Prognostic and immunological implications of a mitophagy-associated gene signature in bladder cancer

Abstract

Background

The pronounced heterogeneity of bladder cancer (BLCA) drives divergent patient outcomes and therapy responses. Mitophagy, a pivotal cellular quality-control and metabolic mechanism, modulates the dynamics of the tumor microenvironment (TME). Its prognostic significance and potential for guiding precision oncology, however, remain incompletely defined. This investigation aimed to evaluate the translational utility of mitophagy-associated signatures for the risk stratification of BLCA, profiling of TME, and prediction of therapeutic response.

Methods

Using The Cancer Genome Atlas (TCGA)-BLCA cohort, a prognostic signature was established via least absolute shrinkage and selection operator (LASSO) and multivariable Cox regression analyses. The robustness of the signature was externally validated using two independent cohorts from the Gene Expression Omnibus (GSE32894, n=224; GSE31684, n=93). A nomogram was established to estimate survival probability. A multi-faceted analysis of TME features was conducted to compare the subgroups, integrating three algorithms: Cell-Type Identification by Estimating Relative Subsets of RNA Transcripts (CIBERSORT), Tumor Immune Dysfunction and Exclusion (TIDE), and Estimation of STromal and Immune cells in MAlignant Tumor tissues using Expression data (ESTIMATE). Drug sensitivity and molecular alteration profiles were interrogated to identify potential personalized therapeutic avenues.

Results

A five-core-gene signature (NTN4, AKR1B1, FBXW7, MYH10, and IGF2BP3) was established. Across all cohorts, high-risk individuals exhibited notably shorter overall survival (OS) (TCGA-BLCA: P<0.001; GSE32894: P=0.003; GSE31684: P=0.03). The areas under the curves (AUCs) for 1-, 3-, and 5-year survival were 0.645–0.654 (TCGA-BLCA), 0.625–0.785 (GSE32894), and 0.567–0.651 (GSE31684). The TME analysis suggested that high-risk tumors were associated with elevated stromal scores, enrichment of M2 macrophages, and a predicted T-cell exclusion phenotype. Conversely, low-risk individuals exhibited an ‘immune-inflamed’ profile, marked by elevated microsatellite instability (MSI) and infiltration of CD8⁺ T cells. The profiling of drug sensitivity indicated that the low-risk subgroup might be more susceptible to cisplatin, whereas high-risk cases showed potential responsiveness to epidermal growth factor receptor (EGFR) and mechanistic target of rapamycin (mTOR) inhibitors.

Conclusions

This research developed a mitophagy-derived signature that predicts survival and is associated with a potential ‘stromal barrier and immune-excluded’ phenotype underlying therapy resistance in high-risk BLCA. The model offers a stratification framework for therapeutic decisions. Standard chemotherapy or immune checkpoint blockade (ICB) may lead to benefits in the low-risk subgroup, whereas high-risk individuals likely require novel combinatorial strategies targeting microenvironmental remodeling. However, as these results represent predictive associations derived from public datasets, further experimental validation is warranted to confirm the underlying biological mechanisms.

Highlight box.

Key findings

• A robust five-gene mitophagy signature (NTN4, AKR1B1, FBXW7, MYH10, and IGF2BP3) was established to predict the prognosis of bladder cancer (BLCA). High-risk tumors exhibited a ‘stromal barrier and immune-excluded’ microenvironment characterized by enrichment of M2 macrophages and dense stroma. While low-risk patients displayed sensitivity to cisplatin and immunotherapy, high-risk cases showed resistance but potential vulnerability to epidermal growth factor receptor (EGFR)/mechanistic target of rapamycin (mTOR) inhibitors.

What is known and what is new?

• Mitophagy regulates cellular metabolism and stress adaptation, yet its specific influence on the remodeling of the tumor microenvironment in BLCA and clinical therapeutic stratification remains insufficiently defined.

• This study systematically links aberrant mitophagy to an ‘immune-excluded’ phenotype and validates a prognostic model that identifies specific drug repurposing opportunities (e.g., rapamycin, erlotinib) for chemo-resistant high-risk patients.

What is the implication, and what should change now?

• The model offers a practical framework for precision oncology. Clinical decision-making should prioritize standard chemotherapy or immune checkpoint blockade for low-risk individuals, whereas high-risk patients warrant novel combinatorial strategies targeting the stromal barrier or alternative EGFR/mTOR-targeted therapies to overcome resistance.

Introduction

Bladder cancer (BLCA), a malignant urinary neoplasm, is marked by high incidence and mortality. The burden of BLCA is progressively increasing worldwide (1,2). While transurethral resection of bladder tumor (TURBT) plus intravesical therapy is effective for non-muscle-invasive BLCA (NMIBC), about 15–20% of cases eventually progress to muscle-invasive BLCA (MIBC)—a stage with an extremely unfavorable prognosis (3,4). For MIBC, cisplatin-based neoadjuvant chemotherapy followed by radical cystectomy constitutes the current standard treatment. Nonetheless, most individuals with BLCA develop primary or secondary resistance to cisplatin-based chemotherapy owing to high heterogeneity of the tumor, leading to treatment failure and shortened survival (5,6). Recently, immune checkpoint blockade (ICB) therapy has provided novel therapeutic prospects for individuals with advanced BLCA. Nevertheless, the overall response rate remains limited and is observed only in some individuals. The majority fail to benefit due to an ‘immune-cold’ or ‘immune-excluded’ tumor status (7). Consequently, a critical unresolved challenge is to clarify the molecular mechanisms underlying BLCA progression and drug resistance, and to identify novel biomarkers that accurately predict prognosis and guide chemotherapy and immunotherapy.

