Epigenomic subtypes and prognostic methylation signatures in clear cell renal cell carcinoma

Abstract

Background

Clear cell renal cell carcinoma (ccRCC) is characterized by marked epigenetic dysregulation, contributing to aberrant gene expression and tumor progression. To expand current knowledge on genome-wide methylation patterns in ccRCC, we performed Methylated DNA Immunoprecipitation sequencing (MeDIP-seq) on tumor samples from 116 ccRCC patients and 34 adjacent normal renal tissues. We identified differentially methylated regions (DMRs) and integrated these findings with genome-wide copy number alterations.

Results

ccRCC tumors exhibited global hypomethylation combined with focal hypermethylation, particularly within intergenic and repetitive genomic regions. Unsupervised clustering revealed three distinct subtypes, including one subtype characterized by chromosomal instability, synchronous metastasis, and poor survival outcomes. We developed a novel methylation score (MethScore), based on hypermethylated regions, which strongly correlated with disease stage and independently predicted overall survival beyond age and clinical stage. Notably, high methylation levels of specific DMRs were linked to recurrence and advanced disease. Validation with TCGA Kidney Renal Clear Cell Carcinoma methylation data supported the prognostic value of these hypermethylated regions. Additionally, MeDIP-seq accurately detected critical copy number alterations, particularly chromosome 3p loss, which were found in 88% of tumors.

Conclusions

Together, our findings demonstrate that genome-wide methylation profiling can resolve clinically relevant ccRCC subtypes, uncover novel biomarkers of disease aggressiveness, and improve patient risk stratification beyond current clinical models.

Graphical Abstract

Supplementary Information

The online version contains supplementary material available at 10.1186/s13148-026-02055-7.

Introduction

Renal cell carcinoma (RCC) accounts for around 2% of all malignancies diagnosed worldwide [1, 2]. Europe shows among the highest age-standardized incidence rates globally and has the highest mortality from RCC [1, 2]. RCC is a highly heterogeneous disease characterized by complex molecular mechanisms, with clear cell RCC (ccRCC) being the most prevalent and lethal subtype, accounting for 75–80% of cases [3]. Despite advances in treatment, ccRCC continues to pose a significant clinical challenge. Current risk stratification models, such as the Leibovich risk score [4] and IMDC [5] systems, rely on clinical and histological factors but often fail to accurately predict outcomes (reviewed in [6]). This underscores the need for novel biomarkers to enhance prognosis and inform patient management.

ccRCC typically arises from a set of homogeneous genetic events, with the inactivation of VHL and one of the chromatin regulators PBRM1, SETD2, or BAP1 allowing the initiation of oncogenesis [7]. This underscores the importance of epigenetic dysregulation in ccRCC development. As a central part of epigenetic regulation, changes in DNA methylation patterns have been linked to ccRCC and associated with aberrant expression of factors influencing cell proliferation, apoptosis, and metastasis [8] (reviewed in [9]). In ccRCC tumors, VHL promoter hypermethylation functions as an alternative, mutually exclusive mechanism to VHL mutation [7], and several other tumor suppressor genes are likewise more often hypermethylated than mutated in RCC [7, 10–12].

Methylation patterns have been shown to provide both prognostic and predictive value [13–16]. Increased promoter methylation correlates with higher stage and tumor grade in ccRCC tumors [7], and methylation of specific genomic sites has been shown to predict immunotherapy resistance [15]. However, none of these markers have yet been validated or clinically integrated.

Similarly, copy number alterations (CNAs) constitute a promising avenue for biomarker development, with recurrent loss observed on chromosome 3p and losses of chromosomes 9p and 14q being associated with poor prognosis and increased metastatic potential, highlighting their relevance as biomarkers for recurrence and treatment response [17, 18]. Since CNAs often co-occur with epigenetic changes, integrating both alterations may improve risk stratification and treatment prediction.

While CNAs are well-studied in ccRCC and previous epigenetic studies have provided valuable insights, the latter have predominantly relied on 1) cell lines, which do not resemble ccRCC tumors [19], and 2) array-based methods, which, although powerful, have limitations in terms of coverage and resolution, potentially overlooking important methylation events across the genome [20]. Overall, methylation-specific ccRCC studies are limited, and more than 50% of studies investigating prognostic methylation markers are based on the TCGA Illumina Infinium HumanMethylation450 BeadChip array (HM450K) data set (reviewed in [16]). Here, we employ the Methylated DNA Immunoprecipitation sequencing method [21] (MeDIP-seq) on 150 ccRCC patient tumor and normal samples to characterize the ccRCC methylation landscape and examine CNAs across all disease stages.

