We have generated DNA methylation profiles of 148 human breast tumors and found significant differences in hormone receptor (HR) status between clusters of DNA methylation profiles. Of 35 DNA methylation markers analyzed, the ESR1 gene, encoding estrogen receptor α, proved to be the best predictor of progesterone receptor status, whereas methylation of the PGR gene, encoding progesterone receptor, was the best predictor of estrogen receptor status. ESR1 methylation outperformed HR status as a predictor of clinical response in patients treated with the antiestrogen tamoxifen, whereas promoter methylation of the CYP1B1 gene, encoding a tamoxifen- and estradiol-metabolizing cytochrome P450, predicted response differentially in tamoxifen-treated and nontamoxifen-treated patients. High levels of promoter methylation of the ARHI gene, encoding a RAS-related small G-protein, were strongly predictive of good survival in patients who had not received tamoxifen therapy. Our results reveal an as yet unrecognized degree of interaction between DNA methylation and HR biology in breast cancer cells and suggest potentially clinically useful novel DNA methylation predictors of response to hormonal and non-hormonal breast cancer therapy.

Breast cancer is the most common malignancy among females in most Western countries, who have an overall lifetime risk of >10% of developing invasive breast cancer (1). The presence of estrogen receptor (ER) and/or progesterone receptor (PR) is an important diagnostic feature of breast cancer, reflective of disease etiology (2) and predictive of response to treatment with the antiestrogen tamoxifen (3, 4). Recent advances in molecular profiling by gene expression (cDNA) microarrays have led to a further refinement of the subclassification of breast cancer into five major subtypes and to the identification of gene expression signatures associated with prognosis (5, 6, 7, 8, 9). Molecular profiling in breast cancer has thus far focused primarily on the use of cDNA microarrays, which are limited by the innate instability of RNA and are poorly compatible with the formalin fixation and paraffin embedding of tumor tissues used in routine histopathology. Therefore, we have explored the use of DNA methylation markers as an alternative approach to molecular profiling (10). Hypermethylation of promoter CpG islands, which is frequently observed in breast cancer (11, 12, 13), is often associated with transcriptional silencing of the associated gene, thus providing a DNA-based surrogate marker for expression status (14). Microarray-based methods of DNA methylation analysis are hampered by modest quantitative accuracy, poor sensitivity to low levels of CpG island hypermethylation, and technical challenges in target DNA preparation, which requires either bisulfite PCR amplification of each individual locus (15) or the use of restriction enzyme digestion (11), which is not consistently reliable with formalin-fixed tissues. As an alternative, we have used a moderate-throughput, fluorescence-based, semiautomated quantitative technique called MethyLight (16) to screen a panel of 35 methylation markers in 148 cases of breast cancer. Interestingly, we found that among these 35 markers, the best predictor of ER status was methylation of the PR gene (PGR). Conversely, the best predictor of PR status was methylation of the ER gene (ESR1).

Hormone receptor (HR) status, defined as ER and/or PR positivity, has been shown to predict response to tamoxifen treatment (3, 4). Interestingly, although tamoxifen is thought to act through the ER, PR status is an independent factor predictive of adjuvant endocrine treatment benefit (4). Tamoxifen, which is a selective ER modulator, has been shown to dramatically reduce the risk of breast cancer (17) and of breast cancer recurrence (18). Since its introduction more than 25 years ago, tamoxifen has been the mainstay of the endocrine adjuvant treatment of breast cancer, has become the most widely used anticancer drug, and may be considered one of the first targeted therapies (18). In this study we found that ESR1 methylation predicts survival only in tamoxifen-treated patients and that ARHI methylation predicts survival only in non-tamoxifen-treated patients, whereas CYP1B1 methylation predicts survival differentially in tamoxifen-treated and nontreated patients. We propose that these differences in DNA methylation profiles reflect alternative pathways of tumorigenesis associated with differences in HR status, possibly due to different originating cell types (9) and/or disease etiology (2).

Tissues.

Tumor samples were retrieved from the tissue bank of the Department of Obstetrics and Gynecology, Innsbruck University Hospital (Innsbruck, Austria). Clinical, pathological, and follow-up data are stored in a database in accordance with hospital privacy rules. Specimens were brought to the pathologist (E. M-H.) immediately after resection, and part of the tissue was placed in liquid nitrogen and stored at −80°C until lyophilization. A total of 148 patients with breast cancer treated at the Department of Obstetrics and Gynecology, University of Innsbruck between 1989 and 2000 were included in this study. Patient characteristics are provided separately in Data Supplement 1.

Histopathological Analyses.

All breast cancer specimens were reviewed by a single pathologist (E. M-H.). HR positivity was defined as presence of ER and/or PR in >10% of tumor cells (immunohistochemistry was done for the 106 breast cancers) or ≥15 fmol/mg protein (biochemical assays were performed for 42 breast cancer specimens, which were obtained before immunohistochemistry had been established in our laboratory).

DNA Methylation Analyses.

Genomic DNA was isolated using a QIAmp tissue kit (Qiagen, Hilden, Germany). Sodium bisulfite conversion of genomic DNA was performed as described previously (16). DNA methylation analysis was performed by MethyLight (16, 19). Three sets of primers and probes designed specifically for bisulfite-converted DNA were used [a methylated set for the gene of interest and two reference sets, β-actin (ACTB) and collagen 2A1 (COL2A1)] to normalize for input DNA. The specificity of the reactions for methylated DNA was confirmed separately using SssI (New England Biolabs)-treated human peripheral blood lymphocyte DNA (Promega), which results in near complete methylation of this reference DNA (16). The percentage of fully methylated molecules at a specific locus was calculated by dividing the GENE:ACTB ratio of a sample by the GENE:ACTB ratio of SssI-treated sperm DNA and multiplying by 100 and calculated separately by dividing the GENE:COL2A1 ratio of a sample by the GENE:COL2A1 ratio of SssI-treated sperm DNA and multiplying by 100. The mean of these two resulting values was used in subsequent statistical analyses. We use the abbreviation PMR (percentage of fully methylated reference) to indicate this measurement (20). The initial 64 methylation markers and the final panel of 35 markers were selected based on published reports demonstrating a role for DNA methylation in breast cancer or due to the fact that they are involved in HR action. Primer and probe sequences are shown in Data Supplement 2.