Metabolic reprogramming is an important feature of cancer. As the central hub for cellular energy production and stress responses, mitochondrial homeostasis is intimately linked to tumorigenesis and progression (8). Mitophagy, a selective quality-control mechanism for clearing damaged mitochondria, plays a dual role in tumor cells by promoting adaptation to metabolic stress, evading apoptosis, and maintaining stemness (9,10). More importantly, emerging evidence indicates a complex interplay between mitophagy and the remodeling of tumor microenvironment (TME). The released mitochondrial DNA (mtDNA) activates innate immune pathways—most notably the cGAS-STING—which in turn regulate downstream immune functions such as antigen presentation by dendritic cells and macrophage polarization. These coordinated changes ultimately determine whether the tumor exhibits a ‘cold’ or a ‘hot’ immune phenotype (11-13). In BLCA, the frequently hypoxic and metabolically stressed urothelial microenvironment renders mitophagy crucial for cancer cell adaptation, survival, and subsequent chemoresistance. Furthermore, BLCA is an inherently immunogenic malignancy, and aberrant mitophagy can drive immunosuppressive TME remodeling—for instance, by promoting the polarization of M2 macrophages. This provides a compelling biological rationale for immunotherapy failure (14). Translating these biological insights into actionable clinical tools requires a robust method to stratify patients. However, systematic investigations into the expression profiles of mitophagy-related genes (MRGs) in BLCA and their impact on patient prognosis and treatment response through the remodeling of the TME remain scarce.

To address this, this study systematically examined the clinical implications of MRGs in BLCA through integrated multi-omics analysis. Leveraging data derived from the Gene Expression Omnibus (GEO) and The Cancer Genome Atlas (TCGA), a robust mitophagy-associated predictive model was constructed via rigorous bioinformatics screening and validated in two independent external cohorts. More innovatively, this study dissected a ‘stromal barrier and immune exclusion’ microenvironment feature in depth in high-risk individuals by integrating three algorithms: Cell-Type Identification by Estimating Relative Subsets of RNA Transcripts (CIBERSORT), Tumor Immune Dysfunction and Exclusion (TIDE), and Estimation of STromal and Immune cells in MAlignant Tumor tissues using Expression data (ESTIMATE). Additionally, targeted strategies for chemotherapy sensitization and combination immunotherapy are proposed. This work provides a novel perspective for the molecular subtyping of BLCA and offers a theoretical framework for achieving precision medicine based on mitophagic features. We present this article in accordance with the TRIPOD reporting checklist (available at https://tau.amegroups.com/article/view/10.21037/tau-2026-1-0147/rc).

Methods

Data collection

This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The TCGA provided the RNA-sequencing dataset and corresponding clinical data for BLCA (https://portal.gdc.cancer.gov/), encompassing 404 BLCA samples and 18 normal bladder tissue samples. The RNA-sequencing data from TCGA were converted into transcripts per million (TPM) format and log₂-normalized prior to downstream analyses. Patients from TCGA were designated as the training set. The GEO, offering RNA-sequencing data and clinical profiles for BLCA patients, furnished two independent validation cohorts: GSE32894 (n=224) and GSE31684 (n=93) (https://www.ncbi.nlm.nih.gov/geo/). For GEO datasets, the log₂-transformed normalized expression matrices were directly downloaded. Gene identifiers were mapped to corresponding gene symbols using the human gene annotation file. The list of MRGs was compiled from a previous study (15).

Differential expression and functional enrichment analyses

The expression matrix of MRGs was derived from the TCGA dataset. Differentially expressed MRGs (DEMRGs) for BLCA were detected using the R package ‘limma’, with thresholds set as P value <0.05 and |log₂fold change (FC)| >0.1. Heatmaps and volcano plots were generated utilizing R packages ‘pheatmap’ and ‘ggplot2’. To clarify the underlying biological functions and pathways, Kyoto Encyclopedia of Genes and Genomes (KEGG) and Gene Ontology (GO) analyses were performed with R packages ‘org.Hs.eg.db’, ‘topGO’, and ‘clusterProfiler’. Enriched categories showing both P value and q value below 0.05 were considered significant.

Detection of prognosis-associated MRGs and prognostic model construction/validation

To detect prognosis-related MRGs among DEMRGs, univariable Cox regression was performed by integrating TCGA survival data with DEMRGs, utilizing publicly available genes from GEO datasets. Genes exhibiting a P value below 0.05 were preliminarily selected. The least absolute shrinkage and selection operator (LASSO) analysis was then performed using the ‘glmnet’ R package to minimize overfitting. Specifically, a 10-fold cross-validation was performed to determine the optimal penalty parameter (λ) that minimized the partial likelihood deviance, ensuring a robust selection of features and preventing over-parameterization. Finally, multivariable Cox regression (R package ’survival’) was conducted to create a prognostic model, applying a significance threshold of P<0.05.

