Integrated genetic and epigenetic analysis identifies
`three different subclasses of colon cancer
`
`Lanlan Shen*, Minoru Toyota†, Yutaka Kondo‡, E Lin§, Li Zhang§, Yi Guo*, Natalie Supunpong Hernandez*, Xinli Chen*,
`Saira Ahmed*, Kazuo Konishi*, Stanley R. Hamilton¶, and Jean-Pierre J. Issa*储
`
`*Departments of Leukemia, §Biostatistics and Applied Biomathematics, and ¶Pathology, University of Texas M. D. Anderson Cancer Center,
`Houston, TX 77030; †Sapporo Medical University, S1W17, Chuo-ku, Sapporo 060-8556, Japan; and ‡Aichi Cancer Center, Division of
`Molecular Oncology, 1-1 Kanakonden, Chikusa-Ku, Nagoya, Japan
`
`Edited by Stephen E. Fienberg, Carnegie Mellon University, Pittsburgh, PA, and approved October 8, 2007 (received for review May 18, 2007)
`
`Colon cancer has been viewed as the result of progressive accu-
`mulation of genetic and epigenetic abnormalities. However, this
`view does not fully reflect the molecular heterogeneity of the
`disease. We have analyzed both genetic (mutations of BRAF, KRAS,
`and p53 and microsatellite instability) and epigenetic alterations
`(DNA methylation of 27 CpG island promoter regions) in 97 primary
`colorectal cancer patients. Two clustering analyses on the basis of
`either epigenetic profiling or a combination of genetic and epige-
`netic profiling were performed to identify subclasses with distinct
`molecular signatures. Unsupervised hierarchical clustering of the
`DNA methylation data identified three distinct groups of colon
`cancers named CpG island methylator phenotype (CIMP) 1, CIMP2,
`and CIMP negative. Genetically, these three groups correspond to
`very distinct profiles. CIMP1 are characterized by MSI (80%) and
`BRAF mutations (53%) and rare KRAS and p53 mutations (16% and
`11%, respectively). CIMP2 is associated with 92% KRAS mutations
`and rare MSI, BRAF, or p53 mutations (0, 4, and 31% respectively).
`CIMP-negative cases have a high rate of p53 mutations (71%) and
`lower rates of MSI (12%) or mutations of BRAF (2%) or KRAS (33%).
`Clustering based on both genetic and epigenetic parameters also
`identifies three distinct (and homogeneous) groups that largely
`overlap with the previous classification. The three groups are
`independent of age, gender, or stage, but CIMP1 and 2 are more
`common in proximal tumors. Together, our integrated genetic and
`epigenetic analysis reveals that colon cancers correspond to three
`molecularly distinct subclasses of disease.
`
`classification 兩 DNA methylation 兩 genetic alterations
`
`Colorectal cancer (CRC) is the second and fourth most
`
`common cancer in men and women, respectively (1). Ap-
`proximately 70% of colorectal cancers are sporadic, with no
`inherited predisposition. A stepwise progression model involv-
`ing two distinct genetic pathways has been proposed to explain
`the etiology of colon cancer from benign neoplasm to adeno-
`carcinoma (2). One class of genetic alterations involves muta-
`tions of oncogenes and tumor-suppressor genes that directly
`control cell birth and death, such as APC, KRAS, and p53.
`Another involves mutations of DNA mismatch repair genes.
`In addition to these genetic alterations, cancer initiation and
`promotion can occur by epigenetic mechanisms (3). CpG meth-
`ylation is the best characterized epigenetic change in the mam-
`malian genome. Whereas CpG dinucleotides are underrepre-
`sented in the mammalian genome, approximately half of all
`human genes contain a CpG-rich region called a ‘‘CpG island’’
`in the 5⬘ area, often encompassing the promoter and transcrip-
`tion start site of the associated gene (4, 5). Gene silencing by
`hypermethylation of CpG islands (including tumor-suppressor
`genes) is a common event in tumors. Further, hypermethylation
`of specific genes such as ER␣, MYOD1, and N33 occurs in the
`normal colon tissue of aging individuals (6, 7), and hypermeth-
`ylation of the secreted frizzled-related gene family (SFRPs) is
`detectable in aberrant crypt foci (8). The early occurrence of
`epigenetic alterations led to a hypothesis that they allow for the
`
`subsequent accumulation of both genetic and epigenetic alter-
`ations that promote tumor development and progression.