Methods

Design and study cohort

This study included 116 ccRCC patients from the Bio- and Genome Bank, Denmark (RBGB), recruited to the Danish Cancer Biobank at three centers between April 2011 and October 2020 (Table 1). To facilitate the investigation of molecular features associated with disease progression, the cohort was deliberately enriched for patients who experienced relapse or developed metastatic disease. Accordingly, patients were categorized into three predefined groups: (i) no recurrence, (ii) recurrence after nephrectomy, and (iii) synchronous metastatic disease at diagnosis. Additionally, inclusion criteria required the availability of both tumor and matched germline samples, ensuring the feasibility of comprehensive methylation and genomic analyses. Patients were followed until death or, if still alive, until the last follow-up date (October 11th, 2023), with a median follow-up time of 8 years (range: 3.2–11.5 years). Fresh frozen tumor samples were available for all patients following surgical resection, with matched adjacent normal tissue from 32 cases and 2 additional normal tissues from ccRCC patients not otherwise included in this study (Supplementary Data S1). This study was approved by the Danish National Center for Ethics (j. no. 2208138).

Table 1.

Summary of patient demographics and clinical characteristics

Characteristic	N = 1161
Sex
Female	41 (35%)
Male	75 (65%)
Age	66 (58, 72)
Stage
I	23 (20%)
II	22 (19%)
III	33 (28%)
IV	38 (33%)
Leibovich
Low	15 (13%)
Intermediate	39 (34%)
High	24 (21%)
Not relevant (metastatic)	38 (33%)
Status
No recurrence	43 (37%)
Recurrence	35 (30%)
Synchronous metastatic	38 (33%)
Follow-up alive, year	7.98 (6.30, 9.67)
Follow-up dead, year	3.41 (1.88, 5.08)

¹n (%); Median (Q1, Q3)

Sample processing

DNA from fresh frozen kidney tissue (normal and tumor) was precipitated using isopropanol. DNA was quantified using Qubit (dsDNA Broad Range, Thermofisher) and sheared for 80s using a Covaris E220 Evolution ultrasonicator (Covaris) system to an average size of 298 bp, evaluated by TapeStation analysis (4200 system, HSD1000, Agilent Technologies).

Methylated DNA immunoprecipitation (MeDIP) and sequencing

MeDIP was performed according to the published protocol by Shen et al. [21] with minor adjustments (see Supplementary Methods). 10% of each sample was stored as an input control (IC sample), whereas the remainder was subjected to immunoprecipitation using 5-mC antibodies (IP sample). Pooled IC and IP libraries were paired-end sequenced on an Illumina NovaSeq X Plus instrument (25B flow cell, 2 × 150 bp), with a median of 14.9 million reads (median of 0.58X coverage depth) for IC libraries and a median of 48.3 million reads (median of 1.58X coverage depth) for IP libraries (see Supplementary Data S1 for quality control measurements).

Copy number alteration analysis

Low-pass whole-genome sequencing profiles of IC libraries were used for copy number alteration (CNA)-based tumor purity estimation by ichorCNA [22]. IchorCNA estimates large-scale CNA segments in 1 Mb bins and tumor fractions using a hidden Markov model while accounting for tumor ploidy. In addition, CNA estimates (log2 ratios) from ichorCNA were used to normalize read counts of IP libraries. AbsCN-seq [23] was used as a comparison tool for estimating tumor purity and was selected for downstream analyses. Chromosome 3p loss was investigated in 1 Mb genomic bins from position 0 to 90,900,000 and defined as median log2 ratio < − 0.1.

Weighted genome instability index and whole genome doubling

The weighted genome instability index (wGII) was calculated using the segmented copy number data [24]. Genome Doubling (GD) was called for a sample if more than half of the genome had a major allele copy number (nA) over 2. The nA was estimated from the ichorCNA corrected copy number calls and corrected copy number (cn) as nA = cn—1, unless the segment call was hemizygous deletion (HETD), in which case nA = 1. The minor allele copy number (nB) was set to 1 unless the segment call was HETD, in which case nB = 0.

