Dissecting Complex Epigenetic Alterations in Breast Cancer Using CpG Island Microarrays¹

Cancer is generally thought to arise after multiple genetic mutations. More recently, a different type of alteration, i.e., CpG island hypermethylation, has been shown to be common in cancers (reviewed in Refs. 1 and 2). Unlike genetic mutations, this alteration is an epigenetic event that does not involve changes in nucleotide sequences. Nevertheless, accumulating evidence indicates that this heritable change has profound effects in tumorigenesis. When CpG island hypermethylation occurs at the 5′-end in the regulatory regions of genes (i.e., the promoter and the first exon), it may result in silencing of the corresponding genes (2). Both critical and noncritical CpG island loci could be affected, and only those providing loss of key biochemical functions would render selective advantages for tumor cells. Thus far, CpG island hypermethylation has been shown to be an important mechanism for the transcriptional inactivation of >90 tumor-related genes in many types of cancers.³

As an additional step toward a comprehensive understanding of this epigenetic event in cancer, it is necessary to develop new techniques for genome-wide methylation analysis. Several approaches are now available for profiling CpG island hypermethylation in human cancers (3, 4). Recently we adapted an array-based strategy and developed DMH⁴ (5) for high-throughput analysis of CpG island hypermethylation. Using a panel of CpG island tags arrayed on nylon membranes, DMH was successfully applied to detect specific methylation profiles in breast and ovarian tumors (6, 7).

To increase DMH throughput, we have made a complete transition from nylon membrane macroarrays to glass slide microarrays. A panel containing 7776 CpG islands was generated and used to analyze 17 paired tissues of breast tumors and normal controls. Close to 6.5% (496 loci) of these tags exhibited hypermethylation at least once in the tumors analyzed. Hierarchical clustering classified breast tumors into groups based on their methylation profiles that correlated with clinically related features.

Tumor tissues were obtained from patients undergoing mastectomy before chemotherapy at the Ellis Fischel Cancer Center, Columbia, MO, in compliance with our Institutional Review Board. Adjacent tumor-free parenchyma 3–5 cm away from the tumor area was obtained from each patient to serve as a paired control. Breast cancer cell lines T47-D, ZR-75, Hs578t, MCF-7, MDA-MB-231, and MDA-MB-468 were acquired from various resources as described (5) and routinely maintained in our laboratory. High-molecular-weight genomic DNA was isolated using the QIAamp Tissue Kit (Qiagen).

The resource material for preparing the microarray panel was derived from a CpG island library, CGI (8). CGI was prepared previously using male genomic DNA restricted with MseI, a four-base cutter known to restrict DNA into small fragments but to retain CpG island fragments largely intact (8). The GC-rich MseI fragments were isolated through an affinity column containing methyl-binding MeCP2 protein and cloned into vector for library construction. A total of 7776 CGI clones were individually organized in 96-well culture chambers as master plates. This included 10 preselected MseI-tags that act as positive controls because they are known to lack the test methylation-sensitive sites. A fraction (∼1 μl) of each clone was transferred to a well of separate 96-well PCR tubes using the MULTI-PRINT replicator (V&P Scientific). CpG island inserts (0.2–2 kb) from these clones were amplified by PCR as described (5). The primers immediately flanking the inserts are: HGMP 3558, 5′-CGG CCG CCT GCA GGT CTG ACC ATA A; and HGMP 3559, 5′-AAC GCG TTG GGA GCT CTC CCA TAA (8). To ensure the reproducibility of each PCR and to prevent cross-contamination among bacterial clones in microplates, amplified inserts were individually verified using a 96-well format gel electrophoresis system (Cascade Biologicals). Our choice of arrayer, namely the Affymetrix/GMS 417 Arrayer, permits the dotting of unpurified PCR products, because its ring-and-pin system is much less susceptible to clogging than the quill-type pen and ink jet-type printing head. Unpurified PCR products (∼0.02 μl/dot and 0.1 μg/μl), in the presence of 20% DMSO, were printed as microdots (150 μm in diameter, spaced at 300 μm) on poly-L-lysine-coated microscope slides as described by DeRisi et al.⁵ Spotted DNA was denatured before use.

The MseI digestion was performed in a 50-μl volume/sample containing ∼2 μg of genomic DNA. The digested fragments are expected to match the MseI-digested inserts originally used in the construction of the CGI genomic library. The digests were purified and their sticky ends ligated with 0.5 nmol of unphosphorylated linkers H-24/H-12, as described (5). The oligonucleotide sequences were as follows: H-24, 5′-AGG CAA CTG TGC TAT CCG AGG GAT; and H-12, 5′-TAA TCC CTC GGA. The ligated DNA was digested with methylation-sensitive endonucleases BstUI and HpaII (New England Biolabs). PCR reactions were performed using the digests as templates and subjected to 20 cycles of amplification. The amplified products were purified, a portion of the products was reserved for Southern analysis, and the rest was used for fluorescence labeling.