`Importantly, certain individuals appear predisposed to aber-
`rant promoter hypermethylation, including at several tumor-
`suppressor genes (9). This phenomenon, termed CpG island
`methylator phenotype (CIMP), provides an alternative pathway
`to promote colon cancer (10). Several independent studies have
`linked CIMP to distinct genetic and clinical features, including
`high rates of BRAF and KRAS mutation, low rates of p53
`mutations, specific histology (mucinous, poorly differentiated),
`familial occurrence, and distinct clinical outcome (11). However,
`the current view of the formation of colon cancer does not fully
`reflect the molecular heterogeneity of the disease. Here, we
`analyzed both genetic and epigenetic alterations in primary
`colorectal cancers and found that, molecularly, colon cancer
`consists of three distinct subclasses, each of which is fairly
`homogeneous.
`
`Results
`Clinical Variables and Epigenetic and Genetic Alterations. We ana-
`lyzed colorectal cancers from 97 individual CRC patients se-
`lected solely based on tissue availability. Clinical characteristics
`of the patients are summarized in Table 1. DNA isolated from
`grossly microdissected cancers was analyzed to determine the
`methylation status of 27 promoter-associated CpG islands se-
`lected based on prior studies. For each gene, the average
`methylation level measured quantitatively and the frequency of
`positive cases (with methylation level greater than ⬎15%) are
`shown in supporting information (SI) Table 6. In an initial
`analysis, we selected 20 cases to compare methylation analysis
`for the same genes by different methods [methylated CpG island
`amplification (MCA) (12), combined of bisulfite restriction
`enzyme amplification (COBRA) (13), or bisulfite pyrosequenc-
`ing (14)] and found excellent correlation in methylation between
`the methods (similar results were observed for 92% cases, using
`MCA or pyrosequencing methods, and 95% cases using COBRA
`or pyrosequencing methods). Therefore, we combined all results
`together for further analysis. Methylation frequencies for the 27
`genes we examined ranged from 5.2 to 98.9%. Five genes, ER␣,
`MyoD1, N33, HPP1, and SFRP1, were hypermethylated in
`⬎80% of cancer cases, suggestive of age-related methylation (6,
`15). Indeed, when we examined the methylation of these genes
`in normal-appearing mucosa from the same patients, we found
`substantial methylation in normal colon and significant corre-
`
`Author contributions: L.S., S.R.H. and J.-P.J.I. designed research; L.S., M.T., Y.K., Y.G., N.S.H.,
`X.C., S.A., and K.K. performed research; L.S., M.T., Y.K., E.L., L.Z., Y.G., and J.-P.J.I. analyzed
`data; and L.S., S.R.H., and J.-P.J.I. wrote the paper.
`
`The authors declare no conflict of interest.
`
`This article is a PNAS Direct Submission.
`储To whom correspondence should be addressed. E-mail: jpissa@mdanderson.org.
`This article contains supporting information online at www.pnas.org/cgi/content/full/
`0704652104/DC1.
`
`© 2007 by The National Academy of Sciences of the USA
`
`18654 –18659 兩 PNAS 兩 November 20, 2007 兩 vol. 104 兩 no. 47
`
`www.pnas.org兾cgi兾doi兾10.1073兾pnas.0704652104
`
`Geneoscopy Exhibit 1038, Page 1
`
`

`

`MEDICALSCIENCES
`
`Comparison of methylation level and frequency for 20 genes be-
`Fig. 1.
`tween CIMP-positive and negative groups. (A) Comparison of the mean
`methylation level of each gene between CIMP-positive group and CIMP-
`negative group. All genes except SFRP1 and SOCS1 showed significantly
`higher methylation level in the CIMP-positive group. (B) Comparison of meth-
`ylation frequency of positive cases for each gene between CIMP-positive and
`-negative groups. Methylation-positive case is defined by methylation level
`⬎15%. All Type-C genes except SOCS1 showed significantly higher frequency
`of methylation in CIMP-positive group, whereas 5 Type-A genes (on the Right)
`showed no difference between these two groups. **, P ⬍ 0.001; *, P ⬍ 0.05
`
`53%, respectively) but few KRAS and p53 mutations (16% and
`11%, respectively). Conversely, CIMP2 was associated with a
`high frequency of KRAS mutations (92%), but MSI and BRAF
`mutation occurred rarely (0% and 4% respectively) with a low
`rate of p53 mutation (31%). CIMP-negative cases had a higher
`rate of p53 mutation (71%) and lower rates of MSI (12%) and
`mutations of BRAF (2%) and KRAS (33%). Thus, each of MSI,
`BRAF, KRAS, and p53 alterations were unevenly distributed
`within the three groups (Fig. 3), and all of the P values were
`statistically significant (⬍0.0001 by Fisher’s exact test).
