Integrated single-cell genomics, transcriptomics, and pathomics to identify potential biomarkers in muscle-invasive bladder cancer

Abstract

Background

The integration of multi-omics approaches provides a powerful strategy for understanding cancer. By combining these methods, researchers gain insights into tumor diversity, gene activity, and the tumor microenvironment, which are essential for advancing cancer biology, improving early detection, refining prognostic tools, and developing targeted treatments. This study aims to explore key biomarkers in muscle-invasive bladder cancer (MIBC) and to develop a predictive model for better understanding disease progression and therapeutic responses

Methods

Single-cell analysis of MIBC samples from public datasets identified basal-related genes. Using Cox analysis of The Cancer Genome Atlas (TCGA)-bladder urothelial carcinoma (BLCA) clinical data and the expression matrix, and combining it with weighted gene co-expression network analysis (WGCNA) of another MIBC dataset, a disintegrin and metalloprotease protein 17 (ADAM17) was identified as a key target gene. We collected BLCA patient samples from TCGA, Shanghai Tenth People’s Hospital (STPH), and Guangdong Second Provincial General Hospital (GD2H) to develop the pathomics model for predicting ADAM17 expression. Single-cell validation of ADAM17 expression was performed using GD2H MIBC samples.

Results

ADAM17 was highly expressed in MIBC and associated with poor prognosis. Its expression correlated with clinical features such as non-papillary subtype, advanced pathological stage, and higher tissue grade. ADAM17 overexpression was linked to immune reprogramming and drug resistance. The pathomics model effectively predicted ADAM17 expression in BLCA samples. Single-cell analysis of GD2H MIBC samples confirmed ADAM17 overexpression in epithelial cells and identified key pathways in cell communication, matrix remodeling, tumor invasion, and immune regulation.

Conclusions

Multi-omics approaches effectively identify biomarkers for MIBC, with ADAM17 emerging as a key biomarker. Further research is needed to clarify its role in MIBC progression.

Highlight box.

Key findings

• A disintegrin and metalloprotease 17 (ADAM17) as a key biomarker: high ADAM17 expression in muscle-invasive bladder cancer (MIBC) correlates with poor prognosis, advanced stage, and drug resistance.

• Pathomics model: a novel pathomics model predicting ADAM17 expression demonstrated strong predictive accuracy across multiple datasets, including both internal and external validations.

• Immune landscape and drug sensitivity: ADAM17 expression alters the immune landscape, decreasing the presence of immune effector cells and enhancing immunosuppressive cells, while influencing sensitivity to chemotherapies and targeted therapies.

What is known and what is new?

• Multi-omics is a well-established approach for identifying cancer biomarkers. While previous research has highlighted biomarkers like ADAM17 in various solid tumors, its role in bladder cancer remains underexplored.

• This study identifies ADAM17 as a key biomarker for MIBC, linked to tumor aggressiveness, immune modulation, and resistance. It also presents a pathomics model predicting ADAM17 expression and its role in drug sensitivity and immune response, providing insights for targeted therapies.

What is the implication, and what should change now?

• ADAM17’s role as a prognostic biomarker in MIBC highlights its potential for early detection and monitoring. Its involvement in immune modulation and drug resistance suggests it could be targeted in combination therapies to overcome resistance.

• Further clinical validation of ADAM17 as a therapeutic target is essential. Future studies should focus on the molecular mechanisms behind ADAM17’s role in immune regulation and tumor progression. Refining the pathomics model could support its use in personalized diagnostics and treatment strategies for MIBC.

Introduction

In recent years, the rapid advancement of multi-omics technologies has transformed cancer research, enabling a more comprehensive and high-resolution understanding of tumor biology (1,2). By integrating data from genomics, transcriptomics, proteomics, and epigenomics, multi-omics approaches provide a unique ability to unravel the complexity and heterogeneity of tumors, identify key molecular drivers, and uncover novel therapeutic targets. These methods are particularly valuable for studying aggressive and heterogeneous cancers, where traditional single-dimensional analyses often fail to capture the full biological context (2).

Bladder urothelial carcinoma (BLCA) includes non-muscle invasive bladder cancer (NMIBC) and muscle-invasive bladder cancer (MIBC), which is a highly heterogeneous and aggressive malignancy (3). Tobacco smoking and occupational exposure remain the most important risk factors for bladder carcinogenesis (4). Despite advancements in surgery, chemotherapy, immunotherapy, and radiotherapy, the prognosis for MIBC remains poor. Micrometastases present at the time of cystectomy lead to recurrence in approximately 50% of patients after surgery (5). A major challenge in managing MIBC is the significant variability in tumor biology and treatment response among patients. This heterogeneity stems from complex interactions between genetic mutations, epigenetic alterations, the tumor microenvironment, and immune responses, all of which influence tumor behavior and therapeutic outcomes (6). Reliable biomarkers are urgently needed to improve risk stratification, predict clinical outcomes, and guide therapeutic decision-making. However, identifying robust biomarkers in such a biologically complex disease requires advanced molecular profiling technologies combined with complementary approaches to address the limitations of current diagnostic and prognostic methods.

Traditional pathology has long been the cornerstone of cancer diagnosis, relying primarily on morphological and histological features observed under the microscope. While this approach has been invaluable, it falls short in capturing the molecular complexity of tumors and their associated biological processes (7). Pathomics, a rapidly evolving field that integrates imaging-based analyses with genomic, transcriptomic, and proteomic data, represents a significant advancement in modern pathology. By combining high-throughput imaging data with molecular insights, pathomics bridges the gap between traditional histopathology and molecular biology. This integrated approach provides unique advantages in diagnosing diseases, assessing prognosis, and informing treatment strategies by offering a holistic view of tumor biology that incorporates both spatial and molecular dimensions (2).

