Both Hypomethylation and Hypermethylation in a 0.2-kb Region of a DNA Repeat in Cancer Free

Both hypermethylation (1, 2) and hypomethylation (3) of DNA have been observed in most tested cancers but in different sequences (4-7). Many specific gene regions become hypermethylated, and some other gene regions and many noncoding DNA repeats become hypomethylated during carcinogenesis (8-10). Nonetheless, hypomethylation and hypermethylation in different parts of the genome in various cancers have been found not to be significantly associated with each other (4, 6, 11). Therefore, cancer-linked DNA hypomethylation is not simply a response to cancer-linked hypermethylation or vice versa.

Here, we show for the first time that, relative to normal somatic tissues, cancers can display both hypomethylation and hypermethylation within the same small region on the same DNA molecule. We analyzed NBL2, a tandem 1.4-kb repeat with a complex sequence. We found recently by Southern blot analysis that NBL2 exhibits either predominant hypermethylation or hypomethylation at HhaI sites in 89% of 18 studied ovarian carcinomas and 84% of 51 Wilms' tumors (12). We also observed hypomethylation at NotI sites in some ovarian carcinomas. Itano et al. (13) showed that hypomethylation at NotI sites was an independent prognostic indicator in hepatocellular carcinoma patients. NBL2 is often hypomethylated at NotI sites in neuroblastomas (hence, the name NBL2) and hepatocellular carcinomas (14, 15). This primate-specific sequence (15) is CpG rich (61% C + G; 5.7% CpG). It is present in ∼200 to 400 copies per haploid human genome, mostly in the vicinity of the centromeres of four of the five acrocentric chromosomes (12), repeat-rich regions for which only little sequence information is available.

The powerful method that we used to analyze NBL2 methylation in ovarian carcinomas, Wilms' tumors, and various control cell populations is the hairpin-bisulfite PCR variant of genomic sequencing (hairpin genomic sequencing) developed by Laird et al. (16). In bisulfite-based genomic sequencing, bisulfite causes deamination of unmethylated C residues, whereas methylated C residues (usually only in CpG sequences) are resistant (17). Hairpin genomic sequencing allows analysis of methylation of every CpG dinucleotide pair (dyad) in a given region on covalently linked DNA strands of a restriction fragment. Analysis of both CpGs of a dyad is of special interest because CpG dyads are generally assumed to be symmetrically methylated or symmetrically unmethylated, except during DNA replication. Hairpin genomic sequencing also unambiguously differentiates naturally occurring sequence variation from bisulfite- and PCR-mediated C-to-T conversions at unmethylated cytosines. Our results show that NBL2, which does not seem to be a gene (12), is especially sensitive to multiple diverse DNA methylation changes during oncogenesis, even within the same 0.2-kb region on the same DNA molecule.

We analyzed the 1.4-kb NBL2 repeat by hairpin genomic sequencing (16) as shown in Fig. 1. In DNA clones resulting from bisulfite treatment and PCR, a genomic m5CpG, the predominant site of vertebrate DNA methylation, will appear as CpG because it escaped bisulfite deamination, and an unmethylated CpG will become TpG due to cytosine deamination followed by amplification. In hairpin genomic sequencing, strand ligation results in the sequence information from both genomic strands of a DNA fragment being present in each strand of the resulting DNA clone (Fig. 1A). Corresponding CpG positions in the two halves of one strand of a DNA clone are compared to determine the methylation status of CpG dyads in the template DNA molecule. For simplicity, we will use the following terms for the DNA clones, which describe the CpG dyad methylation status of the molecule that gave rise to the clone (Fig. 1B): M/M (symmetrically methylated), U/U (symmetrically unmethylated), M/U (hemimethylated with the lower strand unmethylated), and U/M (hemimethylated with the upper strand unmethylated). Not only does hairpin genomic sequencing resolve a symmetrical methylation pattern at a CpG dyad from hemimethylation, but also it allows an unmethylated CpG to be unambiguously distinguished from germ-line C-to-T changes (Fig. 1B). This is especially useful for DNA repeats because of their appreciable sequence variation (16).