`Based on the hierarchical clustering results, we used both
`genetic and epigenetic information to perform K-means clus-
`tering, which identifies the most homogeneous clusters. The
`three groups classified from this analysis (Fig. 4) were largely
`overlapping with the previous classification, with only 17 (18%)
`cases being reclassified. By K-Means clustering, 22 cases were
`classified as CIMP1 (23%), 37 cases (38%) were classified as
`CIMP2, and 38 cases (39%) were classified as CIMP negative.
`To assess the reliability and reproducibility of the classifica-
`
`Table 1. Clinical characteristics of 97 CRC patients
`
`Characteristic
`
`Age
`Median age in years, range
`Missing data
`Gender
`Female
`Male
`Missing data
`Location
`Proximal
`Distal
`Missing data
`Stage
`I or II
`III or IV
`Missing data
`
`n ⫽ 97
`
`68 years (25–98 years)
`2
`
`29 (30%)
`66 (68%)
`2 (2%)
`
`38 (39%)
`44 (45%)
`15 (16%)
`
`43 (44%)
`37 (38%)
`17 (18%)
`
`lation between patient age and methylation of each gene (R ⫽
`0.36, P ⫽ 0.0005 for ER␣; R ⫽ 0.42, P ⬍ 0.001 for MyoD1; R ⫽
`0.45, P ⬍ 0.0001 for N33; R ⫽ 0.45, P ⬍ 0.0001 for SFRP1; and
`R ⫽ 0.33, P ⫽ 0.002 for HPP1; see SI Fig. 6). This was not found
`for any of the other genes examined. Therefore, as previously
`proposed, we called these five genes Type-A genes for age-
`related and all other genes Type-C genes for tumor-specific.
`We next determined the status of BRAF mutation (using
`pyrosequencing), KRAS mutation (using mutant allele specific
`amplification), p53 mutation (using single-strand conforma-
`tional polymorphism and sequencing), and microsatellite insta-
`bility (using the classical panel) in these same cases. BRAF
`mutation was observed in 11 of 87 cancers (12.6%); KRAS
`mutation was found in 43 of 94 cancers (45.7%); and 44 of 93
`patients (47.3%) had p53 mutation. Of the 97 tumors evaluated
`for microsatellite instability, 22 (22.7%) had high levels of
`microsatellite instability (MSI-H).
`
`CIMP Affects Most Genes. It was shown that methylation clusters in
`specific colorectal cancer subsets termed CIMP, and CIMP was
`originally defined based on seven cancer-specific MINT markers
`with hypermethylation at 2 or more loci (9). Using the original
`definition, 49 cases studied here were defined as CIMP-positive
`(51%) and 48 cases were CIMP negative. We compared the
`average methylation measured quantitatively at the additional 20
`genes between these two groups and found that all genes except
`SFRP1 and SOCS1 showed significantly higher methylation
`density in the CIMP-positive group (Fig. 1 A). When we analyzed
`the frequency of methylation-positive cases (methylation density
`⬎15%), we found all 15 Type-C genes except SOCS1 showed
`significantly higher frequency of methylation in the CIMP-
`positive group; the 5 Type-A genes showed no difference be-
`tween these two groups (Fig. 1B).
`
`Three Distinct Clusters of Colon Cancers. To explore the underlying
`patterns of gene-methylation changes, we performed unsuper-
`vised hierarchical clustering analysis, using the methylation of 27
`genes as a continuous variable within primary CRC patients.
`Three separate clusters were identified by this analysis, one of
`which corresponded very closely to the previous CIMP-negative
`group (middle cluster in Fig. 2) showing low or less methylation
`for all genes we examined. Surprisingly, CIMP-positive cases fit
`into two subgroups: CIMP1 (cluster 1 in Fig. 2) and CIMP2
`(cluster 3 in Fig. 2). When we compared the genetic alterations
`within these three clusters, each of them corresponded to very
`distinct genetic profiles (Fig. 3). CIMP1 cases showed a signif-
`icantly higher frequency of MSI and BRAF mutations (80% and
`
`Shen et al.
`
`PNAS 兩 November 20, 2007 兩 vol. 104 兩 no. 47 兩 18655
`
`Geneoscopy Exhibit 1038, Page 2
`
`

`

`in CIMP 2) compared with the CIMP-negative group (24% of
`proximal tumors, P ⫽ 0.004 by Fisher’s exact test).
