Factors affecting the persistence of drug-induced reprogramming of the cancer methylome

ABSTRACT

Aberrant DNA methylation is a critical feature of cancer. Epigenetic therapy seeks to reverse these changes to restore normal gene expression. DNA demethylating agents, including 5-aza-2′-deoxycytidine (DAC), are currently used to treat certain leukemias, and can sensitize solid tumors to chemotherapy and immunotherapy. However, it has been difficult to pin the clinical efficacy of these agents to specific demethylation events, and the factors that contribute to the durability of response remain largely unknown. Here we examined the genome-wide kinetics of DAC-induced DNA demethylation and subsequent remethylation after drug withdrawal in breast cancer cells. We find that CpGs differ in both their susceptibility to demethylation and propensity for remethylation after drug removal. DAC-induced demethylation was most apparent at CpGs with higher initial methylation levels and further from CpG islands. Once demethylated, such sites exhibited varied remethylation potentials. The most rapidly remethylating CpGs regained >75% of their starting methylation within a month of drug withdrawal. These sites had higher pretreatment methylation levels, were enriched in gene bodies, marked by H3K36me3, and tended to be methylated in normal breast cells. In contrast, a more resistant class of CpG sites failed to regain even 20% of their initial methylation after 3 months. These sites had lower pretreatment methylation levels, were within or near CpG islands, marked by H3K79me2 or H3K4me2/3, and were overrepresented in sites that become aberrantly hypermethylated in breast cancers. Thus, whereas DAC-induced demethylation affects both endogenous and aberrantly methylated sites, tumor-specific hypermethylation is more slowly regained, even as normal methylation promptly recovers. Taken together, these data suggest that the durability of DAC response is linked to its selective ability to stably reset at least a portion of the cancer methylome.

Introduction

Although cancer has historically been understood as a disease resulting from genetic mutation, it is now clear that epigenetic alterations are also of critical importance in carcinogenesis. In contrast to genetic mutations, epigenetic changes are potentially reversible, making them provocative therapeutic targets. Among epigenetic therapies, DNA methyltransferase inhibitors in particular have shown clinical activity; 5-aza-2'-deoxycytidine (DAC) and 5-azacytidine (AZA) are used to treat myelodysplastic syndrome as well as advanced acute myeloid leukemia and chronic myeloid monocytic leukemia.¹ These drugs have also exhibited activity in solid tumors of the lung in combination with other therapies, including HDAC inhibitors,² and are in clinical trials for the treatment of breast³ and ovarian cancers.^2,4-6

In spite of the considerable use and study of these drugs, the molecular basis of their efficacy remains incompletely understood. As cytidine analogs, DAC and AZA require intracellular conversion to the triphosphate form and incorporation into DNA, where they covalently trap DNA methyltransferase 1 (DNMT1) and lead to its proteasomal degradation.⁷ Successive rounds of DNA replication in the absence of DNMT1 result in global hypomethylation. However, the means by which this hypomethylation affects gene expression programs and suppresses tumorigenic potential remain unclear. Although tumor-suppressor genes are frequently silenced in cancer through hypermethylation of promoter-associated CpG Islands (CGI),⁸ and DAC is capable of reversing this aberrant methylation to achieve gene reactivation, less than 10% of DAC-induced transcriptional changes are accounted for by relief of promoter hypermethylation.^9,10

Cell culture studies have demonstrated that at low doses, transient exposure to DAC results in a loss of tumorigenic potential that persists for many generations following drug withdrawal,¹¹ suggesting that durable methylation changes underlie its antitumor activity. Leukemia cells from patients undergoing DAC treatment experience rapid remethylation at the end of each treatment cycle, as measured globally,¹² and at LINE elements.¹³ Intriguingly, neither the initial genomic methylation level,¹⁴ nor the amount of global demethylation observed upon treatment predicts patient response,¹³ suggesting that specific, persistent demethylation events may be responsible for outcome.

Our laboratory and others have documented that while some genes remain stably demethylated after drug withdrawal, others rapidly regain their original methylation state after treatment ends.^10,15-17 These studies implicated several factors in the propensity of given CpG sites to remethylate, including proximity to a transcription start site (TSS), occupancy of RNA Polymerase II (RNAP II), and the presence of certain histone modifications.^10,16 Given the widespread use of these agents and the promising clinical trials underway utilizing DAC or related compounds as a synergistic or sensitizing agents,^4-6 it is vital to understand both the extent and stability of DAC-induced demethylation given its immediate potential to improve cancer treatment.

In this study, we sought to identify the locus-specific determinants underlying the genome-wide susceptibility to DAC-induced demethylation and subsequent remethylation after drug withdrawal. We unearth extensive variation in the sensitivity of CpG sites to demethylation as well as in inclination to remethylation following drug withdrawal. We find that DAC affects more highly methylated CpGs, and that CGI-associated CpGs tend to be resistant to demethylation, in keeping with their generally low methylation levels, but those that do undergo demethylation tend to be slower to remethylate upon drug removal. Moreover, chromatin modifications appear to be predictive of remethylation rate, with chromatin features associated with active promoters, strong enhancers, and elongation rate (H3K79me2) correlated with resistance to remethylation, while H3K36me3 predisposed associated sites to rapid remethylation. Strikingly, we find that the more methylated a CpG site is in normal breast tissue, the quicker it is to regain that methylation after drug removal, whereas CpGs exhibiting cancer-specific hypermethylation, once demethylated, are slower to recover. These data suggest that DAC treatment stably resets at least a portion of the cancer epigenome to its original state, and substantiates the continued interest in epigenetic therapy as cancer treatment.