The model was established using the five core genes and their respective coefficients derived from the multivariable Cox regression. Patient-specific risk scores were derived as a coefficient-weighted sum of gene expression levels using the ‘predict’ function in R. Patients within each cohort (TCGA training cohort, GSE32894, and GSE31684 validation cohorts) were stratified into low- and high-risk subgroups per median risk scores. Kaplan-Meier (K-M) survival analysis was implemented to compare inter-group differences in overall survival (OS). The predictive accuracy was evaluated by plotting time-dependent receiver operating characteristic (ROC) curves and computing areas under the curves (AUCs) at 1-, 3-, and 5-year time points using the R package ‘timeROC’.

Multivariable and univariable Cox regressions

Univariable Cox regression was first used to evaluate the association of risk scores and clinical features [sex, age, race, stage, tumor-node-metastasis (TNM) classifications] with prognosis. Multivariable Cox regression was then performed to identify independent prognostic factors for survival in BLCA individuals. Adjusted P values below 0.05 indicated statistical significance.

Nomogram establishment and evaluation for BLCA individuals

A prognostic nomogram was developed using the R package ‘rms’ to predict 1-, 2-, and 3-year survival probabilities for BLCA individuals. Predictive accuracy was assessed using calibration plots, and clinical net benefit was evaluated by decision curve analysis (DCA).

Analysis of TME, immune checkpoints, and human leukocyte antigens (HLAs)

To dissect the differences in the TME across both risk subgroups, multiple bioinformatics algorithms were adopted. Infiltration abundances were estimated for 22 immune cell types via the CIBERSORT tool (with the LM22 signature matrix) and visualized via stacked bar plots. Box plots were used to compare inter-group infiltration levels of immune cells. Correlation heatmaps among the 22 immune cell subtypes were generated using the ‘corrplot’ package. Spearman correlation analysis was employed to examine correlations of genes with immune-infiltrating cells, and the ‘ggplot2’ package was applied for visualizing the results via heatmaps. The ESTIMATE approach was utilized for computing the ImmuneScore, StromalScore, and ESTIMATEScore, from which TumorPurity was inferred. This analysis quantifies the infiltration of non-tumor cellular components (stromal and immune cells) in the TME. To evaluate the potential clinical effectiveness of immune-based therapy, the TIDE approach was employed (http://tide.dfci.harvard.edu/login/). This algorithm integrates features reflecting T-cell exclusion and T-cell dysfunction. A standardized gene expression matrix was uploaded to the TIDE platform to compute the T cell dysfunction score, microsatellite instability (MSI) score, T cell exclusion score, and TIDE score for each patient. Higher TIDE scores typically suggest a higher probability of immune escape and a poorer predicted response to ICB therapy. Inter-group comparisons were also conducted for the expression of immune checkpoints and HLA molecules. A P value below 0.05 indicated statistical significance.

Chemosensitivity evaluation

The ‘pRRophetic’ package was applied for estimating the half-maximal inhibitory concentration (IC₅₀) of drugs for both risk groups. Inter-group comparisons were conducted for drug sensitivity employing the Wilcoxon rank-sum test and visualized via box plots (P<0.05).

Gene mutation analysis

BLCA mutation data were acquired from TCGA with the R package ‘TCGAbiolinks’. Tumor mutation burden was computed using the R package ‘maftools’. Summary and waterfall plots were generated for the two risk groups separately, depicting the top 20 mutated genes in both groups.

Statistical analysis

Data were analyzed and visualized in R (v4.4.3). Univariable and multivariable Cox regressions were implemented to determine independent prognostic factors. K-M analysis was implemented to assess the predictive capability of the model for OS. Time-dependent ROC curves were created to assess the predictive accuracy. Inter-group comparisons were conducted utilizing the Student’s t-test for normally distributed data and the Wilcoxon test for skewed data. Unless otherwise specified, statistical significance was set at P<0.05 (****, P<0.0001; ***, P<0.001; **, P<0.01; *, P<0.05).

Results

Detection of DEMRGs in the TCGA cohort

The analysis of TCGA-derived BLCA and normal tissue specimens detected DEMRGs (Figure 1). In total, 53 DEMRGs (34 upregulated and 19 downregulated) were identified (Figure 2A,2B).

Figure 1.

Study flow diagram. BLCA, bladder cancer; DEMRG, differentially expressed mitophagy-related gene; GEO, Gene Expression Omnibus; GO, Gene Ontology; IHC, immunohistochemistry; KEGG, Kyoto Encyclopedia of Genes and Genomes; LASSO, least absolute shrinkage and selection operator; ROC, receiver operating characteristic; TCGA, The Cancer Genome Atlas; TMB, tumor mutational burden.

Figure 2.

Differential analysis results. (A) Heatmap illustrating the distinct expression patterns of the 53 DEMRGs between BLCA tissues (tumor, red) and normal tissues (normal, blue). (B) Volcano plot of 53 DEMRGs in the TCGA-BLCA cohort. Red dots indicate significantly upregulated DEMRGs, blue dots represent significantly downregulated DEMRGs, and grey dots represent genes with no significant change. BLCA, bladder cancer; DEMRG, differentially expressed mitophagy-related gene; TCGA, The Cancer Genome Atlas.