From hairpin genomic sequencing of each normal tissue or cancer DNA, we obtained a single or predominant PCR band of the expected size (508 bp) from which 12 to 32 clones were generated and sequenced (Fig. 2). Given the originally self-complementary nature of the ligated DNA for bisulfite treatment and the specificity of bisulfite for denatured DNA, various controls were done to make sure that hairpin genomic sequencing did not yield artifacts. First, we should see essentially only CpG methylation (16) because we were examining postnatal tissues (18). As expected, we found only 0% to 0.3% of non-CpG C residues per tissue sample (0.1% overall; data not shown). In addition, only 0.6% of C residues persisted as C in hairpin genomic sequencing clones from a NBL2-containing Escherichia coli plasmid (Fig. 2B). In addition, we checked the completeness of bisulfite modification by digesting all hairpin PCR products with Tsp509I (recognizing AATT). Gel electrophoresis of the digests indicated complete digestion due to bisulfite-mediated C deamination at genomic AACC or AACT in NBL2 (data not shown). Hairpin genomic sequencing of a NBL2 plasmid methylated in vitro at most CpG sites by M.SssI showed that most of the CpG C residues were retained in the clones.

In addition, we also verified that we were not amplifying only a few template molecules by using a 1:20 dilution of the bisulfite-treated DNA instead of the undiluted sample for PCR. A strong PCR product band was obtained with or without dilution from each sample. We conclude that the sequenced molecular clones represent the heterogeneity in the sample DNA, which is consistent with their epigenetic and genetic sequence diversity (Fig. 2). Lastly, we did a mixing experiment with hairpin PCR products amplified as 1:0, 1:3, 1:1, 3:1, and 0:1 mixtures of M.SssI-methylated and unmethylated NBL2 plasmid. We obtained approximately the expected ratios of BstUI-sensitive (CGCG) to BstUI-resistant sites in the PCR product mixtures. Therefore, there was no appreciable selection during the PCR for templates that were methylated or unmethylated as is sometimes found for bisulfite-treated DNA even when methylation-specific primers are avoided (19).

The normal cell subtype of origin for ovarian epithelial carcinomas is an uncertain minor ovarian epithelial component (20) and for Wilms' tumors, an embryonic kidney remnant. Because they are not available, we compared these diverse cancers with various normal postnatal somatic tissues. First, we had to identify CpG sites with invariant methylation status in somatic control tissues (brain, spleen, and lung from different normal individuals). There was a surprisingly high degree of conservation of a complex methylation pattern at NBL2 in the normal somatic tissues (Figs. 2 and 3A) in contrast to the usual findings of either very heterogeneous methylation patterns from molecule to molecule or almost complete methylation or lack of methylation in a given DNA region (21-23). Among the 91 NBL2 DNA clones from somatic controls subject to hairpin genomic sequencing, 7 of the 14 CpG sites were always symmetrically methylated (CpG2, CpG3, CpG5, CpG8, CpG10, CpG11, and CpG12). Two nonadjacent CpGs were never symmetrically methylated (CpG6 and CpG14). One of these, CpG14, was always U/U, and the other, CpG6, was usually U/U but occasionally U/M or M/U. CpG13, which is exactly adjacent to always-unmethylated CpG14 was often replaced by GpG and hence could not be methylated. However, whenever it was not replaced, it was always M/M despite its immediate U/U neighbor (Figs. 2 and 3A). Normal sperm had no M/M sites in this subregion (Fig. 2B), consistent with previous results from various tandem DNA repeats (9).

None of the 146 DNA clones from the 10 cancers had the conserved methylation pattern of normal somatic controls (Fig. 2). Moreover, 56% of the cancer clones had a mixture of both hypomethylated and hypermethylated CpG sites. These methylation changes were defined by the loss of the normally conserved M/M status at CpG2, CpG3, CpG5, CpG8, CpG10, CpG11, or CpG12 and the gain of M/M status at CpG6 or CpG14, sites never normally symmetrically methylated (Figs. 2 and 3). The overall methylation status at each of these nine CpG sites in the cancers was significantly different from that in the somatic controls (P < 0.005, adjusted for multiple comparisons).