`
`Optimal Markers to Predict the Three Groups. We further examined
`in detail the epigenetic signatures among three groups of CRC
`identified (CIMP1, CIMP2, and CIMP negative). By Kruskal–
`Wallis tests, we found that all Type-C genes (except for COX2,
`DAPK, and RASSF1A) showed significant differences among
`these groups (see SI Table 7 for details). The three genes
`showing no difference had very low levels of methylation overall.
`For Type-A genes, only MYOD1 showed a statistically signifi-
`cant difference among the groups. However, there was a non-
`significant trend for increased methylation of ER␣, HPP1, N33,
`and SFRP1 in CIMP2 compared with the other groups. Next, we
`used Z-score method to assign equal weight for methylation of
`each gene by substituting all raw methylation values in each data
`set with their respective Z-scores (see SI Materials and Methods
`for details), and assigned methylation scores for each patient
`based on the average Z-scores of either Type-A genes or Type-C
`genes. As shown in Fig. 5, the methylation score for Type-C
`genes was significantly higher in CIMP1, followed by CIMP2,
`and CIMP-negative cases were the lowest (0.56, 0.06 and ⫺0.38,
`respectively, P ⬍ 0.001). Interestingly, the methylation score for
`Type-A genes was significantly higher among CIMP2 (0.21)
`compared with CIMP1 (⫺0.18) and CIMP-negative (⫺0.15)
`individuals (P ⬍ 0.04).
`To determine which individual genetic or epigenetic alteration
`can best predict these three groups clinically, we calculated the
`sensitivity, specificity, positive and negative predictive values and
`␬ coefficient value (assessment for reliability) for each marker.
`Table 4 shows the top 10 single markers for predicting each
`group. Based on ␬ coefficient, the best single marker to predict
`CIMP1 group is hMLH1 methylation, whereas KRAS mutation
`is the best predictive marker for CIMP2 group, and p53 mutation
`is the best predictive marker for CIMP-negative group. As
`expected, the two genetic markers MSI-H and BRAF mutation
`were also among the best predictors for CIMP1, with a high
`degree of accuracy determined by sensitivity and predictive
`values. Several methylation markers are also on the top of the list
`for predicting each cluster; hMLH1, TIMP3, and MINT17
`methylation were most closely linked to CIMP1, methylation of
`MINT2 and MINT27 were associated with CIMP2, and lack of
`methylation of MINT1, MINT2, MINT27, and MINT31 pre-
`dicted CIMP negativity.
`To explore whether a combination of markers could provide
`greater accuracy than individual markers in predicting subtypes
`of CRC, we selected the top five predictive markers based on
`predictive values and analyzed them together. For the CIMP1
`group, a combination analysis of five markers (BRAF mutation
`and methylation of hMLH1, TIMP3, MINT1, and RIZ1) indi-
`cates that having three positive markers results an excellent
`positive predictive value and negative predictive value (94% and
`94% respectively, Table 5). For the CIMP2 group, no combina-
`tion performs better than KRAS mutation alone. In CIMP-
`negative group, p53 mutation and lack of methylation at
`MINT27, MINT2, MINT31, and MINT1 are the top five best
`markers, and a combination of any three markers gave 73%
`positive predictive value and 100% negative predictive value
`(Table 5). The performance of these markers in classifying CRC
`should, however, be validated in independent studies.
`
`Discussion
`In this study, we show that primary colorectal cancers cluster into
`three distinct subclasses based on epigenetic and genetic profiles:
`CIMP1, intense methylation of multiple genes and MSI and
`BRAF mutations; CIMP2, methylation of a limited group of
`genes, increased methylation level for age-related genes, and
`mutation in KRAS; and CIMP negative, rare methylation with
`
`Unsupervised hierarchical clustering analysis on the basis of 27
`Fig. 2.
`methylation markers. Three separate clusters were generated by this analysis
`with one cluster corresponding very closely to the previous CIMP-negative
`group (middle cluster), and CIMP-positive cases were separated into two
`subgroups, CIMP1 (cluster 1) and CIMP 2 (cluster 3).
