The human colon cancer methylome shows similar
hypo- and hypermethylation at conserved tissue-specific
CpG island shores
Rafael A Irizarry1,2,8, Christine Ladd-Acosta2,3,8, Bo Wen2,3, Zhijin Wu4, Carolina Montano2,3,
Patrick Onyango2,3, Hengmi Cui2,3, Kevin Gabo2,3, Michael Rongione2,3, Maree Webster5, Hong Ji2,3,
James B Potash2,6, Sarven Sabunciyan2,7 & Andrew P Feinberg2,3,8
For the past 25 years, it has been known that alterations in DNA methylation (DNAm) occur in cancer, including
hypomethylation of oncogenes and hypermethylation of tumor suppressor genes. However, most studies of cancer methylation
have assumed that functionally important DNAm will occur in promoters, and that most DNAm changes in cancer occur in CpG
islands. Here we show that most methylation alterations in colon cancer occur not in promoters, and also not in CpG islands, but
in sequences up to 2 kb distant, which we term ‘CpG island shores’. CpG island shore methylation was strongly related to gene
expression, and it was highly conserved in mouse, discriminating tissue types regardless of species of origin. There was a notable
overlap (45–65%) of the locations of colon cancer–related methylation changes with those that distinguished normal tissues, with
hypermethylation enriched closer to the associated CpG islands, and hypomethylation enriched further from the associated CpG
island and resembling that of noncolon normal tissues. Thus, methylation changes in cancer are at sites that vary normally in
tissue differentiation, consistent with the epigenetic progenitor model of cancer, which proposes that epigenetic alterations
affecting tissue-specific differentiation are the predominant mechanism by which epigenetic changes cause cancer.
Since the discovery of altered DNA methylation in human cancer1, the
focus has largely been on specific genes of interest and regions assumed
to be important functionally, such as promoters and CpG islands2,3,
and there has not been a comprehensive genome-scale understanding
of the relationship between DNA methylation loss and gain in cancer
and in normal differentiation. In these experiments, we focused on
three key questions. First, where are DNA methylation changes that
distinguish tissue types? Taking a comprehensive genome-wide
approach, we examined three normal tissue types representing the
three embryonic lineages—liver (endodermal), spleen (mesodermal)
and brain (ectodermal)—obtained from five autopsies. A difference
from previous methylation studies of tissues, besides the genome-wide
design, is that here tissues were obtained from the same individual,
thus controlling for potential interindividual variability. Second, where
are DNAm alterations in cancer, and what is the balance between hypomethylation
and hypermethylation? For this purpose, we examined 13
colorectal cancers and matched normal mucosa from the subjects.
Third, what is the functional role of these methylation changes? To this
end, we carried out a comparative epigenomics study of tissue
methylation in the mouse, as well as gene expression analyses.
To examine DNAm on a genome-wide scale, we carried out
comprehensive high-throughput array-based relative methylation
(CHARM) analysis, which is a microarray-based method agnostic to
preconceptions about DNAm, including location relative to genes and
CpG content4. The resulting quantitative measurements of DNAm,
denoted with M, are log ratios of intensities from total (Cy3) and
McrBC-fractionated DNA (Cy5): positive and negative M values are
quantitatively associated with methylated and unmethylated sites,
respectively. For each sample, we analyzed B4.6 million CpG sites
across the genome using a custom-designed NimbleGen HD2
microarray, including all of the classically defined CpG islands as
well as all nonrepetitive lower CpG density genomic regions of
the genome. We included 4,500 control probes to standardize these
M values so that unmethylated regions were associated, on average,
with values of 0. CHARM is 100% specific at 90% sensitivity for
known methylation marks identified by other methods (for example,
in promoters) and includes the approximately half of the genome not
identified by conventional region preselection4. The CHARM results
were also extensively corroborated by quantitative bisulfite pyrosequencing
analysis.
Received 27 October 2008; accepted 7 November 2008; published online 18 January 2009; doi:10.1038/ng.298
1Department of Biostatistics, Johns Hopkins Bloomberg School of Public Health, Baltimore, Maryland 21205, USA. 2Center for Epigenetics, Institute for Basic
Biomedical Sciences and 3Department of Medicine, Johns Hopkins University School of Medicine, Baltimore, Maryland 21205, USA. 4Center for Statistical Sciences,
Brown University, Providence, Rhode Island, USA. 5Stanley Laboratory of Brain Research, Uniformed Services University of Health Sciences, Bethesda, Maryland
20892, USA. 6Departments of Psychiatry and 7Pediatrics, Johns Hopkins University School of Medicine, Baltimore, Maryland 21205, USA. 8These authors contributed
equally to this work. Correspondence should be addressed to R.A.I. (rafa@jhu.edu) or A.P.F. (afeinberg@jhu.edu).
178 VOLUME 41 [ NUMBER 2 [ FEBRUARY 2009 NATURE GENETICS
A R T I C L E S©2009NatureAmerica,Inc.Allrightsreserved.
RESULTS
Most tissue-specific DNAm occurs in ‘CpG island shores’
Because CHARM is not biased for CpG island or promoter sequences,
we could obtain objective data on tissue-specific methylation. We
identified 16,379 tissue differential methylation regions (T-DMRs),
defined as regions with M values for one tissue consistently different
than that for the others at a false discovery rate (FDR) of 5% (see
Methods). The median size of a T-DMR was 255 bp. Previous studies
of tissue- or cancer-specific DNAm have focused on promoters and/or
CpG islands, which have been defined as regions with a GC fraction
greater than 0.5 and an observed-to-expected ratio of CpG greater
than 0.6 (refs. 2,5). It has previously been reported that the degree of
differences in DNAm of promoters in somatic cells is relatively low in
conventionally defined CpG islands and higher at promoters with
intermediate CpG density6,7. Two recent studies identified a relatively
small fraction, 4–8%, of CpG islands with tissue-specific methylation8,9.
We also found that DNAm variation is uncommon in CpG
islands (Supplementary Fig. 1 online).
The genome-wide approach of CHARM also enabled us to find an
unexpected physical relationship between CpG islands and DNAm
variation, namely that 76% of T-DMRs were located within 2 kb of
islands in regions we now denote as ‘CpG island shores’. For example,
for the T-DMR in the PRTFDC1 gene, which encodes a brain-specific
phosphoribosyltransferase that is relatively hypomethylated in the
brain, the spreading of M values among the tissues begins B200 bp
from the CpG island and at a point where the CpG density associated
with the island has fallen to 1/10 the density in the island itself
(Fig. 1). The association of T-DMRs with CpG island shores is not due
to an arbitrary definition of CpG islands but to a true association of
these DNAm differences near but not in the regions of dense CpG
content (Supplementary Data 1 online describes all T-DMR regions,
and plots similar to those in Fig. 1 for the complete set of T-DMRs,
ordered by statistical significance, are available online; see URLs
section in Methods). The distribution of T-DMRs by distance from
the respective islands shows that DNAm variation is distributed over a
B2 kb shore, and that although CpG islands are enriched on the
arrays, because of their high CpG content (33% of CHARM probes are
in islands), only 6% of T-DMRs are in islands, compared to 76% in
shores; an additional 18% of T-DMRs were located greater than 2 kb
from the respective islands (Fig. 2). The localization of T-DMRs also
occurred largely outside of promoters (96%), as CpG islands are
localized primarily within promoters10. Furthermore, more than half
(52%) of T-DMRs were greater than 2 kb from the nearest annotated
gene. The distribution of the distance to islands remained essentially
unchanged when we used FDR cutoffs of 0.01, 0.05 and 0.10 (data
not shown).