As shown in Figure 1, we utilized a multi-omics approach to systematically identify key driver genes in MIBC. By integrating single-cell genomics and transcriptomics, we characterized the cellular and molecular heterogeneity of MIBC and identified several candidate genes associated with tumor progression and poor prognosis. Among these, a disintegrin and metalloprotease 17 (ADAM17) emerged as a key driver gene with significant functional and prognostic implications. As a member of the ADAM family of proteases, ADAM17 mediates the shedding of membrane-bound proteins, releasing soluble factors such as growth factors, cytokines, and adhesion molecules into the tumor microenvironment (8). Through this activity, ADAM17 modulates several oncogenic pathways, including the epidermal growth factor receptor (EGFR) and Notch signaling cascades (9). These pathways are central to tumor cell proliferation, survival, and differentiation. Its role has been implicated in various malignancies, including colorectal cancer (CRC) (10), breast cancer (11), etc. While scattered reports have linked ADAM17 to other cancers, a systematic investigation of its role in BLCA, especially MIBC, using multi-omics approaches remains lacking.

Figure 1.

A schematic representation of the study design. ADAM17, a disintegrin and metalloprotease protein 17; AUC, area under the curve; BLCA, bladder cancer; CI, confidence interval; GD2H, Guangdong Second Provincial General Hospital; HPA, Human Protein Atlas; HR, hazard ratio; LASSO, least absolute shrinkage and selection operator; MIBC, muscle-invasive bladder cancer; PCD, programmed cell death; RF, random forest; STPH, Shanghai Tenth People’s Hospital; TCGA, The Cancer Genome Atlas; UMAP, uniform manifold approximation and projection; WGCNA, weighted gene co-expression network analysis; WSIs, whole-slide images.

Moreover, we propose a multi-omics framework combining single-cell genomics, transcriptomics, and pathomics for analyzing BLCA, which may also be applicable to other tumor types and offers a promising approach for identifying potential biomarkers and therapeutic targets across cancers. Using this strategy, we explored the expression patterns, functional roles, and prognostic value of ADAM17 in MIBC. Our findings suggest that ADAM17 could serve as a biomarker for poor prognosis and a potential therapeutic target, contributing to a better understanding of MIBC at both molecular and cellular levels. We present this article in accordance with the TRIPOD reporting checklist (available at https://tau.amegroups.com/article/view/10.21037/tau-2025-79/rc).

Methods

Identification of basal subtype signature genes

We used the “Seurat” package to create objects and perform quality control to filter out low-quality cells. Before starting the analysis, we first exclude cells with fewer than 200 or more than 3,000 expressed genes, as well as cells in which the mapped unique molecular identifiers of mitochondrial genes exceed 10%. A Seurat object was clustered with a resolution of 0.01. For cell annotation, we made use of the “singleR” package and CellMarker 2.0 (http://bio-bigdata.hrbmu.edu.cn/CellMarker). This process led to the classification of 8 cell clusters, among which were epithelial cells. Epithelial cells were then re-extracted and reclustered to identify the Basal and Luminal subtypes. The relationships between these subtypes were visualized using a three-dimensional (3D) dimensionality reduction plot to better illustrate their proximity. Basal subtype signature genes were identified based on a fold change (FC) greater than 1, resulting in the identification of 101 genes. Apoptosis and autophagy scores for the Basal and Luminal subtypes were compared using the Seurat, Singscore, Ultra-fast Cell scoring (UCell), and Area Under the Curve for Cells (AUCC) packages.

Identification of key prognostic genes based on basal subtype signatures

We downloaded the expression data for BLCA from The Cancer Genome Atlas (TCGA) based on the 101 genes associated with the Basal subtype. To identify key prognostic genes from the 101 genes, we applied the least absolute shrinkage and selection operator (LASSO) method using the “glmnet” R package for coefficient selection with the log(λ) criterion, which yielded 16 genes. To further validate the prognostic significance of these 16 genes, we conducted Cox proportional hazards regression analysis, calculating hazard ratios (HRs), 95% confidence intervals (CIs), and corresponding P values.

For weighted gene co-expression network analysis (WGCNA), we utilized the de novo MIBC dataset GSE149582 to identify gene modules associated with muscle invasion. Gene expression data were preprocessed to remove outliers and missing values. After data preprocessing, a correlation matrix was generated, and a soft threshold of 0.85 was selected based on the scale-free topology fit index. The topological overlap matrix (TOM) was calculated, and hierarchical clustering was performed to identify gene modules linked to muscle invasion. We selected the MEturquoise module, which is closely associated with muscle invasion, and intersected the genes with those identified by LASSO regression, yielding three overlapping genes: ADAM17, ID2, and ITGB6, which we visualized on a chromosomal map. Based on Cox regression analysis, of these three genes, only ADAM17 exhibited a HR greater than 1 (HR =1.415), indicating that it serves as an independent risk factor. Consequently, we designated ADAM17 as the focal gene for subsequent investigations.

Establishment of prognostic gene signatures

BLCA clinical and prognostic data were obtained from TCGA, the IMvigor210 cohort, and the Gene Expression Omnibus (GEO) database (GSE194582, GSE48075 and GSE32894). Based on the TCGA database, we investigated the expression variations of the ADAM17 gene in relation to various clinical characteristics, including tumor features, muscle invasion, subtype, pathological stage, histologic grade, primary treatment outcome, body mass index (BMI), race, and survival status. The Wilcoxon rank-sum test (WRS Test) was used for comparisons between two groups, while Dunn’s Test was applied for comparisons among multiple groups. Kaplan-Meier survival curves with log-rank tests, performed using the survival analysis R package, were used to compare overall survival (OS), disease-specific survival (DSS), and progression-free interval (PFI) between high and low ADAM17 expression groups. Additionally, a nomogram model was constructed using the RMS package. The prognostic calibration curve of the nomogram was visualized to assess the relationship between predicted probabilities and observed outcomes. Furthermore, decision curve analysis (DCA) was conducted to evaluate the net clinical benefit of the model for OS at 1, 3, and 5 years in BLCA patients.

Functional enrichment and immune landscape analysis

We constructed a protein-protein interaction (PPI) network centered on ADAM17 using the STRING database with a high confidence score of 0.700. Gene Ontology (GO) functional annotation and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analyses were performed on candidate genes using the clusterProfiler package in R (v3.6.3). GO analysis included biological process (BP), molecular function (MF), and cellular component (CC) categories, while KEGG analysis explored signaling pathways involving the candidate genes. The ggplot2 package was used to visualize enrichment results with lollipop plots and cluster maps, highlighting significant functional categories and pathways.