GO and KEGG analyses of DEMRGs

An analysis indicated significant enrichment of DEMRGs in biological processes like regulation of mitophagy, regulation of autophagy of mitochondrion, Fc-epsilon receptor signaling pathway, mitophagy, and negative regulation of macroautophagy (Figure 3A). KEGG analysis showed notable enrichment of DEMRGs in pathways including pathogenic Escherichia coli infection, adipocytokine signaling pathway, alcoholic liver disease, and insulin resistance (Figure 3B).

Figure 3.

Functional enrichment analysis of DEMRGs. (A) Bar graph displaying the findings of GO analysis. (B) Bar graph displaying the results of KEGG analysis. In both panels, the color gradient, ranging from blue to red, indicates the adjusted P value, with red denoting lower P values and higher statistical significance. BP, biological process; CC, cellular component; DEMRG, differentially expressed mitophagy-related gene; GO, Gene Ontology; KEGG, Kyoto Encyclopedia of Genes and Genomes.

Selection of prognosis-related MRGs

Thirty-seven common genes were obtained by intersecting the 53 DEMRGs from TCGA with the GEO datasets. Univariable Cox regression analysis identified 11 prognosis-related genes under the criterion of P<0.05 (Table S1). Then, LASSO regression was applied to reduce multicollinearity among these 11 candidate genes, yielding a refined set of nine genes (Figure 4A,4B). Multivariable Cox regression finally established five MRGs as independent prognostic factors: NTN4, AKR1B1, FBXW7, MYH10, and IGF2BP3. The coefficients for these five genes are listed in Table 1.

Figure 4.

LASSO regression. (A) Coefficient change path plot. The bottom X-axis represents log(λ), and the top X-axis indicates the number of non-zero coefficients in the model. (B) The 10-fold cross-validation curve for selecting the optimal tuning parameter (λ) in the LASSO model. The X-axis represents log(λ), and the Y-axis represents the partial likelihood deviance. The red dots denote the average deviance values, while the grey error bars represent the standard errors. LASSO, least absolute shrinkage and selection operator.

Table 1. Prognosis-associated MRGs and their coefficients from multivariable Cox analysis.

Gene	Coefficient
NTN4	−0.11738445
AKR1B1	0.09681419
FBXW7	−0.60547438
MYH10	0.21876191
IGF2BP3	0.14986097

MRG, mitophagy-related gene.

Establishment of an MRG-associated risk score model

Specimen-specific risk scores were computed according to the five prognostic MRGs (Figure 5). The 422 BLCA individuals in the TCGA (training) cohort were classified as low- (n=211) and high-risk (n=211) categories per median risk scores. K-M analysis indicated that low-risk individuals exhibited notably longer OS (P<0.001, Figure 5A). The distribution of risk scores and survival status is shown in Figure 5D, indicating a trend towards shorter survival and increased mortality with higher risk scores. Time-dependent ROC analysis for the TCGA cohort yielded AUCs of 0.653, 0.654, and 0.645 at 1-, 3-, and 5-year time points (Figure 5G), indicating moderate predictive performance.

Figure 5.

Prognostic evaluation of the five-gene signature in the validation and training cohorts. (A-C) K-M curves for patients’ OS across both risk groups in (A) TCGA, (B) GSE32894, and (C) GSE31684 cohorts. (D-F) Distribution of risk scores and survival status scatter plots in (D) TCGA, (E) GSE32894, and (F) GSE31684 cohorts. (G-I) ROC curves in (G) TCGA, (H) GSE32894, and (I) GSE31684 cohorts. K-M, Kaplan-Meier; OS, overall survival; ROC, receiver operating characteristic; TCGA, The Cancer Genome Atlas.

Validation of the MRG-related risk score model in GEO cohorts

The predictive capacity and accuracy of MRG-associated risk model were evaluated using the GSE32894 dataset. Patients in GSE32894 were also stratified into low- (n=112) and high-risk (n=112) subgroups. Similarly, K-M analysis confirmed significantly worse survival in high-risk individuals (P=0.003, Figure 5B). The distribution of risk scores and survival status for this subgroup is illustrated in Figure 5E. ROC curves at 1, 3, and 5 years produced AUCs of 0.625, 0.770, and 0.785 (Figure 5H), validating its acceptable stratification ability in this cohort.

To further confirm the robustness of this five-gene signature, it was validated in a second independent external cohort, GSE31684 (n=93). In agreement with the results from the first validation and training sets, K-M analysis showed that high-risk individuals exhibited substantially worse prognosis (P=0.03, Figure 5C). Figure 5F illustrates the distribution of risk scores and survival status. Time-dependent ROC curves revealed moderate predictive efficacy, with AUCs being 0.567, 0.639, and 0.651 for survival at 1, 3, and 5 years, respectively (Figure 5I). The consistent results across three cohorts strongly supported the reliability and broad applicability of this MRG-related risk score model.

