Clear cell renal cell carcinoma (ccRCC) accounts for approximately 80% of all kidney cancers and remains one of the most lethal urological malignancies when detected at an advanced stage. With roughly 63,990 new cases and 14,400 deaths in the United States in 2017 alone, accurate staging is critical for guiding treatment decisions and estimating patient survival.
Standard staging relies on imaging techniques such as CT, MRI, and bone scans, combined with the TNM staging system. While widely used, these approaches have significant limitations and cannot always reliably distinguish early-stage disease from late-stage disease at the molecular level, where treatment strategies may differ substantially.
Fatality rates are meaningfully higher when ccRCC is discovered in later stages, making early and precise classification essential. The development of molecular biomarkers derived from gene expression data offers a complementary approach to existing imaging methods, potentially enabling more cost-effective and biologically grounded staging.
Researchers analyzed gene expression data from 523 ccRCC patient samples drawn from The Cancer Genome Atlas (TCGA) to identify genes differentially expressed between early-stage (Stage I-II) and late-stage (Stage III-IV) disease. The dataset contained RSEM quantification values for over 20,000 genes across tumor samples.
The study pursued two complementary strategies: a threshold-based approach that ranks individual genes by their discriminatory power, and a machine learning approach using tools such as Support Vector Machines (SVM), Random Forests, and Naive Bayes classifiers applied to curated gene subsets.
The ultimate output of this work was CancerCSP, a publicly accessible web server that allows researchers and clinicians to input RNA-seq expression values and receive a predicted cancer stage (early or late) along with statistical context for each gene used in the model.
For single-gene classification, researchers developed threshold-based models for each of the 19,166 genes in the dataset, ranking them by the area under the ROC curve (AUC). The gene NR3C2 emerged as the top performer, achieving 71.12% accuracy and an ROC of 0.67, with overexpression in early-stage disease indicating a lower-risk classification.
To move beyond single-gene models, the top 50 genes were selected and filtered by removing highly correlated pairs, yielding a refined set of 28 non-redundant genes (RCSP-set-Threshold). SVM models trained on this set reached 73.27% accuracy and ROC 0.78 during cross-validation, with performance validated on an independent test set at ROC 0.77.
Researchers then split the 28 genes into two functional groups: those overexpressed in early-stage disease and those overexpressed in late-stage disease. Iterative combination of the top candidates yielded two four-gene panels (setA-1 and setB-1). When combined into an eight-gene Combo-1 panel, the SVM model achieved validation ROC of 0.80 with 74% accuracy.
In addition to threshold and SVM-based selection, the study employed Weka software with a Fast Correlation-Based Feature selection algorithm to reduce the gene space. By resampling the dataset 100 times and selecting features with maximum overlap across iterations, a stable set of 64 genes (RCSP-set-Weka) was identified.
SVM models built on this 64-gene set achieved the study's best overall performance: 78.18% accuracy and ROC 0.83 on training data, and 72.64% accuracy with ROC 0.81 on the independent validation set. These results demonstrate that automated feature selection can meaningfully improve classification performance compared to simpler threshold methods.
A biologically motivated subset of 38 genes (RCSP-set-Weka-Hall) was further derived by restricting Weka selection to genes annotated to known cancer hallmark processes such as apoptosis, cell cycle, DNA repair, and immune response. This hallmark-focused set matched the full 64-gene panel in validation accuracy, confirming that cancer-relevant genes are enriched among the most predictive features.
The study examined whether separate male and female models would outperform a combined model. Among the 523 patients (338 males, 185 females), only two genes overlapped in the top ten differentially expressed genes between early and late stage for each gender, indicating fundamentally different molecular stage signatures.
Gender-specific SVM models trained using Weka-selected features achieved notable performance improvements. The female-specific model reached 78.95% validation accuracy and ROC 0.82, while the male-specific model reached 77.14% validation accuracy and ROC 0.80, both outperforming the combined model on their respective subsets.
Network analysis of gender-specific biomarkers revealed distinct biological programs: female-associated proteins clustered around GAPDH as a hub and involved MAP kinase and PI3K/AKT signaling, while male-specific proteins were enriched for intracellular transport functions. These findings underscore the importance of sex-stratified analyses in cancer genomics research.
RSEM gene expression values were log2-transformed and normalized to Z-scores using mean and standard deviation from the training set. Low-variance features (below 0.25) were removed, reducing the feature space from 20,531 to 19,166 genes. The 523-patient dataset was split 80/20 into training (419 patients) and independent external validation (104 patients) sets.
All machine learning models were evaluated using 10-fold cross-validation on training data and assessed independently on the held-out validation set. Performance metrics included sensitivity, specificity, overall accuracy, Matthews Correlation Coefficient (MCC), and the area under the ROC curve (AUC), providing a comprehensive view of classifier quality beyond simple accuracy.
Feature robustness was further tested by resampling the dataset 100 times and selecting features with maximum overlap across all iterations. This bootstrap-like approach guards against overfitting and ensures that the reported gene sets reflect stable signals in the data rather than artifacts of a single random split.
This study successfully compressed a search space of nearly 20,000 genes down to as few as eight genes capable of staging ccRCC with reasonable accuracy. The Combo-1 eight-gene threshold model achieved ROC 0.77, and the Weka-selected 64-gene SVM model reached ROC 0.81 on the validation dataset, establishing competitive performance against prior work.
Several identified biomarkers have known roles in cancer biology: NR3C2 is a recognized tumor suppressor with reduced expression across multiple cancer types; DNASE1L3 is involved in apoptosis induction; and BMP5 shows lower expression in later-stage tumors. Some genes, such as ENAM, represent novel candidates whose roles in renal cancer had not been previously investigated.
The CancerCSP web server enables researchers to submit RSEM gene expression values and receive predictions of early or late cancer stage using any of the validated gene panels. The platform also provides a data analysis module that calculates per-gene p-values, threshold scores, and Z-score normalization, making it a practical tool for both laboratory and clinical research settings.