysis.

Identification of pesticide and HCC-related targets

Relevant pesticide targets were retrieved from the SwissTarget Prediction (http://swisstarget-prediction.ch/) and CTD (https://ctdbase.org/) databases. Targets in the Swiss Target Prediction were chosen if their p-value exceeded 0. Protein names were standardized by using the database- UniProt (https://www.uniprot.org/), and duplicate entries were eliminated. The main HCC-related targets were obtained from GenCards (https://www.genecards.org/) and supplemented by OMIM (https://www.omi- m.org/). The intersection of pesticide and HCC-related targets were identified by VENNY2.1 (https://bioinfogp.cnb.csic.es/tools/venny/index.html) as potential toxicological targets for further investigation.

GO and KEGG functional analysis

The intersected protein names were input into the Metascape database (https://metascape.org/gp/index.html#/main/step1) for GO analysis and KEGG pathway enrichment19. For GO analysis, the top 20 biological processes (BP), cellular components (CC), and molecular functions (MF) were selected according to p-values. The 20 signaling pathways with the smallest p-values were selected in the KEGG analysis. The plot was then performed according to −log10(P).

GEPIA analysis and GEO data collection for validation of expression differences, chromosomal position, and correlation of SDECGs

In order to reveal the core mechanism of THX affecting HCC in more depth, we screened the intersection targets. The intersected targets were input into the GEPIA (http://gepia.cancer-pku.cn/) website to assess significant differential expression (P < 0.01) of each target in HCC and normal groups, and the differential targets were identified as core genes for further analysis. Then a search of the GEO-database (https://www.ncbi.nlm.nih.gov/geo/) for datasets related to “Hepatocellular carcinoma” was conducted. GSE45267, GSE62232, GSE112790 and GSE54236 were used for further study. Gene symbol annotation and data correction were performed using Perl code to extract the expression levels of core genes in normal and HCC samples. Then these data were extracted for differential expression analysis using RStudio packages such as “limma”, “ggpubr” and “pheatmap”. The results were shown using boxplots and heatmaps, and then verified with the results obtained from GEPIA. Significantly differentially expressed core genes (SDECGs) were defined as genes with an adjusted p-value (adjP, FDR-corrected) < 0.01, consistent with GEPIA expression results20. The chromosomal location of core genes was determined using Perl code and visualized as circular plots using the ‘Rcircos’ R package. Additionally, the correlation of each SDECG was computed using the ‘cor’ function and visualized.

Immune cell infiltration, differences, and correlation in HCC samples

Immune cell infiltration was assessed using the CIBERSORT algorithm in R (4.4.1), with 1,000 iterations to estimate the relative abundance of various immune cell types. The results were visualized through bar plots representing immune cell content in each sample. To compare immune cell infiltration between normal and HCC groups, single-sample Gene Set Enrichment Analysis (ssGSEA) was conducted using the “GSVA” and “GSABase” R packages. The ssGSEA results were visualized using box plots to highlight differences in immune cell content. Furthermore, the correlation between SDECGs and ssGSEA scores was analyzed, and the correlation coefficients were visualized for better interpretation21.

Machine learning models for pesticide-induced HCC and feature gene selection with external validation

Four predictive models were constructed using the expression data of significantly SDECGs, including Random Forest (RF), Support Vector Machine (SVM), Generalized Linear Model (GLM), and Extreme Gradient Boosting (XGB). All models were implemented in R (4.4.1) using the caret package, with model-specific algorithms provided by randomForest, kernlab, and xgboost. Each model was trained on a 70% randomly selected training set using five-fold repeated cross-validation to optimize performance and minimize overfitting. To further evaluate the predictive performance and interpretability of each model, model interpretation procedures were constructed using the DALEX package. A unified prediction function was first defined to ensure the compatibility of all models. Based on this function, the prediction residuals of each model for the test set were calculated. The residual distribution was visualized using inverse cumulative distribution plots and box plots. The receiver operating characteristic (ROC) curve and the area under the curve (AUC) were calculated by pROC software package. The receiver operating characteristic (ROC) curve was drawn for comparison visualization, and the AUC value was marked to comprehensively evaluate the discrimination ability of each model by the above method. After determining the optimal model, a nomogram was developed based on the feature genes and their expression levels in both normal and HCC groups. Decision curves and calibration curves were then used to assess the accuracy of the nomogram. For external validation, a separate dataset containing normal and HCC samples was obtained from GEO database. Machine learning models were then applied in R following the same methodology, and ROC curves were plotted to validate the model performance.

