Integrating single-cell transcriptomics and machine learning reveals 4-aminobiphenyl exposure signatures and novel diagnostic biomarkers in bladder cancer

Abstract

Background

Bladder cancer ranks among the top four most common malignancies in men worldwide. Despite therapeutic advancements, metastatic cases remain associated with dismal survival rates, highlighting the critical importance of studying environmental carcinogenic factors such as 4-aminobiphenyl (4-ABP). Classified as a Group 1 carcinogen by IARC, this compound—found in tobacco smoke and industrial chemicals—induces DNA damage through aryl amine metabolism. However, its cell-type-specific molecular mechanisms in bladder carcinogenesis remain poorly characterized, particularly at single-cell resolution.

Methods

Our study combined single-cell RNA sequencing data (n = 95,136 cells) with publicly available bulk transcriptomic datasets from the GEO repository. The computational workflow incorporated Harmony algorithm for batch effect correction, WGCNA for co-expression network construction, and a machine learning framework utilizing LASSO, SVM, and Random Forest algorithms for biomarker identification. Additionally, we performed molecular docking simulations to investigate 4-ABP-protein interactions and employed ssGSEA to characterize immune cell infiltration patterns in the tumor microenvironment.

Results

Single-cell profiling uncovered distinct fibroblast and mast cell subpopulations exhibiting significant 4-ABP-associated transcriptional signatures (adjusted p = 2.22 × 10⁻¹⁵). Cross-platform integration revealed 15 functionally conserved genes significantly enriched in IL-4-mediated signaling pathways (false discovery rate < 5%). Through machine learning-based feature selection, we established a diagnostic panel comprising six core genes (EIF4G2, CA2, CDKN2A, HSP90B1, GOT2, and IL4R), with the random forest classifier demonstrating optimal discriminatory performance (area under curve = 1.00). Structural modeling predicted high-affinity interactions between 4-ABP and both HSP90B1 (binding energy: − 6.7 kcal/mol) and IL4R (− 6.0 kcal/mol). Unsupervised clustering delineated two clinically relevant molecular subtypes characterized by divergent immune microenvironment compositions.

Conclusion

Through integrated multi-omics analyses, this investigation systematically elucidates 4-ABP’s carcinogenic mechanisms while identifying clinically actionable biomarkers and molecular targets for precision medicine applications in bladder cancer prevention and treatment. These findings provide critical insights for developing targeted strategies to mitigate environmental carcinogenesis in susceptible populations. However, this study is limited by its retrospective nature and reliance on public sequencing data, which restricted access to detailed clinical metadata and precluded survival analysis for the identified subtypes.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12672-025-03819-y.

Introduction

Bladder cancer ranked fourth in incidence and mortality among male malignancies, accounting for approximately 6% of new cancer diagnoses and 4% of cancer-associated deaths [1]. This malignancy poses a significant clinical burden due to its high recurrence rates and resistance to therapy, particularly in advanced stages [2]. While surgical resection and immunotherapy have improved outcomes, the five-year survival rate for distant metastatic disease remains below 10% [3], underscoring the need for better early detection strategies and mechanistic insights into environmental drivers of carcinogenesis. Established risk factors such as tobacco smoke and occupational chemical exposure account for ~ 50% of cases, leaving a substantial proportion of etiology unexplained [4]. Growing evidence implicates environmental aromatic amines—notably 4-aminobiphenyl (4-ABP)—as understudied contributors to bladder cancer development [5].

4-Aminobiphenyl (4-ABP), also known as 4-Biphenylamine, a tobacco smoke constituent and industrial byproduct, is classified as a Group 1 human carcinogen by IARC due to its DNA-damaging properties [6]. Metabolic activation of 4-ABP generates reactive intermediates that form mutagenic DNA adducts, preferentially targeting urothelial cells and inducing oncogenic mutations (e.g., TP53) [7, 8]. Epidemiological studies report a two to threefold increased bladder cancer risk in occupationally exposed populations (e.g., dye workers, rubber manufacturers) [9]. Preclinical models further demonstrate that 4-ABP promotes tumorigenesis through oxidative stress, inflammatory pathway activation [10, 11]. However, its cell-type-specific effects within the bladder tumor microenvironment (TME) and potential role in shaping molecular subtypes remain uncharacterized.

This investigation carries significant scientific and clinical importance, given 4-ABP’s confirmed DNA-damaging properties and established association with bladder cancer development. By systematically examining how 4-ABP interacts with bladder cells at the molecular level, we aim to clarify its specific contributions to cancer formation and advancement. Such understanding could uncover previously unrecognized pathways involved in chemical-induced tumor development. Importantly, these mechanistic insights may point to new opportunities for both preventing and treating this malignancy. The findings promise to fill existing knowledge gaps while offering practical applications for protecting individuals regularly exposed to this cancer-causing agent through their work or environment.