Expression Analyses.

RNA isolation and expression analyses were performed as described previously (19). Before cDNA synthesis, RNA samples were treated with DNase to ensure removal of contaminating genomic DNA. TATA box-binding protein served as the reference gene. Primer and probe sequences are shown in Data Supplement 3.

Statistics.

To cluster samples and DNA methylation markers, we used agglomerative hierarchical cluster analysis in SPLUS 2000 (Insightful Corp.) Because many of the CpG regions had undetectable methylation, we categorized the PMR values into quartiles (coded 1–4). If >25% but <50% of the samples had undetectable methylation, this resulted in scores of 1 (undetectable methylation), 2 (detectable methylation and ≤50th percentile), 3 (51st–75th percentile), and 4 (>75th percentile). If >50% but <75% of the samples had undetectable methylation, the scores were 1, 3, and 4. If >75% of the samples had undetectable methylation, the score was either 1 for undetectable methylation or 4 for detectable methylation. Manhattan distance, the sum of the absolute deviations across methylation markers, was used to measure dissimilarity. The dissimilarity between clusters was measured by the group average method (21). We tested the association between (categorized) PMR values and HR status using logistic regression. Separate analyses were conducted for each gene. Multiple linear regression was used to study the relationship between DNA methylation and ESR1 gene expression. A total of 75 samples had ESR1 gene expression measured, but one was omitted from the analyses as an outlier (ESR1 upstream A expression > 3 SDs above the mean). Comparisons between groups of samples and between different genes were simplified by expressing all expression data relative to the mean for the entire set of 74 samples. We used Cox regression to study the association between PMR values and overall (and disease-free) survival, treating PMR quartiles as ordered categorical variables. Using an interaction model, we tested whether the association of PMR values and survival varied by treatment with tamoxifen therapy (received tamoxifen treatment versus did not receive tamoxifen treatment; tamoxifen treatment was defined as 20 mg of tamoxifen daily for 5 years or until recurrence of disease). The analyses were adjusted for nodal status (0, 1–3, and >3) and tumor stage (I, II, and III/IV). Nodal status was coded using indicator variables for two of the three levels. All analyses were age-adjusted.

We prescreened 65 DNA methylation markers on a limited set of pilot samples (8 breast cancer cell lines and 8 breast carcinomas) to identify markers with sufficiently high methylation frequencies and/or methylation levels. From this initial set, 35 informative markers were selected for MethyLight analysis on 148 primary breast carcinomas obtained from the University Hospital, University of Innsbruck (a summary of clinical characteristics is shown in Data Supplement 1). A semiautomated MethyLight platform was used to execute these reactions, and data were obtained for 4978 of the 5180 analyses performed (96% success rate; a summary of the data is shown in Data Supplement 4).

Two-dimensional unsupervised hierarchical clustering analysis of cases versus methylation markers revealed that the tumors segregated naturally into groups of cases with distinct methylation profiles (Fig. 1). These two major clusters differed significantly in their HR status [P = 0.0011 for cluster 1 (indicated in green; n = 87) versus cluster 2 (indicated in red; n = 56) for ER+ versus ER−; P = 0.0013 for PR+ versus PR−; P = 0.0011 for HR+ versus HR−] and in age (P = 0.0080 for cluster 1 versus cluster 2; mean age, 57 versus 63 years, respectively). We adjusted for age in all subsequent analyses. We did not detect significant clustering of cases by HER2 status (22), menopausal status, relapse, death, grade, nodes, stage, or tumor diameter (data not shown).

The hierarchical clustering data suggest that HR status is associated with profiles of multiple methylation markers, but it does not reveal which markers contribute most to the clustering. We therefore investigated the association of each of the 35 markers with HR status individually, and we ranked them according to the strength of their association (Table 1). Fifteen of the 35 genes yielded Ps < 0.05 using methylation values (PMR measurements as described in “Materials and Methods”) as predictors of HR (ER+ and/or PR+) status (Table 1, HR Status Predictors). Since we tested 35 different markers simultaneously, we adjusted for the multiple comparisons by controlling the false discovery rate, which is the expected proportion of false positive tests among all positive tests (see “Materials and Methods”). After this adjustment, three markers (SOCS1, RASSF1A, and BCL2) were significantly associated with HR status (Table 1, HR Status Predictors). Interestingly, SOCS1 deficiency results in accelerated mammary gland development in mice (23) and is known to be methylated in human tumors (24). Promoter methylation of RASSF1A has previously been reported to be methylated in a large percentage of human breast cancers (25), and this methylation can even be detected in epithelial hyperplasia and intraductal papillomas (26). Here, we show that RASSF1A promoter methylation is associated with HR status in advanced breast tumors. Our observation of a significant negative association between BCL2 methylation and HR status is consistent with earlier reports that BCL2 expression is positively associated with HR positivity (27) and that its down-regulation is negatively associated with HR positivity (28).

