Generation of a comprehensive epigenomic atlas in clear cell renal cell carcinoma informs kidney cancer progression and heritability

SUMMARY

By generating 194 epigenomic and transcriptomic datasets from 57 human tissue samples using H3K27ac and HIF2α chromatin immunoprecipitation sequencing (ChIP-seq), assay for transposase-accessible chromatin using sequencing (ATAC-seq), and RNA sequencing, we provide a comprehensive, integrated characterization of clear cell renal cell carcinoma (ccRCC) across normal, tumor, and metastatic states. Our analyses provide several insights into ccRCC biology. First, we demonstrate significant reprogramming of enhancer and HIF2α cistromes as well as chromatin accessibility during the normal-to-tumor transition, whereas localized and metastatic tumors show minimal epigenomic differences. Second, we show reactivation of kidney-specific developmental pathways driving malignancy. Third, we perform a cistrome-wide association study in ccRCC, validating five established RCC risk loci and identifying six novel loci, including a locus at 12q24 linked to SCARB1 that was functionally validated. These datasets provide new perspectives on the role of developmental pathways in ccRCC tumorigenesis, insights into epigenetic mechanisms of ccRCC heritability, and a comprehensive epigenomic atlas for the research community.

Graphical Abstract

In brief

Abou Alaiwi et al. generate a comprehensive epigenomic atlas of clear cell renal cell carcinoma, revealing that enhancer reprogramming and reactivation of fetal kidney developmental programs drive tumorigenesis. The cistrome-wide association study in ccRCC identifies novel HIF2α-mediated risk loci, with functional validation of a SCARB1 regulatory enhancer.

INTRODUCTION

Epigenomic alterations play a central role in cancer biology and are now considered a hallmark of cancer.¹ Epigenomic reprogramming is increasingly recognized as a driver of cancer development and progression by rewiring gene expression programs.^2–8 Such alterations enable the activation of oncogenic pathways and suppression of tumor-suppressive mechanisms, profoundly influencing tumorigenesis and progression. The mapping of epigenomic landscapes has transformed our understanding of cancer biology, revealing distinct enhancer landscapes and transcriptional dependencies across tumor subtypes.^4,9–17

Enhancer reprogramming has been shown to play a central role in renal cell carcinoma (RCC) by activating tumor-specific transcriptional programs. In clear cell RCC (ccRCC), the loss of the von Hippel-Lindau (VHL) tumor suppressor gene activates hypoxia-inducible transcription factors (TFs) (HIF1α and HIF2α) and drives extensive enhancer dysregulation.^11,12 This enhancer reprogramming leads to the activation of ccRCC-specific pathways through mechanisms involving HIF2α stabilization and increased acetylation at tumor-specific enhancers and super-enhancers.¹⁸ Other RCC subtypes exhibit distinct enhancer landscapes driven by subtype-specific TFs, thus reinforcing the role of epigenetic mechanisms in driving RCC heterogeneity.¹⁹

Cancer cells frequently co-opt developmental transcriptional programs to promote tumor progression, reactivating fetal and embryonic pathways that confer properties such as invasion and metastasis.¹⁵ Epithelial-to-mesenchymal transition (EMT) exemplifies this phenomenon, whereby cancer cells adopt migratory mesenchymal characteristics reminiscent of developmental cell migration.²⁰ Epigenomic approaches provide a rational framework for studying these developmental-cancer parallels, as the chromatin modifications governing cell fate decisions during development are often dysregulated in tumorigenesis.

Finally, genome-wide association studies (GWAS) have highlighted the importance of the non-coding genome. Over 90% of GWAS hits are outside of protein-coding regions and are enriched in regulatory elements.²¹ These variants can lead to altered TF binding and chromatin accessibility and subsequently affect target transcript levels. For example, RCC-specific enhancer regions are enriched for RCC risk variants identified through GWAS.¹⁹ To characterize how variants in regulatory DNA confer disease, our group developed a method termed cistrome-wide association studies (CWAS), which links inherited genetic variation to changes in regulatory element activity, allowing direct tests of whether epigenomic features mediate disease risk. CWAS is a powerful and biologically interpretable paradigm for studying variants that influence traits by affecting transcriptional regulation.²²

In this study, we generated large-scale epigenomic datasets from clinically relevant samples, including normal kidney tissue, localized early-stage RCC tumors, and metastatic ccRCC. Our data offer new insights into the epigenetic mechanisms driving tumorigenesis, underlying ccRCC heritability, and more generally, provide a valuable resource for the research community.

RESULTS

Epigenomic analysis of normal, local tumor, and metastatic ccRCC patient samples reveals chromatin remodeling at key oncogenic enhancers during transition to malignant stage and advanced state

We characterized transcript expression using RNA sequencing (RNA-seq), chromatin accessibility using assay for transposase-accessible chromatin using sequencing (ATAC-seq), active regulatory elements, and HIF2α cistromes using H3K27ac and HIF2α chromatin immunoprecipitation (ChIP-seq), respectively, in 57 samples from 32 patients (Figure 1A; Tables S1A–S1C). These samples included normal kidney (N), local tumor (T), and metastatic tumor tissue. 80% of the overall tumor samples, including 100% of the local tumor samples and 62% of the metastatic tumor samples, were obtained from patients who were treatment-naive at the time of nephrectomy or metastasectomy prior to any systemic therapy. We defined a census of 142,546 chromatin accessible sites, 175,254 active regulatory elements, and 133,925 HIF2α binding sites across 194 datasets (48 HIF2α ChIP-seq, 53 H3K27ac ChIP-seq, 51 ATAC-seq, and 42 RNA-seq; Figure 1B; Table S1A). Unsupervised analyses of the H3K27ac and HIF2α cistromes showed distinct clustering of normal kidney (N) tissue with no clear separation of tumor (T) or metastatic (M) samples; however, primary tumors and metastatic samples derived from the same patients clustered closely together (Figures S1A–S1D). There was a similar distinction across the three states (N, T, and M) using ATAC-seq data (Figures S1E–S1F). Unsupervised and supervised clustering of individual metastatic tumor samples did not show a significant difference based on the site of metastatic disease (not shown). We found no clear association between the site of metastasis, time to metastasis, or disease grade and specific epigenetic programs; however, these analyses were not well powered due to a limited sample size (Table S1C).

Figure 1. Epigenetic landscape of ccRCC across different stages.

(A) Overview of the samples and assays used to assess chromatin accessibility and epigenetic modifications across normal kidney tissue, localized kidney cancer, and metastatic kidney cancer. The tracks represent data from H3K27ac ChIP-seq, HIF2α ChIP-seq, and ATAC-seq.

(B) Summary of assays performed on each sample across three states: normal kidney tissue, localized clear cell renal cell carcinoma (ccRCC), and metastatic ccRCC. Numbers in the boxes indicate the number of metastatic samples for which the assays were performed (if more than one sample).

(C) Principal component analysis (PCA) plot based on H3K27ac ChIP-seq data, illustrating the separation of normal, primary tumor, and metastatic samples. PC1 accounts for 15.4% of the variance, while PC2 explains 9.9%.

(D) PCA plot derived from HIF2α ChIP-seq data, with PC1 and PC2 explaining 17.1% and 9.5% of the variance, respectively.

(E) PCA plot of chromatin accessibility (ATAC-seq) data. PC1 and PC2 capture 13.6% and 10.9% of the variance, respectively.

(F) Venn diagram showing the overlap of tumor upregulated peaks identified by H3K27ac ChIP-seq, HIF2α ChIP-seq, and ATAC-seq across the analyzed samples. The numbers indicate the unique and shared peaks among the three datasets.

To control for inter-individual variation across state transitions (N to T and T to M), we defined a sub-cohort consisting of 8 patients, where each patient had available N, T, and M samples (Table S1D). Unsupervised hierarchical clustering analysis of H3K27ac ChIP-seq data showed epigenomic divergence of T and M samples from their normal counterparts (Figure S1G) as well as no distinct clustering of the 8 paired T and M tumors (16 samples total) with one another (Figure 1C). This was similarly observed in the HIF2α cistrome (Figures 1D and S1H) and the global chromatin accessibility landscape across the three clinical states (N, T, and M) (Figures 1E and S1I).