Methylation landscape analysis

Quantitative Sequence Enrichment Analysis (QSEA) [25] (v1.30.0) was performed on enrichment profiles of IP libraries to quantify methylation signals in non-overlapping genome-wide bins of 300 bp, normalized to ichorCNA reported CNAs (nrpm: normalized reads per million mapped reads) and accounting for sequencing depth, library size, background noise, and CpG density. Methylation levels (beta) from TCGA Renal HM450K array data [26] (n = 577) were filtered for highly methylated regions averaged in 300 bp windows (mean beta > 0.85) and used for calibration. The mean difference in nrpm values between tumor and normal samples (tumor mean—normal mean) was visualized genome-wide by a rolling mean of 25 regions and combined neighboring regions if below 0.03 difference. FDR-adjusted p-values for all regions compared between tumor vs normal samples were illustrated in Manhattan plots across chromosomes using the ggmanh R package (v1.10.0). Hypermethylated regions were defined as log2 fold change (log2FC) > 0 and hypomethylated regions as log2FC < 0. Raw methylation nrpm values were plotted for individual chromosomes using the karyoploteR R package (v1.32.0). Chromosome 3p methylation was investigated in 300 bp genomic bins for regions being differentially methylated, ranging from position 219,601 to 88,395,600. Methylation was defined as median nrpm > 0.1 across regions. Regions were annotated using Bioconductor’s AnnotationHub (v3.12.0) and annotatr (v1.30.0) packages by lift over from hg19 to hg38. A 300 bp window was considered a repeat if it overlapped a repeat-annotated region retrieved via AnnotationHub/annotatr. Heatmap was created using the ComplexHeatmap R package (v2.22.0) using the Euclidean Distance method and Ward.D2 clustering.

Differential methylation analysis

Differentially methylated regions (DMRs) were calculated between tumor samples and adjacent normal samples by fitting 300 bp windows to a generalized linear model incorporated in the QSEA [25] (v1.30.0) framework. DMRs were defined based on statistical significance (FDR-adjusted P < 0.05 and log2FC > 1.3|< − 1.3). Principal component analysis was performed using the ggfortify R package (v0.4.17). For each sample, a methylation score (MethScore) was defined as the sum nrpm across all tumor-enriched hypermethylated DMRs (HyperDMRs; n = 2 424) identified in the tumor versus matched normal DMR analysis, divided by the total number of regions (n = 2 424). The genomic coordinates and annotations of these 2 424 HyperDMRs, which together constitute the full input set for MethScore, are provided in Supplementary Data S3. To assess potential cofounding by frequent chromosome 3 alterations, MethScore was recalculated after excluding all the HyperDMRs in MethScore located on chromosome 3 (n = 176/2424). Since DMRs were not normalized to tumor purity, we included tumor purity as a covariate in regression analyses. Gene set enrichment analysis was performed on GO (org.HS.eg.db, BP ont) and KEGG (hsa) gene lists using the clusterProfiler R package (v4.12.6). DMR analysis was also performed using QSEA between advanced (n = 38) and localized (n = 78) patients and between recurrence/metastatic (n = 73) and no-recurrence (n = 43) patients with DMRs defined by P < 0.05 and log2FC > 1|< − 1.

External validation

TCGA Kidney Renal Clear Cell Carcinoma (KIRC) HM450K array data [7] on ccRCC patients with matched clinical data (n = 309) were averaged in 300 bp windows across hg38 (BSgenome.Hsapiens.UCSC.hg38) and used for validation, as adjacent CpG sites tend to be co-methylated. Similarity between TCGA KIRC HM450K data and our data was calculated as 1-abs(rank(nrpm_mean/nrow)—rank(beta_mean/nrow)) for each region. Regions annotated to promoter or 1to5kb regions and shores, islands, or shelves were compared, with the highest log2FC chosen if multiple regions were available per gene.

Statistical considerations

All statistical analyses were conducted using R version 4.4.1 and results visualized using ggplot2 (v3.5.2) unless otherwise stated. Fisher’s exact test was used for analyzing categorical data, and Wilcoxon rank-sum test was used for comparing group medians. Paired t-test was applied to assess mean differences between paired observations. Statistical comparisons, including display, were conducted using ggpubr (v0.6.0), and summary tables were made using gtsummary (v2.2.0). Survival rates were analyzed using Kaplan–Meier curves by the Survival (v3.8–3) and Survminer R packages (v0.5.0), with comparisons calculated by the log-rank test. An overall median nrpm value was calculated across all patients and regions. Patients were then classified into high or low methylation groups based on whether their individual median nrpm (across all regions) was above or below this overall median. Cox proportional hazard regression calculated hazard ratios and 95% confidence intervals using the survival R package (v3.8-3) and were visualized using the forestmodel R package (v0.6.2). Outcome was defined as time from diagnosis to death of any cause (overall survival). Patients alive at the last follow-up were censored at that time point. Statistical significance was considered upon P < 0.05.

