Utilizing a novel mitochondrial-related gene signature for predicting the prognosis and immunological impact in bladder cancer

Abstract

Background

Mounting evidence highlights the critical role of mitochondrial dysfunction, driven by mitochondrial-related genes (MTRGs), in the development, progression, and therapeutic response of cancer. However, a comprehensive analysis linking specific Mitochondria-Related Gene Signature (MTRGS) to Bladder Cancer (BLCA) prognosis and immunotherapy efficacy remains largely unexplored. Therefore, this study aims to investigate the role of MTRGs in BLCA, construct and validate a novel MTRGs-based prognostic signature, and explore its potential for guiding personalized treatment strategies.

Materials and methods

Leveraging transcriptomic and clinical data from The Cancer Genome Atlas (TCGA-BLCA) cohort, we constructed a mitochondrial-related risk score model using LASSO, univariate and multivariate Cox regression analyses. This model was subsequently validated in an independent Gene Expression Omnibus (GEO) dataset. We then employed integrated bioinformatics approaches (implemented in R with online databases) to characterize features of the tumor microenvironment (TME), immune cell infiltration, Gene Set Enrichment Analysis (GSEA), tumor mutational burden (TMB), and drug sensitivity across different risk groups. Additionally, using data from public databases, we further verified our findings through single-cell RNA sequencing (scRNA-seq) analyses.

Results

Using 104 mitochondria-related differentially expressed genes (MTR-DEGs), unsupervised non-negative matrix factorization (NMF) clustering stratified BLCA patients into three molecular subtypes (Clusters 1–3). Survival analysis revealed that patients in Cluster 3 had significantly longer overall survival than those in Clusters 1 and 2. Our mitochondrial-related risk model incorporating six core genes (MAP1B, PYCR1, HSD3B1, KLK6, AKR1B15, and TAT) - exhibited robust prognostic capability (3-years AUC = 0.695 in TCGA-BLCA, 0.798 in GEO-GSE32894, 0.703 in GSE13507). The risk model revealed distinct immune infiltration patterns between high- and low-risk groups. Furthermore, Tumor Immune Dysfunction and Exclusion (TIDE) and immunophenotype score (IPS) analyses demonstrated that integrating risk scores with stromal/immune signatures significantly enhanced immunotherapy benefit prediction across BLCA risk-subgroups. Crucially, the model demonstrated predictive power for therapy response: low-risk patients showed potential benefit from immune checkpoint inhibitors, while high-risk patients exhibited heightened sensitivity to specific chemotherapy agents or targeted therapies (e.g., Tozasertib, Gemcitabine) and may require intensified regimens.

Conclusion

This validated mitochondrial risk model delivers a clinically actionable biomarker for BLCA prognosis stratification and guides personalized therapeutic selection, enabling precision treatment intensification.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12672-025-03927-9.

Introduction

Worldwide, bladder cancer is the ninth most frequently diagnosed cancer, with approximately 614,000 new cases and 220,000 deaths occurring in 2022 [1]. Bladder cancer is classified as either metastatic or non-metastatic. Non-metastatic bladder cancer is subdivided into non-muscle invasive bladder cancer (NMIBC) (75% of cases, high recurrence rate) and muscle invasive bladder cancer (MIBC) (significant risk of metastasis) [2, 3]. For non-muscle-invasive bladder cancer, the mainstay of treatment is complete resection of the tumor followed by induction and maintenance immunotherapy [4]. Whereas the mainstay of treatment for muscle-invasive advanced bladder cancer is cisplatin-based chemotherapy, and some novel therapies [5]. BLCA frequently progresses asymptomatically in early stages. Coupled with limited public awareness and diagnostic challenges, this leads to significant rates of advanced-stage presentation at initial diagnosis. Therefore, recurrence rates remain high after systemic treatment, and suboptimal clinical outcomes for advanced patients [6]. Identifying and exploring novel biomarkers is essential for developing reliable clinical outcome predictors and optimized therapeutic strategies in BLCA, underscoring the critical need for establishing robust risk predictive models. Given the pivotal role of mitochondrial dysfunction in cancer pathogenesis and treatment response, mitochondria-related genes (MTRGs) emerge as a highly promising source for such biomarker discovery.

Although most cancer cells rely on aerobic glycolysis rather than oxidative phosphorylation (the Warburg effect) [7], this metabolic shift paradoxically highlights profound mitochondrial reprogramming as a core component of the deregulated cellular energetics hallmark of cancer [8]. Mitochondria, far beyond being mere “powerhouses,” are master regulators of cellular metabolism and signaling. They orchestrate critical processes including energy production (notably ATP synthesis), lipid metabolism, nucleic acid metabolism, and redox balance [9]. They impart considerable flexibility for tumor cell growth and survival in otherwise harsh environments, such as during nutrient depletion, hypoxia, and cancer treatments, establishing them as key players in tumorigenesis [10]. Mutations in mtDNA and nuclear TCA cycle genes, frequently implicated in cancer, critically drive the metabolic reprogramming recognized as a major cancer hallmark [9, 11, 12]. Furthermore, mitochondria in cancer cells not only directly confer tumor malignant advantages such as increased proliferation rate, enhanced invasive and metastatic capabilities, and resistance to chemotherapy, but also indirectly drive tumor progression by modulating immune cell metabolism and activity within the tumor microenvironment (TME) [13, 14]. A recent study on differentiated thyroid cancer revealed that mitochondrial metabolic reprogramming, mediated by factors such as aspartate β-hydroxylase (ASPH), remodels the TME by reducing CD8+ T cell infiltration and increasing immunosuppressive macrophages, alongside a metabolic shift from oxidative phosphorylation to glycolysis in high-risk subtypes [15]. This underscores the role of MTRGs in orchestrating TME remodeling across cancers, including BLCA, through metabolic-immune crosstalk. Mitochondrial dysfunction and mtDNA mutations induce both increased mitochondrial ROS production (which transmit survival and proliferation signals associated with tumor promotion and maintenance) and low ATP levels with high cytosolic calcium levels, all of which act as apoptosis-inducing signals. Defects in apoptosis are among the major causes of tumorigenesis [16–18]. The TME consists of tumor cells and stromal cells, inflammatory cells, vasculature, and extracellular matrices (ECM) in complex tumor tissues. Tumor response to chemotherapy, immunotherapy, and targeted therapies is often contingent upon dynamic interactions between tumor cells and immunomodulators within the TME [19–21]. Critically, emerging evidence highlights that mitochondrial metabolic reprogramming can significantly regulate T cell function and exhaustion, thereby impacting the efficacy of immunotherapies [22]. In conclusion, mitochondrial dysfunction and the resulting apoptotic process (programmed cell death) will likely influence BLCA’s development and prognosis.