Remarkably, of all 35 markers, DNA methylation of the ESR1 gene encoding the ERα was the least associated with HR status. PGR methylation was also not significantly associated with HR status after adjustment for multiple comparisons. Breast tumors are often concordant for ER and PR status. In our study, 127 of 148 breast cancer specimens were either double receptor (ER and PR) positive (n = 86) or double receptor negative (n = 41), whereas only 21 tumors were positive for either just ER (n = 12) or PR (n = 9). This highly significant association (P = 2.6 × 10−16 by χ2) between ER and PR status is attributed to induction of PGR gene expression by activated ER (29, 30). This makes it difficult to separate the effects of the two receptors. We addressed this problem by investigating which methylation markers best predict the status of ER and PR individually, while adjusting for the status of the other receptor and for age (Table 1, ER Status Predictors and PR Status Predictors). Interestingly, methylation of the 5′ CpG island of the PGR gene turned out to be the best predictor of ER status (Table 1, ER Status Predictors). Positive ER status is inversely associated with PGR CpG island methylation, consistent with the well-established induction of PGR gene expression by activated ER (31). On the other hand, ESR1 methylation turned out to be the best predictor of PR status (Table 1, PR Status Predictors), even though it is the least significant predictor of overall HR status (Table 1, HR Status Predictors). Thus, whereas neither ESR1 nor PGR methylation marker was a good predictor of overall HR status, each is the best predictor of the status of the other receptor, but not of their own cognate receptor. It is not a priori clear why this would be the case, or why there would be a positive association between PR status and ESR1 methylation (Table 1, PR Status Predictors) because activated ER is known to induce PGR gene expression. This raises the question of whether ESR1 methylation is truly reflective of reduced ESR1 expression. We analyzed ESR1 expression by quantitative real-time reverse transcription-PCR in 74 samples for which we had frozen tissue available for RNA analysis. We did not find a clear inverse relationship between ESR1 expression levels and quartiled ESR1 methylation levels when the tumors were analyzed collectively (Fig. 2). However, PR− tumors did show a statistically significant inverse trend between ESR1 gene expression levels and ESR1 methylation levels (Fig. 2). This suggests that PR+ status may confer resistance of the ESR1 gene expression to ESR1 DNA methylation. We have further explored this interesting observation, as described in “Discussion,” but we have not yet resolved the mechanism. It should be noted that the levels of methylation that we observed at the ESR1 locus are quite low, with a median PMR of 0.8 (see Data Supplement 4), and may therefore be more of a reflection of chromatin status at the ESR1 locus rather than a major driving force in silencing ESR1 gene expression. Regardless, it is yet another intriguing finding of an interaction between DNA methylation and HR biology.

Hormone therapy with the antiestrogen tamoxifen is frequently used as an adjuvant treatment in breast cancer. HR status has been shown to predict response to tamoxifen treatment (3, 4). Our patients showed a similar trend using a Cox proportional hazards model to test whether the association between HR status and survival differed by treatment with tamoxifen therapy (mean follow-up, 5.2 years). However, the interaction with treatment response was not statistically significant (Table 2). Our results described above suggest a link between HR status and cellular DNA methylation profiles in breast cancer cells. Therefore, we tested whether any of the 35 DNA methylation markers would be better predictors of response to tamoxifen treatment than HR status. We ranked our 35 DNA methylation markers according to their ability to predict response to tamoxifen therapy as measured by a test for interaction in a Cox model with adjustments for treatment-specific effects of HR status. The model was also adjusted for age, stage, and number of positive nodes. Three markers were statistically significant predictors of response to tamoxifen therapy (Table 3). Interestingly, we had shown above that ESR1 methylation is a good predictor of PR status. Here, we show that ESR1 methylation outperforms PR status as a predictor of response to tamoxifen (Table 3). We further analyzed the relationship between ESR1 methylation and tamoxifen response by comparing the survival curves of patients with above or below median ESR1 methylation levels who either received tamoxifen therapy or did not receive tamoxifen therapy (Fig. 3). High ESR1 methylation was a significant predictor of better survival in the tamoxifen-treated group but showed no significant predictive value for the non-tamoxifen-treated group (Fig. 3).

ARHI promoter methylation was a highly significant predictor of survival in patients who had not received tamoxifen therapy but showed no predictive value for patients treated with tamoxifen therapy (Table 3; Fig. 3). Finally, CYP1B1 methylation was a highly significant predictor of tamoxifen response in the interaction model (Table 3). The survival curves reveal that this is due to a differential predictive behavior of this marker in tamoxifen-treated versus nontreated patients (Fig. 3). These results show that DNA methylation markers can outperform HR status as predictors of response to tamoxifen therapy and suggest that DNA methylation markers may be of clinical use in directing hormone therapy in breast cancer patients.

Significant progress has been made in recent years toward the implementation of DNA methylation markers as clinical tools in cancer detection and diagnosis (10). Here, we explore their utility in classifying breast cancer tumors and in predicting response to hormonal therapy. Although our initial goal was open-ended and was not directed toward HR status or response to hormone therapy, unsupervised clustering of the data indicated a relationship between DNA methylation and HR status. To our surprise, methylation levels of the genes encoding the HRs were not the best predictors of HR status. This suggests that DNA methylation markers are not perfect inverse predictors of gene expression status and also that they may contain relevant information that is independent of gene expression status. Transcriptional repression by promoter DNA methylation is thought to be mediated through changes in chromatin structure. The association between DNA methylation and gene expression may therefore show threshold effects, rather than a simple linear relationship. Indeed, robust ESR1 expression in PR− tumors is seen only in the lowest quartile of ESR1 methylation (Fig. 2). The lack of observed effect of ESR1 methylation on ESR1 expression in PR+ tumors (Fig. 2) is interesting. As mentioned earlier, the PMR values obtained for ESR1 methylation are very low, with a median PMR of 0.8, which may not be sufficiently high to cause gene silencing. Moreover, our ESR1 methylation assay is located immediately downstream of the transcription start site for exon 1A of the ESR1 gene (+14 to +114), rather than in the promoter region itself, due to limitations imposed by MethyLight primer design criteria. However, PR− tumors did show the expected inverse relationship between ESR1 methylation and expression with these same assays (Fig. 2). Others have shown that ESR1 gene expression can be activated in ER− cell lines by DNA methyltransferase and histone deacetylase inhibitors (32). One hypothesis that could reconcile these disparate observations is that PR+ tumors rely more extensively on expression initiated at upstream exons of the ESR1 gene (33). Expression driven from these upstream promoters would not be expected to be affected by the DNA methylation that we measure at exon 1A (34). Indeed, we identified several putative progesterone response elements located near upstream exon 1C (data not shown). However, we found that the relative utilization of ESR1 exons 1A, 1B, 1C, 1D, and 1E was similar in PR+ and PR− tumors. A summary of this analysis is shown in Data Supplement 5. This interesting difference between PR+ and PR− tumors in the relationship between ESR1 methylation and expression will require further investigation.