Spreading of de novo methylation along a DNA region has been observed (24-27). We did not see evidence of predominant spreading of de novo methylation or demethylation because a pairwise comparison of neighboring CpG sites in NBL2 in the cancers indicated that there was no statistically significant bias toward adjacent sites having the same methylation status. Furthermore, at CpG6 and CpG8, which are separated by only 6 bp, there were seven clones from four cancers (OvCaN and WT4, WT9, WT21, and WT67) that exhibited opposite methylation changes [i.e., hypermethylation at CpG6 (M/M) and hypomethylation at CpG8 (M/U); Fig. 2]. Overall, the methylation changes in many of the clones suggest multiple discontinuous hits of demethylation and de novo methylation within a 0.2-kb region during carcinogenesis.

Nonetheless, there were some DNA clones whose methylation patterns suggested spreading of methylation or demethylation. Some had all 14 CpG dyads unmethylated or all methylated (Fig. 2). Others had the first five or six CpG sites unmethylated on at least one strand. The frequency of clones with no methylation in the first five CpG sites was significantly higher than expected if the methylation at each site was independent, as was the combined frequency of fully methylated or fully unmethylated clones (both Ps < 0.0001). In summary, there seems to be spreading of altered DNA methylation patterns in some, but not most, of the copies of NBL2 in the examined cancers.

In the examined NBL2 subregion in the somatic controls, 1.6% of the CpG sites and 15% of the somatic control clones displayed hemimethylation. The hemimethylation frequencies rose in the cancers to 3.4% of the CpG sites in 47% of the ovarian cancer clones and 6.6% of the CpG sites in 71% of the Wilms' tumor clones. Because incomplete bisulfite modification was observed at only ∼0.1% of the non-CpG C residues, we are truly describing hemimethylation at CpGs. There was also a change in the distribution of hemimethylation in the cancer DNAs. Many more of the 14 CpG positions were subject to hemimethylation in the cancers than in the somatic controls. In addition, some of the hemimethylated CpGs at a given position in the cancer clones displayed a strong bias for demethylation of the top or the bottom strand (Table 1). Hemimethylated CpG dyads in cancer and control clones usually did not occur as runs but rather had the closest CpG on either side as a M/M or U/U dyad. Furthermore, of the 27 cancer clones containing more than one hemimethylated CpG site, 15 had hemimethylated dyads of opposite polarity with respect to which strand was unmethylated.

Some normally M/M CpG sites seemed to be more likely to become demethylated in both ovarian carcinomas and Wilms' tumors than others (Fig. 3). To test the significance of this finding, we did a pairwise comparison of methylation changes in cancer clones at the seven normally M/M sites and also at the two normally unmethylated CpG dyads. In both Wilms' tumor and ovarian carcinoma groups, the following significant differences were observed: demethylation at CpG12 was more frequent than at CpG8 or CpG11, demethylation at CpG2 was more frequent than at CpG5, and demethylation at CpG3 was more frequent than at CpG11 (P < 0.05, adjusted for multiple comparisons). With respect to the two positions that were never normally symmetrically methylated, CpG6 was significantly more prone to cancer-associated hypermethylation (conversion to M/M) than CpG14 (P < 0.00001) in ovarian carcinomas although not in the Wilms' tumors.

There was also evidence of preferential epigenetic patterning involving multiple CpG positions in the sequenced NBL2 region from the cancers. Eleven cancer clones derived from four cancers had the following methylation status: CpG4, U/M; CpG12, M/U; CpG1, CpG3, CpG5, CpG6, and CpG10, U/U; CpG7, CpG8, CpG9, CpG11, CpG13, and CpG14, M/M (Fig. 2, first row for WT4, second row for WT9 and OvCaO, and third row for WT67). This methylation pattern constitutes changes from the normally conserved methylation status of the five underlined CpG sites.

To try to explain site preferences for cancer-linked methylation changes or for the conserved methylation patterns in somatic controls, we looked for possible effects of the sequence 1 to 3 bp on either side of each CpG. No rules for predicting the methylation status in the somatic controls or cancers based on adjacent sequences could be deduced just from the region subjected to genomic sequencing. However, Southern blot analysis of NBL2 arrays gave us further insights as described below.

