Genomics 103 (2014) 204–210

Contents lists available at ScienceDirect

Genomics

j ourna l homepage: www.e lsev ie r .com/ locate /ygeno

Direct ChIP-bisulfite sequencing reveals a role of H3K27me3 mediatingaberrant hypermethylation of promoter CpG islands in cancer cells

Fei Gao a,⁎,1, Guanyu Ji a,1, Zhaowei Gao a, Xu Han a, Mingzhi Ye a, Zhimei Yuan a, Huijuan Luo a,Xiaojun Huang a,e, Karthikraj Natarajan a, Jun Wang a,b,c,d, Huanming Yang a, Xiuqing Zhang a

a Science & Technology Department, BGI-Shenzhen, Shenzhen, 518083, Chinab Department of Biology, University of Copenhagen, Copenhagen, 2200, Denmarkc King Abdulaziz University, Jeddah, 21589, Saudi Arabiad The Novo Nordisk Foundation Center for Basic Metabolic Research, University of Copenhagen, 2200, Denmarke College of Life Sciences, Wuhan University, Wuhan, 430072, China

Abbreviations: CGI, CpG island; PRC, Polycomb-rephistone H3 lysine 27; TSSs, transcription start sites; COADprostate adenocarcinoma; STAD, stomach adenocarcinMillion reads.⁎ Corresponding author.

E-mail address: gaofei@genomics.cn (F. Gao).1 Both authors contributed equally to this work.

0888-7543/$ – see front matter © 2014 Elsevier Inc. All rihttp://dx.doi.org/10.1016/j.ygeno.2013.12.006

a b s t r a c t

a r t i c l e i n f o

Article history:Received 10 July 2013Accepted 28 December 2013Available online 7 January 2014

Keywords:DNA methylationChromatin immunoprecipitationCpG IslandGastric cancer

Themodel describing that aberrant CpG island (CGI)methylation leads to repression of tumour suppressor genesin cancers has been influential, but it remains unclear how such aberrancy is induced. Recent studies providedclues indicating that promoter hypermethylation in cancers might be associated with PRC target genes. Here,we used ChIP-BS-seq to examine methylation of the DNA fragments precipitated by the antibodies to bothH3K27me3 and H3K4me3 histone modifications. We showed that, for a set of genes highly enriched withH3K27me3 both in cancer and normal cells, CGI promoters were aberrantly hypermethylated only in cancercells in comparison with normal cells. In contrast, such aberrant CGI hypermethylation in cancer promotersthat were deficient of H3K27me3 was not notable. Furthermore, we confirmed that these genes wereconsistently hypermethylated in TCGA primary cancer cells. These works support the association betweenH3K27me3 andDNAmethylationmarks for specific cancer genes andwill spur futurework on combined histoneand DNA methylation that could define cancer's epigenetic abnormalities.

© 2014 Elsevier Inc. All rights reserved.

1. Introduction

Epigenetic information comprises histone modifications [1] and DNAmethylation [2] that can substantially influence chromatin structureand DNA accessibility [3,4]. In the last decade, the study of epigeneticmechanisms has been highlighted in cancer causation, progression andtreatment as an alternative for genetic defects [5]. Especially, aberrantDNA hypermethylation of CpG island (CGI) promoters is associatedwith transcriptional repression of many tumor suppressor genes thatcan lead to tumor progression in many cancers. Although this modelhas been hugely influential [6], the significance of hypermethylation atCGIs in cancer has long been debated as well [7,8]. Recently, a researchgroup has observed that genes that are hypermethylated and repressedin cancers were also repressed in pre-cancerous tissues even thoughtheir promoters are hypomethylated [9,10]. Another research group

ressive complex; H3K27me3,, colon adenocarcinoma; PRAD,oma; RPKM, Reads Per Kb per

ghts reserved.

found that many genes with de novo promoter hypermethylation incolon cancer were among the subset of genes that were bound withboth the repressingH3K27me3 and the activatingH3K4me3 in embryon-ic stem cells and adult stem/progenitor cells [11,12]. H3K27me3 iscatalyzed by the SET domain histone methyltransferase EZH2, a compo-nent protein of the Polycomb-repressive complex (PRC). It was knownthat genes that are targeted and repressed by PRC are also poised foractivation in pluripotent cells [13]. It was also discovered by usingchromatin immunoprecipitation in a previous study that binding ofDNA methyltransferases (DNMTs) to several EZH2-repressed genesdepended on the presence of EZH2 [14]. These observations suggestedthat promoter hypermethylation in cancers might be associated withPRC target genes. However, de novo DNA methylation without PRCoccupancy or de novo PRC occupancy without DNA methylation didexist as well [15].

Recently, a simple and effective new method termed ChIP-BS-seq(chromatin immunoprecipitation followed by bisulfite sequencing)was developed, enabling direct examination of the methylation statusof DNA sequences immunoprecipitated by ChIP for specific histonemodifications [16,17]. Therefore, we took advantage of this new ChIP-BS-seq technology and applied it to examine trimethylation of histoneH3 lysine 27 (H3K27me3) and histone H3 lysine 4 (H3K4me3) profilesfor a normal cell line (YH lymphoblastoid) and three cancer cell lines(one cervical cancer cell line (HeLa) and two gastric cancer (GC) cell

205F. Gao et al. / Genomics 103 (2014) 204–210

lines (BGC-823 and AGS)). Additionally, we also downloaded the ChIP-BS-seq data for a normal and two cancer cell lines from previous studies[16,17], thus, enabling examination of 7 cell lines in total.

By overall comparison between these cancer and normal cell lines, weshowed that a set of genes commonly enriched with H3K27me3 marksboth in cancer and normal cells presented aberrant hypermethylation ofpromoter CGIs in cancers, but hypomethylated state is maintained inthe normal cells. Gene ontology analyses suggested that these geneswere highly enriched in ion transport or cellular ion homeostasispathway, which were reported frequently in the study of carcinogenesisand cancer metastasis.