Clinical and epidemiological studies in the past have suggested that breast cancer is composed of at least two distinct groups (2, 35). More recently, molecular profiling of breast cancer using gene expression profiles has revealed five distinct clusters composed of one basal-like subgroup, one ERBB2-overexpressing subgroup, two luminal-like subgroups, and one normal breast tissue-like subgroup (9). Because we have not performed gene expression microarray experiments on our group of breast tumors, we cannot directly compare our DNA methylation clustering results with the five major groups identified by gene expression profiles. However, it seems likely that the DNA methylation cluster 2, which contains mostly HR+ tumors (Fig. 1), overlaps with the two luminal-like subgroups, which contain ER+ tumors (5). The DNA methylation cluster 1 contains the majority of HR− tumors and likely overlaps with the other three gene expression subtypes, which tend to be ER− (5). It seems likely that the gene expression profile subgroupings represent a much more stable subgrouping because these analyses are based on a much larger number of samples and genes (9). Nevertheless, the undirected clustering of our methylation data led us to the identification of an interesting link between DNA methylation patterns and HR biology.

We chose to use MethyLight technology for this study, rather than methylation-specific PCR or the methylation microarray technologies currently under development. One of the unique features of MethyLight technology is that the resulting data are composed of a mixture of discrete and variable measures. The discrete measures arise from the large number of data points with undetectable methylation (PMR values of 0) versus the data points with positive detection of methylation. This type of data structure is similar to that obtained with methylation-specific PCR analysis. On the other hand, the quantitative nature of MethyLight also generates continuous measures for the samples with detectable levels of DNA methylation. We show here that useful information can be extracted from both types of measures. For example, among the methylation markers predictive of response to tamoxifen therapy, CYP1B1 was used as a discrete measure of positive versus negative DNA methylation, similar to methylation-specific PCR analysis. However, a methylation-specific PCR-based approach for the other two markers predictive of treatment response would have been noninformative because ESR1 and ARHI are positive in 100% and 99.3% of the samples, respectively (see Data Supplement 4). The quantitative aspect of MethyLight analysis was required to reveal the association of these methylation markers with response to tamoxifen therapy.

Of 35 DNA methylation markers tested, three genes showed the potential to serve as independent predictors of clinical response to systemic hormonal therapy with tamoxifen. Two of the three genes (ESR1 and CYP1B1) are known to be intimately involved in the function and metabolism of estradiol. This lends credence to the biological relevance of DNA methylation changes in breast tumors. The third gene (ARHI) encodes a RAS-related small G-protein, which may play a role in the regulation of breast cancer cell growth (36). We found that patients with high levels of ARHI methylation had better survival than patients with low levels of ARHI methylation. However, this effect was completely obliterated in the tamoxifen-treated group (Fig. 3). This may be due to the ability of antiestrogens such as tamoxifen to block growth factor-induced mitogenesis, possibly involving pathways regulated by ARHI(36, 37).

ESR1 encodes the ERα. Patients treated with tamoxifen who had high levels of tumor ESR1 methylation showed better survival than tamoxifen-treated patients with low levels of ESR1 methylation. The survival benefit in patients with high levels of ESR1 methylation may be due, in part, to the positive association between ESR1 methylation and PR status (Table 1, PR Status Predictors). PR status appears to be a better predictor of response to tamoxifen than ER status (Table 2).

CYP1B1 encodes cytochrome P450 1B1, which catalyzes the conversion of 17-β-estradiol (E2) to the catechol estrogen metabolites 2-OH-E2 and 4-OH-E2. The 2-hydroxylated form of E2 has been shown to have weak ER agonist or antagonist properties (38). CYP1B1 is also the principal catalyst of 4-hydroxytamoxifen trans-cis-isomerization, which converts the primary potent antiestrogen trans-4-hydroxytamoxifen to the weak estrogen agonist cis-4-hydroxytamoxifen (39). We have not investigated gene expression levels of CYP1B1 as a function of CYP1B1 methylation, and our measured levels of methylation are quite low (see Data Supplement 4). Nevertheless, if patients with positive CYP1B1 methylation do indeed have reduced CYP1B1 expression, then these patients would be expected to have lower rates of 4-hydroxytamoxifen trans-cis-isomerization and would thus retain higher levels of active antiestrogen. This would be consistent with the better survival of these patients in the tamoxifen-treated group (Fig. 3). Conversely, in the patients who did not receive tamoxifen therapy, patients with tumor CYP1B1 methylation would have a reduced capacity for conversion of E2 to its weaker catechol derivatives. This would be consistent with the observation that these patients show a worse survival among the group not receiving tamoxifen therapy (Fig. 3).

Our results show a level of interaction between DNA methylation changes in breast cancer and HR status or response to hormonal therapy that was not previously appreciated. Because DNA methylation markers rely on DNA as an analyte, as opposed to the more chemically labile RNA molecule, these results suggest exciting opportunities for the development of robust assays for clinical diagnosis and for predicting response to antiestrogen therapy in the adjuvant setting.

Grant support: Austrian Science Foundation Grants J2024 and P15995-B05 (M. Widschwendter).

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.

Note: Supplementary data for this article are available at Cancer Research Online (http://cancerres.aacrjournals.org).

