Identification of pyroptosis-related molecular subtypes and development of a subtype-derived six-gene prognostic signature for bladder cancer

Abstract

Background

Pyroptosis is an inflammatory form of programmed cell death that may remodel the tumor immune microenvironment and influence clinical outcomes. However, the prognostic value of pyroptosis-related molecular patterns and their translational implications in bladder cancer remain incompletely defined. This study aimed to identify pyroptosis-related molecular subtypes, develop and externally validate a subtype-derived prognostic signature, and explore its associations with immune and therapeutic features in bladder cancer.

Methods

The Cancer Genome Atlas-bladder cancer cohort (TCGA-BLCA) was used as the development cohort, GSE13507 and GSE31684 as external validation cohorts, and IMvigor210 as an exploratory immunotherapy-treated cohort. Pyroptosis-related molecular subtypes were identified by non-negative matrix factorization (NMF), and pathway activity and immune infiltration were assessed by gene set variation analysis (GSVA) and single-sample gene set enrichment analysis (ssGSEA). A subtype-derived six-gene prognostic signature was developed using least absolute shrinkage and selection operator Cox (LASSO-Cox) regression, and performance was evaluated by Kaplan-Meier analysis, time-dependent receiver operating characteristic (ROC) curves, and calibration.

Results

Two pyroptosis-related molecular subtypes (C1/C2) with distinct survival and immune characteristics were identified. A subtype-derived six-gene prognostic signature comprising ECM1, FER1L4, FKBP10, ANXA1, ARL4C, and CTSE stratified patients into high- and low-risk groups in the development and validation cohorts and remained significantly associated with overall survival (OS) after multivariable adjustment in the TCGA-BLCA cohort. Higher risk scores were associated with higher pathological grade and advanced T stage. Risk stratification was also associated with mutational profiles, immune checkpoint expression, and immunophenoscore (IPS) differences, while CellMiner analysis generated exploratory drug-sensitivity clues.

Conclusions

We established pyroptosis-related molecular subtypes and developed a subtype-derived six-gene prognostic signature that was evaluated across multiple independent cohorts. The model is associated with immune microenvironment features and genomic context, and offers translational clues for immunotherapy and targeted strategies, warranting further prospective and experimental validation.

Highlight box.

Key findings

• Two pyroptosis-related molecular subtypes with distinct survival and immune characteristics were identified in bladder cancer.

• A subtype-derived six-gene prognostic signature was developed from genes differentially expressed between pyroptosis-related molecular subtypes and externally validated across multiple independent cohorts.

• The risk score was associated with clinicopathological aggressiveness, immune checkpoint expression, mutational context, and inferred therapeutic sensitivity.

What is known and what is new?

• Pyroptosis has been implicated in tumor immunity and bladder cancer progression, but the prognostic value of pyroptosis-related molecular heterogeneity remains incompletely defined. Previous studies often lacked rigorous multi-cohort validation and did not systematically integrate molecular subtyping, prognostic modeling, immune landscape analysis, and translational clues.

• This study links pyroptosis-related subtypes with immune features and establishes a six-gene prognostic signature derived from subtype-associated transcriptional differences, validated in The Cancer Genome Atlas, Gene Expression Omnibus, and IMvigor210 cohorts.

What is the implication, and what should change now?

• Pyroptosis-related transcriptomic features may refine molecular risk stratification in bladder cancer beyond conventional clinicopathological variables.

• The proposed model may serve as a hypothesis-generating framework for prognostic assessment and immunotherapy-related research.

• Prospective validation and mechanistic studies are needed before clinical application.

Introduction

Bladder cancer is one of the most common malignancies of the urinary system and is characterized by pronounced molecular and clinical heterogeneity, resulting in substantial inter-patient variability in recurrence/progression risk and survival outcomes (1). Although current clinical risk stratification primarily relies on pathological factors such as tumor stage (T stage) and grade, marked outcome heterogeneity persists within the same stage/grade categories, suggesting that conventional clinicopathological features alone are insufficient for precise prognostic assessment and individualized treatment decision-making (2,3). Therefore, robust and reproducible molecular stratification systems and prognostic tools that can be linked to the tumor immune microenvironment and potential treatment responses are needed for precision medicine in bladder cancer.

Pyroptosis is an inflammatory form of programmed cell death, classically characterized by inflammasome activation, caspase-mediated cleavage of gasdermins, pore formation, and release of inflammatory cytokines (4). Pyroptosis plays context-dependent roles in cancer: it may enhance immunogenic cell death and promote antitumor immune responses, whereas sustained inflammatory signaling may facilitate an immunosuppressive microenvironment and tumor progression (5,6). As immunotherapy has increasingly become an important treatment strategy for urothelial carcinoma, the associations between pyroptosis-related molecular events and tumor immune infiltration, immune checkpoint axes, and genomic features such as tumor mutational burden (TMB) are biologically plausible and may have translational implications (7-9).