Results

Characterization of DAC-induced DNA demethylation

To model the dynamics of DNA demethylation and remethylation in response to DNA methyltransferase inhibitors, as it might occur in a solid tumor setting, we treated MDA-MB-231 triple-negative breast cancer cells with 500 nM DAC for 6 d to induce DNA demethylation, and then tracked DNA remethylation kinetics for over 27,000 CpGs over 27 passages (∼3 months) in the absence of drug using Illumina Infinium DNA methylation arrays. DAC-induced demethylation and remethylation kinetics were established for 3 independent time-course experiments. Treatment of MDA-MB-231 cells with DAC elicited a >25% decrease in genome-wide DNA methylation as estimated from the average β value (Fig. 1A). Hierarchical clustering of CpG methylation values segregated untreated and DAC-treated samples by overall methylation levels (Fig. 1B, P < 0.01).¹⁸ Examination of the distribution of DNA methylation levels across CpG sites in treated vs. untreated samples revealed that DAC treatment elicited a systematic shift toward lower methylation levels and a decrease in the frequency of highly methylated CpGs (β > 0.7), complemented with sites of moderate gain in methylation frequency (β = 0.4–0.6) and unmethylated sites (β < 0.2), consistent with global demethylation (Fig. 1C).

Figure 1.

DAC induces global demethylation. A) Average DNA methylation levels after 6 d DAC treatment. Each circle represents a biological replicate, thick horizontal bars represent the means of all replicates, and error bars are the standard deviation. *, P<0.05, Welch's t-test. B) Hierarchical clustering of 3 untreated and 3 treated replicate samples across each assayed CpG site. C) Density plot of DNA methylation levels (β values) across all 27K sites in treated and untreated samples at time zero. D,E) A linear mixed effects model (CpG Assoc) was applied to determine CpG sites that were significantly differentially methylated between untreated samples and immediately following treatment with DAC. A change of at least 0.2 in β and an FDR<0.05 was then imposed to focus on those sites with the greatest biological significance (green). Using these parameters, a total of 5,344 CpGs were significantly demethylated and none were significantly hypermethylated. F) Density plot of DNA methylation levels (β values) across the 5,316 CpG sites found to be significantly demethylated in untreated and post-treatment samples at time zero.

To identify loci that were differentially methylated after DAC treatment, we used a linear fixed-effects model (see methods). This analysis identified 5,316 loci that were significantly changed upon DAC treatment, all of which were hypomethylated [False Discovery Rate (FDR) ≤ 0.05, Δβ ≥ 0.2] (Fig. 1D,E). Analysis of the distribution of pre- and post-treatment DNA methylation levels for these significantly affected CpGs indicated that DAC tended to affect CpGs that had higher initial levels of DNA methylation, with more than 73% of affected sites having a starting methylation level of >0.7 (Fig. 1F). These data demonstrate that DAC induced an overall genome-wide loss in DNA methylation.

Promoters and CpG Islands undergo less demethylation than adjacent genomic regions

Next we determined the spatial relationship between the DAC-affected sites and other genomic features. The position of CpG sites subject to DAC-induced hypomethylation were plotted by their distance to the nearest transcription start site (TSS) and their scaled distance within or to a CGI (Fig. 2A,B). Consistent with the tendency for a higher initial DNA methylation level, demethylated CpGs tended to be more distal, and were enriched upstream of the TSS and in the CGI shores as compared to all CpGs represented on the array (P < 2.2E-16, Chi-squared). As expected, given the relatively unmethylated status of CGI, CpGs undergoing DAC-induced demethylation were underrepresented in CGI.

Figure 2.

Genomic features distinguishing CpG sites significantly demethylated by DAC treatment. A) Distribution of CpGs relative to nearest RefSeq TSS. Demethylated CpGs (N = 5316) (red) were plotted by the distance to the nearest TSS oriented to the direction of transcription relative to all probes on the 27K array as a whole (blue). B) Distribution of CGI-associated CpGs (defined as those with a UCSC CpG island (CGI) within 2 kb of the nearest RefSeq TSS to the CpG) that underwent DAC-induced demethylation (red) were plotted relative to CpGs on the array as a whole (blue) based on their scaled position within the CGI or their absolute distance to the 5′ or 3′ edge of the CGI oriented to the direction of transcription. C-H) ChIP-seq data for the indicated histone modification from HMEC cells (ENCODE) was used determine the relationship between chromatin state and demethylation potential. Average normalized tag densities for the indicated histone modification for demethylated CpG sites (red) or all assayed sites on the 27K array (blue) were compiled in 20 bp bins for 5,000bp centered on the CpG.

To determine what chromatin features might influence DAC-induced demethylation, we examined the density of various histone modifications in and around the demethylation-prone CpG sites, utilizing Chip-seq data from human mammary epithelial cells (HMEC) from the ENCODE project.¹⁹ We found that CpGs that underwent significant demethylation were far less likely to be marked by active modifications (H3K4me2/3, H3K9Ac, H3K27Ac, and H2AZ) in normal cells than the other CpG sites (Fig. 2C). This finding is consistent with the fact that these marks are enriched near active promoters, which also typically exhibit low levels of DNA methylation. In contrast, CpGs subject to DAC-induced demethylation were more enriched than other analyzed sites in H3K27me3, a repressive mark deposited by the Polycomb complex. This may be a reflection of the propensity for H3K27me3-marked CpGs in normal cells to acquire DNA methylation in cancer cells²⁰ and, thus, a high initial methylation level of these sites in general in the MDA-MB-231 cells, or a propensity toward loss of DNA methylation at these sites upon treatment.

Kinetics of CpG remethylation following DAC withdrawal

The above data suggest that both the initial methylation level and the normal underlying chromatin features influence DAC-induced demethylation. We next sought to determine whether, and to what degree, CpG sites differ in their propensity for remethylation after drug withdrawal. To account for differences in initial methylation at each CpG site and amount of methylation lost, we normalized CpGs by their initial and post-treatment DNA methylation levels, such that the degree of demethylation achieved after 6 d of treatment was 100%. DNA methylation changes at later time points (3, 6, 9, and 27 passages) were used to calculate the fractional recovery of methylation over time in the absence of drug. A self-organizing map (SOM)²¹ approach was then used to cluster the CpGs into classes based on their patterns of remethylation.

