DNA methylation associates with survival in non-metastatic clear cell renal cell carcinoma

The study cohort consisted of 115 ccRCC patients, primary treated with radical or partial nephrectomy between 2001 and 2009, and diagnosed at the University hospital in Umeå, Sweden. None of the patients received neoadjuvant or adjuvant therapy. Eighty-seven patients were metastasis-free (M0), while 28 had metastases (M1) at diagnosis. Tumor free (TF) tissue samples were obtained from 12 surgically removed tumor bearing kidneys and were considered histologically normal by a pathologist. The tumor and TF tissue samples obtained were snap-frozen in liquid nitrogen, and stored in − 80 °C until analysis.

Patients were followed-up at least yearly by routine clinical and radiological examination in accordance with a scheduled follow-up program. Clinical follow-up data were extracted in August 2017. All patients have given informed consent and the study was approved by the regional ethical review board in Umeå (Dnr 2011–156-31 M, 20110523).

The publically available TCGA-KIRC dataset was used as a validation cohort and clinical information was downloaded from the Broad Institute’s Genome Data Analysis Center Firehose (http://​gdac.​broadinstitute.​org/​). Only unique non-metastatic (M0) ccRCC samples (technical replicates excluded) analysed with Illumina HumanMethylation450K array were included in the analysis (n = 230). All patients were treated with radical or partial nephrectomy and patients receiving neoadjuvant and/or adjuvant therapy were excluded.

DNA was extracted from the tissue samples as described previously [35] and was subjected to bisulfite conversion (500 ng of each sample) using the EZ DNA Methylation Kit (Zymo Research, Irvine, USA) according to the manufacturer’s protocol. Bisulfite DNA conversion was verified by MethyLight analysis of the ALU gene with the ALU-C4 primer/probe set as described [36]. Genome-wide assessment of DNA methylation was performed using HumanMethylation450K BeadChip arrays (Illumina, San Diego, CA, USA) according to manufacturer’s protocol. To each array, 200 ng of bisulfite-converted DNA was applied, and the arrays were scanned with a HiScan array reader (Illumina). The fluorescence intensities were extracted using the Methylation module (1.9.0) in the Genome Studio software (V2011.1). Pre-filtering and normalization steps are shown in Additional file 1: Table S1 and were performed as previously described [37, 38], which excluded the X and Y chromosomes, CpGs with detection p-value > 0.05 and CpG probes that aligned to multiple loci in the genome or were located less than 3 bp from a known single nucleotide polymorphism [39]. The methylation levels (i.e., the β value) of each CpG sites ranges from 0, corresponding to completely unmethylated DNA, to 1, representing fully methylated DNA. The technical reproducibility of methylation array analysis was monitored by including a replicate sample on each array and the R²-values ranged from 0.995 to 0.997.

The analysis was restricted to the n = 155,931 CpGs located in gene promoter regions, i.e. located within TSS1500, TSS200 and 5’UTR, remaining after the initial filtration steps (Additional file 1: Table S1). Twelve TF tissue samples were included as reference samples. The TF-samples showed a high similarity in promoter methylation, the R²-values ranged from 0.96 to 0.99. A CpG site was determined as differently methylated (DM-CpG) if the absolute value of the difference in beta value between tumor sample and the mean of TF samples (Δβ) was greater than or equal to 0.2. The DM-CpGs in the M0-PF (non-metastatic progression free), M0-P (non-metastatic with later progress) and M1 (metastasis at diagnosis) groups were analysed against known methylation quantitative trait loci (mQTLs) [40] based on middle-aged individuals to evaluate genetic variants versus cancer specific alterations, and these sites were excluded from further analysis.

The epigenetic mitotic clock described by Yang et al., 2016 was used to estimate mitotic age [41]. A prognostic Risk score was calculated for 114 out of 115 tumor samples using the CpG-methylation-based assay previously presented by Wei et al., (2015) [33]. Patients with a Risk score higher than − 0.1 were defined as high risk and lower than − 0.1 were defined as low risk, as previously stated [33].

DM-CpGs within the M0-PF, M0-P and M1 groups were selected for further analysis if differently methylated in at least 70% of samples (in Fig. 7). Hypermethylated CpGs showed increased methylation compared to TF whereas hypomethylated CpGs were less methylated.

