Human cancer cell lines are commonly used in basic cancer research to understand the behavior of primary tumors. Aberrations in the DNA methylation patterns are nowadays recognized as a hallmark of the cancer cell. However, no comprehensive study defines the DNA methylation environment present in the established cancer cell lines used in everyday laboratory-based research. To address this matter, we have analyzed 70 widely used human cancer cell lines of 12 different tumor types for CpG island promoter hypermethylation of 15 tumor suppressor genes, global 5-methylcytosine genomic content, chemical response to the demethylating agent 5-aza-2′-deoxycytidine, and their genetic haplotype for methyl-group metabolism genes. Several conclusions arise from our study: (a) a specific profile of CpG island hypermethylation exists for each tumor type, allowing its classification within hierarchical clusters according to the originating tissue; (b) cancer cell lines generally have higher levels of CpG island hypermethylation than primary tumors, because of the contribution of particular CpG islands and tumor types; and (c) there are no major differences between cell lines in their 5-methylcytosine DNA content, efficacy of 5-aza-2′-deoxycytidine treatment, and distribution of allelotypes of methyl-group metabolism genes. Our data provide a basis for a better use of human cancer cell lines in basic and translational research with respect to their DNA methylation environment.
Human cancer cell lines are a widely used experimental tool, and their use in cancer research has several advantages, e.g., a wide spectrum of tumor types is commercially available, they are easy to culture in vitro, yielding significant amounts of high-quality RNA and DNA, and although primary tumors are heterogeneous and invariably contaminated with normal cells, tumor cell lines are usually free of normal contamination. However, several drawbacks should also be recognized. The first is the problem of cross-contamination, which, ever since the first mammalian cell lines were established, has led to misleading results (1). Other questions about human cancer cell lines concern how closely related they are to the primary tumors from which they originated and, even more importantly, how accurately a particular cell type represents the general population of primary tumors. This latter concern is very important if we wish to understand the genetic and epigenetic factors involved in the development of cancer in the human population. It is necessary to separate what underlies the transformed phenotype and what is only induced by cell culture.
In the last decade, aberrations in DNA methylation patterns have been accepted to be a common feature of human cancer (2, 3). The two major changes occurring are global genomic hypomethylation, probably related to chromosomal instability and reactivation of endogenous parasitic sequences, and transcriptional silencing of tumor suppressor genes (p16INK4a, hMLH1, and BRCA1) associated with the hypermethylation on the CpG islands located in their promoter regions (2, 3). In recent years, a CpG island hypermethylation profile of human primary tumors has emerged, showing specific gene promoter hypermethylation of these genes dependent on the tumor type (4, 5), e.g., the mismatch repair gene hMLH1 is silenced by hypermethylation only in colorectal, endometrial, and gastric tumors, whereas the double-strand break repair gene BRCA1 is methylated in breast and ovarian tumors but not in other types (4). Other genes, such as the cell cycle inhibitor p16INK4a and the DNA repair gene MGMT, are less exclusive and aberrantly methylated over a range of different tumor types, perhaps reflecting their more general role in tumorigenesis. However, no comprehensive profile of CpG island hypermethylation in the established human cancer cell lines has thus far been built up.
The need to resolve this matter is ever more pressing for several reasons: (a) the old axiom that cancer cells are more prone to methylation than primary tumors (6) persists without any satisfactory proof; (b) the exponential use of cancer cell lines in basic research that does not take into account the epigenetic environment of that cell line, e.g., if we wish to study a putative cell cycle gene, it would be worth knowing how many and which cell cycle genes (p16INK4a, p15INK4b, etc.) have hypermethylation-associated silencing in the experimental model; similar cases can be made about apoptosis (e.g., death-associated protein kinase, caspase-8, etc.), DNA repair (e.g., BRCA1, hMLH1, MGMT, etc.), cell adhesion (e.g., CDH1, TIMP-3,3 etc.), and hormonal response (e.g., estrogen and progesterone receptors); and (c) the proposed use of cancer therapeutics based on epigenetics (7, 8, 9) requires a profound knowledge of the DNA methylation patterns of human cancer cell lines in which new or updated chemical agents are going to be tested.