Samples were divided into high and low ADAM17 expression groups. Immune infiltration levels were assessed using the CIBERSORT (12) algorithm with 22 immune cell markers sourced from the CIBERSORTx platform (https://cibersortx.stanford.edu/). Additionally, immune infiltration was evaluated using the single-sample gene set enrichment analysis (ssGSEA) algorithm, implemented in the R package gene set variation analysis (GSVA) (13), based on 24 immune cell markers provided by Bindea et al. (14). Visualization was performed with ggplot2, generating violin and stacked bar plots. Correlation analysis between ADAM17 expression and representative immune cell abundance was presented in scatter plots. Subsequently, we used the “corrplot” R package to conduct a correlation analysis between the risk scores and the immune checkpoint genes (ICGs).

Drug sensitivity analysis and immunotherapy prediction based on ADAM17 expression in BLCA

For the drug sensitivity analysis, we obtained the dataset of commonly used chemotherapeutic drugs for BLCA from the Genomics of Drug Sensitivity in Cancer (GDSC) database (https://www.cancerrxgene.org/). The “pRRophetic” R package was utilized to estimate the half-maximal inhibitory concentration (IC50) values of each chemotherapeutic drug.

Immune phenotype scores (IPS) were retrieved from The Cancer Imaging Archive website (https://dev.cancerimagingarchive.net/). Subsequently, the response to anti-CTLA-4 and anti-PD-1 antibody immunotherapy was predicted for the high and low ADAM17 expression groups.

Methodology for predicting ADAM17 expression using deep learning features from hematoxylin and eosin (H&E) stained whole-slide images (WSIs) in BLCA

We selected WSIs of BLCA patients from the TCGA database based on criteria including confirmed pathological diagnosis, RNA sequencing data, comprehensive clinical and pathological data, and high-quality H&E staining. Exclusion criteria removed WSIs without clear BLCA lesions, poor scan quality, missing feature data, or cases with preoperative treatments. Additional H&E slides and RNA sequencing data were collected from BLCA patients at Shanghai Tenth People’s Hospital (STPH) and Guangdong Second Provincial General Hospital (GD2H) between 2017 and 2024, with ethical approval obtained from the respective committees at STPH (approval No. 24KT68) and GD2H (approval No. 2024-KY-KZ-128-01). Informed consent was collected from all participants. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

For model construction, we employed transfer learning with ResNet50 to classify tumor/normal patches, followed by ADAM17 expression feature extraction using RetCCL, batch correction and feature aggregation, Z-score normalization, and minimum redundancy maximum relevance (mRMR) feature selection. Three models [random forest (RF), support vector machine (SVM), and logistic regression (LR)] were trained and validated using internal and external datasets, with hyperparameter tuning and performance evaluation based on area under the curve (AUC), sensitivity, and specificity as evaluation metrics. Hyperparameter tuning was conducted to optimize model performance, with AUC, sensitivity, and specificity evaluated as performance metrics. Detailed process and characteristics of patients with BLCA from multiple datasets are available in the Appendix 1. Immunohistochemical staining images of bladder tissue and BLCA tissue were retrieved from the Human Protein Atlas (HPA) database (https://www.proteinatlas.org/).

Validation of ADAM17 expression and intercellular communication in MIBC through single-cell analysis

We selected three MIBC patients from the GD2H cohort and conducted single-cell RNA sequencing. The resulting single-cell data on bladder cancer were uploaded to the Genome Sequence Archive for Human (GSA-Human: HRA009938, https://ngdc.cncb.ac.cn/gsa-human) within the National Genomics Data Center (NGDC) at the Beijing Institute of Genomics, Chinese Academy of Sciences. Key steps including quality control, batch correction, resolution selection, clustering, and annotation were performed to analyze ADAM17 gene expression at the single-cell level. Volcano plots were used to visualize the differential expression of ADAM17 across five cell clusters: T cells, epithelial cells, fibroblasts, monocytes, and endothelial cells. We also used “Monocle 3” to analyze the evolutionary trajectories of different subpopulations and presented the pseudotime distribution of the ADAM17 gene. Furthermore, to explore intercellular communication and signaling pathways, and to evaluate ligand-receptor interactions, the “CellChat” tool was employed to model intercellular communication patterns based on ligand-receptor information and predict communication probabilities. Using a P value threshold of <0.01, we visualized interactions between epithelial cells, fibroblasts, and other cell types, including ligand-receptor interactions within cell-cell communication.

Statistical analyses

Statistical analyses were performed using R (version 3.6.3), with the ggplot2 package utilized for visualizing expression differences. A P value of less than 0.05 was considered statistically significant.

Results

Identification of marker genes based on basal subtype signature

As shown in Figure 2A, we demonstrate the correlation between total RNA molecule counts and ribosomal gene proportion, mitochondrial gene proportion, and the number of distinct genes. Notably, the correlation coefficient between RNA molecule counts and the number of distinct genes is 0.93, which aligns with expectations for single-cell sequencing data. Quality control analysis (Figure 2B) confirms the reliability of our results. Using uniform manifold approximation and projection (UMAP) dimensionality reduction techniques, we identified eight distinct cell type clusters, including T cells, monocytes, dendritic cells, epithelial cells, endothelial cells, macrophages, neutrophils, and stromal cells (Figure 2C,2D). Epithelial cells were further reclustered, and we identified two epithelial subtypes: basal and luminal, along with a small subset of immune cells. Figure 2E,2F display UMAP and 3D dimensionality reduction plots, highlighting the spatial distribution of these three clusters. We identified 101 marker genes for the basal subtype (available online: https://cdn.amegroups.cn/static/public/tau-2025-79-1.xlsx). To compare apoptotic and autophagic activity between basal and luminal epithelial cell types, we employed four scoring methods. All results consistently showed that basal cells exhibit lower programmed cell death scores (Figure 2G), indicating that they are less prone to death. This suggests that basal cells have a greater potential for sustained proliferation and invasion.

Figure 2.

