Clear cell renal cell carcinoma (ccRCC) is the most common and lethal form of kidney cancer, driven predominantly by loss of the VHL tumor suppressor gene. VHL loss leads to constitutive activation of hypoxia-inducible factors, particularly HIF-2alpha (encoded by EPAS1), which drives a transcriptional program promoting angiogenesis, metabolic reprogramming, and immune evasion.
Targeted therapies such as VEGF inhibitors and mTOR inhibitors have improved outcomes in ccRCC, but most patients eventually develop resistance. EPAS1/HIF-2alpha has emerged as a promising next-generation drug target, with belzutifan becoming the first approved HIF-2alpha inhibitor, though resistance mechanisms are already being described.
This study combined single-cell RNA sequencing (scRNA-seq) analysis of the ccRCC tumor microenvironment with deep learning-based drug screening to discover novel small molecule compounds targeting EPAS1, potentially expanding the therapeutic arsenal against ccRCC.
scRNA-seq was performed on tumor samples from 6 ccRCC patients, generating high-resolution transcriptomic profiles of 31,625 individual cells. This approach allowed comprehensive characterization of all cell types within the tumor microenvironment (TME) at single-cell resolution, revealing cellular heterogeneity invisible to bulk RNA sequencing.
Unsupervised clustering identified major TME components including malignant epithelial cells, T lymphocytes, tumor-associated macrophages (TAMs), cancer-associated fibroblasts (CAFs), endothelial cells, and other stromal populations. Each cluster was annotated using established marker genes specific to each cell type.
Within the malignant cell population, EPAS1 was confirmed as a highly expressed and functionally active transcription factor, with its target gene network enriched across tumor cells from all six patients. This consistent expression pattern reinforced EPAS1 as the primary therapeutic target for the subsequent drug discovery phase.
The drug discovery pipeline employed a directed message passing neural network (DMPNN), a type of graph neural network that represents molecules as graphs with atoms as nodes and bonds as edges. DMPNN learns molecular features automatically from structural representations, without requiring hand-crafted chemical descriptors.
DMPNN predictions were further refined using gradient boosted decision trees (GBDT) to predict EPAS1 binding affinity and functional activity. This ensemble approach combined the structural learning capabilities of deep learning with the robustness of gradient boosting for final compound ranking.
The screening was applied to a large library of commercially available and FDA-approved compounds, prioritizing compounds that could be repurposed for ccRCC treatment. Repurposing existing drugs accelerates the path to clinical application by bypassing many early safety and formulation hurdles.
The deep learning pipeline identified five candidate compounds with predicted EPAS1 binding and functional activity. Two of these compounds, flufenamic acid and fludarabine, are already FDA-approved for other indications, making them immediate candidates for drug repurposing trials in ccRCC without needing to clear the early phases of drug development.
Flufenamic acid is a non-steroidal anti-inflammatory drug (NSAID) with anti-inflammatory and ion channel modulating properties, while fludarabine is a purine analog used in treating certain leukemias and lymphomas. Their repurposing potential in ccRCC depends on achieving effective concentrations at the tumor site with acceptable toxicity profiles.
The three additional candidate compounds represent novel chemical scaffolds not previously approved for any indication. These compounds would require conventional drug development pathways but potentially offer more optimized EPAS1-targeting properties than the repurposed drugs.
Analysis of the T cell compartment revealed functionally distinct subpopulations including naive, effector, memory, and exhausted T cell states. Exhausted CD8+ T cells expressing high levels of checkpoint receptors such as PD-1, TIM-3, and LAG-3 were enriched in the tumor core, consistent with an immunosuppressive TME promoted by EPAS1-driven transcriptional programs.
Tumor-associated macrophages (TAMs) were predominantly polarized toward the immunosuppressive M2 phenotype, expressing markers such as CD163 and CD204. M2 TAMs contribute to immune exclusion by secreting immunosuppressive cytokines and remodeling the extracellular matrix to impede T cell infiltration.
Cancer-associated fibroblasts (CAFs) and endothelial cells showed transcriptional signatures consistent with EPAS1-driven angiogenic and stromal remodeling programs. These findings suggest that EPAS1 inhibition could have pleiotropic effects on the TME beyond direct anti-tumor activity, potentially restoring immune competence.
The integration of single-cell transcriptomics with deep learning drug discovery represents a powerful new paradigm for cancer drug development. By grounding compound selection in the actual molecular landscape of patient tumors at single-cell resolution, this approach generates candidates with direct mechanistic relevance to the disease.
The identification of FDA-approved compounds as candidates enables rapid preclinical validation and accelerated clinical translation. Fludarabine and flufenamic acid could be tested in combination with existing VEGF or mTOR inhibitors in preclinical ccRCC models, with promising results potentially leading to early-phase clinical trials within a shorter timeframe than de novo drug candidates.
Longer term, this combined approach could be applied systematically across cancer types to discover target-specific compounds guided by patient-derived single-cell data, accelerating the development of precision oncology therapeutics tailored to the unique molecular landscape of individual tumor microenvironments.