We confirmed the array-based result that the differential methylation
was in CpG island shores rather than in the associated islands by
carrying out bisulfite pyrosequencing analysis on over 100 CpG sites
in the islands and shores associated with four genes, three T-DMRs
and one cancer differential methylation region. At all 101 sites, the
DMR was confirmed to lie within the shore rather than the island
(Supplementary Table 1 online). For example, PCDH9, which
encodes a brain-specific protocadherin, was relatively hypomethylated
in the brain at all 6 sites examined in the CpG island shore but
unmethylated in both brain and spleen at all 18 sites examined in the
associated island (Supplementary Table 1). Differential methylation
of an additional four CpG island shores was also confirmed by
bisulfite pyrosequencing of 39 total CpG sites, and all showed
statistically significant differences in DNAm (P o 0.05) (Supplementary
Table 2 online). These data verify the sensitivity of CHARM for
detecting subtle differences in DNAm. Furthermore, they confirm that
most normal differential methylation takes place at CpG island shores.
Similar CpG island shore hypo- and hypermethylation in cancer
We used the same comprehensive genome-wide approach to address
cancer-specific DNA methylation. We focused on colorectal cancer, a
paradigm for cancer epigenetics because of the availability of subjectmatched
normal mucosa, the cell type from which the tumors arise.
We analyzed DNAm on 13 colon cancers and matched normal mucosa
from the same individuals, identifying 2,707 regions showing differential
methylation in cancers (C-DMRs) with an FDR of 5% (Supplementary
Data 2 online, and plots similar to those in Fig. 1 for the
complete set of C-DMRs, ordered by statistical significance, are
available online; see URLs section in Methods). These C-DMRs
were similarly divided between those showing hypomethylation in
the cancer (compared to the normal colon) and those showing
hypermethylation (1,199 (44%) and 1,508 (56%), respectively). The
CHARM arrays, like other tiling arrays, do not contain repetitive
sequences, so the abundance of hypomethylation is not due to
Figure 1 Most tissue-specific differential DNA
methylation is located at CpG island shores.
(a) An example of a T-DMR located at a CpG
island shore in the PRTFDC1 gene. The upper
panel is a plot of M value versus genomic
location for brain (gray), liver (pink) and spleen
(purple). Each point represents the methylation
level of an individual sample for a given probe.
The curve represents averaged smoothed M
values, described in detail in the Methods.
Because of the scale and standardization used,
M values that range from À0.5 to 0.5 represent
unmethylated sites as defined by the control
probes, and values from 0.5 to 1.5 represent
baseline levels of methylation. The middle panel
provides the location of CpG dinucleotides with
black tick marks on the x axis. CpG density was calculated across the region using a standard density estimator and is represented by the smoothed black
line. The location of the CpG island is denoted on the x axis as an orange line. The lower panel provides gene annotation for the genomic region. The thin
outer gray line represents the transcript, the thin inner lines represent a coding region. Filled in gray boxes represent exons. On the y axis, plus and minus
marks denote sense and antisense gene transcription, respectively. (b) An example of a C-DMR that is located in a CpG island shore and that overlaps a
T-DMR. Liver (pink) is hypomethylated relative to brain (gray) and spleen (purple) tissues. Hypomethylation of colon tumor (orange) is observed in
comparison to matched normal colon tissue (green) and overlaps the region of liver hypomethylation.
−0.5
0.5
1.5
2.5
a b
−0.5
0.5
1.5
2.5
M
Brain (normal)
Liver (normal)
Spleen (normal)
0.00
0.10
0.00
0.10
CpGdensity
Shore Island
25280400 25280800 25281200 25281600
Location
Genes
Chr. 10
−
+
PRTFDC1
M
Brain (normal
Liver (normal)
Spleen (normal)
Colon (normal)
CpGdensity
25608000 25609000 25610000 25611000
Location
Genes
Chr. 16
−
+ HS3ST4
IslandShore
Colon (tumor)
NATURE GENETICS VOLUME 41 [ NUMBER 2 [ FEBRUARY 2009 179
A R T I C L E S©2009NatureAmerica,Inc.Allrightsreserved.
enrichment for repetitive DNA, which has been shown to be hypomethylated
in cancer11. This similarity in amount of hypomethylation
and hypermethylation is also shown in a quantile-quantile plot, in
which quantiles for the observed average difference between tumor
and normal sample Ms are plotted against quantiles from a null
distribution constructed with the control (M ¼ 0) regions (Fig. 3a).
Although both hypomethylation and hypermethylation in cancer
involved CpG island shores, there were subtle differences in the precise
regions that were altered. The hypermethylation extended to include
portions of the associated CpG islands in 24% of cases (termed
‘overlap’ in Fig. 2), which could account for the island hypermethylation
frequently reported in cancer, even though that is not the
predominant site of modification. In contrast, the hypomethylation
extended to between 2 and 3 kb from the associated island in 10% of
cases and was not associated with an island in 35% of cases (Fig. 2).
To confirm differential methylation in colon tumors, we carried out
additional bisulfite pyrosequencing validation of nine C-DMRs,
including five regions showing hypermethylation and four regions
with hypomethylation, in an average of 50 primary cancer and normal
mucosal samples per gene. For all of the genes, the pyrosequencing
data matched the CHARM data (P values ranging from 10–4 to 10–17)
(Fig. 3b–j and Supplementary Table 3 online). Thus, CHARM was
precise in identifying both T-DMRs and C-DMRs.
Our screening process was effective at identifying known targets of
altered DNAm in cancer. For example, 10 of the 25 most statistically
significant C-DMRs have previously been reported to show altered
DNAm in cancer, for example, WNK2, hypermethylated in glioblastoma12
and HOXA6, hypermethylated in lymphoid malignancies13.
However, we also identified hundreds of genes not previously
described. For example, for hypermethylation, we identified genes
encoding GATA-2, an important regulator of hematopoetic differentiation14,
and RARRES2, whose expression is decreased in intestinal
adenomas15. For hypomethylation, we identified genes encoding
DPP6, a biomarker for melanoma16, MRPL36, a DNA helicase that
confers susceptibility to breast cancer17, and MEST, a known target of
hypomethylation and loss of imprinting in breast cancer18. Note that
although previous T-DMR screens have focused on CpG islands,
which we show account for only 8% of T-DMRs, our screen did
identify CpG island loci validated by others as well, for example, PAX6,
OSR1 and HOXC12. Thus, cancer, like normal tissues, involves
changes in DNAm in CpG island shores, with comparable amounts
of hypomethylation and hypermethylation but with subtle differences
in the precise distribution of these alterations with respect to the
associated CpG island. These differences will have important functional
implications for gene expression, as discussed later.