Incorporation of amino-allyl dUTP into amplicons (5 μg) was conducted using the BioPrime DNA labeling system (Life Technologies, Inc.). Cy5 and Cy3 fluorescent dyes were coupled to amino-allyl dUTP-labeled tumor and normal amplicons, respectively, and cohybridized to the microarray panel. Hybridization and the posthybridization washing protocols were according to DeRisi et al.⁵ Hybridized slides were scanned with the GenePix 4000A scanner (Axon) and the acquired images were analyzed with the software GenePix Pro3.0. Because Cy5 and Cy3 labeling efficiencies varied among samples, we determined a global normalization factor for each microarray image. The effectiveness of the normalization factor was evaluated using 10 internal positive controls in which their adjusted Cy5:Cy3 ratios were expected to be 1. The adjusted ratio for each CpG island locus was calculated, and the data were analyzed using three statistical algorithms. A hierarchical clustering algorithm⁶ was used to investigate relationships among tumor samples. The complete linkage and the dissimilarity measure (1 minus the Pearson correlation coefficient of the log-adjusted Cy5:Cy3 ratios) were used for the analysis. The resultant dendrogram linked related breast tumors into a phylogenetic tree whose branch lengths represented the degree of similarity between these tumors. A nonhierarchical clustering algorithm, the Fuzzy C-Means protocol (Partek Pro 2000; Partek), was used to analyze the same data set to determine whether similar patterns could be independently identified. Another statistical protocol, multidimensional scaling (a component of Partek Pro 2000), was also used to display the overall similarity in DNA methylation profiles. Using a matrix of Pearson correlation coefficients from the complete pairwise comparison of all of the breast tumors, the multidimensional scaling procedure positioned tumor samples in a three-dimensional space so that the distance between each pair of samples would closely approximate the Pearson correlation coefficient in the matrix for the corresponding sample pair. This three-dimensional approximation of multidimensional relationships produces a visually intuitive pattern of sample relatedness.

One μg of PCR products (or amplicons) or 10 μg of genomic DNA digested with MseI and/or BstUI were separated on 1.0% agarose gels and transferred to nylon membranes. Short fragments (200–250 bp) from selected CpG island clone inserts were PCR-amplified and used as probes for Southern hybridization as described (5).

Total cellular RNA was isolated from cells using the RNAeasy Total RNA System (Qiagen). Ten μg of RNA were electrophoresed on a 1.5% agarose gel and subjected to northern analysis using a GPC3 cDNA probe. RT-PCR was conducted using primers 5′-ATC CTG TAT ACC TCC TCC AG and 5′-ATC CAT GCA AAG AGA GAA CG for GPC3 cDNA. The levels of GPC3 mRNA were normalized with the level of β-actin mRNA as described in an earlier study (5).

Previously we developed the DMH technique for high-throughput screening of CpG island hypermethylation in cancer (5, 6, 7). To increase its throughput, we reconfigured DMH into a microarray-based system containing 7776 short DNA tags prepared from a genomic library, CGI (8). Although the estimated 20,000–30,000 MseI-restricted CGI clones may not represent the entire repertoire of CpG islands in the genome, this library has provided a valuable resource for identifying aberrantly methylated loci in breast and ovarian cancer (6, 7). Approximately 25% of the 7776 CpG island tags have thus far been sequenced, and their sequence information has been deposited in GenBank. On the basis of this sequence information, ∼10–15% of our CpG island panel contains highly repetitive elements (Alu repeats and long interspersed nuclear elements) and other repetitive elements (α-satellite and ribosomal and mitochondrial DNA). About 0.5–1.0% are associated with the X-chromosome regions.

For DMH screening, targets were prepared from a group of 17 paired breast tumors and normal samples. Unlike cDNA microarrays that use targets directly prepared from two different mRNA populations, targets generated for DMH involve a series of steps. One of the key steps requires the methylation-sensitive restriction of MseI-digested, linker-ligated DNA. In contrast to the previous use of a single endonuclease (5, 6, 7), in the present study DNA was digested sequentially with two methylation-sensitive endonucleases, HpaII and BstUI. This new approach enhances detection of CpG loci with extensive methylation in tumors, a situation more likely to result in chromatin condensation and the subsequent silencing of corresponding genes (9). In addition, the approach reduces the risk of incomplete digestion when only one endonuclease is used for probing methylation differences. Low amplification cycles (20 cycles) were used for linker-PCR. This step prevented overamplification of unrestricted repetitive sequences in the ligated DNA and yet yielded sufficient PCR products for single or low-copy number CpG island loci. Genomic fragments containing aberrantly methylated sites were protected from the digestion and could be amplified by linker-PCR in the tumor sample, whereas the same fragments containing the unmethylated sites were cut and could not be amplified in the normal sample. The amplified products (or amplicons) therefore contained different pools of DNA fragments because of the differential methylation status of tumor relative to the control sample.