However, despite the established importance of mitochondrial dysfunction in cancer biology, its specific prognostic value and immunotherapeutic implications in BLCA remain insufficiently explored. Specifically, existing studies have not fully integrated the associations between MTRGs and BLCA prognosis, immune microenvironment characteristics, and treatment response. Therefore, in this study, we constructed a new risk score model using six signature genes to effectively predict the prognosis and immunotherapy effects in BLCA patients. Multiple databases were used to evaluate immune cell infiltration, immune checkpoint responsiveness, and drug sensitivity in different risk groups. Taken together, our risk score model can effectively help provide personalized and precise treatment strategies for BLCA patients.

Materials and methods

BLCA cohort characterization and mitochondrial signature analysis

MTRGs were systematically curated from two principal resources: the MitoCarta 3.0 database (https://www.broadinstitute.org/mitocarta/mitocarta30-inventory-mammalian-mitochondrial-proteins-and-pathways) [23], providing a comprehensive inventory of mammalian mitochondrial proteins across species, and the Molecular Signatures Database (MSigDB) through Gene Set Enrichment Analysis (GSEA, https://www.gsea-msigdb.org/gsea/index.jsp) [24, 25]. We obtained transcriptome profiles and corresponding clinical information from the Cancer Genome Atlas (TCGA) database (https://portal.gdc.cancer.gov), while RNA-sequencing data and survival outcomes for external validation were retrieved from the Gene Expression Omnibus (GEO) database (ID: GSE13507, N = 165; ID: GSE32894, N = 221; https://www.ncbi.nlm.nih.gov/geo/). To ensure data quality, we implemented rigorous filtering criteria: transcripts mapping to the same gene were averaged to obtain gene-level expression. For GEO microarray datasets, raw data were normalized with the limma::normalizeBetweenArrays() function. Probes were converted to gene symbols using platform annotations, and those mapping to the same gene were averaged. Duplicate samples, samples lacking key clinical data, or samples with survival time < 30 days were excluded.

Identification of DEGs and functional enrichment

We applied the “DESeq2” package of R (version 4.4.1) to perform differential expression analysis between normal and tumor samples from the training cohort, with the significance thresholds for differentially expressed genes (DEGs) set as |log₂ fold change| >2 and adjusted P < 0.05. Differential expression patterns were visualized through volcano plots using the “GdcVolcanoPlot” R package. Subsequently, mitochondria-related differentially expressed genes (MTR-DEGs) were identified via Venn diagram analysis, intersecting the DEGs with the comprehensively curated mitochondrial genes. To gain deeper insights into the potential functions of MTR-DEGs, we conducted three-tier Gene Ontology (GO) classification (biological processes, molecular functions, cellular components) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses using R packages “ClusterProfiler” and “org.Hs.eg.db”. For these enrichment analyses, the significance thresholds were set as P < 0.01 to ensure biological relevance.

Non-negative matrix factorization (NMF)-based molecular subtyping of BLCA

Based on the expression of 104 MTR-DEGs, molecular subtyping was performed via the non-negative matrix factorization (NMF) algorithm in the R package “NMF”. The optimal number of clusters (Rank) was determined by systematically evaluating cophenetic correlation coefficients, residual sum of squares (RSS), and dispersion metrics to ensure the stability of the decomposition model. After cluster assignment, principal component analysis (PCA) and t-distributed stochastic neighbor embedding (t-SNE) algorithms were used to reduce dimensionality and visualize inter-cluster divergence. These dimensionality reduction analyses were performed using the R packages “t-SNE”, “limma”, and “ggplot2” to generate spatial distribution plots. Kaplan-Meier (K-M) survival analysis was performed using the “survival” and “survminer” R packages for prognostic stratification, with systematic evaluation of associations between molecular clusters and clinicopathological features (age, gender, tumor stage) as well as survival outcomes. The infiltration level of 23 types of immune cells was assessed using the “GSVA” R package through a single sample gene set enrichment analysis (ssGSEA) [26].

Construction and validation of the prognostic mitochondrial-related risk score model

Stepwise prognostic modeling was conducted through a three-phase analytical cascade: (A) univariate Cox proportional hazards regression for preliminary screening, (B) LASSO regularization for dimension reduction, and (C) multivariate Cox regression for final model optimization, ultimately identifying prognostic-related genes to construct a novel MTRGs prognostic signature. The Mitochondrial-Related Risk Score was mathematically formulated as: Risk Score = ∑Coefi × Xi (Coefi: the regression coefficient, Xi: expression of each signature gene). The BLCA patients from the TCGA cohort were stratified into high- and low-risk groups according to the median risk score cutoff. K-M survival analysis was conducted to evaluate the correlation between different risk groups and overall survival (OS). The prognostic accuracy of the mitochondrial-related risk-scoring model was further evaluated by calculating the area under the receiver operating characteristic curve (AUC) using the R package “timeROC”. To evaluate the predictive performance of our Mitochondria-Related Gene Signature (MTRGS), we conducted a comprehensive comparison with three previously published gene models—Song et al. [27] (a 7-gene apoptosis-related signature), Li et al. [28] (a 5-gene Tumor-Infiltrating Lymphocyte (TIL)-related signature), and Zheng et al. [29] (a 6-gene propionate metabolism-related signature)—using the AUC at 1-, 3-, and 5-year overall survival time points.To ensure generalizability, external validation was performed in two independent BLCA cohorts from the GEO database (GSE13507 and GSE32894), where the risk score formula and median risk score cutoff derived from TCGA were consistently applied.