We found that 4 SOM classes clearly discriminated CpG sites into classes with different remethylation kinetics. We termed these 4 classes Resistant (N = 1,565), Slow (N = 1,991), Moderate (N = 1,244), and Rapid (N = 516) based on their remethylation rates (Fig. 3A). Resistant CpGs represented those relatively resilient to remethylation; most failed to regain even 15% of their lost methylation after 27 passages. Slow CpGs exhibited a sluggish remethylation rate, and on average regained only 30% of their lost methylation, while Moderate CpGs regained approximately half of their initial methylation by 9 passages (∼1 month). The majority of demethylated CpGs belonged to the Resistant and Slow classes, which together encompassed more than two-thirds of demethylated sites (Fig. 3B). Most striking were Rapid CpGs, which regained more than 70% of their lost methylation within just 9 passages, and almost 90% by 27 passages (∼3 months). Comparatively, no other class regained more than 65% of lost methylation even following 27 passages. To validate these findings, we assessed the remethylation kinetics of select CpG sites and genes by COmbined Bisulfite Restriction Analysis (COBRA), a quantitative approach that takes advantage of the methylation-sensitive creation or destruction of restriction enzyme sites after bisulfite conversion to assess DNA methylation.²² In general, there was excellent correlation in remethylation kinetics of CpGs as determined on the 27K array and CpGs in the same genomic vicinity measured by COBRA (Supplemental Fig. 1). Gene ontology and gene set enrichment analysis suggested that whereas the genes associated with Rapidly remethylating CpGs tended toward signal-responsive genes and those involved in immune regulation, those associated with the Resistant class tended toward developmental regulators of cell fate, tissue-specific differentiation, and targets of Polycomb repressive complex-2 (PRC2)-mediated repression (Supplemental Data).

Figure 3.

CpG sites demethylated in response to DAC differ in their remethylation kinetics. A) Self-Organizing Maps were used to group significantly demethylated CpGs (N = 5316) into 4 classes based on their kinetics of remethylation. Fractional recovery of DNA methylation after DAC-induced demethyation for CpGs in each remethylation class; Resistant = 1565; Slow = 1991; Moderate = 1244; Rapid = 516. Plotted is the median (solid line) and first and third quartiles (shadows) across 3 biological replicates. B) Relative distribution of DAC-demethylated CpGs in each remethylation class. C) Remethylation kinetics of the same CpGs in an independent experiment in which HCT116 colon cancer cells were exposed to 300 μM DAC and allowed to recover in the absence of drug for 42 d (Data from Yang et al.¹⁰).

The above data demonstrate the considerable variation in remethylation potential and dynamics among CpGs following DAC-induced demethylation. Recent work by Yang et al.¹⁰ tracked DNA remethylation kinetics following DAC treatment of HCT116 colon cancer cells. An analysis of the remethylation kinetics of the CpGs in our 4 classes in that experiment showed that these CpGs behave similarly (Fig. 3C) in another cell type. These data suggest that the susceptibility of individual CpG sites to demethylation and subsequent remethylation correlates with intrinsic genomic and epigenomic features that are consistent across tissue types.

CpG Islands are resistant to remethylation; shores and gene bodies are prone to remethylation

We next sought to elucidate the factors governing the differences in remethylation rates among CpG sites. First, we addressed the hypothesis that the initial methylation state of CpG sites would correlate with their remethylation rate. Indeed, CpGs in the Rapid class tend to have higher initial methylation levels than Resistant CpGs (Fig. 4A, P < 2.2E-16, Mann-Whitney U). However, there was no difference in the initial DNA methylation levels among the Moderate and Rapid loci (mean β = 0.82 for each class, P = 0.55, Mann-Whitney U) suggesting that there are additional factors that govern remethylation rate. Next, we investigated if these methylation differences were linked to differences in proximity to CGI, given the generally low methylation levels of CGI. First, we defined the subset of CpGs in each class associated with genes whose promoters contained CGI (presence of a UCSC-defined CGI within 2 kb of the nearest RefSeq TSS to the CpG site, N = 3,029). Overall, CpGs in CGI-associated genes represented a greater proportion of the Resistant class (67.9%, Odds Ratio (OR) = 1.29, P = 4.95E-6, Fisher's exact) and there was a trend toward decreasing representation of such sites as remethylation rate increases (Slow, 56.7%, OR = 0.98, P = 0.77, Moderate, 50.3%, OR = 0.85, P = 0.011, Rapid, 40.9%, OR = 0.69, P = 0.00046). We next classified this subset of DAC-demethylated CpG loci by their position within or distance to the CGI. While ∼30% of Resistant CpGs were located in a CGI, only half that portion (∼15%) of Rapid CpGs were in CGI (OR within CGI: Resistant = 1.20, P = 0.023; Rapid = 0.52, P = 0.0015) (Fig. 4B). Indeed, there was a direct relationship between the proximity to the CGI and the remethylation rate. Whereas the Resistant loci tend to be distributed within or immediately downstream of the CGI (the South shore), the Rapid CpGs were found to be more distally distributed (Fig. 4C). Slow and Moderate CpGs, intermediate in their remethylation rate, were also intermediate in their distances to CGI (Fig. 4C).

Figure 4.