Analysis based on the locally infiltrating BLCA single-cell dataset GSE135337. (A) Distribution plots showing the relationship between the number of molecules and ribosomal, mitochondrial, and gene counts. (B) Quality control plots. (C,D) UMAP plots displaying different cell clusters. (E,F) Trajectories of different cell clusters (E) and secondary annotation analysis of epithelial cell subtypes, including basal and luminal cells (F). (G) Comparison of apoptosis and autophagy scores between basal and luminal cell types using four algorithms: AddModuleScore, Singscore, UCell, and AUCC. ****, P<0.0001, basal vs. luminal group. AUCC, area under the curve for cells; BLCA, bladder cancer; UCell, Ultra-fast Cell scoring; UMAP, uniform manifold approximation and projection.

Identification of prognostic biomarkers in BLCA

We extracted comprehensive clinical information and the corresponding expression matrices of BLCA patients from the TCGA database. The extraction was based on 101 genes for screening purposes. First, we applied LASSO regression and 10-fold cross-validation to select the optimal variables. Sixteen genes, including ADAM17, S100A10, TMEM159, and IFNGR1, were identified as significant predictive variables (Figure 3A,3B). Subsequently, we conducted multivariate Cox regression analysis to evaluate the statistical significance of the relationship between the selected genes and patient prognosis. As shown in Figure 3C, the results revealed that high expression of ADAM17 (HR =1.415, P=0.005) and THBS1 (HR =1.142, P=0.03) was associated with an increased risk of death. Conversely, high expression of ITGB6 (HR =0.882, P=0.04), TMEM159 (HR =0.776, P=0.02), CSAG1 (HR =0.886, P=0.002), and IFNGR1 (HR =0.630, P<0.001) was associated with a lower risk of death.

Figure 3.

Identification of prognosis-related genes in basal cells. (A) A coefficient solution path plot of 101 basal marker genes was generated using L1 regularization. (B) Optimal variables were determined through LASSO regression and 10-fold cross-validation. (C) A forest plot presents multivariate Cox regression analysis to evaluate the prognostic significance of basal subtype-related genes. (D) Cluster analysis was performed using WGCNA on the GSE149582 dataset for primary MIBC. (E) The soft-thresholding power in WGCNA was determined. (F) Module-trait relationships in de novo MIBC were analyzed. (G) The intersection of the MEturquoise module gene set from WGCNA with prognostic basal subtype-related genes was visualized with chromosomal localization. CI, confidence interval; HR, hazard ratio; LASSO, least absolute shrinkage and selection operator; MIBC, muscle-invasive bladder cancer; WGCNA, weighted gene co-expression network analysis.

Accompanying this analysis is the WGCNA performed on the MIBC dataset (GSE149582). Figure 3D illustrates the gene dendrogram, with branches representing co-expressed gene groups that were assigned to modules through dynamic tree cutting. The soft-thresholding power (power =12) was selected to achieve a balance between scale-free topology (fit index ≥0.9) and gene connectivity stability (Figure 3E). Figure 3F demonstrates a significant positive correlation between the MEturquoise module and de novo MIBC (correlation coefficient =0.47, P=0.01). Following this analysis, the 433 genes (available online: https://cdn.amegroups.cn/static/public/tau-2025-79-2.pdf) in the MEturquoise module were intersected with 16 LASSO regression-identified genes, yielding three overlapping genes: ADAM17, ID2, and ITGB6. The circos plot highlights their chromosomal localization and adjacent positions. Based on Cox analysis, ADAM17, marked in red, is identified as an independent risk factor for BLCA (Figure 3G).

Prognostic significance and clinical implications of ADAM17 expression in BLCA

Based on the TCGA-BLCA dataset, ADAM17 expression was significantly higher in tumor tissues than in normal tissues (P=0.002, Figure 4A). Additionally, ADAM17 expression was higher in MIBC, the non-papillary subtype, advanced pathological stages, higher histologic grades, poorer therapy outcomes, higher BMI, and non-Asian populations (Figure 4B-4I). Detailed parameters can be found in Appendix 1, Table S1. This indicates that high ADAM17 expression is closely associated with the aggressive features and poor clinical outcomes of BLCA.

Figure 4.

Association between ADAM17 expression and clinical features of BLCA based on TCGA database analysis. (A) Comparison between normal and tumor tissues. (B) Comparison of non-muscle-invasive and muscle-invasive bladder cancer. (C) Comparison across different subtypes. (D) Comparison among different pathological stages. (E) Comparison between low and high histologic grades. (F) Comparison of primary therapy outcomes. (G) Comparison between low and high BMI groups. (H) Comparison across different racial groups. (I) Comparison between survival statuses. The significance levels were denoted as follows: *, P<0.05; **, P<0.01; ***, P<0.001. ADAM17, a disintegrin and metalloprotease protein 17; BLCA, bladder urothelial carcinoma; BMI, body mass index; CR, complete response; OS, overall survival; PD, progressive disease; PR, partial response; SD, stable disease; TCGA, The Cancer Genome Atlas; TPM, transcript per million.

As shown in Figure 5A-5C, Kaplan-Meier survival analysis of the TCGA dataset revealed that high ADAM17 expression was significantly associated with worse OS (HR =1.39, P=0.03), DSS (HR =1.45, P=0.04), and PFI (HR =1.49, P=0.008). This association with worse OS was further validated in the IMvigor210 cohort (HR =1.49, P=0.002, Figure 5D), GSE48075 (HR =2.77, P=0.002, Figure 5E), and GSE32894 (HR =2.76, P=0.01, Figure 5F), establishing high ADAM17 expression as a risk factor for poor prognosis in BLCA. Next, we developed a nomogram model to predict 1-, 3-, and 5-year OS by incorporating ADAM17 expression, pathological T/N/M stages, and histological grade (Figure 5G). The calibration curve demonstrated strong consistency between predicted and observed survival probabilities, highlighting the model’s robust predictive accuracy (Figure 5H). DCA demonstrated the net benefit for patients (Figures 5I-5K). In the 5-year OS analysis, the model’s net benefit (blue line) exceeded those of the treat-all (red line) and treat-none (green line) strategies, confirming its superior clinical utility.

Figure 5.