Independent prognostic factors for OS

Clinical data and risk scores were integrated for subjects in the training cohort. Univariable and multivariable Cox regressions were performed (Figure 6A,6B). Univariable Cox regression revealed significant differences in age [P<0.001; hazard ratio (HR) =1.033; 95% confidence interval (CI): 1.02–1.05], pN stage (P=0.002; HR =1.182; 95% CI: 1.06–1.31), pT stage (P<0.001; HR =1.704; 95% CI: 1.37–2.12), clinical stage (P<0.001; HR =1.74; 95% CI: 1.43–2.12), and risk score (P<0.001; HR =2.197; 95% CI: 1.67–2.88). Subsequent multivariable Cox regression established age (P=0.001; HR =1.027; 95% CI: 1.010–1.044), pT stage (P=0.004; HR =1.535; 95% CI: 1.150–2.048), pN stage (P=0.01; HR =1.192; 95% CI: 1.042–1.363), and risk score (P<0.001; HR =1.916; 95% CI: 1.421–2.585) as independent prognostic factors in BLCA.

Figure 6.

Independent prognostic predictors for the risk evaluation model in the TCGA cohort. (A) Forest plot of univariable analysis. (B) Forest plot of multivariable analysis. (C) Nomogram for predicting 1-, 2-, and 3-year OS of BLCA individuals. (D) Calibration curves for nomogram. (E) Decision curve for nomogram. BLCA, bladder cancer; CI, confidence interval; HR, hazard ratio; M, metastasis; N, node; OS, overall survival; T, tumor; TCGA, The Cancer Genome Atlas.

Construction of a nomogram

A prognostic nomogram, incorporating age, pathologic tumor (pT) stage, pathologic node (pN) stage, and risk scores, was constructed to project the probabilities of OS at 1, 2, and 3 years for BLCA individuals (Figure 6C). Calibration curves revealed high concordance between the nomogram-predicted and actual survival rates, confirming its predictive accuracy (Figure 6D). DCA revealed a net clinical benefit, supporting its clinical reliability and stability (Figure 6E). This nomogram demonstrated good prognostic predictive performance and may aid clinicians in making more accurate and efficient therapeutic decisions.

Association of mitophagy-related signature and immune microenvironment

Immunotherapy has advanced rapidly in recent years, inducing durable responses in cancer patients that are often superior to those achieved with conventional chemotherapy (16). To investigate the link between immune microenvironment and mitophagy-associated signature, the CIBERSORT approach was adopted to deconvolute the immune infiltration profile in the TCGA-BLCA cohort. The stacked bar plot was created to visualize the abundance of 22 immune cell types, revealing that macrophages and T cell subsets (CD4⁺ and CD8⁺) were the predominant infiltrating populations in the microenvironment of BLCA (Figure 7A). Correlation analysis among immune cell types uncovered patterns of co-existence or mutual exclusion between different immune cell subsets within the TME (Figure 7B). To investigate the possible influence of MRGs on reshaping the immune landscape of BLCA, the correlations of the five prognostic MRGs with 22 immune cell types were analyzed (Figure 7C). Notably, high-risk genes (e.g., MYH10) exhibited a notable positive link with M2 macrophages—a prototypical immunosuppressive cell type—while protective genes (e.g., FBXW7) exhibited a notable inverse correlation with M2 macrophages. This finding suggested that aberrant expression of MRGs may remodel the TME by recruiting or inducing specific immune cell subsets. Subsequent differential analysis revealed distinct immune landscapes across both risk groups (Figure 7D). Significant enrichment of immunosuppressive M0 and M2 macrophages was noted in high-risk cases, consistent with worse prognoses. Conversely, low-risk cases displayed significantly higher infiltration of various T cell subsets (CD8⁺, CD4⁺ memory resting, follicular helper, gamma delta), regulatory T cells, NK cells activated, and dendritic cells resting.

Figure 7.

Immune infiltration analysis. (A) Stacked bar plot showing the relative abundances of 22 immune cell subtypes. (B) Correlation among the 22 immune cell subtypes. (C) Correlation of prognosis-related MRGs with the 22 immune cell subtypes. (D) Analysis comparing the proportions of immune-infiltrating cells across both risk subgroups. (E) Expression of immune checkpoints across both risk subgroups. (F) Expression of HLA genes across both risk subgroups. *, P<0.05; **, P<0.01; ***, P<0.001; ****, P<0.0001. HLA, human leukocyte antigen; MRG, mitophagy-related gene; NK, natural killer.

Transcriptional expression of 47 immune checkpoints was also compared across the two groups, revealing significant differential expression for 25 of them (Figure 7E). Most of those differentially expressed showed upregulation in the high-risk cohort. Analysis of HLA gene expression further showed that the high-risk cohort exhibited notably decreased expression of several HLA class I (HLA-V, HLA-A, HLA-J, HLA-E, HLA-F-AS1, HLA-T, HLA-K, HLA-U) and class II (HLA-DMA, HLA-DRB1) genes, suggesting a higher likelihood of immune evasion (Figure 7F).

