Discovery of Novel Methylation Biomarkers in Cervical Carcinoma by Global Demethylation and Microarray Analysis

A genome-wide screening study for identification of hypermethylated genes in invasive cervical cancer (ICC) was carried out to augment our previously discovered panel of three genes found to be useful for detection of ICC and its precursor neoplasia. Putatively hypermethylated and silenced genes were reactivated in four ICC cell lines by treatment with 5-aza-2′-deoxycytidine and trichostatin A and identified on expression microarrays. Thirty-nine of the 235 genes up-regulated in multiple ICC cell lines were further examined to determine the methylation status of associated CpG islands. The diagnostic use of 23 genes that were aberrantly methylated in multiple ICC cell lines were then analyzed in DNA from exfoliated cells obtained from patients with or without ICC. We show, for the first time, that aberrant methylation of six genes (SPARC, TFPI2, RRAD, SFRP1, MT1G, and NMES1) is present in a high proportion of ICC clinical samples but not in normal samples. Of these genes, SPARC and TFPI2 showed the highest frequency of aberrant methylation in ICC specimens (86.4% for either) and together were hypermethylated in all but one ICC cases examined. We conclude that expression profiling of epigenetically reactivated genes followed by methylation analysis in clinical samples is a powerful tool for comprehensive identification of methylation markers. Several novel genes identified in our study may be clinically useful for detection or stratification of ICC and/or of its precursor lesions and provide a basis for better understanding of mechanisms involved in development of ICC. (Cancer Epidemiol Biomarkers Prev 2006;(15)1:114–23)

The control of ICC is based upon detection and treatment of the cervical intraepithelial neoplasia stage III (CIN-3) precursor lesions and carcinoma in situ (CIS). In the United States and other developed countries, the incidence of ICC has dramatically decreased as a result of the identification and ablation of the precursor CIN-3/CIS lesions achieved via regular Papanicolaou screening with referral of all women with abnormal Papanicolaou smears (atypical squamous cells of undetermined significance and greater) to repeat cytology or colposcopy and biopsy. This approach is costly and extremely burdensome to the health care system but deemed necessary in light of the low reproducibility of cytology and the low sensitivity of a single cytologic smear (range, 30-60%) to detect CIN-3 (1). More recently, testing for oncogenic human papillomavirus (HPV) types has been introduced to help triage women with atypical squamous cells of undetermined significance. Nevertheless, current recommendations (2) based on a screening approach, which include HPV-based triage of atypical squamous cells of undetermined significance, will send over three million women each year for colposcopy and biopsy, although most are at little risk of ICC. Furthermore, the cost and infrastructure requirements of primary screening based on cytology (with or without HPV testing) make this approach impractical in resource poor settings. Lastly, our studies showed that vaccines that prevent HPV infection will eventually be central to cervical cancer control (3); however, because such vaccines will not offer protection to most of the millions of women already infected with oncogenic HPVs, screening will be necessary for the foreseeable future. Thus, development of more accurate molecular screening methods that could be automated would be of significant interest.

Epigenetic silencing of a variety of genes by hypermethylation of promoter-associated CpG islands is now recognized as a frequent and early event in the pathogenesis of many cancers, including ICC, making such changes of potential interest as biomarkers for identification of individuals with or at risk for cancer (4). Currently, relatively little is known about specific patterns of CpG island hypermethylation in ICC. Most previous studies examining DNA hypermethylation in ICC have generally focused on testing of genes known to be aberrantly methylated in other cancers (the candidate gene approach; refs. 5-7). In the largest such study thus far, we studied 21 candidate hypermethylated genes and identified a panel of three (DAPK1, RARB, and TWIST1), which were sensitive and specific for ICC and its immediate precursor CIS. We estimated that detection of hypermethylation of these three genes in exfoliated cell samples would identify histologically confirmed CIN-3 or worse with a sensitivity of 60% and a specificity of 95% in unscreened women over age 35, presenting to community based clinics for reasons other than cervical neoplasia (8). This study also showed that exfoliated cell samples could be used to identify women who had acquired hypermethylated changes in their cervical epithelial cells that are highly associated with CIS/ICC (8). However, whereas the performance of this panel of hypermethylated genes compared favorably with the performance of exfoliated cervical cytology and detection of high-risk types of HPV DNA, hypermethylated genes with increased sensitivity and specificity for ICC and precursor neoplasia will need to be identified if detection of DNA hypermethylation in exfoliated cell samples is to have clinical utility for cervical cancer screening. The present study was undertaken to find such genes using a genome-wide examination of the methylation pattern of cervical cancer cell lines.

