Cancer-associated fibroblasts are among the most abundant non-tumor cells in the bladder cancer microenvironment, yet their prognostic role has been largely overlooked. These cells remodel the extracellular matrix, recruit immunosuppressive cells, and secrete growth factors that promote tumor invasion and therapy resistance.
Single-cell RNA sequencing has made it possible to dissect the tumor microenvironment at unprecedented resolution, identifying distinct fibroblast subpopulations with different functional states. Prior bulk RNA studies could not distinguish these subpopulations because their signals were averaged together with thousands of other cell types.
Bladder cancer presents a particular clinical challenge because patient outcomes vary widely even within the same pathologic stage, and no reliable multi-gene signature exists to predict which patients will respond to immune checkpoint inhibitors. A fibroblast-based prognostic model could fill this gap by capturing microenvironmental features that tumor-intrinsic signatures miss.
This study combined single-cell and bulk transcriptomic data across seven independent cohorts with machine learning-based gene selection to identify a minimal fibroblast gene signature capable of stratifying prognosis and predicting immunotherapy response.
The analysis began with a publicly available single-cell RNA sequencing dataset of 10 bladder cancer samples (GSE135337), from which fibroblast clusters were identified using canonical marker genes including ACTA2, TAGLN, FAP, and COL1A1. Genes specifically upregulated in fibroblast clusters compared to all other cell types were extracted as candidate fibroblast-related genes.
This initial fibroblast gene list was then cross-referenced with bulk TCGA-BLCA gene expression data to retain only genes whose expression correlated with overall survival. This filtering step reduced the candidate pool to 54 fibroblast-related genes with confirmed prognostic relevance at the bulk transcriptome level.
Ten machine learning algorithms were then applied in combination to identify the most predictive gene subset. The algorithms included LASSO regression, ridge regression, elastic net, random forest, gradient boosting, survival support vector machines, CoxBoost, stepwise Cox regression, partial least squares, and principal component analysis-based regression. Each algorithm nominated gene subsets, and the three genes appearing most consistently across methods were selected.
The final three-gene fibroblast-related gene index (FRGI) was constructed as a linear combination: FRGI = 0.02982 multiplied by COL6A1 expression plus 0.07886 multiplied by CRYAB expression plus 0.08066 multiplied by FN1 expression. Patients were stratified into high and low FRGI groups at the median, with high FRGI associated with worse outcomes.
The FRGI was validated in six independent bladder cancer datasets beyond the TCGA-BLCA training set. In each cohort, high FRGI consistently predicted significantly shorter overall survival and recurrence-free survival, demonstrating that the index generalizes across different patient populations and sequencing platforms.
A nomogram was constructed by combining FRGI with clinical variables including tumor stage, grade, and age. The combined nomogram achieved AUC values up to 0.96 for predicting recurrence-free survival at one, three, and five years, outperforming FRGI alone and each clinical variable individually.
Multivariable Cox regression confirmed that FRGI remained an independent prognostic factor after adjusting for stage, grade, and other clinical covariates. The hazard ratios were consistent across cohorts, indicating that the fibroblast signature captures biologically meaningful information not encoded in standard pathologic variables.
Calibration curves for the nomogram showed close agreement between predicted and observed survival probabilities, confirming that the model produces reliable probability estimates rather than purely relative rankings.
Samples were classified into fibroblast-hot and fibroblast-cold subtypes based on the overall expression level of the 54 fibroblast-related genes. Fibroblast-hot tumors showed high FRG expression and markedly different immune cell infiltration patterns compared to fibroblast-cold tumors.
Single-sample gene set enrichment analysis revealed that fibroblast-hot tumors were enriched for regulatory T cells and M2 macrophages, both of which suppress anti-tumor immunity. Pathway analysis showed activation of TNFA/NFKB signaling, KRAS signaling, interferon gamma response, and epithelial-to-mesenchymal transition pathways in the high-FRGI group.
Fibroblast-cold tumors had higher infiltration of CD8-positive cytotoxic T cells and natural killer cells, consistent with a more immunologically active microenvironment. These differences suggest that fibroblast activity directly shapes the composition of the immune infiltrate in bladder cancer.
Tumor mutation burden and microsatellite instability scores did not fully account for the immune differences between FRGI groups, indicating that the fibroblast signature captures stromal regulation of immunity that is independent of tumor mutational landscape.
Combining FRGI with CD8-positive T cell infiltration levels produced a four-subtype classification system with distinct prognostic and therapeutic implications. The four subtypes were: CD8-positive/fibroblast-hot, CD8-positive/fibroblast-cold, CD8-negative/fibroblast-hot, and CD8-negative/fibroblast-cold.
The CD8-negative/fibroblast-hot subtype had the worst prognosis of all four groups, characterized by simultaneous absence of cytotoxic T cell immunity and dominance of pro-tumorigenic fibroblast activity. This combination represents a maximally immunosuppressed microenvironment.
The CD8-negative/fibroblast-cold subtype paradoxically showed the best prognosis among the four groups, suggesting that in the absence of active fibroblast-driven immunosuppression, tumors with low CD8 infiltration may still be controlled through other mechanisms.
FRGI also predicted immune checkpoint inhibitor response. In the TCGA-BLCA cohort, the AUC for predicting non-response to immunotherapy was 0.8642. High-FRGI patients showed lower predicted response rates to PD-L1 blockade, consistent with the immunosuppressive microenvironment driven by active fibroblast signaling.
Drug sensitivity analysis using the GDSC database identified compounds predicted to be more effective in high-FRGI tumors. Drugs targeting pathways enriched in fibroblast-hot tumors, including EMT and KRAS signaling, showed differential sensitivity between FRGI groups, providing potential therapeutic hypotheses for patients with poor-prognosis fibroblast signatures.
The FRGI could serve as a patient stratification tool before initiating immune checkpoint inhibitor therapy. Patients with high FRGI scores would be predicted as poor responders and could be prioritized for alternative or combination strategies targeting the fibroblast-driven immunosuppressive microenvironment.
CRYAB, FN1, and COL6A1 each have known biological roles relevant to cancer progression. CRYAB is a small heat shock protein that promotes cell survival and invasion. FN1 encodes fibronectin, a major extracellular matrix glycoprotein involved in cell adhesion and migration. COL6A1 encodes a collagen subunit associated with stromal stiffness and tumor invasion.
The three-gene signature is simple enough to be implemented in clinical laboratories using standard mRNA quantification assays, potentially enabling practical deployment as a prognostic test without requiring full transcriptome sequencing.