Molecular docking

The 3D structures of the selected pesticide components and feature genes were obtained from the PubChem, UniProt database and PDB database (http://www.rcsb.org/). If the required structure was not obtained in the PDB database, the AlphaFold structure was obtained in the UniProt database. Molecular docking was carried out using AutoDock Vina (1.1.2), and visualization was done with Pymol (2.6).

Results

Toxicity analysis of common neonicotinoid pesticides and pesticide selection

The chemical details of five plasticizers (IMI, ACE, THX, CLO, and DIN) are provided in Table 1 in Supplementary File, which includes their chemical formulas, molecular weights, and SMILES structures. Toxicity information related to NEOs was retrieved using ADMETlab 3.0 and Protox 3.0, as shown in Table 1. On the basis of carcinogenic and hepatotoxic properties, THX was selected for further investigation.

Table 1 The prediction toxicity of Thiamethoxam.
Full size table

Identification of THX and HCC-related targets

The core biological significance of gene crossover in network toxicology lies in its ability to decipher the molecular interconnections between compound targets and disease mechanisms. This process provides a critical mechanistic foundation for elucidating how THX exerts its effects in the context of this study, thereby offering deeper insights into the underlying action pathways and contributing to a more comprehensive understanding of the therapeutic or toxicological mechanisms involved. A total of 11 and 124 THX-related targets were obtained from Swiss Target Prediction and CTD, respectively. After merging and removing duplicates, 84 THX-related targets were identified. A total of 11,938 and 192 HCC-related targets were retrieved from Genecards and OMIM, respectively. To ensure that the genes obtained were highly correlated with HCC, the GeneCards “score” threshold was set to ≥ 3 (determined by rounding the median score value upward22). Following the merging and removal of duplicates, 5672 targets were identified. A Venn diagram was constructed, resulting in the identification of 61 common targets (Fig. 1), which were subjected to further analysis to elucidate the molecular mechanisms underlying THX-induced HCC.

Fig. 1
figure 1

Venn diagram of the intersection between HCC and THX target sets.

Full size image

GO and KEGG functional analysis

The 61 common targets were imported into Metascape for GO and KEGG enrichment analysis23. The significant biological processes identified based on p-values included response to xenobiotic stimulus, cellular response to chemical stress, cellular response to biotic stimulus and response to toxic substance. The mainly involved cellular components were peroxisome, organelle outer membrane, endoplasmic reticulum lumen. The primary molecular functions included protein domain specific binding, protein homodimerization activity, protease binding. The KEGG enrichment analysis identified several key pathways, including IL-17, FoxO, HIF-1 signaling pathway, cell cycle and Glycolysis / Gluconeogenesis (Fig. 2).

Fig. 2
figure 2

GO and KEGG analysis of THX and HCC-Related Targets (A–C) Diagrams of BP, CC, and MF results in GO analysis; (D) KEGG analysis diagram.

Full size image

GEPIA analysis, GEO data collection, and validation of expression differences, chromosome location, and correlation of SDECGs

Following the importation of the 61 common targets into the GEPIA website, 15 targets exhibited significant differential expression (P < 0.01) (Figs. 12 in Supplementary File). We combined three datasets (GSE45267, GSE62232, and GSE112790) for analysis and model construction. After removing the batch effect using the “ComBat” function, the combined dataset included 41, 10, and 15 normal samples, and 46, 80, and 183 hepatocellular carcinoma (HCC) samples from each respective dataset21. And GSE54236, which contained 80 normal and 81 HCC samples was used for external validation. A differential analysis between the HCC group and the normal group from the combined GEO dataset was performed. The results revealed the following genes were differentially expressed (adjP < 0.01): ANXA10, ATF3, BAX, BCL6, CDC25C, CDK1, CYP2C19, CYP3A4, CYP7A1, ESR1, FBP1, GADD45B, IMPDH2, S100P and THBS4. The aforementioned genes were identified as SDECGs, these core genes selected through human sample validation, and demonstrate high accuracy and clinical relevance. The genes BAX, CDC25C, CDK1, CYP7A1, IMPDH2, S100P and THBS4 exhibited high expression levels in HCC group, whereas the remaining genes showed low expression in HCC group, as illustrated in Figs. 3A and B. The expression patterns of these genes were consistent with the results obtained from GEPIA analysis. The exact chromosomal locations of the THX core genes are shown in Fig. 3C. As shown in Fig. 3D and E, the correlation analysis between each pair of SDECGs in the HCC samples demonstrated a robust, predominantly positive correlation between them.