Relationship between genomic features and remethylation potential. A) Boxplots representing the distribution of starting methylation levels of CpGs in each remethylation class relative to all demethylated CpGs. The line represents the median, hinges bound the first and third quartiles, and the whiskers represent the maximum or minimum values within 1.5x the interquartile range. B) Fraction of CpGs in and around CGI. Shown is the fraction in each class lying within or relative to a CGI. Shores are defined as 1 kb from the edge of the CGI, the North shelf as 1–2.5 kb upstream of the CGI, and South Shelf as gene body regions distal to the island (>1 kb downstream of the CGI). C) Density plot of CpGs in each remethylation class relative to CGI. Frequency of CpGs were plotted based on the scaled position within a CGI or by absolute distance from either edge of the nearest CGI oriented to the direction of transcription of the nearest TSS. D, E) Density plot of CGI-associated (D) and non-CGI-associated (E) CpGs relative to the position of the nearest TSS. Frequency of CpGs in each class were plotted based on their position to the nearest TSS and oriented to the direction of transcription. CGI- associated sites are defined here as those with a CGI within 2 kb of the nearest RefSeq TSS.

To determine whether the differences in remethylation kinetics were driven by the CGI itself or its embedded TSS, we plotted the position of each class of CpGs relative to the nearest TSS. There was no correlation between remethylation rate and distance to the nearest TSS (Fig. 4D). Consistent with this finding, there was also no relationship between TSS proximity and remethylation rate among CpGs whose nearest TSS is not CGI-associated (Fig. 4E). Together these data suggest that it is characteristics of the CGI domain itself, rather than the embedded TSS, that are responsible for the differences observed in remethylation rate.

Chromatin features correlate strongly with remethylation potential

The above data suggest that one determinant of remethylation potential is the proximity to CGI, which are characterized by a unique chromatin domain.²³ We therefore examined the relationship between remethylation kinetics and local chromatin environment for CpGs in each remethylation class. We first utilized an annotation of chromatin states derived from a Hidden Markov Model (ChromHMM) that uses ChIP-seq data of 9 chromatin marks from HMECs to partition the genome into discrete functional states (Fig. 5).²⁴ We considered both the fraction of sites in each remethylation class that fell into broad categories of chromatin HMM states, as well as the Odds Ratio (OR) of enrichment of CpGs in individual chromatin states. As a second approach, we directly examined ChIP-Seq data for specific histone marks derived from HMECs¹⁹ to investigate the spatial relationship between the enrichment of a given mark around the CpG sites in each remethylation class (Fig. 6).

Figure 5.

Relationship between chromatin functional states and remethylation potential. A) CpGs were assigned to functional chromatin states based on ChromHMM, a Hidden Markov Model based on 9 chromatin modifications. Shown is the fraction of CpG sites in each remethylation class in the indicated ChromHMM category derived from HMEC. B) Odds ratios of enrichment, relative to all demethylated sites, in each chromatin state (*, P<0.05, Fisher's exact). C) Proportion of enhancer associated-CpG sites that represent HMM-defined strong or weak enhancers (Resistant = 231, Slow = 366, Moderate = 209, Rapid = 75).

Figure 6.

Histone modifications and remethylation potential. Chip-seq tag densities for the indicated histone modifications were used to determine the relationship between histone modifications associated with (A) active promoters, (B) gene bodies/transcribed regions, (C) repressed chromatin, and (D) active enhancers in the regions surrounding CpGs in normal cells and the remethylation potential of the CpG sites. Average normalized tag densities for the indicated histone modification in HMEC cells for CpGs in each remethylation class were compiled in 20 bp bins for 5,000 bp centered on the CpG.

Consistent with the enrichment of Resistant loci in CGI (Fig. 4B), Resistant and Slowly remethylating CpGs were far more likely to be associated with ChromHMM states associated with promoter activity in HMECs (Fig. 5A, B) and to be enriched in histone marks indicative of active promoters, including H3K4me2/3 and H3K9Ac (Fig. 6A), relative to Moderate or Rapidly remethylating loci. In contrast, Moderate and Rapidly remethylating CpGs were depleted in promoter features and display only modest (Moderate) or no (Rapid) enrichment for promoter-associated histone modifications (Fig. 6A). As a class, Resistant sites were especially enriched in strong promoters, whereas Slow CpGs were found within many weak and poised promoters (Fig. 5C), suggesting that among promoter-associated CpGs the relative resistance to remethylation may be correlated with promoter activity.

CpGs classified as Rapidly remethylating were strongly enriched in transcribed areas of the genome (Fig. 5A) and, in particular, those marked with features of transcriptional elongation (Fig. 5B). Indeed, Rapid CpGs are preferentially enriched in H3K36me3 and H4K20me1 in normal cells (Fig. 6B). Rapid CpGs are also notably enriched in the weak transcription/transcriptional transition ChromHMM category (Fig. 5B). Interestingly, the Resistant class is particularly enriched in the gene body mark H3K79me2 (Fig. 6B), the levels of which are correlated with RNAP II elongation rate.^25,26

We next considered the remethylation of CpG sites associated specifically with enhancers, as defined by the ChromHMM classification (Resistant N = 231, Slow N = 366, Moderate N = 209, Rapid N = 75). While a similar fraction of CpGs in each class were annotated to enhancers (Fig. 5A), Resistant CpGs that are enhancer-associated are proportionally overrepresented in strong versus weak enhancers, whereas Rapid CpGs displayed the opposite pattern and were proportionally associated with more weak than strong enhancers (Fig. 5C). Enhancer activity has been correlated with the extent of local histone acetylation, in particular H3K27Ac levels.²⁷ Analysis of this mark demonstrated that it is inversely correlated with remethylation rate (Fig. 6D) with striking enrichment among Resistant CpGs relative to the other classes. Together, these data suggest that among enhancer-associated CpGs, enhancer activity correlates with resilience against remethylation, mimicking the pattern observed at promoters.