For all these reasons, we have completed a systematic study of the DNA methylation profiles in a large collection of human cancer cell lines. We have discovered the existence of a tumor type-specific pattern of CpG island hypermethylation of tumor suppressor genes that allows their classification by cell type that is not essentially different to that observed in primary tumors. This local hypermethylation occurs in the context of generalized global genomic demethylation, with similar responses to the administration of demethylating drugs between different human cancer cell lines.
MATERIALS AND METHODS
Human Cancer Cell Lines and Demethylating Treatments.
The 70 human cancer cell lines used in this study were obtained from the American Type Culture Collection (Rockland, MD), the German Collection of Microorganisms and Cell Cultures (DSMZ, Braunschweig, Germany), or were kindly provided by Dr. David Sidransky (Johns Hopkins Otolaryngology Department, Baltimore, MD), Dr. Nancy Davidson (Johns Hopkins Breast Cancer Research Program), and Dr. Scott Lowe (Cold Spring Harbor, NY). Twelve different tumor types were represented in our set of 70 tumoral cell lines: breast (MCF-7, T47D, SK-BR-3, BT-20, CAMA1, Hs578t, ZR-75-1, MDA-MB-231, MDA-MB-468, MDA-MB-453, and MDA-MB-435), colon (Colo-205, SW48, SW480, SW837, DLD-1, HCT-15, HCT-116, CACO2, LoVo, RKO, and HT-29), lung (A549, DMS53, U1752, NCI-H358, NCI-H1299, NCI-H249, NCI-H209, NCI-H157, NCI-H460, NCI-H69, and NCI-H1618), prostate (PC-3, LNCaP, DuPro-1, and Du145), leukemia (KG1a, HL-60, ML-1, K562, A3.01, and Jurkat), melanoma (G-361, Malme-3M, M14-Mel, UACC-62, UACC-257, LOX-LIVM, SK-MEL-2, SK-MEL-5, SK-MEL-19, SK-MEL-28, SK-MEL-29, SK-MEL-94, SK-MEL-100, SK-MEL-103, SK-MEL-147, SK-MEL-173, SK-MEL-187, and SK-MEL-197), bladder (T24 and RT-4), head and neck (UMSCC1 and HN12), glioma (U87MG and U373), kidney (RXF-393), thyroid (TT), and lymphoma (RAJI). Two immortalized, but not transformed, cell lines from normal HMEC and normal human diploid fibroblasts (IMR90) were also included in the study. Cell lines were maintained in appropriate media and treated with 5-aza-2-deoxycytidine (Sigma) at a concentration of 1 μm for 3–5 days to achieve demethylation.
Human Primary Tumors.
To compare overall DNA methylation data from cancer cell lines with the patterns observed in human primary tumors, we have used 233 primary tumors from the colon (n = 118), breast (n = 84), and lung (n = 31) that we have described previously and for which the CpG island hypermethylation patterns were known (10, 11). CpG island hypermethylation frequencies of particular genes were compared with those described previously in our CpG island hypermethylation profile of human primary tumors (4).
Analysis of CpG Islands Methylation Status.
DNA methylation patterns in the CpG islands of tumor suppressor genes were determined by chemical conversion of unmethylated, but not methylated, cytosines to uracil and subsequent PCR using primers specific for either the methylated or modified unmethylated DNA (12). The genes selected were p16INK4a, p14ARF, p15INK4b, O6-methylguanine DNA methyltransferase (MGMT), hMLH1, GSTP1, BRCA1, E-cadherin (CDH1), TIMP-3, death-associated protein kinase, thrombospondin-1, p73, retinoic acid receptor β2, Ras association domain family 1, and LKB/STK-11. The primers and PCR conditions for the methylation-specific PCR analysis have been described previously (4, 10, 11). Placental DNA treated in vitro with Sss I methyltransferase was used as a positive control for methylated alleles, and DNA from normal lymphocytes was used as a negative control for methylated alleles. Each PCR reaction (12 μl) was directly loaded onto nondenaturing 6% polyacrylamide gels, stained with ethidium bromide, and visualized under UV illumination.
