Identification of loci differentially methylated between pancreatic cancer and normal pancreatic tissues
We used Agilent 244K CpG island microarrays to perform 9 MCAM experiments comparing pancreatic cancer cell lines to normal pancreas. The pancreatic cancer cell lines were compared to pooled laser-captured microdissected normal pancreatic ductal tissues (A32-1, A38-5, Panc215, Panc2.5, Panc2.8 and Panc3.014), to normal pancreatic tissue (MiaPaca2 and Capan2), or to the non-neoplastic pancreatic ductal line HPDE (AsPC1).
To identify differentially methylated probes we used the previously validated cutoff (13
) of ~4-fold (normalized log2
signal ratios (NLR) ≥2 or ≤−2) differential methylation between pancreatic cancers and normal pancreata (). This cut-off was established using the same MCAM strategy comparing the pancreatic cancer cell line Panc-1 to the epithelial line, HPDE and validating the methylation status of 31 genes using bisulfite sequencing and MSP. By this criterion, there were 41,606 probes, representing 10,012 loci with differential methylation in at least one of the nine MCAM experiments, a considerably more extensive list than we identified in our pilot MCA experiment using the 44K array (13
). To identify from this list genes commonly methylated in pancreatic cancers, we then calculated the mean of differential methylation ratios of the nine cancer/normal pairs and the combined Z-scores for each of these 41,606 probes and filtered out probes for which the mean of the nine NLRs was ≥ −2 but ≤2 (Note Z scores reflects how far a probe log-ratio value is in relation to the Gaussian distributions of probes with similar melting temperatures on the Agilent array). This criterion allowed us to identify 4672 probes for which the mean combined Z-score was either ≥ −4.67 or ≤ 5.1. To refine our gene list, we then filtered out outlier probes for which the mean combined Z-scores was ≥ −5 but ≤ 5 which left 4,662 probes, corresponding to 1,658 known loci (Supplemental Table 1
). Thus, the gene lists identified using the mean NLR of ≥ −2 but ≤ 2 and the mean combined Z-scores was ≥ −5 but ≤ 5 were almost identical. Among these 1,658 loci, 1,226 were hypermethylated (1,206 genes, Supplemental Table 2
) and 423 were hypomethylated (379 genes, Supplemental Table 3
) in the pancreatic cancer cell lines compared to normal pancreatic tissue.
An overview of experimental strategy to evaluate differential DNA methylation in pancreatic cancer cells compared to normal pancreas. NP: Normal Pancreas.
We found that 60% of hypermethylated probes were located inside known gene bodies, while 31.3% were located in their promoters, and 8.42% were located downstream of known genes (, left panel). These proportions were very similar among the hypermethylated genes whose expression was downregulated (, right panel) as well as to the overall distribution of probes on the Agilent chip (). We also found a very similar distribution of hypomethylated probes (, left panel). However, when we included only hypomethylated probes that were associated with upregulated genes, we found that 48.5% of probes were located in the promoter region (, right panel). This analysis is consistent with the observation that hypermethylation within the gene body as well as the promoter can result in downregulation, whereas gene upregulation is associated promoter hypomethylation but not gene body hypomethylation.
Distribution of Agilent Human CpG island 244K probes covering known genes A, global distribution. B, Distribution of hypermethylated probes. C, Distribution of hypomethylated probes.
Since cancer-associated CpG island hypermethylation tends to extensively involve the affected CpG island we used MSP to confirm differential CpG island methylation identified by MCAM. Specifically, to validate our MCAM results we used TMSP to examine the methylation status of 24 aberrantly hypermethylated genes in 69 pancreatic cancers (43 xenografts of primary pancreatic cancers, 15 primary pancreatic cancer tissues, and 11 cell lines (i.e.
BxPC3, Capan1, CFPAC1, HS766, Panc1, Su8686, A2-1, A6L, Panc198, Panc486, and Panc8.13), and 16 normal pancreatic tissue. We confirmed the differential methylation of 21 of the 24 genes tested (Supplemental Table 4
), demonstrating the robustness of our MCAM strategy.
Several genes were methylated in >70% of pancreatic cancers and 10% or less of normal pancreata by MSP including BMI1, ID4, CACNA1H, IRF5 and SMOC2.