`
`tion, first we performed bootstrap analysis (resampling with
`replacement method) (16) to determine the level of confidence
`of the clustering. As shown in SI Fig. 7, we observed three main
`blocks robustly clustered in bootstrap datasets, suggesting that
`each of these three classes is fairly stable. Interestingly, the
`middle cluster (CIMP2) shows more heterogeneity than the
`other two clusters. We also compared the current classification
`with our classification in ref. 9 in 49 CRC patients. We found that
`44 cases (90%) remained in the same groups, with only 5 cases
`being reclassified (Table 2). These results show that these three
`newly identified clusters largely overlap with the previous clas-
`sification. Together, our results suggest that combined genetic
`and epigenetic characteristics subclassify colorectal cancer into
`three distinct groups.
`We next analyzed whether the different CRC subclasses
`identified correlated to distinct clinical characteristics. Among
`the three groups, there was no significant difference in age,
`gender, or stage (Table 3). However, a significantly higher
`incidence of proximal colon cancer was found in both CIMP1
`and CIMP2 groups (63% of proximal tumors in CIMP 1 and 60%
`
`Fig. 3. Comparison of the genetic alterations among the three clusters. Each
`cluster corresponds to very distinct genetic profiles. CIMP1 is characterized by
`high frequency of MSI (80%) and BRAF mutations (53%), CIMP2 is character-
`ized by a higher rate of KRAS mutations (92%), and CIMP negative is charac-
`terized by high frequency of p53 mutations (71%).
`
`18656 兩 www.pnas.org兾cgi兾doi兾10.1073兾pnas.0704652104
`
`Shen et al.
`
`Geneoscopy Exhibit 1038, Page 3
`
`

`

`MEDICALSCIENCES
`
`K-means clustering analysis on the basis of both genetic and epigenetic markers. K-means clustering including genetic information yielded very
`Fig. 4.
`homogenous groups. Twenty-two cases were classified as CIMP1 (23%), 37 cases (38%) were classified as CIMP2, and 38 cases (39%) were classified as CIMP
`negative. This clustering largely overlaps with the previous hierarchical clustering with only 17 cases (18%) reclassified.
`
`p53 mutation. These three groups are relatively homogeneous on
`a molecular level and likely representative of three different
`subclasses of disease.
`These data suggest that colon cancer can be divided into
`substantially distinct groups in a way similar to breast cancers,
`where hormone status and HER2 amplification define distinct
`groups (17), and to leukemias, where specific chromosomal
`changes define very different diseases (18). The three colorectal
`cancer groups also differ clinically; CIMP1 and CIMP2 are more
`often proximal; CIMP1 has a good prognosis because it consists
`mostly of MSI-H cancers (19, 20), whereas CIMP2 has a poor
`prognosis (21). Moreover, they may have distinct precancerous
`lesions such as HPP/serrated adenomas for CIMP1 (22, 23), and
`villous adenomas for CIMP2 (24). It is unclear whether these
`three groups reflect initiations of cancer in distinct precancerous
`cells (as hypothesized for breast cancer), or reflect entirely
`different diseases (with a different cause/epidemiology) that
`affect the same precancerous cells. Nevertheless, they are suf-
`ficiently distinct to merit consideration in clinical trials and
`clinical management of colorectal cancer.
`The mechanistic basis of these two CIMP in colon cancer
`remain unknown. One possibility is that genetic events that
`activate methylases or inactivate methylation-protection factors
`explain CIMP1, where increased methylation degree and fre-
`quency is observed for multiple CpG islands including a number
`of tumor-suppressor genes, such as hMLH1, p16, p14, etc.
`Another possibility is environmental exposure-related CIMP
`etiology, possibly explaining CIMP2,
`in which methylation
`spreading could be a molecular signature of environmental
`exposure by targeting age-related genes (25). In this case,
`
`Table 2. Comparison between previous and current
`classifications based on 49 CRC patients
`
`Previous study*
`
`Current study
`
`Group
`
`Old
`categories
`
`Cases,
`no.
`
`New
`categories
`
`1
`2
`
`3
`
`4
`
`MSI⫹/CIMP⫹
`MSI⫺/CIMP⫹
`
`MSI⫹/CIMP⫺
`
`MSI⫺/CIMP⫺
`
`12
`16
`
`4
`
`17
`
`*Based on refs. 9 and 10.
`
`Shen et al.
`
`CIMP1
`CIMP2
`CIMP1
`CIMP-negative
`CIMP1
`CIMP-negative
`CIMP2
`
`No. of cases,
`%
`agreement
`
`12 (100%)
`14 (88%)
`2
`3 (75%)
`1
`15 (88%)
`2
`
`increased methylation may not be directly linked to the meth-
`ylation machinery, but to a constitutional predisposition to
`environment-DNA interactions, such as chronic inflammation
`or an exaggerated response to tissue injury (25, 26).