A number of different approaches to genome-wide identification of cancer-associated hypermethylated genes have been developed, including restriction landmark genome scanning, methylation-sensitive arbitrarily primed PCR, and methylated CpG island amplification (4). An alternative approach is to treat cells with epigenetic modifying drugs to allow expression of hypermethylated/silenced genes and to then compare the expression profile by microarray analysis of untreated and drug-treated cells to identify putatively hypermethylated genes. The methylation status of such genes can then be confirmed in untreated cells using methylation-specific PCR analyses. Global demethylation and expression microarray analysis has been successfully used to identify a number of novel genes hypermethylated in association with colorectal and pancreatic cancer (9, 10). In the present study, we treated four ICC cell lines (C4I, CaSki, HeLa, and SiHa) with 5-aza-2′-deoxycytidine (DAC), a DNA methyltransferase inhibitor, and trichostain A, a histone deacetylase inhibitor and compared the pretreatment and post-treatment expression profiles to identify putatively hypermethylated genes. Genes with low initial expression that were strongly up-regulated after treatment in the cancer cell lines were considered of potential interest as ICC-associated hypermethylated genes. For such potentially hypermethylated genes, methylation-specific PCR analyses were carried out to examine the presence of aberrantly methylated CpG dinucleotides in cell lines. Twenty-three genes, thus identified as hypermethylated in cell lines, were assessed for the diagnostic use in clinical samples. Using this approach, we identified six genes whose methylation has not been previously described in ICC but which seem to be of a great interest as methylation biomarkers for detection of this disease.

HeLa, SiHa, CaSki, and C4I were obtained from the American Type Culture Collection (Manassas, VA) and grown in DMEM (HeLa, SiHa, and C4I) or RPMI (CaSki) supplemented with penicillin (0.5 units/mL), streptomycin (0.5 μg/mL), Glutamax (2 mmol/L), and 10% (v/v) fetal bovine serum. All media and supplements were from Invitrogen (Carlsbad, CA). Cervical exfoliated cell samples were obtained as previously described (11) from women with and without cervical neoplasia in a study in Dakar, Senegal, approved by the University of Washington Internal Review Board. Median age and high-risk and low-risk HPV prevalence in patient population is listed in Table 1. DNA from 43 exfoliated cell samples was isolated using the QiaAmp DNA Mini kit (Qiagen, Valencia, CA).

The four cervical carcinoma cell lines listed above were seeded in 10-cm culture dishes at a low density and treated with 200 nmol/L DAC for 3 days and with 300 nmol/L trichostain A for the last 24 hours. Medium containing fresh drugs was replaced daily. Controls consisted of cells treated identically with drug solvents (PBS and 100% ethanol), and all experiments were done in duplicate or triplicate (see below). After treatment, cells were removed from dishes using trypsin/EDTA solution (Invitrogen) and rinsed with media containing 10% fetal bovine serum. Total RNA from mock-treated and DAC/trichostain A–treated HeLa cells was isolated using the RNEasy Mini kit (Qiagen), and 30 μg were labeled by incorporating amino-allyl dUTP in cDNA using reverse transcriptase (Invitrogen) according to the manufacturer's instructions. For C4I, CaSki, and SiHa cells, total RNA was amplified and labeled with amino-allyl dUTP using the Amino Allyl MessageAmp^TM aRNA amplification kit (Ambion, Austin, TX). Genomic DNA from untreated cells lines was isolated using the QiaAmp DNA Mini kit (Qiagen).