Furthermore, we obtained data from primary cancers in the CancerGenome Atlas (TCGA), and confirmed that some highly methylatedgenes in cancer cells were also significantly hypermethylated in TCGAdata for primary cancers. By combining our cell line methylation datawith TCGA data, we discovered new genes that possess significantly dif-ferent methylation patterns between cancer and normal tissues.

2. Materials and methods

2.1. Cell acquisition

A lymphoblastoid cell line was generated from a male Han Chineseindividual (YH), whose genome sequence was reported previously [18].Two human gastric cancer cell lines (BGC-823 and AGS) were providedby Beijing Tumor Hospital; the HeLa cell line was purchased from theAmerican Type Culture Collection (ATCC). Cells were cultured withRPMI1640 (Gibco C22400500BT) supplemented with 10% fetal bovineserum (Gibco 12657–029) in a humidified incubator with 5% CO2 at37 °C.

2.2. Chromatin immunoprecipitation sequencing (ChIP-seq)

Cells were precipitated by centrifugation, and the supernatant wasremoved. For each ChIP assay, approximately 5 × 106 cells were used.DNA and proteins were cross-linked with 1% formaldehyde in 10 mlPBS at 37 °C for 10 min, and then the cells were washed with pre-cooled PBS with 0.5% bovine serum and by PBS supplemented with pro-tease inhibitor compound (PIC). The cells were collected by centrifuga-tion at 850 rpm for 3 min after each wash. The cells were resuspendedin 200 μl ice-cold lyses buffer (1% SDS, 10 mM EDTA, 50 mM Tris–HCl,plus PIC) and then thawed on ice for 10 min to allow cell lyses. Thecell lysate was sonicated for 180 s using BioruptorTM 200 (pulses: 30 son/30 s off) to generate chromatin fragments of size range from 100 to700 bp. As input, one-tenth of the sonicated chromatin sample wasseparated. The remaining chromatin was immunoprecipitated in ChIPdilution buffer (1% Triton, 2 mM EDTA, 150 mM NaCl, 20 mM Tris–HCl)with 4 μg of antibody against H3K4me3 (Millipore; 17–614) orH3K27me3 (Millipore; 17–622) that has beenpre-incubatedwithproteinA/G magnetic beads (Invitrogen; 10003D). The immunoprecipitationreaction was incubated overnight at 4 °C and the beads were washedtwice with each of the following buffers at 4 °C: RIPA buffer (10 mMTris, 1 mM EDTA, 0.1% SDS, 0.1% Na-deoxycholate, 0.1% Triton X-100);RIPA buffer plus 0.3 M NaCl; LiCl buffer (0.25 M LiCl, 0.5% NP-40, 0.5%Na-deoxycholate) and TE buffer. The reaction tubewas placed in amag-netic rack to capture the beads. The bounded DNA was eluted from thebeads with elution buffer (1% SDS, 0.1 M NaHCO3) at 65 °C for 2–3 h.The same procedure was performed for the input sample. The immu-noprecipitated DNA was purified by phenol-chloroform extraction andprecipitated with ethanol and glycogen. Successful immunoprecipitationwas verified by qPCR using input as background. The obtained high-quality DNA was subjected to library preparation and sequenced onIllumina Hiseq 2000 using a standard pair-end 50 bp (PE50) sequencingprotocol.

2.3. Library construction for bisulfite-sequencing of ChIP-ed DNA

The ChIP DNA (60 ng) ends was repaired in a 100 μl reactioncontaining 1XT4 PNK buffer, 3 units T4 DNA polymerase (ENZYMATICS;P708), 0.5 unit Klenow enzyme (ENZYMATICS; P7060), 10 units T4-PNK(ENZYMATICS; Y9040) and dNTP (0.125 mM for each) for 30 min at20 °C. To the end repaired ChIP DNA, 15 unit Klenow fragment (3′–5′exo, ENZYMATICS; P7010-LC) in a 50 μl reaction containing 1X bluebuffer (ENZYMATICS; B011) and 0.2 mM dATP at 37 °C for 30 min togenerate protruding 3′Abase. Methylated pair-end adapters wereligated to the DNA fragments using 2400 unit of rapid T4 DNA ligase(ENZYMATICS; L6030-HC) at 37 °C for 15 min. After purification,200 ng exogenous λ-DNA fragments were added to the samples, andthe sodium bisulfite conversion assay (ZYMO D5006) was performed,followed by 16 cycles of PCR amplification that was consisted ofdenaturation (94 °C for 15 s), annealing (60 °C for 30 s), extension(72 °C for 30 s). Then, the PCR products were size-selected on 2%agarose gel, retaining 250–350 bp DNA fragments. The purified DNA(ChIP-BS libraries) was used for cluster generation and standard PE50sequencing using Illumina Hiseq 2000.

2.4. Processing of ChIP-seq reads

After PE50 sequencing, ChIP-seq readswere processed by the Illuminabase-calling pipeline. Low quality reads that containmore than 30% of ‘N’or over 10 % of the sequence with low quality per reads were omittedfrom the data analysis. The reads were alignedwith the human referencegenome (UCSC hg18) using SOAP (Short Oligonucleotide AnalysisPackage) 2.01 [19] with default parameters. Reads that were mapped tomore than one position in the genome were filtered out. Multiple readsmapping to the same position were counted once to avoid potentialbias from PCR. RSEG algorithm was applied for identification ofH3K27me3- and H3K4me3-enriched regions [20]. RSEG is based onhidden Markov model (HMM) framework including the Baum–Welchtraining and posterior decoding, modeling the read counts with a nega-tive binomial distribution. Subsequently, it uses a two-state HMM forsegmentation of the genome into fore-ground domains and backgrounddomains. To define an enriched region of H3K27me3 or H3K4me3,default RSEG settings was used based on bin size of 500 bp, includingthat the posterior probability of each bin obtained by HMM decoding islarger than 0.95 and that the mean of read counts within a region isabove the top 90th percentile of foreground emission distribution. Theadjacent enriched bins were merged.