`Our data also confirm that CIMP affects many genes, not just
`a subset of genes, and show that there are two distinct CIMPs
`with potentially different causes. The optimal markers for CIMP
`remain unclear. A recent article by the Laird group (27) used a
`panel of five-markers by MethyLight method, and concluded
`that a new panel of genes outperforms the classic panel. How-
`ever, this study possibly focused mainly on the CIMP1 group and
`largely underestimated the CIMP2 group. In our study, we also
`included three of the five genes (Neurog1, RUNX3, and SOCS1)
`from the previous report. All three markers performed well to
`identify CIMP1 confirming the previous study, but they did not
`perform well for identifying the CIMP2 group. Among all of the
`methylation markers we analyzed, the original markers (all
`MINT markers) still show the best predictive values, and the
`combination of them could best define CIMP2. However, in this
`study, genetic markers performed equally well or better than
`epigenetic markers in some cases, highlighting the importance of
`integrated genetic and epigenetic analysis to resolve the heter-
`ogeneity in cancers.
`
`Table 3. Patient clinical characteristics in each cluster
`
`Characteristic
`
`Age
`Median, years
`Range, years
`Missing data (N)
`Gender (N)
`Female
`Male
`Missing data
`Location (n)
`Proximal
`Distal
`Missing data
`Stage (n)
`1 or 2
`3 or 4
`Missing data
`
`CIMP1
`(n ⫽ 22)
`
`CIMP2
`(n ⫽ 37)
`
`68
`25–88
`1
`
`69
`26–85
`1
`
`CIMP
`negative
`(n ⫽ 38)
`
`67
`29–98
`0
`
`6
`15
`1
`
`12
`7
`3
`
`8
`6
`8
`
`14
`22
`1
`
`18
`12
`7
`
`18
`14
`5
`
`9
`29
`0
`
`8
`25
`5
`
`17
`17
`4
`
`P
`
`0.90
`
`0.40
`
`0.004
`
`0.84
`
`PNAS 兩 November 20, 2007 兩 vol. 104 兩 no. 47 兩 18657
`
`Geneoscopy Exhibit 1038, Page 4
`
`

`

`Table 4. Predictive values of each marker to identify
`three clusters
`
`Marker
`
`Genetic
`BRAF-MT
`MSI-H
`Epigenetic
`hMLH1-M
`TIMP3-M
`MINT17-M
`MINT1-M
`RIZ1-M
`SOCS1-M
`MINT12-M
`RUNX3-M
`P16-M
`MINT31-M
`P14-M
`
`Genetic
`KRAS-MT
`P53-WT
`MSS/MSI-L
`Epigenetic
`MINT27-M
`MINT2-M
`Neurog1-M
`MINT31-M
`Megalin-M
`MINT1-M
`hMLH1-UM
`
`Genetic
`P53-MT
`KRAS-WT
`Epigenetic
`MINT27-UM
`MINT2-UM
`MINT31-UM
`MINT1-UM
`MINT12-UM
`Neurog1-UM
`P16-UM
`
`␬
`coefficient
`
`Specificity,
`%
`
`Sensitivity,
`%
`
`PPV,
`%
`
`NPV,
`%
`
`CIMP1
`
`99
`96
`
`100
`84
`84
`75
`93
`86
`76
`85
`81
`71
`83
`CIMP2
`
`88
`60
`37
`
`63
`68
`71
`68
`72
`67
`32
`CIMP-negative
`
`81
`63
`
`81
`68
`63
`63
`56
`60
`47
`
`0.60
`0.82
`
`0.91
`0.62
`0.52
`0.49
`0.48
`0.47
`0.44
`0.43
`0.41
`0.38
`0.36
`
`0.85
`0.30
`0.31
`
`0.39
`0.30
`0.30
`0.22
`0.21
`0.15
`0.26
`
`0.70
`0.41
`
`0.71
`0.58
`0.53
`0.55
`0.46
`0.45
`0.41
`
`53
`86
`
`86
`86
`73
`86
`50
`63
`77
`60
`64
`77
`55
`
`100
`72
`100
`
`78
`62
`59
`54
`49
`49
`100
`
`92
`81
`
`92
`95
`95
`97
`95
`89
`100
`
`91
`86
`
`100
`61
`57
`50
`69
`55
`49
`52
`50
`44
`48
`
`84
`53
`49
`
`57
`55
`56
`51
`51
`47
`47
`
`75
`59
`
`76
`65
`62
`63
`58
`59
`55
`
`88
`96
`
`96
`95
`91
`95
`86
`90
`92
`89
`88
`91
`86
`
`100
`77
`100
`
`83
`75
`73
`71
`69
`68
`100
`
`94
`84
`
`94
`95
`95
`97
`94
`90
`100
`
`PPV, positive predictive value; NPV: negative predictive value; MT, muta-
`tion; M, methylated; UM, unmethylated.