Human cDNA microarrays consisting of 17,568 sequence-verified clones from the Research Genetic IMAGE (Integrated Molecular Analysis of Genomes and their Expression) clone set, available through Invitrogen (http://clones.invitrogen.com) were printed by the Fred Hutchinson Cancer Research Center's DNA microarray facility. Labeled cDNA or aRNA were coupled from both directions with Cy3 and Cy5 dyes according to the manufacturer's instructions (Amersham Biosciences, Piscataway, NJ). Replicate hybridizations to microarrays were done with three independent RNA isolates for C4I, HeLa, and SiHa and two independent RNA isolates for CaSki; each isolate was hybridized twice with a dye swap. Microarray slides were scanned at wavelengths of 532 and 635 nm using a dual-laser GenePix4000A microarray scanner (Axon Instruments, Union City, CA), and fluorescence data were extracted using the GenePix Pro 3.0 program.

Raw hybridization signals were normalized using the locfit (LOWESS) method in MIDAS program (12). The two numbers for each spot, representing its normalized median fluorescence minus the median background fluorescence in both scanning wavelengths, were used to calculate the log₂-transformed ratios of normalized intensities. The ratios reflected the expression changes of individual transcripts upon DAC/trichostain A treatment. The basal expression of genes was estimated from fluorescence signals of hybridized cDNA or aRNA targets generated from mock-treated cells; the signal values were log₂ transformed and mean centered. For hierarchical clustering analysis, expression ratios from 11 experiments in cervical carcinoma cell lines were combined and filtered to include only genes that were up-regulated at least 2-fold in five or more individual experiments (i.e., up-regulated in at least two cell lines; 235 genes). To include basal expression pattern in our analysis, expression ratios of 235 up-regulated genes were combined with basal expression values and clustered such that basal expression values were given twice the weight of ratio values. Clustering and visualization were done in Cluster and TreeView (13). To ascertain the reproducibility of array data obtained in repeated experiments, we calculated the SD of expression ratio and coefficient of variance. Data summarizing expression ratios have been deposited in GEO database (http://www.ncbi.nlm.nih.gov/projects/geo/) in series GSE2097, in accordance with MIAME standards (14). A probability associated with methylation pattern of candidate genes in ICC-positive and ICC-negative exfoliated cervical cell clinical samples was calculated using Fisher's exact test.

Five micrograms of total RNA were reverse-transcribed using Superscript II (Invitrogen). Each cDNA sample was additionally prepared in the absence of enzyme and used as a negative control to estimate the amount of contaminating DNA. From a resulting 20-μL reaction mix, 0.17 μL was used as a template in triplicate PCR reactions in a Prism HT7900 Sequence Detection System (ABI, Foster City, CA) under the following conditions: (a) 95°C for 10 minutes; (b) 40 cycles of 95°C for 20 seconds, 60°C for 20 seconds, and 72°C for 45 seconds; (c) final extension for 2 minutes at 72°C; and (d) denaturation/renaturation step for melting curve analysis. The TMP21 gene that has previously been found to be expressed in most tissues (15) was used as an external standard for each plate. The primers for genes examined in our study are listed in Supplementary Table S01. The change in gene expression was calculated using the 2^-ΔΔCt method, in which the ΔC_t value was calculated by subtracting the C_t value of TMP21 from a gene-specific C_t value. The ΔΔC_t value was calculated by subtracting the ΔC_t value, obtained from mock-treated cells, from the ΔC_t value obtained from drug-treated cells. Presence of single amplification products in reactions was ascertained by melting curve analysis and gel electrophoresis.