Prognostic predictive value of ADAM17 in BLCA. (A-C) Kaplan-Meier survival analysis of ADAM17 expression in BLCA patients from the TCGA cohort, including OS (A), DSS (B), and PFI (C). (D-F) Kaplan-Meier analysis of OS for ADAM17 expression in BLCA patients from the IMvigor210 cohort (D), GSE48075 (E), and GSE32894 (F). (G) Nomogram model predicting 1-, 3-, and 5-year OS for BLCA patients. (H) Calibration curve for evaluating the accuracy of the prognostic model. (I-K) DCA for the 1-year (I), 3-year (J), and 5-year (K) OS of the risk model. ADAM17, a disintegrin and metalloprotease protein 17; BLCA, bladder urothelial carcinoma; CI, confidence interval; DCA, decision curve analysis; DSS, disease-specific survival; OS, overall survival; PFI, progression-free interval.

The dual role of ADAM17 in cancer-related functions and immune landscape remodeling in BLCA

As shown in Figure 6A, we analyzed the correlation between ADAM17 and its related genes, such as ADAM10, FURIN, TNF, and NOTCH1. Functional enrichment analysis was then performed on these genes. Figure 6B,6C illustrate the results, showing MF including Notch binding and protease binding, CC such as cell-cell junctions and adherens junctions, and BP like wound healing and inflammatory response to antigenic stimulus. KEGG pathway analysis revealed significant enrichment in pathways such as the Notch signaling pathway, endocrine resistance, microRNAs in cancer, and proteoglycans in cancer.

Figure 6.

Functional enrichment and immune infiltration analysis of ADAM17 was conducted to elucidate its biological significance and immune-related associations. (A) The PPI network identifies the top ten genes closely associated with ADAM17. (B,C) GO and KEGG analyses are presented using bar plots (B) and clustering plots (C), respectively, highlighting enriched biological processes and pathways. (D) A stacked bar plot displays the proportions of 22 immune cell types in high and low ADAM17 expression groups, analyzed using CIBERSORT, while a violin plot (E) depicts the expression levels of 24 immune cell types between these groups based on ssGSEA. (F) A scatter plot demonstrates the correlation between ADAM17 expression and representative immune infiltrating cells. The significance levels were denoted as follows: *, P<0.05; **, P<0.01; ***, P<0.001. ADAM17, a disintegrin and metalloprotease protein 17; BP, biological process; CC, cellular component; GO, Gene Ontology; KEGG, Kyoto Encyclopedia of Genes and Genomes; MF, molecular function; PPI, protein-protein interaction; ssGSEA, single-sample gene set enrichment analysis; TPM, transcript per million.

Next, we used the CIBERSORT and ssGSEA algorithms to explore the impact of ADAM17 on the immune landscape in BLCA. As shown in Figure 6D, there are notable differences in the overall distribution of immune cell populations between high and low ADAM17 expression groups. Violin plots further illustrate the group comparisons more intuitively (Figure 6E). For results with statistical significance, we selectively visualized them using scatter plots. We observed that ADAM17 expression is negatively correlated with the enrichment of immune effector cells, such as CD8⁺ T cells, cytotoxic cells, and NK cells, while positively correlated with the enrichment of immunosuppressive cells, such as Tgd cells, T helper cells, and Th2 cells (Figure 6F). As shown in Appendix 1 (Figure S1), ADAM17 exhibited significant correlations with multiple immune checkpoints, including CD276, TIGIT, and CD160, among others.

Therapeutic implications of ADAM17 expression

As shown in Figure 7, we present the sensitivity analysis of chemotherapeutic and targeted drugs in relation to ADAM17 expression levels. The results indicate that the IC50 values of drugs such as cisplatin, gemcitabine, vinorelbine, paclitaxel, and docetaxel are higher in the ADAM17 high-expression group, suggesting that high ADAM17 expression may contribute to resistance to these drugs. However, drugs such as afatinib, dabrafenib, and dactolisib show no significant difference in their IC50 values between the high- and low-expression groups. Overall, high ADAM17 expression may reduce sensitivity to certain chemotherapeutic agents, such as platinum-based and taxane-based drugs, but does not significantly affect sensitivity to specific targeted therapies, such as afatinib and dabrafenib. As shown in Figure 7Q-7T, we further predicted immune therapy responses to anti-CTLA-4 and anti-PD-1 antibodies. Unfortunately, ADAM17 expression levels did not significantly influence patients’ responses to CTLA-4 and PD-1 immune checkpoint therapies.

Figure 7.

Predicting treatment responses in BLCA patients was performed by comparing the half-maximal IC50 values of various therapeutic agents between different ADAM17 expression groups. The agents analyzed included cisplatin (A), gemcitabine (B), vinorelbine (C), paclitaxel (D), docetaxel (E), cyclophosphamide (F), 5-fluorouracil (G), IGF1R inhibitors (H), JAK inhibitors (I), GSK inhibitors (J), elephantin (K), linsitinib (L), afatinib (M), dabrafenib (N), dactolisib (O), and lapatinib (P). (Q-T) IPS score prediction for response to anti-CTLA-4 and anti-PD-1 antibodies. The significance levels were denoted as follows: *, P<0.05; **, P<0.01; ***, P<0.001; ****, P<0.0001; ns, P>0.05. ADAM17, a disintegrin and metalloprotease protein 17; BLCA, bladder urothelial carcinoma; CTLA-4, cytotoxic T lymphocyte-associated protein 4; IC50, half-maximal inhibitory concentration; IPS, immune phenotype score; PD-1, programmed cell death protein 1.

Pathomics based RF model for predicting ADAM17 expression in BLCA

To investigate the correlation between ADAM17 expression and pathological tissues in clinically diagnosed BLCA patients, we developed a Resnet50 model to distinguish tumor and normal tissue patches. As shown in Figure 8A, the workflow includes key steps such as model training, prediction, and feature extraction. The confusion matrices demonstrate the novel classification performance of the Resnet50 model across different datasets, including the TCGA training set, TCGA validation set, STPH external validation set, and GD2H external validation set (Figure 8B). Detailed information can be found in Appendix 1, Table S2.

Figure 8.