Gene expression is linked to non–CpG-island methylation
Because the identification of CpG island shores was unexpected, we
explored the functional relationship between their differential methylation
and the expression of associated genes. To address tissue- and
cancer-specific DNAm, we analyzed gene expression across the
genome in five primary brains and livers from the same autopsy
specimens and in four colon cancers and subject-matched normal
mucosa; all samples were from subjects for whom we had genomewide
methylation analysis data. Methylation of T-DMRs showed a
strong inverse relationship with differential gene expression, even
though these DMRs were not CpG islands but rather CpG island
shores. The relationship between DNAm and gene expression was
greater for DMRs in which one of the two measured points had
approximately no methylation (‘none-to-some’ methylation compared
to ‘some-to-more’ or ‘some-to-less’ methylation), particularly for
hypomethylation (Fig. 4). The significant association of gene expression
with T-DMRs was true even when the DMR was 300–2,000 bp
from the transcription start site, for example, average log-ratio values
of 0.84 and 0.35 (P o 10–37 and 10–4) for some-to-none and someto-more/less
methylation, respectively, comparing liver to brain
expression (Fig. 4). Moreover, when we related T-DMRs to changes
in gene expression from over 242 samples, representing 20 different
tissue types, we found that 5,352 of the 8,910 genes that were
differentially expressed across the 20 tissues were within 2 kb of
a T-DMR, much more than expected by chance (P o 10–15). For
C-DMRs as well, even though there were fewer of them than T-DMRs,
there was a significant association of gene expression with DNAm:
P o 10–6 and P o 10–3 for hypermethylation and hypomethylation,
respectively; again, the relation was much more marked when one
of the two measured points had no methylation (Supplementary
Fig. 2 online).
Islands Overlap (0, 500) (500,
1,000)
(1,000,
2,000)
(2,000,
3,000)
>3,000
Percent
0
10
20
30
40
Percent
0
10
20
30
40
Percent
0
10
20
30
40
Percent
0
10
20
30
40
Percent
0
10
20
30
40
CpG regions
Shore
Islands Overlap (0, 500) (500,
1,000)
(1,000,
2,000)
(2,000,
3,000)
>3,000
T−DMR
Shore
Islands Overlap (0, 500) (500,
1,000)
(1,000,
2,000)
(2,000,
3,000)
>3,000
C−DMR
Shore
Islands Overlap (0, 500) (500,
1,000)
(1,000,
2,000)
(2,000,
3,000)
>3,000
Hypermethylation C−DMR
Shore
Islands Overlap (0, 500) (500,
1,000)
(1,000,
2,000)
(2,000,
3,000)
>3,000
Hypomethylation C−DMR
Shore
Figure 2 Distribution of distance of T-DMRs and C-DMRs from CpG islands.
Islands (teal) are regions that cover or overlap more than 50% of a CpG
island. Overlap (orange) are regions that overlap 0.1–50% of a CpG island.
Regions denoted by (0, 500] (purple) do not overlap islands but are located
r500 bp of islands. Regions denoted by (500, 1,000] (magenta) are
located 4500 and r1,000 bp from an island. Regions denoted by (1,000,
2,000] (green) are regions 41,000 bp and r2,000 bp from an island.
Regions denoted by (2000, 3000] (yellow) are located 42,000 bp and
r3,000 bp from an island. Regions denoted 43,000 (brown) are
43,000 bp from an island. The percentage of each class is provided for
CpG regions (the CHARM arrays themselves, null hypothesis), tissue-specific
differentially methylated regions (T-DMRs), cancer-specific differentially
methylated regions (C-DMRs), and the latter subdivided into regions of
cancer-specific hypermethylation and hypomethylation.
180 VOLUME 41 [ NUMBER 2 [ FEBRUARY 2009 NATURE GENETICS
A R T I C L E S©2009NatureAmerica,Inc.Allrightsreserved.
We validated the inverse relationship between DNAm and transcription
at eight CpG island shores, two T-DMRs and six
C-DMRs in tissues and colon cancers, respectively, using quantitative
real-time PCR. Both of the T-DMRs were in shores, one located 844
bp upstream of the promoter and one within the gene body. Similarly,
all six of the C-DMRs assayed were in shores, with five located in the
gene promoter and one within the gene body (Supplementary Table 4
online). These quantitative data provide additional support for a
strong relationship between differential methylation in CpG island
shores and transcription of associated genes. This functional relationship
between gene expression and shore methylation applies to shores
located within 2 kb of an annotated transcriptional start site but
leaves open the possibility of additional regulatory function for shores
located in intragenic regions or gene deserts.
Shore-linked silencing reversed by methyltransferase inhibition
The previous data, although compelling, are associative in nature. For
a more functional analysis, we therefore compared DNA methylation
and gene expression data from tissues studied in the current work
to a rigorous analysis using hundreds of expression microarray
experiments published earlier19, which tested the effects on gene
expression of 5-aza-2¢-deoxycytidine (AZA), and also to double
DNA methyltransferase 1 and 3B somatic cell knockout (DKO)
experiments. We compared genes from the present study that had
DMRs meeting an FDR o 0.05 and that showed differential expression
in the tissues at P o 0.05 to genes that had significant P values
after AZA or DKO. Of 27 DMRs that showed relative hypermethylation
with gene silencing in tissues, 23 were activated by AZA (Fig. 5a
and Supplementary Data 3 online). Similarly, of 25 DMRs that
showed relative hypermethylation with gene silencing in tissues, all
25 were activated by DKO (Fig. 5b and Supplementary Data 3). Thus,
both chemical and genetic demethylation cause changes in gene
expression similar to those associated with increased methylation of
CpG island shores.