Genes selected for methylation analysis were evaluated using methylation specific PCR (MSP), methylation-specific melting curve analysis (MS-MCA), and MethyLight assay (16-18). CpG island sequences were obtained from the University of California Santa Cruz genome browser (http://genome.ucsc.edu/) or from genomic information using CpGPlot (http://www.ebi.ac.uk/emboss/cpgplot/). MSP and MS-MCA primers were designed using MethPrimer (19). MethyLight primers and MGB probes were designed using Primer Express (ABI). Sequences of primers and probes are listed in Supplementary Table S01. For methylation analysis, DNA from cultured cells or exfoliated cervical epithelial cells isolated using the QiaAmp DNA Mini kit (Qiagen) was first converted by sodium bisulfite method (20). Briefly, a saturated solution of sodium bisulfite containing 100 mmol/L hydroquinone was added to denatured DNA, and the resulting mixture was incubated at 55°C for 4 hours. DNA was subsequently recovered using the Qiaex II Gel Extraction kit (Qiagen), air-dried, eluted, and desulfonated, and finally ethanol precipitated in the presence of 1.93 mol/L NH₄OAc and coprecipitant (Pellet Paint, Novagen, San Diego, CA) and dissolved in 100 μL TE [2 mmol/L Tris-HCl (pH 8), 0.3 mmol/L EDTA]. For MSP or MS-MCA, 1 μL of this solution was used as a template for PCR in a Prism HT7900 instrument using 1 unit AmpliTaq Gold, in a mix containing 1.5 mmol/L MgCl₂, 62.5 μmol/L each deoxynucleotide triphosphate, 150 nmol/L each primer, and 0.134 μL of 1,000× diluted SYBR Green I (Molecular Probes, Eugene, OR), with PCR cycling conditions as above. A DNA melting curve was acquired by measuring fluorescence of SYBR Green I during a linear temperature transition from 60°C to 95°C at 0.2°C/s, and fluorescence data were converted into melting peaks by the HT7900 instrument SDS 2.0 software. In MS-MCA, the extent of methylation (T_M index) in cell lines was calculated based on the increase of melting temperature using the formula (T_M,M − T_M,SAMPLE) / (T_M,M − T_M,U) × 100, where T_M denotes the melting temperature of the amplicon obtained from the melting peak plot, and U and M denote the fully methylated or fully unmethylated templates, respectively. For MethyLight, 0.5 μL of bisulfite-converted DNA solution was combined with 600 nmol/L each primer and 100 nmol/L probe in ABI Universal PCR Master Mix and amplified under the default cycling conditions for Prism HT7900 instrument. Human sperm DNA and human sperm DNA methylated in vitro using the SssI (CpG) methylase (New England Biolabs, Beverly, MA), respectively, were used as U (unmethylated) and M (fully methylated) control DNA. Size of PCR products in representative samples was ascertained on agarose gels. Copy number of specific genes in samples was enumerated using the standard curve generated from known amount of bisulfite-converted β-actin sequence. The “percent of methylated reference” (PMR) values were calculated using a formula [N_gene / N_actb(sample)] / [N_gene / N_actb(M)] × 100%, where N denotes copy number for a gene or β-actin amplicon in a clinical sample or in fully SssI-methylated DNA (M). A cutoff PMR value, above which samples were considered methylated and below which samples were considered unmethylated, was determined to discriminate best between histologically normal and ICC clinical samples. The cutoff value was calculated as a median PMR value of histologically normal samples (if >0) plus 1 percentage point. If none or only one PMR value among negative samples was >0, cutoff was automatically adjusted to 1.

To identify hypermethylated genes associated with cervical cancer, we first treated four cervical carcinoma cell lines (C4I, CaSki, HeLa, and SiHa) with DAC and trichostain A. The treatment conditions were chosen such that they would induce low toxicity but cause widespread reactivation of genes silenced by aberrant methylation of their CpG islands (21). A cDNA expression microarray hybridization comparing the mock and drug treatment expression patterns in these cultures was then done to identify reactivated genes (summarized in Fig. 1). The number of genes differentially expressed (>2-fold) in specific cell lines varied from high 1,136 genes in SiHa to low 126 genes in C4I. As expected in tumor-derived cell lines containing many hypermethylated and silenced genes (22), the number of up-regulated genes was higher than the number of down-regulated genes in all cell lines except in CaSki (Fig. 1A). Overall, 1,068 genes were up-regulated ≥2-fold in at least one cervical carcinoma cell line after treatment (Fig. 1B). The specific sets of up-regulated genes varied in individual cell lines, and only 15 genes were up-regulated in all cell lines simultaneously (Fig. 1B). We reasoned that genes hypermethylated in primary ICC tumors would be hypermethylated and reactivated by drug treatment in multiple cervical carcinoma cell lines. To select such genes, we carried out a hierarchical clustering analysis of array data obtained from mock-treated and drug-treated cells that essentially selected genes that were up-regulated ≥2-fold simultaneously in two or more cell lines (data not shown). This analysis identified a cluster of 235 genes (data are summarized in Supplementary Table S02). Within this cluster, we were especially interested in genes with low expression in mock-treated cells that separated into a subcluster of 125 genes. Sixty-three such genes were selected for further analysis (Fig. 1C). Among genes that decreased their transcriptional activity after treatment, 134 were down-regulated in multiple cell lines (Supplementary Table S02). Down-regulated genes were not the focus of this study (see Discussion).