2.5. Processing of ChIP-BS reads

For histone modification signal, same processing was performedas above for ChIP-seq reads. For DNA methylation signal, ChIP-BSsequencing reads were aligned to the human reference genome (UCSChg18) using an algorithm adapted from the procedure described by Listeret al. [21]. Because DNAmethylation has strand specificity, all cytosines inthe reference sequences (“original form”) were replaced in silico bythymines (“alignment form”) to allow alignment after bisulfite conver-sion. The “original forms” of the sequencing reads were transformed tocope with the BS treatment nucleotide conversion during the alignmentprocess using the following criteria: (1) observed cytosines in theforward read of each read pair were replaced by thymines in silico and(2) observed guanines in the reverse read of each read pairwere replacedby adenosines in silico. We then mapped the “alignment form” sequenc-ing reads to the “alignment form” reference sequences using SOAP 2.01[19] with default parameters. After mapping, the number of methylated(C) and unmethylated (T) basecalls at each CpG site within the genomewas used to determine themethylation status of each sequenced cytosinewithin a CpG context, both on the forward strand aswell as on the reversestrand. TheDNAmethylation level of each genomic regionwas defined asthe ratio of supported methylated reads with the sum of methylated and

206 F. Gao et al. / Genomics 103 (2014) 204–210

non-methylated reads. CGIs were defined as regions greater than 200 bpwith a GC fraction greater than 0.5 and an observed-to-expected ratio ofCpG greater than 0.65 as annotated in UCSC genome browser.

2.6. RNA-seq and data analysis

Total RNA was isolated from BGC-823, AGS cells and YH cells usingthe miRNeasy Kit (Qiagen 74104) according to the manufacturer'sprotocol. An additional DNase I digestion step was performed to ensurethat the samples were not contaminatedwith genomic DNA. ThemRNAwas enriched by using oligo(dT) magnetic beads. The mRNA wasfragmented into small fragments around 200 bp by mixing withfragmentation buffer. Then, the first strand of cDNA was synthesizedby using random hexamer-primer. The second strand was synthesizedby adding buffer, dNTPs, RNase H and DNA polymerase I. The doublestrand cDNA was purified with magnetic beads. End reparation, 3′-endaddition of adenine (A) nucleotide followed by ligation of adaptorsto the fragment was performed. The fragment was enriched by PCRamplification. The RNA purity was assessed using Agilent bioanalyzer.Total RNA was converted to cDNA using the NuGEN Ovation RNA-SeqSystem according to the manufacturer's protocol (NuGEN, San Carlos,CA, USA). The cDNA was used for Illumina sequencing library prepara-tion. DNA fragments were end-repaired to generate blunt ends with 5′phosphatase and 3′ hydroxyls, and adapters were ligated for paired-end sequencing on an Illumina Hiseq 2000. The reads were cleaned byremoving low quality data and 50 bp contamination. RNA sequencingreads were aligned to the reference genome (UCSC hg18) using theSOAP aligner with the same parameters that were used to processChIP sequencing reads. We collated a set of 27,071 RefSeq genes (USNational Center for Biotechnology Information, 20 February 2012update). Gene expression was calculated using RPKM (Reads Per Kbper Million reads) [22].

3. Results and discussions

3.1. Variable DNA methylation patterns for H3K27me3-enriched regions

Using ChIP-BS-seq technology, we first determined the H3K27me3status of a normal cell line (YH lymphoblastoid) and three cancer celllines (HeLa, BGC-823 and AGS), which are catalyzed by the SET domainhistone methyltransferase EZH2 and have a repressive function. TheChIP protocol was modified to obtain the DNA methylation status ofthe H3K27me3-bound DNA fragments by using 50-bp pair-endsequencing instead of conventional single-end sequencing. As a result,approximately 100 million uniquely mapped reads were obtained foreach of the original four cell lines (Table S1 in additional file 1). Wefound that CpG islands, (+/−500) transcription start sites (TSSs) andexons were preferentially enriched with H3K27me3 marks in two GCcell lines (BGC-823 and AGS). In YH cells, H3K27me3 marks were onlypreferentially enriched at CpG islands. In contrast, HeLa cells presenteda reverse pattern with highest enrichment of H3K27me3 in theintergenic regions (Fig. 1A). To confirm this result in HeLa cells and tocheck the reproducibility of the ChIP-BS-seq approach, we performedtwo independent replicates of ChIP-seq and ChIP-BS-seq and foundthat the read densities were correlated well (Pearson R2 = 0.997 forChIP-seq; Pearson R2 = 0.994 for ChIP-BS-seq) (Figures S1A and S1Bin additional file 2). Further analysis on a set of previously publishedChIP-BS-seq data confirmed the observation of distinct H3K27me3pattern in HeLa cells as well (Figure S2A in additional file 2) [23].

By applying a RSEG algorithm with default settings [20], wedefined H3K27me3-enriched genomic regions for each cell line(Table S2, Figure S3 in additional file 2). In order to guarantee an accuratemeasurement of DNA methylation, we calculated methylation levels forcytosine sites within H3K27me3-enriched regions, requiring at least 4×depth of deep sequencing. Thereby a high concordance between tworeplicates of HeLa samples (Pearson R2 = 0.914) was manifested

(Figure S1C in additional file 2). Similar to the histone modification pat-terns, we observed variable DNA methylation patterns for H3K27me3-enriched regions in these four cell lines (Fig. 1B). Thus, variable patternsof both H3K27me3 marks and DNA methylation in different cell popula-tions were revealed. To testmore cell types, we downloaded ChIP-BS-seqdata profiling H3K27me3 marks in another three cell lines (a prostatecancer cell line LNCaP, a normal prostate epithelial cell line PrEC and acolon cancer cell line HCT116). These data were previously reported, inwhich H3K27me3 pattern of HCT116 cell line was similar to HeLa but itwas different from LNCaP and PrEC [16,17] (Figure S2B in additional file2). We then used these data in the subsequent analyses.