Establishment and evaluation of a nomogram

We performed univariate and multivariate COX regression analyses of risk scores and clinicopathological features to evaluate the potential of mitochondria-associated risk-scoring model as independent prognostic factors. A nomogram was constructed using the R package “rms” by integrating significant variables, and calibration curves were generated to assess the accuracy of the nomogram predictions.

Tumor microenvironment and immunity reaction analysis

To characterize the heterogeneity of the TME between the high- and low-risk groups, we employed the CIBERSORT algorithm to quantify infiltration levels of 22 immune cell subtypes and the ESTIMATE algorithm to calculate stromal, immune, ESTIMATE scores and tumor purity in each risk group. Furthermore, the likelihood of tumor immune escape was evaluated using the TIDE algorithm (http://tide.dfci.harvard.edu/), where elevated TIDE scores (validated predictors of immune escape) were associated with resistance to tumor immunotherapy and were less effective in immune checkpoint inhibition. Subsequently, we analyzed the expression levels of PD1 and CTLA4 between the high- and low-risk groups. The IPS of patients with BLCA from The Cancer Immunome Atlas (TCIA) (https://tcia.at/) were used to explore the role of the high- and low-risk groups in immunotherapy responses [30], thus constructing a comprehensive immunological profiling framework.

Drug sensitivity

To assess the model’s predictive capacity for therapy response, we performed a pharmacogenomic analysis using the Genomics of Drug Sensitivity in Cancer (GDSC) database (GDSC2 release) as a reference. Drug sensitivity was estimated via half-maximal inhibitory concentration (IC50) predictions generated by the oncopredict R package, which applies ridge regression models pre-trained on GDSC2 to infer IC50 values in our cohort based on gene expression-drug sensitivity associations. The GDSC2 IC50 matrix was first exponentially transformed to approximate normality. Our BLCA expression data were then integrated with the GDSC2 training set using Empirical Bayes batch correction (batchCorrect = ‘eb’) to minimize technical variation. For each drug, differential sensitivity between risk groups was assessed using a two-sided Student’s t-test, with a significance threshold of P < 0.05. Results were visualized using box plots.

Mutation analysis

To characterize mutational heterogeneity between the high- and low-risk groups, somatic mutation data were retrieved from the cBioPortal (cBio Cancer Genomics Portal) database (https://www.cbioportal.org/). Then, mutational profiling was performed using the R package “maftools” to generate waterfall plots for visualizing distinct mutation spectra.

GSEA enrichment analysis

DEGs between the high-risk and low-risk groups were identified by the differential analysis method described above. GO and KEGG enrichment analyses were then performed on these DEGs. For GSEA, we used the c2.cp.reactome.v7.0 gene set from the MSigDB as the target set (using 1000 permutations). Statistical significance was defined as meeting all three criteria: P < 0.05, adjusted P < 0.05, and q-value < 0.05.

Single-cell analysis

In our single-cell studies, we employed two datasets from the Tumor Immune Single-Cell Hub (TISCH) database - tumor samples (GSE130001) and normal samples (GSE129845) - to further validate the six signature genes and assess the differences in their expression levels and distribution patterns between tumor tissue microenvironment-related immune cells and various other stromal cell types versus those in normal tissue.

Statistical analysis

Data analysis was conducted using R (version 4.4.1). K-M survival curves were generated, and intergroup survival differences were evaluated using the log-rank test. Continuous variables were compared between low-risk and high-risk cohorts via parametric t-test or Wilcoxon rank-sum test, depending on normality assessment. Categorical variables were analyzed using chi-square test. Spearman’s correlation assessed variable associations, and prognostic model performance was quantified with the concordance index (C-index). Statistical significance was defined as *P < 0.05, **P < 0.01, and ***P < 0.001 (two-tailed).

Results

MTR-DEGs between normal and tumor samples and functional enrichment analysis in BLCA

The general workflow of this study is illustrated in Fig. 1. MTRGs were systematically curated by integrating the MitoCarta 3.0 database and MSigDB, and 2,030 candidate targets were identified for subsequent analysis (Additional file 2: Table S1). Differential expression analysis between tumor and normal specimens from TCGA identified 1,955 protein-coding genes (Fig. 2A) (Additional file 2: Table S2). Systematic integration of these DEGs with the mitochondrial genes revealed 104 MTR-DEGs (Fig. 2B) (Additional file 2: Table S3). GO and KEGG enrichment analyses were used to explore the mitochondrial regulatory landscape of 104 MTR-DEGs in BLCA (Additional file 2: Table S7). GO enrichment analysis demonstrated specific organization (Fig. 2C). In terms of cellular components, they are mainly related to the mitochondrial matrix, inner and outer mitochondrial membranes, and neuronal cells, etc. in terms of biological processes, they are related to membrane potential regulation and response to oxidative and chemical stress, etc. and in terms of molecular functions, they are related to tubulin binding, inhibitory phosphate-related functions, and binding of various small molecules, etc. KEGG analysis showed (Fig. 2D) that these genes are also involved in the IL-17 signaling pathway, apelin signaling pathway, cGMP-PKG signaling pathway, and lipid and atherosclerosis signaling pathway, etc.

Fig. 1.

The general workflow of this study

Fig. 2.