The commonly DM-CpGs (n = 172) in the M0-P and M1 samples were defined as a Promoter Methylation Classification (PMC) panel. The hypomethylated CpGs (n = 51) were mirrored (1 – average beta) and thereafter an average beta of the 172 DM-CpGs were calculated for each sample. A ROC-curve was constructed with PMC average beta as test variable and tumor progression within five years as state variable. Youden index was used to determine the cut off for PMC groups (PMC high/low), and was set to PMC average beta 0.688 (specificity 0.85 and sensitivity 0.55).

The prognostic relevance of PMC classification was confirmed in a separate ccRCC cohort (n = 230) within the KIRC project of TCGA [42]. Clinical information along with methylation raw data (Illumina HumMeth450Karrays) were downloaded from the Broad Institute’s Genome Data Analysis Center Firehose (http://​gdac.​broadinstitute.​org/). Beta values were constructed using the genome studio definition and was normalized for different bead types using BMIQ. Missing values among the 172 CpG-sites in the PMC panel were imputed using the k-nearest neighbours method. Deaths without prior progression were counted as non-events and were censored.

The distribution of hyper- and hypomethylated CpGs within twelve genomic regions with frequent gain/loss in ccRCC (as defined by Köhn et al. 2014 [9] and listed in Additional file 2: Table S3) was analysed to identify potential overrepresentation of hyper- and/or hypomethylated CpGs within these regions.

In six individuals, multiple tumor samples were taken from different locations within the same kidney to study intratumoral heterogeneity. To confirm that the collected samples originated from the same individual, the methylation levels of the 65 built in SNP probes (in the methylation array) were analysed. One of the 6 patients (number 3) was excluded at this step due to signs of contaminated DNA (Additional file 3: Figure S1).

The genetic aberrations were profiled using the total intensity signals of the raw data from the HumanMethylation450K arrays [43, 44]. Briefly, copy number variation (CNV) analysis was performed in R (v3.4.1) using the Conumee package (v1.9.0) [45] with data imported through minfi (v1.18) [46]. Parameters and limits for calling deletion and gain were set for each sample individually through visual inspection.

The validity of this method to identify CNV was confirmed by comparing the methylation results with genetic aberrations identified by HumanCytoSNP-12 v2.1 arrays (Illumina) in a subset of samples (n = 57 ccRCC) [9]. The CNV analysis was performed in Genome Studio v1.8 using the Genotyping Module (Illumina). Cohen’s kappa test was used to compare the results from the two array types, which were significantly overlapping (p ≤ 0.001 for all analysed aberrations), with quality of agreement moderate or good (Additional file 4: Table S2).

The genomic aberrations for all ccRCCs were summarized by investigating the minimal overlapping regions of commonly occurring CNVs in ccRCC [9]. The CNVs are presented as percentage of samples where alterations across the entire regions were found (Additional file 5: Figure S2 and Additional file 2: Table S3).

RNA was extracted from 28 tumors using MagAttract RNA Universal Tissue M48 Kit (Qiagen, Hilden, Germany) according to manufacturer’s protocol using BioRobot M48. RNA concentrations were determined by spectrophotometry (NanoDrop, Thermo Scientific, Wilmingron, DE, USA) and quality was analysed using the 2100 Bioanalyzer (Agilent technologies, Santa Clara, CA, USA).

Two hundred ng of total RNA from each sample was used for cRNA production by the Illumina TotalPrep RNA amplification kit (Ambion Inc., St. Austin, TX, USA) according to the provided protocol. The quality of purified cRNA was evaluated using the RNA 6000 p kit (Agilent Technologies) in the Agilent 2100 Bioanalyzer (Agilent Technologies). A total of 750 ng biotinylated cRNA was hybridized to the human HT12 Illumina Beadchip gene expression array (Illumina, San Diego, CA, USA) according to manufacturer’s protocol and scanned using the Illumina Bead Array Reader (Illumina). Expression array data was analysed using the Illumina BeadStudio V2011.1 software and samples were normalized using the cubic spline algorithm. Gene expression levels of the MX2 (ILMN_2231928), SMAD6 (ILMN_1767068) and SOCS3 (ILMN_1781001) genes were extracted from the arrays.

For statistical analysis, the Statistical Package for the Social Sciences (SPSS Inc., Chicago, IL) software version 24, was used. The chi-square/Fisher’s exact test was used to compare differences between subgroups among categorical variables and the Mann–Whitney U test was used for continuous variables. Kruskal Wallis test was used for continuous variables when comparing three groups of samples. Mann–Whitney U test with Bonferroni correction was used in Additional file 6: Figure S4.