Pathomics-based RF model for predicting ADAM17 expression in BLCA. (A) Overview of the ResNet50 model for classifying tumor and normal tissue patches based on ADAM17 expression, using hematoxylin and eosin stained (H&E) staining. (B) Confusion matrices for the TCGA, STPH, and GD2H datasets, showing classification performance for tumor and normal tissues. (C-F) ROC curves for RF, SVM, and LR models using features from all patches (C,D) and tumor patches (E,F), with the RF model outperforming others. (G) Radar plot of key performance metrics (accuracy, sensitivity, specificity, PPV, NPV) across datasets. (H) Accuracy values for the RF model in predicting ADAM17 expression across internal and external datasets. (I) Waterfall plot showing predicted ADAM17 expression probabilities across patients and datasets. (J) IHC images of ADAM17 expression in tumor tissues (https://www.proteinatlas.org/ENSG00000151694-ADAM17/cancer/urothelial+cancer#img) and normal tissues (https://www.proteinatlas.org/ENSG00000151694-ADAM17/tissue/urinary+bladder#img) from the HPA database. ADAM17, a disintegrin and metalloprotease protein 17; BLCA, bladder urothelial carcinoma; GD2H, Guangdong Second Provincial General Hospital; HPA, Human Protein Atlas; IHC, immunohistochemistry; LR, logistic regression; NPV, negative predictive value; PPV, positive predictive value; RF, random forest; ROC, receiver operating characteristic; STPH, Shanghai Tenth People’s Hospital; SVM, support vector machine; TCGA, The Cancer Genome Atlas; WSI, whole-slide image.

In this study, RF, SVM, and LR models were constructed using features from both tumor and normal patches. Figure 8C-8F show AUC values for the models on both patch types. The RF model, using tumor patch features, outperformed the others in all datasets, especially in internal and external validations. A radar plot (Figure 8G) evaluates key metrics [accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV)], showing optimal performance on the training set, with a slight decline in external validation datasets. The RF model achieved accuracy between 60.38% and 66.88% for ADAM17 prediction (Figure 8H), with the highest sensitivity (78.26%) in external dataset 2. The SVM and LR models had lower sensitivity, particularly in external datasets 1 and 2. RF also had good specificity, notably in the training (72.37%) and validation (68.12%) datasets, though slightly lower than that of LR. A waterfall plot (Figure 8I) visually confirms RF’s effectiveness in distinguishing ADAM17 expression levels. The HPA database shows that ADAM17 expression is much higher in tumor tissues than in normal tissues (Figure 8J).

Single-cell analysis validates ADAM17 expression and signaling pathways in MIBC

As shown in Figure 9A-9D, we performed single-cell analysis on MIBC patient samples collected from GD2H. Both t-distributed stochastic neighbor embedding (t-SNE) and UMAP analyses revealed the cellular heterogeneity of MIBC and identified five major cell clusters, including epithelial cells. The volcano plot illustrates the differential expression of the ADAM17 gene across different cell clusters, showing high expression in epithelial cells and endothelial cells but low expression in T cells (Figure 9E). As shown in Figure 9F, trajectory analysis revealed the developmental trajectory of epithelial cells, indicating their evolutionary path in pseudotime. The pseudotime distribution (Figure 9G) further demonstrates that the expression of ADAM17 in epithelial cells gradually increases along pseudotime. Taken together, these findings suggest that the expression pattern of ADAM17 may be closely associated with the dynamic changes or specific developmental stages of tumor cells.

Figure 9.

Single-cell analysis of MIBC patients from GD2H. (A,B) t-SNE plots showing five cell clusters (A), including T cells, epithelial cells, fibroblasts, monocytes, and endothelial cells (B). (C,D) UMAP plots displaying the same five cell clusters (C), with T cells, epithelial cells, fibroblasts, monocytes, and endothelial cells (D). (E) Volcano plot illustrating the differential expression of ADAM17 across different cell clusters, with overexpression in epithelial and endothelial cells, and low expression in T cells. (F) Trajectories of epithelial cells. (G) Pseudotime distribution of the ADAM17 gene. ADAM17, a disintegrin and metalloprotease protein 17; GD2H, Guangdong Second Provincial General Hospital; MIBC, muscle-invasive bladder cancer; t-SNE, t-distributed stochastic neighbor embedding; UMAP, uniform manifold approximation and projection.

Figure 10 presents the cell-cell communication analysis based on single-cell data from MIBC patients. As shown in Figure 10A,10B, epithelial cells, endothelial cells, and T cells exhibited extensive communication connections, highlighting their importance in signal transduction. Based on these observations, we constructed the hierarchical map of the amyloid precursor protein (APP)-collagen-laminin-macrophage migration inhibitory factor (MIF) multi-signaling pathway (Figure 10C). This map illustrates the hierarchical relationships among different cell clusters acting as signal senders (sources) and receivers (targets). Heatmaps for the APP, COLLAGEN, LAMININ, and MIF signaling pathways further visualize the activity of these networks (Figure 10E). Specifically, the APP signaling pathway involves T cells and epithelial cells as active signal senders and receivers, highlighting their significant roles. In the COLLAGEN signaling pathway, strong communication was observed between fibroblasts and epithelial cells, indicating critical interactions. The LAMININ signaling pathway demonstrated active exchanges between epithelial cells and monocytes, emphasizing their importance in cell-cell communication. Meanwhile, the MIF signaling pathway identified endothelial cells as key signal senders, underscoring their potential regulatory role in this network. As shown in Figure 10F, bar plots illustrate the functional roles of different cell clusters in each signaling pathway, including their roles as senders, receivers, mediators, and influencers. The dot plot (Figure 10F) summarizes the communication probability and statistical significance of ligand-receptor interactions across different cell clusters.

Figure 10.

Cell-cell communication analysis. (A) Number of cell-cell communication network connections between cells. (B) Strength of cell-cell communication network connections between cells. (C) Hierarchical plot under the APP-COLLAGEN-LAMININ-MIF multi-signal pathway. (D) Heatmap of the APP, COLLAGEN, LAMININ, and MIF signaling pathways. (E) Roles of different cell clusters in the APP, COLLAGEN, LAMININ, and MIF signaling pathways. (F) An overview of cell-cell communication, illustrating ligand-receptor interactions. APP, amyloid precursor protein; MIF, migration inhibitory factor.