Identification and functional enrichment analysis of mitochondrial-related differentially expressed genes (MTR-DEGs) in TCGA-BLCA cohort. A Volcano plot of 1955 DEGs in BLCA tumors and normal groups, Significantly dysregulated genes were defined by |log₂ fold change| >2 and adjusted P < 0.05. B The Venn diagram showed that the overlap of 1955 DEGs and 2030 mitochondrial genes resulted in the identification of 104 hub genes. C, D The GO analysis for MTR-DEGs across cellular component, molecular function and biological process dimensions. The most enriched KEGG pathways of MTR-DEGs. Adjusted. Significant terms were selected with P < 0.01

Analysis of three molecular subtypes regarding clinicopathology and immune infiltration

Using expression profiles of 104 MTR-DEGs, we performed clustering analysis on 391 BLCA samples using the NMF algorithm to investigate MTR-DEGs expression patterns and their biological implications in BLCA. Clustering revealed that rank = 3 was optimal (Fig. 3A-B). Three molecular subtypes were subsequently identified (Additional file 2: Table S8): Cluster 1 (n = 91), Cluster 2 (n = 147), and Cluster 3 (n = 153), with corresponding MTR-DEGs expression heatmaps showing distinct transcriptional profiles among the subtypes (Fig. 3C). K-M survival analysis demonstrated significantly superior OS in cluster 3 compared to cluster 1 and 2 groups (Fig. 3D). Multivariate analysis revealed significant associations between molecular subtypes and clinicopathological parameters, including age stratification, gender distribution, and stage (Fig. 3E). Cluster 3 exhibited unique characteristics of the immune microenvironment, showing a marked reduction in 19/23 immune cell infiltrates (e.g., M2 macrophages, regulatory T cells (Tregs)) alongside a prominent enrichment of CD56dim natural killer (NK) cells compared to other subtypes (Fig. 3F). The favorable prognosis of Cluster 3, its predominant distribution in the low-risk group, and its distinct immune infiltration profile are highly consistent with a previous study [31]. To further validate cluster separation, Dimensionality reduction visualization by PCA and t-SNE confirmed distinct spatial segregation patterns (Fig. 3G-H). While clusters 1 and 2 showed partial overlap in the low-dimensional embedding space, both showed significant separation from cluster 3, confirming the molecular distinctness identified by clustering analysis.

Fig. 3.

Construction of molecular subtypes of MTR-DEGs in BLCA. A NMF clustering evaluation (rank = 2–10) via cophenetic, dispersion, residuals, RSS, and silhouette score identifies optimal subtyping dimensions. B Consensus map of NMF clustering at rank = 2–10. C Basis-pheatmap profiles gene expression signatures across subtypes. D Kaplan–Meier (KM) survival analysis of BLCA in molecular subtypes. P = 0.00031. E The clinicopathological heatmap includes age, survival state, stage and gender. F Boxplot comparing immune cell infiltration between Clusters. G, H PCA and t-SNE plots visualizing of the expression profile of MTR-DEGs among the three clusters. ***P < 0.001

Construction and validation of the prognostic mitochondrial-related risk-scoring model

Subsequently, to enhance the prognostic accuracy of MTR-DEGs and to adequately account for the differences between tumor and normal samples, we developed a model termed the “Mitochondrial-Related Risk Score”. Initial univariate Cox regression analysis of 104 MTR-DEGs identified 29 genes significantly associated with BLCA prognosis (P < 0.05). Subsequent dimension reduction through LASSO regression and multivariate Cox analysis refined the panel to six optimal predictors: MAP1B, PYCR1, HSD3B1, KLK6, AKR1B15, and TAT (Fig. 4A-C). The finalized prognostic model calculated individual risk scores using the formula: Risk score = (0.243 × MAP1B) + (0.254 × PYCR1) + (-0.217 × HSD3B1) + (0.148 × KLK6) + (0.260 × AKR1B15) + (-0.301 × TAT). Applying the median risk score as a cutoff, TCGA-BLCA patients were stratified into distinct high- and low-risk groups. K-M analysis revealed significantly poorer overall survival (OS) in high-risk patients ( P < 0.001) (Fig. 4D). Notably, risk score distribution and survival status plots demonstrated superior survival outcomes in low-risk patients (Fig. 4E-F). Time-dependent ROC analysis demonstrated predictive accuracy with area under the curve (AUC) values of 0.662, 0.695, and 0.714 for 1-, 3-, and 5-year overall survival, respectively (Fig. 4G). Furthermore, we performed correlational analysis through comparative expression box plots of the six signature genes between tumor and normal tissues (Fig. 4H), coupled with integrated analysis including clinicopathologic characteristics, signature gene expression patterns, and survival outcomes across high- and low-risk groups (Fig. 4I). Moreover, there were significant differences in the risk scores of the different clusters, as shown by the fact that cluster 3 has a significantly lower risk score than clusters 1 and 2 (Fig. 4J). Most patients in clusters 1 and 2 belonged to the high-risk group, whereas the majority in cluster 3 belonged to the low-risk group (Fig. 4K).

Fig. 4.

Construction of the prognostic mitochondrial-related risk-scoring model. A LASSO regression algorithm performs dimensionality reduction on 29 prognostic-related DEGs, showing coefficient decay dynamics. B Partial likelihood deviance plot pinpoints the optimal regularization parameter. C Forest plot displaying hazard ratios and confidence intervals for the signature genes in the prognostic model derived from multivariate Cox regression analysis. D Kaplan–Meier (KM) survival curves in the training set show significant stratification between high- and low-risk groups. P < 0.001. E, F Distribution of risk scores and survival status for patients in the internal training set. G Time-dependent ROC curves assess 1-, 3-, and 5-year OS prediction accuracy in the training set. H Gene expressions of the 6 signature genes in TCGA-BLCA. I Comparison of clinicopathologic characteristics and gene expression of the 6 signature genes between the high- and low-risk groups. ns: not significant, *P < 0.05, ****P < 0.0001. J Boxplot comparing risk scores across the three clusters. K Sankey diagram between clusters, risk groups, and survival outcomes. ns: not significant, ****P < 0.0001

We compared our MTRGS with three existing BLCA signatures. As shown in Additional file 1: Figure S1A-C, ROC curves were compared across the four Signatures at 1, 3, and 5 years. AUC values were summarized in a combined bar-line chart (Additional file 1: Figure S1D) and a heatmap (Additional file 1: Figure S1E) for visual comparison. At 1 year (Additional file 1: Figure S1A), Li’s Signature [28] performed best (AUC = 0.703), followed by Song [27] (0.683), MTRGS (0.662), and Zheng [29] (0.618). By 3 years (Additional file 1: Figure S1B), MTRGS achieved the highest AUC (0.695), exceeding Zheng [29] (0.660), Song [27] (0.659), and Li [28] (0.612). This advantage persisted at 5 years (Additional file 1: Figure S1C), with MTRGS remaining top (AUC = 0.714), followed by Song [27] (0.677), Zheng [29] (0.636), and Li [28] (0.594). The bar-line chart (Additional file 1: Figure S1D) reveals distinct temporal trends: MTRGS improved consistently over time, Li [28] declined steadily, Song [27] remained stable, and Zheng [29] peaked at 3 years before declining. The heatmap (Additional file 1: Figure S1E) further highlights that MTRGS not only achieved the highest AUC at 3 and 5 years but also showed the most robust time-dependent improvement.