We searched for a few CpG sites whose methylation status could be used to distinguish all the cancer-derived molecular clones from all the somatic control clones. We found such sites with 100% predictive power by generating a classification tree from the data. All but two of the cancer-derived clones displayed symmetrical methylation at CpG6 (M/M) or demethylation at CpG10 (U/U or U/M); none of the somatic control-derived clones had these epigenetic attributes. The two exceptional tumor clones could not display hypomethylation because CpC or CpT replaced CpG6 (Fig. 2B, WT67 and WT21, hyphen at the CpG6 position). Those two clones exhibited hypermethylation at CpG14, which distinguishes them from all somatic control-derived clones. Our ability to distinguish all NBL2 cancer clones from all NBL2 somatic control clones also shows the purity of the cancer DNA samples used for this analysis.

All cancers in this study and an additional 13 ovarian carcinomas and 46 Wilms' tumors had been examined by Southern blot analysis for methylation at HhaI and NotI sites with a 1.4-kb NBL2 probe (12). HhaI digests of DNAs from various postnatal somatic control tissues from 15 individuals gave very similar distributions of intermediate molecular weight hybridizing fragments (e.g., see Fig. 4A), whereas NotI digests all gave very high molecular weight hybridizing fragments (12). A comparison of cancers and somatic controls revealed predominant hypermethylation at HhaI sites in 81% of the 69 cancers and hypomethylation in 4% of them (e.g., Fig. 4A). Advantages of Southern blot analysis are that it can show long-range methylation patterns not identifiable by genomic sequencing, especially in tandem repeats, and it provides results from the population average of all the copies of the examined sequence.

In this study, we compared the above Southern blot results describing methylation at all HhaI sites throughout the NBL2 arrays and the hairpin genomic sequencing results for methylation at every CpG in a 0.2-kb NBL2 subregion (Table 2). HhaI site methylation scores were approximated from phosphorimager quantitation of Southern blot results (+1 to +3, increasing hypermethylation relative to somatic controls; −1 to −3, increasing hypomethylation). The genomic sequencing data for each cancer clone were quantified as the weighed average of hypermethylation at the two normally unmethylated CpGs and hypomethylation at the seven normally methylated CpGs. There was a significant association between NBL2 methylation changes in the cancers determined by these two assays (P < 0.001). Therefore, both are monitoring similar methylation changes. In addition, a comparison of the overall m5CpG content of the DNA (by high-performance liquid chromatography analysis) and the total proportion of methylated sites in the NBL2 0.2-kb subregion indicated a significant association between these two epigenetic variables (P = 0.001) as well as between global DNA methylation and NBL2 HhaI site methylation. Therefore, NBL2 methylation changes in cancer are linked to global DNA methylation changes.

To enable us to analyze methylation changes at HpaII, AvaI, HpyCH4IV, or BstUI sites in NBL2 arrays from cancer samples, we first characterized methylation of normal somatic DNAs digested with these CpG methylation-sensitive enzymes and hybridized to a NBL2 probe. All somatic controls from various postnatal somatic tissues derived from different individuals gave very similar results with a given enzyme (Fig. 4; data not shown). However, there was much more resistance of NBL2 arrays in all these controls to cleavage by some of these enzymes than by others. These results could not be explained by the frequency of the restriction sites in NBL2. There is an average of ∼9 to 10 HpaII sites versus 5 to 6 HhaI, 2 to 4 AvaI, 3 to 5 HpyCH4IV, and 6 BstUI sites per NBL2 monomer (Genbank Y10752 and AC0128692). Nonetheless, NBL2 arrays in somatic controls were much more resistant to digestion by HpaII than by the other enzymes, and HpyCH4IV gave more cleavage than the other enzymes (Fig. 4). We disproved the possibility that the low extent of digestion of NBL2 arrays by HpaII in somatic control DNAs was due to sequence variation by showing complete digestion of all tested samples to <0.4-kb fragments by MspI, an isoschizomer of HpaII. MspI is resistant to CpG methylation, except at GGCCGG sites (28). In addition, internal controls for all digests showed that no inhibitors were present. The preferential methylation of HpaII sites in NBL2 from somatic controls observed in these Southern blot assays was consistent with genomic sequencing data (Fig. 2, CpG2, CpG5, and CpG11).