Aberrantly methylated and differentially expressed genes
We have previously examined gene expression in a set of six pancreatic cancer cell lines (A32-1, A38-5, Panc215, Panc2.5, Panc2.8 and Panc3.014) and the pancreatic non-neoplastic line, HPDE, before and after epigenetic treatment (5-Aza-dC, TSA, and the combination) using Affymetrix ST1.0 Exon Arrays (18
). Here, we merged this analysis with the results of our MCAM experiments and compared the baseline and post-treatment gene expression profiles of these six pancreatic cancer cell lines focusing on genes shown to be hypermethylated or hypomethylated in our MCAM experiments. We used the two-way ANOVA analysis to identify expression changes between cancer cells and the non-neoplastic line HPDE. We defined genes as differentially expressed based on a fold change criteria of ±2-fold. Among the hypermethylated genes, 96 genes were underexpressed in pancreatic cancer cells compared with HPDE cells and the expression of 675 were induced by 5-Aza-dC and/or TSA treatment either alone or in combination (p-value <0.05) (Supplemental Table 2
). 77 genes were both underexpressed and responsive to epigenetic treatment (). If we employed a more stringent cut-off for underexpressed genes (>3-fold reduction in expression), 37 genes were both underexpressed and responsive to epigenetic treatment (, bold genes). Among the hypomethylated genes, 24 genes were overexpressed in the pancreatic cancer cell lines compared with non-neoplastic pancreatic duct HPDE cells (Fold Change ≥2) () and 243 genes were significantly induced by epigenetic treatment but not necessarily differentially expressed (p-value <0.05 with a Fold-Change >2; Supplemental Table 3
List of hypermethylated and downregulated candidate genes, responsive to epigenetic drug treatment.
List of hypomethylated and upregulated candidate genes.
Even by a modest criteria of ±2-fold for differential expression, only a minority of genes displayed methylation-associated alterations in expression compared to those that underwent aberrant methylation ( and ). We previously made this same observation in an MCAM analysis of the Panc-1 vs. HPDE cell lines (13
). However, when we next examined the relationship between global methylation and gene expression by comparing mean expression levels in the 6 pancreatic cancer cell lines, we found that hypermethylated genes (Mean of NLR >1) were significantly less expressed in the pancreatic cancer cells than non-hypermethylated genes (Mean of NLR <1, ). And conversely, we found that the expression of hypomethylated genes (Mean of NLR <-1) was significantly higher than non-hypomethylated genes (Mean of NLR >-1, ). These results support the general role of hypermethylation as a mechanism of aberrant silencing in cancer as well as reinforcing the association between localized promoter hypomethylation and gene overexpression.
Figure 3 Correlation between gene methylation and gene expression levels in pancreatic cancer samples (A32-1, A38-5, Panc215, Panc2.5, Panc2.8 and Panc3.014). A, Boxplots stratifying mean gene expression (intensities values from Affymetrix Exon Array ST1.0) by (more ...)
Additionally, we checked the expression profile of our newly identified hypermethylated and hypomethylated genes previously analyzed by the SAGE (serial analysis of gene expression) method (17
). The advantage of this SAGE analysis is that it quantified 1–2 million SAGE tags in each of the 24 pancreatic cancer samples 2 laser capture microdissected normal pancreatic duct samples and the non-neoplastic pancreatic duct cell line, HPDE. We defined silenced genes, neo-expressed genes, underexpressed, and overexpressed genes. Silenced genes were defined as genes for which the number of tags was simultaneously <1 in more than 50% of the 24 pancreatic cancer samples and >10 in both normal samples and were added to . Underexpressed genes were those with a 2-fold decrease in the number of tags in pancreatic cancer compared to either one of the two normal samples (the percentage of pancreatic cancers with underexpression is noted in the SAGE columns of Supplemental Table 2
). Neo-expressed genes were defined as genes for which the number of tags was simultaneously >10 in more than 50% of the 24 pancreatic cancer samples and <1 in both normal samples and were added to . Overexpressed genes were defined as those with a 2-fold increase in the number of tags in pancreatic cancer compared to either one of the two normal samples (the percentage of pancreatic cancer samples with overexpression is noted in Supplemental Table 3
Genes known to be hypermethylated in pancreatic cancer
Among the list of 1206 candidate hypermethylated genes, were virtually all of the genes we previously identified using MCAM on the Agilent 44K microarray and validated by bisulfite sequencing and MSP in Panc-1 cancer cells, i.e. BAI1
, EYA4, BNC1, HOXA5, PAX7, SOX14, TLX3, NRXN1, CNTNAP2, PKP1, ACTA1, MDFI
, EVC2, LIN28, NRN1, PENK
, m FAM84A
, and ZNF415
We also confirmed the differential methylation profiles of genes whose epigenetic silencing was revealed by our group and others. These genes include NPTX2
, FOXE1, PAX6, BNIP3, GADD45B
), HIC1, HS3ST2
), IRF7, CCNA1, ALPP, CEBPA
Our list included the tumor suppressor genes STK11, and WT1, the transcription factor RUNX3, the transcriptional repressor CBFA2T3, the secreted glycoprotein and chemorepulsive factor SLIT2 and the aryl-hydrocarbon receptor repressor, AHRR.