Materials and methods

Figure 1.

Fig. 1.

The flowchart of the research

Acquisition of transcriptomic data and 4-ABP molecular targets

Here, we collected data from the GEO datasets GSE3167, GSE13507, GSE31684, GSE48075 and GSE48276. By merging these datasets, we obtained a new combined dataset, including 76 normal samples and 865 bladder cancer patients. These data were used to analyze bladder cancer-related gene expression patterns and differences. We obtained the SMILES number of 4-ABP from the PubChem (https://pubchem.ncbi.nlm.nih.gov/) database and predicted its molecular targets using the SwissTargetPrediction (http://www.swisstargetprediction.ch/) and ChEMBL (https://www.ebi.ac.uk/chembl/) databases. These targets are proteins or genes that 4-ABP may interact with. Finally, we removed duplicates and merged the 4-ABP molecular targets to obtain a more accurate set of targets. Single-cell RNA sequencing (scRNA-seq) data of bladder cancer were obtained from the Gene Expression Omnibus (GEO) database under accession number GSE222315. The dataset comprises transcriptomic profiles from 4 normal bladder tissue samples and 9 tumor tissue samples derived from patients with bladder carcinoma. This study utilized exclusively de-identified, publicly available data from the GEO database. According to our institutional guidelines and national regulations, research based on such pre-existing, anonymized public data does not require separate ethical approval or informed consent. We confirm that all original studies from which these data were derived were conducted in accordance with the ethical standards of their respective institutions and the Declaration of Helsinki.

Single-cell RNA sequencing data processing and analysis

The scRNA-seq dataset was processed using Seurat v4 following established computational pipelines. Initial quality control measures retained cells expressing 200–7500 genes while excluding those with mitochondrial gene content exceeding 15%. Following quality filtering, raw counts were normalized using the LogNormalize method (scale factor = 10,000) and the top 2000 highly variable genes (HVGs) were selected through variance stabilization transformation to mitigate technical variability. To address potential confounding factors, batch effects were corrected using Harmony integration and cell cycle effects were regressed out based on phase-specific scoring. Dimensional reduction was performed using principal component analysis (25 PCs), followed by graph-based clustering at a resolution of 0.05, which identified 9 transcriptionally distinct cell populations. These clusters were annotated based on established marker genes: T cells (CD3D, CD3G, CD3E), epithelial cells (EPCAM, KRT18, KRT19), fibroblasts (FBLN1, LUM, DCN), macrophages (APOE, C1QB, C1QA), mast cells (IGHG1, IGLC2, IGKC), endothelial cells (VWF, CLDN5, EGFL7), smooth muscle cells (MYL9, TAGLN, ACTA2), and B cells (CD37, MS4A1, CD79A). Subsequent analysis employed AddModuleScore to evaluate enrichment patterns of 4-ABP-associated gene signatures across cell types, revealing significant differences between tumor and normal samples (Wilcoxon rank-sum test, p < 0.05). Single-cell populations were stratified into high- and low-scoring groups according to median expression values of the target signature. Subsequently, differential gene expression analysis was performed using the FindMarkers function (Seurat v4.3.0) with stringent thresholds (absolute log2 fold-change > 0.1 and adjusted p-value < 0.05). This analytical approach enabled identification of transcriptionally distinct features between scoring-defined subpopulations while controlling for multiple testing effects through Bonferroni correction. Cell-cell communication networks were reconstructed using CellChat, demonstrating distinct interaction patterns among identified populations. Finally, differential expression analysis of network hub genes highlighted patient-specific variations in intercellular signaling pathways. All statistical analyses were performed in R (v4.2.2) with visualization implemented using ggplot2 and patchwork packages.

Differential gene analysis in GEO and target intersection analysis

The present study implemented an integrated bioinformatics workflow to delineate molecular perturbations in bladder carcinoma. We conducted differential expression analysis with the limma package in R and integrated surrogate variable analysis (SVA) to adjust for batch effects across datasets. Subsequently, we applied linear modeling via the lmFit function and performed empirical Bayes moderation using eBayes to improve detection sensitivity. To identify significant transcriptional alterations between malignant and normal urothelial specimens, stringent thresholds were employed (adjusted p-value < 0.05 and |log₂ fold-change| >0.2). These differentially expressed genes (DEGs) were then subjected to hierarchical clustering with the pheatmap package and visualized through volcano plots. Furthermore, to assess pathway-level dysregulation, GSVA enrichment scores for the 4-ABP gene set were calculated and compared across clinical subgroups. For co-expression network construction, we first clustered samples by hierarchical clustering with average linkage and excluded outliers. An optimal soft-thresholding power was determined to meet scale-free topology criteria, after which module detection was carried out through dynamic tree cutting. Among the identified modules, four clinically significant ones—blue, green, magenta, and turquoise—exhibited robust correlations with pathological phenotypes and were retained for functional annotation. All analytical procedures were executed in R , and results were visualized using both ggplot2 and WGCNA.