3.2. H3K27me3-bound DNA from normal and cancer cell lines diverges inmethylation

Because of the variable patterns were observed in both H3K27me3marks and methylation of its bound DNA, an intriguing question isaroused whether cell-type-specific epigenetic signatures exist in thesecell lines. As majority of H3K27me3 marks were enriched in promoterscontaining CGIs in most cell lines, we focused our analysis on thepromoter regions: the upstream 2200 bp and the downstream 500 bpregions crossing the transcriptional start site (TSS) (as defined in[24]). In agreement with the variable DNA methylation patternsamong cell lines for all H3K27me3-enriched regions (Fig. 1B), the distri-bution curve for methylation levels of enriched promoters as well asexons, introns and intergenic regions showed divergence among thesefour cell lines too (Fig. 1C, Figure S4 in additional file 2). We furtherperformed clustering analysis for the seven cell lines based on theaverage values of cytosine methylation levels in all promoter regionsenriched with H3K27me3 marks. As the result, two normal cell lines(YH and PrEC) were clustered together and separated from five cancercell lines (Fig. 2A). Moreover, all these cell lines were originated fromdifferent cell lineages. Thus, despite the distinct cellular developmentof these cell lines, the above clustering results might suggest thatonco-epigenomic signatures of the DNA methylation bound to theH3K27me3 marks are different from normal epigenomic signatures.

3.3. Aberrant hypermethylation occurs inH3K27me3-enriched promoter-CGIregions in cancer cell lines

We next asked what characteristics are common for H3K27me3-bound DNA methylation among different cancer cell lines that differfrom normal cell lines. Thereby we first screened H3K27me3-enrichedgenomic regions thatwere overlapped across all seven cell lines, namely228 commonly enriched genes. We also screened genomic regionswhere no H3K27me3 marks were enriched across the seven cell lines,resulting 5143 corresponding genes. We then analyzed the RNA-seqdata of the four cell lines generated in our study (Table S1 in additionalfile 1) and the downloaded RNA-seq data for HCT116, LNCaP and PrECcell lines from previous studies [16,17]. By comparing the expressionlevels for these two categories of genes,we found that the transcriptionsof the 228 genes were highly repressed in comparison with the 5143genes (Figure S5 in additional file 2), indicating a strong repressioneffect on gene transcription exerted by H3K27me3 marks.

We further categorized these genes based on whether CGIs wereincorporated in their promoter or not, i.e. the upstream 2200 bp anddownstream 500 bp around the transcriptional start sites (TSS) [24].146 (64%) out of the 228 H3K27me3-enriched genes contained CGIsin their promoters. In comparison, this percentage number is smaller(45%, 2288 genes) for the 5143 H3K27me3-deficient genes. Sufficientcoverage can be achieved for H3K27me3-enriched genes, ensuring cor-rect examination of DNA methylation, but not for H3K27me3-deficientgenes. Next, we compared the methylation status between cancer andnormal cell lines with respect to presence of CGIs. A boxplot indicatedthat most of the CGIs were highly methylated in five cancer cell lines(median methylation level = 47.2%) than in two normal cell lines

010

2030

0.0

0.5

1.0

1.5

CGI TSS Exon Intron Intergenic

BGC_823 AGS YH HELA

Obs

erve

d/E

xpec

ted

Regions Of H3K27me3 EnrichmentA

B C

Methylation levels of enriched promoters0 0.2 0.4 0.6 0.8

Pro

port

ion

(%)

Methylation levels of all enriched regions

• •

•

•

••

•

•

•

•

1

010

2030

40

•

••

•

•

•

•

•

• •

•

•

•

•

•

•

•

•

•

•

•

•

•

•

•

•

•

••

•

•

•

•

•

•

•

•

•

•

•

•

•

•

•

•

•

•

•

•

•

•

•

•

•

••

•

•

•

•

•

•

•

•

•

•

•

•

•

•

AGS BGC_823 HeLaYH

0 0.2 0.4 0.6 0.8 1

Fig. 1. ChIP-BS-seq DNA methylation profiles of H3K27me3-enriched DNA from BGC-823, AGS, YH and HeLa cells. (A) Distribution of H3K27me3-enrichment genome-wide relativeto observed over expected. The expected reads are calculated in a region, e.g. CGI, as (the length of CGI/the length of whole genome)*(total reads). The observed reads are the real readcounts in this CGI region. (B) Distribution proportion of CpGmethylation levels at H3K27me3-marked regions from low (0%) to high (100%)methylation (0.0–1.0). (C) Distribution of CpGmethylation levels at H3K27me3-enriched promoters (TSS Upstream 2 k and downstream 500 bp) from low (0%) to high (100%) methylation (0.0–1.0).

207F. Gao et al. / Genomics 103 (2014) 204–210

(median methylation level = 10.7%) in the 146 genes. In contrast, the82 H3K27me3-enriched genes without CGIs promoters were not nota-bly different in both cancer and normal cell lines. The normal cell lineseven presented slightly higher methylation levels than cancer celllines (Fig. 2B).We further performedpair-wise comparison for the aver-age values of methylation levels of the 146 CGI-containing promoters,resulting in 127 promoters significantly hypermethylated in cancercells in comparison with normal cells, while only 1 promoter signifi-cantly hypomethylated (Chi-square test and Benjamini & HochbergFDR-adjusted p-value b 0.05).

We also screened out 98 genes and 58 genes that were enrichedwith H3K27me3 marks in their promoter CGIs only in two normal celllines and five cancer cell lines, respectively. Again, the methylationlevels for the 58 genes in cancer cells were significantly higher thanthe 98 genes in normal cells (Fig. 2C). These observations conflictedwith previous observation on mutual exclusion between H3K27me3and DNA methylation for CGIs [16]. Although their statement might belargely true for normal cells, we found that some genes enriched withH3K27me3 marks in diverse cancer cells were aberrantly methylated intheir promoter CGIs.