Results

Clinical characteristics and copy number alterations of ccRCC

To characterize the CNA and methylation landscape of ccRCC, 116 ccRCC tumor samples and 34 matched adjacent normal samples were processed with the MeDIP-seq method [21] (Table 1). The median age at diagnosis was 66 years, with 65% males. Patients were distributed across clinical stages and Leibovich risk scores (Table 1). At diagnosis, 38 patients (33%) had synchronous metastases, whereas 35 patients (30%) experienced a recurrence in the follow-up period. Patients were followed until death or for a median of 8 years (IQR: 6.3–9.7 years, Table 1).

Using bulk low-pass whole genome sequencing data, ichorCNA [22] and AbsCN-seq [23] tools were run to estimate the fraction of tumor-derived DNA in the samples based on CNAs. Tumor samples were estimated to have a mean tumor purity of 40.8% (6.7–84.8%) by ichorCNA and 45.9% (6.8–100.0%) by AbsCN-seq (R² = 0.56, Supplementary Data S2, Supplementary Fig. S1). Tumor purity from AbsCN-seq was used as a covariate in downstream analyses due to a stronger correlation to pathology tumor purity estimates (R² = 0.10 vs R² = 0.07, Supplementary Data S2). Four of the 116 tumor samples were whole-genome doubled (one Stage I, one Stage II, and two Stage IV). The weighted genome instability index (wGII) had a median value of 0.2 for the tumor samples, with increasing wGII at the higher stage (Supplementary Fig. S2).

Global methylation in ccRCC reveals extensive hypomethylation and regional hypermethylation

We examined global methylation changes in ccRCC tumors compared to the matched adjacent normal tissue by analyzing methylated regions in 300 bp bins. These showed consistent variance across samples in a principal component analysis, indicating no batch effects (Supplementary Fig. S3). Methylation estimates (nrpm) were normalized for local copy number and sequencing depth and adjusted for library size, background signal, and CpG density, revealing global hypomethylation in ccRCC tumors accompanied by focal hypermethylation at specific genomic positions (Fig. 1a). Pronounced hypermethylation was detected on chromosome arms 3p, 4p, 6p, 7p, 8p, 9p, 13q, 14q, 17p, 18p, 18q, 19p, 20q, and 21q (Fig. 1a).

Fig. 1.

Global methylation differences in clear cell renal cell carcinoma (ccRCC). a The mean difference in nrpm values between tumor and normal samples is shown, calculated as a rolling mean of 25 regions. Neighboring regions were combined if their values were within 0.03 of each other. Red: Hypermethylated regions. Blue: Hypomethylated regions. Yellow: Centromeric positions. b, c Manhattan plots of all methylated regions. FDR adjusted P from Tumor vs Normal analysis of all regions across chromosomes. Cytobands for the most significant regions above the threshold (adjusted P < 1 × 10^− 20) are shown (the most significant region is annotated if multiple regions are located in the same cytoband). Peak is defined as above the threshold. b Hypermethylated regions; log2FC > 0 (3,594,282 regions). c Hypomethylated regions; log2FC < 0 (4,686,739 regions). d Methylation of chromosomes 3, 5, 6, and 7. Top 5000 methylated regions (adjusted P < 0.05 and top absolute log2FC) are shown with raw nrpm values on y-axis. Each chromosome is represented as an ideogram with cytobands defining CG-rich/gene-rich regions (light), AT-rich/gene-poor (dark) regions, and centromeric regions (yellow). Red: Tumor. Blue: Normal. CGI: Positions of CpG Islands. Genes from [16] annotated. Nrpm: normalized reads per million mapped reads

Hypermethylated (tumor-enriched; Fig. 1b) and hypomethylated (normal-enriched; Fig. 1c) regions were distributed across the entire genome. Hypermethylated regions formed distinct significant peaks on chromosomes including 3, 5, 6, and 7 (Fig. 1b + d), whereas hypomethylation was more diffusely spread, with isolated regions reaching significance (Fig. 1c). The most significant hypermethylated region on chromosome 3 was 3p25.3, encompassing the hallmark ccRCC gene VHL, frequently lost in ccRCC (Supplementary Data S4) [7]. Chromosome 5q had hypermethylation at 5q32, covering the CSF1R gene, prognostic for poor cancer-specific survival in ccRCC upon high expression [27, 28]. On chromosome 7, hypermethylation was detected at 7p14.3 and 7q11.22, which have been found amplified in ccRCC [29, 30]. Furthermore, hypermethylation at 9p21.3 includes the well-known tumor suppressor gene CDKN2A, implicated in several cancer types, including ccRCC [31]. Overall, the elevated methylation was concentrated at CpG islands and aligned with key ccRCC-associated genes, suggesting a potential functional role in tumor biology (Fig. 1d). Together, these data underscore the tumor-specific methylation patterns in key chromosomal regions associated with ccRCC progression and oncogenesis.