Determination of 5-Methylcytosine DNA Content.
The 5-methylcytosine DNA content of the primary tumors and their normal counterparts was determined by high-performance capillary electrophoresis as described previously (10, 11, 13). DNA (0.5–1 μg) was incubated in 20 μl of 88% (volume for volume) formic acid at 140°C for 90 min. After hydrolysis, samples were dried by speed-vac concentration (Savant SC-200). Finally, dried hydrolyzed samples were redissolved in 2 μl of H2O Milli-Q grade and stored at −20°C until their analysis. An uncoated fused silica capillary (600 mm × 0.075 inside diameter, effective length 540 mm; Waters Chromatography S.A.) was used in a capillary electrophoresis system (Capillary Ion Analyzer; Waters Chromatography S.A.) connected to a data processing station. Three replicates of each sample analyzed were carried out. The relative degree of methylation in the DNA samples was measured as the percentage of the mC of the total cytosines, calculated as: (5mdC peak area × 100)/(dC peak area + 5mdC peak area).
Genotype Analysis of Methyl-Group Genes.
Three methyl-group metabolism genes (MTHFR, MS, and cystathionine B-synthase) were genotyped. The MTHFR-677T and MS-2756G alleles have been associated previously with global genomic hypomethylation and lower CpG island hypermethylation, respectively (11). Genotyping was carried out according to the PCR/RFLP methods described previously (11).
Homozygous Deletion Analysis.
Homozygous deletion analysis of the p16INK4a (exon 1A), p14ARF (exon 1B), and p15INK4b genes was performed by comparative multiplex PCR using β-actin fragment as the internal control, as described previously (14).
Clustering of Human Cancer Cell Lines According to CpG Island Promoter Hypermethylation.
We performed cluster analysis using the Self-Organizing hierarchical Neural Network SOTA (15), an unsupervised neural network with a binary tree topology, combining the advantages of divisive and customizable methods. The analysis was performed blind, whereby the tumor type corresponding to each sample was decoded at the end of the process. Cell lines with correlated methylation profiles and genes with similar patterns across cell lines were identified using the Pearson’s correlation coefficient (r), which was taken as a measure of similarity or distance between values. We initially used unrestricted growth; then we set down the training conditions to a defined number of cycles, until we achieved the level of variability desired to be able to single out clusters of interest within the dendrogram at a higher hierarchical level. The output file was visualized as a binary tree with the TREEVIEW program. The matrix values were transformed into a graded color pattern (red through black to green), representing the mean adjusted ratio of each point of the matrix. We assigned a predetermined color to the terminal branches according to a criterion that reflects the assumed tissue of origin of the cell line. The scale below the dendrogram depicts the correlation coefficient represented by the length of the branches connecting pairs of nodes. The programs for the analytical tools are available on the Internet.4
Statistical significance of quantitative variables was assessed using nonparametric procedures, namely the Spearman correlation coefficient and the Kruskal-Wallis test, whereas categorical variables were compared using Fisher’s exact test. For continuous variables, differences between independent samples were evaluated parametrically (t test) and nonparametrically (Wilcoxon’s test). For matched samples, i.e., levels before and after demethylant treatment, Wilcoxon’s nonparametric test for matched pairs was applied. Homogeneity of variances was investigated using the statistic test proposed by Brown and Forsythe (16), which has proved to be robust under non-normality.
CpG Island Hypermethylation Profile of Human Cancer Cell Lines.