GO and KEGG analysis

To systematically delineate core molecular signatures, we integrated and analyzed four distinct gene sets, including WGCNA module genes, GEO-DEGs, single-cell-based median-split differential genes, and 4-ABP targets. Using the VennDiagram package, we identified consensus genes shared among these datasets and visualized the intersections through four-way Venn diagrams. Subsequently, we conducted functional annotation of these overlapping genes by performing Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analyses with the clusterProfiler tool. The GO analysis comprehensively assessed three ontological categories: biological processes such as cell proliferation and immune regulation, cellular components including membrane structures and organelles, and molecular functions like catalytic and binding activities. In parallel, KEGG pathway analysis revealed significantly enriched signaling pathways (adjusted p < 0.05, q < 0.05) potentially involved in bladder carcinogenesis. We visualized the enrichment outcomes using bubble plots generated by ggplot2, wherein the x-axis indicated functional terms, the y-axis reflected gene counts, and bubble properties—size and color intensity—represented statistical significance and enrichment magnitude. Finally, to evaluate expression differences of the consensus genes between bladder cancer patients and normal controls, Wilcoxon rank-sum tests were applied, followed by multiple testing correction using the Benjamini-Hochberg method.

Machine learning for key gene screening

Initially, we conducted least absolute shrinkage and selection operator (LASSO) regression with the glmnet package (4.1–8), employing cross-validation to identify the optimal regularization parameter λ for feature selection. This step facilitated the screening of biologically relevant genes under constrained conditions. Subsequently, support vector machine (SVM) modeling was carried out using the e1071 package (1.7–16). A model incorporating 14 discriminative genes was established, achieving a predictive accuracy of 0.959. In parallel, we implemented a random forest algorithm via the randomForest package (4.7–1.6) to further evaluate feature importance. Genes with an importance score exceeding 9 were retained for downstream analysis. Finally, to consolidate the feature selection outcomes derived from these complementary machine learning approaches, we extracted the intersecting gene set using the VENN tool (1.12). This integrative strategy allowed us to identify a robust and consensus set of key genes with high confidence.

SHAP model construction

To elucidate the functional mechanisms of key genes in bladder cancer pathogenesis, we developed a SHapley Additive exPlanations (SHAP) interpretability model using a suite of R packages, including kernelshap (0.7.0), ggplot2 (3.5.2), ranger (0.17.2), and shapviz (0.9.7). This framework quantifies the contribution of individual genes to predictive outcomes through the computation of SHAP values, thereby enhancing the interpretability of model decisions. As an initial step, the GEO dataset was partitioned randomly into training and validation subsets at a ratio of 7:3. Using the training subset, a Random Forest model was constructed with the ranger package. Subsequently, SHAP values for each feature were derived via the kernelshap package, which provides an efficient implementation of the Kernel SHAP algorithm for attributing predictive influence to input variables. To ensure computational accuracy and robustness, the background dataset size was fixed at 200 samples, and the exact calculation mode was employed during SHAP value estimation. This configuration facilitates a reliable assessment of gene-wise importance within the model. Finally, multiple visualization methods were generated to interpret the model’s behavior, including ROC curves, beeswarm plots, feature dependence plots, and feature contribution waterfall diagrams, each serving to illustrate different aspects of model interpretability and gene influence.

Immune cell infiltration analysis

To systematically characterize immune cell infiltration patterns in bladder cancer, single-sample gene set enrichment analysis (ssGSEA) was implemented using the GSVA package (v1.46.0) in R. This algorithm enables quantification of immune cell abundance in individual tumor samples by calculating enrichment scores for predefined immune gene signatures. Transcriptomic data from GEO datasets were processed through the limma pipeline (v3.54.0) for normalization and quality control. Subsequently, correlation analysis between key genes and immune infiltration scores was performed, with results visualized using ggplot2 to generate comprehensive association heatmaps. All statistical analyses were conducted with FDR correction (q-value < 0.05) to account for multiple hypothesis testing.