We also observed somedifference in the cancer-specificmethylationprofiles for the above 127 genes. Most of the genes were highly methyl-ated in AGS and HCT116 cell lines, but showed low methylation inLNCaP. In contrast, the BGC-823 and HeLa cell lines showed medianmethylation levels in the majority of these genes (Fig. 2D). Therefore,themethylation frequency of H3K27me3-boundCGIs varied in differentcancer cell populations. For each specific cancer, there might be only aset of H3K27me3-bound genes aberrantly hypermethylated in promot-er CGI, suggesting that other regulatory factors might be important inthis process. For instance, H3K9me3 might be another histone markthat converses with DNA methylation, probably also with H3K27me3[25–27]. Furthermore, recent advance indicated that both DNA

methylation and histone modifications can be directed and mediatedby small RNAs (sRNAs) [28,29], suggesting for potential sequence-specificity for the mediation on these epigenetic modifications. Despitesuch complexity of the crosstalk between histonemodification andDNAmethylation, these cancer-specific methylation patternsmight raise thepossibility of usingmethylated DNA bound to the specific histone mod-ifications as an epigenetic signature in cancer studies.

3.4. Hypermethylation occurred more frequently in CGIs only bound withH3K27me3 but not with H3K4me3

We further performed ChIP-BS-seq to assess the presence of theactivating H3K4me3 in YH, AGS and BGC-823 cell lines. As expected,the H3K4me3 marks were mostly located around the transcriptionstart sites (TSSs) of known genes in all these cell lines (Figure S6A inadditional file 2). At the genomic regions enriched with H3K4me3marks, the DNA methylation levels were extremely low. This resultwas not surprising as previous studies supported a strong negativecorrelation between DNA methylation and the presence of H3K4me inseveral cell types [30,31]. We also found that H3K4me3 marks werenegatively correlated with DNA methylation levels and positivelycorrelated with the expression of genes that were richly marked byH3K4me3 (Figure S6B in additional file 2). Our results reinforcedthe observations of antagonistic relationship between H3K4me3 andDNA methylation for the regulation of gene expression by detectingmethylation status of DNA bound with histone modification marks.

A wealth of evidences suggest that PcG-repressed genes are generallymarked by both H3K27me3 and H3K4me3 modifications [32–34] andin addition to stem cells, bivalent-modified domains can also be foundin differentiated cells and cancer cells [35]. Here, we also found thatthe co-occurrence of H3K4me3 and H3K27me3 marks in genic regionsis prevalent in our two cancer cell lines and the normal cell line

BG

C_8

23

HeL

a

AG

S

HC

T11

6 LNC

aP

YH

PrE

C

0.25

0.35

0.45

0.55

Hei

ght

100

100

100

100

100

au

100

100

100

100

100

bp

1

2

3

4

5

edge #

A

DC

PrE

C

YH

LnC

aP

HC

T11

6

AG

S

HeL

a

BG

C-8

23

0 1Value

Color Key

B

Cancer Normal Cancer Normal

0.0

0.2

0.4

0.6

0.8

1.0

Met

hyla

tion

Leve

l

CGI CGI Promoter-

******

0.0

0.2

0.4

0.6

0.8

1.0

NormalCancer

Met

hyla

tion

Leve

l

***

Fig. 2. Hypermethylation occurs inH3K27me3-enriched CGI promoters in cancer cells. (A) Hierarchical clustering of DNAmethylation profiles for five cancer and two normal cell linesby Pvclust R package. Two types of p-values (%) were provided, AU (approximately unbiased) p-values are computed by multiscale bootstrap resampling and shown in red, while BP(bootstrap probability) p-values are computed by normal bootstrap resampling and shown in green. Clusters with AU larger than 95% are strongly supported by the data. 1000 timesof bootstrapping was applied. (B) The CGIs of the 146 genes containing CGIs in their promoters were hypermethylated in cancer versus normal cells. In contrast, the 82 genes withoutCGIs showed opposite trend. (C) CGI methylation levels of 58 genes enriched with H3K27me3 only in five cancer cells were significantly higher than the 98 genes enriched withH3K27me3 marks only across two normal cells. Both in (B) and (C), statistical significance is indicated by triple asterisks (t-test p-value b 0.001). (D) Heatmap of average methylationlevels of promoter CGIs showing cancer-specific methylation profiles for 127 genes.

208 F. Gao et al. / Genomics 103 (2014) 204–210

((Table S2 in additional file 1)). Therefore we categorized all knowngenes in RefSeq into four categories: 1) only H3K4me3 highly enriched(H3K4me3 high), 2) only H3K27me3 highly enriched (H3K27me3high), 3) both H3K4me3 and H3K27me3 highly enriched (Bivalent)and 4) neither H3K4me3 nor H3K27me3 highly enriched (Table S2 inadditional file 1).

We found that themethylation levels of DNA bound with H3K27me3marks within promoters were considerably divergent among the fourcategories of genes. On average, the DNA methylation levels aroundTSS remained in a hypomethylated state in the “H3K4me3 high” or the“Bivalent” category genes both in cancer and normal cells. But in cancercells, “H3K27me3 high” category presented a much higher level of DNAmethylation than the other two categories, while in the normal YH cellsthe methylation level was limited to a relatively low level (less than20%) (Fig. 3A), suggesting much less genes were hypermethylated inH3K27me3-bound regions.

We then examined the distribution of methylation levels of all CGI-containing promoters in comparison with all CGI-deficient promotersto the amount of methylated DNA bound to H3K27me3. Comparisonof these two promoter types showed that hypermethylation primarilyoccurred in the CGI-containing promoters in the cancer cells, whilein YH cells we found that the CGI-containing promoters were morecommonly hypomethylated (Fig. 3B). Thus, this result again supportedthe notion that the increased methylation of H3K27me3 enriched

region at some promoters in cancer cells was correlated with hyperme-thylation that occurred in these CGI regions.

3.5. H3K27me3-bound genes with CGI promoter hypermethylation in celllines were also hypermethylated in primary cancer tissues

Taken above results together, CGI hypermethylation is associatedwith H3K27me3 marks for some genes in cancer cells. However, theCGI hypermethylation did not exert considerably further repressiveeffects on these genes that were already repressed by H3K27me3marks (Figure S7 in additional file 2). What role the CGI hypermethyla-tion had played in the cancer cells is not clear in this stage. However, thepresence of DNA methylation can attract specialized methyl-DNAbinding factors that can then recruit chromatin modifiers [36], thusmay change the cell fate extensively. We further performed GeneOntology analyses on the 127 genes to see whether these genes wereenriched in particular pathways. As a result, most of these genes wereannotated as genes encoding membrane protein and were enriched inion transport or cellular ion homeostasis pathways (Fisher test and BHadjusted p-value b 0.05) (Table S3 in additional file 1), which are highlyrelevant to carcinogenesis or metastasis as previously reported [37,38].