Our analysis of 15 CpG islands of tumor suppressor genes reported previously to undergo promoter hypermethylation in primary tumors (2, 3, 4) in 70 human cancer cell lines provides us with a unique view of the epigenetic lesions underlying malignant transformation. Fig. 1 illustrates the type of DNA methylation data obtained.
From the quantitative standpoint, there was considerable variation in the percentage of CpG island hypermethylation, from the highly hypermethylated Burkitt’s lymphoma cell line RAJI (73% CpG island hypermethylation), the breast cancer cell line MDA-MB-231 (67% hypermethylation), and the leukemia cell line KG1a (67% hypermethylation) to the less hypermethylated lung cancer cell lines H460 (8% hypermethylation) and A549 (8% hypermethylation) and thyroid cancer cell line TT (7% hypermethylation). The median value of CpG island hypermethylation in the cancer cell lines was 26.6% (Percentile 25-Percentile 75: 20–33.3). The individual cancer cell lines studied ranked by degree of CpG island hypermethylation are presented in Fig. 2,A. The untransformed cell lines HMEC and IMR90 did not show any CpG island hypermethylation. The percentage of CpG island hypermethylation was not homogeneously distributed among cell types and indeed showed considerable variation (Fig. 2 C). Hematological cancer cell lines (lymphoma and leukemia) had the highest levels of CpG island hypermethylation, whereas renal and thyroid cancer cell lines had the lowest.
Qualitatively, CpG island hypermethylation of particular genes was not homogeneous across all cell lines (P < 0.05; Kruskal-Wallis test). The CpG island hypermethylation of specific genes is remarkably specific in human cancer cell lines. The most striking cases were those of hMLH1 and LKB1, which were restricted to colorectal cancer cells, p15INK4b, which was confined to leukemia cell lines (our set included two acute myeloblastic leukemias, two acute lymphoblastic leukemias, one acute promyelocytic leukemia, and one chronic myelogenous leukemia), and GSTP1, which was particularly prevalent in prostate and breast cancer cell lines (P < 0.05, comparing particular types; Fisher’s exact test). Fig. 3 shows the distribution of each particular hypermethylated CpG island in the four commonest tumor types: (a) colon; (b) breast; (c) lung; and (d) leukemia (mainly acute lymphoblastic leukemia and acute myeloblastic leukemia).
Clustering of Human Cancer Cell Lines According to their CpG Island Methylation Profile.
The molecular classification of human cancers either by allelotyping, as was done in the past, or by cDNA microarray expression profiles, as it has been done in recent years, has attracted much attention. We wondered whether the DNA methylation profile obtained in the current study would allow the classification of human cancer cell lines. Thus, the CpG island hypermethylation profile described above was converted into an anonymous file in which cell lines and cellular types were not known to the researcher. We then proceeded to cluster the human cancer cell lines as described in “Materials and Methods” by their CpG island hypermethylation profiles and subsequently decoded the identity of the samples (Fig. 4). The main finding of this experiment is that CpG island hypermethylation alone allowed the classification of most of the human cancer cell lines according to their tissue type, e.g., 100% of leukemia (6 of 6), melanoma (18 of 18), glioma (2 of 2), head and neck (2 of 2), thyroid (1 of 1), and lymphoma (1 of 1) cell lines were clustered into independent terminal branches specific to their respective organ types. The clustering classification based on CpG island hypermethylation was almost perfect in the case of colon (91%, 10 of 11) and lung (91%, 10 of 11) cancer cell lines (Fig. 4). Another interesting observation was that the only steroid hormone-dependent cells in our study tumors (breast and prostate) were also grouped into the same terminal branch (Fig. 4). Finally, the program assigned the nontransformed cell lines IMR90 and HMEC to the same terminal branch (Fig. 4).