DMRs are associated with alternative transcription
What might be the function of differential methylation at CpG island
shores? One possibility is alternative transcription. Both the T-DMRs
and C-DMRs often involved alternative transcripts, as defined by cap
analysis gene expression (CAGE)20,21: 68% and 70% of the T-DMRs
−0.4 −0.2 0 0.2 0.4
−1.0
−0.5
0
0.5
1.0
Null distribution quantiles
Observeddistributionquantiles
Hypermethylated
Hypomethylated
100
80
60
40
20
0
Methylationlevel(%)
100
80
60
40
20
0
Methylationlevel(%)
100
80
60
40
20
0
Methylationlevel(%)
100
80
60
40
20
0
Methylationlevel(%)
100
80
60
40
20
0
Methylationlevel(%)
100
80
60
40
20
0
Methylationlevel(%)
100
80
60
40
20
0
Methylationlevel(%)
100
80
60
40
20
0
Methylationlevel(%)
100
80
60
40
20
0
Methylationlevel(%)
cb
a
e
h
d
f g
i j
Normal
(n = 53)
Tumor
(n = 61)
Normal
(n = 58)
Tumor
(n = 51)
Normal
(n = 42)
Tumor
(n = 48)
Normal
(n = 57)
Tumor
(n = 64)
Normal
(n = 30)
Tumor
(n = 34)
Normal
(n = 38)
Tumor
(n = 36)
Normal
(n = 48)
Tumor
(n = 41)
Normal
(n = 61)
Tumor
(n = 65)
Normal
(n = 52)
Tumor
(n = 59)
P = 1.03E–07 P = 2.15E–04 P = 1.33E–17
P = 9.03E–04 P = 3.13E–15 P = 1.42E–15
P = 4.01E–04 P = 2.31E–04 P = 3.22E–07
Figure 3 Similar numbers of sites of hypomethylation and hypermethylation
in colon cancer. (a) A quantile-quantile plot shows a similar number of sites
of hypomethylation and hypermethylation in colon cancer. The quantiles of
the differences in M values between tumor and normal colon tissues are
plotted against the quantiles of a null distribution formed using the
differences seen in the control regions. Points deviating from the diagonal
are not expected by chance, and a similar proportion is seen for
hypomethylation and hypermethylation in cancer. (b–j) Bisulfite
pyrosequencing confirms the prevalence of five hypermethylated and four
hypomethylated C-DMR shores in a large set of colon tumor and normal
mucosa samples. Box plots represent degree of DNA methylation as
measured using bisulfite pyrosequencing. (b) DLX5 (distal-less homeobox 5).
(c) LRFN5 (leucine-rich repeat and fibronectin type III domain containing
5). (d) HOXA3 (homeobox A3). (e) SLITRK1 (SLIT and NTRK like family,
member 1). (f) FEZF2 (FEZ family zinc finger 2). (g) TMEM14A
(transmembrane protein 14A). (h) ERICH1 (glutamate-rich 1). (i) FAM70B
(family with sequence similarity 70, member B). (j) PMEPA1 (prostate
transmembrane protein, androgen induced 1). n, number of samples
analyzed by pyrosequencing.
−1.5 −1.0 −0.5 0 0.5 1.0 1.5
−5
0
5
Differential methylation
Differentialexpression
Figure 4 Gene expression is strongly correlated with T-DMRs at CpG island
shores. For each brain versus liver T-DMR, we found the closest annotated
gene on the Affymetrix HGU133A microarray, resulting in a total of 2,041
gene–T-DMR pairs. Plotted are log (base 2) ratios of liver to brain expression
against DM values for liver and brain DNAm. Orange dots represent T-DMRs
located within 300 bp from the corresponding gene’s transcriptional start
site (TSS). Green dots represent T-DMRs that are located from 300 to
2,000 bp from the TSS of an annotated gene. Black dots, in the middle,
represent log ratios for all genes further than 2 kb from an annotated TSS.
NATURE GENETICS VOLUME 41 [ NUMBER 2 [ FEBRUARY 2009 181
A R T I C L E S©2009NatureAmerica,Inc.Allrightsreserved.
and C-DMRs, respectively, were not within 500 bp of an annotated
transcriptional start site but were within 500 bp of an alternative
transcriptional start site. By chance, we expect only 58% to have this
relationship (P o 10–15). These results suggest that DNA methylation
might regulate alternative transcription in normal differentiation and
cancer. We therefore carried out rapid amplification of cDNA ends
(RACE) experiments in order to confirm the presence of alternative
transcripts and their differential expression in cancer. We examined
three colon tumor and subject-matched normal mucosa at the
PIP5K1A locus, a C-DMR that is hypomethylated in colon tumors,
and confirmed that an alternative RNA transcript is produced in colon
tumors compared to their matched normal counterparts (Supplementary
Fig. 3 online). Thus, a key function for differential methylation
during differentiation may be alternative transcription, and the
role of altered DNAm in cancer may in part be disruption of the
regulatory control of specific promoter usage.
Mouse DNAm discriminates human tissues, even far from genes
A compelling argument for the functional importance of differential
DNAm of CpG island shores would be their conservation across
species. One might expect DMRs near transcriptional start sites to be
conserved because the genes are conserved. However, when we
examined the relationship between gene-distant T-DMRs (2–10 kb
away from an annotated gene) and sequence conservation using the
phastCons28way table from the University of California Santa Cruz
genome browser, we found that 48% of differentially methylated
regions showed sequence conservation. Furthermore, 91% of DMRs
were located within 1 kb of a highly conserved region (P o 0.001).
To address whether the DNA methylation itself is conserved across
species, we created a mouse CHARM array with B2.1 million features
independently of the human array. We then isolated tissue replicates
from each of three mice, corresponding to the tissues examined in the
human T-DMR experiments, and then mapped these methylation
data across species using the UCSC LiftOver tool. The interspecies
correspondence of tissue-specific methylation was notable, and unsupervised
clustering perfectly discriminated among the tissues, regardless
of the species of origin (Fig. 6; P o 10–9). Perfect discrimination
among the tissues was found even when we limited the analysis
to gene-distant DMRs (Supplementary Fig. 4 online). Thus, DNAm
itself is highly conserved across 50 Myr of evolution (approximately
51% of mapped DNAm sites were conserved). We also noticed
relatively little heterogeneity in tissue-specific methylation in the
mouse compared to the human (height of the cluster bars in Fig. 6),
suggesting a genetic-epigenetic relationship, as the mice are inbred.
The location of C-DMRs overlaps that of T-DMRs
Because both C-DMRs and T-DMRs were located at CpG island
shores, we then asked whether they occurred in similar locations. We
–6
–4
–2
0
2
4
6
8
10
12
Gene
Relativeexpression(log2)
–6
–4
–2
0
2
4
6
Gene
Relativeexpression(log2)
DKO
Hypermethylated T-DMR
a
b
W
R
B
SH
C
1
D
U
SP6
AC
AD
VL
SO
X9
SN
C
A
SD
C
4
S100A10
D
YN
LL1
N
TR
K2
N
FKBIA
M
TM
R
6
M
T2A
M
T1X
M
EF2C
KC
N
J8
KAL1
IG
FBP3
IFITM
3
H
SPA8
G
LD
C
G
C
H
1
FYN
EG
FR
C
D
14
C
C
K
ABC
C
5
Hypermethylated T-DMR
AZA treatment
YW
H
AQ
YW
H
AE
TU
BB4
TU
BB2B
TU
BA1A
TPI1
STM
N
1
SG
C
E
R
AD
51C
PR
KAR
2B
PPP2C
A
PPP1C
B
D
YN
LL1
N
TR
K2
N
D
U
FA5
M
APR
E2
H
SPA8
H
2AFX
FG
F12
D
C
K
C
AP1
BASP1
APPBP2
AP2B1
AP1S2
Brain(mouse)
Brain(mouse)
Brain(mouse)
Brain(human)
Brain(human)
Brain(human)
Brain(human)
Brain(human)
Liver(mouse)
Liver(mouse)
Liver(mouse)
Liver(human)
Liver(human)
Liver(human)
Liver(human)
Liver(human)
Spleen(human)
Spleen(human)
Spleen(human)
Spleen(human)
Spleen(human)
Spleen(mouse)
Spleen(mouse)
Spleen(mouse)
Figure 6 Clustering of human tissue samples using mouse T-DMRs results
in perfect discrimination of tissues. The M values of all tissues from the
1,963 regions corresponding to mouse T-DMRs that mapped to the human
genome were used for unsupervised hierarchical clustering. By definition,
the mouse tissues are segregated. Notably, all of the human tissues are also
completely discriminated by the regions that differ in mouse tissues. The
three major branches in the dendrograms correspond perfectly to tissue
type regardless of species. Columns represent individual samples, and rows
represent regions corresponding to mouse T-DMRs. The heat map shows
M values, with red being more methylated and blue less.