In recent years, pyroptosis-related gene signatures have been increasingly explored for tumor subtyping and prognostic prediction in bladder cancer, but several limitations remain. Some studies lack rigorous multi-cohort external validation, raising concerns regarding generalizability; others do not systematically connect pyroptosis-related molecular patterns with the immune microenvironment and translational therapeutic clues (10,11). In addition, the relationship between molecular subtype assignment, quantitative risk stratification, and clinicopathological aggressiveness remains insufficiently defined. Therefore, in this study, we integrated The Cancer Genome Atlas (TCGA) and multiple external cohorts to identify pyroptosis-related molecular subtypes, characterize their biological and immune features, develop a subtype-derived six-gene prognostic signature, and further explore its associations with mutational context, immune checkpoint expression, and inferred therapeutic sensitivity in bladder cancer. We present this article in accordance with the TRIPOD reporting checklist (available at https://tau.amegroups.com/article/view/10.21037/tau-2026-0327/rc).

Methods

Data sources and study design

This retrospective multi-cohort study was based on publicly available databases. Transcriptomic profiles and corresponding clinical follow-up data from TCGA-bladder cancer cohort (TCGA-BLCA) were used as the development cohort (12), while Gene Expression Omnibus (GEO) datasets GSE13507 and GSE31684 served as external validation cohorts (13,14); IMvigor210 was included as an exploratory immunotherapy-treated cohort (15). The GEO validation cohorts (16) included GSE13507 (platform GPL6102, Illumina Human-6 v2.0 Expression BeadChip) and GSE31684 (platform GPL570, Affymetrix Human Genome U133 Plus 2.0 Array). The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. Because all data were de-identified and publicly available, additional ethical approval and informed consent were not required. For prognostic analyses, only tumor samples with available transcriptomic data, OS information, and genes required for risk-score calculation were retained. After matching expression and survival annotations, the final analytic cohorts included 406 patients from TCGA-BLCA, 165 from GSE13507, 93 from GSE31684, and 348 from IMvigor210. These numbers represent the analysis-ready cases used in the corresponding cohort-level analyses, and the final cohort sizes and death events are summarized in Table 1. No formal sample size calculation was performed, and all eligible cases were included. TCGA-BLCA served as the training set (n=406), and GSE13507 served as the primary external validation set (n=165). The primary endpoint was OS, defined as the time from diagnosis or enrollment to death; patients alive at the last follow-up were censored. Disease-specific survival (DSS), progression-free survival (PFS), and disease-free survival (DFS) were evaluated only as supportive secondary endpoints when corresponding public annotations were available and were not used for model development. In the TCGA cohort, the status variable was coded as fustat =1 for death events and fustat =0 for censoring, and follow-up time was recorded as futime (days).

Table 1. Characteristics of the development and validation cohorts.

Characteristic	TCGA-BLCA	GSE13507	GSE31684	IMvigor210
Cohort role	Development cohort	Primary external validation cohort	Additional external validation cohort	Exploratory immunotherapy-treated cohort
Data source	TCGA	GEO	GEO	IMvigor210CoreBiologies
Platform/data type	RNA-seq	GPL6102, Illumina Human-6 v2.0 Expression BeadChip, microarray	GPL570, Affymetrix Human Genome U133 Plus 2.0 Array, microarray	RNA-seq
Disease setting	Bladder urothelial carcinoma	Primary bladder cancer	High-risk bladder cancer managed by radical cystectomy	Metastatic urothelial carcinoma treated with atezolizumab
Sample size, n	406	165	93	348
Death events, n (%)	177 (43.6)	69 (41.8)	66 (71.0)	232 (66.7)
Primary endpoint used in this study	OS	OS	OS	OS
Treatment context	Not uniformly available	Not uniformly available	Radical cystectomy cohort	Anti-PD-L1 immunotherapy

GEO, Gene Expression Omnibus; OS, overall survival; PD-L1, programmed death-ligand 1; RNA-seq, RNA sequencing; TCGA, The Cancer Genome Atlas; TCGA-BLCA, The Cancer Genome Atlas-bladder cancer cohort.

Pyroptosis gene set and expression data preprocessing

A pyroptosis-related gene set was curated from core molecules frequently reported in published literature and reviews (17), and the resulting gene list is provided in Table S1. TCGA-BLCA RNA sequencing (RNA-seq) expression data were converted to transcripts per million (TPM) and transformed as log₂(TPM +1) before downstream analyses. For GEO microarray cohorts, publicly available processed expression matrices were used directly for analysis. Probes were mapped to official gene symbols according to the corresponding platform annotation files, and when multiple probes mapped to the same gene, the mean expression value was used to generate a single gene-level estimate. Genes without matched official symbols were removed. Gene identifiers were harmonized across cohorts to official gene symbols, and only genes available in the corresponding dataset (or intersecting genes for cross-cohort analyses) were retained for downstream analyses.