To rigorously validate the prognostic robustness, we applied the same risk score formula and median cutoff to two independent validation cohorts (GSE32894 and GSE13507). External validation in both cohorts successfully replicated the prognostic stratification, with high-risk patients exhibiting significantly worse survival outcomes (P = 0.001 for GSE32894, Fig. 5A; P < 0.001 for GSE13507, Fig. 5C). Time-dependent ROC analysis demonstrated consistent predictive accuracy across time points. In the GSE32894 cohort, the AUC values for 1-, 3-, and 5-year survival were 0.719, 0.798, and 0.755, respectively (Fig. 5B). Similarly, the GSE13507 cohort yielded strong and consistent AUC values of 0.713, 0.703, and 0.700 for the same time points (Fig. 5D). The concordance in risk patterns across both independent cohorts, combined with the temporal AUC stability, conclusively validates the model’s clinical utility as a stable and generalizable prognosticator for BLCA outcomes.

Fig. 5.

Nomogram to predict the survival probability of BLCA patients. A, B Kaplan–Meier (KM) survival curves and ROC curves in the internal test set (GSE32894). P = 0.001. C, D Kaplan–Meier (KM) survival curves and ROC curves in the internal test set (GSE13507). P < 0.001 E, F Forest plot showing the results of Uni- and multiCox analyses of clinical features, revealing that the risk score was a critical independent prognostic factor among clinicopathological factors for overall survival in BLCA. G Nomogram integrating age, gender, stage, T/N stage, and risk score enables visualizable individual survival probability prediction. H The calibration diagram for 1-, 3-, and 5-year survival showing high consistency between predicted and observed outcomes. **P < 0.01, ***P < 0.001

Establishment and evaluation of a nomogram

Independent prognostic analysis was conducted by comprehensively integrating clinical characteristics and risk scores to evaluate the prognostic value of the risk-scoring model. Univariate COX regression analysis showed that except for gender, other characteristics significantly impacted OS (Fig. 5E). Multivariate COX regression analysis further identified age (P = 0.001, HR = 1.030), T stage (P = 0.044, HR = 1.396) and risk score (P < 0.001, HR = 2.461) as independent prognostic indicators for BLCA (Fig. 5F). Thus, we developed a nomogram to precisely predict 1-, 3-, and 5-year survival probabilities in BLCA patients (Fig. 5G). The calibration curve (Fig. 5H) revealed a good predictive potential of the nomogram.

Tumor microenvironment and immunity reaction analysis

A comprehensive analysis was conducted to elucidate the intricate interplay between risk stratification and the immunological landscape of BLCA. This analysis linked the tumor’s immune microenvironment to the patient’s prognosis and the therapeutic efficacy of immunotherapy. Analysis of the ESTIMATE database revealed a substantial augmentation in Immune, Stromal, and Estimate scores (P < 0.001), concomitant with a decline in tumor purity in the high-risk group(P < 0.001) (Fig. 6A-D). This finding underscores the intricate dynamics of the immune microenvironment that influence BLCA prognosis. The present study utilized the CIBERSORT database to analyze the infiltration levels of 22 distinct immune cell subtypes within the high- and low-risk groups (Fig. 6E) (Additional file 2: Table S9). The findings of this analysis are consistent with those of preceding studies, which demonstrated a marked predominance of M0, M1, and M2 macrophages in the high-risk group. Elevated macrophage infiltration has been observed to be associated with a poor prognosis in patients with BLCA. Conversely, the low-risk group exhibited elevated levels of Tregs, CD8 + T cells, and activated dendritic cells (DCs). The relatively higher Treg infiltration within the low-risk group may reflect a TME that is poised to mitigate an overexuberant immune response, thereby potentially curbing tumor aggressiveness through an immunosuppressive mechanism. The heatmap of the six signature genes with immune cell subtypes is shown in Fig. 6F. Correlation analysis revealed significant associations between risk scores and immune infiltration patterns. Positive correlations were identified with Macrophages M0 (R = 0.25, P = 0.00075), resting Mast cells (R = 0.17, P = 0.019), and resting CD4 + memory T cells (R = 0.15, P = 0.048) (Fig. 6G-I). Conversely, inverse relationships emerged with resting NK cells (R = -0.16, P = 0.03), Plasma cells (R = -0.23, P = 0.0016), activated CD4 + memory T cells (R = -0.16, P = 0.033), and CD8 + T cells (R = -0.22, P = 0.003) (Fig. 6J-M).

Fig. 6.

TME signatures and immune profiles between the low- and high-risk groups. A-D Results of the analysis of the correlation between risk scores and TME signatures in the high-risk and low-risk groups. E CIBERSORT analysis: An illustration of the infiltration levels of 22 types of immune cells for high- and low-risk groups. F Heatmap of correlations between immune cell infiltration levels and the expression of 6 signature genes. G-M Correlation analysis between the infiltration levels of different immune cells and risk scores. *P < 0.05, **P < 0.01, ***P < 0.001

The TIDE score (Additional file 2: Table S10) was significantly higher in the high-risk group than in the low-risk group (Fig. 7A), and the response rate to immunotherapy was significantly lower (19%) than in the low-risk group (57%) (Fig. 7B). An increase in risk score was correlated with higher proportions of response failures, and the above results suggest a positive correlation between the low-risk score and response to immunotherapy (Fig. 7C). Furthermore, higher IPS (a reliable indicator of response to immunotherapy) scores indicate better sensitivity to immunotherapy. The IPS was significantly higher in the low-risk group, particularly in response to CTLA-4 immunotherapy (Fig. 7D-G). This elevation suggests that low-risk patients may derive substantial benefit from CTLA-4 targeted therapy, highlighting the potential for personalized treatment strategies. Collectively, our findings robustly establish a compelling correlation between risk scores and the propensity to respond to immunotherapy. These insights advocate for the integration of risk stratification in the formulation of immunotherapy protocols and the prognostication of patient outcomes, paving the way for more tailored and efficacious therapeutic approaches in BLCA management.