Estimates of 5-year cancer specific survival (pCSS_5yr) rates and Cumulative incidence of Progress (CIP_5yr) in subgroups of ccRCC were obtained from Kaplan–Meier survival tables, and the equality of survival distributions for the groups was compared using the log rank test. The significance level used in all tests was 0.05. In CIP analysis local, regional or distant metastatic progression was the endpoint. In the CSS analysis, ccRCC specific death was the endpoint.

Hierarchical clustering was performed using the Ward’s method and a Euclidean distance metric for clustering samples.

Principal Component Analysis (PCA) was performed in SIMCA version 14 (Umetrics, Umeå, Sweden) after centering the average beta values of 155,931 promoter associated CpG sites.

WebGestalt [47] was used to analyse functional enrichment of the genes associated with progress of disease (specified in Fig. 7a), using all genes represented by the 155,931 promoter associated CpGs as the background list. The gene functions of potential relevance for ccRCC pathology were selected among the 20 most significant GO Terms (Additional file 7: Table S8).

One hundred and fifteen tumor tissue samples from ccRCC patients along with 12 adjacent tumor free (TF) kidney-cortex tissue samples were analysed using the genome-wide HumanMethylation450K arrays.

We have previously shown that genome-wide promotor methylation status based on Illumina Human Methylation 27 K arrays can predict survival in ccRCC [34]. To further examine the prognostic relevance of promoter methylation status with regard to cancer specific survival and progress free survival we focused the analysis to 155,931 promoter associated CpGs on the HumanMethylation450K array, in metastatic and non-metastatic ccRCCs. Hierarchical clustering of the 155,931 promoter associated CpGs divided the ccRCCs (n = 115) and the TF samples (n = 12), into two clusters; A (n = 81 tumor tissue samples and n = 12 TF samples) and B (n = 34 tumor tissue samples) (Fig. 1). Cluster A contained a high fraction of samples from patients with non-metastatic disease at diagnosis (M0), as well as the TF tissue samples, whereas samples from patients with metastasis at diagnosis (M1) were enriched in cluster B (p < 0.001). Cluster B tumors had larger tumor diameter (p < 0.001), higher morphological grade (p < 0.001), higher TNM stage (p < 0.001) and poorer outcome (Table 1, Fig. 2). There was no difference in gender or age at diagnosis between cluster A and B (Table 1).

Cancer specific survival (pCSS) analysis confirmed the prognostic relevance of methylation cluster classification in ccRCC (pCSS_5yr 80% for cluster A vs 27% for cluster B, p < 0.001) (Fig. 2a). As previously known, the difference in CSS is large between patients with distant metastasis (M1, n = 28) at diagnosis and patients with non-metastatic disease (M0, n = 87) (pCSS_5yr 7% vs. 84%, p < 0.001) (Fig. 2b). We found no difference in pCSS_5yr in the M1 patients regardless of cluster status (pCSS_5yr 9% vs 6%, p = 0.840) (Fig. 2c), whereas cluster status clearly separated survival in the M0 patients (pCSS_5yr 92% vs 50%; p < 0.001) (Fig. 2c). Importantly, the cumulative incidence of progress (CIP) in M0 patients showed a significant difference between cluster A and B (p < 0.001), with lower incidence in cluster A patients (pCIP_5yr 15%) compared to cluster B patients (pCIP_5yr 63%) (Fig. 2d). However, methylation cluster status did not remain as an independent prognostic marker for neither CSS (n = 115) nor progression free survival (PFS) (n = 87) in multivariate Cox regression analysis including cluster status, TNM-stage, morphological grade, age and gender. In this analysis, only TNM stage, grade, and gender remained as independent markers (Additional file 8: Table S4 and Additional file 9: Table S5).

To relate cluster A/B status to previously defined prognostic panels in ccRCC, we classified the tumor samples based on the five CpG-site Risk classification panel described by Wei et al., (2015) [33]. Risk score could be calculated for 114 out of 115 tumor samples and were classified as Low (n = 63) or High risk (n = 51). There was no significant correlation between cluster A/B status and Wei Risk score (p = 0.745, Table 1). Risk classification according to Wei predicted CSS (pCSS_5yr 72% for Low risk vs 55% for High risk, p = 0.027) (Additional file 10: Figure S3A), but not as strong as genome-wide promoter methylation cluster classification (Fig. 2a). When stratifying patients into non-metastatic (n = 86) and metastatic (n = 28) groups according to Wei there was neither any statistical difference in CSS (pCSS_5yr 88% for Low risk vs 78% for High risk p = 0.113) in non-metastatic tumors nor between Low risk vs High risk in metastatic tumors (pCSS_5yr 0% vs 12%, p = 0.288, Additional file 10: Figure S3B). Further there was no difference in CIP between non-metastatic tumors with low risk versus high risk (pCIP_5yr 20 and 31% respectively, p = 0.099) (Additional file 10: Figure S3C).