Next, we identified state-enriched elements. Our analysis identified 9,969 sites with increased H3K27ac in normal samples compared to primary tumor (N-H3K27ac) and 14,240 sites with increased acetylation in tumors compared to normal kidney (T-H3K27ac) (log₂ fold change [FC] > 1 and adjusted p [padj] value < 0.01; Table S2A). T-H3K27ac sites were most enriched for the AP-1 (p = e⁻⁶¹) and SPI1 (p = e⁻²⁰) motifs, whereas N-H3K27ac sites were most enriched for the NR4A2 (p = e⁻⁵¹) and THAP1 (p = e⁻³⁷) motifs. We also identified 4,341 regions with increased acetylation in metastatic samples compared to normal samples (M-H3K27ac), 2,252 (52%), of which overlapped with T-H3K7ac sites. These acetylation changes occurred in a recurrent manner across the eight independent patients, irrespective of the anatomic site of the metastatic lesion. A similar analysis of the HIF2α cistrome identified 10,969 regions with increased HIF2α signal in normal samples compared to tumor (N-HIF2α) and 5,574 regions with increased HIF2α signal in tumors compared to normal (T-HIF2α). We identified 3,961 regions with increased chromatin accessibility in normal samples compared to tumor (N-ATAC), and 12,043 regions with increased accessibility in local tumor (T-ATAC). We observed between 20% and 24% overlap in the H3K27ac and HIF2α cistromes in tumor-enriched peaks and normal-enriched peaks, respectively (Figures 1F and S2A). These overlaps reflect only sites showing differential activity (|log₂FC| > 1, padj < 0.01) rather than all detected peaks. To verify that this observed overlap was a consequence of our differential peak-calling strategy rather than technical variance, we assessed the genome-wide concordance of all peaks. The genome-wide overlap confirmed high concordance between assays (87% of ATAC peaks were acetylated in normal samples and 82% in tumor samples (Figures S2B and S2C).

Analysis of tumor-enriched active regulatory elements (H3K27ac) identified upregulation of pathways related to immune cell regulation and embryonic development (Figure S2D). These pathways implicated CAV1 (Figure 2A), a gene also implicated in fibrosis-mediated kidney transplant rejection and VEGF-R signal transduction.^23,24 Another example included upregulated regulatory element activity in DUXAP8 (Figure 2B), which is a long non-coding RNA upregulated in multiple tumor types, including RCC, associated with worse overall survival, and thought to promote EMT as well as tumor progression at least partly by binding to the polycomb complex 2 and the downstream epigenetic mechanisms.^25,26 Lastly, we identified increased acetylation of a regulatory element distal to the gene NETO2 (Figure 2C), which has been described to promote malignant cellular invasion and metastasis.²⁷ To assess whether inter-patient variability influences our findings, we re-analyzed differential peaks using a paired design that includes the patient identifier as a blocking factor (see STAR Methods). Log₂ fold changes from the paired model showed strong concordance with the original unpaired analysis across all three modalities (Figure S2E).

Figure 2. Chromatin landscape and gene expression profiles ccRCC-enriched H3K27ac peaks.

(A–C) Examples of genomic loci with differential H3K27ac enrichment, including CAV1/2, DUXAP8, and NETO2. These loci show tumor-specific activation of enhancers in ccRCC.

(D) Volcano plot showing the relationship between RNA-seq log2 fold change and H3K27ac ChIP-seq log2 fold change for differentially expressed genes and their associated chromatin marks.

(E) Gene Ontology (GO) enrichment analysis associated with tumor-specific H3K27ac sites. The bar graph shows −log10 (binomial p value) for the top enriched GO terms.

(F) Heatmap showing H3K27ac ChIP-seq signal in different fetal H3K27ac datasets, including embryonic stem cells, endoderm, ectoderm, mesoderm, and specific tissues such as kidney, retina, and mesenchyme, within tumor-specific and normal-specific H3K27ac sites.

(G) Cell-type-specific accessibility scores across normal kidney and ccRCC samples. Cell types are ordered by the highest mean tumor accessibility scores. Red bars indicate mean tumor accessibility for each cell type. Right, a comparison of mean accessibility scores between the normal kidney and ccRCC groups for each cell type. Values are displayed within cells (3 decimal places). SCTB, Syncytiotrophoblasts and bronch, bronchiolar.

We independently performed RNA expression analysis on samples taken from the same patients. Unsupervised hierarchical clustering of RNA-seq data revealed clear separation of normal kidney samples from ccRCC tumors (Figure S3A). Tumor samples showed upregulation of pathways related to immune regulation compared to normal (Table S2B). Differential transcript expression correlated moderately with differential peak intensity across the H3K27ac cistrome (Spearman’s correlation coefficient R = 0.40, p value < 2e⁻¹⁶; Figure 2D) as well as the HIF2α cistrome (Figure S3B). Of the pathways upregulated in tumor samples, we noted significant upregulation of a pathway related to EMT (Figures S3C and S3D).

Mutation-stratified analysis reveals minimal cistrome alterations

To investigate whether genetic drivers influence chromatin landscapes, we leveraged OncoPanel sequencing data available for 15 samples from 7 patients (see STAR Methods for mutation calling criteria). Because patients had multiple samples profiled (local tumor and/or metastases with shared mutations), our analysis included 22 total samples with high-confidence pathogenic mutations in four canonical ccRCC driver genes: VHL (n = 18 samples), SETD2 (n = 7), BAP1 (n = 5), and PBRM1 (n = 3) (Table S2C). Given the high frequency of VHL alterations in ccRCC (>80% of cases) and the likelihood that VHL-wild-type samples harbor alternative mechanisms of VHL pathway inactivation (e.g., promoter hypermethylation), we restricted our analysis to the 3 remaining genes. Unsupervised hierarchical clustering of H3K27ac ChIP-seq, HIF2α ChIP-seq, and ATAC-seq profiles did not reveal distinct clusters based on mutation status for any of these genes (Figures S4A–S4C), indicating that mutational profiles are not the main drivers of global chromatin reorganization patterns.

Supervised differential analysis comparing mutant versus wild-type samples for each gene identified limited cistrome alterations. PBRM1-mutant samples showed increased activity at 333 H3K27ac peaks (padj < 0.01, |log₂FC| >1), 505 HIF2α peaks, and 592 ATAC-seq peaks compared to wild-type samples. While H3K27ac and HIF2α differential peaks did not show significant pathway enrichment, PBRM1-mutant ATAC-seq sites were enriched for genes involved in organic acid and carboxylic acid transport (Figure S4D). BAP1-mutant samples showed increased H3K27ac activity at 381 sites, but these changes were not accompanied by concordant alterations in HIF2α binding or chromatin accessibility. SETD2 mutations were associated with fewer than 300 differential peaks across all three chromatin modalities. These findings demonstrate that despite differences in driver mutations, samples exhibit largely similar chromatin accessibility and histone modification patterns.

ccRCC-specific regulatory elements reactivate embryonic pathways during tumorigenesis

Next, we asked how specific the T-H3K27ac peaks were to ccRCC. To address this question, we investigated this set of peaks against H3K27ac data from 745 samples obtained from 76 distinct healthy adult and fetal tissues and 238 samples derived from 26 distinct cell lines, all of non-kidney origin (STAR Methods). Of 3,260 T-H3K27ac (LFC 1.58, p < 0.001), 427 (13%) were ccRCC-specific (Tables S2D and S2E). A pathway analysis of these ccRCC-specific sites reveals enrichment for pathways related to urogenital development, including nephron, ureteric bud, and renal tubule development (Figure 2E).