Interestingly, we found that chromatin features associated with the Polycomb complex were also inversely correlated with the propensity for remethylation. The Resistant class exhibited enrichment in Polycomb-silenced regions as defined by the ChromHMM state while the Rapid class was strongly depleted in such regions (Fig. 5A, B). This trend was also observed upon analysis of the local enrichment of H3K27me3, the mark deposited by the Polycomb repressive complex, at CpGs of the various classes with greater enrichment of the histone modification in the Resistant class, but little at sites in the Rapid class (Fig. 6C). Consistent with this, gene ontology and gene set enrichment analysis showed that the genes associated with CpGs in the Resistant class were strongly enriched in developmental regulators and targets of Polycomb-mediated repression in embryonic stem cells relative to the other classes (Supplemental Data). We also find that enrichment in H3K9me3, a prominent marker of heterochromatin and gene repression, correlates with resistance to remethylation when considered independently (Fig. 6C).

DAC induces stable reversal of cancer-specific hypermethylation

Previous work has shown that low dose DAC treatment of cancer cells leads to a durable loss of tumorigenic potential for many passages following treatment even when global methylation levels appear largely restored.¹¹ This suggests that there may be some component of the cancer methylome whose demethylation is preferentially reset to a more normal pattern. At the same time, one concern is the potential impact on normal methylation patterns. To address these questions, we defined a set of normally methylated CpGs as those highly methylated in HMEC cells (β > 0.7; N = 2,326) (Resistant = 429, Slow = 794, Moderate = 711, Rapid = 392). We then determined the relative enrichment for these sites in each remethylation class (Fig. 7A). We found that these ‘normally’ methylated sites were much more likely to Rapidly remethylate compared to other sites, indicating that where endogenous methylation is lost during DAC treatment, it is quickly regained. We further defined a set of CpG sites that undergo ‘cancer-specific’ hypermethylation as those that were hypermethylated in MDA-MB-231 cells relative to HMEC cells (change in β ≥ 0.2, MDA-MB-231 cells vs. HMEC). This identified 2,464 CpG sites (Resistant = 944, Slow = 1,011, Moderate = 428, Rapid = 81). Strikingly, the Resistant class was heavily enriched in such sites, whereas the Rapid class was markedly depleted, suggesting that methylation gained during tumorigenesis is less likely to be regained after DAC-induced demethylation (Fig. 7B).

Figure 7.

CpG sites that undergo cancer-specific hypermethylation versus those normally methylated exhibit distinct remethylation kinetics. A) Enrichment of CpG sites that are endogenously methylated in HMEC cells (β > 0.7) in the 4 remethylation classes. Plotted is the odds ratio (Fisher's exact test). P-values for Resistant = 8.58E-25, Slow = 0.0057, Moderate = 6.48E-11, and Rapid = 4.91E-18. B) Enrichment of CpG sites hypermethylated in MDA-MB-231 cells relative to HMEC (change in β > 0.2). Plotted is the OR (Fisher's exact) P-values for Resistant = 1.1E-14, Slow = 0.003, Moderate = 1.13E-9, Rapid = 4.69E-26. C) Methylation status of CpGs in each remethylation class reliably segregate tumor from normal tissues in human breast cancers. The β values for CpGs in each remethylation class were extracted for in 90 matched primary tumor/normal paired breast tissues (TCGA) for the subset of CpGs in each class represented on the Illumina 450K array (Resistant = 1,213, Slow = 1,527, Moderate = 906, Rapid = 342, total = 3,988) and used in a hierarchical clustering analysis (agglomerative, complete linkage). D) Box plot representation of the distribution of methylation levels for CpGs in each remethylation class in normal breast tissues (N = 90, TCGA). The β values for CpGs in each remethylation class were extracted for the subset of CpGs in each class represented on the Illumina 450K array (Resistant = 1213; Slow = 1527; Moderate = 906; Rapid = 342). The line represents the median, hinges bound the first and third quartiles, and the whiskers represent the maximum or minimum values within 1.5x the interquartile range. E) CpGAssoc was used to define CpGs significantly hypermethylated in breast cancer (FDR<0.05, see methods) and whose methylation status is also represented on the 27K array. This identified 440, 595, 328, and 98 CpG sites in the Resistant, Slow, Moderate, and Rapid classes, respectively. Shown are the ORs of enrichment for these hypermethylated sites in each class. Note the underrepresentation of cancer-specific hypermethylation among CpGs in the Rapid class (*, P = 0.028, Fisher's exact).

We next examined the methylation status of CpG sites in each remethylation class among a collection of 90 matched primary breast tumor-normal pairs (N = 180 samples) for whom DNA methylation data was collected as part of the TCGA project.³⁵ We limited our analysis to the subset of our significantly demethylated CpGs common between the platform used here (Illumina 27K array) and that used for the TCGA samples (Illumina 450K array). This resulted in a total of 3,989 CpG sites (Resistant = 1,213, Slow = 1,527, Moderate = 906, Rapid = 342). An unsupervised hierarchical clustering approach showed that the methylation status of CpG sites in each class was independently capable of segregating tumor from normal samples (Resistant, P < 1.25E-37; Slow, P < 8.22E-37; Moderate P < 1.82E-38; Rapid, P < 1.25E-37) and each outperformed 1,000 randomly-selected sets of the same size in this regard (P < 0.0001) indicating that the CpG sites in each remethylation class capture at least a portion of the cancer-specific methylome (Fig. 7C). Intriguingly, an examination of the distribution of DNA methylation levels among CpGs in each class in normal breast tissue samples showed that there was a significant difference in the normal methylation levels among CpGs in the different remethylation classes, with the Rapidly remethylating CpGs having twice the normal DNA methylation levels of Resistant loci (Resistant: median = 0.41, Rapid: median =0.85, P<2.2E-16, Mann-Whitney U) (Fig. 7D). The relationship between remethylation rate and normal tissue methylation levels was even more striking than that observed in untreated MDA-MB-231 cells (compare Fig. 4A and Fig. 7D). These data confirm the above finding that the more methylated a CpG site is in normal tissue, the more likely it is to quickly regain this methylation after DAC-induced demethylation, and further suggest that the methylation state of a CpG in normal tissue is more predictive of its resilience to remethylation than its pretreatment methylation state in cancer cells.