Figure 5 Genes downregulated in association with T-DMR shore
hypermethylation are activated by 5-aza-2¢-deoxycytidine treatment of colon
cancer cell line HCT116 and knockout of DNA methyltrasferase 1 and 3b
in HCT116. (a) Genes significantly upregulated (P o 0.05) after treatment
of HCT116 cells with 5-aza-2¢-deoxycytidine (AZA) (black) that are also
associated with a relatively hypermethylated T-DMR showing a significant
change in gene expression (P o 0.05) (gray): 23/27 genes are activated by
AZA. (b) Genes significantly upregulated (P o 0.05) after knockout of DNA
methyltransferases 1 and 3b (DKO) in HCT116 cells (black) that are also
associated with a relatively hypermethylated T-DMRs showing a significant
change in gene expression (P o 0.05) (gray): 25/25 genes are activated
by DKO. Plotted are log (base 2) ratios of expression of AZA/untreated,
DKO/HCT116 and relatively hypermethylated/hypomethylated tissue.
182 VOLUME 41 [ NUMBER 2 [ FEBRUARY 2009 NATURE GENETICS
A R T I C L E S©2009NatureAmerica,Inc.Allrightsreserved.
focused on DMRs in which the methylation difference was from
no methylation to some methylation, that is, those DMRs for
which the gene expression data above showed a strong relationship
between ‘none-to-some’ methylation and gene silencing. Notably,
we found that 52% of the C-DMRs overlapped a T-DMR,
compared to only 22% expected by chance (P o 10–14), when
using an FDR of 5% for defining T-DMRs. Although these data
are significant, the definition of a T-DMR based on FDR of 5%
is conservative. We therefore also asked directly whether C-DMRs
are enriched for tissue variation in DNAm by computing an
averaged F-statistic (comparison of cross-tissue to within-tissue
variation) at each C-DMR. The cross-tissue variation in normal
tissues was significant at 64% of the C-DMRs, compared to 20%
of randomly selected CpG regions on the array matched for size
(P o 10–143). When we defined DMRs using an FDR of 5%, 1,229 of
2,707 C-DMRs overlapped a T-DMR, of which 265, 448 and 185 are
brain-, liver- and spleen-specific, respectively, and 331 show variation
among all of the tissues (Supplementary Data 4 online). The colon CDMRs
were highly enriched for overlap with liver T-DMRs (P o 10–15),
and liver is embryologically closest to colon of the autopsy tissues
studied. For example, the C-DMR located in the CpG island shore
upstream of the HS3ST4 (heparan sulfate D-glucosaminyl 3-Osulfotransferase
4) gene is hypomethylated in colon cancer compared
to normal colon and coincides with a T-DMR that distinguishes liver
from other tissues (Fig. 1b). The correspondence between C-DMRs
and T-DMRs was so marked that when we carried out unsupervised
clustering of the normal brain, liver and spleen, using the M values
from the C-DMRs, there was perfect discrimination of the tissues
(Fig. 7).
Most tissue-specific methylation difference more commonly involves
hypomethylation, although this varies by tissue type with 50% of liver,
62% of spleen, and 79% of brain DMRs representing hypomethylation,
and cancer-specific methylation differences slightly more frequently
involve hypermethylation (56%:44%). For both
T-DMRs and C-DMRs, when there was differential methylation, it
was common that at least one of the tissues was completely unmethylated
(68% and 37%, respectively). Furthermore, hypomethylated
C-DMRs were twice as likely to resemble another tissue type, such
as liver, than were hypermethylated C-DMRs (82% versus 61%,
P o 10–31), even though hypermethylated C-DMRs overlapped
T-DMRs 1.5-fold more frequently than did hypomethylated
C-DMRs (54% versus 35%, P o 10–21).
To further explore the relationship between differentiation and type
of methylation change, we carried out Gene Ontology (GO) analysis
for both hypomethylated and hypermethylated C-DMRs in the
cancers (see Methods). The GO analysis showed enrichment for
development and pluripotency-associated genes for both hyper- and
hypomethylated C-DMRs (P o 0.01) (Supplementary Table 5
online). Hypomethylated C-DMRs were also enriched for genes
associated with differentiated cellular functions for lineages other
than the colon (P o 0.01) (Supplementary Table 5). Thus, cancerspecific
DNA methylation predominantly involves the same sites that
show normal DNAm variation among tissues, particularly at genes
associated with development.
Next, we examined the magnitude of differential methylation and
variation in C-DMRs and T-DMRs. The DM values for tissue and
cancer DMRs differed markedly from nonmethylated controls or
randomly selected regions (the latter have an average value
Colon
Colon
Colon
Colon
Colon
Colon
Colon
Colon
Colon
Colon
Colon
Colon
Colon
Colon(tumor)
Colon(tumor)
Colon(tumor)
Colon(tumor)
Colon(tumor)
Colon(tumor)
Colon(tumor)
Colon(tumor)
Colon(tumor)
Colon(tumor)
Colon(tumor)
Colon(tumor)
Colon(tumor)
Brain
Brain
Brain
Brain
Brain
Liver
Liver
Liver
Liver
Liver
Spleen
Spleen
Spleen
Spleen
Spleen
a b
Figure 7 Clustering of normal tissue samples using C-DMRs results in perfect discrimination of tissues. The M values of all tissues from the 2,707 regions
corresponding to C-DMRs were used for unsupervised hierarchical clustering. (a) By definition, the colon tumors and matched normal mucosa are segregated.
The two major branches in the dendrograms correspond perfectly to tissue type. (b) Notably, all of the normal brains, spleens and livers are also completely
discriminated by the regions that differ in colon cancer. The three major branches in the dendrograms correspond perfectly to tissue type. Columns represent
individual samples, and rows represent regions corresponding to C-DMRs. The heat map shows M values, with red being more methylated and blue less.