To further dissect microenvironmental differences, the ESTIMATE method was first applied for assessing the composition of the TME. Results revealed that high-risk individuals had markedly elevated StromalScores (P<0.001, Figure 8A) and ESTIMATEScores (P<0.001, Figure 8B), accompanied by significantly lower TumorPurity (P<0.001, Figure 8C). Notably, the two groups did not exhibit notable differences in ImmuneScores (P=0.11, Figure 8D).

Figure 8.

Comprehensive evaluation of TME purity and immunotherapy response. (A-D) Evaluation of TME landscape using the ESTIMATE approach. Comparisons across both risk groups for (A) stromal score, (B) ESTIMATE score, (C) tumor purity, and (D) immune score. (E-H) Prediction of immunotherapy response using the TIDE algorithm. Comparisons across the two risk groups for (E) TIDE, (F) T cell dysfunction, (G) T cell exclusion, and (H) MSI scores. Data are presented as box plots and violin plots. Statistical significance was determined by the Wilcoxon rank-sum test (***, P<0.001; ****, P<0.0001; ns, not significant). ESTIMATE, Estimation of STromal and Immune cells in MAlignant Tumor tissues using Expression data; MSI, microsatellite instability; TIDE, Tumor Immune Dysfunction and Exclusion; TME, tumor microenvironment.

Subsequent analysis employing the TIDE approach indicated that high-risk individuals had notably higher TIDE scores (P<0.001, Figure 8E), predicting a poorer response to ICB therapy. Further dissection revealed that this immune resistance was not driven by the dysfunction of T cells (no difference in dysfunction score, Figure 8F), but was strongly linked to a significantly elevated exclusion score of T cells (P<0.001, Figure 8G). This finding aligned with the ‘high-stroma’ characteristic observed in the ESTIMATE analysis, suggesting that dense stromal components may collaborate with M2 macrophages to exclude effector T cells from the tumor core. Low-risk individuals exhibited a substantially higher MSI score (P<0.001, Figure 8H), highlighting their potential as a favorable candidate population for immunotherapy.

Chemosensitivity evaluation

To improve the benefit of chemotherapy for BLCA individuals, the prognostic utility of the MRG-related predictive model for the effectiveness of commonly used agents was evaluated. The IC₅₀ values for 251 drugs from the ‘cgp2016’ dataset were predicted using the ‘pRRophetic’ algorithm and compared across the two risk groups. The analysis identified 97 drugs with significantly different IC₅₀ values between groups (Figure S1). The results for two conventional BLCA chemotherapy drugs (cisplatin and vinblastine) and four additional agents with notably low P values (axitinib, erlotinib, lapatinib, and rapamycin) are visually presented in Figure 9. The high-risk cohort exhibited significantly lower sensitivity to cisplatin, vinblastine, and axitinib, indicating that low-risk patients had comparatively greater chemosensitivity to these drugs. Conversely, the high-risk cohort exhibited greater sensitivity to erlotinib, lapatinib, and rapamycin. These results provide potential insights into the selection of drug regimens for BLCA individuals; however, as these findings are derived from in silico estimations, they represent preliminary associations that require further validation.

Figure 9.

Inter-group comparison of drug sensitivity. (A) Cisplatin, (B) erlotinib, (C) vinblastine, (D) lapatinib, (E) axitinib, (F) rapamycin. IC₅₀, half-maximal inhibitory concentration.

Analysis of somatic mutations

The profile of somatic mutation in BLCA was further analyzed. A waterfall plot was created for visualizing the mutation landscape across the TCGA-BLCA cohort (Figure 10A). The five most frequently mutated genes were TP53 (49%), TTN (43%), KMT2D (26%), ARID1A (25%), KDM6A (25%), and MUC16 (25%). The most frequent variant classification was missense mutation, predominantly comprising single nucleotide polymorphisms (SNPs); among SNP classifications, C>T transitions occurred most frequently (Figure 10B). Waterfall plots were generated to visualize the mutation landscapes for both risk cohorts separately (Figure 10C,10D). In the high-risk cohort, the top mutated genes were TP53 (61%), TTN (42%), ARID1A (28%), KMT2D (26%), RB1 (24%), and MUC16 (24%). In the low-risk cohort, the top mutated genes were TTN (43%), TP53 (36%), KDM6A (28%), MUC16 (27%), KMT2D (26%), and PIK3CA (26%) (Figure 10C,10D).

Figure 10.

Somatic mutation analysis. (A) Waterfall plot depicting the mutational landscape of the top 20 most frequently mutated genes in the overall BLCA cohort. Each column represents an individual patient, and each row represents a specific gene. The top barplot displays the TMB for each sample, while the right barplot indicates the mutation frequency of each gene. Distinct colors represent different types of genetic alterations. (B) Statistical summary of the somatic mutation data, illustrating the distribution of variant classifications, variant types, SNV classes, variants per sample, variant classification summary, and the top 10 mutated genes. (C) Waterfall plot of mutated genes in the high-risk cohort. (D) Waterfall plot of mutated genes in the low-risk cohort. BLCA, bladder cancer; SNV, single nucleotide variant; TMB, tumor mutational burden.