Discussion

As shown in Figure 1, we initially selected a single-cell dataset of MIBC for analysis due to its high invasiveness, metastatic potential, and poor prognosis, making it a critical subtype for understanding BLCA. Investigating the molecular mechanisms of MIBC is essential for improving clinical diagnosis and treatment. On the other hand, single-cell datasets offer high-resolution insights into cellular heterogeneity, enabling us to explore the roles of various cell types and their subpopulations in tumor development and identify potential driver genes (15).

In BLCA, the epithelial basal cell subtype, characterized by stem cell-like properties, cell proliferation, and drug resistance, plays a critical role in tumor invasiveness, metastasis, and immune evasion (16). By analyzing genes related to basal cells, we can better understand the characteristics and functions of these cells in BLCA and uncover their impact on tumor progression and clinical prognosis. As shown in Figure 2E, we identified basal cells in BLCA and evaluated apoptosis and autophagy scores between basal and luminal cells. The results confirmed that basal cells had lower scores, indicating stronger anti-apoptotic and anti-autophagic abilities.

To address the limitations of simple differential gene analysis, we integrated multi-factor Cox regression analysis with WGCNA. Differential gene analysis alone typically identifies genes that are significantly altered under specific conditions but ignores gene interactions and complex network relationships. WGCNA can identify gene modules associated with specific disease phenotypes, providing a more comprehensive molecular network perspective (17). By using multi-factor Cox regression models for survival analysis, we can assess the independent contribution of each gene to clinical prognosis and reveal the associations between genes and disease progression under the combined influence of multiple factors. As shown in Figure 3G, we focused on the ADAM17 gene, which may act as a driver gene in the pathogenesis of MIBC. ADAM17 is a metalloprotease involved in a variety of biological processes, particularly in tumorigenesis and progression. ADAM17 regulates the tumor microenvironment by shedding cytokine receptors, cell adhesion molecules, and growth factors, thereby promoting tumor cell proliferation, migration, invasion, and immune evasion (8). Its aberrant expression in various malignancies is closely linked to tumor invasiveness and metastasis, influencing immune responses, inflammation, and tumor cell growth and differentiation through cleavage of key protein substrates (18). Research has reported that exosomal ADAM17 facilitates the metastasis of CRC. In nude mice, it does so by inducing vascular leakage and mediating the formation of the pre-metastatic niche (10). Moreover, employing anti-ADAM17 monoclonal antibody as a carrier could represent a promising novel strategy for selective anti-breast cancer drug delivery (11).

As shown in Figure 4, clinical feature analysis reveals that ADAM17 is associated with various clinical indicators, including muscle invasion and non-papillary subtypes, suggesting the possibility of differential expression of ADAM17. This suggests that aberrant ADAM17 expression may be associated with broader genomic instability. Similar to studies that identify loss of Y chromosome (LoY) as an early marker of genomic instability (19), ADAM17 may play an analogous role as an early indicator of genomic instability, influencing the initiation and progression of BLCA. Furthermore, ADAM17 is correlated with higher pathological stage, tissue grade, and poorer treatment outcomes, which explains why its high expression is associated with worse OS, PFI, and DSS, as shown in Figure 5A-5C. We developed an ADAM17-based Nomogram model, and the DCA indicated that patients could benefit from this model (Figure 5G-5K), which is a promising outcome. As illustrated in Figure 6A, ADAM17 and its co-expressed genes play key roles in various cancers. For instance, furin promotes the proliferation, invasion, and metastasis of endometrial adenocarcinoma by activating ADAM17, transforming growth factor-β (TGF-β), and other factors (20). Moreover, increased ADAM17 expression may drive gastric cancer progression through the activation of Wnt signaling pathways (21). Gene knockdown of ADAM17 and short-term inhibition of ADAM17 prevent the formation of long-term metastatic tumors in the lungs (22). As illustrated in Figure 6B, KEGG pathway analysis revealed that ADAM17 is closely associated with the Notch signaling pathway and endocrine resistance. Studies indicate that inhibition of ADAM17 activity inhibits the activation of the Notch pathway (23), while the role of NOTCH signaling in tumorigenesis in bladder cancer has been well documented (24). Moreover, a study emphasized that ADAM17 regulates angiogenesis through a mechanism independent of the Notch signaling pathway (25). Additionally, overexpression of certain ADAM family proteins may contribute to endocrine resistance. For example, LGI1, a neuropeptide that binds to ADAM proteins, inhibits cell migration and proliferation (26). Furthermore, GO analysis showed that ADAM17 is functionally linked to wound healing, adherens junction organization, and cell-cell junction assembly, suggesting its potential role in tumor cell invasion, migration, and tumor microenvironment remodeling. As shown in Figure 6C, ADAM17 was also associated with inflammatory responses to antigenic stimuli. Previous studies have demonstrated that systemic inflammation markers, including the systemic inflammation response index (SIRI), pan-immune inflammation value (PIV), and systemic immune-inflammation index (SII), serve as robust prognostic indicators in bladder cancer, correlating with aggressive tumor features and poorer survival outcomes (27,28). These results may significantly influence ADAM17 in the progression of MIBC and tumor microenvironment remodeling. Subsequently, we proceeded to analyze the immune landscape of ADAM17, which further confirmed our hypothesis (Figure 6F). In a high ADAM17 expression state, immune effector cells, such as CD8⁺ T cells and cytotoxic cells, were reduced, while immune-suppressive cells, such as Tgd and Th2 cells, were increased. This suggests that changes in ADAM17 expression are associated with immune reprogramming. As shown in Figure 7, drug sensitivity analysis revealed that high expression of ADAM17 is linked to resistance to various chemotherapy agents, including cisplatin, paclitaxel, and 5-fluorouracil. Other targeted drugs, such as Linsitinib, showed resistance in only a small subset of cases, while most targeted therapies did not demonstrate significant differences in response. Although ADAM17 does not directly affect the response to anti-CTLA-4 and anti-PD-1 therapies, it shows significant correlations with several immune checkpoints, such as CD276 and TIGIT. This suggests that ADAM17 may influence immune evasion mechanisms by regulating immune checkpoint expression. TIGIT, when overexpressed in the tumor microenvironment, is often associated with immune evasion, particularly in the evasion of T cell-mediated immune surveillance (29). CD276, also known as B7-H3, is an immune-suppressive molecule expressed in various cancers, promoting tumor progression by inhibiting T cell activity (30). ADAM17 may play a role in the progression of MIBC by modulating immune checkpoints like CD276 and TIGIT, facilitating tumor immune evasion. Moreover, ADAM17, responsible for shedding membrane-anchored proteins involved in both health and disease, has emerged as a potential target for tumor-targeted therapies. Cleavage of erbB2/erbB4 by ADAM17 is associated with poor prognosis and reduced efficacy of receptor-targeted therapies. Therefore, monoclonal antibodies targeting ADAM17 could benefit patients with solid tumors, such as triple-negative breast cancer and ovarian cancer (31). The anti-ADAM17 inhibitory monoclonal antibody inhibits the proliferation and migration of head and neck squamous cell carcinoma cells in combination with the EGFR tyrosine kinase inhibitor (TKI) gefitinib (32). Additionally, overactivation of ADAM17 suppresses NK cell function by downregulating CD16, and ADAM17 antagonists help restore NK cell activity (33). In bladder cancer, a study suggested that ADAM17 inhibitors combined with immune checkpoint inhibitors, such as anti-programmed cell death ligand 1 (anti-PD-L1), may enhance the efficacy of immunotherapy (34), although the underlying mechanism still requires further investigation, particularly in MIBC.