Note that our selection criterion was designed to identify genes that are commonly differentially methylated in pancreatic cancer. Thus, several important genes that are infrequently methylated in pancreatic cancer did not meet our commonly methylated filter including p16/CDKN2A
since these genes are uncommon targets in pancreatic cancer (24
). Among our nine pancreatic cancer/normal pancreas methylomes examined, three had one hypermethylated probe in the CDKN2A
promoter. Other genes hypermethylated in pancreatic cancer (RELN
)) did not meet our filter for commonly methylated genes, although several MCAM probes covering these genes were significant for hypermethylation. Finally, some genes hypermethylated in pancreatic cancer were not identified because their promoters were not interrogated by the MCAM method (e.g. SPARC
, MMP3, MMP2, MMP9
Genes involved in stem cell pluripotency are hypermethylated in pancreatic cancer
We found DNA hypermethylation of genes involved in stem cell pluripotency, such as the intestinal stem cell marker and chromatin structure regulator BMI1
, the genes encoding bone morphogenetic proteins BMP3, BMP6, the transcription factors FOXD3, CDX2, UTF1, and T, as well as NR5A1
genes. Additionally, hypermethylation of FOXA1
was associated with gene relative underexpression in pancreatic cancer by SAGE analysis (Supplemental Table 2
The WNT signaling pathway is a target of hypermethylation in pancreatic cancer
Interestingly, numerous actors and mediators of the WNT signaling pathway were found to be hypermethylated in the majority of our pancreatic cancer samples and account for an enrichment ratio of 0.167 in our Ingenuity pathway analysis. As expected, we found hypermethylation involving genes previously shown to be epigenetically silenced both in pancreatic and other types of cancer (14
) and encoding either the WNT ligands WNT5A, WNT7A,
, or the cell-surface receptor FZD9
, as well as the cytoplasmic transducer APC2,
the nuclear factors SOX1, SOX7, SOX14,
and the pathway inhibitors FRZB, SFRP1, SFRP2, KREMEN2, NKD2
. Finally, we also found hypermethylation of the tumor suppressor candidate HIC1
, which encodes a transcriptional repressor responsible for abnormal survival circuits through nuclear release of WNT signaling pathway transcription factors (28
We also found DNA hypermethylation surrounding genes of the WNT pathway and overexpressed in different types of cancer, either in the promoter region (WNT3 and SOX3), within the gene body (FZD1, FZD2, FZD7, SOX15 and SOX18), or both (FZD10, SOX11, and MDFI).
Several hypermethylated WNT pathway genes demonstrated differential gene expression in multiple pancreatic cancer cell lines versus HPDE either by Affymetrix Exon Array (WNT5A
, FZD2, SFRP1
or by SAGE analysis (WNT10A
) (Supplemental Table 2
Eleven other WNT pathway genes showed no differential expression by Affymetrix Microarray or SAGE analysis (Supplemental Table 2
)), although 5-Aza-dC treatment increased expression of 7 of these 11 genes (FZD9
, and KREMEN2, Supplemental Table 2
), suggesting that DNA hypermethylation influences their expression.
Among the WNT Pathway hypermethylated gene members WNT5A
are of particular interest because they had evidence of epigenetic silencing (loss of expression in pancreatic cancers relative to normal pancreas and induced expression with 5-Aza-dC treatment (Supplemental figure 1
is epigenetically inactivated in colorectal and other cancers (29
) and suspected to facilitate tumorigenesis by abrogating WNT signaling.