Discussion

The pronounced heterogeneity of BLCA constitutes a principal factor underlying the substantial variability in clinical outcomes and inconsistent response rates to therapeutic interventions. While neoadjuvant chemotherapy regimens centered on cisplatin and ICB have emerged as standard treatments for MIBC, a considerable proportion of patients still experience primary resistance or disease recurrence (17-19). Mitophagy is a crucial process governing cellular metabolic reprogramming and stress adaptation. Although its role in remodeling the TME is increasingly recognized, the potential utility of mitophagy for prognostic stratification and personalized treatment guidance remains inadequately explored (20,21). This investigation addressed this gap by integrating multi-omics data to establish and verify a stable predictive signature derived from MRGs. This model delineated a ‘high-risk-immune-excluded’ microenvironmental profile. Based on this characterization, stratified therapeutic strategies were proposed for chemotherapy and immunotherapy. These findings furnish a theoretical framework for precision oncology in BLCA.

Initial analysis identified 53 DEMRGs between BLCA and normal tissues from TCGA data. Subsequent univariable Cox, LASSO, and multivariable Cox analyses refined this gene set to a five-MRG signature. A notable strength of this work is its rigorous multi-cohort validation, which distinguishes it from studies reliant on a single validation dataset. The mitophagy-related signature was externally validated using two independent cohorts: the larger-scale GSE32894 dataset, followed by the GSE31684 cohort. Across all three datasets—encompassing varying sample sizes and potential heterogeneity—K-M analyses consistently demonstrated significantly superior survival prognosis for low-risk individuals. The moderate predictive performance was consistently demonstrated by time-dependent ROC curves. This cross-platform and cross-population validation underscores the robustness of our model, effectively mitigating concerns of overfitting or batch effects associated with single-dataset studies and enhancing its translational potential. Both univariable and multivariable Cox regressions confirmed the independent prognostic value of MRG-based risk scores. Subsequently, a nomogram was constructed for projecting 1-, 2-, and 3-year probabilities of OS for BLCA individuals. Calibration curves confirmed its accuracy, while DCA demonstrated its superior clinical net benefit. Collectively, these results affirm the moderate but statistically significant prognostic capability of the signature. Despite its moderate AUC, the signature-based nomogram serves as an effective supplement to existing staging, enhancing the robustness of clinical decision-making.

The five MRGs in the final model—NTN4, AKR1B1, FBXW7, MYH10, and IGF2BP3—are implicated in a molecular network that may drive the progression of tumors through metabolic reprogramming, therapy resistance, and modulation of the TME. Regarding metabolic adaptation and drug resistance, the risk gene AKR1B1 encodes the key rate-limiting enzyme of the polyol pathway. Its overexpression might promote malignant proliferation via metabolic shifts—such as enhanced fructose metabolism—and potentially confers resistance to cisplatin by increasing cellular tolerance to oxidative stress and then suppressing mitophagy-mediated cell death (22,23). This hypothesis aligns with our drug sensitivity analysis, which showed cisplatin resistance in the high-risk cohort. Conversely, the protective gene FBXW7 encodes a core component of an E3 ubiquitin ligase complex that targets oncoproteins such as c-Myc for degradation. Loss of FBXW7 function—a characteristic of the high-risk cohort—is speculated to compromise ubiquitin-proteasome system activity. This process is closely linked to mitochondrial quality control via mitophagy and may foster uncontrolled proliferation and immune evasion (24,25). In the realm of epitranscriptomic regulation and the remodeling of the TME, IGF2BP3, an N⁶-methyladenosine (m6A) reader protein, stabilizes oncogenic messenger RNAs (mRNAs) (e.g., CDK6) in an m6A-dependent manner. This effect has been suggested to contribute to cisplatin resistance in BLCA (26) and may modulate inflammatory infiltration within the TME, potentially by stabilizing factors like HMGB1 to promote invasion and metastasis (27). Concerning the formation of a physical barrier and immune exclusion, MYH10 regulates cytoskeletal rearrangement and cell motility. Emerging evidence links it to disulfidptosis, suggesting a putative role in stromal remodeling of MIBC (28,29). NTN4, an extracellular matrix protein, primarily modulates angiogenesis and tissue homeostasis (30,31). In high-risk individuals, aberrant expression of MYH10 and NTN4 may promote a dense stromal matrix and aberrant vasculature, consistent with the high stromal scores from our ESTIMATE analysis. This physical barrier likely excludes effector T cells from the tumor core, culminating in the ‘immune-excluded’ phenotype predicted by the TIDE algorithm. However, it is essential to emphasize that these mechanistic links within the specific context of BLCA mitophagy are exploratory at this stage and require rigorous experimental validation.