Consensus clustering analysis

To delineate transcriptionally distinct molecular subtypes associated with hub gene expression in the GEO cohort, we employed consensus clustering via the “ConsensusClusterPlus” R package (version 1.60.0). This unsupervised learning procedure was repeated across 100 iterations to reinforce classification reliability. The optimal number of clusters, k, was determined through a comprehensive assessment of consensus matrices alongside analytical evaluation of cumulative distribution function (CDF) profiles. Upon subtype establishment, we utilized principal component analysis (PCA) to graphically represent transcriptional heterogeneity among the identified subgroups. Subsequent comparative evaluations encompassed two major aspects: first, immune infiltration patterns were assessed using ssGSEA; second, subtype-specific biological mechanisms were annotated via GO enrichment profiling. All statistical computations and visualizations were conducted in R , with an adjusted p-value threshold of FDR < 0.05 deemed statistically significant.

Molecular docking

To investigate the binding capacity of 4-ABP with key genes, we obtained the corresponding protein molecular structures of these genes from the PDB database in this study. When selecting protein structures, we prioritized X-ray crystallography structures with higher resolution, as they provide more accurate atomic position information. After obtaining the protein structures, we visualized the protein molecular structures using PyMOL software. Subsequently, we performed molecular docking experiments using AutoDock software. In the molecular docking experiments, we employed a binding energy threshold of − 5.0 kcal/mol—a widely recognized benchmark in computational docking studies—to identify ligand-receptor complexes exhibiting strong and potentially biologically meaningful interactions. Finally, we obtained the binding energies of the key genes with 4-ABP, identifying key proteins that can bind 4-ABP effectively.

Statistical analysis

Statistical analyses in this work were carried out predominantly within the R environment, supported by a range of bioinformatics tools. Initially, differential gene expression analysis was performed with the limma package to identify statistically significant genes. Subsequently, functional profiling of these genes was conducted through GO and KEGG enrichment analyses implemented in the clusterProfiler package. To further refine feature selection and evaluate predictive importance, we applied multiple machine learning algorithms, including random forests and extreme gradient boosting, facilitated by the randomForest, xgboost, and caret packages. The contribution of each key gene to model predictions was quantitatively interpreted using SHAP. In addition, immune infiltration levels were estimated via ssGSEA. Lastly, molecular docking simulations with AutoDock were employed to assess the binding potential between identified key genes and the ligand 4-ABP.

Result

Single-cell transcriptomic profiling reveals cell-type specific 4-BPA signature in bladder cancer

Following rigorous quality control, high-quality transcriptomic data were obtained from 95,136 cells (Fig. 2A). Batch effects were corrected using Harmony integration, revealing nine distinct clusters through UMAP visualization. Systematic cell annotation identified these populations as epithelial cells, endothelial cells, fibroblasts, macrophages, plasma cells, T cells, smooth muscle cells, B cells, and NK cells, based on canonical marker expression (Fig. 2B). Dot plot analysis demonstrated cell-type specific expression patterns of these markers across clusters (Fig. 2C). Comparative analysis of 4-ABP signature scores showed significant elevation in bladder cancer patients versus controls (Wilcoxon rank-sum test, p < 2.22e−15; Fig. 2D), with fibroblast and mast cell populations exhibiting particularly strong enrichment (Figs. 2E, F). Differential expression analysis between high- and low-scoring cells identified 2022 significantly dysregulated genes (absolute log2FC > 0.1, FDR < 0.05), suggesting potential mechanistic links between 4-ABP exposure and bladder carcinogenesis.

Fig. 2.

Screening of 4-ABP-related genes in bladder cancer single-cell data. A UMAP plot of single-cell clusters. B UMAP plot of annotated cell subpopulations. C Bubble plot showing marker genes for distinct cell subpopulations. D Violin plot comparing 4-ABP scores between bladder cancer and normal patients. E Dot plot of 4-ABP scores across cell subpopulations. F UMAP visualization of 4-ABP scores by patient group

Integrated transcriptomic analysis reveals 4-ABP-associated molecular signatures in bladder cancer

Differential gene expression analysis was performed using limma on bladder cancer and control samples from GEO, identifying 1,114 downregulated and 817 upregulated genes (absolute log2FC > 0.2, p < 0.05), with results visualized through hierarchical clustering heatmaps and volcano plots (Fig. 3A, B). Subsequent GSVA revealed significantly elevated 4-ABP pathway activity in tumor samples (Wilcoxon p < 0.001; Fig. 3C). Weighted gene co-expression network analysis (WGCNA) was implemented with the following parameters: (1) sample clustering threshold (r > 0.85), (2) soft-thresholding power (β = 8) confirming scale-free topology (Fig. 3D), and (3) module-trait correlation identifying four clinically relevant modules (blue: r = − 0.64, p = 8e−88; green: r = − 0.6, p = 3e-74; magenta: r = 0.41, p = 1e−31; turquoise: r = − 0.71, p = 2e−118) as 4_ABP-associated gene clusters (Fig. 3E). All analyses were conducted in R with multiple testing correction using the Benjamini-Hochberg method.