NATURE GENETICS VOLUME 41 [ NUMBER 2 [ FEBRUARY 2009 183
A R T I C L E S©2009NatureAmerica,Inc.Allrightsreserved.
comparable to controls but with significant tails, as by definition they
may contain DMRs themselves) (Fig. 8a–d). The DM values for normal
tissues were comparable across the tissues, but the DM values between
normal and cancer tissues were on average approximately half the DM
between normal tissue pairs (Fig. 8e), which is logical given that the
cancers are compared with their tissue of origin. Another difference
between cancer and normal tissues was an increase in the interindividual
variation in M among the colon cancers, which was on
average B50% greater than the interindividual variation among the
normal colons (Fig. 8f), a result which may help to explain tumor
cell heterogeneity. Given the strong interindividual variability we
found in cancer, we identified 205/2,707 C-DMRs that were consistently
differentially methylated between the colon tumor and matched
normal mucosa from all 13 individuals examined (Supplementary
Data 4). These regions, heavily over-represented for development and
morphogenesis genes, provide a smaller, more focused set of regions
for biomarker discovery and carcinogenesis studies.
DISCUSSION
We have carried out a genome-wide analysis of DNA methylation
addressing variation among normal tissue types, variation between
cancer and normal, and variation between human and mouse,
revealing several surprising relationships among these three types
of epigenetic variation, supported by extensive bisulfite pyrosequencing
and functional analysis. First, most tissue-specific DNAm
occurs not at CpG islands but at CpG island shores. The identification
of these regions opens the door to functional studies, such as
those investigating the mechanism of targeting DNAm to these regions
and the role of differential methylation of shores. Supporting a
functional role for shores, gene expression was closely linked to
T-DMR and C-DMR methylation, particularly for switches from
‘none’ to ‘some’ methylation. The relationship between shore methylation
and gene expression was confirmed by 5-aza-2¢-deoxycytidine
and DNA methyltransferase knockout experiments altering expression
of the same genes. Another mechanism for shores supported by this
study is regulation of alternative transcripts, supported by mapping
and RACE experiments.
Although 76% of T-DMRs were in CpG island shores, at least for
the three tissues examined here, 24% were not adjacent to conventionally
defined CpG islands. However, many of these regions were
nevertheless shores of CpG-enriched sequences (for an example, see
Supplementary Fig. 5 online). We are currently developing a novel
algorithm for CpG island definition based on hidden Markov modeling
that will likely increase the fraction of T-DMRs in CpG island
shores. The ‘CpG clusters’ recently identified22 are not CpG island
shores (only 4% of shore DMRs map to them), although the shores of
these clusters, like the shores of CpG islands, are enriched for DMRs.
Note that the variation in DNAm is still not within the dense CpG
regions as defined by any of these definitions but in CpG shores.
The second key finding of the study is that T-DMRs are highly
conserved between human and mouse, and the methylation itself is
sufficiently conserved to completely discriminate tissue types regardless
of species of origin. This was true even for T-DMRs located 42 kb
from transcriptional start sites. The incorporation of epigenetic data,
such as DNAm, in evolutionary studies as done here, should greatly
enhance the identification of conserved elements that regulate differentiation.
We also found greater DNAm heterogeneity in human than
in mouse (at least in an inbred strain), even for DMRs located 42 kb
from a gene promoter. This result suggests that the conservation of
DNAm between human and mouse may have a strong genetic basis,
consistent with a greater degree of tissue DNAm homogeneity in the
inbred mouse strain.
The third key finding of the study is that most cancer-related
changes in DNAm, that is, C-DMRs, at least for colon cancer,
correspond to T-DMRs, and that these changes are similarly divided
between hypomethylation and hypermethylation and also involve CpG
island shores. Thus, epigenetic changes in cancer largely involve the
same DMRs as epigenetic changes in normal differentiation. These
results have important implications for studies such as the Cancer
Genome Atlas, in that most altered DNA methylation in cancer
does not involve CpG islands, and thus these studies would benefit
from analysis of CpG island shores. Similarly, high-throughput
sequencing efforts based on reduced representation analysis of CpG
islands per se are unlikely to identify most DNAm variation in normal
tissues or in cancer.
Random Control Hyper Hypo
−1.5
−1.0
−0.5
0
0.5
1.0
1.5
a
c
e f
d
bLiver vs. brain Spleen vs. brain
Spleen vs. liver Cancer
−1.5
−1.0
−1.0
−0.5
−0.5
0 0
0.5
0.5
1.0
1.01.5
−1.5
−1.0
−0.5
0
0.5
1.0
1.5
Random Control Hyper Hypo
Random Control Hyper Hypo Random Control Hyper Hypo
Brainvs.liver
Brainvs.spleen
Livervs.spleen
Hypomethylated
Hypermethylated
0.2
0.4
0.6
0.8
1.0
1.2
1.4
∆M
Normal Cancer
0.05
0.10
0.15
0.20
Interindividualvariation
Figure 8 Magnitude of differential methylation and variation in C-DMRs
and T-DMRs. (a–d) Box plots of average DM values over all DMRs,
compared to randomly chosen regions and unmethylated control regions,
matched for length. (a) Liver versus brain. (b) Spleen versus brain.
(c) Spleen versus liver. (d) Colon cancer versus normal colonic mucosa.
(e) Differences in DNA methylation are greater in magnitude among
normal tissues than are differences between colon tumors and matched
normal mucosa. For all DMRs we computed the average DM. We then
stratified these values into T-DMRs, hypermethylated C-DMRs and
hypomethylated C-DMRs. T-DMRs were further stratified according to
brain versus liver, brain versus spleen and liver versus spleen pairwise
comparisons. The box plots represent absolute values of the DMs.
(f) Interindividual variation in M is larger among colon tumors than matched
normal mucosa. For each C-DMR we computed the average interindividual
s.d. of the M values. The box plots represent these values for normal colon
mucosa and colon tumors.
184 VOLUME 41 [ NUMBER 2 [ FEBRUARY 2009 NATURE GENETICS
A R T I C L E S©2009NatureAmerica,Inc.Allrightsreserved.
Finally, GO annotation analysis suggests that DNAm changes in
cancer reflect development and pluripotency-associated genes, and
differentiated cellular functions for lineages other than the colon.
These data are consistent with the epigenetic progenitor model of
cancer23, which proposes that epigenetic alterations affecting tissuespecific
differentiation are the predominant mechanism by which
epigenetic changes cause cancer. The genes identified in this analysis
will themselves be of considerable interest for further study, as will be
the potential regulatory regions that did not lie in close proximity to
annotated genes.
METHODS
Samples. We obtained snap-frozen colon tumors and dissected normal mucosa
from the same subjects courtesy of B. Vogelstein (Johns Hopkins University
School of Medicine). Human postmortem brain, liver and spleen tissues, from
the same individual, were donated by The Stanley Medical Research Institute
brain collection.
Genomic DNA isolation and McrBC fractionation. We carried out genomic
DNA isolation using the MasterPure DNA purification kit (Epicentre) as
recommended by the manufacturer. For each sample, 5 mg of genomic DNA
was digested, fractionated, labeled and hybridized to a CHARM microarray.
CHARM microarray design. CHARM microarrays were prepared as previously
described4 and additionally included a set of 4,500 probes, totaling 148,500 base
pairs across 30 genomic regions as controls. As these control probes represent
genomic regions without CpG sites and hence can not be methylated, we used
them to normalize and standardize array data. We standardized the observed
M values so that the average in the control regions was 0. Therefore, M values
of 0 for other probes on the array are associated with no methylation.