Additionally, given the upregulation of kidney developmental pathways during malignant ccRCC transition, we compared ccRCC-enriched and normal kidney-enriched active regulatory elements to embryonic and fetal epigenetic datasets from Epimap.²⁸ T-H3K27ac most resembled fetal kidney tissue and fetal retina.²⁹ On the other hand, N-H3K27ac most resembled mesodermal tissue (Figure 2F). To determine whether these ccRCC tumor-enriched active regulatory elements were activated de novo during malignant transition or were already accessible during the normal state, we looked at the chromatin accessibility of the tumor-enriched active regulatory elements in their paired normal samples, characterized by ATAC-seq. The ccRCC tumor-enriched active regulatory elements were relatively less accessible during the normal state, supporting the shift in accessibility landscape between normal kidney and kidney tumor states and supporting de novo activation of tumor-enriched active regulatory elements (Figure S5).

To further investigate the developmental context of chromatin accessibility changes in ccRCC, we examined fetal cell type-specific accessibility using single-cell ATAC-seq (scATAC-seq) specificity scores in paired normal and tumor samples, derived by Domcke et al.³⁰ (Table S2F). scATAC-seq analysis revealed that fetal accessibility sites showed markedly higher overlap with accessibility in tumor samples compared to normal samples across nearly all cell types, with the highest score observed in metanephric cell type signatures—embryonic precursors of kidney development (Figure 2G). This indicates that ccRCC tumors undergo de novo access to developmental regulatory programs, particularly those from metanephric mesenchyme.

Consistent with prior observations, these data collectively support that ccRCC tumor cells exploit developmental pathways that are physiologically active during embryogenesis to promote their growth and formation. These pathways, which are tightly regulated during normal development, become aberrantly reactivated during tumorigenesis and metastasis.

CWAS identifies genetic determinants of regulatory element activity in RCC

Our epigenomic datasets from 32 distinct individuals allowed us to study the effects of common genetic variants on the regulatory element activity in RCC. After imputing germline single-nucleotide polymorphisms (SNPs) from ChIP-seq reads, we tested how SNP genotypes associate with chromatin accessibility at nearby regulatory elements. We measured two types of genetic effects on chromatin signal: allelic imbalance (AI), defined as differential representation of heterozygous alleles within the same individual, and chromatin quantitative trait loci (cQTLs), where minor allele dosage correlates with ChIP-seq signal across individuals (Figure 3A). Combining these approaches, we identified 11,793 (12%) H3K27ac peaks (out of 100,934 total consensus peaks, STAR Methods) and 5,202 HIF2α peaks (out of 72,924 total consensus peaks, 7%) with significant AI or chromatin QTLs, termed cPeaks (Figures 3B and 3C). There were 2,294 overlapping cPeaks, containing 12,953 SNPs; there was a strong correlation in alternate allele fractions between the H3K27ac and HIF2α datasets for the common SNPs (rho = 0.712, p < 2.2e⁻¹⁶; Figure 3D). To test whether cPeaks specifically capture regulatory elements affecting kidney-specific biological processes, we applied the GREAT analysis on H3K27ac cPeaks, with the H3K27ac consensus peak set used as background.³¹ This analysis showed substantial enrichment for kidney-related Gene Ontology (GO) biological processes, including amino acid transport, renal cell development, as well as immune cell maturation pathways (Figure S6A). In contrast, HIF2α cPeaks showed substantial enrichment for pathways related to antigen processing and T cell function (Figure S6B). These results demonstrate that genetic variants influence regulatory element activity in RCC by disrupting both core kidney functions—including amino acid transport essential for renal filtration and renal cell development critical for nephron architecture—and anti-tumor immune responses through impaired antigen processing and T cell function. This suggests that inherited RCC susceptibility operates via mechanisms affecting kidney development, kidney physiology, and immune surveillance.

Figure 3. CWAS identifies ccRCC cancer risk mediated by genetic variation in HIF2α binding and active regulatory element activity.

(A) Schematic representation of allelic imbalance and chromatin quantitative trait loci (cQTL).

(B and C) Portion of all H3K27ac and HIF2α peaks with evidence of genetic determination, defined as a significant combined test for allelic imbalance and cQTLs with Q < 0.05 (STAR Methods).

(D) Overlap between the H3K27ac cPeaks and HIF2α cPeaks and correlation in alternate allele fractions at overlapping sites.

(E) For all heterozygous SNPs overlapping the indicated motif, the difference in the motif position weight matrix score for alternate versus reference alleles is plotted against the allele fraction observed in HIF2α ChIP-seq reads. The top five motifs inferred de novo from 10,000 randomly selected HIF2α binding peaks are shown. Two-sided p values and Pearson correlation coefficient are indicated. AF, allele fraction.

(F) Boxplot showing the fold enrichment of GWAS risk heritability (versus baseline expected based on proportion of SNPs in peak set) across different cancer types, with a focus on RCC. Error bars represent jackknife SE around the estimates of enrichment.

(G and H) Manhattan plot showing significant genetic associations with kidney cancer for H3K27ac CWAS and HIF2α CWAS. Red lines indicate genome-wide significance thresholds. The associations are annotated with the nearest genes.

(I) Overlap of GWAS risk loci identified using different GWAS and TWAS methods, as labeled in the figure. ccRCC cancer risk loci were defined as genome-wide-significant SNPs ±1 Mb and assessed for overlap with a high-confidence CWAS or TWAS peak.

To determine whether AI sites represent functionally relevant regulatory variants that alter TF binding, we evaluated whether changes in alternate allele fraction at SNP sites correlated with differences in the TF motif scores using position-weighted matrices. This analysis revealed specific and significant disturbance of motifs for the top 4 most enriched TF motifs in HIF2α binding sites: AP1, ETS, HNF1β, IRF, as well as the HIF2α/EPAS1 motif itself but not for NANOG (Figure 3E), which was included as a negative control. These results demonstrate that AI is, in part, due to disrupted TF binding.

We applied stratified linkage disequilibrium score regression³² to calculate the heritability enrichment of cPeaks in multiple cancer types with publicly available summary statistics from GWAS.^33–39 We observed more than 70× enrichment of RCC GWAS SNPs when we intersected H3K27ac, HIF2α, and ATAC cPeaks (Figure 3F), with incremental enrichment as we overlayed cPeaks from additional epigenetic assays. In contrast, RCC cPeaks did not show significant enrichment in lung cancer, esophageal, cervical cancer, breast cancer, bladder cancer, or colorectal cancer GWAS loci (4.8 ± 1.6, p = 0.01) (Figure 3F; Table S3A). These results showed that cPeaks are enriched for kidney-specific regulatory activities and RCC cancer risk heritability.

Next, we applied the CWAS framework to identify individual H3K27ac and HIF2α peaks whose genetic determinants were associated with the risk of RCC.²² Briefly, CWAS involves constructing genetic predictive models of chromatin features based on cis-SNPs and imputing them into GWAS summary statistics.²² We created six types of models, which predicted the chromatin peak intensity from all SNPs within a 25-kb window or the top SNP with the highest correlation based on allelic, linear, or combined signals.⁴⁰ By integrating the summary statistics from a large-scale RCC GWAS meta-analysis with 29,020 cases and 835,670 controls of multiple ancestral backgrounds,¹⁶ CWAS identified 15 H3K27ac loci and 18 HIF2α loci that were significantly associated with kidney cancer risk after Bonferroni correction (Figures 3G and 3H; Tables S3B and S3C), with 6 loci overlapping between datasets for a total of 27 unique risk loci. Four H3K27ac sites and four HIF2α sites were novel and not previously identified in GWAS studies¹⁶. Among those, we identified a CWAS peak in the 12p13.33 locus within FOXM1, a driver of proximal tubule proliferation that has not been previously linked to ccRCC⁴¹ as well as another peak at the TULP3 promoter, a cystogenesis inhibitor during kidney tubule development.⁴² We also identified a CWAS hit in the 9p locus close to IL17RC/IL17RE, coding for IL-17 receptors involved in glomerular disease and response to acute and chronic kidney injury.^43,44

To compare the relative contributions of transcriptome-wide association studies (TWAS) and CWAS in accounting for GWAS risk loci, we compared CWAS associations with results from TWAS (an integrative analysis of eQTL-trait associations) that used reference gene expression data from 45 tissues (4,448 individuals), including benign kidney tissue and kidney cancer. Ten CWAS peaks colocalized with genes identified by TWAS, but nine did not (Figure 3I). CWAS-only (non-TWAS) associations included a 1q43 locus that is ~350 kb upstream of FH (HIF2α CWAS), a 2p21 locus upstream of ZFPL36L2 (H3K27ac CWAS), a known hypoxia regulator,⁴⁵ as well as a 14q32 locus nearby AKT1, and VESTAR, a long-non-coding RNA contributing to VEGF-C stability.⁴⁶ These results show how chromatin QTLs identified from RCC epigenomic profiles provide biological context to risk SNPs and identify chromatin-mediated disease associations in the absence of detectable changes in steady-state gene expression levels, though differences in the statistical power between the two approaches may play a role.