To more directly address the relationship between cancer-specific hypermethylation and remethylation potential, we defined a set of CpGs that were significantly hypermethylated in the breast tumors relative to the matched normal tissues using a linear mixed-effects model approach.²⁸ Among the 25,978 CpG sites common to the 2 analytic platforms (Illumina 27K vs. 450K), a total of 2,356 sites were identified as significantly hypermethylated using this approach, including 440 Resistant, 595 Slow, 328 Moderate and 98 Rapid. Interestingly, we found that CpGs hypermethylated in cancer were depleted in Rapidly remethylating CpGs relative to other classes (Fig. 7C). Together, these data suggest that cancer-specific hypermethylation is more likely to be regained at a slower rate than normal methylation, perhaps contributing to the lasting effect of demethylating agents.

Discussion

DNA methyltransferase inhibitors are currently the standard of care for certain myeloid malignancies and are showing promise alone and in combination with other therapies in the treatment of solid tumors.^29,30 A number of mechanisms have been proposed to account for the antitumor activity of DNA methyltransferase inhibitors, including the reactivation of cancer testes antigens stimulating an immune response,³¹ reactivation of silenced tumor suppressor genes, repression of oncogenes through loss of gene body methylation,¹⁰ and most recently, the activation endogenous retroviruses, leading to the cytoplasmic accumulation of double-stranded RNAs and the triggering of an antiviral interferon response.^32,33 In solid tumor model systems low-dose DAC treatment has been shown to result in prolonged demethylation and a sustained, heritable inhibition of the tumorigenic potential of cancer-initiating/stem-like populations.^10,11,33 Thus, it is vital to understand the site-specific and global factors influencing DNA demethylation and remethylation kinetics as this may provide insight into the mechanisms of drug action and potentially differential responses, which have thus far been difficult to predict.³⁰

Our data indicate that CpGs across the genome differ in both their susceptibility to DAC-induced demethylation as well as their propensity for remethylation after drug removal. We find that CpG islands are not the major targets of DAC-induced demethylation, in line with their low initial methylation, and consistent with the observation that few transcriptional changes in DAC treatment are explained by relief of promoter CGI hypermethylation.⁹ Rather, we find that upstream shore and shelf regions are much more likely to undergo demethylation than CGI. Recent work has shown that much of the cancer-specific variation in DNA methylation patterns occurs in these shore regions.³⁴ Reversal of cancer-specific methylation in these areas in particular may thus play an important role in determining the therapeutic activity.

Chromatin states appear to have profound influence on both DAC-induced demethylation and remethylation rate. Regions that are marked by active chromatin marks (e.g., H3K4me2/3, H3K9/27Ac, and H2AZ) in normal mammary epithelial cells, such as CGI, tend to be resistant to demethylation, and those that do undergo demethylation are resistant to remethylation upon drug removal. Indeed, remethylation-prone CpGs tend to be excluded from CGI. Whether it is transcription itself, RNAP II occupancy, or the chromatin architecture of the CGI that is the primary deterrent is difficult to uncouple, but it is noteworthy that remethylation rate appears to be more tightly linked to the proximity to or presence within the CGI domain than to the TSSs within those CGI, suggesting that it is the CGI structure rather than transcription per se that plays a key role in determining propensity to remethylation. Indeed, CpG site proximity to the TSS of non-CGI associated genes had little impact on remethylation rate. CGIs represent unique chromatin environment in the genome, and are characterized by constrained divergent (bidirectional) transcription, marking by H3K4me3, and high levels of GC-skew, a sequence-based feature associated with the formation of unusual secondary structures.³⁵ Transcription through such regions leads to the formation of R-loops formed by the pairing of a G-rich nascent RNA to its C-rich template behind the progressing polymerase. R-loops have been suggested to play a key role in preventing DNA methylation at CGI.³⁶ Our group has recently shown that GC skew additionally defines key points of RNAP II pausing in CGI promoters,³⁷ regulating its release into elongation. In previous work, we showed that the re-establishment and persistence of RNAP II, even in its paused state impedes remethylation and gene re-silencing after DAC-induced demethylation.¹⁶

A similar trend was observed at enhancers: CpGs associated with strong enhancers and high levels of H3K27Ac in normal cells were found to be more resistant to remethylation. Like CGI, active enhancers are associated with bidirectional transcription, though of non-coding enhancer RNAs (eRNAs), the levels of which are correlated with enhancer activity.³⁸ Transcription or RNAP II occupancy in these regions may also contribute to remethylation resistance. Interestingly, a significant proportion of the gene expression changes induced by severe hypomethylation in DNA methyltransferase-deficient cells has been linked not to the demethylation of promoters, but rather to the activation of intergenic enhancer regions.³⁹ The resistance of such demethylated and hyperacetylated loci to remethylation could in part contribute to the synergistic activation of gene expression observed with the combination of DNA methyltransferase inhibitors and HDAC inhibitors and perhaps the treatment benefit observed in some clinical settings.^40-42

We find that CpGs embedded in regions of H3K27me3 in normal cells are more sensitive to demethylation, yet resistant to remethylation. While there is a great deal of literature suggesting that Polycomb occupancy predisposes to DNA methylation in a developmental context⁴³ and to aberrant hypermethylation in cancer cells,²⁰ the 2 marks are largely mutually exclusive genome-wide in differentiated cells, especially in CGI.^45,46 Indeed, recent work has demonstrated that genomic regions that lose DNA methylation after azacytidine treatment or deletion of Dnmt1, tend to gain H3K27me3.⁴⁷ This implies a model in which loss of DNA methylation enables restoration of PRC2 occupancy and deposition of H3K27me3, which protects from DNA remethylation. A similar phenomenon may be occurring with H3K9me3. While studies have indicated that H3K9me2/3 is required for the maintenance of DNA methylation⁴⁴ and participates in the direction of DNMT1 during replication,⁴⁸ recent work has demonstrated that H3K9me3 and DNA methylation generally repress distinct sets of genes.⁴⁹ That DAC-induced loss of H3K9me3 from repressed areas often leads to its replacement by H3K27me3⁹ suggests that the relationship between H3K9me3 and remethylation rate may be indirect, and a function of the accumulation of H3K27me3 in hypomethylated regions.