Fig. 3
figure 3

Differential gene expression profiling and interaction analysis. (A) Box plot showing the expression differences of core genes between normal and HCC samples; (B) Heatmap of SDECG expression in normal and HCC samples; (C) Circular plot depicting the chromosomal locations of core genes; (D) Correlation network of SDECGs; (E) Correlation analysis between the two SDECGs. Tip: ** is p < 0.01; *** is p < 0.001.

Full size image

Analysis of immune cell infiltration, differences, and correlation between normal and HCC samples

To investigate the underlying mechanisms distinguishing HCC from normal tissues, immune cell infiltration was analyzed from multiple perspectives. Initially, immune cell infiltration analysis was conducted, followed by ssGSEA to identify significantly altered immune cells in both normal and HCC samples. The results revealed that monocytes were predominantly expressed in the normal group, while T cells (CD8), T cells regulatory(Tregs), M1 macrophages, and activated dendritic cells were significantly upregulated in the HCC group (Fig. 4A). Further correlation analysis of immune cells demonstrated significant associations between immune cell infiltration and SDECGs (P < 0.05) (Fig. 4B). Specifically, resting dendritic cells, M0 macrophages, M1 macrophages, resting mast cells, and activated CD4 memory T cells showed a primarily negative correlation with SDECGs. In contrast, activated dendritic cells, M2 macrophages, activated mast cells, monocytes, activated NK cells, follicular helper T cells, and regulatory T cells (Tregs) exhibited a positive correlation with SDECGs.

Fig. 4
figure 4

Immune cell infiltration analysis of SDECGs in HCC. (A) Box plot comparing immune cell fractions between normal and HCC samples; (B) Heatmap of the correlation analysis between SDECGs and immune cells. Tip: * is p < 0.05; ** is p < 0.01; *** is p < 0.001.

Full size image

Machine learning model for THX-induced HCC and feature gene selection with external validation

Four machine learning models—SVM, RF, XGB, GLM—were constructed using SDECGs data. Among these models, the SVM model demonstrated superior performance, as evidenced by its better ROC curves, residual box plots, and inverse cumulative distribution plots. Specifically, the SVM model exhibited the highest Area Under the Curve (AUC) and the lowest residuals and inverse cumulative values (Fig. 5A–C). Using the importance scores calculated from the SVM model, ten feature genes were identified from the SDECGs, including CYP2C19, CYP3A4, FBP1, THBS4, CYP7A1, ESR1, CDK1, BCL6, IMPDH2, and CDC25C (Fig. 5D). The top five feature genes were then used to construct nomograms, allowing for the derivation of individual score scales (Fig. 5E). These scores were employed to evaluate sensitivity to THX-induced HCC and predict the risk of developing HCC due to prolonged exposure to THX. Calibration curves and decision curves showed that the proximity of solid lines and dotted lines, as well as the distances between the red and grey lines, indicated a high degree of accuracy in the model’s predictions (Fig. 5F–G). The ROC curve further confirmed the model’s reliability, with an AUC of 0.951 (95% CI > 0.903) (Fig. 5H). To validate the model externally, the top five identified genes were incorporated into the SVM-based prediction model. The resulting ROC curve for the validation set demonstrated an AUC of 0.902, highlighting the model’s high accuracy when applied to the GEO dataset (Fig. 5I).

Fig. 5
figure 5

Machine learning for screening the key genes. (A) ROC curves for SVM, RF, XGB, GLM; (B) Box plots of residuals for SVM, RF, XGB, GLM; (C) Reverse cumulative distribution of residuals for SVM, RF, XGB, GLM machine learning models; (D) Bar plot showing the feature importance of SVM, RF, XGB, GLM machine learning models; (E) Nomogram was constructed based on five feature genes; (F) Calibration curve of the feature gene nomogram for THX-induced HCC; (G) Decision curve for the feature gene nomogram for THX-induced HCC; (H) ROC curve for the test GEO dataset; (I) ROC curves for external validation using the GEO dataset.

Full size image

Molecular Docking

In order to verify whether the proteins encoded by the top five THX feature genes are the optimal targets for inducing HCC, molecular docking was performed. The Autodock Vina docking results showed that the binding energy of the complexes formed by THX and the target proteins were (−6.2, −6.9, −6.8, −5.7, −6.0 kcal/mol) (Table 2), respectively, which were all less than -5 kcal/mol. These data suggested strong interactions between THX and the target proteins. Visualization with PyMOL revealed the formation of hydrogen bonds between the target proteins and THX which confirmed the stability of these interactions (Fig. 6).