Chromosome 3p methylation correlates with recurrence and metastatic disease

Patients who experienced recurrence or had metastatic disease at diagnosis exhibited an enrichment of hypermethylation on chromosome 3p compared to patients without recurrence (Table 2). At the same time, no differences were observed for chromosome 3p loss or combined chromosome 3p loss and methylation across disease groups (Table 2). This supports that genomic loss of chromosome 3p is a hallmark of ccRCC (88% of cases) [7], whereas methylation in this region is associated with poor prognosis, and may represent a biallelic hit in tumor suppressor gene inactivation.

Table 2.

Chromosome 3p characteristics are linked to recurrence and metastatic disease

Characteristic	Overall N = 1161	No recurrence N = 431	Recurrence N = 351	Metastatic N = 381	p-value2
Loss	102 (88%)	39 (91%)	32 (91%)	31 (82%)	0.4
Methylation	81 (70%)	24 (56%)	27 (77%)	30 (79%)	0.041
Both	72 (62%)	23 (53%)	24 (69%)	25 (66%)	0.3

Chromosome 3p (chr3p) regions included in differentially methylated regions (n = 2717), ranging from chr3p position 219,601 to 88,395,600 for 300 bp methylated regions and the entire chr3p region, ranging from chr3p position 0 to 90,900,000 for 1 Mb genomic bins. Loss is defined as median log2ratio < − 0.1 across chr3p. Methylation is defined as median nrpm > 0.1 across chr3p regions. Metastatic = synchronous metastatic. Both = loss and methylation at chr3p regions

¹n (%)

²Fisher's exact test; Pearson's Chi-squared test

Validation of ccRCC methylated regions in existing literature and beyond

To further validate previously published literature, we analyzed CpG-rich promoter and regulatory hypermethylated regions from our MeDIP-seq data and compared them to TCGA Kidney Renal Clear Cell Carcinoma (KIRC) HM450K data [7] (Supplementary Fig. S4). Some of the most concordant genes between the two datasets were AP1G1, PIP4K2B, MFGE8, and PRKRA, which are prognostic markers in TCGA KIRC associated with shorter overall survival upon low expression [32], as well as MIR4534, which has been found elevated in advanced RCC [33]. Notably, the well-associated, prognostic ccRCC-associated marker genes, BAP1, SETD2, and PBRM1, as well as TP53 target genes, also demonstrated strong similarity in methylation patterns between TCGA KIRC data and our MeDIP-seq data, with BAP1 being highly methylated in both data sets (Supplementary Fig. S4). Furthermore, the MeDIP-seq data validates methylation in previously reported genes, including SFRP1, HOXA9, DPP6, THBS4, SLC17A1, and FOXO3 [16].

Our genome-wide approach also reveals methylation in repetitive and transposable elements (~ 96% are covered), which are poorly covered by array-based methods (~ 2%) [34]—only 1.2% of repetitive elements found in the MeDIP-seq data were covered in the TCGA KIRC data. Notably, hypermethylation at regions linked to repetitive elements showed a correlation with advanced stage and metastatic disease (Supplementary Fig. S5), highlighting their potential clinical relevance and shedding light on the epigenomic features of ccRCC that were not previously captured by array-based methods.

Methylation defines clinically relevant subtypes

Hierarchical clustering of the top 5000 most differentially methylated regions between tumor and normal, separated samples into one normal (Cluster1) and two tumor clusters (Cluster2 and 3, Fig. 2). Most tumors (Cluster2, n = 98) split into two subgroups: one (n = 20) with high genome instability (wGII, median: 0.4) and low chromosome 3p loss (8/20), and another (n = 78) with widespread chromosome 3p loss (76/78) and low wGII (median: 0.1). A smaller tumor cluster (Cluster3, n = 16) showed extensive hypermethylation, high wGII (median: 0.4), frequent losses of chromosome 3p (16/16), 9p (11/16), and 14q (12/16), and was associated with poor prognosis—median survival time of 2.01 years vs. 3.42 years in Cluster2 (HR: 3.13 (1.67–5.87), p < 0.001, Supplementary Fig. S6). Cluster3 tumors were more aggressive, with most patients experiencing recurrence (5/16) or synchronous metastases (9/16).