To confirm expression data obtained from array experiments, we did quantitative real-time reverse transcription-PCR (RT-PCR) for 46 of the 63 genes of interest, using RNA isolated from mock-treated and DAC/trichostain A–treated cell lines. Quantitative RT-PCR allowed us to estimate the pretreatment expression of these genes and the post-treatment ratio reflecting their up-regulation or down-regulation independently of array results (Table 2). Both the pretreatment (mock) expression levels and expression changes induced by treatment were estimated using the reference housekeeping gene TMP21 (Table 2) that is constitutively expressed in most tissues (15). In our experiments using cervical carcinoma cell lines, its expression was determined to be slightly above the median level of all genes present on the array, and it did not change after DAC and trichostain A treatment (data not shown). The correlation between ratios obtained from quantitative RT-PCR and array for 42 genes in each of the four cell lines is shown in Fig. 2. For all but a few genes in each cell line, the direction of expression change was the same in both assays, as indicated by the clustering of data points in the top right quadrant of the plots. In addition, higher ratios were consistently obtained by quantitative RT-PCR (slopes > 1). Based on these results, we were confident that the identification of reexpressed genes from array data was valid, and that a 2-fold cutoff for array ratios was sufficient for identifying differentially regulated and potentially methylated genes.

The methylation of CpG islands associated with the majority of candidate genes that we selected for further analysis has not been analyzed previously. Therefore, we employed a method shown earlier to reliably score methylated CpG content in target DNA sequences. The assay, MS-MCA, relies upon determination of melting temperature of a PCR product derived from bisulfite-converted DNA amplified from methylation-insensitive primers (16). Methylated CpG dinucleotides are in this assay amplified as CpGs, whereas unmethylated CpGs are amplified as TG dinucleotides, consequently lowering the melting temperature in a typical amplicon by several degrees Celsius (16). We first confirmed in DNA isolated from HeLa, SiHa, and CaSki that the amplicon-specific melting temperature shift (T_M index, see Materials and Methods) corresponded to CpG methylation data obtained by bisulfite sequencing of target areas and found an excellent match. An example is shown in Fig. 3 for SFRP1-associated amplicon. The 275-bp amplicon located within the CpG island associated with this gene contains 24 CpG dinucleotides, 61.5% of which were methylated in HeLa on average and 93.8% and 97.2% in SiHa and CaSki, respectively (Fig. 3A; sequence data for C4I was not obtained). The position of cell line–specific melting peaks relative to U and M DNA (Fig. 3B) matched well with the relative proportion of methylated CpGs as the T_M index, indicating this density was 75 in HeLa and 95 in CaSki and SiHa, respectively. For a number of genes, a classic MSP assay (17) was done and found to largely confirm data obtained by MS-MCA (data not shown). Subsequently, 39 genes (21 of the 46 genes whose up-regulation was confirmed by quantitative RT-PCR, and an additional 18 genes selected based on the array data but not examined by quantitative RT-PCR) were examined by MS-MCA to determine the presence and extent of CpG dinucleotide methylation in their promoter-associated CpG islands (refer to Fig. 1C). These analyses showed that 33 of 39 genes were extensively methylated (T_M index > 50) in at least one cell line, and 17 of these genes showed extensive methylation in three or four cell lines (Table 2). Specifically, 26 genes were extensively methylated in CaSki, 19 in SiHa, 18 in C4I, and 15 in HeLa cells (Table 2).