Fig. 3.

Transcriptase screening for genes related to 4-ABP. A Heatmap of differentially expressed genes (DEGs) from the GEO database. B Volcano plot of bladder cancer differential gene expression analysis (GEO dataset). C Boxplot of GSVA-calculated 4-ABP scores comparing bladder cancer vs. control groups. D WGCNA scale independence analysis plot. E. Heatmap of module-trait correlations between WGCNA modules and 4-ABP-related scores

Integrative functional analysis of consensus genes in bladder carcinogenesis

Venn analysis identified 15 consensus genes intersecting across four datasets: GEO-derived differentially expressed genes, single-cell differential expression results, WGCNA module genes, and 4-ABP targets (Fig. 4A). Functional enrichment analysis using clusterProfiler revealed these genes were significantly associated with interleukin-4 response and peroxisomal membrane organization in biological process ontology (Fig. 4B). KEGG pathway analysis demonstrated predominant involvement in pancreatic secretion, renal bicarbonate reclamation, and endoplasmic reticulum protein processing (Fig. 4C). Comparative expression analysis showed significant dysregulation of all 15 genes in bladder cancer versus controls (Wilcoxon p < 0.05 after Benjamini-Hochberg correction), with 4 genes upregulated and 12 downregulated in tumor samples (Fig. 4D). These findings suggest 4-ABP may promote bladder carcinogenesis through modulation of interleukin signaling and cellular organelle functions.

Fig. 4.

Screening of multiple genes. A Venn diagram of 4-ABP-associated genes, single-cell markers, GEO DEGs, and WGCNA module genes. B GO enrichment analysis of the 15 overlapping genes. C Bar plot of KEGG pathway enrichment for the 15 candidate genes. D Boxplot of differential expression for the 15 genes in the GEO dataset

Machine learning identifies six key genes

We screened the remaining genes using LASSO regression analysis (Fig. 5A). The Support Vector Machine (SVM) analysis showed that when the number of genes was 2–7, the model’s accuracy was as high as 0.959. We selected the model with 14 genes for subsequent analysis (Figs. 5B, C). Subsequently, The Random Forest analysis results indicated that EIF4G2, ATG4B, CDKN2A, CA2, IL4R, GOT2, and HSP90B1 had the highest importance. We selected genes with importance greater than 9 for further analysis (Figs. 5D, E). Finally, by combining the results of the three machine learning methods, we obtained six key genes: EIF4G2, CA2, CDKN2A, HSP90B1, GOT2, and IL4R (Fig. 5F).

Fig. 5.

Model gene screening. A LASSO Regression Analysis. The coefficient shrinkage paths are shown with the log-transformed regularization parameter (λ) on the x-axis and standardized regression coefficients on the y-axis. Variable selection was performed through 10-fold cross-validation to identify optimal λ. B, C Support Vector Machine (SVM) Performance. B Model accuracy (y-axis) versus number of selected genes (x-axis). C Receiver operating characteristic (ROC) curve for the optimal 15-gene SVM model. D Random Forest Feature Importance. Genes were ranked by mean decrease in Gini index, with E highlighting the 9 most influential features (importance score > 9). F Integrative Analysis of Machine Learning Approaches. Venn diagram illustrates consensus genes identified by all three methods (LASSO, SVM, and random forest), revealing six core biomarkers: EIF4G2, CA2, CDKN2A, HSP90B1, GOT2, and IL4R

Machine learning-based evaluation of diagnostic biomarkers in bladder cancer

Six candidate genes (EIF4G2, CA2, CDKN2A, HSP90B1, GOT2, and IL4R) were subjected to comprehensive SHAP analysis for diagnostic model development. The GEO dataset was randomly partitioned into training (70%) and validation (30%) cohorts, with ten distinct machine learning algorithms evaluated for predictive performance. Among these, the random forest (RF) model demonstrated optimal diagnostic accuracy (AUC = 1.00), prompting its selection for subsequent analyses. Feature importance assessment revealed EIF4G2 as the most influential predictor, as evidenced by both bar plot ranking (Fig. 6A) and beeswarm plot distribution (Fig. 6B). Partial dependence analysis indicated a positive correlation between CA2 expression levels and SHAP values (Fig. 6C), a finding further corroborated by waterfall plot visualization of individual feature contributions (Fig. 6D). Model performance metrics, including ROC curves, recall (sensitivity), precision, and F1 scores, were systematically compared across all algorithms in both training and validation sets (Fig. 6E, F), confirming the superior robustness of the RF approach.

