DNA hypermethylation-induced gene silencing is a common event in many malignancies and serves as an alternative mechanism to genetic mutation for the loss of tumor suppressor functions [1, 2]. Although the mechanisms that underlie aberrant DNA methylation in cancer cells remain to be elucidated, current evidence suggests that it may be an early and possibly even an initiating event in the development of colorectal cancer (CRC).

A subset of CRC has been shown to exhibit frequent and concurrent hypermethylation at specific gene promoters and is referred to as the CpG island methylator phenotype (CIMP+) [3]. CIMP+ CRC is associated with distinct clinicopathological and molecular features including proximal tumor location, preponderance in elderly females, poorly differentiated and mucinous tumor histology, microsatellite instability (MSI) and frequent BRAF V600E mutation [3–10]. CIMP+ CRC often lack the hallmark genetic alterations in APC, p53 and 18q that characterize the classic adenoma-carcinoma sequence. Instead, CIMP+ tumors are thought to develop along an alternate serrated adenoma pathway in which hypermethylation rather than mutation is used to inactivate tumor suppressor genes [11].

In an effort to establish CIMP+ CRC as a distinct subgroup of CRC, Laird and colleagues analysed the methylation of 195 individual gene promoter regions in 295 CRC using the quantitative MethyLight assay [10]. From their results, they proposed a panel of 5 CpG island methylation markers to standardize the classification of CIMP+ CRC. However, different groups have continued to use a variety of methylation markers to define CIMP+ CRC [7, 12–14]. The lack of consensus markers has led to reports of several CIMP subgroups according to the frequency of CpG island methylation [13–15]. The investigators who originally proposed CIMP recently described two subgroups of CIMP+, termed CIMP-1 and CIMP-2, that displayed increased frequencies of BRAF and KRAS mutations, respectively [14]. Similarly, Nagasaka et al described two distinct patterns of gene methylation in CRC that also segregated with BRAF and KRAS mutations [13, 16]. Using a panel of 8 methylation markers, Ogino et al identified a CRC subgroup which they termed CIMP-low that was associated with frequent KRAS mutation, MGMT methylation and occurrence in males [17].

Most previous studies of CIMP+ CRC have investigated a relatively small number of CpG island markers for methylation. The GoldenGate Methylation BeadArray (Illumina, Inc.) technology provides the opportunity for high-throughput methylation analysis of a large number of CpG sites. In the present study the GoldenGate Methylation Cancer Panel I containing 1,505 CpG loci within 807 cancer-related genes was used to study methylation patterns in 91 unselected CRC. These genes were selected based on their involvement in cell growth control, differentiation, migration, apoptosis, DNA damage repair and oxidative metabolism. The GoldenGate technology allowed us to identify three distinct CRC subgroups according to their methylation pattern which showed distinctive clinicopathological and molecular characteristics and differed in their frequencies of BRAF and KRAS mutation.

The current study is the first to use array-based technology to enable comprehensive methylation profiling of CRC. A total of 1,505 CpG sites contained within 807 genes were assessed in 91 consecutive cases of CRC. The GoldenGate^® arrays employed here were recently used to profile methylation in head and neck cancer [31], renal cancer [32], glioblastoma [33] and hematological neoplasms [24, 27]. The validity of these arrays for the quantitative assessment of methylation was shown in several previous studies by comparison with other quantitative methods [23, 26, 34]. The finding that methylation of CpG sites in X-linked genes correlated with gender provided further validation [Additional file 1]. Many of the genes found to be hypermethylated in this study were previously reported to be methylated in CRC [Additional file 5]. Finally, in agreement with earlier work on cancer [30], many of the genes showing de novo hypermethylation in this study of CRC (cluster A and C genes, Fig. 1) are known targets for PRC2 [29].

Similar to earlier studies in CRC that evaluated a limited number of methylation markers [3–10], comprehensive methylation profiling in the present study revealed the existence of distinct tumor subgroups (Fig. 1). The three major subgroups identified by unsupervised hierarchical clustering were classified as CIMP-H, CIMP-M and CIMP-L according to the level and frequency of methylation. In agreement with previous studies, CIMP-H tumors were associated with older patient age, proximal site and BRAF mutation (Table 1). All 16 tumors identified as CIMP^W-high using a proposed consensus panel of 5 markers were contained within the CIMP-H group, as well as all 15 tumors containing a BRAF mutation. Using small numbers of methylation markers in unselected CRC, the original studies by Toyota et al reported CIMP+ frequencies of 62% [3] and 51% [35] whereas subsequent studies reported lower frequencies of 15-32% [5, 7–10], [14]. In contrast, by investigating a large number of methylation sites and using unsupervised hierarchical clustering to analyze the results, we observed a relatively high proportion (65%) of CIMP-H tumors in the present study.