Fig. 2.

Heatmap of top 5000 differentially m ethylated regions. Heatmap of top 5000 methylated regions (adjusted P < 0.05 and top absolute log2FC). Nrpm: normalized reads per million mapped reads (methylation values). Euclidean distance and ward.D2 clustering. Columns splitted by the three main clusters; cutree (k = 3). Left annotations: Chr: chromosomes. DMR: significant differentially methylated regions (hyperDMR: log2FC > 1.3, adjusted P < 0.05; hypoDMR: log2FC < − 1.3, adjusted P < 0.05). Top annotations: Chr3p, Chr9p, and Chr14q: loss (log2ratio < 0.1), gain (log2ratio > 0.1) or no CNA of chromosome arms. wGII: weighted genome instability index. Age: age groups <  = 55, > 55–65, > 65–80, and > 80 years. Sex: male or female. Leibovich: Leibovich groups for localized patients. Stage: TNM stage. Methylation: median split on median nrpms across regions. Status: no recurrence, recurrence, or synchronous metastatic disease. Sample type: ccRCC tumor or normal tissue

No clear associations were found with sex, stage, or Leibovich risk scores across clusters. However, chromosome 9p/14q loss or high methylation appeared in all low-stage cases (stage I and low Leibovich risk score) that experienced recurrence.

Differentially methylated region analysis highlights importance beyond coding regions

A differentially methylated region (DMR) analysis identified 2424 significant hypermethylated DMRs (HyperDMRs) and 293 significant hypomethylated DMRs (HypoDMRs) in ccRCC tumors compared to normal samples (Fig. 3a), distributed across all chromosomes (Fig. 3b). DMRs were predominantly annotated to open-sea regions (89.4%), with 4.1% annotated to CpG shores and 3.3% annotated to CpG shelves (Supplementary Fig. S7a). This indicates that epigenomic dysregulation in RCC extends beyond traditionally studied promoter regions, potentially affecting other regulatory elements. 70.9% of regions were annotated to repetitive elements, underscoring their significance and the added value of MeDIP-seq compared to array-based methods. Many of these repeat-associated regions contributed to the broad hypermethylation peaks observed on chromosomes 3p, 6p, 9p, and 14q (Fig. 1b), and the most strongly hypermethylated repeat regions showed higher methylation in advanced and recurrent tumors (Supplementary Fig. S5). Furthermore, 47.9% of DMRs were intergenic, and 39.7% were within intronic regions (Supplementary Fig. S7a), highlighting the fact that methylation changes in ccRCC might influence gene regulation through enhancer or silencer modulation, alternative splicing, or chromatin accessibility rather than direct promoter silencing. The genomic distribution of HyperDMRs and HypoDMRs was largely consistent (Supplementary Fig. S7b).

Fig. 3.

Differentially methylated region analysis between tumor and normal. ccRCC tumor vs ccRCC normal. a, b Red: Tumor-enriched HyperDMRs (log2FC > 1.3, adjusted P < 0.05; n = 2424). Blue: Normal-enriched HypoDMRs (log2FC < − 1.3, adjusted P < 0.05; n = 293). Grey: Not significant. The most significant annotated genes are shown. a Volcano plot (− log10(adjusted P) to log2FC) of all methylated regions. b Log2FC of all regions across chromosomes. c, d MethScore is calculated by the sum of methylation values across HyperDMRs (n = 2424) per sample, divided by the total number of regions (n = 2424). c MethScore across tumor stages. d MethScore comparison by recurrence status (Metastatic: synchronous metastatic). Statistical comparisons were calculated by Wilcoxon tests

Gene set enrichment analysis was performed on HyperDMRs to gain insights into epigenetically driven dysregulation in ccRCC. Genetic regions were enriched in gene expression regulation, signal transduction pathways, and the Hippo signaling pathway (Supplementary Fig. S8), which has been described as dysregulated in ccRCC [35].