Requests for reprints: Peter W. Laird, University of Southern California/Norris Cancer Center, Room 6418, 1441 Eastlake Avenue, Los Angeles, CA 90089-9176. Phone: (323) 865-0650; Fax: (323) 865-0158; E-mail: plaird@usc.edu

Fig. 1.

Unsupervised hierarchical clustering of DNA methylation markers and breast carcinomas. Agglomerative hierarchical analysis was used to cluster samples (columns) and DNA methylation gene markers (rows) as listed on the right. Quartiles of DNA methylation measures (percentage of fully methylated reference values) are indicated by color, ranging from dark green indicating the lowest methylation (first quartile), and light green, light red, and dark red indicating the second, third, and fourth (highest methylation) quartiles, respectively. Failed analyses are indicated in white. The two major sample clusters are shown in green (cluster 1) and red (cluster 2). Hormone receptor (HR) status is indicated by black ovals (HR−) or white ovals (HR+).

Fig. 1.

Unsupervised hierarchical clustering of DNA methylation markers and breast carcinomas. Agglomerative hierarchical analysis was used to cluster samples (columns) and DNA methylation gene markers (rows) as listed on the right. Quartiles of DNA methylation measures (percentage of fully methylated reference values) are indicated by color, ranging from dark green indicating the lowest methylation (first quartile), and light green, light red, and dark red indicating the second, third, and fourth (highest methylation) quartiles, respectively. Failed analyses are indicated in white. The two major sample clusters are shown in green (cluster 1) and red (cluster 2). Hormone receptor (HR) status is indicated by black ovals (HR−) or white ovals (HR+).

Close modal
Fig. 2.

A, ESR1 gene expression and ESR1 DNA methylation. ESR1 gene expression from exon 1A was measured by real-time quantitative reverse transcription-PCR. The resulting values were divided by the expression values for the control gene TATA box-binding protein and normalized to a mean ratio of 1.0 across all 74 samples. Methylation levels were measured immediately downstream of the transcription start site of exon 1A (amplicon located from +14 to +114) by MethyLight, and the resulting percentage of fully methylated reference values was divided into quartiles. In the two panels on the right, the four quartiles were further subdivided into progesterone receptor-positive and progesterone receptor-negative tumors.

Fig. 2.

A, ESR1 gene expression and ESR1 DNA methylation. ESR1 gene expression from exon 1A was measured by real-time quantitative reverse transcription-PCR. The resulting values were divided by the expression values for the control gene TATA box-binding protein and normalized to a mean ratio of 1.0 across all 74 samples. Methylation levels were measured immediately downstream of the transcription start site of exon 1A (amplicon located from +14 to +114) by MethyLight, and the resulting percentage of fully methylated reference values was divided into quartiles. In the two panels on the right, the four quartiles were further subdivided into progesterone receptor-positive and progesterone receptor-negative tumors.

Close modal
Fig. 3.

Survival curves. Kaplan-Meier survival curves are shown for the three genes that were statistically significant predictors of response to tamoxifen therapy in a Cox proportional hazards interaction model (Table 2). Above median methylation levels are indicated by HIGH mC for ESR1 and ARHI, whereas below median methylation levels are indicated by LOW mC for these two genes. For CYP1B1 methylation, POS mC refers to any detectable methylation, whereas NEG mC refers to a lack of detectable methylation. The high or positive methylation survival curves are shown in red, with squares indicating censored events. The low or negative methylation survival curves are shown in green, with circles indicating censored events.

Fig. 3.

Survival curves. Kaplan-Meier survival curves are shown for the three genes that were statistically significant predictors of response to tamoxifen therapy in a Cox proportional hazards interaction model (Table 2). Above median methylation levels are indicated by HIGH mC for ESR1 and ARHI, whereas below median methylation levels are indicated by LOW mC for these two genes. For CYP1B1 methylation, POS mC refers to any detectable methylation, whereas NEG mC refers to a lack of detectable methylation. The high or positive methylation survival curves are shown in red, with squares indicating censored events. The low or negative methylation survival curves are shown in green, with circles indicating censored events.

Close modal
Table 1

DNA methylation markers as predictors of HRa status

The 35 DNA methylation markers are ranked according to the strength of their association with HR as determined by logistic regression. HR status (HR, ER, and PR, in which HR refers to ER and/or PR positive) was used as the outcome, and quartiled DNA methylation (PMR) values were used as predictors, with adjustment for age. Significant associations after adjustment for multiple comparisons are indicated in bold. In the ER analysis, the data were adjusted for PR status, whereas in the PR analysis, the data were adjusted for ER status. “Association” refers to the direction of the association (+ indicates a positive relationship between DNA methylation and HR status, and − indicates an inverse relationship).