It is also worth mentioning the cases of the cancer cell lines that were not grouped in their theoretical cell type branches, e.g., this was the case with the cell line MDA-MB-435 (Fig. 4), which has long been believed to be a breast cancer cell line. Our analysis situates this cell line in a different branch to that typifying breast cancer. These results are supported by microarray expression analysis, which also strongly suggests that MDA-MB-435 is not a breast cancer cell line (17). Another illustrative example is provided by the putative breast cancer cell line Hs578t. Our analysis clearly clustered this cell line outside the terminal branch of the classical epithelial mammary cancer cell lines (Fig. 4). Supporting again our DNA methylation-based clustering, the microarray expression data also indicate that Hs578t is in fact a stromal-like cell line (17). Thus, the development of CpG island hypermethylation profiles is also a good tool for clarifying the origin of misplaced cancer cell lines. Another excellent example of the possible uses of CpG island hypermethylation profiles is provided by the colon cancer cell lines DLD-1 and HCT-15. Originally believed to be two completely different cell lines, an increasing amount of genetic evidence (18, 19) suggests that, in fact, they are one cell line driven from the same patient. Our finding that they have an identical CpG island hypermethylation pattern (Fig. 4), whereby they cluster in the same branch of the hierarchical clustering, strongly supports this idea. Finally, our observation that the cancer cell lines DMS53 (lung), SW48 (colon), and BT-20 (breast) are clustered outside their typical cell type terminal branches should stimulate the re-examination of the cell type origin of these cancer cell lines. The case of BT-20 is particularly intriguing because it was isolated in the same laboratory than the putative mammary cancer cell line BT-549, which microarray expression pattern also differs from a typical epithelial breast cancer cell line, and instead, it fits with a stromal-like cell line (17).
DNA Content of 5-Methylcytosine of Human Cancer Cell Lines and the Effect of 5-Aza-2′-Deoxycytidine.
Transformed cells also undergo massive genomic hypomethylation in the context of the CpG island hypermethylation described previously. Therefore, we wanted to analyze the content of 5-methylcytosine DNA among the 70 established human cancer cell lines of our study.
There was a lot of 5-methylcytosine DNA content in the human cancer cell lines assessed, the lowest levels being found in the glioma cell lines U373 (1.3%) and Du145 (1.4%) and the highest in the colorectal cell line SW48 (5.6%) and melanoma cell line SK-MEL-28 (5.2%). The mean 5-methylcytosine DNA content was 3.3%. No significant differences in the amount of 5-methylcytosine DNA were found when different tissue types were considered (P = 0.329; Kruskal-Wallis test). Fig. 2 B illustrates the gradient of values obtained.
Finally, we wondered whether all cell lines are equally sensitive to the demethylation of their DNA when chemical inhibitors of DNA methylation, such as 5-aza-2′-deoxycytidine, are used. This matter could be of crucial clinical importance in the future if the experimental treatments of neoplasms with demethylating agents one day become widely accepted in oncology. To address this issue, we treated 16 human cancer cell lines (MDA-MB-435, MCF-7, HCT-116, RKO, SW48, COLO-205, HT-29, DuPro-1, Dn145, PC3, H1618, H69, DMS53, KG1a, HL-60, and RXF-393) with the demethylating drug 5-aza-2′-deoxycytidine and measured the degree of genomic hypomethylation produced by this chemical inhibitor. The 5-methylcytosine DNA content was significantly reduced in an average of 49% of the cell lines after treatment with 5-aza-2′-deoxycytidine (P = 0.0004; Wilcoxon matched pair rank test). The key point of the analysis, as shown in Fig. 2,D, is that despite there being many initial 5-methylcytosine contents (from 1.4 to 5.6%) with a mean of 3.3% in untreated cancer cell lines, after the use of the demethylating agent 5-aza-2′-deoxycytidine, values of the remaining 5-methylcytosine all clustered within a very narrow range (from 1.96 to 0.66%; mean, 1.45%). The comparison of the variances of the SDs of 5-methylcytosine DNA content before (SD = 1.208) and after (SD = 0.397) treatment with 5-aza-2′-deoxycytidine demonstrates that all treated cancer cell lines are homogeneously identical with respect to their 5-methylcytosine DNA content (P = 0.00047). From the gene re-expression point of view, 5-aza-2′-deoxycytidine was able to reactivate all methylation-silenced genes analyzed, as we have demonstrated previously (reviewed in Ref. 4), exemplified in the cases of the MGMT and retinoic acid receptor β2 genes across different colon and breast cell lines (data not shown). Fig. 2 D illustrates the results.