We next sought to determine whether or not the hypermethylated genes identified in the preceding experiments in cell lines were also hypermethylated in primary cervical cancers. This was examined in exfoliated cervical cell samples obtained from patients with and without ICC, as confirmed by a same-day biopsy and histology examination. We first used MS-MCA and MSP to estimate the prevalence of methylation in CpG islands associated with 23 genes identified as hypermethylated in cervical cancer cell lines (Table 2). Of these genes, nine (DKK3, HPCAL4, MT1G, NMES1, RRAD, SFRP1, SPARC, TFPI2, TPPP) were more often methylated in ICC-positive clinical samples (n = 22) than in ICC-negative (normal) samples (n = 21; Table 2), whereas two genes (FLJ20152 and GSN) were not methylated in any clinical sample, and one gene (SLC7A8) was methylated only in one ICC-positive sample and in none of the normal samples. CpG islands associated with additional nine genes (BRDT, C8ORF4, FLJ13937, FLJ36166, GAGED2, GPNMB, HSPA2, NALP2, and SSX4) displayed partial or complete methylation in most or all negative as well as positive samples (Table 2).

Among nine genes hypermethylated in association with ICC, six (MT1G, NMES1, RRAD, SFRP1, SPARC, and TFPI2) were analyzed in clinical samples by quantitative MethyLight assay. Numbers of negative and positive results of MethyLight assay (i.e., when the PMR value was either 0 or >0, respectively) in these samples for each gene are shown in Table 3. Results show that the aberrant methylation of studied genes was present in ICC-positive samples with high (SPARC and TFPI2_2) to modest (MT1G and NMES1) frequency. At the same time, these genes were methylated in ICC-negative samples with generally low frequency, with RRAD and TFPI2 being the only exception as its PMR value was >0 in nine and eight normal samples, respectively (Table 3).

We noticed that PMR values for all genes were low in the vast majority of normal samples, indicating that the level of methylation within DNA sequences targeted by MethyLight or the proportion of cells containing the methylated alleles (or possibly both) were low in these samples. This is illustrated in Fig. 4, in which the PMR values derived from normal and ICC samples are plotted separately for each gene. Plotted are also median PMR values for negative and ICC samples (median PMR values for negative samples plus 1 percentage point were used to set a PMR cutoff, above which samples were considered methylated; such cutoffs resulted in optimal specificity and sensitivity of our assays). Consequently, if the cutoff value was taken into account, only three normal samples scored as methylated for TFPI2_2 and none for RRAD. On the other hand, 13 ICC samples scored positive for TFPI2_2 and nine for RRAD. We subsequently generated a heatmap showing the log₁₀-transformed PMR values for all samples and genes, if they exceeded the cutoff (Fig. 5A).

The data obtained thus far (Table 3; Fig. 4 and Fig. 5A) suggested that some of the novel genes (e.g., SPARC and TFPI2) may be aberrantly methylated in ICC with the highest frequency ever observed for any single gene in cervical cancer, while at the same time, displaying very low frequency of methylation in normal cervical tissue. We have previously identified three genes (DAPK1, RARB, and TWIST1) as frequently hypermethylated in ICC (8) and wished to compare the rate of methylation of these “benchmark” genes with our new genes. The methylation status of these three genes was determined by MethyLight in the same set of samples as above. Resulting PMR values, if above the cutoff, were plotted in Fig. 5 along with six newly identified genes (SPARC, TFPI2, SFRP1, RRAD, MT1G, NMES1). For both DAPK1 and RARB, 13 ICC samples scored above the cutoff. SPARC and TFPI2 compared favorably with these markers (18 and 19 scored above cutoff, respectively), suggesting the potential value of these two novel genes as methylation markers for detection of cervical cancer and its precursor lesions.