Table 2 Results of molecular docking binding energy.
Full size table
Fig. 6
figure 6

The molecular docking results of THX with the top five genes. The molecular docking analysis of THX binding to CYP2C19 (A), CYP3A4 (B), FBP1 (C), THBS4 (D) and CYP7A1 (E) was conducted using AutoDock Vina software,

Full size image

Discussion

As a second-generation NEO featuring a mechanism analogous to its predecessor, THX has emerged as a staple in global agriculture—particularly across key farming nations including China, the United States, and numerous European countries—due to its exceptional efficacy, minimal toxicity, and broad-spectrum pesticidal activity. This versatility is further underscored by its dual application in seed treatment protocols and foliar spray formulations, enabling comprehensive pest management across diverse cropping systems such as grain, fruit, vegetables and so on24,25. The use of THX has also generated some controversy and challenges. In recent years, several countries and regions have implemented strict regulatory measures, or have limited its use on certain crops26. However, there is limited research on thiamethoxam’s potential to cause disease, particularly cancer. Accordingly, in this study, we first predicted the toxicity of commonly used neonicotinoid pesticides through the toxicology prediction platform. The findings of this study indicated that the carcinogenic potential and liver injury toxicity of THX were higher than those of the other four commonly used neonicotinoid pesticides. Subsequently, the mechanism of thiamethoxam-induced hepatocellular carcinoma was studied using network toxicology methods. Additionally, ACE was found to have the lowest carcinogenic potential, while CLO and DIN exhibited lower hepatotoxicity, providing valuable insights for future pesticide usage.

GO analyses of the 61 intersecting targets revealed key biological processes, including cellular response to xenobiotic stimulus, chemical stress, biotic stimulus, and toxic substances. The result of KEGG enrichment analysis mainly included cell cycle and Glycolysis / Gluconeogenesis. The development of HCC is associated with chronic liver diseases such as chronic hepatitis, cirrhosis, and liver fibrosis. Prolonged exposure to external stimuli, including viral infections and chemicals, can lead to chronic liver damage, increasing the risk of HCC27. The onset of HCC is often accompanied by dysregulation of the cell cycle. HCC cells frequently accelerate the cell cycle by modulating cyclin-dependent kinases (CDKs), promoting tumor proliferation and transformation28. The up-regulation of CDK1 in HCC samples, as observed in our differential gene analysis, lends further support to support this conclusion. As shown in Fig. 3E, there was a strong positive correlation between CDK1 and CDC25C. The metabolic pathways of glycolysis and gluconeogenesis are vital in the development and progression of hepatocellular carcinoma (HCC)29. Gluconeogenesis is the reverse process of glycolysis and is inhibited in tumorous liver30. In clinical studies, synthetic glucocorticoid drugs such as dexamethasone have been reported to have therapeutic effects on HCC. The underlying mechanism is to enhance the gluconeogenic pathway to inhibit tumor growth and angiogenesis31. This is consistent with the results obtained from our KEGG enrichment analysis.

We identified differentially expressed core genes through integrated analysis of GEPIA and GEO datasets, revealing a consistent trend of dysregulation in 15 candidate targets (SDECGs). Leveraging these data-driven observations, we developed a Support Vector Machine (SVM) model to prioritize key diagnostic features, which identified 10 high-impact genes (CYP2C19, CYP3A4, FBP1, THBS4, CYP7A1, ESR1, CDK1, BCL6, IMPDH2, and CDC25C). Based on their ranked importance in the SVM model, the top five genes were selected to construct a clinical nomogram. This data-derived diagnostic model for chronic THX exposure gains biological plausibility from the established roles of its key genes (e.g., CYP2C19, CYP3A4, CYP7A1, FBP1) in xenobiotic metabolism and cellular stress responses. Critically, these genes converge on liver metabolic pathways-a well-established hub influencing HCC pathogenesis32. Cytochrome P450 (CYP) is essential in the metabolism of drugs, chemicals, and endogenous compounds33. CYP2C19 is a drug-metabolizing enzyme, the metabolites produced during its metabolism may exert a significant influence on the liver’s microenvironment34. Abnormal CYP2C19 metabolism may contribute to the exacerbation of liver inflammation, thereby promoting the development of HCC35. CYP3A4 expression has been found to be significantly reduced in liver tumor tissues of patients with hepatocellular carcinoma36. Several studies have indicated that reduced CYP3A4 expression is associated with a poorer prognosis in patients with hepatocellular carcinoma (HCC)37. Dysregulation of bile acid metabolism is involved in the development of liver cancer38. Multiple studies have suggested that decreased CYP3A4 expression correlates with worse prognosis in patients with HCC, this enzyme is expressed only in the endoplasmic reticulum of hepatocytes and is regulated by the end product bile acids39. Studies have shown a correlation between elevated levels of major circulating bile acids and an increased risk of HCC, which is consistent with our results of increased CYP7A1 expression40. Fructose-1,6-bisphosphatase (FBP1) is a crucial enzyme in the process of gluconeogenesis41. Experimental studies have shown that low expression of FBP1 promotes HCC progression42. At the same time, loss of FBP1 disrupts liver metabolic homeostasis and promotes tumor progression43. This verified the accuracy of our KEGG analysis results. Studies have demonstrated that overexpression of THBS4 can mediate the proliferation and metastasis of HCC44. The expression of THBS4 have been linked to the invasion and metastasis of HCC45. Additionally, in the liver cancer microenvironment, THBS4 may promote tumor growth and metastasis by regulating intercellular interactions and angiogenesis46. The results of molecular docking showed that THX had a stable binding to the five signature genes, which revealed the accuracy of the above machine learning results and provided certain evidence for our study, and elucidated that THX-induced HCC may be mainly through the five signature genes.