Previous studies have reported inconsistent results for the association between CpG island methylation and KRAS mutation [8, 13, 17, 35, 36], probably because of the different methylation markers used in each study. Analysis of a large number of CpG sites in the present study revealed that CIMP-H tumors showed a significantly higher KRAS mutation frequency compared to both CIMP-M and CIMP-L tumors (Table 1). This result agrees with some studies [4, 5, 8, 35] but not others that found an inverse association between KRAS mutation and CIMP+ [7, 9, 10].

Since BRAF mutations are strongly associated with CIMP and mutually exclusive to KRAS mutations ([10]; Fig. 1), a point of interest is whether methylation patterns differ between tumors with BRAF and KRAS mutations. Supervised clustering analysis with Bonferroni correction revealed that only 1 of the 202 tumor-specific CpG sites was differentially methylated between these tumor groups (HTR1B_P222_F, upregulated in BRAF mutant tumors, p = 8.1 × 10^-6). HTR1B (5-hydroxytryptamine (serotonin) receptor 1B) is a G protein-coupled multi-pass membrane protein involved in regulation of the serotonin system [37]. The gene is hypermethylated in lung cancer and its chromosome locus (16q14.1) is frequently deleted in a number of cancer types [38]. However, no links with BRAF or RAS mutations or signaling have been reported.

A novel finding of this array-based analysis was the existence of an apparent CIMP-M group (Fig. 1). These tumors displayed a higher frequency of EMVI compared to both CIMP-L and CIMP-H, and a significantly higher stage compared to CIMP-H (Table 1). CIMP-M tumors were almost exclusively located in the distal colon or rectum (12/13, 92%). MSI and KRAS and BRAF mutations were notably absent in these tumors, although this may be due to reportedly lower frequencies of these alterations in distal tumors [39]. Taken together, these results suggest CIMP-M tumors could be a distinct clinical and molecular entity, although confirmation in larger, independent tumor series is required.

After adjustment for multiple testing, 170 CpG sites were hypermethylated in CIMP-H compared to CIMP-L. The 112 genes containing these CpG sites are ranked according to significance in Additional file 5. Of these, 54 were previously reported as methylated in cancer, 38 as methylated in gastrointestinal cancers and 30 in CRC [Additional file 5]. Of the top 10 genes that were hypermethylated in CIMP-H compared to CIMP-L tumors, 5 have previously been implicated in the pathogenesis of gastrointestinal tumors (NTRK3, HS3ST2, TWIST1, CD40 and EYA4). Somatic mutation of NTRK3 has been reported in human colon cancer [40], while methylation of EYA4 has been documented previously in ulcerative colitis-associated dysplasia [41] and CRC [42].

CIMP-M tumors were found to have a relatively high incidence of EMVI (38%) compared to CIMP-H and CIMP-L tumors (Table 1). Supervised analysis revealed that HS3ST2, also known as 3-OST-2, was the only gene to be differentially methylated between tumors showing presence or absence of EMVI. Methylation-associated silencing of HS3ST2 expression has been demonstrated in breast, lung, pancreatic and colon cancers [43]. This gene encodes an enzyme that modifies heparin sulfate proteoglycans [44] involved in cell adhesion and migration [45], thus suggesting a possible mechanistic link between HS3ST2 methylation and EMVI.

The use of Illumina GoldenGate^® Beadarray technology in this study allowed a large number of CpG sites to be evaluated for methylation in an unbiased fashion. However, there are several limitations with this approach for the characterization of CIMP subgroups in CRC. Firstly, only a small fraction of all genes were investigated for methylation and in 70% of these just one CpG site per gene was evaluated. Secondly, it is unclear whether the methylation level at these sites relates to expression of the genes. Thirdly, some of the probes used in this assay contain single nucleotide polymorphisms (SNPs) or repetitive elements that could influence methylation analysis [25]. The cost effectiveness of using arrays to characterize CIMP-H is questionable, given the strong concordance between CIMP-H from this study and CIMP^w. Further studies should clarify if the additional information provided by methylation arrays is worth the complexity and expense.

Methylation profiling of 807 cancer-related genes revealed the presence of three CRC subgroups with distinct clinicopathological and molecular features. Similar to earlier studies that investigated fewer methylation markers, CIMP-H CRC were associated with older patient age, proximal location and mutations in BRAF and KRAS. Further investigations in large and independent population-based series are required to validate these findings and to assess the clinical utility of CIMP subgroups.