WNT5A is a member of the non-transforming WNT protein family frequently inactivated by tumor-specific methylation in colorectal cancer (30
), although opposing roles of WNT5A in cancer have been described (31
). For instance, one study found WNT5A
overexpression in pancreatic cancers (32
) whereas our gene expression profiling found underexpression of WNT5A
compared to HPDE cells (2.9-Fold by Affymetrix Exon Array) and compared to normal pancreatic ductal epithelium by SAGE. Another identified hypermethylated and downregulated gene implicated in pancreatic differentiation was SOX15
). To our knowledge, hypermethylation and silencing of SOX15
have not been previously reported in pancreatic cancer.
The finding that the WNT signaling pathway is a common target of DNA methylation in colon and other cancers led Jones and Baylin (34
) to propose that loss of epigenetic gatekeepers such as WNT signaling pathway inhibitors locks cells into stem-cell like states that foster abnormal clonal expansion. The pancreatic cancer genome project provided evidence for aberrant WNT pathway signaling during pancreatic tumorigenesis as somatic mutations in the pathway were occasionally identified in pancreatic cancers (17
). In this study, the enrichment of aberrant methylation of the WNT pathway gene members supports a significant role for alterations of this pathway during pancreatic tumorigenesis.
Hypermethylation of cell adhesion molecules during pancreatic carcinogenesis
Numerous genes involved in cell adhesion and junctions were also hypermethylated in pancreatic cancer samples, including members of the cadherin superfamily CDH2
, and RET
. Interestingly, the pancreatic cancer genome project identified occasional somatic mutations in cadherins and protocadherins in pancreatic cancers (17
). Of these genes, hypermethylation of CDH4
was associated with gene underexpression in pancreatic cancers by Affymetrix Array or SAGE data, respectively ( and Supplemental table 2
Hypermethylation of gene clusters and families as a common feature of human cancer
We found that numerous families of genes were simultaneously hypermethylated in our pancreatic cancer samples compared to normal pancreas (Supplemental Table 2
, bold genes). These include families of transcription factors and, in particular, homeobox genes clustered on chromosome 7p15.2 (EVX1
, on 17q21.32 (HOXB1
, on 12q13.13 (HOXC11
, on 2q31.1 (EVX2, HOXD1, HOXD10
(Supplemental Figure 2
). These also include chromobox genes on 17q25.3 (CBX2
, Iroquois homeobox genes on 16q12.2 (IRX3
, IRX5, IRX6)
and on 5p15.33 (IRX1
, IRX2, IRX4)
, and SIX homeobox genes on 2p21 (SIX2
Hypermethylation of clusters of homeobox genes has been reported in other types of cancer, such as lung cancer and astrocytomas (35
), but not in pancreatic cancers. Hypermethylation of a gene region such as the homeobox genes is most likely related to deregulation of the well orchestrated epigenetic machinery involved in the spatiotemporal regulation of these developmental genes. This finding supports other evidence that gene neighborhood is required for an efficient coordination of chromatin modifications during carcinogenesis as well as developmental processes (37
). An important contributor to these epigenetic marks is the Polycomb group of proteins, which has been shown to be responsible for the concerted silencing of homeobox gene clusters. From our hypermethylated candidate gene list, we found 288 known targets of Polycomb proteins (Supplemental Table 2
), including families of highly homologous genes that do not belong to clusters, such as LIM homeobox, NK2/3/6 homeobox, Paired box, POU class homeobox, sex determining region Y-box and T-box transcription factors, which confirms the very tight link between specific hypermethylation and Polycomb protein complexes occupancy in cancer. However, we found aberrant hypermethylation of several other gene family members including members of the G-protein-coupled receptor superfamily, adenylate cyclases, and voltage-dependant calcium and potassium channels. Sequence similarities between these family members are likely responsible for their similar cancer epigenetic patterns. For example, SINE and LINE repetitive elements are enriched at sites of aberrant hypermethylation (38
). These sequence-based contributions to aberrant methylation indicate that selection driven by gene silencing need is not responsible for causing methylation marks to arise during tumorigenesis. The fact that only a relatively small proportion of hypermethylated genes showed evidence of downregulation or silencing (HOXD9
, Supplemental Table 2
) supports this hypothesis.