The frequency of 12 common genetic aberrations in ccRCC as defined by Köhn et al. 2014 [9] (listed in Additional file 2: Table S3 and Additional file 5: Figure S2), were related to methylation status in the 115 ccRCCs. Among the 12 genomic regions, aberrations were found in 15 to 84% of the ccRCCs (Fig. 1, Additional file 2: Table S3, Additional file 5: Figure S2).

The most common genetic aberration in ccRCC is loss of chromosome 3p (including the VHL gene) [48]. In our cohort, 84% of the ccRCCs showed a loss of 3p, and deletions were similarly distributed in clusters A and B (Fig. 1, Additional file 2: Table S3). Loss of 9p (p < 0.001), 9q (p < 0.001) and 14q (p = 0.005) were more frequently observed in cluster B samples, whereas the other aberrations were similarly distributed among the clusters (Fig. 1, Table 1).

The chromosomal distribution of hyper- and hypomethylated CpGs was analysed in relation to the genomic aberrations in the 12 previously defined regions (Additional file 2: Table S3 and Additional file 6: Figure S4). Although genetic aberrations on chromosome 7p and 9q were associated with overrepresentation of hyper- and/or hypomethylated CpGs, there were no general enrichment of DM-CpGs within genomic regions harboring aberrations (Additional file 6: Figure S4). Importantly, loss of 3p which is frequently observed in ccRCC was not associated with significant accumulation of neither hyper- nor hypomethylated CpGs (Additional file 6: Figure S4).

We observed an intra-individual heterogeneity in the number of aberrations within the M0-PF (progress-free), M0-P (progress) and M1 (distant metastasis at diagnosis) group of patients (Fig. 3). M1 patients had generally more genomic aberrations compared with the M0-PF and M0-P patients (p = 0.024 and p = 0.050, respectively) (Fig. 3 and Table 2). The only genetic aberration that was more frequent in the M0-P compared to M0-PF patients, was the loss of 9q (43% vs. 20%, p = 0.031), which was even more frequent in the M1 patients (61%) (Table 2).

Genetic and epigenetic aberrations may accumulate during extended replication. In order to evaluate such possible associations in ccRCC, we analysed the number of hyper- and hypomethylated CpG sites in relation to number of genetic alterations and estimated mitotic age of the tumor samples. The number of genetic aberrations were positively correlated to number of hypermethylated CpGs (p < 0.001), but not to the number of hypomethylated CpGs in promoters (Fig. 4a, Additional file 11: Figure S5A). The number of genetic aberrations and number of hypermethylated CpGs were both strongly positively associated with estimated mitotic age [41] (R = 0.407, and R = 0.641, respectively, p < 0.001) (Fig. 4b-c), but there was no correlation between number of hypomethylated CpGs and mitotic age (Additional file 11: Figure S5B).

Since cluster status was of prognostic relevance in M0 patients, we focused the analysis on identifying methylation alterations associated with clinical progress.

M0-PF, M0-P and M1 patients showed gradually larger median tumor diameter (p < 0.001), higher TNM stage (p < 0.001), higher morphological grade (p < 0.001), increased number of hypermethylated CpGs (p = 0.015), higher mitotic age (p = 0.047), and a higher correlation to cluster B status (p < 0.001) (Table 2). The M0-PF and M0-P groups differed significantly in all these clinicopathological parameters (except for mitotic age) (Table 2). In contrast, the M0-P and M1 groups differed in TNM (p < 0.001) and gender distribution only (p = 0.046) (Table 2).

With the aim to distinguish TF, M0-PF, M0-P and M1 subgroups of patients based on the promoter associated methylation profiles, we used principal component analysis (PCA) (Fig. 5). All ccRCC and TF samples were included in the PCA, and different sample categories were highlighted to identify similarities and/or differences. TF tissue samples showed a homogenous methylation pattern (Fig. 5a-b), and cluster A and cluster B patients could be separated (Fig. 5a). However, M0-PF, M0-P and M1 patients could not be clearly separated by PCA since their methylation patterns were partly overlapping (Fig. 5b).