To avoid information leakage, feature selection, parameter estimation, and coefficient fitting for the prognostic model were performed exclusively in the TCGA training cohort. External cohorts were used only for validation and did not contribute to model training. For cross-platform merged visualization or non-modeling association analyses, batch effects were corrected after log transformation using the ComBat function in the sva package (18), with batch defined by data platform. In this study, subtype discovery, differentially expressed gene (DEG) screening between C1 and C2, and subsequent LASSO-Cox model construction were performed sequentially in the TCGA cohort, whereas all external cohorts were reserved for validation only.

Pyroptosis molecular subtyping, pathway activity, and immune infiltration analyses

In the TCGA cohort, differential expression of pyroptosis-related genes between tumor and normal tissues was assessed, followed by multiple testing correction using the Benjamini-Hochberg (BH) method [threshold: false discovery rate (FDR) <0.05; |log₂fold change (FC)| >1]. Univariate Cox regression was then performed to identify prognostically significant pyroptosis-related candidate genes with respect to OS. Based on the expression profiles of these candidate genes, non-negative matrix factorization (NMF) was applied for unsupervised clustering (19), and the optimal number of clusters (K) was determined by evaluating consensus matrix stability and rank survey metrics across candidate K values. Pathway activity was quantified using gene set variation analysis (GSVA) with the Kyoto Encyclopedia of Genes and Genomes (KEGG) gene set from the Molecular Signatures Database (MSigDB) (c2.cp.kegg.v7.4.symbols.gmt) (20). Immune infiltration was estimated using single-sample gene set enrichment analysis (ssGSEA) to compute enrichment scores for immune cell-related gene sets and compare differences across subtypes/risk groups (21). For analyses involving multiple comparisons, BH correction was applied (FDR <0.05).

Differential expression and functional enrichment analyses

Genome-wide differential expression analysis between the C1 and C2 subtypes was conducted using the limma package (22), with BH adjustment for multiple testing. DEGs were defined as those with FDR <0.05 and |log₂FC| >1. Gene Ontology (GO) and KEGG over-representation analyses were performed using the clusterProfiler package, with BH correction applied; FDR <0.05 was considered significant. The background gene universe was defined as all genes detectable in the cohort. The resulting DEG set was further used as the candidate feature pool for downstream prognostic model construction.

Construction, stratified validation, and extended analyses of the six-gene risk model

Using the DEGs identified between C1 and C2 as candidate input features, A least absolute shrinkage and selection operator Cox (LASSO-Cox) model was built in the TCGA training cohort using glmnet (23), with 10-fold cross-validation to determine the penalty parameter λ; the final model was selected using the minimum cross-validation error criterion (lambda.min). The risk score was calculated as Σ βi × Expi, where βi represents the LASSO-Cox coefficient and Expi represents the corresponding gene expression level. The full risk score equation was as follows: risk score = 0.006382 × ECM1 − 0.027190 × FER1L4 + 0.038683 × FKBP10 + 0.040513 × ANXA1 + 0.007344 × ARL4C − 0.022874 × CTSE. Patients were dichotomized into high- and low-risk groups using the median risk score of the TCGA training cohort as the predefined cutoff, which was directly applied to the external validation cohorts. This fixed-threshold strategy was intended to provide a stringent external evaluation of whether the training-derived model could retain prognostic stratification across independent datasets and platforms, rather than to assume strict numerical equivalence of absolute risk-score values across different expression platforms. No recalibration or model updating was performed in the validation cohorts. Survival differences between groups were assessed by Kaplan-Meier analysis and the log-rank test. Model discrimination was evaluated using time-dependent receiver operating characteristic (ROC) curves (at 1/3/5 years or specified time points), and calibration curves were used to assess agreement between predicted and observed outcomes. Associations between risk score and clinicopathological features (e.g., pathological grade, T stage) were evaluated using non-parametric tests, and univariate and multivariate Cox regression analyses were performed to examine the independent prognostic value of the risk score.

Somatic mutation profiles and TMB were analyzed based on TCGA mutation annotation format (MAF) files and visualized using maftools (24). TMB was defined as the number of non-synonymous mutations per megabase (mutations/Mb), using a fixed effective exome length of 38 Mb as the denominator. Immune checkpoint gene expression differences between risk groups were assessed with BH correction. Immunophenoscore (IPS) scores from The Cancer Immunome Atlas (TCIA) were used to infer potential benefit differences under distinct immunotherapy scenarios (10). CellMiner (NCI-60) was used to explore correlations between key model genes and drug sensitivity (25); Pearson correlation was applied with BH correction (FDR <0.05).