Among the most striking findings was the relationship between CpG remethylation rate and chromatin features associated with gene bodies and transcribed sequences. We find that H3K36me3 correlates well with the rate of remethylation and that the rapidly remethylating class is strongly enriched in chromatin features associated with transcriptional elongation. One possible mechanism for this is the recognition of H3K36me3 by the PWWP domain of the de novo DNA methyltransferases DNMT3A or DNMT3B.⁵⁰ Recent work suggests that DNMT3B1 in particular is selectively targeted to the bodies of transcribed genes in an H3K36me3- and PWWP-dependent manner.⁵¹ In this way, H3K36me3 may serve as a redundant memory of the lost DNA methylation, allowing rapid remethylation of these genomic regions following drug removal. This suggests that inhibition of SETD2,⁵² the methyltransferase responsible for H3K36me3,⁵³ or blockade of the DNMT3 PWWP domain may synergize with DAC in cancer cell reprogramming. Interestingly, while H3K36me3 and the density of DNA methylation in gene bodies correlates with steady-state transcript levels, they do not appear to be related (H3K36me3) or are negatively correlated (DNA methylation) with the rate of RNAP II elongation within gene bodies.^25,26 In contrast, we find that H3K79me2, which has been directly linked to RNAP II elongation rate²⁶ correlates with resistance to remethylation. Together, these data support the hypothesis that the function of DNA methylation and H3K36me3 in gene bodies is to provide a heritable memory of prior transcription and to suppress RNAP II initiation at cryptic start sites,⁵⁴ rather than a direct role in RNAP II elongation.

It has been proposed that cancer cells become addicted to at least some oncogenic driver methylation events acquired during tumorigenesis, and that the demethylation of such sites might contribute to the anticancer activity of DAC.⁵⁵ Critically, we find that the DAC-induced demethylation of CpGs that undergo cancer-specific hypermethylation, whether in the breast cancer cell line or in primary tumors, have a tendency to be more resilient against remethylation, whereas sites that are heavily methylated in normal HMEC cells tend to rapidly recover their lost methylation. One potential implication is that remethylation is directed specifically toward those CpGs that are normally methylated, while the aberrant methylation acquired in cancer cells is slower to recover or largely forgotten. The propensity for such “slow” sites, which included sites associated with known tumor suppressors and other genes that undergo cancer-specific hypermethylation (e.g., BRCA2, APC, TSC1/2, HIC1, TIMP3, HOX genes, SFRP4/5, POU3F3) to remain demethylated may contribute to the antitumor effects of demethylating agents. The swift remethylation of endogenous methylation may ensure that DAC treatment does not permanently disrupt the normal epigenome, and may explain in part why cancer cells are more sensitive to demethylating agents.¹¹ Indeed, it is interesting to note that the Rapidly remethylating sites tend to be associated with signal-responsive genes and genes involved in immune regulation (e.g., IL4, JAK, STAT5A, IFI35, IFITM1), the re-methylation of which may ensure their continued potential for rapid activation. Alternatively, in cancer cells, Rapid remethylation may prevent the persistent activation of genes whose expression is actively selected against. Consistent with this idea, CpGs found to be strongly resistant to demethylation and to be required for cell survival in a genetic model of DNMT1/3b depletion⁵⁵ also tended to be resistant to DAC-induced demethylation in our study, and the few that were demethylated are strongly enriched in the rapidly remethylating class (OR = 7.5, P = 4E-16, data not shown).

In sum, these finding suggest that similar genomic factors may govern sensitivity to programmed de novo methylation in development and remethylation in DAC-treated cancer cells. While promoters are rarely methylated in normal cells, gene bodies are dynamically and heavily methylated throughout development, and CpGs that specifically gain aberrant methylation in cancer cells are unlikely to regain that methylation following DAC treatment, even while normal methylation is immediately replaced. This restoration of the endogenous methylome is the express goal of epigenetic therapy, and thus our work reinforces the idea of epigenetic reprogramming and sustained interest in the application of demethylating agents in cancer treatment.

Methods

Cell culture and DAC treatment

Human mammary epithelial cells (HMECs) were obtained from Clonetics and maintained in Mammary Epithelial Cell Growth Medium (Lonza). MDA-MB-231 breast cancer cells were obtained from ATCC. Cells (4.5×10⁵) were plated in 10 cm plates and treated the following day with 500 nM DAC every other day for 6 d.¹⁶ Cells were then maintained in the absence of DAC for 27 passages, with cells being passed 1:10 every 3 d. DNA was harvested with the Qiagen DNA extraction kit before treatment, immediately following treatment, and at 3, 6, 9, and 27 passages in drug-free media. Three independent biological time-course experiments were performed.

DNA methylation analysis

Genomic DNA was bisulfite converted and hybridized to the Illumina Infinium HumanMethylation27 BeadChip by the Emory Integrated Genomics Shared Resource. β values were exported from Genome Studio and analyzed in R / Bioconductor.⁵⁶ Loci differentially methylated in response to DAC treatment were determined using a linear fixed-effects model as previously described.⁵⁷ Briefly, the “lm” function of the stats package in R / Bioconductor was used to determine the significance of DNA methylation changes. For clustering of 27K arrays, the Pvclust R package was used¹⁸ with the average agglomerative method of hierarchical clustering and correlation distance measure with 10,000 bootstrap replications. Pvclust calculates the approximately unbiased (AU) P-value via multiscale bootstrap resampling and the bootstrap probability (BP) P-value based on normal bootstrap resampling. Both were <0.01 in parsing untreated from DAC-treated methylation arrays. P-values were corrected for multiple hypothesis testing using Benjamini-Hochberg False Discovery Rate (FDR) correction.⁵⁸ Additionally, a minimum change in DNA methylation of β ≥ 0.2 was imposed. Data was analyzed using the “heatmap.2” function of the gplots package of R / Bioconductor. Genome-wide (tag) density plots and boxplots were generated using the ggplot2 package in R.

