One of the defining characteristics of clear cell renal cell carcinoma (ccRCC) is that tumor cells behave as if they are starved of oxygen - even when oxygen is plentiful. This is driven by mutations in the VHL gene, which normally helps cells respond to oxygen levels. When VHL is lost, proteins called HIFs (hypoxia-inducible factors) become permanently activated, switching on hundreds of genes that would normally only turn on when oxygen is scarce.
The resulting shift in gene expression - called the hypoxia response - fundamentally rewires the tumor's metabolism and drives aggressive behavior. Several specific genes tied to this hypoxic state have been identified as prognostically important in ccRCC, including KLF6, BCL2, ETS1, PLOD2, and PPARGC1A. Knowing whether these genes are overexpressed or underexpressed in a patient's tumor can inform prognosis and treatment decisions.
Currently, measuring gene expression requires tumor tissue, obtained either through surgery or biopsy. But what if CT scan images - already routinely collected before and after treatment - could reveal information about a tumor's gene expression profile without requiring tissue? Radiogenomics is the field exploring exactly this question, seeking to link imaging features with genomic characteristics of tumors.
The study analyzed CT scans and genetic data from 190 patients in the TCGA-KIRC (The Cancer Genome Atlas Kidney Renal Clear Cell Carcinoma) database. From each CT scan, researchers extracted 2,824 radiomic features - quantitative measurements of texture, shape, and intensity patterns within the tumor region identified on the scan. These features capture subtle properties of how the tumor appears on imaging that cannot be readily perceived by eye.
Because many of these 2,824 features are redundant or irrelevant, a selection process was applied to identify the 388 most reproducible and informative features. This filtered set was then used to train a Random Forest (RF) machine learning model - an approach that builds many decision trees and combines their predictions to produce a robust final result.
The model was trained to predict whether each of five hypoxia-related genes (KLF6, BCL2, ETS1, PLOD2, PPARGC1A) was highly or lowly expressed in the patient's tumor, using only CT scan features as input. The model's predictions were validated and the most predictive imaging features were analyzed to understand which aspects of tumor appearance best reflected underlying gene expression.
The Random Forest model successfully predicted the expression levels of all five hypoxia-related genes from CT radiomic features, with the best performance achieved for KLF6, ETS1, and BCL2. Two types of radiomic features were most consistently predictive across the genes analyzed: GLDM (Gray Level Dependence Matrix) features extracted in 2D and GLCM (Gray Level Co-occurrence Matrix) features extracted in 3D. These feature types capture how pixel intensities relate to one another locally within the tumor, reflecting tissue heterogeneity and internal structure.
Among the five genes, KLF6 showed the most notable relationship with clinical tumor characteristics. KLF6 expression decreased significantly in higher tumor grades and later disease stages - meaning that as the tumor becomes more aggressive and advanced, this gene becomes less active. This pattern makes KLF6 a potentially valuable prognostic marker, and the ability to estimate its expression from CT scans could add clinical value at the time of initial diagnosis.
The successful prediction of PLOD2 and PPARGC1A underexpression from imaging features is also noteworthy. PLOD2 is involved in collagen modification and has been linked to tumor invasiveness, while PPARGC1A is a key regulator of mitochondrial function. Both are relevant to ccRCC biology and therapy development, and their imaging correlates may reflect structural or metabolic properties of the tumor visible on CT.
The clinical appeal of radiogenomics is straightforward: CT scans are already a standard part of kidney cancer diagnosis and follow-up, while genomic testing requires tissue samples that carry procedural risks and are not always feasible. If validated, a tool that estimates key gene expression patterns from CT scans could provide molecular information about a tumor without additional invasive procedures.
For KLF6 specifically, this could be particularly useful. Since KLF6 expression correlates with tumor grade and stage, an imaging-based estimate of KLF6 activity could complement conventional staging information to provide a more complete picture of tumor aggressiveness at the time of initial diagnosis. Patients with CT features suggesting low KLF6 expression might warrant closer surveillance or more aggressive treatment.
More broadly, this research supports the concept of imaging biomarkers - specific measurable properties of tumor images that serve as proxies for underlying molecular characteristics. As radiomics technology matures and more radiogenomic correlations are discovered and validated, CT scans may become much richer sources of prognostic and predictive information than they are today.
This study establishes that CT-based radiomic features can predict the expression level of hypoxia-related genes in ccRCC with meaningful accuracy. The finding that texture features - particularly GLDM and GLCM metrics - are most predictive suggests that the internal structural complexity of kidney tumors on imaging reflects real differences in their molecular biology.
For patients and families, this research represents a step toward a future where a standard CT scan provides not just anatomical information about a tumor's size and location, but also estimates of its molecular properties - all without additional biopsies or laboratory tests. This could accelerate treatment planning and improve the precision of initial risk assessment.
The next steps include validating these radiomic-genomic associations in independent patient cohorts, ideally across multiple institutions with different CT scanner types and protocols. If the associations are robust across diverse settings, this approach could eventually be incorporated into clinical workflow as part of a comprehensive pre-treatment tumor profiling process for kidney cancer patients.