Clear cell renal cell carcinoma (ccRCC) originates from malignant transformation of renal tubular epithelial cells, yet the molecular characteristics that define the malignant epithelial cell (MEC) population within the complex tumor microenvironment (TME) have not been comprehensively characterized. Understanding MECs at single-cell resolution is essential for discovering prognostic biomarkers grounded in tumor cell biology.
The TME of ccRCC is highly immunosuppressive, comprising tumor-associated macrophages (TAMs), regulatory T cells (Tregs), cancer-associated fibroblasts, and other stromal populations that interact dynamically with malignant cells. These interactions shape tumor progression, metastatic competence, and response to immunotherapy.
This study leveraged single-cell RNA sequencing (scRNA-seq) to identify malignant epithelial cell-related genes (MECRGs) and then applied a comprehensive machine learning framework to build a robust prognostic signature (MECRGS) validated across multiple independent patient cohorts.
scRNA-seq data from 8 RCC tumors and 6 benign kidney samples (GSE159115 dataset) were analyzed to profile the cellular composition and transcriptional states of renal tissue at single-cell resolution. This comparison of malignant versus normal epithelial cells enabled identification of 219 MECRGs specifically upregulated or defining of the malignant epithelial cell population.
Cell type annotation was performed using established marker gene panels, and malignant epithelial cells were distinguished from normal tubular epithelial cells based on copy number variation inference and transcriptional state analysis. The 219 MECRGs represent genes with consistent differential expression in malignant compared to benign epithelial cells across all analyzed patients.
These MECRGs were then integrated with clinical outcome data from multiple bulk RNA sequencing cohorts to build and validate the prognostic signature. The transition from single-cell discovery to bulk cohort validation ensures that the selected genes are both biologically specific to malignant epithelial cells and clinically prognostic across diverse patient populations.
To select the optimal prognostic signature from the 219 MECRG candidates, an exhaustive machine learning framework tested 101 model combinations derived from 10 different algorithms including CoxBoost, Lasso, Ridge regression, stepwise Cox, random forest, gradient boosting, XGBoost, and others. This systematic comparison ensures the selected model is not dependent on algorithm choice.
The optimal model was identified as the combination of CoxBoost with Ridge regression, achieving the highest concordance index (C-index) across all validation cohorts. The C-index quantifies how accurately the model ranks patients by predicted survival, with values above 0.7 considered strong in oncology prognostics.
Validation was performed across 6 independent cohorts, including both publicly available datasets (TCGA, ICGC) and institutional patient series. This six-cohort validation strategy is unusually comprehensive for a single study and significantly strengthens confidence in the signature's generalizability.
In a head-to-head benchmarking comparison against 92 previously published prognostic signatures for ccRCC, the MECRGS consistently achieved superior or equivalent performance in the majority of validation cohorts. This systematic benchmarking provides strong evidence that the MECRGS represents a genuine advance over existing tools rather than an incremental improvement.
The MECRGS stratified patients into high-risk and low-risk groups with significantly different overall survival in all six validation cohorts, demonstrating exceptional cross-cohort consistency. Such consistency across geographically, temporally, and demographically diverse patient populations is rare in cancer genomics and substantially increases clinical credibility.
Multivariable Cox regression analyses confirmed that MECRGS risk score is an independent prognostic factor after adjusting for age, sex, tumor stage, and histological grade, confirming that the signature provides information beyond what is available from standard clinical parameters alone.
Single-cell analysis identified a specific malignant epithelial cell subtype co-expressing PLOD2 (procollagen-lysine 2-oxoglutarate 5-dioxygenase 2) and SAA1 (serum amyloid A1), termed the PLOD2+SAA1+ subtype. This subtype was enriched in high-MECRGS-risk tumors and was associated with an immunosuppressive tumor microenvironment.
PLOD2+SAA1+ tumor cells were found to recruit CD163+ M2-polarized tumor-associated macrophages (TAMs) and regulatory T cells (Tregs) into the tumor microenvironment through secreted signaling molecules. This active recruitment of immunosuppressive immune cells by the malignant epithelial subtype represents a specific mechanism by which high-MECRGS tumors evade immune destruction.
The identification of a specific tumor cell subtype that orchestrates its own immunosuppressive microenvironment is a significant mechanistic finding, suggesting that PLOD2 and SAA1 may be functional drivers of immune evasion rather than merely markers of aggressive disease, opening potential therapeutic avenues targeting these interactions.
The MECRGS was validated at the protein level in the Renji Hospital tissue microarray (TMA) cohort using immunohistochemistry (IHC) and multiplex immunofluorescence. Protein-level validation is a critical step that demonstrates the signature genes translate from RNA expression differences to functional protein abundance differences detectable in routine pathology specimens.
Multiplex immunofluorescence allowed simultaneous visualization of multiple MECRGS proteins and immune cell markers in the same tissue section, confirming the spatial relationship between PLOD2+SAA1+ tumor cells and CD163+ TAMs and Treg infiltration. This spatial co-localization provides direct in situ evidence for the TME recruitment mechanism identified computationally.
The MECRGS also showed significant correlation with predicted response to immune checkpoint inhibitor therapy, with low-risk patients more likely to respond to anti-PD-1 and anti-CTLA-4 therapies. This predictive capability, combined with its prognostic power, positions MECRGS as a dual prognostic and predictive biomarker with broad clinical utility in the emerging era of ccRCC immunotherapy combinations.