Instead, we aimed to identify specific CpGs associated with progress after primary treatment. DM-CpGs were identified in each tumor sample using the average methylation for TF tissue samples as a reference. Large inter-individual variations in the number of hyper/hypomethylated DM-CpGs was observed in the tumor samples (Fig. 6a). The number of hypermethylated DM-CpGs were significantly higher in the M0-P and M1 patients compared with M0-PF (p = 0.025 and 0.001, respectively) (Fig. 6a, left chart). The number of hypomethylated DM-CpGs did not differ significantly between the groups of patients (Fig. 6a, right chart).

To further dissect possible variations in methylation profiles at intra-tumoral level, we analysed multiple samples taken from various parts of the tumor in five individuals (Additional file 12: Table S6). We found that samples from the same tumor generally had similar methylation profiles (R² correlations 0.97 to 0.99; p ≤ 0.001). Accordingly, the number of hyper- and hypomethylated CpGs were similar (Fig. 6b), and the individuals´ samples clustered together in PCA-analysis (Fig. 6c).

DNA methylation alterations have been described as early events in tumor progression. Therefore, we focused on common DM-CpGs in the M0-P and M1 samples, with potential mQTLs excluded [40], which might represent epigenetic alterations of relevance for the formation of metastasis. CpGs that showed either hyper- or hypomethylation in at least 70% of samples in the M0-PF, M0-P and M1 groups of patients were selected and combined in a Venn diagram (Fig. 7a). Most of the DM-CpGs in the M0-PF patients were also present in the M0-P and M1 patients, indicating that methylation alterations accumulated during progression and was not dominated by stochastic events. Only a low percentage of DM-CpGs were unique for M0-PF and M0-P (13 and 17% respectively) whereas M1 had 56% unique DM-CpGs. 172 DM-CpGs were common for the M0-P/M1 patient groups (121 hyper- and 51 hypo-methylated) (Fig. 7a). These 172 DM-CpGs were defined as a Promoter Methylation Classifier (PMC) panel, and was used to classify samples as PMC low and high as described in the materials and methods section. In the non-metastatic tumors of our cohort, 40 samples were classified as PMC low and 47 samples as PMC high. The pCIP_5yr survival analysis showed poorer outcome in the PMC high subgroup (PMC low pCIP_5yr 8% vs. PMC high pCIP_5yr 38%, p = 0.001; Fig. 7b). The PMC panel (PMC high Hazard ratio (HR) 4.4 (1.3–15.8)) and TNM stage (TNM III HR 3.9 (1.4–11.0)) remained significant prognostic markers for PFS in a Cox regression analysis including PMC status, TNM, morphological grade, age and gender (Additional file 13: Table S7). The prognostic relevance of PMC was confirmed in 230 ccRCC tumor samples from the TCGA-KIRC data set (PMC low (n = 144) pCIP_5yr 16% vs. PMC high (n = 86) pCIP_5yr 39%, p < 0.001; Fig. 7c).

The functional relevance of the PMC associated genes were analysed for the most significant gene ontology categories (GO terms) (Fig. 7a, Additional file 7: Table S8). The commonly hypermethylated genes were enriched for GO terms including SMAD protein complex assembly, RNA polymerase II regulation and response to pH (Fig. 7a). The hypomethylated genes were enriched for GO terms associated with external stimulus-, immune- and defence- response (Fig. 7a). Three genes were of special interest with regard to ccRCC progress; i.e. SMAD family member 6 (SMAD6, CpG cg10402698 and cg12476188), Suppressor of cytokine signaling 3 (SOCS3, CpG cg10279487 and cg27637521) and MX dynamin like GTPase 2 (MX2, CpG cg05656374 and cg15281283). These genes showed significant altered methylation levels between TF/M0-PF and the M0-P/M1 samples in more than one promoter associated CpG site (Fig. 7d-f).

Methylation alterations in these genes were analysed in relation to mRNA gene expression alterations in 28 samples with available RNA (M0-PF = 13; M0-P = 5 and M1 = 10). A significant increased gene expression was observed in M0-P and M1 samples compared to M0-PF samples in the SOCS3 (ILMN_1781001; p = 0.019 and 0.001) and MX2 (ILMN_2231928; p = 0.014 and 0.002) genes, whereas no significant difference in expression for SMAD6 (ILMN_1767068) was seen (Fig. 7g-i).