Statistical analysis and software environment

All analyses were conducted in R (version 4.3.2). Continuous variables were compared using the Wilcoxon rank-sum test (two groups) or the Kruskal-Wallis test (multiple groups). Categorical variables were compared using the Chi-squared test or Fisher’s exact test. Correlations were assessed using Pearson correlation. Survival analyses included Kaplan-Meier curves, the log-rank test, and Cox proportional hazards regression. Time-dependent ROC analyses were reported as areas under the curve (AUCs) with corresponding 95% confidence intervals (CIs). Unless otherwise specified, all tests were two-sided, with P<0.05 considered statistically significant; BH adjustment was applied for multiple testing, and FDR <0.05 was considered significant. Key R packages included limma, sva, survival, survminer, glmnet, timeROC, rms, NMF, GSVA, clusterProfiler, and maftools. Unless otherwise indicated, ns, not significant (P≥0.05); *, P<0.05; **, P<0.01; ***, P<0.001.

Results

Screening pyroptosis-related candidates and establishing NMF-based molecular subtypes

After applying the predefined inclusion criteria, the TCGA-BLCA cohort was used for model development and subtype discovery, GSE13507 served as the primary external validation cohort, and GSE31684 and IMvigor210 were used for additional validation; the cohort characteristics are summarized in Table 1. We first compared the expression of pyroptosis-related genes between tumor and normal tissues in TCGA-BLCA. Pyroptosis-related genes exhibited marked dysregulation overall (Figure 1A), and the curated gene list is provided in Table S1. We next performed univariate Cox regression for OS and identified five pyroptosis-related candidates significantly associated with prognosis [Figure 1B; hazard ratios (HRs), 95% CIs, and P values are summarized in Table S2], with their expression patterns visualized in Figure 1C.

Figure 1.

Screening of pyroptosis-related candidate genes and construction of NMF-based molecular subtypes. (A) Differential expression of pyroptosis-related genes between tumor and normal tissues in the TCGA-BLCA cohort. (B) Forest plot of five prognostic pyroptosis-related genes identified by univariate Cox regression for OS in TCGA-BLCA. (C) Heatmap showing expression patterns of the five prognostic genes across TCGA-BLCA samples. (D) Consensus matrix derived from NMF clustering based on the expression profiles of the five genes (K=2), indicating two stable pyroptosis-related molecular subtypes (C1 and C2). *, P<0.05; **, P<0.01; ***, P<0.001. CI, confidence interval; NMF, non-negative matrix factorization; OS, overall survival; TCGA-BLCA, The Cancer Genome Atlas-bladder cancer cohort.

Based on these five prognostic pyroptosis-related genes, we conducted unsupervised clustering using NMF. To reduce subjectivity in selecting the number of clusters (K), we compared clustering stability across candidate K values using consensus matrix consistency and rank survey metrics. K=2 yielded the most stable clustering solution (Figure 1D), with supporting evidence across K values shown in Figure S1A,S1B. Accordingly, patients were classified into two pyroptosis-related molecular subtypes (C1 and C2) for subsequent analyses of prognosis, immune microenvironment features, and biological mechanisms.

Pyroptosis subtypes show robust prognostic stratification

To evaluate the clinical relevance of the pyroptosis subtypes, we compared survival outcomes between C1 and C2. Kaplan-Meier analyses demonstrated significant differences in the primary endpoint OS and in the supportive secondary endpoints DSS and PFS between the two subtypes (Figure 2A-2C; log-rank P<0.001, P<0.001, and P=0.009, respectively), whereas DFS was not significantly different (Figure 2D; log-rank P=0.88). In addition, principal component analysis (PCA) indicated a degree of separation between the two subtypes at the transcriptome level (Figure 2E), supporting consistency of subtype-specific expression patterns. Given the significant prognostic differences, we next explored biological underpinnings by examining pathway activity and immune infiltration.

Figure 2.

Survival differences and transcriptomic separation of pyroptosis molecular subtypes. (A-D) Kaplan-Meier survival curves of C1 and C2 for OS, DSS, PFS, and DFS. (E) PCA showing transcriptome-level separation between C1 and C2. DFS, disease-free survival; DSS, disease-specific survival; OS, overall survival; PC, principal component; PCA, principal component analysis; PFS, progression-free survival.

Systematic differences in pathway activity and immune infiltration between subtypes

To investigate mechanisms underlying the survival divergence between C1 and C2, we quantified sample-level pathway activity using GSVA and compared pathway differences between the subtypes. GSVA based on the KEGG gene set (c2.cp.kegg.v7.4.symbols) revealed systematic differences in multiple tumor-related and immune-related pathways (Figure 3A; FDR <0.05). We further assessed immune cell infiltration using ssGSEA and observed significant differences across multiple immune cell enrichment scores between the two subtypes (Figure 3B; FDR <0.05), indicating close coupling between pyroptosis-related subtypes and the tumor immune microenvironment. These findings motivated further subtype-specific differential expression and enrichment analyses to provide gene-level evidence.

Figure 3.

Differences in pathway activity and immune infiltration between subtypes. (A) Heatmap of KEGG pathway activity scores calculated by GSVA, showing systematic differences in pathway activity between C1 and C2. (B) Differences in ssGSEA-derived immune cell infiltration scores between C1 and C2. ns, not significant (P≥0.05); **, P<0.01; ***, P<0.001. GSVA, gene set variation analysis; KEGG, Kyoto Encyclopedia of Genes and Genomes; ssGSEA, single-sample gene set enrichment analysis.