All 18 ovarian cancer DNAs and 13 of the 15 Wilms' tumor DNAs examined with at least three of the above enzymes exhibited altered Southern blot patterns of NBL2 methylation relative to somatic controls (Table 2; data not shown). We compared Southern blot data from cancer DNAs digested with different enzymes (Fig. 2) with the caveats that HpaII digests give an underestimate of hypermethylation and HpyCH4IV digests give an underestimate of hypomethylation. Importantly, HhaI sites seemed to undergo de novo methylation during carcinogenesis more frequently than AvaI, HpyCH4IV, and BstUI sites despite all of these enzymes giving mostly intermediate molecular weight NBL2-hybridizing bands in somatic controls (Fig. 4; Table 2; data not shown). This suggests some sequence specificity to cancer-linked hypermethylation. In addition, the distribution of NBL2-containing restriction fragments in HhaI digests and AvaI digests of OvCaD and OvCaE indicated that NBL2 arrays can be bifurcated into two epigenetic components differing in the extent of methylation at a given restriction site in several of the cancers (Fig. 4A and C, brackets). Long tandem regions of hypermethylation at these two kinds of restriction sites were observed as increases in NBL2 signal in >10-kb fragments, although those tumors also displayed increases in low molecular weight signal relative to the somatic controls. Individual fractions of NBL2 repeats with respect to long-range methylation patterns might correspond to NBL2 arrays on different acrocentric chromosomes.

Immunodeficiency, centromeric region instability, and facial anomalies (ICF) syndrome patients usually have missense DNA methyltransferase 3B (DNMT3B) mutations in both alleles (29), which greatly reduce enzymatic activity (30). To examine the involvement of DNMT3B in methylation of NBL2, we did Southern blot analysis of DNA digests from six ICF B-cell lines known to have DNMT3B mutations and 10 control B-cell lines. Relative to normal somatic tissues, hypomethylation at NBL2 HhaI sites was seen in four of the six ICF lymphoblastoid cell lines (LCL) but none of the 10 control LCLs. Instead, the control LCLs were hypermethylated in NBL2 arrays compared with normal somatic tissues, including leukocytes (Fig. 5A; data not shown). This indicates that NBL2 underwent de novo methylation at HhaI sites during generation or passage of LCLs only if the LCLs had normal DNMT3B activity.

All but one of the ICF LCLs displayed hypomethylation at NotI sites in NBL2 arrays, whereas none of the control LCLs did (data not shown). In addition, the ICF LCLs showed hypomethylation at HpaII sites compared with control LCLs and control somatic tissues (Fig. 5B). However, both control and ICF LCLs exhibited hypomethylation at AvaI and HpyCH4IV sites compared with control somatic tissues, although the extent of AvaI site hypomethylation was greater for the ICF LCLs (Fig. 5C and D). By hairpin genomic sequencing of NBL2, there was hypomethylation at some CpGs and hypermethylation at others in LCLs from ICF patients B and C and in a control LCL (Fig. 5E; data not shown). Nonetheless, there was more hypomethylation in the ICF cells and more hypermethylation in the control cells relative to normal somatic tissues (Table 2).

We showed previously that eight diverse somatic tissues and 16 of 20 cancers (ovarian carcinomas and Wilms' tumors) were negative by reverse transcription-PCR (RT-PCR) for NBL2 transcripts (12). The four positive cancers and one tested LCL (ICF B) evidenced transcripts by both real-time and semiquantitative RT-PCR but only at low levels. There was no relationship to hypomethylation at HhaI sites in NBL2, and NBL2 RNA was shown to probably result from run-through transcription. Here, we tested 5 more ICF LCLs and 10 control LCLs for NBL2 transcripts by real-time RT-PCR with GAPDH transcripts as the internal standard. Low levels of NBL2 RNA were seen in four of the ICF LCLs that displayed hypomethylation at HhaI sites (data not shown). Neither of the other two ICF LCLs and none of the control LCLs gave a signal appreciably above background, and also none of these displayed hypomethylation at HhaI sites. Duplicate cDNAs prepared from each LCL with random primers or oligo(dT) gave similar results in real-time RT-PCR. In addition, semiquantitative RT-PCR confirmed that the correct size product was obtained from an ICF LCL (ICF C) using either oligo(dT) or random priming, and no product was obtained from a control LCL (patient C). Product formation from ICF LCLs was shown to be dependent on reverse transcription. An unspecified promoter adjacent to one of the NBL2 arrays might be hypomethylated in the NBL2 RNA-positive ICF LCLs and cancers and thereby activated for run-through transcription.