Prognostic significance of primary tumor methylation

To explore the relationship between methylation, disease stage, and aggressiveness, a methylation score (MethScore) was calculated for each sample. MethScore is the sum of methylation values across all tumor-enriched HyperDMRs (Tumor vs Normal: n = 2424; Supplementary Data S3) divided by the total number of regions for each sample (Materials and Methods). The MethScore showed a progressive increase with advancing stage, metastatic disease (Fig. 3c, d, Supplementary Data S4), and in the aggressive methylation Cluster3 (P < 0.0001, Supplementary Fig. S10).

Furthermore, MethScore was strongly associated with overall survival in both univariate (HR = 1.47, P < 0.001) and multivariate (HR = 1.24, P = 0.038) Cox regression analyses, alongside age and Stage IV (Fig. 4a, b). In patients with localized disease, MethScore remained a robust independent predictor of poor outcome (HR = 1.65, P = 0.003, Fig. 4c, d). This association remained robust after exclusion of HyperDMRs on chromosome 3 (Supplementary Fig. S9).

Fig. 4.

MethScore is associated with a higher risk of death. Cox proportional hazards regression of overall survival with MethScore scaled to represent 0.1-unit increments. a Univariate for all patients. b Multivariate: MethScore, age, sex, stage, wGII, and tumor purity for all patients. c Univariate for localized patients. d Multivariate: MethScore, age, sex, stage, Leibovich risk score, wGII, and tumor purity for localized patients. wGII: weighted genome instability index

High methylation (median split) of MethScore regions (n = 2424) was linked to significantly shorter survival, both in localized and advanced cases (Fig. 5a, b). Similarly, high methylation of the top 100 HyperDMRs enriched in advanced or recurrent tumors predicted worse survival (p < 0.0001, Supplementary Fig. S11).

Fig. 5.

Prognostic value of MethScore. Kaplan–Meier analysis of MethScore regions (n = 2424) for tumor samples. a Overall survival with median split on nrpm of HyperDMRs. b Overall survival with median split on nrpm of HyperDMRs and stratified by localized (stage I, II, and III) and advanced (stage IV) disease. Nrpm: normalized reads per million mapped reads

To validate our findings, we compared hypermethylated regions from MeDIP-seq with TCGA KIRC HM450K data [7]. Of the 2424 HyperDMRs, only 51 regions (71 probes) overlapped with TCGA and showed only a non-significant trend toward poorer survival (Supplementary Fig. S12a). In contrast, 114 regions (182 probes) of the top 5000 HyperDMRs overlapped with TCGA, and high methylation in these regions was significantly associated with worse overall and progression-free survival (Supplementary Fig. S12b–d). Regions linked to stage or recurrence could not be validated due to limited overlap with array data.

Discussion

In this study, we present the first large-scale, genome-wide methylation profile of ccRCC using MeDIP-seq, revealing global hypomethylation and regional hypermethylation, supported by selected previous studies [36, 37]. Owing to our genome-wide approach, we identify hypermethylation at chromosome arm level, including chromosomes 3p, 4p, 6p, 8p, 9p, and 14q, which, to our knowledge, has not previously been investigated. Methylation was enriched in intergenic regions and gene bodies, which supports the region's functional relevance in ccRCC and is consistent with the pilot observations by Malouf et al. [38]. Further, we identify hypermethylation in repetitive and transposable elements, which have not previously been shown.

Compared with array-based platforms such as the 450K/EPIC arrays, which primarily target CpG islands and promoters, MeDIP-seq provides genome-wide coverage and captures methylation changes in enhancers, intergenic regions, and repetitive elements, areas that are particularly dysregulated in ccRCC. This broader coverage allows the detection of global hypomethylation patterns and regulatory-element alterations that would be missed by array-based profiling. Relative to bisulfite-based methods such as WGBS, MeDIP-seq also requires substantially less input DNA and sequencing depth, enabling high-throughput profiling of clinical samples while simultaneously providing copy-number information. A limitation of MeDIP-seq is its lower resolution, as methylation is inferred from enriched fragment density (~ 300 bp) rather than at single-CpG resolution. However, this trade-off is balanced by increased sensitivity and genome-wide breadth, allowing inclusion of a larger number of DMRs and improved detection of widespread methylome alterations in ccRCC.

The predominance of open-sea and intragenic HyperDMRs suggests that many of these methylation changes may occur at distal regulatory elements such as enhancers or silencers, although a systematic assessment of enhancer overlap would require integration with RCC-specific chromatin datasets (e.g., H3K27ac or ATAC-seq), which were not generated for this cohort. These changes go beyond the scope of array-based methods, underscoring the advantage of comprehensive profiling. Moreover, using a single tumor biopsy, we recapitulate key genomic features, such as chr3p loss in 88% of cases, consistent with RenalTRACERx data [18], reinforcing its role as an early and ubiquitous event in ccRCC tumorigenesis. Further, methylation of chromosome 3p was associated with recurrence and metastatic disease, potentially demonstrating a biallelic inactivation mechanism.