The effectiveness of immune-based therapy is largely determined by the immune phenotype of TME. Integrated analysis using CIBERSORT, ESTIMATE, and TIDE algorithms suggested a potential microenvironmental basis for unfavorable prognosis in high-risk individuals, highlighting a critical ‘immunotherapy paradox’. Although these patients exhibited elevated expression of most immune checkpoint molecules—theoretically making them ideal candidates for ICB—their high TIDE scores predicted poor response to ICB. Multi-faceted analysis suggests that this phenomenon might be attributed to immune exclusion. One factor is a markedly elevated stromal score, reflecting a dense extracellular matrix barrier potentially driven by high expression of MYH10 and NTN4 (29,30). The other is a marked enrichment of M2-polarized macrophages, which are known to promote tumor progression and invasion (32). The risk gene IGF2BP3 may regulate immune cell infiltration by stabilizing mRNAs of pro-inflammatory cytokines such as HMGB1 (27), while aberrant mitophagy metabolites may also drive the polarization of M2 macrophages, fostering an immunosuppressive milieu (33). Furthermore, significant downregulation of HLA class I/II molecules in the high-risk group is hypothesized to impair tumor antigen recognition by T cells, aligning with the T cell exclusion signature captured by TIDE (34). These insights inform potential therapeutic strategies. For high-risk patients, ICB monotherapy may prove insufficient. Combinatorial approaches targeting the depletion of M2 macrophages or disrupting the stromal barrier to convert ‘cold’ tumors into ‘hot’ ones warrant investigation. Conversely, the low-risk group, characterized by an ‘immune-inflamed’ phenotype with high MSI, low stromal content, and enrichment of CD8⁺ T cells, may represent a more optimal population for ICB monotherapy. Notably, these characterizations of the TME are derived from computational inference algorithms and warrant further experimental validation.

Chemotherapy remains a cornerstone of BLCA treatment. Cisplatin and vinblastine are integral components of first-line regimens [e.g., GC (gemcitabine + cisplatin), MVAC (methotrexate, vinblastine, doxorubicin, cisplatin)] for muscle-invasive disease, proven to improve patient survival (35). The drug sensitivity analysis indicated greater sensitivity in low-risk individuals to these agents, implying that standard chemotherapy might remain a favorable choice for maximizing survival benefit in these patients. The high-risk group, however, was predicted to exhibit pronounced resistance to first-line chemotherapeutics, potentially explaining the disease progression in some patients despite adequate treatment. Mechanistically, we hypothesize that this resistance could be mediated by AKR1B1-driven suppression of chemotherapy-induced apoptosis via enhanced antioxidant capacity (23) and IGF2BP3-mediated, m6A-dependent stabilization of CDK6 mRNA (26). For these refractory patients, continued use of cisplatin might incur toxicity without proportional benefit. Our model identified potential alternative agents for the high-risk group, such as rapamycin [a mechanistic target of rapamycin (mTOR) inhibitor] and erlotinib [an epidermal growth factor receptor (EGFR) inhibitor]. A hallmark of this group is low expression of FBXW7. As an E3 ligase targeting mTOR for degradation, the loss of FBXW7 leads to aberrant accumulation of mTOR proteins and hyperactivation of pathways, which may theoretically confer specific sensitivity to mTOR inhibitors like rapamycin (24,36,37). Thus, for high-risk patients, exploring mTOR or EGFR-targeted therapies within clinical trials or as second-line options might present a plausible strategy to overcome chemotherapy resistance. These findings suggest potential avenues for personalizing chemotherapy and targeted therapies. Given the in silico nature of these estimations, they are best understood as hypothesis-generating insights that require prospective validation before being translated into actionable clinical recommendations.

In recent years, various prognostic gene signatures based on diverse biological processes, such as efferocytosis and glycolysis, have been developed for BLCA (38,39). Compared with these existing models, the current mitophagy-associated signature demonstrates comparable, moderate predictive accuracy. Rather than claiming absolute statistical superiority, the primary value of this signature lies in its distinct biological perspective. Specifically, it uncovers the potentially critical linkage between aberrant mitophagy, the enrichment of M2 macrophages, and the formation of a dense stromal barrier. This provides a novel mechanistic hypothesis for immunotherapy resistance in BLCA that differs from the focus of previous metabolism- or cell-death-related signatures.

While this study establishes a compelling logical framework through multi-dimensional bioinformatics analyses, certain limitations persist. First, the current conclusions rely on retrospective public datasets, which inevitably introduces inherent biases, such as patient selection bias, heterogeneous treatment protocols, and missing clinical confounders across different databases. Despite robust multi-cohort validation, support from large-scale prospective clinical trial data is still lacking. Second, the immune infiltration and stromal content profiles discussed herein are computational estimations. Specifically, bulk RNA-based deconvolution algorithms like CIBERSORT and TIDE cannot fully capture the spatial transcriptomic architecture of the TME or the dynamic phenotypic plasticity of individual immune cells. The precise molecular mechanisms by which these MRGs regulate the polarization of M2 macrophages and stromal remodeling remain correlative and lack direct biological verification. Future work will involve validating these associations in clinical tissue samples via immunohistochemistry and elucidating the underlying mechanisms through in vitro co-culture experiments, aiming to facilitate the translation of this signature into clinical practice.

Conclusions

Through in-depth characterization of the TME, our research not only constructs a robust mitophagy-associated prognostic signature but also elucidates a ’stromal barrier and immune exclusion’ mechanism driving therapeutic resistance in the high-risk cohort. It provides a clear clinical stratification strategy: standard chemotherapy or immunotherapy monotherapy is recommended for low-risk BLCA individuals, whereas those at high risk may benefit from combinatorial strategies such as stromal disruption or tumor-associated macrophage targeting combined with immunotherapy, or from EGFR/mTOR-targeted therapies.

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.