We noticed that an alternative amplicon TFPI2_4 placed outside of original MethyLight amplicon TFPI2_2 in the CpG island associated with TFPI2 gene (see map in Fig. 5C) was able to score additional ICC samples as positive while remaining undetected in all but one negative samples (Fig. 5B). Interestingly, detection of TFPI2_4 was observed in 13 samples, where TFPI2_2 was positive but also in six samples with undetectable or below-cutoff TFPI2_2 result (Fig. 5B). On the other hand, whereas TFPI2_2 was not positive in any samples where TFPI2_4 scored negative among ICC samples described in this study, we have detected TFPI2_2 in DNA derived from CIN3/CIS precursor lesions, where TFPI2_4 was negative.³

This suggests a variable methylation pattern along TFPI2-associated CpG island and also provides a rationale for designing multiple MSP or MethyLight assays per gene to achieve maximum detection sensitivity. Altogether, the two TFPI2 assays (TFPI2_2 and TFPI2_4) scored in 20 of 22 ICC samples as positive while being so in only three normal samples. Both SPARC and TFPI2 were methylated in all but one ICC samples. However, the sample negative for these two genes was positive for RARB (Fig. 5). Thus, methylation of CpG islands associated with three genes (SPARC, TFPI2, and RARB) was detected in 100% of cancers and in only 19% of normal cervical samples. These three CpG islands together may therefore represent the most sensitive and specific methylation markers for detection of cervical cancer identified to date.

In this study, we did a systematic global search for aberrantly methylated CpG islands to identify potential novel methylation biomarkers for the detection of ICC. We used DNA methyltransferase and histone deacetylase inhibitors, DAC and trichostain A, to demethylate/reactivate hypermethylated/silenced genes in four cervical carcinoma cell lines and then followed by their identification on expression microarray and by examination of their methylation status in cell lines and in clinical samples. The impetus for our investigation has been provided by encouraging results of published studies using methylation biomarkers for detection of ICC and its precursor stages (5-8).

Treatment of cell lines with DNA methyltransferase and histone deacetylase inhibitors induced widespread expression changes in hundreds of genes in all cell lines. A total of 235 genes were up-regulated after treatment in at least two cell lines. These 235 up-regulated genes included genes with known or putative tumor suppressor activity, such as CDH1, DKK3, PTEN, SFRP1, TFPI2, and TIMP3 (9, 23-25). Some of the genes that were up-regulated in cervical carcinoma cell lines have been previously reported to be up-regulated by epigenetic reactivation in pancreatic (10), colorectal (9), and the head and neck squamous cell cancer cell lines (26). We examined the methylation status of CpG islands in 39 of the 235 putatively hypermethylated genes by MS-MCA and MSP and found 33 (85%) to be hypermethylated in at least one of the four cervical carcinoma cell lines, and 17 (44%) to be hypermethylated in three or more of the four cell lines (Table 2). Treatment-induced up-regulation of some genes was not a direct consequence of CpG island demethylation, as 6 of 39 analyzed genes (15%) were not methylated in cell lines at all (Table 2). Up-regulation of such genes could have been a secondary effect caused by an induced upstream factor, or an effect of histone reacetylation induced by trichostain A. A substantial number of putatively hypermethylated genes remain unexamined at present (Supplementary Tables S02 and S03). A considerable number of genes were down-regulated rather than up-regulated by the drug treatment (Supplementary Table S02). Although these genes were not the main focus of this study, we examined methylation of CpG islands associated with 10 genes down-regulated in multiple cervical carcinoma cell lines (SERPINH2, GBE1, HMG17, NUCKS, LY6K, CCNA2, PLK1, DDIT4, PSMB10, and NMI). Among these, only a CpG island associated with SERPINH2 was found hypermethylated in cell lines as well as in DNA derived from normal cells (data not shown).

Nine genes that were shown to be hypermethylated in multiple cervical cancer cell lines were found to be hypermethylated in clinical samples (exfoliated cervical cell samples) in association with biopsy-confirmed ICC. Because neither MS-MCA nor MSP assays provide quantitative information about the relative proportion of methylated alleles in clinical material, we developed the MethyLight assay for six of these genes (MT1G, NMES1, RRAD, SFRP1, SPARC, and TFPI2). Using MethyLight, we identified, most notably, SPARC and TFPI2 as aberrantly methylated in all but 2 of 22 ICC-positive samples (90.9%) and only in 3 of 21 normal samples (Table 3; Figs. 4-5). These hypermethylated genes are therefore of interest as biomarkers for cervical cancer either alone or together with our recently described “cervical cancer screening panel” (8) or panels reported by other authors (5-7).