Newly identified hypermethylated and underexpressed candidate genes
Of the 81 hypermethylated and underexpressed genes identified by our analysis, 33 have, to our knowledge, not been previously identified as hypermethylated either in pancreatic or other cancers. The differential methylation of several of these genes including ID4, FOXF2, FOXG1, IRF5, SIM1, SMOC2,
was confirmed by MSP (Supplemental table 4
). Twelve genes demonstrated gene silencing in pancreatic cancer compared to normal pancreatic duct by SAGE (including HOXD8, BNIP3, FOXF2
many of which have not been identified as differentially methylated in cancer ().
Many hypermethylated and underexpressed genes identified have been implicated in tumorigenesis including tumor suppression (RINT1
)), proliferation (GAS6
)), apoptosis (IRF5
)), cell adhesion (CLDN10
, LAMA1, AMOTL1
), or transcriptional regulation (FOXG1
). Interestingly, although SEMA5A was recently proposed as a marker for aggressive pancreatic cancers we find most pancreatic cancers underexpressed SEMA5A. (43
). Other genes of note include the Type IV collagen genes COL4A1
which have not been described as aberrantly methylated in cancer.
Finally, to further evaluate genes identified as both hypermethylated and underexpressed by array analysis, we searched for commercially available antibodies to proteins corresponding to genes on our list () and selected an antibody to ZBTB16 for further analysis. The transcription factor ZBTB16 (a.k.a PLZF), has been shown to regulate apoptosis and proliferation, as well as the Hox family of genes during development, and has been proposed as a candidate tumor suppressor gene in acute promyelocytic leukemias. By MSP, we confirmed ZBTB16
promoter methylation in 35.2% of the pancreatic cancer samples tested, while none of the DNAs from 16 normal pancreas samples showed methylation. By immunohistochemistry, ZBTB16 protein expression was not detectable in 38 (15.4%) of 247 pancreatic cancers evaluated. Reduced expression relative to normal pancreas was found in the majority of pancreatic cancers (193/211, 94%) (Mean histo-score 2.7 vs 6.9, p<0.0001; Supplemental Figure 3
). Interestingly, aberrant methylation and silencing of ZBTB16
in pancreatic cancers may be functionally significant as it functions to repress miR-221 and miR-222 which is overexpressed in pancreatic cancer cells (45
). MiR-221/222 expression is associated with enhanced proliferation and dedifferentiation in melanoma cells (46
Newly identified hypomethylated and overexpressed candidate genes
Since hypomethylation is associated with gene overexpression in pancreatic and other cancers (9
), we also identified candidate hypomethylated genes using the MCAM assay. Of the 379 unique hypomethylated candidate genes identified, 31 have been described as aberrantly methylated in different cancers (Supplemental Table 3
). This list confirmed the aberrant hypomethylation of NDN
genes, previously described in pancreatic cancer (14
). Similarly to hypermethylated genes, some known hypomethylated genes did not meet our criteria for frequent hypomethylation. However, when we examined the NLR for probes covering MUC4
, genes previously identified as hypomethylated in pancreatic cancer (47
), there was evidence for hypomethylation (NLR ≤ −2) in at least one of the nine differential methylomes of pancreatic cancer/normal pancreas studied. The promoter regions of other hypomethylated genes in pancreatic cancers (S100A4
, MSLN CLDN4
) were not interrogated by our MCAM method.