The tandem 1.4-kb NBL2 repeat provided new insights into several aspects of epigenetics in normal tissues and cancers. In the 0.2-kb subregion of NBL2 from diverse control somatic tissues that was examined by hairpin genomic sequencing, there was a completely conserved pattern of undermethylation at two nonadjacent CpGs and full methylation at seven other CpGs (Fig. 3A). This methylation pattern was lost in all 146 DNA clones from 10 cancers (ovarian carcinomas and Wilms' tumors). Moreover, all but two of the cancer clones were hypomethylated at one specific CpG dyad or hypermethylated at another dyad only 14 bp away (CpG6 and CpG10). None of the normal DNA clones had this epigenetic signature. The two exceptional cancer clones lacked one of these CpGs because of sequence variation, but the methylation status of a third CpG (CpG14) allowed those two clones to be distinguished from all normal clones. Hypermethylation at CpG6 and/or CpG14 as well as hypomethylation at CpG2, CpG3, CpG5, CpG8, CpG10, CpG11, and/or CpG12 were seen in the majority of cancer DNA clones. Although hypermethylation of CpG6 or hypomethylation of CpG10 were the most diagnostic epigenetic changes for cancer, there was only one cancer DNA clone (in WT67; Fig. 2) that displayed both hypermethylation and hypomethylation at precisely these positions. This observation and the nonrandom nature of many of the other DNA methylation changes observed by hairpin genomic sequencing and Southern blot analyses indicate that losses and gains of methylation in NBL2 during carcinogenesis are often targeted to specific CpG positions and in specific patterns within the repeat.

The targeting of NBL2 for nonrandom hypermethylation and hypomethylation cannot be explained by transcription-related binding of sequence-specific DNA binding proteins as is the case for certain promoters (31). NBL2 underwent extensive cancer-linked alterations in methylation despite its lack of transcription in normal tissues and in most analyzed cancers and the absence of an in silico–predicted gene structure (12). Therefore, silencing of transcription is not necessary for all cancer-associated DNA hypermethylation, although it has been implicated in promoter hypermethylation (32). Moreover, an in silico search for consensus sites for sequence-specific DNA-binding proteins in NBL2 (http://www.cbil.upenn.edu/tess) did not yield putative sites that could explain the observed methylation patterns. Whether gene regions and other DNA repeats incur the same type of interspersed hypermethylation and hypomethylation, which might occur especially in the early stages of tumorigenesis, remains to be determined.

Cancer-linked demethylation of NBL2 was often observed in more than one of the seven normally methylated CpG positions with intervening CpGs that retained methylation. With respect to hemimethylation, cancer clones had a higher frequency of hemimethylated CpG sites than somatic control clones, and 15 clones had two hemimethylated sites with opposite strands unmethylated. These results indicate that demethylation by inhibition of maintenance methylation after DNA replication is not the major source of cancer-linked hypomethylation. Instead, they suggest some kind of active demethylation. However, there was a small, but statistically significant, percentage of cancer DNA clones containing runs of unmethylated CpGs on one or both strands. An overall decrease in fidelity of maintenance methylation could contribute to this cancer-associated demethylation, but the finding that most hypomethylation was intermittent along the examined cancer DNA molecules and the considerable sequence specificity observed in the cancer-linked CpG hypomethylation suggests more than just the previously reported inaccurate methylation maintenance (33). The mechanism for demethylation in cancer is uncertain. However, it is clear that mammals have the capacity for active demethylation as seen in the male pronucleus of the mouse zygote (34). Hemimethylated sites in vertebrate DNAs were described previously (16, 35). There is evidence for demethylation specifically of one strand in a transcription regulatory region followed by demethylation of the other during normal vertebrate development (36, 37). Our results suggest that hemimethylated sites are a rather stable intermediate in demethylation of DNA during carcinogenesis and that this demethylation of one strand of a CpG dyad occurs with preferences for certain CpGs and with strand preferences at some of these CpGs.