To validate our findings, we prioritized GWAS risk loci where conditioning on the CWAS peak eliminated the association signal (P_{GWAS cond. CWAS peak} = 1), indicating that the chromatin peak fully accounts for the genetic risk at that locus and that functional perturbation of this peak is likely to affect the functional impact of the risk SNP. Among 10 loci with 100% risk signal captured by CWAS, we focused on the 12q24 locus for SCARB1, which contained the SNP with the strongest GWAS association (rs4765623), was identified as a cPeak in both HIF2α and H3K27ac datasets (Figure 4A) and was additionally supported by an independent ccRCC cohort.⁴⁷ SCARB1, encoding the SR-BI receptor, is an HIF target linked to lipid metabolism in RCC. High SR-BI expression correlates with poor outcomes and promotes tumor proliferation via HDL uptake and metabolism.^48–53

Figure 4. Functional validation of the SCARB1 cPeak in the 786-O cell line.

(A) GWAS p values across the locus before and after conditioning on the highlighted cPeak.

(B) Multi-omics tracks (ATAC-seq, H3K27ac ChIP-seq, and HIF2α ChIP-seq) and CRISPRi design. Enhancer-targeting guides (Enh1–Enh4) were placed around rs4765623; promoter-targeting guides (Prom1–Prom2) served as the positive controls. A distal intronic region lacking open chromatin (~2.5 kb away) served as a negative-control region (CTRL1–CTRL4).

(C) Effects of guides on SCARB1 mRNA (RT-qPCR, ΔΔCt normalized to UBC). CTRL: pooled control bar (CTRL1-4). Bars show mean ± SEM for biological replicates.

(D) ATAC-qPCR at the enhancer and promoter after CRISPRi with the most effective enhancer guide. Bars show mean ± SEM for biological replicates.

(E) Knockdown of EPAS1 and ARNT reduces SCARB1 expression (three independent shRNAs each; ACTB control). Bars show mean ± SEM for biological replicates; two-sided Student’s t test with false discovery rate correction. CRISPRi/shRNA assays follow Gusev et al.⁴⁷ with the indicated modifications.

To assess enhancer activity on SCARB1, we performed CRISPR interference (CRISPRi) with dCas9-KRAB in the RCC cell line 786-O (Figure 4B). We targeted the SNP with the strongest GWAS association (rs4765623) with four sgRNAs (Enh1–Enh4) (Table S4), all of which significantly suppressed SCARB1 expression, similar to positive control sgRNAs (Prom1 and Prom2) targeting the SCARB1 promoter (Figure 4C). To determine an appropriate stringent negative control, we performed joint fine-mapping of eQTL and GWAS signals and identified 39 variants in the joint 95% credible set; we then targeted the two most significant SNPs based on joint fine-mapping (rs10846748 and rs10846750) using four sgRNAs (CTRL1–4). These SNPs were located within the same intron and super enhancer, approximately 2.5 kb from Enh1-Enh4 but showed no DNase I hypersensitive site (DHS) activity in the kidney samples in Epimap²⁸ (Figure 4B). CRISPRi targeting this region did not affect SCARB1 expression (Figure 4C). To assess the impact of CRISPRi on chromatin accessibility at the transcriptional start site (TSS) and rs4765623 enhancer site, we performed ATAC-qPCR and revealed a significant decrease in relative chromatin accessibility at the enhancer site after CRISPRi (Figure 4D). To further validate that HIF2α/EPAS1 is the trans-acting factor binding to this cis-element, we used short hairpin RNAs (shRNAs) to suppress HIF2α/EPAS1 mRNA, along with its known co-acting TF, HIF1B/ARNT (Figure 4E). The reduction in SCARB1 expression with three different shRNAs confirmed that HIF2α regulates SCARB1 via this enhancer in conjunction with ARNT, establishing a molecular pathway linking hypoxia signaling to lipid metabolism in ccRCC. This functional validation of a top CWAS pipeline hit demonstrates that our approach can pinpoint specific regulatory elements responsible for GWAS associations, enabling targeted experimental validation of the specific enhancers driving disease susceptibility. Chromatin profiling of three additional cell lines (UMRC6, 769-P, and A498) revealed minimal enhancer activity at this locus (Figure S7A) with relatively lower SCARB1 expression (Figure S7B), precluding validation in alternative cellular contexts.

DISCUSSION

This work creates an atlas of epigenetic regulation governing local tumor and metastasis in ccRCC. By systematically profiling the transcriptome, active regulatory elements, accessible chromatin, and HIF2α cistromes in human samples obtained from different metastatic sites—including multiple samples from individual patients—we have defined the epigenetic landscape of ccRCC metastasis and its divergence from normal and early-stage RCC. While prior efforts have focused on gene mutations and chromosomal complexity driving mechanisms of ccRCC metastasis,⁵⁴ less is understood about epigenetic mechanisms underlying chromatin state transitions. We observed thousands of recurrent regulatory elements that associate with tumorigenesis and progression, independent of tumor stage, grade, or time of metastasis. While the majority of samples were obtained from patients who were treatment-naive, a subset of the metastatic samples had received prior systemic therapy. Given the heterogeneity of treatment regimens in ccRCC (including immune checkpoint inhibitors, VEGF-targeted therapies, and mTOR inhibitors), each of which may differentially affect chromatin landscapes, we cannot exclude the possibility that treatment history may have contributed to some of the observed variability. Despite this, our analysis revealed a notable stability in the global chromatin accessibility landscape across normal, tumor, and metastatic states compared to the more dynamic enhancer reprogramming driving tumorigenesis. This is similar to prior work highlighting that the similarity between the accessible chromatin landscape of different cancer types is defined by the cell of origin.^15,17 Moreover, while previous studies highlighted enrichment of EMT pathways during metastatic transition,^17,55 our data show that ccRCC activates EMT earlier during normal to tumor transition. We also show that the epigenetic landscape of ccRCC during the localized tumor to metastatic transition, irrespective of metastatic sites, was relatively preserved across the accessible chromatin landscape, active enhancer, and HIF2α cistrome.

We show that ccRCC activates fetal kidney developmental programs during normal to tumor development. This is consistent with prior work in prostate cancer, highlighting that epigenetic memory serves as a blueprint that cancer cells recommission during the transition to malignant states and that repurposing genes expressed during prior developmental stages endow tumor cells with metastatic capabilities.¹⁵ It is thought that inciting triggers earlier in time during cellular life, such as inflammatory injuries, allow epigenetic regulators to prime oncogenic pathways that are later utilized by malignant cells during their progression.^56,57

Furthermore, we observed that ccRCC-specific H3K27ac peaks with significant AI or cQTLs (cPeaks) are substantially enriched for inherited ccRCC genetic risk, and that this enrichment increases as we layer HIF2α cPeaks and chromatin accessible cPeaks (~70× enrichment compared to the ccRCC-specific H3K27a cistrome). At HIF2α-bound sites with AI, there is significant disruption of TFs, such as ETS-1, HNF1b, Jun-AP1, and IRF, which are important for ccRCC tumorigenesis.^18,19,58 This supports that risk alleles, through the creation or disruption of disease-specific TF binding sites, preferentially alter downstream gene transcription, leading to increased cancer risk.

