Clear cell renal cell carcinoma (ccRCC) is the predominant histological subtype of kidney cancer, responsible for the majority of kidney cancer-related deaths. The molecular drivers of ccRCC are heterogeneous, involving loss of VHL tumor suppressor function, hypoxia-inducible factor activation, and widespread transcriptomic reprogramming. Identifying reliable biomarkers that distinguish ccRCC from normal kidney tissue and predict clinical outcomes remains an active area of investigation.
Gene expression profiling through public databases such as the Gene Expression Omnibus (GEO) and The Cancer Genome Atlas (TCGA) provides large-scale transcriptomic datasets that enable systematic identification of differentially expressed genes (DEGs) between tumor and normal tissue. Combining multiple independently generated expression datasets increases statistical power and reduces dataset-specific noise, improving the reliability of identified DEGs as true biological signals.
Artificial neural networks offer a complementary analytical approach for biomarker discovery by learning non-linear relationships between gene expression patterns and clinical outcomes without requiring prior assumptions about gene interactions. Integrating computational network biology with neural network modeling can identify hub genes that are both transcriptionally dysregulated in ccRCC and strongly associated with patient prognosis and tumor stage.
Two independent GEO expression datasets (GSE105288 and GSE40435) containing ccRCC tumor and adjacent normal kidney tissue samples were analyzed using the limma package in R to identify differentially expressed genes in each dataset separately. The 251 genes that were differentially expressed in the same direction across both datasets were designated as common DEGs and used for downstream analysis, ensuring that retained genes represent robust signals reproducible across independent experimental platforms.
The 251 common DEGs were mapped onto the STRING protein-protein interaction database to construct a PPI network comprising 189 nodes and 406 edges. Network topology analysis was used to identify hub genes based on the degree of connectivity within the network, reflecting the number of interaction partners each protein has. High-connectivity hub genes in cancer-associated PPI networks often represent central regulators of tumor biology with elevated relevance as therapeutic targets or biomarkers.
The top 10 hub genes ranked by network connectivity were identified as VEGFA, AURKB, CCNA2, MCM2, MCM7, SMC4, TPX2, SLC2A1, MCM5, and NCAPG. These genes were then subjected to clinical correlation analysis to assess associations with pathological stage and overall survival in both the GEO datasets and the independent TCGA ccRCC cohort comprising 537 tumor samples and 407 normal kidney samples.
Among the 10 hub genes, four showed statistically significant associations with both higher pathological tumor stage and worse overall survival in ccRCC patients: AURKB, CCNA2, TPX2, and NCAPG. All four genes were significantly overexpressed in higher-stage tumors compared to lower-stage tumors, and elevated expression was associated with shorter overall survival in Kaplan-Meier analysis with log-rank p-values below 0.05.
ROC curve analysis assessed the ability of individual hub gene expression levels to distinguish ccRCC from normal kidney tissue in the GSE105288 dataset. NCAPG achieved the highest AUC at 0.940, followed by AURKB at 0.937, CCNA2 at 0.933, and TPX2 at 0.895. These AUC values indicate that individual gene expression levels can reliably discriminate tumor from normal tissue, supporting their value as diagnostic biomarkers.
VEGFA, while a well-established oncogene in ccRCC linked to tumor angiogenesis and the VHL-HIF pathway, was not significantly associated with overall survival in this analysis, possibly reflecting the complex prognostic landscape of angiogenic signaling and the effects of anti-VEGF therapies in the study population. The four prognostically significant hub genes (AURKB, CCNA2, TPX2, NCAPG) all participate in cell cycle regulation and mitotic apparatus function.
A multilayer artificial neural network was trained using MATLAB to classify samples as ccRCC or normal kidney tissue based on gene expression features derived from the hub gene set. The neural network architecture was optimized for classification performance, with hyperparameters including the number of hidden layers, neurons per layer, and training algorithm selected through iterative evaluation on the training dataset.
The trained neural network achieved a correlation coefficient (R) of 0.9906 between predicted and actual class labels on the training set and 0.9977 on the validation set. These high R values indicate near-perfect fitting of the neural network model to the training data, while the maintained performance on the validation partition suggests that generalization was preserved without severe overfitting.
The neural network model provided an independent line of evidence for the discriminatory value of the hub gene expression signature, complementing the univariate ROC analysis by demonstrating that a multivariate nonlinear model trained on the same gene set achieves strong classification accuracy. The high validation R value is particularly notable given the small number of input features derived from the hub gene analysis.
The four prognostically significant hub genes (AURKB, CCNA2, TPX2, NCAPG) share functional roles in cell cycle progression and mitotic regulation. AURKB (Aurora Kinase B) is a serine-threonine kinase essential for chromosome alignment and segregation during mitosis, and its overexpression in cancer promotes chromosomal instability. CCNA2 encodes Cyclin A2, a core regulator of S-phase entry and mitotic initiation.
TPX2 (Targeting Protein for Xklp2) is required for mitotic spindle assembly and is transcriptionally regulated by the E2F family of transcription factors, which are frequently activated downstream of RB pathway disruption in ccRCC. NCAPG (Non-SMC Condensin I Complex Subunit G) is essential for chromosome condensation during mitosis and has been associated with aggressive phenotypes across multiple cancer types.
The convergence of prognostic relevance among genes governing cell cycle entry and mitotic fidelity suggests that cell cycle dysregulation is a dominant contributor to ccRCC aggressiveness and clinical outcome. This biological coherence strengthens confidence that the identified hub genes represent true drivers of aggressive tumor behavior rather than statistical artifacts of the analytical pipeline.
This study identifies AURKB, CCNA2, TPX2, and NCAPG as candidate biomarkers with dual utility for ccRCC diagnosis and prognosis, supported by multi-dataset transcriptomic analysis, PPI network topology, ROC curve performance, survival analysis, and independent neural network validation. The convergent evidence across multiple analytical frameworks strengthens the candidacy of these genes for further clinical investigation.
The availability of Aurora Kinase inhibitors in clinical development provides a potential therapeutic angle for AURKB-overexpressing ccRCC tumors. If high AURKB expression identifies a subgroup of ccRCC patients with particular sensitivity to Aurora Kinase B inhibition, the biomarker could serve both as a prognostic stratification tool and a predictive biomarker for targeted therapy selection.
Prospective validation of the four-hub-gene signature in independent ccRCC cohorts with standardized expression profiling platforms is needed before clinical implementation. Development of immunohistochemistry-based protein expression assays for AURKB, CCNA2, TPX2, and NCAPG applicable to formalin-fixed paraffin-embedded tumor specimens would enable integration of this biomarker panel into routine pathological evaluation of surgically resected ccRCC.