Differential expression and functional enrichment analyses between subtypes

To characterize molecular differences between C1 and C2, we performed genome-wide differential expression analysis and identified DEGs between the two subtypes. The top 30 upregulated and top 30 downregulated expression pattern of DEGs is shown in Figure 4A, and the full DEG list (log₂FC and FDR) is provided in table available at https://cdn.amegroups.cn/static/public/tau-2026-0327-1.xlsx. GO and KEGG enrichment analyses indicated that DEGs were enriched in biological processes and pathways related to tumor progression, extracellular matrix remodeling, inflammatory responses, and immune regulation (Figure 4B,4C; FDR <0.05). Collectively, these results support that pyroptosis subtyping is not only prognostically informative but also reflects distinct immune-related biological states. Given the need for clinically applicable quantitative risk assessment, we next developed and validated a subtype-derived six-gene prognostic signature based on genes differentially expressed between the pyroptosis-defined subtypes.

Figure 4.

DEGs and functional enrichment analyses between subtypes. (A) Heatmap of the top 30 upregulated and top 30 downregulated DEGs between C1 and C2. (B) Dot plot of GO enrichment analysis for DEGs (BH-adjusted; significant terms are shown). (C) Dot plot of KEGG enrichment analysis for DEGs (BH-adjusted; significant terms are shown). *, P<0.05; **, P<0.01. BH, Benjamini-Hochberg; BP, biological process; CC, cellular component; DEG, differentially expressed gene; GO, Gene Ontology; KEGG, Kyoto Encyclopedia of Genes and Genomes; M, metastasis; MF, molecular function; N, node; T, tumor.

Construction of a subtype-derived six-gene prognostic signature and stratified validation in external cohorts

To construct a subtype-derived prognostic model, we applied LASSO-Cox regression to the DEG candidate pool in the TCGA training cohort for feature shrinkage and selection. The cross-validation curve was used to determine the optimal penalty parameter λ (Figure 5A), and the coefficient path plot illustrates feature retention as λ varies (Figure 5B). Using 10-fold cross-validation and the λ_min rule, we obtained a final set of genes with non-zero coefficients and constructed a six-gene prognostic signature comprising ECM1, FER1L4, FKBP10, ANXA1, ARL4C, and CTSE.

Figure 5.

Construction and external validation of the six-gene prognostic risk model. (A) Ten-fold cross-validation curve of the LASSO-Cox model for selecting the penalty parameter λ. (B) LASSO coefficient profiles, with non-zero coefficients retained to define the prognostic signature. (C) Kaplan-Meier curve of OS for high- vs. low-risk groups in the TCGA training cohort. (D) Kaplan-Meier curve of OS for high- vs. low-risk groups in the GSE13507 external validation cohort (the training-cohort cutoff was fixed and applied; log-rank P values are shown in the figure). LASSO, least absolute shrinkage and selection operator; OS, overall survival; TCGA, The Cancer Genome Atlas.

The risk score was calculated as a weighted sum of the six gene expression levels, with model coefficients (β) provided in Table S3. The complete univariate Cox results of candidate genes are provided in Table S2. Using the prespecified median cutoff derived from the TCGA training cohort, patients were stratified into high- and low-risk groups. In the training cohort, the high-risk group showed significantly worse OS than the low-risk group (Figure 5C; log-rank P<0.001). Consistent stratification was observed in the GEO external validation cohort (Figure 5D; log-rank P=0.007), indicating that the six-gene model retained external prognostic stratification ability. We then systematically evaluated discrimination, calibration, and expression pattern consistency.

Predictive performance, calibration, and expression pattern validation of risk stratification

Time-dependent ROC analyses showed that the model achieved AUCs of 0.690, 0.665, and 0.655 at 1, 3, and 5 years in the training cohort, respectively, and calibration curves indicated good agreement between predicted and observed survival probabilities (Figure 6A). In the GSE13507 external validation cohort, the model yielded AUCs of 0.706, 0.650, and 0.615 at 1, 3, and 5 years (Figure 6B). The corresponding 95% CIs are provided in Table S4. Risk score distribution, survival status, and the six-gene expression heatmap further supported stratification consistency: mortality events increased with higher risk scores, and the six genes exhibited stable risk-associated expression patterns (Figure 6C). Similar trends were observed in the external cohort (Figure 6D), providing additional support for the reproducibility of risk stratification.

Figure 6.