It is noteworthy that considerable discrepancies in ssGSEA were discerned between the HCC and normal groups with respect to various immune cell types, including Monocytes, T cells CD8, Tregs, Macrophages M1 and Dendritic cells activated. Furthermore, correlation analysis revealed significant associations between some SDECGs and immune cell types. These include M1 macrophages, Tregs, Monocytes. M1 macrophages are known to promote the activation of T cells and NK cells by secreting pro-inflammatory cytokines and chemical factors, and have been shown to enhance tumour-specific immune responses. These cytokines can induce tumor cell apoptosis, inhibit tumor angiogenesis, and even enhance immune surveillance of tumor cells47,48. In the tumor immune microenvironment, a higher proportion of M1 macrophages is often correlated with improved prognosis. Additionally, monocyte count has been identified as an independent prognostic factor for overall survival (OS) in patients with HCC. In the HCC microenvironment, monocytes can be polarized into tumor-promoting macrophages (M2 macrophages). M2 macrophages foster tumor growth and metastasis by enhancing angiogenesis and fibrosis within the tumor microenvironment49. However, in our study, there was no significant difference in M2 macrophages, which may be related to the number of samples. In addition, monocytes can trigger autophagy, induce epithelial-mesenchymal transition (EMT) of cancer cells, and promote tumor metastasis50. Tregs are the key mediators of HCC-related immunosuppression. Tregs can help cancer cells escape immune surveillance by inhibiting anti-tumor immune responses51. Our completed molecular docking results further imply potential mechanistic intersections: for instance, if THX binds to key regulatory proteins associated with these immune cells (such as receptors or signaling molecules on M1 macrophages, Tregs, or monocytes), it might modulate their functions. Binding to M1 macrophage-related targets could either enhance their anti—tumor cytokine secretion or disrupt it, while interactions with Treg-associated targets might impair their immunosuppressive capacity. These potential bindings, if validated, would deepen our understanding of THX’s immunomodulatory role in HCC, aligning with KEGG identified IL-17 mediated immune microenvironment alterations and the observed immune cell discrepancies from ssGSEA.

This study elucidates the potential molecular mechanisms of THX-induced HCC using network toxicology and machine learning methods, providing support for disease diagnosis in populations exposed to THX or long-term use of this pesticide. However, it should be noted that this study is not without limitations. For example, this study solely investigates the molecular mechanisms underlying THX-induced HCC from a data analysis perspective. Therefore, additional experiments (such as animal studies) are required to validate these findings. Furthermore, due to the small number of clinical liver cancer samples, errors may have been introduced into the study results.

Conclusion

This study systematically investigated THX, a NEO with notable carcinogenic and hepatotoxic potential, by identifying 61 intersection genes between THX targets and hepatocellular carcinoma (HCC)-related genes. Through integrated bioinformatics analyses and machine learning modeling, five core feature genes (CYP2C19, CYP3A4, FBP1, THBS4, CYP7A1) were screened, and molecular docking confirmed strong binding affinities (≤ −5 kcal/mol) between THX and these genes. These findings provide mechanistic insights into THX-induced HCC, establishing a theoretical foundation for understanding its hepatotoxic and carcinogenic mechanisms.