We performed the first CWAS in ccRCC by utilizing genetic and epigenetic interactions across different stages of tumor development to validate and discover novel risk associations using a fraction of the samples required for traditional GWAS studies (32 cases versus 29,000 cases).¹⁶ CWAS uncovered many chromatin-kidney cancer risk associations at the eQTL-negative loci, where genetic effects on steady-state gene expression are not observed. By focusing on SNPs with a high prior likelihood of affecting phenotypes, CWAS enabled these discoveries by testing tens of thousands of genetically determined epigenomic features rather than millions of unselected SNPs. CWAS identified risk loci associated with regulatory elements controlling developmental programs, particularly those related to nephron development and tubule formation (FOXM1/TULP3), paralleling our earlier observation that ccRCC recapitulates kidney developmental pathways through epigenetic mechanisms. Importantly, CWAS identified chromatin-mediated disease associations that operate independently of detectable changes in steady-state gene expression levels, as demonstrated by the 9 CWAS-only loci that did not colocalize with TWAS signals, highlighting the power of chromatin-based approaches to uncover regulatory mechanisms missed by expression-based studies. This aligns with findings in prostate cancer, where CWAS has similarly proven effective in identifying genetic determinants of disease.²² Finally, our experimental validation of the SCARB1 locus demonstrated that CWAS-identified peaks represent functionally relevant regulatory elements, as CRISPRi targeting of the rs4765623 enhancer directly suppressed SCARB1 expression and reduced chromatin accessibility. This functional validation confirms that genetic risk variants can modulate ccRCC susceptibility through HIF2α-mediated enhancer activity, establishing a more direct mechanistic link between inherited genetic variation and cancer risk. A limitation of our functional validation approach is that among commonly available ccRCC cell lines, only 786-O exhibited the chromatin profile (H3K27ac, HIF2α binding, and accessibility) at the SCARB1 enhancer that we observed in primary tumors. This highlights the challenge of translating primary tissue chromatin findings to cell line models and underscores the importance of direct profiling in patient samples.

We anticipate this study will serve as an atlas and further enrich the field for genomic and epigenomic investigators. It also serves to shed light on the understanding of the chromatin organization landscape in tumors; how it changes at different cancer stages; and investigating the interplay between chromatin accessibility, genetic mutations, and transcriptional activity are key for advancing cancer research and its clinical applications.

Limitations of the study

Several limitations should be considered when interpreting our findings. First, while our cohort of 57 samples from 32 patients represents one of the largest epigenomic profiling efforts in ccRCC to date, the sample size limited our statistical power for certain analyses, particularly mutation-stratified comparisons and site-specific metastatic patterns. The number of samples from individual metastatic sites was insufficient to definitively characterize site-specific epigenomic programs, and our analyses of associations between epigenetic features and clinical variables (grade, time to metastasis) were underpowered. Second, although 80% of tumor samples were treatment-naive, 38% of metastatic samples had received prior systemic therapy with heterogeneous regimens; we cannot exclude treatment-related effects on chromatin landscapes in this subset. Third, our CWAS approach, while powerful for identifying chromatin-mediated risk associations, relies on cis-regulatory models within 25-kb windows and may miss longer-range enhancer-promoter interactions. Fourth, functional validation was limited to the 786-O cell line, as other ccRCC lines lacked the requisite chromatin profile at the SCARB1 locus, highlighting the challenge of translating primary tissue findings to available cellular models.

RESOURCE AVAILABILITY

Lead contact

Requests for further information and resources should be directed to and will be fulfilled by the lead contact, Matthew L. Freedman (matthew_freedman@dfci.harvard.edu).

Materials availability

The 786-O-dCas9-KRAB stable cell line generated in this study is available from the lead contact upon request. Plasmids for sgRNA (pXPR_BRD003) and shRNA (pLKO.1) constructs are available from Addgene. All oligonucleotide sequences are provided in the supplementary table.

Data and code availability

• All sequencing data generated in this study (RNA-seq, ATAC-seq, H3K27ac ChIP-seq, and HIF2α ChIP-seq) have been deposited at Gene Expression Omnibus (GEO) and are publicly available as of the date of publication. The accession number is GEO: GSE309697.

• This paper does not report original code. All computational analyses were performed using publicly available tools as described in the STAR Methods.

• Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

STAR★METHODS

Detailed methods are provided in the online version of this paper and include the following:

EXPERIMENTAL MODEL AND STUDY PARTICIPANT DETAILS

All tumor samples were obtained at the time of resection at Brigham and Women’s Hospital/Dana Farber Cancer Institute. Normal kidney tissue samples were obtained from adjacent non-tumorous tissue at the time of resection. The samples were then stored fresh and frozen in the Gelb Center biobank at the Dana-Farber Cancer Institute, under protocol #01-130 approved by the Mass General Brigham Institutional Review Board. Informed consent was obtained from all subjects whose samples were included in the study. All specimens were reviewed by pathologists with expertise in genitourinary malignancies (SS, MF, MSA) to confirm the diagnosis, histological subtype, tumor grade, and stage. Clinical data around diagnosis, stage, survival, and site of metastasis are provided in the supplementary data.

The 786-O RCC cell line was obtained from ATCC and maintained in RPMI-1640 (Invitrogen) supplemented with 10% FBS. All media was supplemented with 100 U/mL penicillin, and 100 mg/mL streptomycin (Invitrogen) and all cell lines were cultured at 37°C with 5% CO2. The cell line was authenticated by short tandem repeat profiling and tested negative for mycoplasma (ABMG) in every month.

METHOD DETAILS

Chromatin immunoprecipitation for histone and TF marks

Histone (H3K27ac) and transcription factor (HIF2α) Chromatin immunoprecipitation (ChIP) in primary human tumors was conducted following previously established protocols.^14,19 Briefly, the CryoPREP dry impactor system (Covaris) was used to pulverize 20–30 mg of frozen tissue. The tissue was fixed with 0.2 mM DSG (Thermo Fisher, Catalog number: 20593) for 25 min followed by 1% fixation with FA for 20 min at room temperature and then quenched with 125 mM glycine. Ice-cold lysis buffer (50mM Tris, 10mM, EDTA, 1% SDS with protease inhibitor) was then used to lyse the tissue which was afterward sheared to 300–800 nt using the Covaris E220 sonicator (105-watt peak incident power, 5% duty cycle, 200 cycles/burst). Each sample was added to five volumes of dilution buffer (1% Triton X-100, 2 mM EDTA, 150 mM NaCl, 20 mM Tris HCl pH 8.1), and a fraction do the sample was saved for later DNA preparation. 5–20 micrograms DNA equivalent was then incubated with antibodies (H3K27ac, Diagenode, C15410196; HIF2α, abcam, ab199), along with protein A and protein G beads (Life Technologies) with constant rotation at 4°C overnight. The beads subsequently underwent three wash cycles, each time utilizing different buffers: Low Salt Wash Buffer (0.1% SDS, 1% Triton X-100, 2 mM EDTA, 20 mM Tris-HCl pH 7.5, 150 mM NaCl), High Salt Wash Buffer (0.1% SDS, 1% Triton X-100, 2 mM EDTA, 20 mM Tris-HCl pH 7.5, 500 mM NaCl), and LiCl Wash Buffer (10 mM Tris pH 7.5, 250 mM LiCl, 1% NP-40, 1% Na-Doc, 1 mM EDTA). The beads were afterward rinsed once with TE buffer (pH 8.0). Sequencing libraries were produced from isolated IP sample DNA utilizing the ThruPLEX-FD Prep Kit (Rubicon Genomics). The produced libraries the underwent sequencing using 150-base paired-end reads conducted on an Illumina platform (Novogene).