Fig. 1,A shows representative DMH microarrays cohybridized with fluorescently labeled tumor and normal amplicons. CpG island tags whose signal intensities were slightly above the background or were devoid of hybridization signals represent the unmethylated loci in both tumor and normal samples; their genomic fragments were restricted away by the methylation-sensitive endonucleases before linker-PCR. Yellow spots (Cy5:Cy3 = 1) represent equal amounts of bound DNA from each amplicon, indicating no methylation differences between tumor and normal genomes. In some instances, these yellow spots represented CpG island tags that do not contain the internal HpaII or BstUI recognition sites and have equal copy numbers in both tumor and control DNA. CpG island tags hybridized predominately with the tumor amplicon, but not with the normal amplicon, appear as red spots. We set an arbitrary cutoff of 1.5 for the Cy5:Cy3 ratio, i.e., loci with ratios ≥1.5 were identified as hypermethylated in tumors. It should be noted that the magnitude of the Cy5:Cy3 ratio does not necessarily reflect the extent of hypermethylation, because the target preparation is PCR-based. An example is shown in Fig. 1, B—D: the SC76F1 locus seemed to be hypermethylated in the breast tumor of patient 241, but not in patients 151 and 109. Although less frequently, we also encountered green spots (Cy5:Cy3 ≤ 0.5) by DMH, denoting the presence of hypomethylated sequences in the tumor genome (see Fig. 1 C). Sequence analysis indicated that most of the green spots in the microarray panel are repetitive elements, which are often methylated in nontumor cells (10). Hypomethylation is observed in heterochromatin or α-satellite DNA in cancer cells, but has not been seen commonly in single-copy CpG islands (10).

We conducted a confirmation study to determine whether the cutoff ratio (≥1.5) could accurately identify hypermethylated loci. This was performed by Southern analysis using amplicons, originally prepared for DMH, as hybridization templates (Fig. 2). For example, the SC76F1 probe detected a DNA fragment (770 bp) in the tumor amplicons 65, 119, 167, and 241 but not in their paired normal amplicons. The presence of this amplified fragment in tumor amplicons was attributable to methylated CpG sites within the SC76F1 locus that were insensitive to restriction by endonucleases. This 770-bp fragment was not detected in tumor amplicons 83 and 151 or in their controls, indicating this locus was not hypermethylated in these tumors. In some cases (e.g., CpG5H12, SC13B2, and CpG15E5), normal breast samples had detectable methylation, but methylation of the loci was more extensive in their tumors. It is possible that this low level of preexisting methylation is present in normal samples (5, 6). The results of seven positive loci based on this Southern approach as well as their Cy5:Cy3 ratios derived from DMH are summarized in Table 1. The Southern findings seemed to be consistent with DMH results in the majority (95%) of the loci analyzed. The inconsistency between these methods is, in part, attributable to the imprecise nature of an overall Cy5:Cy3 normalization factor for a given DMH image. As a result, the cutoff ratio set for the identification of DNA hypermethylation may be too high to exclude some positive loci and too low to include some unmethylated loci in a given tumor. One remedy to this problem is to perform sufficient Southern amplicon analyses and then to readjust the normalization factor to clearly define the hypermethylation boundary for most of the CpG island loci. As with other microarray approaches, we expect low levels of false-positive and -negative findings in DMH. This should have only a minimal impact on the subsequent data analysis because of a large number of hypermethylated loci identified by DMH.

Of the 7776 CpG island loci probed by our DMH amplicons, loci having a normalized Cy5:Cy3 ratio ≥1.5 were selected for cluster analyses. Across the 17 tumors studied, the average number of hypermethylated loci was 83 (ranging from 15 to 207), consistent with the notion that hypermethylation is infrequent and occurs in ∼1% of CpG islands in the breast tumor genome (3). Because tumors exhibited hypermethylation in some shared loci while having other uniquely methylated CpG islands, the total number of loci identified in this patient group was 496. The Stanford hierarchical algorithm (11) was then used to extract maximum information from these positive loci for tumor classification (Fig. 3,A). Loci with little or no methylation differences between tumor and normal DNA were left blank, whereas hypermethylated loci in tumors were colored red. The resulting dendrogram divided 15 of these 17 tumors into three clusters, which also were confirmed using a separate clustering algorithm, the multidimensional 3-D scatter plot (Fig. 3,B). Sections of the dendrogram that contribute significantly to segregating these clusters are shown in Fig. 3 C. Two of these clusters were made up of tumors largely with negative ER/PR status, a group commonly less responsive to certain forms of chemotherapy; CpG island loci in block 2 were frequently methylated in these two patient groups. Block 1 represented loci predominately methylated in one of these two groups (patients 241, 31, 243, 91, 159, and 155), whereas loci in block 3 were methylated more frequently in the other group (patients 45, 43, 65, 167, and 119). The remaining cluster, comprised of ER/PR-positive tumors (patients 157, 151, 129, and 109), was largely devoid of methylation in these CpG island loci.