Predictive performance, calibration, and visualization of risk stratification. (A) Time-dependent ROC curves (1/3/5-year AUC) and calibration curves in the TCGA training cohort. (B) Time-dependent ROC curves and calibration curves in the GSE13507 external validation cohort. (C) Risk score distribution, survival status, and heatmap of the six-gene signature in the TCGA training cohort. (D) Risk score distribution, survival status, and heatmap of the six-gene signature in the GSE13507 external validation cohort. *, P<0.05; ***, P<0.001. AUC, area under the curve; M, metastasis; N, node; OS, overall survival; ROC, receiver operating characteristic; T, tumor; TCGA, The Cancer Genome Atlas.

As supplementary evidence, PCA/t-distributed stochastic neighbor embedding (t-SNE) analyses showed separability between high- and low-risk groups in both the TCGA and GEO cohorts (Figure S2A-S2D), suggesting transcriptome-wide differences between risk strata. Given the stable stratification and predictive performance, we next examined clinical consistency and independent prognostic value of the risk score.

Clinical relevance of the risk score and its adjusted prognostic association in the TCGA cohort

To assess clinical consistency, after constructing the model in the TCGA cohort, we performed external validation in the GEO cohort and compared risk score distributions across pathological grade and T stage. In both TCGA and GEO, risk scores were significantly higher in high-grade tumors than in low-grade tumors. In the TCGA cohort (Figure 7A, left), risk scores were higher in T3–4 tumors than in T1–2 tumors; in the GEO cohort (Figure 7A, right), risk scores were higher in the more advanced stage group (T2–4) than in the early-stage group (T0–1), supporting consistent associations between the risk score and aggressive clinicopathological phenotypes. Univariate (Figure 7B) and multivariate Cox analyses (Figure 7C) in the TCGA cohort showed that the risk score remained significantly associated with overall survival (OS) after adjustment for clinicopathological covariates (multivariate HR =13.695; 95% CI: 3.339–56.170; P<0.001). Given the potential collinearity among covariates, this result should be interpreted cautiously. The variance inflation factor (VIF) values for age, gender, grade, stage, T, metastasis (M), node (N), and risk score were 1.150, 1.053, 1.375, 4.354, 2.148, 1.206, 3.019, and 1.291, respectively.

Figure 7.

Clinical relevance and independent prognostic value of the risk score. (A) Differences in risk scores across pathological grade and T stage in the TCGA cohort (left) and the GSE13507 cohort (right). (B) Univariate Cox regression analysis evaluating associations of the risk score and clinical variables with OS. (C) Multivariate Cox regression analysis evaluating the independent prognostic value of the risk score after adjusting for clinicopathological covariates. CI, confidence interval; M, metastasis; N, node; OS, overall survival; T, tumor; TCGA, The Cancer Genome Atlas.

To clarify the correspondence between the pyroptosis subtypes (C1/C2) and risk stratification (high/low risk), we used a Sankey diagram to visualize transitions among cluster assignment, risk score group, stage, and outcome status (Figure S3A). In addition, the correlation matrix between the risk score and ssGSEA immune infiltration scores is shown in Figure S3B, and the risk score differed significantly between C1 and C2 (Figure S3C). Together, these supplementary analyses suggest structural coupling among subtypes, risk, immune infiltration, and clinical progression, motivating further exploration of translational associations.

Genomic and therapeutic associations of risk stratification: TMB/mutation profiles, immune checkpoints, IPS, and drug-sensitivity clues

We first evaluated associations between risk stratification and genomic features, including TMB and somatic mutation profiles. The TMB-by-risk stratified survival analysis revealed significant prognostic differences across combined groups (Figure 8A; log-rank P<0.001), suggesting complementary prognostic information from the risk score and mutational background. Somatic mutation landscapes for high- and low-risk groups are shown as waterfall plots (Figure 8B,8C), indicating differences in the composition and frequency of frequently mutated genes.

Figure 8.

Genomic and translational extension analyses of risk stratification. (A) Joint survival analysis stratified by TMB and risk groups, including H-TMB/H-risk, H-TMB/L-risk, L-TMB/H-risk, and L-TMB/L-risk subgroups. (B,C) Waterfall plots of somatic mutation landscapes in the high- and low-risk groups. (D) Differential expression of 38 immune checkpoint genes between high- and low-risk groups in TCGA. (E) TCIA-derived IPS comparisons under four immunotherapy scenarios: top-left (CTLA4⁻/PD-1⁻), bottom-left (CTLA4⁺/PD-1⁻), top-right (CTLA4⁻/PD-1⁺), and bottom-right (CTLA4⁺/PD-1⁺). (F) CellMiner (NCI-60) drug-sensitivity exploration showing the top 20 candidate drugs with the strongest associations with model genes/risk score. *, P<0.05; **, P<0.01; ***, P<0.001. Cor, correlation; H, high; IPS, immunophenoscore; L, low; TCGA, The Cancer Genome Atlas; TCIA, The Cancer Immunome Atlas; TMB, tumor mutational burden.

In terms of immune regulation, 38 immune checkpoint genes showed broad differential expression between high- and low-risk groups in TCGA (Figure 8D; FDR <0.05). A similar trend was observed in the external GEO cohort as a replication analysis (Figure S4A).