Fig. 6.

SHAP analysis of candidate biomarkers. A Feature importance bar plot. The bar plot quantifies the relative importance scores of six candidate genes (EIF4G2, CA2, CDKN2A, HSP90B1, GOT2, and IL4R) identified through integrative machine learning approaches. B Beeswarm Plot of SHAP Values. Each point represents the SHAP value for an individual sample, with color intensity indicating feature importance magnitude. The plot confirms EIF4G2 and CA2 as dominant predictors. C Partial Dependence Plots. Nonlinear relationships between gene expression levels (x-axis) and their SHAP value contributions (y-axis) are shown. All six genes exhibit monotonic positive correlations, suggesting dose-dependent effects on the predictive model. D SHAP Waterfall Plot. Illustrates instance-specific feature contributions, ranked by absolute SHAP values. EIF4G2 and CA2 consistently appear in the top. E, F Model Performance Metrics. ROC curves and classification metrics (recall, accuracy, F1-score) for both training (E) and validation (F) datasets demonstrate robust predictive capacity

Cell-type specific expression patterns and intercellular communication analysis

Dot plot visualization revealed distinct expression profiles of hub genes across cellular subpopulations, with EIF4G2 and CA2 predominantly expressed in endothelial cells, while CDKN2A showed malignant cell-specific expression (Fig. 7A). Comparative analysis demonstrated significant upregulation of all hub genes in bladder cancer versus controls (Wilcoxon rank-sum test, FDR < 0.01; Figs. 7B–G). CellChat analysis of intercellular communication networks identified enhanced signaling strength and interaction complexity in high-scoring populations (Fig. 7H). Notably, fibroblasts exhibited the most extensive communication networks, particularly with immune cell subsets (Fig. 7I).

Fig. 7.

The hub gene in the single-cell expression pattern. A Dot plot showing hub gene expression across different cell types. B–G Violin plots displaying differential expression analysis of CA2, CDKN2A, EIF4G2, IL4R, HSP90B1, and GOT2 between bladder cancer and control groups. H Bar plot comparing the number and strength of cell–cell communication events in bladder cancer versus normal samples at single-cell resolution. I Heatmap visualizing the quantitative differences in cell-cell communication patterns between bladder cancer and normal patients

Immune cell infiltration and significant correlations with key genes

In this study, immune cell infiltration in bladder carcinoma patients and healthy controls from the GEO database was evaluated using ssGSEA. Notably, significant disparities in immune cell composition were observed between the two cohorts (P < 0.05). Specifically, tumor tissues exhibited elevated infiltration levels of activated CD4+ T cells, activated CD8+ T cells, memory CD8+ T cells, type 2 T helper (Th2) cells, and natural killer (NK) cells (Fig. 8 A). Subsequently, correlation analyses were performed to assess associations between six pivotal genes (EIF4G2, CA2, CDKN2A, HSP90B1, GOT2, and IL4R) and immune cell subsets. Intriguingly, EIF4G2 demonstrated robust correlations with activated B cells, effector memory CD8 + T cells, immature B cells, NK cells, Th17 cells, Th1 cells, myeloid-derived suppressor cells (MDSCs), and central memory CD8+ T cells (Fig. 8B, P < 0.001). Similarly, CA2 was significantly linked to activated CD4+ T cells, Th2 cells, effector memory CD4+ T cells, and memory B cells (Fig. 8C, P < 0.001). Furthermore, CDKN2A, HSP90B1, GOT2, and IL4R displayed pronounced associations with CD56bright NK cells, activated CD4+ T cells, and γδ T cells (Figs. 8D–G, P < 0.001).

Fig. 8.

Immune cell infiltration analysis in bladder cancer. A Differential immune infiltration landscape. The bar plot compares immune cell infiltration levels (y-axis) between bladder cancer patients and controls in the GEO dataset, stratified by immune cell subtypes (x-axis). Significant elevations were observed in bladder cancer samples for: Activated CD4+ T cells, Activated CD8+ T cells, Memory CD8+ T cells, Th2 cells, Natural killer cells (p < 0.001). B–G Immune-Gene Correlation Networks. Scatterplots demonstrate significant associations (Pearson’s r) between expression levels of: EIF4G2 (B), CA2 (C), CDKN2A (D), HSP90B1 (E), GOT2 (F) and IL4R (G) and infiltrating immune cell abundances. Asterisks denote statistically robust correlations (***p < 0.001; **p < 0.01)