HR status predictorsER status predictorsPR status predictors
GeneAssociationPGeneAssociationPGeneAssociationP
SOCS1 0.0001 PGR − 0.0010 ESR1 0.0118 
RASSF1A 0.0002 TFF1 − 0.0035 TGFBR2 0.0218 
BCL2 − 0.0009 CDH13 − 0.0043 PTGS2 0.0295 
PGR − 0.0022 TIMP3 − 0.0063 CDH13 0.0326 
TGFBR2 0.0023 HSD17B4 − 0.0110 SOCS1 0.0581 
GSTP1 0.0033 ESR1 − 0.0330 TFF1 0.0626 
PTGS2 0.0037 BCL2 − 0.0496 GSTP1 0.1238 
HSD17B4 − 0.0053 APC 0.0552 TWIST 0.1287 
ARHI 0.0085 CDH1 − 0.0653 RASSF1A 0.1491 
APC 0.0151 TERT − 0.0671 PGR 0.1763 
TIMP3 − 0.0177 MCJ 0.0772 CDH1 0.1948 
TWIST 0.0224 RASSF1A 0.1223 MLH1 0.2000 
MLH1 0.0369 SOCS1 0.1463 ESR2 0.2685 
ESR2 0.0414 CDKN2A − 0.1748 ARHI 0.2954 
TNFRSF12 0.0452 RNR1 0.2322 HRAS − 0.2962 
TFF1 − 0.0564 ARHI 0.2470 CCND2 0.2976 
SYK − 0.0812 ABCB1 0.3798 HSD17B4 0.3781 
CDH13 − 0.0853 TGFBR2 0.4330 TNFRSF12 0.4491 
RNR1 0.1088 CALCA − 0.4691 TERT 0.4558 
CDKN2A − 0.1282 CCND2 − 0.4915 TIMP3 0.4590 
MCJ 0.1324 MGMT − 0.5157 DAPK1 0.4996 
TERT − 0.1465 HRAS 0.5212 MCJ − 0.5811 
MGMT − 0.2350 GSTP1 0.5219 BCL2 − 0.5980 
CDH1 − 0.2674 SYK − 0.5333 THBS1 − 0.6297 
HRAS − 0.3155 TNFRSF12 0.5517 ABCB1 − 0.6560 
CALCA − 0.3796 BRCA1 − 0.5741 SYK − 0.7178 
CYP1B1 − 0.3954 MLH1 0.6753 TYMS − 0.7404 
THBS1 − 0.4210 ESR2 0.7660 CYP1B1 − 0.7750 
BRCA1 − 0.5164 TYMS 0.8311 CALCA − 0.8015 
DAPK1 0.5700 PTGS2 0.8563 MGMT − 0.8400 
CCND2 0.7342 CYP1B1 − 0.8907 CDKN2A 0.8522 
TYMS 0.7771 DAPK1 − 0.9232 BRCA1 0.8717 
ABCB1 0.7861 THBS1 − 0.9411 APC − 0.9011 
MYOD1 0.8096 MYOD1 0.9441 RNR1 0.9213 
ESR1 − 0.8737 TWIST − 0.9671 MYOD1 − 0.9751 
HR status predictorsER status predictorsPR status predictors
GeneAssociationPGeneAssociationPGeneAssociationP
SOCS1 0.0001 PGR − 0.0010 ESR1 0.0118 
RASSF1A 0.0002 TFF1 − 0.0035 TGFBR2 0.0218 
BCL2 − 0.0009 CDH13 − 0.0043 PTGS2 0.0295 
PGR − 0.0022 TIMP3 − 0.0063 CDH13 0.0326 
TGFBR2 0.0023 HSD17B4 − 0.0110 SOCS1 0.0581 
GSTP1 0.0033 ESR1 − 0.0330 TFF1 0.0626 
PTGS2 0.0037 BCL2 − 0.0496 GSTP1 0.1238 
HSD17B4 − 0.0053 APC 0.0552 TWIST 0.1287 
ARHI 0.0085 CDH1 − 0.0653 RASSF1A 0.1491 
APC 0.0151 TERT − 0.0671 PGR 0.1763 
TIMP3 − 0.0177 MCJ 0.0772 CDH1 0.1948 
TWIST 0.0224 RASSF1A 0.1223 MLH1 0.2000 
MLH1 0.0369 SOCS1 0.1463 ESR2 0.2685 
ESR2 0.0414 CDKN2A − 0.1748 ARHI 0.2954 
TNFRSF12 0.0452 RNR1 0.2322 HRAS − 0.2962 
TFF1 − 0.0564 ARHI 0.2470 CCND2 0.2976 
SYK − 0.0812 ABCB1 0.3798 HSD17B4 0.3781 
CDH13 − 0.0853 TGFBR2 0.4330 TNFRSF12 0.4491 
RNR1 0.1088 CALCA − 0.4691 TERT 0.4558 
CDKN2A − 0.1282 CCND2 − 0.4915 TIMP3 0.4590 
MCJ 0.1324 MGMT − 0.5157 DAPK1 0.4996 
TERT − 0.1465 HRAS 0.5212 MCJ − 0.5811 
MGMT − 0.2350 GSTP1 0.5219 BCL2 − 0.5980 
CDH1 − 0.2674 SYK − 0.5333 THBS1 − 0.6297 
HRAS − 0.3155 TNFRSF12 0.5517 ABCB1 − 0.6560 
CALCA − 0.3796 BRCA1 − 0.5741 SYK − 0.7178 
CYP1B1 − 0.3954 MLH1 0.6753 TYMS − 0.7404 
THBS1 − 0.4210 ESR2 0.7660 CYP1B1 − 0.7750 
BRCA1 − 0.5164 TYMS 0.8311 CALCA − 0.8015 
DAPK1 0.5700 PTGS2 0.8563 MGMT − 0.8400 
CCND2 0.7342 CYP1B1 − 0.8907 CDKN2A 0.8522 
TYMS 0.7771 DAPK1 − 0.9232 BRCA1 0.8717 
ABCB1 0.7861 THBS1 − 0.9411 APC − 0.9011 
MYOD1 0.8096 MYOD1 0.9441 RNR1 0.9213 
ESR1 − 0.8737 TWIST − 0.9671 MYOD1 − 0.9751 
a

HR, hormone receptor; ER, estrogen receptor; PR, progesterone receptor; PMR, percentage of fully methylated reference.