We further assessed potential immunotherapy benefit differences using TCIA-derived IPS under distinct immunotherapy scenarios. IPS was significantly higher in the low-risk group under CTLA4⁻/PD-1⁻ and CTLA4⁺/PD-1⁻ settings (Figure 8E, left; P<0.001), whereas no significant difference was observed under CTLA4⁻/PD-1⁺ and CTLA4⁺/PD-1⁺ settings (Figure 8E, right; P=0.99 and P=0.43). In addition, using CellMiner (NCI-60), we evaluated correlations between the six model genes and drug sensitivity and presented the top 20 candidate drugs with the strongest associations (Figure 8F). The complete drug association results are provided in Table S5. Notably, TCIA/IPS and CellMiner analyses provide inferential evidence based on public scoring systems and cell-line pharmacogenomic resources and are presented here to generate therapeutic hypotheses and guide further validation.

Additional evaluation in GSE31684 and IMvigor210 cohorts

To further assess external generalizability, we performed additional validation in GSE31684 (n=93) and IMvigor210 (n=348). In both cohorts, the high-risk group exhibited significantly worse OS than the low-risk group (Figure 9A,9B; log-rank P≤0.001). Additional time-dependent ROC analyses indicated limited discriminative performance in these two cohorts. In GSE31684, the AUCs at 1, 3, and 5 years were 0.528, 0.590, and 0.579, respectively, while in IMvigor210 the AUCs at the prespecified 10- and 20-month time points were 0.543 and 0.583; the corresponding 95% CIs are summarized in Table S4. As supplementary evidence, risk score distributions, survival status, and the six-gene expression heatmaps for GSE31684 and IMvigor210 are shown in Figure S4B, complementing the Kaplan-Meier, ROC, and calibration results presented in Figure 9.

Figure 9.

Additional external validation. (A) Kaplan-Meier survival curves, time-dependent ROC curves, and calibration curves for risk stratification in the GSE31684 cohort. (B) Kaplan-Meier survival curves, time-dependent ROC curves, and calibration curves for risk stratification in the IMvigor210 cohort. AUC, area under the curve; OS, overall survival; ROC, receiver operating characteristic.

Discussion

In this study, we established an evidence chain linking “subtyping-mechanisms-modeling-translational clues” based on pyroptosis-related transcriptomic features in bladder cancer. First, using unsupervised clustering of prognostic pyroptosis-related genes, we classified patients into two molecular subtypes (C1/C2) and observed significant differences in OS, DSS, and PFS, whereas DFS was not significant. Pyroptosis is an inflammatory form of programmed cell death characterized by inflammasome activation, caspase cleavage, and gasdermin-mediated pore formation (17). Recent evidence in bladder cancer further suggests that induction of pyroptosis can activate antitumor immune responses, although persistent inflammatory signaling may also promote immunosuppression and tumor progression (26). Thus, it is biologically plausible that pyroptosis-related transcriptomic features delineate subgroups with distinct clinical outcomes. Regarding the non-significant DFS findings, we interpret it as potentially reflecting heterogeneity in endpoint definitions, event counts, and follow-up strategies that may limit statistical power (27); therefore, we place greater emphasis on OS/DSS/PFS as more stable endpoints.

To explain subtype-associated outcome differences, we further compared pathway activity and immune infiltration between C1 and C2. GSVA and ssGSEA results indicated systematic differences in immune-related pathways and immune-cell infiltration scores, and enrichment analyses of DEGs also implicated immune regulation, inflammatory responses, and tumor progression processes. These findings support a framework in which pyroptosis-related programs influence disease course by remodeling the tumor immune microenvironment (including immune recruitment, activation, and inhibitory axes) (28). Importantly, increased immune infiltration does not necessarily translate into improved prognosis (29); when immunosuppressive programs dominate or T-cell exhaustion is prevalent, higher immune-cell abundance may still accompany unfavorable outcomes (30).

Building on the subtype findings, we constructed and externally validated a six-gene prognostic signature derived from genes differentially expressed between the two pyroptosis-defined subtypes. Although the six model genes were not directly selected from the curated pyroptosis gene set, the signature still originated from subtype-specific transcriptional differences associated with pyroptosis-related molecular heterogeneity. Therefore, it may be better understood as a subtype-derived prognostic signature linked to pyroptosis-related biology, rather than a model composed exclusively of canonical pyroptosis genes. Although ECM1 and ARL4C had relatively small coefficients in the final model, coefficient magnitude alone does not necessarily reflect the overall contribution of a feature in a penalized Cox model, which may also depend on expression range and correlations among predictors. For this reason, the six-gene signature was retained as the original LASSO-selected solution. The model consistently stratified patients into high- and low-risk groups in the TCGA training cohort and GEO external validation cohort, with acceptable time-dependent ROC and calibration performance. Additional evaluation in GSE31684 and IMvigor210 showed that the signature retained some survival stratification ability in independent datasets, although the time-dependent ROC analyses indicated limited discriminative performance in these cohorts. Moreover, risk scores were significantly higher in high-grade and advanced T stage tumors and remained significant in multivariate Cox regression, suggesting that the risk score captures not only statistical prognostic signal but also a biological gradient related to aggressive clinicopathological phenotypes. For biological plausibility, prior studies provide supportive context: ECM1 has been linked to extracellular matrix remodeling and migration (31); ANXA1 is implicated in inflammatory regulation and tumor-related signaling (32); CTSE may relate to immune-associated processes such as antigen processing; and FER1L4, as a long non-coding RNA (lncRNA), may participate in transcriptional regulation or competing endogenous RNA (ceRNA) networks (33,34). Nevertheless, these arguments remain correlative, and mechanistic experiments are required to establish causality for key genes in the pyroptosis-immunity interplay.