Analysis of the Relevance of Germ-line Variants in Methyl-Group Genes in the DNA Methylation Profiles of Human Cancer Cell Lines.
SAM is the universal methyl-donor for all methylation reactions, including DNA methylation, and so the existence of genetic difference in the activity of the genes involved in their metabolism may affect the DNA methylation patterns. We have recently confirmed the principle of this assumption in human primary tumors (11), whereby the analysis of a large collection of human primary tumors revealed that germ-line variants in MTHFR and MS genes influence the 5-methylcytosine DNA content and number of hypermethylated CpG islands, respectively (11). To establish whether the presence of these genetic variants in the cultured cancer cell lines has any effect on DNA methylation patterns, we studied four known polymorphic alleles of the methyl-group genes MTHFR, MS, and cystathionine β-synthase in our collection of human cancer cell lines. The four germ-line variants analyzed are relatively common in the general population, yield less active proteins, and thus reduce the availability of the universal methyl-donor SAM (11). The genotyping of each allele singly or in combination with others in different haplotypes revealed no relation either with CpG island hypermethylation or 5-methylcytosine DNA content. These findings are consistent with the observation that human cancer cell lines are cultured with media that contain huge amounts of folates, the precursor of SAM, and are less dependent on external methylation-related environmental or dietary factors than are primary tumors.
Comparison of DNA Methylation Profiles in Human Cancer Cell Lines versus Primary Tumors.
We also compared the degree of CpG island hypermethylation in these 70 established cancer cell lines with those we observed previously in a large collection of human primary tumors (n = 233) from the breast, colon, and lung using identical methodology (10, 11). The median level of CpG island hypermethylation in this set of human primary tumors was 17% (Percentile 25-Percentile 75: 0–33). An increased frequency of overall CpG island hypermethylation in established human cancer cell lines compared with human primary tumors is thus observed and significant using parametric (t test) and nonparametric (Wilcoxon) tests (P = 0.0001 in both cases; Fig. 3 B).
To determine whether the 5-methylcytosine DNA content of cultured cancer cells reflects the values found in patients with neoplasms, we also compared the levels of 5-methylcytosine DNA in our set of cancer cell lines with values in human primary tumors obtained using the same technology (10, 11). Cancer cell lines, on average and by tumor type, were more hypomethylated overall than were their primary tumor counterparts (means of 3.3 and 4.5%, respectively; t test: P = 0.003; Wilcoxon: P = 0.027). The existence of contaminating normal tissue with higher 5-methylcytosine DNA content in the primary tumors, which the cultured cancer cell lines do not have, may in part explain these results, but the possibility that cell culture itself leads to higher levels of global genomic hypomethylation cannot be excluded (Fig. 3 B).
Finally, we compared the rates of each particular hypermethylated CpG island in the cancer cell lines with the frequency of aberrant methylation in each particular tissue type of our set of human primary tumors, where their CpG island methylation patterns were known previously (10, 11). Most of the CpG islands studied showed a similar distribution between cancer cell lines and primary tumors. However, significant statistical differences were found between cultured cells and primary tumors for a few CpG islands, consistently whereby a few genes were more hypermethylated in cancer cell lines than in primary tumors. One example is that of the cell cycle inhibitor p16INK4a, which was hypermethylated in all of the colon cancer cell lines studied but was only hypermethylated in ∼37% of primary colorectal tumor cases. Fig. 3 A compares the percentage of CpG island hypermethylation of particular genes in the cancer cell lines and primary tumors.