`
`Comparison of methylation for Type-C genes and Type-A genes
`Fig. 5.
`among the three clusters. A Z-score method was used to standardize the
`methylation level of each gene and each patient was assigned methylation
`scores based on the average Z-scores of either Type-C genes or Type-A genes.
`(Left) For Type-C genes, the average methylation Z-score was significantly
`higher in CIMP1 compared with other two groups (P ⬍ 0.001). (Right) For
`Type-A genes, the average methylation Z-score was significantly higher in
`CIMP2 group (P ⫽ 0.04).
`
`In summary, by integrating genetic and epigenetic analysis, we
`show that colon cancers correspond to three molecularly distinct
`subclasses of disease. Further studies will be needed to quantify
`the prognostic utility of our findings. It will also be important to
`study the epidemiology and clinical courses of these three
`subclasses of colon cancers. We suggest that molecular classifi-
`cation of all cancers by combined genetic and epigenetic analyses
`will improve our understanding of the diseases and the selection
`of optimal therapy.
`
`Materials and Methods
`Further details of tissue samples, DNA methylation analysis,
`mutation analysis, and statistical analysis used in this study are
`described in SI Materials and Methods.
`
`Tissue Samples. We collected samples of primary colorectal
`tumors and adjacent normal-appearing tissues from 97 patients
`selected solely on the basis of availability.
`
`DNA Methylation Analysis. We used different methods (MCA,
`COBRA, MSP, and bisulfite-pyrosequencing) to study the meth-
`ylation status of 27 promoter region CpG islands (see also details
`in SI Table 8).
`
`Mutation Analysis. Mutations of KRAS and p53 were determined
`by mutant allele specific PCR for KRAS codons 12 or 13 and
`single-strand conformational polymorphism and sequencing for
`p53 (10, 28). BRAF mutations at exon 11 and 15 were deter-
`mined by the pyrosequencing method.
`
`Statistical Analysis. Correlation between methylation and clini-
`cal variables were analyzed by Fisher’s exact test for categorical
`variables and Spearman correlation analysis for continuous
`
`variables. Unsupervised hierarchical clustering and K-means
`clustering analyses were used to identify potential distinct
`subgroups among CRC patients based on either epigenetic or
`
`Table 5. Predictive values for combination markers to identify each of the three clusters
`
`Positive, no.
`
`␬ coefficient
`
`Specificity, %
`
`Sensitivity, %
`
`PPV, %
`
`NPV, %
`
`2 of 5
`3 of 5
`
`CIMP 1 (BRAF-Mutation, hMLH1-Meth, Timp3-Meth, MINT1-Meth, and RIZ1-Meth)
`0.80
`95
`86
`83
`0.81
`99
`77
`94
`CIMP2 (KRAS-Mutation, MINT27-Meth, MINT2-Meth, MINT31-Meth, and Megalin-Meth)
`86
`0.40
`60
`84
`56
`2 of 5
`87
`0.30
`45
`89
`50
`3 of 5
`CIMP Negative (p53-Mutation, MINT27-Unmeth, MINT2-Unmeth, MINT31-Unmeth, MINT1-Unmeth)
`2 of 5
`0.55
`61
`100
`62
`100
`3 of 5
`0.72
`76
`100
`73
`100
`
`96
`94
`
`18658 兩 www.pnas.org兾cgi兾doi兾10.1073兾pnas.0704652104
`
`Shen et al.
`
`Geneoscopy Exhibit 1038, Page 5
`
`

`

`combined of genetic and epigenetic profiling. Bootstrapping
`cluster analysis (16) was performed to assess the reliability of
`clustering results. The difference of molecular and clinical
`variables among each cluster was analyzed by the Kruskal–
`Wallis test. Sensitivity, specificity, positive and negative pre-
`dictive values, and ␬ coefficient values were calculated to
`determine the sensitivity and specificity of either single mo-
`
`lecular marker or combination of markers to predict each
`subgroup of CRC patients.
`
`This work was supported in part by National Institutes of Health Grants
`CA098006 and CA105346. J.-P.J.I.
