Clear cell renal cell carcinoma (ccRCC) accounts for 80% of all renal cell carcinoma cases and is particularly aggressive, with over 30% of patients developing metastatic disease. Despite this prevalence, the complete landscape of ccRCC risk genes remains poorly understood, limiting the development of effective diagnostic markers and therapeutic targets.
Traditional approaches to cancer risk gene identification rely either on mutation frequency alone or on supervised machine learning models trained on known risk gene labels. Frequency-based methods miss genes with low mutation rates despite high biological impact (for example, PIK3CA is mutated in only 5% of ccRCC patients yet is a validated driver). Supervised methods require large labeled training sets, but only 44 ccRCC risk genes are currently catalogued in the IntOGen database, far too few to train reliable classifiers.
Protein-protein interaction (PPI) networks offer a complementary approach by enabling identification of risk genes through their functional proximity to known drivers, even when those genes have low mutation frequencies. The challenge is designing a method that effectively integrates mutation data and network structure without relying on large labeled training sets of confirmed risk genes.
The authors developed RL-GenRisk (Reinforcement Learning-based GENe RISK), which frames ccRCC risk gene identification as a Markov Decision Process. In this formulation, the PPI network is the environment, nodes are genes, and the agent sequentially adds genes to a growing subgraph by selecting which neighboring gene to include next. This enables exploration of the gene interaction landscape without requiring large labeled training sets.
The policy network of RL-GenRisk combines a Graph Convolutional Network (GCN) to encode state representations capturing both network topology and gene mutation information, with a Deep Q-Network (DQN) that maps states to action values (Q values) for each gene. Training uses an epsilon-greedy exploration strategy. After training, genes are ranked by the Q values assigned to each node starting from an empty subgraph, with higher Q values indicating greater predicted risk.
A key innovation is the data-driven reward function, designed to address the scarcity of labeled risk genes. Rather than relying on binary labels (known risk gene or not), the reward integrates information from both the PPI network structure and mutation frequency distributions across the sampled subgraph, enabling the model to learn meaningful signals from the data itself without requiring extensive labeled supervision.
RL-GenRisk was benchmarked against eight existing methods across five PPI networks (HPRD, STRING, Multinet, IRefIndex, HumanNet) and three reference gene sets (IntOGen, NCG, and a merged dataset). Evaluation used discounted cumulative gain (DCG), normalized DCG, DCG-AUC, and average precision metrics. RL-GenRisk achieved the highest scores on all four metrics across all evaluated datasets, demonstrating consistent superiority over frequency-based methods like MutSigCV, network propagation methods like HotNet2 and HierarchicalHotNet, and supervised classifiers like SVM and Random Forest.
PPI network perturbation experiments confirmed that the network structure is essential for detecting low-frequency risk genes. When edges were randomly swapped at increasing rates (25% to 100%), three known ccRCC risk genes with mutation frequencies below 5% (TP53, PIK3CA, SPEN) dropped out of the top 20 rankings as perturbation increased. This demonstrated that RL-GenRisk uses PPI network information rather than mutation frequency alone to elevate these biologically important but rarely mutated genes.
Among 411 ccRCC patients from TCGA, 77.13% (317 patients) carried somatic mutations in at least one of the top 20 high-confidence risk genes (HRGs) identified by RL-GenRisk, with missense mutations predominating. The most significantly enriched pathway for the top 20 HRGs was the clear cell renal cell carcinoma pathway (FDR p-value = 1.48e-8), and enrichment in renal neoplasm phenotypes was confirmed by the Human Phenotype Ontology database.
Among the top 20 HRGs, 8 genes were not present in existing ccRCC risk gene databases. Differential expression analysis in TCGA identified EGFR and PCLO as significantly upregulated in ccRCC tumor tissue compared to normal kidney tissue (EGFR FDR p-value = 2.07e-38, PCLO FDR p-value = 1.04e-14). Validation in two independent GEO datasets (GSE46699 and GSE36895) confirmed differential expression for 5 of the 8 novel HRGs.
EGFR was upregulated in tumor cells compared to normal cells at both bulk and single-cell levels, with significantly higher expression specifically in tumor cells from single-cell RNA-seq data (FDR p-value less than 1.21e-4). EGFR encodes the epidermal growth factor receptor, which activates RAS/RAF/MAPK, AKT, and JAK/STAT signaling pathways involved in cell proliferation, survival, and migration. Protein-level analysis using TCGA reverse-phase protein array data showed that the top 25% of patients by EGFR protein expression had significantly worse progression-free survival (log-rank p = 0.0018) and disease-specific survival (log-rank p = 0.021).
Pathway enrichment analysis of the top 20 RL-GenRisk HRGs showed significant enrichment in cancer-related biological processes including cell adhesion, regulation of cell proliferation, and cell migration. Among all competing methods, the top 20 genes from RL-GenRisk showed the most significant enrichment in the clear cell RCC pathway, and RL-GenRisk's top 20 genes included the highest number (14) of genes significantly differentially expressed across both independent GEO validation datasets.
shRNA-mediated EGFR knockdown in ACHN and 786-O ccRCC cell lines significantly inhibited cell viability (CCK-8), reduced colony formation, suppressed migration (transwell assay), and promoted apoptosis (flow cytometry). EGFR overexpression in 786-O cells had the opposite effect, significantly increasing migration. These experiments establish EGFR as a functional driver of ccRCC progression rather than simply a biomarker.
The EGFR inhibitor erlotinib was tested on ccRCC cells and tumor-bearing mice. In vitro, erlotinib inhibited cell migration and growth and promoted apoptosis. In vivo xenograft mouse models showed that both erlotinib treatment and EGFR knockdown significantly repressed tumor growth, demonstrating that EGFR is pharmacologically actionable in ccRCC and that its inhibition has measurable antitumor effects in a preclinical setting.
PCLO knockdown in 786-O cells using shRNA significantly inhibited ccRCC cell progression in vitro, including reduced viability and impaired migration. PCLO encodes a component of the presynaptic cytomatrix, a protein whose role in kidney cancer had not been previously explored. These functional experiments validate PCLO as a genuine regulator of ccRCC cell behavior beyond its known neurological functions.
RL-GenRisk addresses a fundamental challenge in cancer genomics: identifying biologically important risk genes that occur at too low a mutation frequency to be detected by frequency-based approaches. The reinforcement learning framework does not require large labeled training sets, making it applicable to cancer types where few confirmed driver genes are known, a situation that applies across many rare cancers.
A practical limitation is that the approach requires high-quality PPI network data, and network perturbation experiments showed that accuracy degrades as network integrity is disrupted. The quality and completeness of the underlying PPI network therefore influences which low-frequency genes can be recovered, and different networks produced somewhat different rankings across comparison methods.
The 6 other novel HRGs beyond EGFR and PCLO were identified bioinformatically but not yet experimentally validated, representing hypotheses for future study. The authors note that the RL-GenRisk framework is generalizable to other diseases where the number of confirmed risk genes is limited, and propose it as a template for applying reinforcement learning to network-based gene prioritization problems beyond cancer.