Fig. 7.

Prediction of immunotherapy response based on risk scores between high- and low-risk groups. A Differences in TIDE score between the two risk groups. B Distribution of TIDE score for patients in BLCA. Proportion of the patients who non-respond and respond to immunotherapy within the high- and low-risk groups. C Association between immunotherapy response success and risk scores. D-G The difference in the IPS between different risk groups: CTLA4- and PD1-negative, CTLA4-negative and PD1-positive, CTLA4-positive and PD1-negative, and both CTLA4- and PD1-positive. ns: not significant, ***P < 0.001

Drug sensitivity

We predicted IC50 values for 198 agents in both high- and low-risk groups of BLCA patients (Additional file 2: Table S11). After excluding eight agents that showed no significant difference (P ≥ 0.05), 173 drugs were found to exhibit greater sensitivity in the low-risk group. In contrast, high-risk patients showed significantly lower IC50 values for 17 compounds (Fig. 8), including Gemcitabine, Sepantronium bromide, AZD7762, AZD6482, AZD8055, Foretinib, BI.2536, Staurosporine, ULK1_4989, BMS.754,807, GNE.317, SB216763, RO.3306, GSK269962A, SB505124, Tozasertib, and Pictilisib. These results suggest that high-risk patients may have increased therapeutic responsiveness to these agents.

Fig. 8.

Drug sensitivity analysis in high- and low-risk groups. Boxplots showing the IC50 values for 17 therapeutic agents in two risk groups. *P < 0.05, **P < 0.01, ****P < 0.0001

Mutation analysis

To understand the relationship between risk scores and gene mutations, we mapped mutations in the top 15 most frequently mutated genes in the high- and low-risk groups. TP53 and TTN were the most frequently mutated genes in the high- and low-risk groups, with mutation rates of 56% and 47%, and 49% and 43%, respectively (Fig. 9A-B). These findings highlight the critical role of TP53 in bladder cancer, consistent with its recognized tumor suppressor function, and the high prevalence of TTN mutations suggests a potential role in tumorigenesis and warrants further investigation. Notably, the higher mutation rate of TP53 in the high-risk group suggests its potential synergistic effect with mitochondrial dysfunction in promoting tumor progression [32]. In addition, genes such as MUC16 and KMT2D showed high mutation frequencies, suggesting that they may be involved in the pathogenesis of bladder cancer. The mutational spectrum, which includes nonsense, missense, splice site, frameshift and multiple hit events, provides a comprehensive genetic map that is essential for understanding the molecular basis of the disease and therapeutic strategies.

Fig. 9.

Mutation status in the high- and the low-risk groups in BLCA. The top 15 genes, ranked by mutation frequency, in the high- and low-risk groups, respectively

GSEA enrichment analysis

The GSEA uncovered significant dysregulation of MTRGs and pathways in BLCA, emphasizing their likely contribution to tumor development and immune evasion (Additional file 2: Table S12). The 224 DEGs in high- and low-risk groups are involved in essential biological processes, including epidermal cell differentiation, skin development, keratinization, drug metabolism, chemical carcinogenesis, and retinol metabolism (Fig. 10A-B). These processes are fundamental to tumor growth and progression. For example, alterations in epidermal cell differentiation and keratinization might disrupt the structural and functional integrity of the bladder epithelium, thereby promoting tumor invasion. Additionally, these core genes are implicated in tumor immune evasion. Changes in drug metabolism and chemical carcinogenesis pathways, for instance, might alter the TME, affecting immune cell infiltration and function. The up-regulated pathways related to extracellular matrix remodeling and collagen metabolism, notably influenced by genes such as KLK6, may enhance tumor cell migration and metastasis (Fig. 10C). The down-regulated pathways associated with mitochondrial function, protein synthesis, and RNA processing suggest a disruption in cellular homeostasis, potentially facilitating tumor development (Fig. 10D). Dysregulation of these genes and pathways could suppress antitumor immune responses, enabling tumor cells to evade immune surveillance.

Fig. 10.

Functional and pathway enrichment analysis in the high-and low-risk groups. A Circle Map (GO): The right semicircle uses color-coded bands to represent six significantly enriched Gene Ontology (GO) categories, encompassing Biological Processes (BP), Cellular Components (CC), and Molecular Functions (MF). Genes listed in the left semicircle were the source of enrichment for these categories. B Circle Map (KEGG): Similarly structured, the right semicircle displays the top six significantly enriched KEGG pathways, identified through analysis of the gene set presented in the left semicircle. C, D These panels illustrate the top 5 enriched KEGG pathways specific to each of the two risk groups, as determined on GSEA. E Distinct pathway activation profiles of the high- and low-risk groups by GSEA

Further GSEA analysis of the high- and low-risk groups revealed distinct pathway activation profiles (Fig. 10E). The low-risk group showed significant enrichment in homeostatic regulatory pathways, including HALLMARK_BILE_ACID_METABOLISM, HALLMARK_OXIDATIVE_PHOSPHORYLATION, HALLMARK_FATTY_ACID_METABOLISM, and HALLMARK_PEROXISOME. In contrast, the high-risk group showed significant enrichment in cancer-associated pathways such as HALLMARK_WNT_BETA_CATENIN_SIGNALING, HALLMARK_HEDGEHOG_SIGNALING, HALLMARK_EPITHELIAL_MESENCHYMAL_TRANSITION, HALLMARK_HYPOXIA, and HALLMARK_ANGIOGENESIS, along with inflammatory pathways including HALLMARK_INFLAMMATORY_RESPONSE and HALLMARK_IL2_STAT5_SIGNALING.