Then, by looking at the genes that had 65% or greater methylationlevel in at least 4 of the 5 cancer cell lines, we found 12 protein-encoding genes that were relatively conserved for hypermethylation

YH AGS

H3k4me3 only Bivalent H3k27me3 only

BGC−823

Met

hyla

tion

leve

l

00.

20.

40.

60.

8

+2k +1k TSS -1k -2k

CGI CGI

Met

hyla

tion

leve

l

+Promoter- Promoter

A

B

00.

20.

40.

60.

81

YH AGSBGC−823

+2k +1k TSS -1k -2k +2k +1k TSS -1k -2k

CGI CGI+Promoter- Promoter CGI CGI+Promoter- Promoter

*** *** ***

*********

Fig. 3. Distribution of methylation levels at promoters in BGC-823, AGS and YH cells. (A) Distribution of average methylation levels detected by H3K27me3-BS in promoters of eachcategory of genes in three cell lines. The genes were categorized into “H3K27me3 high”, “H3K4me3 high”, “Bivalent” or “Neither” genes, the “Neither” category was not shown. At theleft bottom, box plots show distribution of methylation levels for “H3K27me3 high” and “Bivalent” categories with same color. (B) Distribution of methylation levels of CGI-containing(CGI+) promoters in comparison with CGI-deficient (CGI-) promoters of H3K27me3 high genes in three cell lines. The bold line is the median value of methylation levels, and theerror bars represent the interquartile range above and below the mean value. Statistical significance is indicated by triple asterisks (t-test p-value b 0.001).

209F. Gao et al. / Genomics 103 (2014) 204–210

in different cancer types (Table S4 in additional file 1). ZRANB2-AS1,CR1L, CH25H and CCDC140 were highly methylated in cell lines ofcolon (HCT116), prostate (LNCaP) and stomach (both BGC-823 andAGS). Other 8 genes (SHE, POU4F1, SYT14, GLT25D2 (COLGALT2),LHFPL4, NPAS4, PTPRN and NPTX2) were highly methylated in colonand stomach cancer cell lines, but only moderately methylated in pros-tate cancer cell line LNCaP. As these 8 genes were nearly non-methylated in normal prostate PrEC cell line, the possibility of therebeing specificmethylation differences between primary prostate cancerand normal tissues was still raised (Table S4 in additional file 1).

Due to the heterogeneity of cancer tissues, diverse cell populationsin cancer tissue might vary in the DNA methylation status. However,we reasoned that, if the cancer cells comprised of major populationthat contain genes with hypermethylated DNA bound to H3K27me3mark in primary cancer tissues, it could lead to an outcome of abnormalhypermethylation in this cancer tissue in comparison with normaltissues. We thus downloaded methylation data for primary colonadenocarcinoma (COAD), prostate adenocarcinoma (PRAD), stomachadenocarcinoma (STAD) and from their paired normal controls fromThe Cancer Genome Atlas (TCGA) (Table S5 in additional file 1), inorder to expand our observations from cancer cell lines to primarycancer tissues. By comparing our cell line methylation data to that ofTCGA patient data, we indeed found that majority of the 12 genes aresignificantly hypermethylated in primary cancers. The methylationstatus was generally consistent between our cell lines and TCGA primarycancer results. (Figure S8 in additional file 2).

We also assessed data on themethylation status of these genes fromliterature. Of note, aberrant methylation of POU4F1, PTPRN, LHFPL4 andNPTX2 genes was previously reported in breast cancer [39], ovariancancer [40], cervical cancer [41] andpancreatic cancer [42], respectively.Among the rest of the genes that haven't been reported as havingaberrant methylation in cancer, CCDC140, which also hasn't beenshown to have strong association with cancer, had a very consistentand significant hypermethylation in the TCGA cancer data as comparedwith normal controls; raising a high possibility that aberrant methyla-tion of this gene may be important in cancer as well. For the remaininggenes, therewas also indication that theymight be aberrantlymethylat-ed in TCGA primary cancers. However, given the cell complexity ofprimary cancer tissues, current results from TCGA data might be biased.

Further studies applying ChIP-BS-seq technology are needed to com-pare primary cancer with adjacent normal tissues to further confirmthe observations of aberrant methylation for these genes in cancer celllines.

4. Conclusions

Current study by using ChIP-BS-seq revealed the correlationbetween H3K27me3 and DNA methylation marks in diverse cancercells, as some genes enriched with H3K27me3 marks can be aberrantlymethylated in their promoter CGIs in comparison with normal cells.Despite of the complexity of the crosstalk between histonemodificationand DNA methylation, the H3K27me3-bound hypermethylated CGIs inspecific genes might be raised as potential new epigenetic signature infuture cancer studies on biomarker discovery and targeted therapeutics.

Acknowledgments

We thank Laurie Goodman for help in editing the manuscript andthank Youyong Lv for kindly providing the two gastric cancer cell lines.This study was supported by grants from the National High TechnologyResearch and Development Program of China (863 Program, No.2009AA022707).

Competing interests

The authors declare that they have no competing interests.

Source of downloaded data

Extra data from previously published sources were included in ouranalysis: (a) H3K27me3 ChIP-seq data for HeLa cell line (GSM325898)[23]. (b) H3K27me3 ChIP-BS-seq data for HCT116 (GSM699327) [16],LNCaP (GSM758360) and PrEC (GSM758361) [17] cell lines. (c) RNA-seq data for HeLa (GSM958739), HCT116 (GSM699333) [16], LNCaPand PrEC (GSE29155) [17] cell lines. (d) Infinium methylation arraydata for primary colon adenocarcinoma (COAD), prostate adenocarcino-ma (PRAD), stomach adenocarcinoma (STAD) and paired normal

210 F. Gao et al. / Genomics 103 (2014) 204–210

controls were downloaded from the Cancer Genome Atlas (TCGA;http://tcga.cancer.gov/dataportal; file names are provided in Table S5in additional file 1).