From a translational perspective, we further evaluated associations between risk stratification and mutational background, immune regulation, and potential treatment responses. The combined TMB-by-risk stratification showed significant prognostic differences, and high/low-risk groups exhibited distinct mutational landscapes, suggesting that the risk score and genomic context may provide complementary prognostic information. Given that the relationship between TMB and prognosis or immunotherapy benefit is highly context-dependent across tumor types and immune states (35), we interpret these findings as enhanced explanatory power via joint stratification rather than as a unidirectional causal conclusion. We also observed broad differential expression of immune checkpoint genes between risk groups with replication in an external cohort, supporting the possibility that the risk score reflects changes in immunosuppressive programs and provides a background for immunotherapy response inference. Using TCIA-derived IPS, IPS was higher in the low-risk group under CTLA4⁻/PD-1⁻ and CTLA4⁺/PD-1⁻ scenarios, whereas differences were not significant under the remaining two scenarios. Because IPS is an inferential transcriptome-based score, we interpret it as generating hypotheses regarding context-specific immunotherapy benefit, and the enhanced validation in IMvigor210 provides additional support; nonetheless, prospective cohorts or real-world data are still needed for confirmation. In addition, CellMiner-based correlation analyses yielded exploratory drug-gene association clues linked to the model genes. These findings are hypothesis-generating only, particularly because several top-ranked agents do not have established clinical relevance in bladder cancer. Moreover, cell-line systems differ from the in vivo tumor context and cannot be directly equated with clinical drug sensitivity. Their value lies in providing testable directional hypotheses, which can be further evaluated in bladder cancer cell lines, organoids, or other pharmacogenomic resources that better approximate tumor biology.

Several limitations should be acknowledged. First, this is a retrospective analysis of public datasets, and cohort heterogeneity and missing clinical annotations may introduce bias. Second, cross-platform integration and expression standardization may affect absolute cutoff transferability. In addition, the analytical pipeline was carried out sequentially within the TCGA cohort, including subtype identification, DEG screening between C1 and C2, and subsequent prognostic model construction. Although the external cohorts were not involved in model development and were used only for validation, this design may still introduce within-cohort dependency and contribute to optimistic feature selection or performance estimation. Therefore, the six-gene signature should be interpreted as a subtype-derived exploratory model that requires further validation in fully independent datasets. Third, TCIA/IPS and CellMiner analyses provide inferential evidence rather than direct clinical outcomes. Fourth, no in vitro/in vivo experiments were performed to validate causal roles of key genes in pyroptosis-immune interactions. Because American Joint Committee on Cancer (AJCC) stage is derived in part from the T/N/M components, simultaneous inclusion of these variables in the same multivariable Cox model may introduce collinearity and affect coefficient stability. In addition, we did not formally compare the six-gene signature with a clinical-variables-only model or established molecular classifiers; therefore, its incremental prognostic value over existing approaches remains to be determined. Therefore, the full multivariable model in this study should be interpreted primarily as an exploratory adjustment analysis. Future studies should integrate single-cell/spatial transcriptomics to localize gene expression sources, perform functional experiments to test causal effects on pyroptosis and immunosuppressive axes, and validate the risk score in prospective clinical cohorts to assess its utility for guiding immunotherapy or combination strategies.

Overall, this study proposes pyroptosis-related molecular subtypes associated with prognosis in bladder cancer and establishes a six-gene risk score model validated across multiple independent cohorts. The model enables reproducible prognostic stratification and is linked to immune microenvironment features, mutational background, and potential therapeutic response clues, providing reproducible molecular evidence and testable hypotheses for risk assessment and individualized management of bladder cancer.

Conclusions

We identified two pyroptosis-related molecular subtypes with distinct survival and immune characteristics in bladder cancer and developed a six-gene prognostic signature derived from transcriptional differences between these subtypes. The signature was externally validated across multiple independent cohorts and was associated with clinicopathological aggressiveness, immune checkpoint expression, mutational context, and inferred therapeutic sensitivity.

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. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. Because all data were de-identified and publicly available, additional ethical approval and informed consent were not required.