In de novo methylation of NBL2 in cancer, DNMT3B is likely to be the main enzyme involved as determined by our analysis of B-cell LCLs from controls and from ICF patients who have inactivating mutations in DNMT3B that eliminate most DNMT3B activity (30). The much lower levels of NBL2 methylation in ICF LCLs than in control LCLs indicate that DNMT3B is necessary for establishing the normal NBL2 methylation pattern during development. The hypermethylation of NBL2 at HhaI sites in control LCLs relative to somatic control tissues could be explained by overexpression of DNMT3B (as well as DNMT3A and DNMT1) during transformation with EBV (38) and/or by the oncogenic transformation-associated loss of fidelity of DNA methyltransferases (33). In vitro transformation of lymphocytes by EBV may provide a model for understanding NBL2 methylation changes during malignant transformation in vivo because both hypomethylation and hypermethylation relative to control somatic tissues were observed in NBL2 in normal LCLs. In the two ICF LCLs subject to genomic sequencing, despite the overall hypomethylation of NBL2, some hypermethylation was observed at CpG6, although not at CpG14, the other site at which we could analyze cancer-linked hypermethylation. In addition, the control LCL displayed more methylation at CpG6 than CpG14. Similarly, CpG6 was hypermethylated significantly more frequently than CpG14 in ovarian carcinomas. Moreover, CpG6 was occasionally hemimethylated in somatic controls, whereas CpG14 was always symmetrically unmethylated. These findings might be related to the dynamic system of normal maintenance methylation and de novo methylation proposed by Pfeifer et al. (39). At NBL2, there may be infrequent de novo methylation of CpG6 in one strand in normal cells, which is not followed by maintenance methylation. In contrast, there may be frequent hemimethylation at this site with subsequent maintenance methylation on oncogenic transformation.

An in vitro study of methylation by DNMT3B indicated strong sequence preferences for de novo methylation (40). Our findings, especially from Southern blot analysis, support this idea, although the specificities that we found do not match the in vitro ones. This enzyme may have its sequence preferences strongly altered in vivo. Both our genomic sequencing and Southern blot analysis indicate that HpaII sites (CCGG) have an especially high level of methylation in NBL2 in normal somatic tissues. In addition, our Southern blot analysis suggests that HhaI sites (GCGC), which were missing from the bisulfite-sequenced region, were more frequently de novo methylated in the cancers than HpyCH4IV (ACGT), AvaI (CYCGRG), and BstUI (CGCG) sites.

Evidence had been provided for cross-talk between demethylation and de novo methylation pathways in tumorigenesis (41) and in Arabidopsis containing an antisense DNA methyltransferase transgene (42). However, hypermethylation of 5′ regions of tumor suppressor genes and hypomethylation of LINE1 interspersed repeats, satellite DNA, and promoter regions of cancer-testes antigen genes (4, 6, 11, 43) have been shown to be statistically independent of each other, although all such changes are linked to cancer. Although hypermethylation at certain DNA sequences and hypomethylation at others in cancer are not associated with one another, this study shows that both cancer-linked hypomethylation and hypermethylation are somehow targeted to NBL2 repeats. We hypothesize that a chromatin structure change in NBL2 arrays occurs during oncogenesis that predisposes to both demethylation and de novo methylation in cis. Alternatively, NBL2 arrays, which have a high overall m5CpG content, might first be demethylated during tumorigenesis and the resulting chromatin structure change might favor further demethylation as well as de novo methylation.

Previous bisulfite-based genomic sequencing studies of cancer DNA usually involved unmethylated CpG-rich promoter regions that become hypermethylated mostly homogeneously (22, 23, 44). There may be several reasons for NBL2 displaying surprisingly complex, nonrandom patterns of altered methylation during carcinogenesis. It is apparently not a gene, and its methylation status probably confers no selective advantage to a developing tumor. This is unlike the situation with promoters of tumor suppressor genes whose almost complete methylation can benefit the growing tumor by repressing transcription or stabilizing this repression. In addition, unlike most DNA regions from cancers analyzed by genomic sequencing, NBL2 normally has very low levels of methylation at some CpGs and complete methylation at many others so that both cancer-linked increases and decreases of DNA methylation can be observed. Furthermore, it seems to be an unusually frequent target for multiple methylation changes during carcinogenesis. As such, it is a good candidate for a cancer marker as well as a source of insight into cancer-linked epigenetic alterations without the skewing of DNA methylation patterns by oncogenic selection pressures.