Hierarchical clustering of methylation patterns identified three distinct methylation subtypes: One normal (Cluster1) and two tumor clusters (Cluster2 and 3). Within Cluster2, two subgroups emerged—one with low chr3p loss and high wGII, another with high chr3p loss and low wGII, potentially reflecting the VHL-wildtype and VHL-monodriver subtypes described in RenalTRACERx [18], with the latter typically linked to lower-stage tumors, aligning with our data. Cluster3 showed extensive losses at chromosomes 3p, 9p, and 14q, high wGII, elevated methylation, and poor prognosis consistent with the aggressive, multi-clonal driver subtype also reported in TRACERx.

We derived a methylation score (MethScore) based on 2424 HyperDMRs, which independently predicted overall survival, particularly in localized cases, and outperformed the traditional Leibovich risk model. As adjuvant immunotherapy (IO) has recently been introduced to high-risk localized cases [39], improved risk stratification is essential to identify patients most likely to benefit. Integrating methylation-based scores could refine patient selection and information in the adjuvant setting, potentially distinguishing those at higher risk of recurrence who may benefit from IO from those with more indolent diseases who may be spared unnecessary treatment.

While TCGA validation was limited by differences in disease stage (TCGA KIRC data are enriched for low stage disease, whereas our cohort is enriched for advanced disease) and platform coverage, the top-ranking methylated regions remained prognostic. This indicates the need for future whole-genome methylation studies in larger, more balanced cohorts. External validation using contemporary EPIC methylation arrays was not feasible, as currently available RCC EPIC datasets lack either clinical outcome information or sufficient cohort size to support prognostic validation.

Although our study comprises a large cohort of patients and long-term follow-up, it is limited by the single-biopsy analysis, which may overlook regional variation given the high level of epigenetic intratumoral heterogeneity [8]. However, identifying markers detectable from a single sample aligns with clinical feasibility. In addition, our method entails tumor purity below 100% due to the bulk analysis technique, which we accounted for in our Cox regression analysis. In addition, our low observed rate of whole genome doubling in our cohort compared to other studies [40] may reflect limited sequencing depth rather than true biological absence, constraining some genomic analyses. Moreover, while MeDIP-seq offers cost-effective, genome-wide coverage, it is less sensitive to hypomethylation and lacks single-base resolution. Another consideration is the use of adjacent renal tissue as the “normal” comparator. Epigenetic field effects, where histologically normal tissue exhibits early molecular alterations related to carcinogenesis, have been reported in several tumor types. Such field effects could introduce subtle methylation shifts in adjacent samples; therefore, our findings should be interpreted with this possibility in mind. Future studies incorporating non-cancer renal tissue will help more clearly define the true baseline renal epigenome. Finally, since whole-genome methylation data, extending from array-based coverage, in ccRCC is rare, we have found it difficult to validate our findings. Future genome-wide methylation studies in large cohort sizes are needed to validate our findings further.

Author contributions

Conception and design of study: LI, IL; Acquisition and analysis of data: LI, IL, CL, JA, LZ, AKK, MGT, KDS, NF; Interpretation of data: LI, IL, JA, JJ, KK; First draft of paper: LI, IL, JJ; Approval of submitted version: All authors.

Funding

The study was funded by the Independent Research Fund Denmark (2100-00017B), The Danish Cancer Society (R413-2022-606), The Lundbeck Foundation (R413-2022-606), Fabrikant Einar Willumsens Fond, and Harboefonden. IL was funded by a fellowship grant from the Lundbeck Foundation (R413-2022-606). The funding agencies had no role in the design, analysis, or reporting of the study.

Availability of data and materials

Data are available through controlled access from GenomeDK ([https://genome.au.dk/library/]) (https://genome.au.dk/library/GDK000019/).

Declarations

Ethics approval and consent to participate

This study was approved by the Danish National Center for Ethics (j. no. 2208138).

Consent for publication

Not applicable.

Declaration of generative AI and AI-assisted technologies in the writing process

During the preparation of this work, the authors used ChatGPT-4 to optimize language. After using this tool/service, the authors reviewed and edited the content as needed and take full responsibility for the content of the publication.

Competing interests

All the authors declared no competing interests.