Table 2

Cox regression analysis of the association between HRa status and overall survival or disease-free survival in patients treated with TAM or not treated with TAM

TAMNot TAMInteraction
Hazard ratiob95% CIHazard ratiob95% CIP
Survival      
 HR 0.7 0.3–1.8 1.0 0.4–2.7 0.54 
 ER 0.8 0.3–2.1 0.9 0.3–2.4 0.87 
 PR 0.5 0.2–1.1 1.1 0.4–2.8 0.22 
Disease-free survival      
 HR 0.6 0.2–1.3 1.1 0.4–2.7 0.31 
 ER 0.7 0.3–1.6 0.8 0.3–2.2 0.71 
 PR 0.4 0.2–0.9 0.9 0.4–2.3 0.19 
TAMNot TAMInteraction
Hazard ratiob95% CIHazard ratiob95% CIP
Survival      
 HR 0.7 0.3–1.8 1.0 0.4–2.7 0.54 
 ER 0.8 0.3–2.1 0.9 0.3–2.4 0.87 
 PR 0.5 0.2–1.1 1.1 0.4–2.8 0.22 
Disease-free survival      
 HR 0.6 0.2–1.3 1.1 0.4–2.7 0.31 
 ER 0.7 0.3–1.6 0.8 0.3–2.2 0.71 
 PR 0.4 0.2–0.9 0.9 0.4–2.3 0.19 
a

HR, hormone receptor; TAM, tamoxifen; Not TAM, not treated with TAM; CI, confidence interval; ER, estrogen receptor; PR, progesterone receptor.

b

Adjusted for age, stage (I, II, and III/IV) and nodes (0, 1–3, and >3). Age and stage are coded as continuous variables.

Table 3

Cox regression analysis of the association between DNA methylation PMRa values (grouped by quartiles) and overall survival or disease-free survival

Significant associations are indicated in bold.

TAMNot TAMInteraction
Hazard ratiob95% CIHazard ratiob95% CIP
Survival      
ESR1 0.7 0.5–1.0 1.5 1.0–2.4 0.0073 
ARHI 1.2 0.9–1.7 0.6 0.3–0.9 0.0103 
CYP1B1 0.7 0.5–1.0 1.5 1.1–2.2 0.0046 
Disease-free survival      
ESR1 0.7 0.5–1.0 1.5 0.9–2.3 0.0134 
ARHI 1.2 0.9–1.6 0.5 0.3–0.8 0.0015 
CYP1B1 0.8 0.6–1.1 1.5 1.1–2.1 0.0081 
TAMNot TAMInteraction
Hazard ratiob95% CIHazard ratiob95% CIP
Survival      
ESR1 0.7 0.5–1.0 1.5 1.0–2.4 0.0073 
ARHI 1.2 0.9–1.7 0.6 0.3–0.9 0.0103 
CYP1B1 0.7 0.5–1.0 1.5 1.1–2.2 0.0046 
Disease-free survival      
ESR1 0.7 0.5–1.0 1.5 0.9–2.3 0.0134 
ARHI 1.2 0.9–1.6 0.5 0.3–0.8 0.0015 
CYP1B1 0.8 0.6–1.1 1.5 1.1–2.1 0.0081 
a

PMR, percentage of fully methylated reference; TAM, tamoxifen; Not TAM, not treated with tamoxifen; CI, confidence interval.

b

Covariates include age, stage (I, II, and III/IV), nodes (0, 1–3, and >3), hormone receptor status among TAM-treated patients (1, hormone receptor positive and treated with TAM, 0, otherwise), and hormone receptor status among those not treated with TAM (1, hormone receptor positive and not treated with TAM; 0, otherwise). Age and stage are coded as continuous variables.

We thank Drs. Mihaela Velicescu and Daniel J. Weisenberger for critical evaluation of the manuscript. We are grateful to Tiffany I. Long for technical advice and assistance.