This methylation profile analysis has led to the identification of CpG island clusters that can distinguish ER/PR-negative from ER/PR-positive tumors. To determine whether a single CpG island locus within these clusters contributes to this discrimination, we examined SC76F1 in greater detail. The SC76F1 sequence matched a CpG island region located within the promoter of GPC3 gene at chromosome Xq26 (12). This gene encodes a glypican integral membrane protein and is mutated in Simpson-Golabi-Behmel syndrome, an X-linked condition involving multiple embryonic neoplasms and pre- and postnatal overgrowth (13). Interestingly, hypermethylation of this gene has recently been reported in several ovarian cancer cell lines (14). We determined the methylation status of GPC3 promoter in 45 breast tumors and 10 normal breast tissue samples by Southern analysis (Fig. 4,A). Both methylated (490 bp) and unmethylated (130 and 360 bp) alleles of GPC3 were observed in the normal breast samples analyzed, consistent with an X-linked status of this gene (12). Various degrees of hypermethylation (570, 640, and 770 bp) were detected in 38 of 45 (84%) breast tumors, whereas the remaining 7 tumors showed no detectable methylation changes. This result also confirms the DMH microarray findings. The exception is two tumors (patients 151 and 129) in which a diminutive level of methylation was detected by Southern analysis but not by DMH. Statistical analysis (t test) reveals a significant difference (P = 0.005) in GPC3 promoter hypermethylation between the hormone receptor-positive tumors and the hormone receptor-negative tumors. We further segregated these tumors according to the levels of GPC3 hypermethylation (Fig. 4,B). It is apparent that high levels of GPC3 methylation are predominantly associated with the hormone receptor-negative group, consistent with the earlier finding based on the cluster analysis. Hypermethylation of the GPC3 promoter resulted in loss of its expression in vitro, suggesting a functional consequence of this methylation-silenced gene in breast tumor growth (Fig. 4, C and D).

The present study is the first to provide such a detailed analysis of the complex methylation alterations in the human breast tumor genome. This type of microarray-based analysis is expected to move this field into the epigenomics era. It is now clear that aberrant DNA methylation has a causative role, rather than a bystander effect, in the development of cancer (1, 2). Many promoter CpG island sequences are susceptible to this epigenetic alteration. As a result, the expression of genes that govern key functions of the cell may become silent, leading to clonal proliferation of tumor cells. Differential susceptibility of critical CpG island loci to DNA hypermethylation may therefore influence the development of different tumor types and produces unique molecular signatures that are associated with clinicopathological characteristics of cancer patients. Dissecting these complex epigenetic profiles can also lead to the identification of tumor-suppressors that are silenced via DNA hypermethylation in particular tumor subtypes. Thus, this analysis has diagnostic potential and provides patient-specific methylation profiles that may predict those responsive to demethylation treatment (15). Gene expression profiles produced by cDNA microarrays have already been used to differentiate different types of lymphoma (16), leukemia (17), melanoma (18), and breast cancer (19). In this regard, DMH not only provides additional information regarding potential mechanisms underlying gene expression, but also offers an alternative to cDNA microarrays for molecular classification of tumors. The cDNA microarray approach requires targets derived from mRNA, which is more labile and difficult to isolate directly from clinical specimens. By contrast, our CpG island microarray uses DNA targets, which are more stable and easier to isolate from patients’ specimens.

In summary, we have presented data from our newly developed DMH microarray method that illustrate its utility both in the basic understanding of CpG island hypermethylation in cancer and also in a translational role as a potential method for molecular classification of tumors and prediction of responsiveness to demethylating agents. Studies such as this may open up a whole new era of individually designed pharmacological agents for each patient based on solid molecular evidence of efficacy.

We thank Diane Peckham, Jeffrey Lin, and Adam Asare for their assistance in the preparation of this manuscript and Dr. Sally H. Cross and her colleagues from the United Kingdom Genome Mapping Project Center for providing the CGI library.