Data access

All data generated by ChIP-BS-Seq and RNA-seq were deposited inthe National Center for Biotechnology Information Gene ExpressionOmnibus database (GEO accession number GSE43096). All datagenerated by liquid hybridization capture-based bisulfite sequencing(LHC-BS) technology for AGS, BGC-823 and YH cell lineswere depositedin GEO (GSE44866).

Author contributions

FG conceived and supervised the study, interpreted data, and wrotethe manuscript. ZWG, XH, MZY, ZMY, HJL, XJH and KN performedexperiments. GYJ conducted bioinformatics analysis. JW, HMY andXQZ helped in data interpretation and manuscript revision. All authorsread and approved the final manuscript.

Appendix A. Supplementary data

Supplementary data to this article can be found online at http://dx.doi.org/10.1016/j.ygeno.2013.12.006.

References

[1] T. Kouzarides, Chromatin modifications and their function, Cell 128 (2007) 693–705.[2] A. Bird, DNAmethylation patterns and epigeneticmemory, GenesDev. 16 (2002) 6–21.[3] M.G. Goll, T.H. Bestor, Eukaryotic cytosine methyltransferases, Ann. Rev. Biochem.

74 (2005) 481–514.[4] R. Margueron, P. Trojer, D. Reinberg, The key to development: interpreting the histone

code? Curr. Opin. Genet. Dev. 15 (2005) 163–176.[5] M. Berdasco, M. Esteller, Aberrant epigenetic landscape in cancer: how cellular identity

goes awry, Dev. Cell 19 (2010) 698–711.[6] P.A. Jones, S.B. Baylin, The epigenomics of cancer, Cell 128 (2007) 683–692.[7] T.H. Bestor, Unanswered questions about the role of promoter methylation in

carcinogenesis, Ann. N. Y. Acad. Sci. 983 (2003) 22–27.[8] A.M. Deaton, A. Bird, CpG islands and the regulation of transcription, Genes Dev. 25

(2011) 1010–1022.[9] D. Sproul, C. Nestor, J. Culley, J.H. Dickson, J.M. Dixon, D.J. Harrison, R.R. Meehan, A.H.

Sims, B.H. Ramsahoye, Transcriptionally repressed genes become aberrantlymethylated and distinguish tumors of different lineages in breast cancer, Proc.Natl. Acad. Sci. U. S. A. 108 (2011) 4364–4369.

[10] D. Sproul, R.R. Kitchen, C.E. Nestor, J.M. Dixon, A.H. Sims, D.J. Harrison, B.H.Ramsahoye, R.R. Meehan, Tissue of origin determines cancer-associated CpG islandpromoter hypermethylation patterns, Genome Biol. 13 (2012) R84.

[11] J.E. Ohm, K.M. McGarvey, X. Yu, L. Cheng, K.E. Schuebel, L. Cope, H.P. Mohammad,W.Chen, V.C. Daniel, W. Yu, et al., A stem cell-like chromatin pattern may predisposetumor suppressor genes to DNA hypermethylation and heritable silencing, Nat.Genet. 39 (2007) 237–242.

[12] H. Easwaran, S.E. Johnstone, L. Van Neste, J. Ohm, T. Mosbruger, Q.Wang, M.J. Aryee,P. Joyce, N. Ahuja, D. Weisenberger, et al., A DNA hypermethylation module for thestem/progenitor cell signature of cancer, Genome Res. 22 (2012) 837–849.

[13] B.E. Bernstein, T.S. Mikkelsen, X. Xie, M. Kamal, D.J. Huebert, J. Cuff, B. Fry, A.Meissner, M. Wernig, K. Plath, et al., A bivalent chromatin structure marks keydevelopmental genes in embryonic stem cells, Cell 125 (2006) 315–326.

[14] E. Vire, C. Brenner, R. Deplus, L. Blanchon, M. Fraga, C. Didelot, L. Morey, A. VanEynde, D. Bernard, J.M. Vanderwinden, et al., The Polycomb group protein EZH2directly controls DNA methylation, Nature 439 (2006) 871–874.

[15] E.N. Gal-Yam, G. Egger, L. Iniguez, H. Holster, S. Einarsson, X. Zhang, J.C. Lin, G. Liang,P.A. Jones, A. Tanay, Frequent switching of Polycomb repressive marks and DNAhypermethylation in the PC3 prostate cancer cell line, Proc. Natl. Acad. Sci. U. S. A.105 (2008) 12979–12984.

[16] A.B. Brinkman, H. Gu, S.J. Bartels, Y. Zhang, F. Matarese, F. Simmer, H. Marks, C. Bock,A. Gnirke, A. Meissner, H.G. Stunnenberg, Sequential ChIP-bisulfite sequencingenables direct genome-scale investigation of chromatin and DNA methylationcross-talk, Genome Res. 22 (2012) 1128–1138.

[17] A.L. Statham, M.D. Robinson, J.Z. Song, M.W. Coolen, C. Stirzaker, S.J. Clark,Bisulphite-sequencing of chromatin immunoprecipitated DNA (BisChIP-seq) directlyinforms methylation status of histone-modified DNA, Genome Res. 22 (2012)1120–1127.

[18] G. Li, L. Ma, C. Song, Z. Yang, X. Wang, H. Huang, Y. Li, R. Li, X. Zhang, H. Yang, J.Wang, The YH database: the first Asian diploid genome database, Nucleic AcidsRes. 37 (2009) D1025–D1028.

[19] R. Li, C. Yu, Y. Li, T.W. Lam, S.M. Yiu, K. Kristiansen, J. Wang, SOAP2: an improvedultrafast tool for short read alignment, Bioinformatics 25 (2009) 1966–1967.

[20] Q. Song, A.D. Smith, Identifying dispersed epigenomic domains from ChIP-Seq data,Bioinformatics 27 (2011) 870–871.

[21] R. Lister, M. Pelizzola, R.H. Dowen, R.D. Hawkins, G. Hon, J. Tonti-Filippini, J.R. Nery,L. Lee, Z. Ye, Q.M. Ngo, et al., Human DNA methylomes at base resolution showwidespread epigenomic differences, Nature 462 (2009) 315–322.