1
Feuer EJ, Wun LM, Boring CC, et al The lifetime risk of developing breast cancer.
J Natl Cancer Inst (Bethesda)
,
85
:
892
-7,  
1993
.
2
Potter JD, Cerhan JR, Sellers TA, et al Progesterone and estrogen receptors and mammary neoplasia in the Iowa Women’s Health Study: how many kinds of breast cancer are there?.
Cancer Epidemiol Biomark Prev
,
4
:
319
-26,  
1995
.
3
Anonymous. Tamoxifen for early breast cancer: an overview of the randomised trials. Early Breast Cancer Trialists’ Collaborative Group.
Lancet
,
351
:
1451
-67,  
1998
.
4
Bardou VJ, Arpino G, Elledge RM, Osborne CK, Clark GM. Progesterone receptor status significantly improves outcome prediction over estrogen receptor status alone for adjuvant endocrine therapy in two large breast cancer databases.
J Clin Oncol
,
21
:
1973
-9,  
2003
.
5
Perou CM, Sorlie T, Eisen MB, et al Molecular portraits of human breast tumours.
Nature (Lond)
,
406
:
747
-52,  
2000
.
6
Sorlie T, Perou CM, Tibshirani R, et al Gene expression patterns of breast carcinomas distinguish tumor subclasses with clinical implications.
Proc Natl Acad Sci USA
,
98
:
10869
-74,  
2001
.
7
van’t Veer LJ, Dai H, van de Vijver MJ, et al Gene expression profiling predicts clinical outcome of breast cancer.
Nature (Lond)
,
415
:
530
-6,  
2002
.
8
van de Vijver MJ, He YD, van’t Veer LJ, et al A gene-expression signature as a predictor of survival in breast cancer.
N Engl J Med
,
347
:
1999
-2009,  
2002
.
9
Sorlie T, Tibshirani R, Parker J, et al Repeated observation of breast tumor subtypes in independent gene expression data sets.
Proc Natl Acad Sci USA
,
100
:
8418
-23,  
2003
.
10
Laird PW. Early detection: the power and the promise of DNA methylation markers.
Nat Rev Cancer
,
3
:
253
-66,  
2003
.
11
Yan PS, Chen CM, Shi H, et al Dissecting complex epigenetic alterations in breast cancer using CpG island microarrays.
Cancer Res
,
61
:
8375
-80,  
2001
.
12
Yang X, Yan L, Davidson NE. DNA methylation in breast cancer.
Endocr Relat Cancer
,
8
:
115
-27,  
2001
.
13
Widschwendter M, Jones PA. DNA methylation and breast carcinogenesis.
Oncogene
,
21
:
5462
-82,  
2002
.
14
Jones PA, Baylin SB. The fundamental role of epigenetic events in cancer.
Nat Rev Genet
,
3
:
415
-28,  
2002
.
15
Adorjan P, Distler J, Lipscher E, et al Tumour class prediction and discovery by microarray-based DNA methylation analysis.
Nucleic Acids Res
,
30
:
E21
2002
.
16
Eads CA, Danenberg KD, Kawakami K, et al MethyLight: a high-throughput assay to measure DNA methylation.
Nucleic Acids Res
,
28
:
E32
2000
.
17
Powles TJ. Anti-oestrogenic prevention of breast cancer: the make or break point.
Nat Rev Cancer
,
2
:
787
-94,  
2002
.
18
Jordan VC. Tamoxifen: a most unlikely pioneering medicine.
Nat Rev Drug Discov
,
2
:
205
-13,  
2003
.
19
Eads CA, Danenberg KD, Kawakami K, et al CpG island hypermethylation in human colorectal tumors is not associated with DNA methyltransferase overexpression.
Cancer Res
,
59
:
2302
-6,  
1999
.
20
Eads CA, Lord RV, Wickramasinghe K, et al Epigenetic patterns in the progression of esophageal adenocarcinoma.
Cancer Res
,
61
:
3410
-8,  
2001
.
21
Kaufman L, Rousseeuw PJ. .
Finding groups in data: an introduction to cluster analysis
, Wiley Interscience New York  
1990
.
22
Slamon DJ, Clark GM, Wong SG, et al Human breast cancer: correlation of relapse and survival with amplification of the HER-2/neu oncogene.
Science (Wash DC)
,
235
:
177
-82,  
1987
.
23
Lindeman GJ, Wittlin S, Lada H, et al SOCS1 deficiency results in accelerated mammary gland development and rescues lactation in prolactin receptor-deficient mice.
Genes Dev
,
15
:
1631
-6,  
2001
.
24
Yoshikawa H, Matsubara K, Qian GS, et al SOCS-1, a negative regulator of the JAK/STAT pathway, is silenced by methylation in human hepatocellular carcinoma and shows growth- suppression activity.
Nat Genet
,
28
:
29
-35,  
2001
.
25
Dammann R, Yang G, Pfeifer GP. Hypermethylation of the cpG island of Ras association domain family 1A (RASSF1A), a putative tumor suppressor gene from the 3p21.3 locus, occurs in a large percentage of human breast cancers.
Cancer Res
,
61
:
3105
-9,  
2001
.
26
Lehmann U, Langer F, Feist H, et al Quantitative assessment of promoter hypermethylation during breast cancer development.
Am J Pathol
,
160
:
605
-12,  
2002
.
27
Castiglione F, Sarotto I, Fontana V, et al Bcl2, p53 and clinical outcome in a series of 138 operable breast cancer patients.
Anticancer Res
,
19
:
4555
-63,  
1999
.
28
Park SH, Kim H, Song BJ. Down regulation of bcl2 expression in invasive ductal carcinomas is both estrogen- and progesterone-receptor dependent and associated with poor prognostic factors.
Pathol Oncol Res
,
8
:
26
-30,  
2002
.
29
Savouret JF, Bailly A, Misrahi M, et al Characterization of the hormone responsive element involved in the regulation of the progesterone receptor gene.
EMBO J
,
10
:
1875
-83,  
1991
.
30
Clark GM, Osborne CK, McGuire WL. Correlations between estrogen receptor, progesterone receptor, and patient characteristics in human breast cancer.
J Clin Oncol
,
2
:
1102
-9,  
1984
.
31
Nardulli AM, Greene GL, O’Malley BW, Katzenellenbogen BS. Regulation of progesterone receptor messenger ribonucleic acid and protein levels in MCF-7 cells by estradiol: analysis of estrogen’s effect on progesterone receptor synthesis and degradation.
Endocrinology
,
122
:
935
-44,  
1988
.
32
Yang X, Phillips DL, Ferguson AT, et al Synergistic activation of functional estrogen receptor (ER)-alpha by DNA methyltransferase and histone deacetylase inhibition in human ER-alpha-negative breast cancer cells.
Cancer Res
,
61
:
7025
-9,  
2001
.
33
Kos M, Reid G, Denger S, Gannon F. Minireview: genomic organization of the human ERalpha gene promoter region.
Mol Endocrinol
,
15
:
2057
-63,  
2001
.
34
Jones PA. The DNA methylation paradox.
Trends Genet
,
15
:
34
-7,  
1999
.
35
Fox MS. On the diagnosis and treatment of breast cancer.
JAMA
,
241
:
489
-94,  
1979
.
36
Luo RZ, Fang X, Marquez R, et al ARHI is a Ras-related small G-protein with a novel N-terminal extension that inhibits growth of ovarian and breast cancers.
Oncogene
,
22
:
2897
-909,  
2003
.
37
Clarke R, Leonessa F, Welch JN, Skaar TC. Cellular and molecular pharmacology of antiestrogen action and resistance.
Pharmacol Rev
,
53
:
25
-71,  
2001
.
38
Gupta M, McDougal A, Safe S. Estrogenic and antiestrogenic activities of 16alpha- and 2-hydroxy metabolites of 17beta-estradiol in MCF-7 and T47D human breast cancer cells.
J Steroid Biochem Mol Biol
,
67
:
413
-9,  
1998
.
39
Crewe HK, Notley LM, Wunsch RM, Lennard MS, Gillam EM. Metabolism of tamoxifen by recombinant human cytochrome P450 enzymes: formation of the 4-hydroxy, 4′-hydroxy and N-desmethyl metabolites and isomerization of trans-4-hydroxytamoxifen.
Drug Metab Dispos
,
30
:
869
-74,  
2002
.

Supplementary data