Two subtypes classified based on hub genes for bladder cancer

To further elucidate molecular heterogeneity, consensus clustering analysis was performed on the GEO training cohort of bladder carcinoma patients based on six candidate genes (EIF4G2, CA2, CDKN2A, HSP90B1, GOT2, and IL4R). Notably, two distinct molecular subtypes (C1 and C2) were identified (Figs. 9A, B and Fig S1). Subsequent analysis revealed differential expression patterns of these genes between subtypes, with CA2 and CDKN2A exhibiting significantly higher expression in the C2 subgroup (Fig. 9C). Moreover, comparative immune profiling demonstrated marked disparities in immune cell infiltration between the two subtypes, with the majority of immune subsets showing statistically significant variations (Fig. 9D). To investigate functional implications, Gene Set Variation Analysis (GSVA) was conducted, revealing pronounced enrichment of upregulated genes in oncogenic pathways (Fig. 9E). Finally, PCA confirmed clear stratification of bladder carcinoma patients into the two predefined subtypes (Fig. 9F), underscoring the robustness of the classification.

Fig. 9.

Consistency clustering analysis of the hub gene in bladder cancer. A, B Bladder cancer patients were stratified into two distinct molecular subtypes (C1 and C2) using ConsensusClusterPlus. C Boxplot showing differential expression analysis of the 6 hub genes (CA2, CDKN2A, EIF4G2, IL4R, HSP90B1, and GOT2) between C1 and C2 subtypes. D Boxplot comparing the infiltration levels of 28 immune cell types between C1 and C2 subtypes. E GSVA enrichment analysis of biological pathways in C1 versus C2 subtypes. F PCA plot demonstrating the distribution of patients across C1 and C2 subtypes

Strong binding affinity of 4-ABP with key proteins

We investigated the binding capacity of 4-ABP with the three key gene proteins (EIF4G2, CA2, CDKN2A, HSP90B1, GOT2, and IL4R) through molecular docking experiments. The results showed that the binding energies of 4-ABP with CA2, CDKN2A, EIF4G2, GOT2, HSP90B1, and IL4R were − 6.0, − 6.3, − 6.1, − 6.2, − 6.7, and − 6.0 kcal/mol, respectively, indicating that 4-ABP can bind well with these proteins (Figs. 10A–F).

Fig. 10.

Molecular Docking Analysis of 4-ABP with Candidate Biomarkers. A–F Predicted Ligand-Protein Interaction Patterns. The panels illustrate computational docking models of 4-ABP (carcinogenic metabolite) with six putative molecular targets

Discussion

Bladder cancer represents a significant global health burden [12], with growing evidence implicating environmental carcinogens in its pathogenesis [13]. 4-ABP, a well-characterized aromatic amine, has been strongly associated with bladder cancer development in epidemiological and experimental studies [14–16]. Our comprehensive study provides novel mechanistic insights into 4-ABP -induced bladder carcinogenesis through an integrated multi-omics approach combining single-cell transcriptomics, machine learning algorithms, and molecular docking simulations. The key findings demonstrate cell-type specific vulnerability to 4-ABP exposure, particularly in fibroblast and mast cell populations, while identifying six robust molecular signatures (EIF4G2, CA2, CDKN2A, HSP90B1, GOT2, and IL4R) with exceptional diagnostic performance (AUC = 1.00). Importantly, molecular docking revealed strong binding affinities between 4-ABP and these key proteins (particularly HSP90B1 at − 6.7 kcal/mol and IL4R at − 6.0 kcal/mol), suggesting direct molecular interactions that may underlie 4-ABP’s carcinogenic potential. These findings significantly advance our understanding of environmental bladder carcinogenesis by bridging the gap between population-level epidemiological observations and molecular-level mechanistic insights.

The single-cell analysis demonstrated that 4-ABP signatures were particularly enriched in fibroblast and mast cell populations, suggesting these cell types may play crucial roles in mediating 4-ABP’s carcinogenic effects. This cell-type specificity aligns with emerging evidence that environmental carcinogens can selectively target specific cellular components of the tumor microenvironment [17]. The identification of 15 consensus genes involved in IL-4 signaling and peroxisome pathways provides mechanistic insights into how 4-ABP may disrupt critical cellular processes in bladder tissue. Notably, the IL4R receptor showed strong molecular binding with 4-ABP(− 6.0 kcal/mol), suggesting potential JAK–STAT pathway activation leading to tumor immune evasion [18], while abnormal peroxisome-related genes indicated possible carcinogenic mechanisms through oxidative stress [19] or lipid metabolism dysregulation [20]. These findings not only provide novel evidence for environmental carcinogen-mediated interference with cytokine networks and organelle function, but also offer potential targets for precision prevention in occupationally exposed populations.