Self-organizing maps

To define patterns of re-methylation kinetics, the raw methylation data (β values) from each of the three time course experiments were first combined to obtain the average methylation levels for each CpG site at each time point. The recovery of lost methylation was then extrapolated from the average β values by setting the average initial value for each CpG site to 1, the average post-treatment value to 0, and determining the fraction of the initial methylation recovered at each passage after drug removal. Self-organizing maps were defined using GenePattern⁵⁹ with the following parameters: seed range= 42, 100,000 iterations, random vectors initialization, Gaussian neighborhood function, α_initial = .1, α_final = .005, σ_initial=5.0, σ_final = 0.5.

Analysis of DNA methylation in primary breast tumor-normal pairs

Level 3 Illumina Infinium HumanMethylation450 (450 K) BeadChip methylation data for the 4 SOM classes: Resistant, Slow, Moderate, and Rapid was extracted for 90 matched tumor-normal pairs of breast carcinoma (BRCA) from the TCGA data portal (https://tcga-data.nci.nih.gov/tcga/). The methylation data are preprocessed and normalized β values are available from TCGA. In total, 3,988 CpG sites were considered across the SOM classes that were present on both the Illumina 450K and 27K array. The analysis was restricted to CpG sites with a detection P-value > 0.05 across the dataset (N = 386,512), which included Resistant (N = 1,202), Slow (N = 1,517), Moderate (N = 901), and Rapid (N = 336). Unsupervised hierarchical clustering was performed using the Pearson correlation distance with agglomerative complete linkage for CpG sites in each SOM class across all 180 samples. We assessed the statistical significance of the SOM-defined CpG sites in separating breast tumors from normal vs. a randomly-defined set of CpG sites of the same number using a bootstrap approach. Specifically, we randomly sampled M = 1,000 times the same number of CpG sites in each remethylation class from among the 386,512 CpG sites remaining after filtering for low quality probes. For each re-sampling, dendrograms were constructed and cut based on a fixed number of k = 2 clusters. An association analysis was performed based on a Chi-Square test for each resampling and p-values obtained. A Monte Carlo P-value was defined by comparing the P-value obtained from an association analysis of tumor and normal samples to the distribution of P-values obtained based on the 2 clusters formed using randomly-sampled CpG sites.

CpG sites hypermethylated in primary breast tumors relative to normal tissue were identified from among the probes present on both arrays (N = 25,978) using a linear mixed effects model as implemented in CpGAssoc²⁸ with an FDR cutoff of 0.05. This analysis resulted in 2,356 hypermethylated sites, of which 440 were Resistant, 595 Slow, 328 Moderate and 98 Rapid. The remaining 895 were not among those significantly demethylated in our study. The odds ratio of enrichment of normally methylated CpGs, or those hypermethylated in breast cancers, among the 4 remethylation classes was determined by Fisher's exact. CpG sites with a β value > 0.7 in HMEC cells were considered to represent normally methylated sites (Total = 2,326, Resistant = 429, Slow = 794, Moderate = 711, Rapid = 392). Sites were considered to have undergone cancer-specific hypermethylation if the average β value in untreated MDA-MB-231 cells was greater than that observed in HMEC by at least 0.2 (2,464 Total CpG sites, Resistant = 944, Slow = 1,011, Moderate = 428, Rapid = 81)

Chromatin analyses

HMEC ChIP-Seq data sets were obtained from the ENCODE project. Accession numbers are GSM646374 (H3K27Ac), GSM646376 (H3K27me3), GSM646378 (H3K36me3), GSM646380 (H3K4me1), GSM646382 (H2K4me2), GSM646384 (H3K4me3), GSM646386 (H3K9Ac), and GSM646388 (H4K20me1). Tag densities for each ChIP-Seq were calculated in 20 bp bins for the 5,000 bp surrounding the CpG sites of interest using the GenomicRanges R package,⁶⁰ and normalized to total read count in each data set.

ChromHMM annotations were based on the classifications reported by Ernst et al.²⁴ for HMECs and were downloaded from the UCSC genome browser. For some analyses, ChromHMM classes of similar function were collapsed; “Promoter” was defined as ChromHMM classes 1–3, “Enhancer” as classes 4–7, “Heterochromatin” as classes 8 and 13–15, “Transcribed” as classes 9–11, and “Polycomb” as class 12. For the odds ratios calculations, the “Strong Enhancer” (4 and 5) and “Weak Enhancer” (6 and 7) categories were merged, as were the “Transcriptional Elongation” (9), and “Transcriptional Transition” (10) categories. No CpGs in the list of significantly demethylated sites were found in the Repetitive/CNV ChromHMM categories.

Disclosure of potential conflicts of interest

No potential conflicts of interest were disclosed.

Acknowledgments

The authors wish to thank Michael Nichols for helpful discussions. Supported by NIH grants R01-CA077337, R01-CA132065 (to PMV), NIH NRSA pre-doctoral fellowships F31-CA186676 (to JSKB) and F31-AI112261 (to BGB), a Department of Defense/CDMRP Breast cancer pre-doctoral fellowship CDMRP BC073543 (to JDK), and an American Cancer Society postdoctoral fellowship PF-07-130-01-MGO (to MTM).

Funding

This research project was supported in part by the Winship Biostatistics and Bioinformatics Shared Resource of Emory University and the Winship Cancer Center Support grant P30 CA0138292.