Interestingly, while approximately half of the hypermethylated candidate genes had multiple probes meeting criteria for hypermethylation, the vast majority (84.7%) of hypomethylated gene candidates harbored only one hypomethylated probe (mean of the nine NLR ≤ −2 and mean combined Z-scores was ≤ −5) (Supplemental Table 3
Our hypomethylated gene list contained chromatin modifiers, including the histone methyltransferases SETD8
(lysine) and PRMT1
(arginine), the histone ubiquitin ligase CTR9
, the histone demethylase KDM6A
, the chromatin assembly protein HIRIP3, and the histone acetyltransferase EP400. These genes were overexpressed in pancreatic cancer samples compared to HPDE cells and/or normal ductal epithelium by SAGE, and HIRIP3
was found to be neo-expressed in 41% of pancreatic cancer samples (Supplemental Table 3
Three oncogenes, JUNB
, and FOS
were hypomethylated in our analysis. While FOS
undergoes hypomethylation during liver tumorigenesis, JUNB
have not been previously reported to be hypomethylated. Interestingly, MYB was found to be neo-expressed in 31% of pancreatic cancer compared to normal samples by SAGE (Supplemental Table 3
Of the 44 hypomethylated genes identified in our analysis and found to be overexpressed by both ExonArray and SAGE methods, 34 have, to our knowledge, not been described as aberrantly methylated in cancer. Among them, DDHD2, CREB1 and PIM2 have been proposed as oncogenes, while ITGB2 (CD18) and MAPRE2 are involved in cell invasion and metastasis. HELLS is a putative stem cell marker, while MTRR restores methionine synthase for folate metabolism and polymorphisms have been associated with pancreatic cancer risk.
Differentially Methylated loci of pancreatic cancer samples compared to normal lymphoid tissues
To further validate our list of hyper- and hypomethylated candidate markers, we performed three additional MCAM experiments, in which we compared the methylation profiles of pancreatic cancer cell samples with matched normal DNA either from peripheral blood lymphocytes (A32-1 and Panc215) or from normal spleen (A38-5). We used the cutoffs described above for the NLR and combined Z-score to select probes that were hyper- or hypomethylated in pancreatic cancer cells compared to matched lymphocytes or spleen, respectively ().
An overview of experimental strategy to evaluate differential DNA methylation in pancreatic cancer cells compared to normal lymphoid tissue. NP: Microdissected normal pancreatic ductal epithelium; ly: peripheral blood lymphocytes.
We found 11,029 and 14,863 differentially methylated probes in Panc215 and A32-1 samples compared to the patient’s matched lymphocytes DNA, respectively. We found 13,680 differentially methylated probes in A38-5 cancer cells compared to their matched normal spleen. We then compared these probes to those differentially methylated in Panc215, A32-1 and A38-5 cancer cells, vs. the same pooled microdissected normal pancreatic ductal tissue. Since some methylation patterns reflect tissue specific differences, we found more differentially methylated probes when comparing pancreatic cancer DNA to a patient’s matched lymphoid tissue than to unmatched normal pancreatic duct. We found 4,434 and 6,200 differentially methylated probes both in cancer compared to lymphocytes and in cancer compared to normal pancreatic tissue, for Panc215 and A32-1 samples, respectively. Moreover, we found differentially methylated 2,956 probes in both Panc215 and A32-1 samples, and both in cancer compared to lymphocytes and cancer compared to normal pancreatic tissue, corresponding to 786 known loci (). Among these loci, 669 unique genes were part of our hypermethylated and hypomethylated candidate lists (Supplementary Tables 2
). These data indicate that while more than half of our differentially methylated candidate genes in pancreatic cancers relative to normal pancreata are also differential methylated relative to peripheral blood lymphocytes, many genes that are methylated in pancreatic cancers are normally methylated in other tissues.
We found differentially methylated 6,858 probes that were both in cancer compared to spleen and cancer compared to normal pancreatic tissue (). Among these probes, 1,896 were also differentially methylated in pancreatic cancer cells compared to normal pancreatic tissue and in pancreatic cancer cells compared to lymphocytes. These corresponded to 530 known loci, and 489 genes present in our hypermethylated and hypomethylated gene lists (Supplemental tables 2
Among these 489 genes, 32 genes show DNA hypermethylation associated with downregulation in cancer cells compared to normal cells (). These latter genes, with a more specific hypermethylation profiles in pancreatic cancer tissues are likely to be better candidates as pancreatic cancer markers. Indeed, several of these genes have already been proposed as candidate biomarkers of prognosis (BNC1
)), or response to treatment (WNT5A
)) in other types of cancer.
In summary, we provide a comprehensive list of candidate aberrantly methylated in pancreatic cancers relative to normal pancreatic duct samples, including novel aberrantly methylated and epigenetically silenced genes validated in pancreatic cancer tissue samples, as well as commonly methylated signaling pathways and gene families. We anticipate that our integrated analysis of pancreatic cancer gene methylation and transcription will facilitate the further investigation of pancreatic cancer biomarkers for diagnostic, prognostic and therapeutic applications.