[22] A. Mortazavi, B.A. Williams, K. McCue, L. Schaeffer, B. Wold, Mapping andquantifying mammalian transcriptomes by RNA-Seq, Nat. Methods 5 (2008)621–628.

[23] S. Cuddapah, R. Jothi, D.E. Schones, T.Y. Roh, K. Cui, K. Zhao, Global analysis of theinsulator binding protein CTCF in chromatin barrier regions reveals demarcationof active and repressive domains, Genome Res. 19 (2009) 24–32.

[24] M. Weber, I. Hellmann, M.B. Stadler, L. Ramos, S. Paabo, M. Rebhan, D. Schubeler,Distribution, silencing potential and evolutionary impact of promoter DNA methyl-ation in the human genome, Nat. Genet. 39 (2007) 457–466.

[25] F. Fuks, DNA methylation and histone modifications: teaming up to silence genes,Curr. Opin. Genet. Dev. 15 (2005) 490–495.

[26] P.L. Severson, E.J. Tokar, L. Vrba, M.P. Waalkes, B.W. Futscher, Coordinate H3K9 andDNA methylation silencing of ZNFs in toxicant-induced malignant transformation,Epigenetics (2013) 8.

[27] C.Moison, C. Senamaud-Beaufort, L. Fourriere, C. Champion, A. Ceccaldi, S. Lacomme, A.Daunay, J. Tost, P.B. Arimondo, A.L. Guieysse-Peugeot, DNAmethylation associatedwithpolycomb repression in retinoic acid receptor beta silencing, FASEB J. 27 (2013)1468–1478.

[28] M. Gonzalez, F. Li, DNA replication, RNAi and epigenetic inheritance, Epigenetics 7(2012) 14–19.

[29] S.A. Simon, B.C. Meyers, Small RNA-mediated epigenetic modifications in plants,Curr. Opin. Plant Biol. 14 (2011) 148–155.

[30] D. Jia, R.Z. Jurkowska, X. Zhang, A. Jeltsch, X. Cheng, Structure of Dnmt3a bound toDnmt3L suggests a model for de novo DNA methylation, Nature 449 (2007)248–251.

[31] S.K. Ooi, C. Qiu, E. Bernstein, K. Li, D. Jia, Z. Yang, H. Erdjument-Bromage, P. Tempst,S.P. Lin, C.D. Allis, et al., DNMT3L connects unmethylated lysine 4 of histone H3 to denovo methylation of DNA, Nature 448 (2007) 714–717.

[32] T.S. Mikkelsen, M. Ku, D.B. Jaffe, B. Issac, E. Lieberman, G. Giannoukos, P.Alvarez, W. Brockman, T.K. Kim, R.P. Koche, et al., Genome-wide maps ofchromatin state in pluripotent and lineage-committed cells, Nature 448(2007) 553–560.

[33] G. Pan, S. Tian, J. Nie, C. Yang, V. Ruotti, H. Wei, G.A. Jonsdottir, R. Stewart, J.A.Thomson, Whole-genome analysis of histone H3 lysine 4 and lysine 27methylationin human embryonic stem cells, Cell Stem Cell 1 (2007) 299–312.

[34] X.D. Zhao, X. Han, J.L. Chew, J. Liu, K.P. Chiu, A. Choo, Y.L. Orlov,W.K. Sung, A. Shahab,V.A. Kuznetsov, et al., Whole-genome mapping of histone H3 Lys4 and 27trimethylations reveals distinct genomic compartments in human embryonic stemcells, Cell Stem Cell 1 (2007) 286–298.

[35] X.S. Ke, Y. Qu, K. Rostad, W.C. Li, B. Lin, O.J. Halvorsen, S.A. Haukaas, I. Jonassen, K.Petersen, N. Goldfinger, et al., Genome-wide profiling of histone h3 lysine 4 andlysine 27 trimethylation reveals an epigenetic signature in prostate carcinogenesis,PloS One 4 (2009) e4687.

[36] B.A. Buck-Koehntop, P.A. Defossez, On how mammalian transcription factorsrecognize methylated DNA, Epigenetics 8 (2013) 131–137.

[37] S.F. Pedersen, C. Stock, Ion channels and transporters in cancer: Pathophysiology,regulation and clinical potential, Cancer Res. 73 (2013) 1658–1661.

[38] N. Prevarskaya, R. Skryma, Y. Shuba, Targeting Ca(2+) transport in cancer: closereality or long perspective? Expert Opin. Ther. Targets 17 (2013) 225–241.

[39] M. Faryna, C. Konermann, S. Aulmann, J.L. Bermejo, M. Brugger, S. Diederichs, J. Rom,D. Weichenhan, R. Claus, M. Rehli, et al., Genome-wide methylation screen inlow-grade breast cancer identifies novel epigenetically altered genes as potentialbiomarkers for tumor diagnosis, FASEB J. 26 (2012) 4937–4950.

[40] D.O. Bauerschlag, O. Ammerpohl, K. Brautigam, C. Schem, Q. Lin, M.T. Weigel, F.Hilpert, N. Arnold, N. Maass, I. Meinhold-Heerlein, W. Wagner, Progression-freesurvival in ovarian cancer is reflected in epigenetic DNA methylation profiles,Oncology 80 (2011) 12–20.

[41] S.S. Wang, D.J. Smiraglia, Y.Z. Wu, S. Ghosh, J.S. Rader, K.R. Cho, T.A. Bonfiglio, R.Nayar, C. Plass, M.E. Sherman, Identification of novel methylation markers in cervi-cal cancer using restriction landmark genomic scanning, Cancer Res. 68 (2008)2489–2497.

[42] K. Brune, S.M. Hong, A. Li, S. Yachida, T. Abe, M. Griffith, D. Yang, N. Omura, J.Eshleman, M. Canto, et al., Genetic and epigenetic alterations of familial pancreaticcancers, Cancer Epidemiol. Biomarkers Prev. 17 (2008) 3536–3542.