Established human cancer cell lines are an extremely useful tool for the cancer researcher. The ease of management and possibility of subjecting them to different chemical treatments or DNA/RNA engineering processes open up excellent experimental avenues that are precluded in the study of human primary tumors. In fact, most of our current knowledge of the genetic alterations present in human cancer is derived from in vitro established human cancer cell lines. The recent emergence of epigenetics as a force in the genesis and development of tumors means that the careful characterization of the DNA methylation patterns in human cancer cell lines used around the world is now essential. In this study, we have provided a comprehensive overview of this field.
The first conclusion derived from our study is that every cancer cell line has at least one hypermethylated gene in each major cellular pathway, and the target gene may differ between tumor types, e.g., genomic damage can arise in colorectal cancer cell lines by hypermethylation-associated silencing of the mismatch repair gene hMLH1 and/or the DNA repair MGMT, but breast tumors present only MGMT hypermethylation and never hMLH1. An additional example from another pathway: the capacity of cancer cells to spread is in part mediated by two genes, the intercellular adhesion molecule E-cadherin (CDH1) and TIMP-3. Colorectal cancer cell lines show epigenetic silencing of TIMP-3 but not of CDH1, whereas mammary cancer cell lines show no TIMP-3 hypermethylation but frequent CDH1 aberrant methylation.
A second conclusion from our profiling is that human cancer cell lines closely reflect the primary tumors of that type, e.g., only colorectal cancer cell lines are methylated at hMLH1 in the same way that only colorectal primary tumors are hypermethylated at hMLH1 (4). Endometrial and gastric tumors are also hypermethylated at hMLH1, but cell lines from these tumor types were not part of our panel. Another example is that of the detoxifier gene GSTP1; in primary neoplasms, this gene is mainly methylated in prostate, breast, and kidney tumors, and again, when we analyzed our set of cell lines, GSTP1 hypermethylation was almost exclusive to prostate, breast, and kidney cancer cells (4). Our CpG island profile may provide important insights into the cancer epigenetics of as yet unstudied primary tumors, e.g., the high rate of Ras association domain family 1 hypermethylation in melanoma cell lines may indicate an important role for this tumor suppressor gene in melanoma patients. Additional work is now needed to study the tumor types that we have not extensively covered in our study, such as bladder, hepatocellular, or gynecological cancer.
The enormous specificity of CpG island methylation patterns we have found in human cancer cell lines has allowed us to establish a hierarchical clustering of them. Using a blind model, cancer cell samples matched tumor cell types satisfactorily when their identities were decoded. We were even able to “catch” the cancer cell lines incorrectly classified by their cell type in the literature (i.e., MDA-MB-435) or reveal the cell lines that are just duplicates (i.e., DLD-1 and HCT-15). As the “prediction” of the cancer type was not the main aim of our study of DNA methylation patterns of cancer cell lines, there is still much scope for improving our clustering, eliminating the less informative CpG islands, and adding other hypermethylated CpG islands with marked tumor type specificity. Compared with laborious and expensive expression microarrays, this approach is a good alternative for tumor classification, using nonradioactive user-friendly technology that can be used with paraffin-embedded material. Proof of principle was also provided recently by the use of DNA methylation markers in a hierarchical clustering of non-small lung cancer cells versus small cancer cells (20). In a similar fashion, different automated CpG island arrays or global-genomic DNA methylation assays published previously may also be useful to elaborate more subtle patterns of clustering based on DNA methylation (5, 21, 22, 23).