Pathomics integrates traditional pathology with modern multi-omics to analyze tumor biology. This approach is essential for early tumor diagnosis, prognosis assessment, and treatment strategy selection, particularly in the field of urological tumors. One study reported a machine learning-based multimodal model that combines preoperative magnetic resonance imaging, WSIs, and clinical variables to predict biochemical recurrence following radical prostatectomy in prostate cancer, offering a promising tool for personalized postoperative treatment (35). Another study demonstrated that pathological features could effectively predict the prognosis and CTLA4 expression levels in clear cell renal cell carcinoma (ccRCC) patients, suggesting that high CTLA4 expression is associated with poor prognosis in ccRCC (36). As shown in Figure 8, a RF model, which integrates pathomics deep learning features of both tumor and normal tissue slices, had novel performance in ADAM17 expression prediction in BLCA patients. Immunohistochemistry (IHC) remains the gold standard for protein detection. Our pathomics model is not a replacement but a complementary strategy. It enables rapid WSIs prescreening to prioritize high-probability cases for IHC validation, reducing test volumes (37). Unlike localized IHC sampling, our whole-slide deep learning model provides a better evaluation of spatial heterogeneity (38). These findings highlight the transformative potential of pathomics in advancing precision medicine, offering novel insights into tumor biology and facilitating the development of personalized therapeutic strategies.

Subsequently, we collected fresh MIBC samples for single-cell analysis to validate the expression of the ADAM17 gene (Figure 9). This analysis involved identifying the key cell populations that express ADAM17, exploring its role within the immune microenvironment, and investigating the involvement of receptor-ligand interactions and signaling pathways. As shown in Figure 10A,10B, epithelial cells emerge as the central cell type within the communication network. Figure 10D,10E illustrate that pathways such as APP, COLLAGEN, LAMININ (39), and macrophage MIF play significant roles in cell communication (40), suggesting their involvement in matrix remodeling, tumor invasion, and immune regulation. The study has shown that when ADAM17 is silenced, the adhesion of lymphatic endothelial cells to collagen is enhanced, indicating that ADAM17 plays an important role in regulating the interaction between cells and collagen (41). APP, a cell adhesion molecule (42), can be cleaved by ADAM17 in the amyloid β peptide region (43). This suggests that overexpression or abnormal activity of ADAM17 may lead to abnormal cleavage of APP, generating fragments that promote tumor invasion, thereby facilitating the infiltration of MIBC cells into bladder tissue. However, the specific mechanisms remain unclear. Figure 10F presents receptor-ligand analysis, revealing potential molecular targets like APP-CD74 and COL1A1-ITGA1, which could serve as a foundation for future studies on tumor microenvironment regulation and targeted therapies.

Conclusions

In summary, Bladder cancer requires early diagnosis and comprehensive treatment, including radiotherapy, chemotherapy, surgical intervention, and immunotherapy. Notably, with the evolution of the biopsychosocial model, there has been increasing attention to the quality of life in post-operative bladder cancer patients (44). We anticipate that our multi-omics framework will ultimately identify the key driving role of ADAM17 in MIBC. These findings will not only facilitate the identification of biomarkers for BLCA research but also have potential for extension to other cancer types. This approach may aid early diagnosis and refine personalized therapeutic strategies, ultimately achieving precision medicine aligned with contemporary medical paradigms.

Limitation

Several limitations should be acknowledged. Our study samples were collected from southern China, and their selection was influenced by geographical location, patient population, treatment regimens, and data collection methods, which may impact the representativeness and external validity of the results. The biological characteristics of patients from different regions may affect ADAM17 expression and its association with MIBC. Additionally, this study employed a retrospective design, limiting our ability to control for variables and making it susceptible to missing data and bias. Clinical features and treatment regimens may not have been fully documented, and patient selection could have been influenced by clinical workflows, leading to potential biases in the results. Although the reliability of ADAM17 as a biomarker for MIBC has been validated in multiple datasets, prospective studies are needed to further validate these findings, providing more stringent control and data collection standards to reduce bias. Finally, our study focused mainly on data-driven approaches, which may not fully capture the complex biological interactions in vivo.

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. Ethical approval was obtained from the respective committees at STPH (approval No. 24KT68) and GD2H (approval No. 2024-KY-KZ-128-01). Informed consent was collected from all participants. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Data Sharing Statement

Available at https://tau.amegroups.com/article/view/10.21037/tau-2025-79/dss

tau-14-06-1610-dss.pdf ^{(76.7KB, pdf)}

DOI: 10.21037/tau-2025-79