With institutional review board approval, primary tumor samples were obtained from surgery patients before chemotherapy or radiation therapy. Informed consent was given by all patients or unlinked samples were used. The LCLs were described previously (45-47) or available from the Coriell Institute (Camden, NJ; GM17900, AG14836, AG14953, and AG15022). Control somatic tissues were autopsy specimens of trauma victims (individuals A, B, and C, all males of 56, 19, and 68 years, respectively). DNA was purified as described previously (11).

Hairpin-bisulfite PCR was done basically as described by Laird et al. (16) using a NBL2 sequence (Y10752, Genbank) to design primers and the hairpin linker. Human DNA (0.5 μg) or NBL2-containing pDMHD-1 (50 ng; ref. 14) plus 450 ng λDNA carrier were digested with 10 units BsmAI and ligated to 5′-CCCTAGCGATGCGTTCGAGCATCGCT-3′. The DNA was denatured with 0.6 mol/L NaOH at 37°C for 15 minutes followed by incubation in boiling water for 1 minute. At hourly intervals during the 5-hour bisulfite treatment, the sample was incubated four times in boiling water for 1 minute. In an ultrafiltration device (Microcon-100, Millipore), bisulfite-modified DNA was washed thrice with water, desulfonated with 0.3 mol/L NaOH at 37°C for 15 minutes, and eluted in 50 μL of 10 mmol/L Tris-HCl, 1 mmol/L EDTA (pH 7.5). The primers for subsequent PCR had a 3′ T or A corresponding to deamination products from a non-CpG C residue or its complement. The primers were F2-1, 5′TTTTTGTGGGTTTGTGTTAGT-3′, and R2-2, 5′-CAAAAACATCTTTATTCCTCTA-3′. F2-1 was replaced by F2-2, 5′-AYGTGGTTTGGGTTAGGTAT-3′, in the second round of PCR. Only the F2-2 primer had a CpG in the analogous unmodified genomic sequence (at positions 2 and 3). After denaturation at 94°C for 15 minutes, PCR was done (Hotstar, Qiagen) for 30 cycles on 2 μL of the bisulfite-treated DNA (94°C for 15 seconds, 52°C for 15 seconds, 72°C for 1 minute, and a final extension at 72°C for 5 minutes). Then, 1 μL of the product was amplified analogously for an additional 35 cycles. Purified fragments obtained by electrophoresis in a 1.5% agarose gel were used for cloning (TA Cloning kit, Invitrogen), transformation (E. coli, Top10F), and sequencing (Translational Genomics Research Institute, Phoenix, AZ). Due to much sequence variation, there were three CpG positions that were much less frequently present in the clones than the 14 CpG positions referred to above. They are omitted for simplicity, but their methylation status did not change any of the conclusions of this study.

For Southern blot analysis, 1.5 μg human DNA was digested with 15 to 30 units restriction endonuclease overnight according to the manufacturer's procedures (New England Biolabs), all with parallel internal controls as described previously (12). At least three diverse somatic control tissues and sperm DNA were included as references in each blot.

Using random primers in one set and oligo(dT) in a duplicate set, cDNA was synthesized from 3 μg total RNA that had been treated with 3 units DNase I (amplification grade, Invitrogen) for 45 minutes at room temperature. Real-time PCR (SYBR Green PCR Master Mix, Applied Biosystems) was done with previously described primers and conditions (12). Semiquantitative RT-PCR with evaluation of the product by gel electrophoresis was also done as described previously (12).

Genomic sequencing methylation data were analyzed using R version 2.0.1 (http://www.rproject.org). χ² test statistics were used to assess differences of proportions, and strengths of association for continuous and ordinal variables were evaluated using the standard Pearson's correlation and Kendall's τ statistics, respectively. Where appropriate, Ps were adjusted for multiple tests using the Holm procedure. Classification trees were generated using the RPART library (48).

We thank Drs. Louis Dubeau, Murali Chintagumpala, Jeffrey Dome, Lisa Diller, Martin Champagne, and Hisaki Nagai for supplying the cancer samples and the NBL2 plasmid and Kathryn E. Donahoe for valuable editorial assistance.