A third observation rekindles an old controversy: the axiom that cancer cell lines have more CpG island hypermethylation than do primary tumors. We have found that, on average, these 70 widely used and long-established human cancer cell lines have more CpG island hypermethylation for the 15 genes selected in our study than our panel of human primary tumors. However, as mentioned in the “Results,” this significant finding is attributable mainly to the contribution of particular CpG islands in particular tumor types. A noteworthy example is provided by the cell cycle inhibitor p16INK4a, which is hypermethylated in all colorectal cancer cell lines studied, but only in 35% of primary colorectal tumors (4), implying an essential role in overcoming senescence in this cultured cell type. Another example is that of the DNA repair gene MGMT, which was aberrantly methylated in half of the mammary cancer cell lines studied but in none of the primary breast tumors (4). In fact, our findings reconcile those reports that describe cultured cancer cell lines as being more hypermethylated than primary tumors (6, 24) and others that show the cancer cell lines and primary tumors from which they were established are methylated to a similar extent (25). Most frequently, the differences found relate to the CpG islands studied and methodology used. In our study, all of the genes included share the properties of having unmethylated CpG islands in normal tissues, hypermethylation of the islands leading to transcriptional silencing, and a biologically plausible role in cancer, e.g., permanent lines of cultured cells can have increased hypermethylation of CpG islands associated with tissue-specific genes, whereas in normal primary tissues, they remain unmethylated, regardless of their expression status. Most important, although the cancer cell lines have an increased rate of hypermethylation of certain CpG islands, they retain the specific CpG island hypermethylation profile of each tumor cell type, as demonstrated by our comparison with the primary tumors and in our hierarchical cluster analysis.
Another “quantitative” conclusion that can be drawn from our results is the existence of a range of CpG island hypermethylation. We observed examples across the entire spectrum, from the least hypermethylated cancer cell lines to the most hypermethylated cell line (the Burkitt’s lymphoma cell line RAJI). This pattern does not provide any evidence for the existence of any CpG island methylator phenotype at all. Instead, some cancer cell lines are simply more hypermethylated than others, as we would expect by chance, as we have demonstrated previously for primary human tumors (4). This reasoning can also be applied to the total 5-methylcytosine DNA content of the cancer cell lines, which also has a wide gradient range.
Finally, it is worth mentioning that established human cancer cell lines can be excellent models for finding and testing new and more specific demethylating agents. These types of drugs, specially 5-aza-cytidine and 5-aza-2′-deoxycytidine (decitabine), have been used in the treatment of several types of neoplasms (8), but toxicity and the possibility of affecting the normal (“good”) methylation have always been a concern of researchers and clinicians. Our panel of hypermethylated CpG islands from 12 different tumor types provides new markers that can be tested for putative demethylation events after in vitro assays with new chemical or genetically engineered inhibitors of DNA methylation. Our demonstration that cancer cell lines behave in a similar manner than human primary tumors, with respect to their CpG island methylation patterns, total 5-methylcytosine genomic content, and hypomethylation response to 5-aza-2′-deoxycytidine, indicates that they are valuable tools for the standardization of new drugs based on epigenetics. These studies should now be extended to other available demethylating agents, such as procainamide or nebularine.
In summary, we have provided the first systematic profile of the DNA methylation patterns in the human cancer cell lines most frequently used throughout the world. This study may serve as a starting point for extending our knowledge of the field of epigenetics in established cancer cell lines. It is necessary not only to increase the number of lines and loci analyzed but also to study the different patterns of gene expression of the DNA methylation (DNA methyltransferases and methyl-binding proteins) and chromatin (histone deacetylases and histone acetyltransferases) machinery elements present in each one. The DNA methylation profile currently provided is the first step toward reaching that goal.
The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.
Supported by Grants I+D+I SAF2001-0059, CAM08.1/0010.1/2001 and the International Rett Syndrome Association.
The abbreviations used are: TIMP-3, tissue inhibitor of metalloproteinases-3; MTHFR, methylene-tetrahydrofolate reductase; HMEC, human mammary epithelial cell; MS, methionine synthase; SAM, S-adenosylmethionine; GSTP1, glutathione S-transferase P1.
Internet address: http://bioinfo.cnio.es/sotarray.
We thank Dr. Michel Herranz for helpful comments and assistance with the graphic designs.