ATAC methods

We conducted the assay for transposase accessible chromatin in human normal and tumor tissue as previously described.⁴ 20 mg of frozen tissue were pulverized and then resuspended in 1000 μl of HB. 70 μm Flowmi strainer was used to filter nuclei, which were then isolated using iodixanol Density Gradient Centrifugation method and washed with RSB buffer. 50,000 nuclei were counted and reconstituted in 50 μL of transposition mix (2.5 μL transposase (100 nM final), 16.5 μL PBS, 0.5 μL 1% digitonin, 0.5 μL 10% Tween 20, and 5 μL water) by pipetting up and down six times. Transposition reactions were agitated at 37 C for 30 min in a thermomixer set to 1,000 r.p.m. Qiagen columns were then used to isolate DNA. As previously detailed, libraries were amplified and sequenced on an Illumina platform (Novogene) using 150-base paired end reads.⁸³

RNA methods

RNA was extracted from fresh frozen normal and tumor samples using standard Qiagen RNeasy kits according to the manufacturer’s protocol (catalog number 74104). Briefly, 15–20 mg of tissue was pulverized and homogenized in 600 μL buffer RLT. After adding one volume of 70% ethanol, the sample was transferred to a spin column and centrifuged for 15 s at ≥8000 ×g. Per supplier’s protocol, a DNase digestion step was performed prior to washing with buffers RW1 once and RPE twice. RNA was eluted in 20 μL RNase-free water and stored at −80C. RNA-seq libraries were generated from 1 μg total RNA following the manufacturer’s instructions for the Illumina TruSeq Stranded mRNA LT Sample Prep Kit. Barcoded libraries were combined and sequenced on the Illumina HiSeq 2500 generating 50 bp paired end reads. The FASTQ files were processed using the VIPER workflow.⁸⁴ STAR was utilized to perform read alignment to human genome build hg19.⁶⁷ Transcript-level expression data from filtered alignments was assembled though Cufflinks.⁶⁹ DESeq2 was used to perform differential gene expression analysis.⁶⁶

ChIP-seq data analysis

Quality control and chromatin profiling data processing were done using the CHIPS Snakemake pipeline (Supplementary Document).⁶³ Binding sites were identified using MACS v.2.1.1.20160309 with a false discovery rate threshold of 0.01⁶⁴. The union set of sites independently for HIF2α, H3K27ac histone mark and the ATAC-seq data was generated by merging the peaks from all study samples. The read densities for each peak in each sample were computed and subsequently utilized to assess the ChIP-seq and ATAC-seq signals among the various samples. Only the peaks which were present in more than two samples at least in one group (normal kidney, local tumor and metastatic samples as groups) were kept and defined the list of high-confidence peaks for downstream analysis. The determination of sample similarity was conducted through PCA and hierarchical clustering. Spearman correlation was utilized for sample-sample correlations.

Renal cell carcinoma (RCC) cell lines

Functional assays related to SCARB1 were previously described in Gusev et al.⁴⁷. Here, we validate and place those results in the context of our multi-omics atlas.

Lentivirus production and establishment of stable cell lines

For lentivirus production transfer vectors and packaging mix (Cellecta, cat.no.: CPCP-K2A) were delivered by BTX (Harvard Apparatus) electroporation. To generate stable 786-O-dCas9-KRAB expressing cell line the Lenti-dCas9-KRAB-blast (Addgene #89567) was used. Virus transduced cells were maintained for 2 weeks under blasticidin (10 μg/mL, Invivogen) selection. For stable sgRNA and shRNA expression 3 days puromycin (2 μg/mL, Invivogen) selection was applied.

Design and creation of gRNA constructions

SgRNAs were designed using a previously published algorithm 5 5. For CRISPRi experiments, gRNA sequences were synthesized as complementary single stranded oligonucleotides with sticky ends (detailed protocol also available at https://media.addgene.org/data/plasmids/43/43860/43860-attachment_T35tt6ebKxov.pdf). gRNA target sequences are listed in Table S4. Following annealing, oligonucleotides were cloned into pXPR_BRD003 lentiviral gRNA expression vector by Golden Gate cloning. Stbl3 chemically competent cells (Invitrogen) were transformed with the gRNA constructs. Clones were verified by Sanger sequencing (Genewiz). Clones were propagated in LB broth under carbenicillin (100 μg/mL, Gibco) selection over night at 37°C and plasmid DNA was purified using QIAGEN Plasmid Plus Midi kit.

Design and creation of shRNA plasmids

shRNAs were designed using the GPP web portal (https://portals.broadinstitute.org/gpp/public/). Hairpin oligo nucleotides (Table S4) were ordered from Integrated DNA Technologies and after phosphorylation and annealing were cloned into pLKO.1 - TRC shRNA cloning vector (Addgene #10878) into AgeI/EcoRI site. Clones verified by Sanger sequencing (Genewiz). Clones were propagated in LB broth under carbenicillin (100 μg/mL, Gibco) selection over night at 37°C and plasmid DNA was purified using QIAGEN Plasmid Plus Midi kit.

Quantitative RT-PCR

Total RNA was isolated using NucleoSpin (Macherey-Nagel) RNA isolation kit. cDNA was synthesized from 1 μg total RNA using the High-Capacity cDNA Reverse Transcription Kit (Applied Biosystems). Quantitative real-time PCR was performed on Light Cycler 480 II instrument (Roche) using Power SYBRTM Green PCR Master Mix (Applied Biosystems). Oligo nucleotides were synthetized (Integrated DNA Technologies) with standard purification. Primer sequences used for qRT-PCR are listed in Table S4. Relative gene expression was calculated based on the ddCt method. Each sample was measured by three biological and technical replicates; average and standard deviation were calculated. Statistical tests used and p-values are described in the figure legends.

Interrogation of chromatin accessibility (ATAC-qPCR)

Open chromatin regions were revealed by applying Omni-ATAC method 5 8, using 50 000 cells/samples. Briefly, cells were lysed in ATAC-Resuspension Buffer, after washing out lysis reagents, nuclei were collected and resuspended in transposition mixture containing Tn5 transposase enzyme. Following 30 min incubation at 37°C, DNA was extracted by using the Zymo DNA Clean and Concentrator-5 Kit (cat. no.: D4014) and used for amplification. Primers used are listed in Table S4.

QUANTIFICATION AND STATISTICAL ANALYSIS

General statistical framework

All statistical analyses were performed using R version 4.4.2 unless otherwise specified. For differential expression and differential peak analyses, significance was defined as adjusted p-value (FDR) < 0.05 for RNA-seq and adjusted p-value <0.01 with |log₂ fold change| > 1 for ChIP-seq and ATAC-seq. For chromatin QTL analyses, significance was set at adjusted p-value <0.05 after FDR correction. Correlations were assessed using Spearman’s or Pearson’s correlation coefficients as indicated, with two-sided p-values reported. Pathway enrichment analyses report binomial p-values. For experimental validation assays, data are presented as mean ± SEM, with significance determined by two-sided Student’s t test. Sample sizes (n) for sequencing experiments represent independent patient tissue samples and are detailed in Tables S1A–S1D; for cell-based assays, n represents biological replicates. Statistical details including sample sizes, statistical tests, and significance thresholds for computational analyses (Figures 1, 2, and 3) are provided in the Results text and the analysis-specific paragraphs below; details for experimental validation (Figure 4) are provided in the figure legend.

Gene expression (RNA-Seq data) analysis

The quality of the sequencing reads was evaluated using FastQC v0.11.9.⁷⁰ Adapters and low-quality bases were trimmed from the reads using Cutadapt v2.10, with a quality score threshold of 20 and a minimum length of 25 bp retained.⁷¹ Reads passing these criteria were aligned to the GRCh37 reference genome using Kallisto v0.46.0 for pseudoalignment.⁶⁸ Gene expression levels were quantified in transcripts per million (TPM) and raw number of reads.

Differential expression analysis between groups (Tumor vs. Normal samples, Metastasis vs. Tumor) was conducted using the DESeq2 R package v1.36.0.⁶⁶ Genes were considered differentially expressed at an adjusted p-value (False Discovery Rate, FDR) less than 0.05. Principal Component Analysis (PCA) was performed using the prcomp function in R to identify and exclude outlier samples prior to differential expression analysis.

