Clear cell renal cell carcinoma (ccRCC) accounts for the majority of kidney cancer cases and presents significant clinical challenges due to its heterogeneous behavior and variable response to systemic therapies. Accurate prognostic tools are urgently needed to stratify patients and guide treatment decisions.
The F-box protein family, comprising over 60 members in humans, serves as substrate-recognition subunits of SCF ubiquitin ligase complexes. These proteins regulate cell cycle progression, apoptosis, and oncogenic signaling by controlling targeted protein degradation, making them biologically relevant candidates for cancer prognosis modeling.
Prior research has established individual F-box genes as tumor suppressors or oncogenes in various cancers, but no study had systematically evaluated the entire F-box family's collective prognostic power in ccRCC using deep learning approaches before this work.
This study aimed to construct a deep learning risk score (FB-risk) using F-box gene expression profiles from multiple ccRCC patient cohorts to predict overall survival and immunotherapy response with greater robustness than traditional regression-based biomarkers.
The FB-risk model was constructed using a deep neural network trained on RNA sequencing expression data from three independent ccRCC cohorts: TCGA-KIRC (n=530), ICGC (n=91), and an external validation dataset. The multi-cohort approach was designed to ensure generalizability across different patient populations and sequencing platforms.
Feature selection first identified F-box genes significantly associated with overall survival using univariate Cox regression, filtering the input space before training. The neural network was then optimized to output a continuous risk score, with patients subsequently dichotomized into high-risk and low-risk groups based on median cutoff values derived from the training cohort.
Model performance was assessed using Kaplan-Meier survival analysis, log-rank tests, and time-dependent receiver operating characteristic curves. The prognostic independence of the FB-risk score was confirmed through multivariate Cox regression adjusting for age, sex, tumor grade, and TNM staging.
Pan-cancer validation was performed across multiple TCGA cancer types to assess whether the F-box-based deep learning signal extended beyond ccRCC, providing broader biological context for the identified gene signatures.
Patients classified as high-risk by the FB-risk model showed significantly worse overall survival compared to low-risk patients across all three validation cohorts. The model consistently achieved high C-index values, demonstrating robust discrimination ability that was reproducible across independent datasets.
Time-dependent AUC analysis confirmed that FB-risk outperformed traditional clinicopathological variables such as tumor stage and grade at one-, three-, and five-year survival prediction timepoints. This improvement in predictive accuracy was statistically significant and clinically meaningful.
Multivariate Cox regression analysis confirmed that FB-risk remained an independent prognostic factor after adjusting for standard clinical variables, establishing its potential utility as a standalone biomarker rather than merely a surrogate for known prognostic features.
Pan-cancer analysis revealed that several F-box genes driving the model's performance showed consistent prognostic associations across multiple TCGA cancer types, suggesting the model captures fundamental oncogenic mechanisms rather than ccRCC-specific artifacts.
Unsupervised consensus clustering of ccRCC patients using F-box gene expression profiles identified two novel molecular subtypes designated C3A and C3B, which showed distinct survival trajectories and biological characteristics not captured by standard clinical classification.
The C3A subgroup was characterized by better overall survival, lower immune infiltration scores, and a tumor microenvironment profile suggesting reduced immunosuppression, while C3B patients exhibited worse prognosis with higher expression of immune checkpoint molecules and greater immune cell infiltration dominated by suppressive cell types.
Gene ontology and pathway enrichment analyses revealed that C3A tumors were enriched for metabolic and cell cycle regulation pathways, whereas C3B tumors showed upregulation of epithelial-to-mesenchymal transition, angiogenesis, and inflammatory response gene sets consistent with a more aggressive phenotype.
These phenotypic distinctions were validated across independent cohorts, supporting the biological and clinical relevance of C3A and C3B as reproducible molecular subtypes that could inform patient stratification beyond conventional staging systems.
Among the F-box genes driving prognostic performance, FBXL3 and FBXO3 emerged as particularly significant markers linked to immunotherapeutic response prediction. Both genes showed differential expression between responders and non-responders to immune checkpoint inhibitor therapy in available clinical datasets.
FBXL3, which regulates circadian clock protein degradation and has known roles in immune cell activation, was expressed at higher levels in tumors from patients who responded favorably to anti-PD-1/PD-L1 therapy, suggesting it may serve as a companion diagnostic marker for immunotherapy selection.
FBXO3 expression correlated with immune cell infiltration patterns and PD-L1 expression levels, indicating a potential mechanistic link between F-box-mediated ubiquitination pathways and the tumor's immunological status that could be therapeutically exploited.
These findings position FBXL3 and FBXO3 as actionable biomarkers that bridge prognostic risk stratification with treatment selection, offering a pathway toward more personalized immunotherapy decisions in ccRCC patients.
The FB-risk deep learning model represents a significant advance in ccRCC prognosis by leveraging the collective biological signal of the F-box gene family rather than individual biomarkers, yielding a more robust and generalizable risk stratification tool.
The identification of C3A and C3B molecular subtypes alongside actionable immunotherapy biomarkers demonstrates that transcriptomic deep learning can simultaneously address multiple clinical needs: survival prediction, molecular classification, and therapy guidance within a unified analytical framework.
Future work should focus on prospective validation of FB-risk in clinical trial cohorts and integration with liquid biopsy or imaging data to create multimodal prognostic systems that are more practical for routine clinical deployment.
The study also highlights the broader applicability of the F-box family as a prognostically relevant gene set across multiple cancer types, suggesting that similar deep learning approaches may prove valuable in other malignancies where F-box proteins play oncogenic roles.