Single-cell analysis

Comparative analysis of the two single-cell RNA sequencing datasets (Fig. 11) revealed tumor-specific dysregulation of MAP1B and PYCR1 expression profiles. In normal tissue, MAP1B was ubiquitously expressed in fibroblasts (FBs) and smooth muscle cells (SMCs) compartments, with attenuated signals detected in epithelial and endothelial lineages. Notably, tumor tissues exhibited substantial global downregulation of MAP1B, although expression persisted in endothelial cells, FBs, and SMCs. PYCR1 showed fibroblast-enriched expression in normal tissue, with low expression in epithelial populations. In contrast, malignant transformation triggered robust PYCR1 upregulation in both epithelial and FBs compartments, with epithelial upregulation being particularly pronounced. AKR1B15 and KLK6 showed low expression in normal tissue FBs, while TAT and HSD3B1 were undetectable. In the TME, AKR1B15, KLK6, displayed trace-level expression specifically within epithelial cell populations. Interestingly, the TME supported ectopic activation of TAT and HSD3B1, albeit at levels restricted to epithelial subpopulations.

Fig. 11.

Expression of the 6 signature genes in two single-cell datasets, BLCA tumor (GSE130001) and normal (GSE129845) tissues

Discussion

Mounting evidence substantiates that mitochondrial dysfunction is strongly associated with various human malignancies and that mitochondrial dysfunction may be involved in tumor metastasis and drug resistance through pathways activated during the early stages of oncogenesis and progression [33, 34]. Notably, recent investigations have confirmed that dysregulation of MTRGs directly mediates mitochondrial dysfunction, driving cancer progression through mechanisms such as metabolic reprogramming [35]. MTRGs have emerged as key players in cancer proliferation and migration, holding significant promise as diagnostic and prognostic biomarkers and therapeutic targets [36–38]. Current investigations into the specific role of MTRGs in BLCA pathogenesis and their utility for clinical prediction remain notably underdeveloped. In particular, the construction of robust prognostic models and personalized therapeutic strategies leveraging MTRGS for BLCA patients is lacking.Therefore, our study aims to establish novel diagnostic-therapeutic frameworks for BLCA patients by identifying mitochondrial biomarkers and developing a validated prognostic signature.

With the advancement and development of high-throughput sequencing and bioinformatics technologies, the study of the mechanisms of tumor development, proliferation and metastasis and the construction of predictive models has entered a completely new phase [39–41]. In this study, we screened 104 MTR-DEGs from the TCGA BLCA cohort, most of which were enriched in ‘membrane potential regulation, response to oxidative and chemical stress ’, ‘inhibitory phosphate-related functions’, ‘IL-17 signalling pathway’ and ‘cGMP-PKG signalling pathway’, further verifying that MTR-DEGs play an important regulatory role in the homeostatic maintenance and disorders of mitochondrial function. Genes associated with survival in BLCA patients were screened by univariate Cox regression analysis. Subsequently, six mitochondria-related prognostic genes (MAP1B, PYCR1, HSD3B1, KLK6, AKR1B15, and TAT) were further identified using Lasso regression and stepwise multifactorial COX and used to construct prognostic models. By analysing the clinicopathological features, prognosis and the expression differences of six signature genes between the high and low risk groups, it was concluded that MAP1B, PYCR1, KLK6 and AKR1B15 were highly expressed in the high risk group, which may be tumor promoting factors, and HSD3B1 and TAT were downregulated in the low risk group, which may be tumor suppressor factors.

The core genes involved in building the model: MAP1B is a neuron-specific microtubule-associated protein (MAP) involved in cross-talk between microtubules and actin filaments [42]. Research by Inoue et al. found the high expression of MAP1B is closely associated with invasion, metastasis, and tumorigenesis in breast cancer cells and may be a potential diagnostic and therapeutic target for TNBC [43]. Pyrroline-5‐carboxylate reductase 1 (PYCR1) is a key biosynthetic enzyme for proline synthesis [44]. High expression of PYCR1 is associated with poor prognosis in hepatocellular carcinoma, breast cancer and prostate cancer [45–47]. Overexpression of PYCR1 promotes bladder cancer cell proliferation and invasion through modulation of Akt/Wnt/β-catenin signaling [48]. Kinin-releasing enzyme-related peptidase (KLK) 6, a member of the serine protease family, has trypsin- or chymotrypsin-like activities [49]. KLK6 mediates E-cadherin shedding and enhances cellular proliferation, migration, and invasive capacity [50]. The HSD3B1 gene encodes 3β-hydroxysteroid dehydrogenase-1 (3βHSD1), an enzyme critical for converting adrenal androgen precursors to dihydrotestosterone (DHT), and can serve as a potential predictive biomarker in advanced prostate cancer [51, 52]. Tyrosine aminotransferase (TAT) is an enzyme pivotal in the catabolism of tyrosine. It predominantly localizes to the mitochondria and exerts tumor-suppressive effects by inducing cell cycle arrest in breast cancer cells [53–55]. The biological functions of AKR1B15 in cancer remain poorly characterized, with no established role in tumorigenesis reported to date.