Through comprehensive machine learning-based feature selection and validation, we robustly identified six pivotal molecular drivers (EIF4G2, CA2, CDKN2A, HSP90B1, GOT2, and IL4R) demonstrating exceptional diagnostic capability (AUC = 1.00) for 4-ABP-associated bladder cancer. Of particular mechanistic significance, EIF4G2 was established as the predominant biomarker, a finding that aligns with and extends current understanding of its oncogenic role in dysregulated translation initiation and malignant progression [21, 22]. Molecular docking analyses revealed remarkably strong binding interactions between 4-ABP and these molecular targets, with particularly notable affinities observed for HSP90B1 (− 6.7 kcal/mol) and CDKN2A (− 6.3 kcal/mol). These structural insights provide compelling evidence that 4-ABP may exert its carcinogenic effects through direct perturbation of these critical protein networks. Our findings significantly advance the mechanistic paradigm by demonstrating that environmental carcinogens like 4-ABP can selectively target and functionally compromise specific oncogenic pathways through direct molecular interactions, as substantiated by emerging literature on xenobiotic-protein interactions in carcinogenesis [23–25].

Our comprehensive single-cell analysis reveals the complex cellular dynamics underlying 4-ABP-associated bladder carcinogenesis. The distinct cell-type specific expression patterns of hub genes—particularly endothelial-enriched EIF4G2/CA2 and tumor-specific CDKN2A—corroborate recent findings by Toxicol Pathol on chemical carcinogen-induced vascular niche remodeling [26]. The discovery of two molecular subtypes with divergent immune profiles, especially the C2 subtype characterized by elevated CA2/CDKN2A expression and robust immune infiltration, provides critical insights into 4-ABP’s heterogeneous effects. This aligns with emerging evidence that molecular subtypes can predict immunotherapy response in bladder cancer [27–29], while extending these observations to environmental carcinogen-specific contexts.

The enhanced intercellular communication networks, particularly fibroblast-immune cell interactions, support the paradigm of cancer-associated fibroblasts (CAFs) as central immune regulators [30, 31]. Notably, our immune profiling reveals a paradoxical co-enrichment of effector (activated T/NK cells) and immunosuppressive (Th2/MDSCs) populations, mirroring the “immune checkpoint blockade resistance” phenotype in smoking-related cancers [32]. The strong correlations between hub genes and specific immune subsets, suggest 4-ABP modulates anti-tumor immunity through these molecular targets. These findings collectively demonstrate that 4-ABP promotes bladder cancer through coordinated multi-compartment effects, creating a unique tumor ecosystem that may require subtype-specific therapeutic strategies.

Our comprehensive analysis has uncovered new biological markers and promising treatment possibilities for bladder cancer linked to 4-ABP exposure. Beyond revealing how this environmental toxin triggers cancer development, these discoveries pave the way for tailored prevention approaches, particularly for workers and others facing regular exposure to this chemical.

Our findings shed new light on 4-ABP’s cancer-causing mechanisms, but we should consider some important caveats. The computer-based predictions, while promising, need confirmation through lab experiments using cell cultures and animal models. We mainly looked at gene activity changes in this study—checking protein levels would give us a more complete picture of what’s happening in cells. Most importantly, we’ll need follow-up studies with patients to determine how these discoveries can actually help in real-world cancer prevention and treatment.

Conclusion

Our comprehensive study paints a detailed picture of how 4-ABP drives bladder cancer development. We’ve pinpointed six crucial molecules (EIF4G2, CA2, CDKN2A, HSP90B1, GOT2, and IL4R) that show perfect accuracy in detecting this cancer type. These molecules appear to work through IL-4 signaling and peroxisome-related processes. The strong chemical attraction we observed between 4-ABP and two of these molecules (HSP90B1 and IL4R) helps explain how this environmental toxin directly contributes to cancer formation. Equally important, we found clear differences in the cellular environment of tumors exposed to 4-ABP. These discoveries give us new tools to protect workers and others regularly exposed to this cancer-causing chemical.

Author contributions

Yangyang Xu. and Xiujia Wang. : investigation, methodology, validation, writing, original draft. Xiuwang Wei.: funding acquisition, writing. review and editing.Xinxin Wang.: review and editing. Kaiqiang Li.: resources, supervision.

Funding

This study was partially funded by the The Development, Promotion, and Application Project of Appropriate Medical and Healthcare Technologies of Guangxi. (Grant No.S2022022).

Data availability

All data generated or analyzed during this study can be obtained directly by contacting the corresponding author or https://github.com/xyyko123/BCA_code.git.

Code availability

Further enquiries can be directed to the corresponding author.

Declarations

Ethics approval and consent to participate

The data used in this study were obtained from public databases, therefore no additional ethical certification was required.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.