CHARM DNA methylation analysis. We conducted McrBC fractionation
followed by CHARM array hybridization for all human tissue samples as
previously described4. For each probe, we computed average M values across
the five samples in each tissue type. Differential methylation was quantified for
each pairwise tissue comparison by the difference of averaged M values (DM).
Replicates were used to estimate probe-specific s.d., which provided standard
errors (s.e.m.) for DM. We calculated z scores (DM/s.e.m.(:DM)) and grouped
contiguous statistically significant values into regions. Because millions of
z scores are examined, statistical confidence calculation needed to account
for multiple comparisons. We therefore computed false discovery rates (FDR)
and reported a list with an FDR of 5%24,25. Statistical significance of the
regions was assessed as described below. C-DMRs were determined using the
same procedure described above with the following exception: because we
observed greater heterogeneity in the cancer samples (Fig. 8f), we did not
divide DM by the standard errors, as this would penalize regions of highly
variable M values. For all expression microarray analysis, we used RMA for
processing26 and then averaged the samples in each tissue, and computed the
difference (equivalent to average log ratio). Mouse T-DMRs were determined
using the same statistical procedures as described above for the T-DMRs and
were then mapped to the human genome using the UCSC liftOver tool. To
correct for possible ‘array’ effects, we standardized each T-DMR by subtracting
the mean of M across all samples within a species and divided by s.d. across all
samples within a species. A list of all mouse T-DMRs is provided in
Supplementary Data 5 online. Overlap of C-DMRs with T-DMRs was
determined by adding the number of regions.
Statistical significance of DMRs. Contiguous regions composed of probes with
z scores associated with P values smaller than 0.001 were grouped into regions.
We used the area of each region (length multiplied by DM) to define statistical
significance27. We used a permutation test to form a null distribution for these
areas and the empirical Bayes approach described28. We tested for effects of
fragment length on M values observed using CHARM by computing the
expected DNA fragment size based on McrBC recognition sites. Next, we
stratified the DM values for each probe, from the colon tumor and normal
mucosa comparison, by fragment size. The results show no relationship
between fragment size and DM (data not shown).
Bisulfite pyrosequencing. Isolation of genomic DNA for all bisulfite pyrosequencing
validation was done using the MasterPure DNA purification kit
(Epicentre) as recommended by the manufacturer. For validation of shore
regions, 1 mg of genomic DNA from each sample was bisulfite-treated using an
EpiTect kit (Qiagen) according to the manufacturer’s specifications. We PCRamplified
converted genomic DNA using unbiased nested primers and carried
out quantitative pyrosequencing using a PSQ HS96 (Biotage) to determine
percentage methylation at each CpG site. For bisulfite pyrosequencing of
C-DMRs in 34–65 colon tumor and 30–61 normal mucosa samples, 500 ng
of genomic DNA was bisulfite-treated using the EZ-96 DNA Methylation Gold
kit (Zymo Research) as specified by the manufacturer. Converted genomic
DNA was PCR-amplified using unbiased nested primers followed by pyrosequencing
using a PSQ HS96 (Biotage). Bisulfite pyrosequencing was done as
previously described29. We determined percentage methylation at each CpG site
using the Q-CpG methylation software (Biotage). Supplementary Table 6
online provides the genomic location of CpG sites measured in the CpG shore
and associated CpG island bisulfite pyrosequencing assays. Genomic coordinates
for all CpG sites measured in the set of B50 colon tumor and normal
samples are provided in Supplementary Methods online. Primer sequences
and annealing temperatures for all bisulfite pyrosequencing reactions are
provided in Supplementary Table 7 online.
Total RNA isolation. We isolated total RNA for Affymetrix microarray analysis
from all human tissues using the RNeasy Mini kit (Qiagen) as specified by the
manufacturer. All samples were DNase treated using the on-column DNase
digestion kit (Qiagen) as recommended. We measured total RNA concentration
and determined RNA quality by using an RNA 6000 Nano Lab chip kit and
running the chip on a 2100 Bioanalyzer (Agilent).
Affymetrix microarray expression analysis. Genome-wide transcriptional
analysis was done on a total of five liver and five brain samples from the
same individuals using Affymetrix U133A GeneChip microarrays. We
obtained the raw microarray gene expression data for the brain and liver tissue
from The Stanley Medical Research Institute (SMRI) online genomics database
(see URLs section below)30. The five individuals selected were unaffected
controls from the SMRI Array collection. We also carried out genome-wide
transcriptional profiling on four colon tumor and four matched normal
mucosa using Affymetrix U133 Plus 2.0 microarrays. One microgram of
high-quality total RNA was amplified, labeled and hybridized according to
the manufacturer’s (Affymetrix) specifications and data was normalized as
previously described26,31.
Quantitative real-time PCR. We prepared total RNA for quantitative real-time
PCR using the Trizol method (Invitrogen) for FZD3, RBM38, NDN
and SEMA3C and we used the RNeasy mini kit (Qiagen) to isolate total
RNA for ZNF804A, CHRM2 and NQO1. We prepared cDNA for quantitative
real-time PCR using the QuantiTect RT kit (QIAGEN). We used TaqMan assays
(Applied Biosystems) to determine relative gene expression and analyzed
experiments on a 7900HT detection system. Taqman assay identification
numbers are provided in the Supplementary Methods. We used human ACTB
as an endogenous control. Relative expression differences were calculated using
the DDCt method32.
5¢ RACE PCR. 5¢ RACE experiments were done using a second generation
RACE kit (Roche Applied Science) as specified by the manufacturer’s protocol.
RACE PCR products were directly sequenced with 3100 Genetic Analyzer
(AB Applied Biosystems). Gene specific primer sequences are provided in
Supplementary Methods.
GO annotation. We analyzed GO annotation using the Bioconductor33 Gostats
package34 to find enriched categories (P o 0.01).
URLs. Complete set of T-DMR plots, http://rafalab.jhsph.edu/t-dmr3000.pdf;
complete set of C-DMR plots, http://rafalab.jhsph.edu/c-dmr-all.pdf. The
Stanley Medical Research Institute (SMRI) online genomics database,
www.stanleygenomics.org.
NATURE GENETICS VOLUME 41 [ NUMBER 2 [ FEBRUARY 2009 185
A R T I C L E S©2009NatureAmerica,Inc.Allrightsreserved.
Accession codes. NCBI GEO: Gene expression microarray data has been
submitted under accession number GSE13471.
Note: Supplementary information is available on the Nature Genetics website.
ACKNOWLEDGMENTS
We thank B. Volgelstein (Johns Hopkins University School of Medicine) for
providing colon tumors and matched normal mucosa samples. Postmortem brain,
liver and spleen tissue was donated by The Stanley Medical Research Institute
collection courtesy of M.B. Knable, E.F. Torrey and R.H. Yolken, whom we also
thank for making available gene expression data for the brain and the liver tissue.
We thank B. Carvalho for help with statistical software and C. Crainiceanu for
advice with statistical methods. This work was supported by US National
Institutes of Health grants P50HG003233 (A.P.F.), R37CA54358 (A.P.F.)
and 5R01RR021967 (R.A.I.).