Immunotherapy has emerged as a pivotal component of cancer treatment. A critical determinant of therapeutic efficacy lies in elucidating the complex interactions between tumor cells, stromal cells and immune cell infiltration within the TME, particularly the essential role of immune infiltration in driving tumor progression and mediating immune escape mechanisms. These dynamic cell interactions fundamentally shape immunotherapy outcomes across cancer subtypes. Given the critical importance of immunotherapy, we systematically compared the differences in tumor tissue composition and immune cell infiltration between high-risk and low-risk groups. In the low-risk group, increased infiltration of immune cells and stromal cells was observed, accompanied by decreased tumor purity, which is associated with a better prognosis. Notably, the infiltration levels of M0, M1, and M2 macrophages were significantly elevated in the high-risk group. It has been shown that the substantial accumulation of M0 macrophages predicts a worse prognosis in BLCA [56]. M1 macrophages (classically activated macrophages) are generally recognized as tumor suppressor cells and are usually positively associated with favorable clinical outcomes in many cancers [57]. M2 macrophages (alternatively activated macrophages or tumor-associated macrophages) promote tumor progression through integrated mechanisms: they exert immunosuppressive effects by secreting immunoregulatory molecules into the TME, influence cancer cell proliferation and metastasis, via high CD68 expression and loss of E-cadherin expression to guide epithelial-mesenchymal transition (EMT), and induce tumor angiogenesis, collectively driving their multifaceted pro-tumorigenic functions [58–61]. The concurrent elevation of both M1 and M2 subtypes in high-risk tumors reflects an immune dysregulation driven by mitochondrial dysfunction. Mitochondrial impairment induces dual polarization effects: (a) Oxidized mtDNA activates cGAS-STING signaling, recruiting M1 macrophages via pro-inflammatory cytokines (TNF-α/IL-12) as compensatory anti-tumor response [62, 63]; (b) Fatty acid accumulation enhances CPT1A-mediated FAO, activating PPARγ to upregulate M2 markers (Arg1/IL-10) and immunosuppression [64]. Concurrently, mtROS exhibits dual roles: acute mtROS stabilizes HIF-1α/NF-κB maintaining M1 polarization [65], while chronic mtROS triggers ERO1α-UPRER stress increasing PD-L1 to enhance M2 immunosuppression [66, 67]. The mixed M1/M2 phenotype observed in BLCA may necessitate further elucidation through comprehensive analysis using specific M1 and M2 phenotype markers. In the low-risk group, there is a high level of infiltration of follicular helper T cells (Tfh), CD8 + T cells, dendritic cells (DCs), and plasma cells. Previous studies have demonstrated that in synergy dendritic cells (DCs), increased infiltration levels of CD8 + T cells are associated with a favorable prognosis in bladder cancer. It can also modulate B cell-mediated humoral immune responses by promoting the induction of Tfh cells [68, 69]. Additionally, regulatory T cells (Tregs) exhibit high infiltration in the low-risk group, indicating that Tregs may contribute to microenvironment regulation via a dynamic balance mechanism, favoring the maintenance of immune homeostasis rather than globally suppressing anti-tumor immunity. Our findings reveal complex heterogeneity within TME, indicating that distinct risk groups may originate from differential infiltration patterns of tumor-infiltrating immune cells (TIICs) and distinct TME remodeling states. Furthermore, our results show the pivotal role of TIICs in determining the response to immunotherapy. Our TIDE and IPS analyses showed that combining the MTRG-based risk score with stromal/immune signatures improved immunotherapy benefit prediction in BLCA. This is consistent with a lung adenocarcinoma study, where integrating a PTM-related signature featuring B4GALT2 with immune phenotypes enhanced response prediction [70].

This study is subject to several limitations. Although an independent training–validation approach was employed to reduce technical bias by building and tuning the model exclusively in the TCGA cohort before external validation in GEO datasets, inherent batch effects between sequencing and microarray platforms may still affect gene expression quantification. While the model showed consistent performance across cohorts, more refined calibration and normalization methods could enhance its cross-platform robustness. Furthermore, despite encouraging internal validation results, the potential for overfitting remains, and future studies involving independent cohorts of diverse ethnic backgrounds and disease stages are essential to confirm generalizability. Additionally, as all analyses relied on retrospective publicly available data, prospective multicenter studies will be necessary to validate clinical utility and reduce selection bias. Finally, although computationally validated, the MTRGs-based prognostic signature requires experimental and functional studies to clarify the mechanistic contributions of these genes to BLCA progression. Notwithstanding these limitations, our work provides a conceptual foundation for the future development of mitochondrial-targeted therapies in BLCA.

Conclusion

This study identifies a mitochondrial-related gene signature (MAP1B, PYCR1, HSD3B1, KLK6, AKR1B15, and TAT) that stratifies BLCA patients into distinct risk groups with divergent prognoses, immune microenvironment characteristics, and therapeutic responsiveness. High-risk patients exhibit mitochondrial dysfunction linked to immune suppression and chemotherapy resistance, while low-risk patients display active immune infiltration and stromal interactions, suggesting immunotherapy response. The model underscores mitochondrial pathways as central drivers of BLCA progression and highlights their dual role in tumor biology and treatment resistance. Clinical translation requires prospective validation and functional studies to refine personalized strategies targeting mitochondrial dysregulation.

Electronic Supplementary Material

Below is the link to the electronic supplementary material.

Author contributions

- (I) Conception and design: G Wu, C Fan, J Tian; (II) Administrative support: All authors; (III) Provision of study materials or patients: G Wu, C Fan, J Tian; (IV) Collection and assembly of data: G Wu, J Tian; (V) Data analysis and interpretation: All authors; (VI) Manuscript writing: All authors; (VII) Final approval of manuscript: All authors.

Funding

The author(s) declare that financial support was received for the research, authorship, and/or publication of this article. This study was supported by [1] Science and Technology Project of Chengguan District, Lanzhou City, Gansu Province Science and Technology Bureau (Project Number 2017KJGG0052); [2] Cuiying Graduate Supervisor Applicant Training Program of Lanzhou University Second Hospital (Project Number 201704); [3] Lanzhou City Talent Innovation and Entrepreneurship Project (Project Number 2019-RC-37).

Data availability

The data are provided within the Supplementary Information files. The public data used in this study are available from the TCGA (https://portal.gdc.cancer.gov) and GEO (https://www.ncbi.nlm.nih.gov/geo/) databases. Detailed information can be retrieved through the official portals of these databases. For further inquiries regarding data usage, please contact the corresponding author.

Declarations

Ethics approval and consent to participate

This study was based on publicly available transcriptomic and clinical data from the TCGA-BLCA and GEO-GSE13507, GSE32894 datasets. As per the guidelines for using public datasets, formal ethical approval from an institutional review board and individual patient consent were not required for this study. Clinical trial number: not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.