Clear cell renal cell carcinoma (ccRCC) staging determines treatment selection: early-stage disease (stages I and II) is typically managed with nephron-sparing surgery or active surveillance, while late-stage disease (stages III and IV) requires more aggressive systemic intervention.
Accurate staging is critical because misclassification in either direction leads to harm: under-staging may delay necessary treatment, while over-staging may expose patients to surgical risks or therapies they do not need. Standard clinical staging based on imaging and pathology has known limitations.
Gene expression data from tumor samples captures molecular aspects of disease biology that correlate with stage but are not directly measured by imaging or pathological examination. Integrating genomic features could improve staging precision beyond conventional approaches.
This study focused on developing a gene set-based machine learning classifier to distinguish early (stages I+II) from late (stages III+IV) ccRCC using TCGA gene expression data, with emphasis on feature engineering strategies that improve both accuracy and model stability.
The TCGA-KIRC dataset provided gene expression profiles for 604 ccRCC samples: 482 were used for training and 122 for testing. The binary staging target was defined as early stage (I+II) versus late stage (III+IV) based on clinical annotations.
Raw gene expression values are continuous and often noisy. This study applied Chi-merge binning, a supervised discretization technique that groups continuous expression values into bins based on their association with the class label. This converts continuous features into categorical intervals that capture the most informative expression thresholds.
Weight of Evidence (WOE) encoding was then applied to convert the discretized bins into a single numerical value reflecting the log odds ratio of the positive class within that bin. WOE encoding makes the relationship between gene expression level and staging outcome explicit and monotonic, which benefits many classifiers.
Multiple gene selection strategies were compared, with the final FJL-set comprising 23 genes identified by combining FCBFSearch (Fast Correlation-Based Filter), joint statistical measures, and logistic regression feature importance. This ensemble of selection methods reduces the chance that any single selection artifact drives the result.
FCBFSearch is a filter-based feature selection method that uses symmetrical uncertainty to identify features that are highly correlated with the class label while minimally correlated with each other. It efficiently handles the high dimensionality of gene expression data without a classifier in the loop.
Joint statistical measures incorporate multiple statistical tests simultaneously, selecting genes that are consistently significant across different distributional assumptions. This multi-test approach guards against spurious selections that happen to pass one particular statistical threshold.
Logistic regression with regularization was applied as a wrapper step to refine the gene set further, retaining only genes that contribute independently to classification accuracy. The combination of filter (FCBFSearch), statistical (joint measures), and wrapper (logistic regression) selection constitutes the FJL pipeline.
The resulting 23-gene FJL-set was significantly smaller than the initial gene space, demonstrating substantial dimensionality reduction while preserving the most stage-informative biological signal. Smaller gene sets also reduce the cost of potential clinical translation to targeted assays.
Among the classifiers evaluated (including SVM, random forest, naive Bayes, and others), the Support Vector Machine (SVM) achieved the best performance on the test set using the FJL-set: 81.15% accuracy and an AUC of 0.86 for early versus late stage classification.
The AUC of 0.86 indicates strong discriminative ability, significantly better than chance and competitive with other molecular staging approaches reported in the literature for ccRCC. Sensitivity and specificity were balanced, confirming the model generalizes across both staging classes.
Compared to models using raw gene expression values without Chi-merge and WOE preprocessing, the preprocessed models showed improved stability across cross-validation folds. This demonstrates that the feature engineering steps contribute directly to model reliability, not just nominal accuracy.
The 23-gene FJL-set outperformed larger gene sets selected by simpler methods, suggesting that aggressive feature engineering combined with principled selection produces more generalizable models than simply including more features with raw expression values.
Among the 23 selected genes, several are known regulators of cancer hallmarks in ccRCC: genes involved in hypoxia response, metabolic reprogramming under VHL mutation, immune evasion, and cell cycle control collectively appear in the FJL-set.
The presence of immune-related genes in the stage predictor aligns with the known importance of the tumor immune microenvironment in ccRCC progression and with the clinical success of immune checkpoint inhibitors in advanced disease stages.
Metabolic genes reflect the profound rewiring of glucose and lipid metabolism that occurs as ccRCC transitions from localized to invasive disease, consistent with the metabolic underpinnings of VHL-loss-driven tumorigenesis.
The convergence of genes from these diverse pathways into a single 23-gene classifier suggests that late-stage ccRCC is not driven by a single pathway but by the simultaneous engagement of multiple oncogenic programs, each contributing a dimension of predictive information.
This study demonstrates that thoughtful feature engineering, specifically Chi-merge binning and WOE encoding, meaningfully improves gene expression classifiers for ccRCC staging beyond what raw expression values alone can achieve.
The 23-gene FJL-set achieves strong performance (AUC 0.86) with a gene count small enough to be practically converted into a targeted molecular assay, making clinical translation feasible without whole-transcriptome profiling.
Prospective validation in cohorts with matched gene expression and staging data from independent institutions is needed before clinical deployment. Comparison with established genomic panels and integration with imaging-based staging variables could further improve accuracy.
The methodological framework, combining multi-strategy gene selection with supervised feature engineering, provides a template applicable to other cancer staging and prognosis problems where raw omics data requires transformation to achieve stable, interpretable classifiers.