AUTHOR CONTRIBUTIONS
R.A.I. and A.P.F. designed the study and interpreted the results; R.A.I. designed
new CHARM arrays and statistical methods with Z.W.; C.L.-A. performed
bisulfite pyrosequencing, real-time quantitative PCR and sample preparation
with C.M., K.G., M.R. and H.J.; B.W. and S.S. performed CHARM assays with
sample preparation from M.W. and advice from J.B.P.; P.O. and H.C. performed
functional assays; A.P.F. supervised the laboratory experiments and wrote the
paper with R.A.I. and C.L.-A.
Published online at http://www.nature.com/naturegenetics/
Reprints and permissions information is available online at http://npg.nature.com/
reprintsandpermissions/
1. Feinberg, A.P. & Vogelstein, B. Hypomethylation distinguishes genes of some human
cancers from their normal counterparts. Nature 301, 89–92 (1983).
2. Feinberg, A.P. & Tycko, B. The history of cancer epigenetics. Nat. Rev. Cancer 4,
143–153 (2004).
3. Baylin, S.B. & Ohm, J.E. Epigenetic gene silencing in cancer - a mechanism for early
oncogenic pathway addiction? Nat. Rev. Cancer 6, 107–116 (2006).
4. Irizarry, R.A. et al. Comprehensive high-throughput arrays for relative methylation
(CHARM). Genome Res. 18, 771–779 (2008).
5. Gardiner-Garden, M. & Frommer, M. CpG islands in vertebrate genomes. J. Mol. Biol.
196, 261–282 (1987).
6. Eckhardt, F. et al. DNA methylation profiling of human chromosomes 6, 20 and 22.
Nat. Genet. 38, 1378–1385 (2006).
7. Weber, M. et al. Distribution, silencing potential and evolutionary impact of promoter
DNA methylation in the human genome. Nat. Genet. 39, 457–466 (2007).
8. Illingworth, R. et al. A novel CpG island set identifies tissue-specific methylation at
developmental gene loci. PLoS Biol. 6, e22 (2008).
9. Shen, L. et al. Genome-wide profiling of DNA methylation reveals a class of normally
methylated CpG island promoters. PLoS Genet. 3, 2023–2036 (2007).
10. Antequera, F. & Bird, A. CpG islands as genomic footprints of promoters that are
associated with replication origins. Curr. Biol. 9, R661–R667 (1999).
11. Ehrlich, M. DNA methylation in cancer: too much, but also too little. Oncogene 21,
5400–5413 (2002).
12. Hong, C. et al. Epigenome scans and cancer genome sequencing converge on WNK2, a
kinase-independent suppressor of cell growth. Proc. Natl. Acad. Sci. USA 104,
10974–10979 (2007).
13. Strathdee, G. et al. Inactivation of HOXA genes by hypermethylation in myeloid and
lymphoid malignancy is frequent and associated with poor prognosis. Clin. Cancer Res.
13, 5048–5055 (2007).
14. Cantor, A.B. et al. Antagonism of FOG-1 and GATA factors in fate choice for the mast
cell lineage. J. Exp. Med. 205, 611–624 (2008).
15. Segditsas, S. et al. Putative direct and indirect Wnt targets identified through
consistent gene expression changes in APC-mutant intestinal adenomas from humans
and mice. Hum. Mol. Genet. 17, 3864–3875 (2008).
16. Jaeger, J. et al. Gene expression signatures for tumor progression, tumor subtype, and
tumor thickness in laser-microdissected melanoma tissues. Clin. Cancer Res. 13,
806–815 (2007).
17. Seal, S. et al. Truncating mutations in the Fanconi anemia J gene BRIP1 are lowpenetrance
breast cancer susceptibility alleles. Nat. Genet. 38, 1239–1241 (2006).
18. Pedersen, I.S. et al. Frequent loss of imprinting of PEG1/MEST in invasive breast
cancer. Cancer Res. 59, 5449–5451 (1999).
19. Gius, D. et al. Distinct effects on gene expression of chemical and genetic manipulation
of the cancer epigenome revealed by a multimodality approach. Cancer Cell 6,
361–371 (2004).
20. Shiraki, T. et al. Cap analysis gene expression for high-throughput analysis of
transcriptional starting point and identification of promoter usage. Proc. Natl. Acad.
Sci. USA 100, 15776–15781 (2003).
21. Wakaguri, H., Yamashita, R., Suzuki, Y., Sugano, S. & Nakai, K. DBTSS: database of
transcription start sites, progress report 2008. Nucleic Acids Res. 36, D97–D101
(2008).
22. Glass, J.L. et al. CG dinucleotide clustering is a species-specific property of the
genome. Nucleic Acids Res. 35, 6798–6807 (2007).
23. Feinberg, A.P., Ohlsson, R. & Henikoff, S. The epigenetic progenitor origin of human
cancer. Nat. Rev. Genet. 7, 21–33 (2006).
24. Benjamini, Y. & Hochberg, Y. Controlling the false discovery rate: a practical and
powerful approach to multiple testing. J Royal Statist Soc B (Methodol.) 57, 289–300
(1995).
25. Storey, J.D. The positive false discovery rate: a Bayesian interpretation and the q-value.
Ann. Statist. 31, 2013–2035 (2003).
26. Irizarry, R.A. et al. Exploration, normalization, and summaries of high
density oligonucleotide array probe level data. Biostatistics 4, 249–264 (2003).
27. Bullmore, E.T. et al. Global, voxel, and cluster tests, by theory and permutation, for a
difference between two groups of structural MR images of the brain. IEEE Trans. Med.
Imaging 18, 32–42 (1999).
28. Efron, B., Tibshirani, R., Storey, J.D. & Tusher, V. Empirical Bayes analysis of a
microarray experiment. J. Am. Stat. Assoc. 96, 1151–1160 (2001).
29. Tost, J. & Gut, I.G. DNA methylation analysis by pyrosequencing. Nat. Protocols 2,
2265–2275 (2007).
30. Higgs, B.W. et al. An online database for brain disease research. BMC Genomics 7, 70
(2006).
31. Bolstad, B.M., Irizarry, R.A., Astrand, M. & Speed, T.P. A comparison of normalization
methods for high density oligonucleotide array data based on variance and bias.
Bioinformatics 19, 185–193 (2003).
32. Livak, K.J. & Schmittgen, T.D. Analysis of relative gene expression data using real-time
quantitative PCR and the 2(-Delta Delta C(T)). Methods 25, 402–408 (2001).
33. Gentleman, R.C. et al. Bioconductor: open software development for computational
biology and bioinformatics. Genome Biol. 5, R80 (2004).
34. Falcon, S. & Gentleman, R. Using GOstats to test gene lists for GO term association.
Bioinformatics 23, 257–258 (2007).
186 VOLUME 41 [ NUMBER 2 [ FEBRUARY 2009 NATURE GENETICS
A R T I C L E S©2009NatureAmerica,Inc.Allrightsreserved.