Gene Set Enrichment Analysis (GSEA) was carried out with the clusterProfiler R package v4.4.2 against the Molecular Signatures Database (MSigDB), identifying pathways significantly associated with sample type using a nominal p-value cutoff of 0.05 after FDR correction.⁷² All analyses were performed on an R environment version 4.2.0.

Differential peak analysis

For the differential peak analysis, DEseq2 (v1.36.0) was used to identify peaks that were specific to certain sample groups.⁶⁶ The criteria for identification were an adjusted p-value of less than or equal to 0.01 and a log₂ fold change of at least 1 (fold change of at least 2). The DEseq2 software was utilized to apply the total number of reads in each sample as a size factor. This approach enables the normalization of sequencing depth across the various samples.

Paired differential analysis accounting for patient effects

To account for inter-patient variability, we performed paired differential analysis for samples from patients with matched tissue types (normal kidney, local tumor, and/or metastasis). For H3K27ac ChIP-seq, HIF2α ChIP-seq, and ATAC-seq data, we used DESeq2 with a design formula that includes patient as a blocking factor: ~ patient_id + state, where state represents tissue type (Normal_Kidney, Local_Tumor, or Metastasis). This paired approach controls for baseline differences between patients and provides more precise estimates of tissue state effects. Differential peaks were identified using thresholds of |log₂ fold change| >1 and adjusted p-value <0.01. We compared results from the paired model to our primary unpaired analysis (~ state only) by calculating Pearson correlations of log₂ fold changes between the two approaches.

Mutation-stratified chromatin analysis

A subset of samples (n = 15) underwent targeted-panel sequencing through our institutional OncoPanel platform.⁸⁵ We identified high-confidence pathogenic alterations using the following criteria: (1) deep (2-copy) deletions in tumor suppressor genes, (2) truncating mutations (frameshift or nonsense) in tumor suppressor genes, (3) high-level amplifications in oncogenes, and (4) missense or splice-site mutations classified as pathogenic by at least two of the following prediction algorithms on the SpliceAI platform⁸² (https://spliceailookup.broadinstitute.org/): AlphaMissense, PolyPhen (max), PrimateAI-3D, or SIFT. For splice-site mutations, we additionally required SpliceAI delta scores >0.5. We included all samples (local tumor and metastasis) from patients with OncoPanel data, hypothesizing that shared mutations across multiple sites represent bona fide driver events rather than passenger alterations.

We focused our mutation-stratified analysis on genes altered in at least three samples, yielding four canonical ccRCC driver genes: VHL, PBRM1, BAP1, and SETD2. For each gene, we performed unsupervised hierarchical clustering of H3K27ac ChIP-seq, HIF2α ChIP-seq, and ATAC-seq profiles and supervised differential peak analysis comparing mutant versus wild-type samples using DE-Seq2. Differential peaks were identified using thresholds of |log₂ fold change| >1 and adjusted p-value <0.01.

ENCODE H3K27ac analysis

We conducted a comparative analysis of intervals within the H3K27ac bed file derived from the previously described differential peak analysis of H3K27ac between local tumors and healthy tissue samples. This comparative investigation juxtaposed each interval within the bed file mentioned above against intervals obtained from all available H3K27ac assays within the ENCODE (https://www.encodeproject.org/) database available at the date of our analysis.^60,61 It comprised 745 samples obtained from 76 distinct healthy tissues and 238 samples derived from 26 distinct cell lines that are available in supplement material. We merged the BED files creating one file per tissue containing all the regions and one file per cell line with the information from all related samples. To facilitate this comparative assessment, a Python script was developed, leveraging the ‘count_hits’ function from the ‘pybedtools’ package (https://daler.github.io/pybedtools/).⁷⁹ This script systematically determined, for each interval within the H3K27ac bed file, the extent to which it shared hits with each corresponding interval in the files acquired from the ENCODE repository. The output of this script resulted in a matrix that delineated intervals within the H3k27ac bed file, accompanied by a corresponding list of ENCODE bed files, along with the respective counts of shared hits for each interval-ENCODE sample pairing. To be able to process in parallel the script program was added to a nextflow (https://github.com/nextflow-io/nextflow/tree/master) workflow and ran in a computer cluster.⁶⁴ To refine our analysis and focus on unique intervals, we selectively retained intervals that exhibited a count of zero shared hits. However, in specific analyses, we relaxed this criterion, opting for intervals that demonstrated a minimum of three shared hits. This strategic modification allowed for a more nuanced exploration of interval interactions within the context of H3K27ac, thereby enhancing the robustness and flexibility of our analytical approach.

Enrichment analysis of fetal accessibility sites

To assess the enrichment of fetal accessibility sites within kidney tumor-specific peaks, we performed an overlap analysis with statistical background correction. Fetal specificity scores were derived from a single-cell ATAC-seq analysis.³⁰ Cell type specificity scores for each peak were calculated using Jensen-Shannon divergence, as described in Domcke et al.³⁰ This approach quantifies how specific a peak’s accessibility pattern is to a particular cell type compared to all other cell types in the dataset. The top 10,000 most specific peaks per cell type were selected for enrichment analysis.

For each sample in our matched cohort (n = 8 normal, n = 8 localized ccRCC, n = 8 metastatic ccRCC) we calculated cell type-specific accessibility scores using the fetal kidney specificity signatures. Raw ATAC-seq signal from bigWig files was extracted for genomic regions corresponding to each cell type’s specific peaks using the rtracklayer package in R, v.1.70. For each cell type, we computed a weighted mean accessibility score where individual peak signals were weighted by their respective specificity scores (Jensen-Shannon divergence values). Cell types with fewer than 1,000 peaks passing the minimum specificity threshold (0.0005) were excluded from analysis.

Cell type scores were calculated as: $Score=\Sigma \left({signal}_i\times \left.{weight}_i\right)/\Sigma \left({weight}_i\right.)\right.$, where ${\mathrm{s}\mathrm{i}\mathrm{g}\mathrm{n}\mathrm{a}\mathrm{l}}_{\mathrm{i}}$ represents the mean ATAC-seq signal for peak $i$ and ${\mathrm{weight}}_{\mathrm{i}}$ represents the specificity score for that peak. Z score normalization was performed across samples for each cell type to enable cross-cell type comparisons. Cell types were ordered by mean tumor accessibility score for visualization. Heatmaps were generated using the ComplexHeatmap package, with raw scores displayed alongside Z score patterns to provide both relative and absolute measures.⁸¹

SNP imputation

We aggregated individual off-target reads of epigenomic sequencing (ChIP-seq and ATAC-seq) and used merged BAM files to impute SNP genotypes using STITCH at approximately 5.5 million common single nucleotide polymorphisms (SNPs) with a minor allele frequency of ≥5% in the Haplotype Reference Consortium (HRC) v1.191.^22,73 Haplotype phasing is performed using EAGLE implemented on the TOPMED imputation server (https://imputation.biodatacatalyst.nhlbi.nih.gov) with mixed-population TOPMED reference panel.^62,74

Consensus peak calling

Consensus peak sets used for subsequent analyses were created by dividing the genome into 50bp windows and including regions with peaks in >5% samples. Windows were buffered by 100bp and merged to create a set of 72,924 HIF2α peaks and 100,934 H3K27ac peaks.