`is an American Cancer Society
`Clinical Research professor supported by a generous gift from the F. M.
`Kirby Foundation.
`
`1. Jemal A, Murray T, Ward E, Samuels A, Tiwari RC, Ghafoor A, Feuer EJ,
`Thun MJ (2005) CA Cancer J Clin 55:10–30.
`2. Kinzler KW, Vogelstein B (1996) Cell 87:159–170.
`3. Jones PA, Baylin SB (2002) Nat Rev Genet 3:415–428.
`4. Bock C, Walter J, Paulsen M, Lengauer T (2007) PLoS Comput Biol 3:e110.
`5. Jones PA, Takai D (2001) Science 293:1068–1070.
`6. Ahuja N, Li Q, Mohan AL, Baylin SB, Issa JPJ (1998) Cancer Res 58:5489–
`5494.
`7. Issa JPJ, Ottaviano YL, Celano P, Hamilton SR, Davidson NE, Baylin SB
`(1994) Nat Genet 7:536–540.
`8. Suzuki H, Watkins DN, Jair KW, Schuebel KE, Markowitz SD, Chen WD,
`Pretlow TP, Bin Y, Akiyama Y, van Engeland M, et al. (2004) Nat Genet
`36:417–422.
`9. Toyota M, Ahuja N, Ohe-Toyota M, Herman JG, Baylin SB, Issa JPJ (1999)
`Proc Natl Acad Sci USA 96:8681–8686.
`10. Toyota M, Ohe-Toyota M, Ahuja N, Issa JPJ (2000) Proc Natl Acad Sci USA
`97:710–715.
`11. Issa JP (2004) Nat Rev Cancer 4:988–993.
`12. Toyota M, Ho C, Ahuja N, Jair KW, Li Q, Ohe-Toyota M, Baylin SB, Issa JPJ
`(1999) Cancer Res 59:2307–2312.
`13. Xiong ZG, Laird PW (1997) Nucleic Acids Res 25:2532–2534.
`14. Colella S, Shen L, Baggerly KA, Issa JPJ, Krahe R (2003) Biotechniques
`35:146–150.
`15. Issa JP (1999) Crit Rev Oncol Hematol 32:31–43.
`
`16. Kerr MK, Churchill GA (2001) Proc Natl Acad Sci USA 98:8961–8965.
`17. Martin M (2006) Clin Trans Oncol 8:7–14.
`18. Mrozek K, Heerema NA, Bloomfield CD (2004) Blood Rev 18:115–136.
`19. Issa JPJ (2003) Clin Cancer Res 9:2879–2881.
`20. Van Rijnsoever M, Elsaleh H, Joseph D, McCaul K, Iacopetta B (2003) Clin
`Cancer Res 9:2898–2903.
`21. Shen L, Catalano PJ, Benson III, A., O’Dwyer P, Hamilton SR, Issa JPJ (2007)
`Clin Cancer Res 13:6093–6098.
`22. Iino H, Jass JR, Simms LA, Young J, Leggett B, Ajioka Y, Watanabe H (1999)
`J Clin Pathol 52:5–9.
`23. Minoo P, Baker K, Goswami R, Chong G, Foulkes WD, Ruszkiewicz AR,
`Barker M, Buchanan D, Young J, Jass JR (2006) Gut 55:1467–1474.
`24. Chirieac LR, Shen LL, Catalano PJ, Issa JP, Hamilton SR (2005) Am J Surg
`Pathol 29:429–436.
`25. Issa JPJ, Ahuja N, Toyota M, Bronner MP, Brentnall TA (2001) Cancer Res
`61:3573–3577.
`26. Issa JP, Shen L, Toyota M (2005) Gastroenterology 129:1121–1124.
`27. Weisenberger DJ, Siegmund D, Campan M, Young J, Long TI, Faasse MA,
`Kang GH, Widschwendter M, Weener D, Buchanan D, et al. (2006) Nat Genet
`38:787–793.
`28. Shen L, Kondo Y, Rosner GL, Xiao L, Hernandez NS, Vilaythong J, Houlihan
`PS, Krouse RS, Prasad AR, Einspahr JG, et al. (2005) J Natl Cancer Inst
`97:1330–1338.
`
`MEDICALSCIENCES
`
`Shen et al.
`
`PNAS 兩 November 20, 2007 兩 vol. 104 兩 no. 47 兩 18659
`
`Geneoscopy Exhibit 1038, Page 6
`
`