Five of the six newly identified genes hypermethylated in ICC (SPARC, TFPI2, SFRP1, MT1G, and RRAD) have previously been reported to be hypermethylated in cancer, although not in ICC. SPARC has been found hypermethylated in pancreatic carcinoma cell lines, tumor xenografts, and in primary pancreatic carcinoma (27). The protein product of this gene (secreted protein acidic and rich in cysteine, also known as osteonectin or BM-40) belongs to a group of matricellular proteins that mediate cell/extracellular matrix interactions and is involved in cell proliferation, spreading, adhesion, motility, and invasion (28). TFPI2 encodes a Kunitz-type serine protease inhibitor and has been previously reported to be hypermethylated in choriocarcinoma, gliomas, and lung cancer (24, 29, 30). Tumor suppressor and Wnt antagonist SFRP1 has been shown hypermethylated and silenced in colorectal cancer (9). Expression of metal ion-binding protein MT1G has also been shown silenced by aberrant methylation in thyroid cancer (31). RRAD, encoding a protein belonging to a subfamily of Ras-related GTP-binding proteins (32), has been recently found aberrantly methylated in mesotheliomas in relationship with SV40 infection (33). Hypermethylation of NMES1 has not been previously noted in any tissue, although this gene has been reported to be expressed in normal but not in malignant esophageal tissues (34).

Nine genes (BRDT, C8ORF4, FLJ13937, FLJ36166, GAGED2, GPNMB, HSPA2, NALP2, and SSX4) were found to be hypermethylated at high frequencies in both benign and malignant cervical clinical samples (Table 2). This suggests either that they are hypermethylated in all cells in cervical epithelium (normal and neoplastic) or that they are hypermethylated in a subset of cells (e.g., inflammatory cells, fibroblasts, or endothelial cells) that were present in exfoliated cell samples from women with and without cancer. Of these, SSX4 and GAGED2 reside on the X chromosome and belong to a group of cancer/testis antigens and therefore may be hypermethylated due to X chromosome inactivation or as a general phenomenon related to silencing of many cancer/testis antigen genes in normal somatic tissue (35). Similarly, BRDT, located on chromosome 1, is a cancer/testis antigen gene (36).

It has been shown previously that the methylation of certain CpG islands increases with advanced age (4). We did not find any association between age and the overall prevalence of methylation among the nine genes we examined. More specifically, among ICC patients, 31 to 40 and 41 to 50 years age groups had almost identical average number of methylated genes per patient (4.6 and 4.0, respectively). Among patients with normal diagnosis, all DNA samples with methylated genes were in the 41 to 50 years age group (average, 0.55 methylated gene/patient), whereas no methylation was found in almost identical number of patients in 31 to 40 and 51 to 60 years age groups. It was impossible to assess the correlation between HPV type (low risk versus high risk) because of the small number of cases examined.

In our study, we have carried out the first step in the discovery of potential highly sensitive and specific methylation markers for detection of cervical cancer. We have identified a panel of novel genes with ICC-associated methylation occurring at the highest frequency ever observed. To establish the use of these markers for early detection of cervical cancer, the performance of these markers needs to be examined in large independent sets of clinical samples collected from cervix (exfoliated cells or biopsy) and from remote sites (blood or urine) of women with ICC and with early stages of this disease. Such studies are currently under way in our laboratory. It would also be interesting to examine the role of silencing of SPARC and TFPI2 in cervical cancer, as these two genes mediate the relationship of tumor cells with their environment (24, 28) and may therefore represent therapeutic targets.

We thank Jhon-Chun Ho, Desiree Willis, Cordelie E. Witt, and Yuanjun Deng for technical help and Andy Lin, Daniel Stone, and Randolph L. Woltjer for critical review of the article.