Allelic imbalance analysis

H3K27ac ChIP and HIF2α ChIP data were analyzed for allelic imbalance using stratAS (https://github.com/gusevlab/stratAS). SNPs were initially run through WASP.⁷⁵ to allow for unbiased read mapping. ChIP reads are then run through gatk ASEReadCounter (GATK v4.0.1.2) to calculate allele-specific read counts for patients who were heterozygous for each tested SNP.⁷⁶ stratAS also leveraged linkage disequilibrium (LD) to increase the power by accounting for all SNPs in the same LD block. As described earlier,⁴⁷ reads were modeled for each individual $i$ as drawn from a beta-binomial distribution $\left\{{\mathit{X}}_{\mathit{A}\mathit{L}\mathit{T},\mathit{i}},X_{REF,i}\right\}\sim \mathrm{B}\mathrm{e}\mathrm{t}\mathrm{a}\mathrm{B}\mathrm{i}\mathrm{n}\mathrm{o}\mathrm{m}\mathrm{i}\mathrm{a}\mathrm{l}\left({\mathit{\pi }}_{\mathit{X}},{\mathit{\sigma }}_{\mathit{i}}\right)$, where ${\mathit{\pi }}_{\mathit{X}}$ is the allelic fraction for variant X across all individuals and ${\mathit{\sigma }}_{\mathit{i}}$ is the local overdispersion parameter for individual $i$ and the null hypothesis tested against being ${\mathit{\pi }}_{\mathit{X}}=0.5$. Allele fractions were also compared across different cell states (normal kidney tissue, primary tumors, and metastatic specimens) by testing the model ${\mathit{\pi }}_{\mathit{X}\text{,}\mathit{state}1}={\mathit{\pi }}_{\mathit{X}\text{,}\mathit{state}2}$ against the model ${\mathit{\pi }}_{\mathit{X}\text{,}\mathit{state}1}\ne {\mathit{\pi }}_{\mathit{X}\text{,}\mathit{state}2}$ using a one-degree of freedom likelihood ratio test.

Chromatin quantitative trait loci (cQTL) detection

Chromatin accessibility was quantified by calculating Reads Per Kilobase of transcript per Million mapped reads (RPKM) at HIF-2a and H3K27ac consensus peaks for each sample using QTLtools v1.287 quan (parameters: –filter-mismatch 5 –filter-mismatch-total 5 –filter-mappingquality 30).⁷⁷ Peaks with summed RPKM <10 across all samples were excluded from analysis. To account for potential confounders, a covariate matrix was generated by QTLtools v1.287 pca (–scale –center) and the first 6 principal components were included in the cis-QTL model. Permutation-based p-values for SNP-peak associations within 1Mb were computed by QTLtools v1.287 cis (–normal –permute 1000) after regressing out principal components. p-values were adjusted by FDR correction and SNP-peak pairs with q < 0.05 were considered significant cQTLs for downstream analysis.

GWAS enrichment analysis

Genetically determined peaks, i.e., peaks that contained either an allelicly-imbalanced SNPs or a cQTL were denoted ‘cPeaks’. Using stratified LD-score regression,³² we estimated partitioned heritability of different traits using summary statistics output of GWAS studies. We then calculated enrichment of trait heritability for cPeaks.

CWAS analysis

Cistrome-wide association studies employ statistical methods to identify epigenomic features that are genetically associated with a trait. Normalized read abundance and allele-specific reads at heterozygous SNPs are modeled as a function of cis-SNP genotypes to estimate the genetic determinants of peak intensity as previously described.^40,47 Allelic imbalance analysis was used to capture heterozygous SNPs that exhibited differential allelic representation in ChIP-seq reads and chromatin QTLs captured loci where the dosage of a minor allele correlated with ChIP-seq signal.^47,77 The genetic association between GWAS risk and AI/cQTL activity was modeled using FUSION after accounting for linkage disequilibrium. Significance level was set at p < 0.05 after Bonferroni correction. CWAS was run independently for H3K27ac ChIP-seq and HIF2α ChIP-seq data.

Supplementary Material

SUPPLEMENTAL INFORMATION

Supplemental information can be found online at https://doi.org/10.1016/j.celrep.2026.116968.

KEY RESOURCES TABLE.

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
Anti-H3K27ac	Diagenode	Cat# C15410196; RRID: AB_2637079
Anti-HIF2α (EPAS1)	Abcam	Cat# ab199; RRID: AB_301687
Bacterial and virus strains
Stbl3 chemically competent cells	Invitrogen	Cat# C737303
Biological samples
Human ccRCC tumor tissue	Brigham and Women's Hospital/DFCI	N/A
Human normal kidney tissue	Brigham and Women's Hospital/DFCI	N/A
Chemicals, peptides, and recombinant proteins
DSG (disuccinimidyl glutarate)	Thermo Fisher	Cat# 20593
Formaldehyde	Sigma-Aldrich	Cat# F8775
Blasticidin	Invivogen	Cat# ant-bl-1
Puromycin	Invivogen	Cat# ant-pr-1
Critical commercial assays
ThruPLEX-FD Prep Kit	Rubicon Genomics	Cat# R400428
TruSeq Stranded mRNA LT Sample Prep Kit	Illumina	Cat# 20020594
RNeasy Mini Kit	Qiagen	Cat# 74104
High-Capacity cDNA Reverse Transcription Kit	Applied Biosystems	Cat# 4368814
Power SYBR Green PCR Master Mix	Applied Biosystems	Cat# 4367659
NucleoSpin RNA isolation kit	Macherey-Nagel	Cat# 740955
Zymo DNA Clean and Concentrator-5 Kit	Zymo Research	Cat# D4014
Deposited data
Raw and processed sequencing data	This paper	GEO: GSE309697
RCC GWAS summary statistics	Purdue et al.16	Planchard et al.59
ENCODE H3K27ac data	ENCODE Project60,61	https://www.encodeproject.org/
Fetal single-cell ATAC-seq	Domcke et al.30	GEO: GSE149683
TOPMED Imputation Server	Das et al.62	https://imputation.biodatacatalyst.nhlbi.nih.gov
Experimental models: Cell lines
786-0	ATCC	Cat# CRL-1932; RRID: CVCL_1051
786-0-dCas9-KRAB	This paper	N/A
UMRC6	Dr. W. Kaelin laboratory	N/A
769-P	ATCC	Cat# CRL-1933; RRID: CVCL_1050
A498	ATCC	Cat# HTB-44; RRID: CVCL_1056
Oligonucleotides
sgRNA and shRNA sequences	This paper	Table S4
qRT-PCR primers	This paper	Table S4
ATAC-qPCR primers	This paper	Table S4
Recombinant DNA
Lenti-dCas9-KRAB-blast	Addgene	Cat# 89567; RRID:Addgene_89567
pXPR_BRD003	Broad Institute	N/A
pLKO.1 - TRC cloning vector	Addgene	Cat# 10878; RRID:Addgene_10878
Software and algorithms
CHIPS Snakemake pipeline	Taing et al.63	https://f1000research.com/articles/10-517
Nextflow	Di Tomasso et al.64	https://github.com/nextflow-io/nextflow
MACS	Zhang et al.65	v2.1.1.20160309
DESeq2	Love et al.66	v1.36.0
STAR	Dobin et al.67	v2.7.9a
Kallisto	Bray et al.68	v0.46.0
Cufflinks	Trapnell et al.69	v2.2.1
FastQC	Babraham Bioinformatics; Trivedi70	v0.11.9
Cutadapt	Martin71	v2.10
clusterProfiler	Wu et al.72	v4.4.2
STITCH	Davies et al.73	v1.6.6
EAGLE	Loh et al.74	v2.4.1
stratAS	Gusev et al.47	https://github.com/gusevlab/stratAS
WASP	van de Geijn et al.75	v0.3.4
GATK ASEReadCounter	Broad Institute; DePristo et al.76	v4.0.1.2
QTLtools	Delaneau et al.77	v1.287
FUSION	Gusev et al.78	http://gusevlab.org/projects/fusion/
LDSC	Bulik-Sullivan et al.32	v1.0.1
GREAT	McLean et al.31	v4.0.4
pybedtools	Dale et al.79	v0.9.0
Rtracklayer	Lawrence et al.80	v1.70
ComplexHeatmap	Gu et al.81	v2.12.0
SpliceAI	Jaganathan et al.82	https://spliceailookup.broadinstitute.org
R	R Core Team	v4.4.2

Highlights.

• Enhancer reprogramming drives clear cell renal cell carcinoma tumorigenesis

• Modest epigenomic remodeling occurs during metastatic progression

• Clear cell renal cell carcinoma reactivates fetal kidney developmental programs

• Cistrome-wide association study nominates HIF2α-associated enhancer at a SCARB1 locus