Reprogramming the Methylome: Erasing Memory and Creating Diversity

1Cellular identity is maintained by epigenetic memory, which in part reflects the developmental history of a cell. Epigenetic memory relies upon the faithful inheritance of epigenetic marks (such as DNA methylation or histone modifications) during cell division. This heritability is a hallmark of epigenetic regulation and, in the case of DNA methylation, is brought about by a well-understood mechanism.

2In mammals, cytosine methylation occurs mostly in the context of palindromic CpG dinucleotides, where methylation occurs on both DNA strands in an antiparallel fashion. During DNA replication, the CpG methylation pattern can be copied from the template strand to the newly synthesized DNA strand. The UHRF1 protein recognizes hemimethylated sites and recruits the maintenance DNA methyltransferase, DNMT1, to methylate the newly synthesized strand. In contrast, the de novo methyltransferases DNMT3A/B, and their cofactor DNMT3L, do not depend on a hemimethylated template and can establish new patterns of DNA methylation. As well as methylating CpG dinucleotides, these enzymes are thought to be responsible for the low, but consistent, levels of non-CpG methylation observed in oocytes, prospermatogonia, embryonic stem cells (ESCs), and neural cells. There is no apparent epigenetic memory of non-CpG methylation, as there is no mechanism to maintain these marks following replication.

3Epigenetic memory is thought to be robust in somatic tissues, where levels of CpG methylation are consistently high (70%–85%) (Figure 1). Methylation is mainly located in transposons, gene bodies, and intergenic regions, while regions of high CpG density (known as CpG islands, CGIs), often located at gene promoters, are generally kept free of methylation. A minority of CGIs (approximately 10% of a total 23,000 in mouse) are differentially methylated between tissues. In addition, distal regulatory elements, such as enhancers (defined by DNase I hypersensitivity, transcription factor binding, and chromatin modifications), show differential methylation across tissues. Thus, tissues and developmental lineages can be robustly delineated by the extent and pattern of DNA methylation at relatively few genic and nongenic CGIs and enhancers. This epigenetic patterning probably provides a framework for the stability of the 10,000–13,000 genes that are differentially expressed between somatic tissues. These patterns first arise during early postimplantation development and are dependent on the de novo methyltransferases DNMT3A and B, but the precise mechanisms by which they are generated are not known.

4In marked contrast to somatic tissues, mammalian primordial germ cells (PGCs), early embryos, and naive embryonic stem cells (ESCs) have methylation levels between 5% and 30% (Figure 1), meaning that in the extreme case they have lost up to 15 million methylated CpGs per haploid genome. Since these hypomethylated cell types are pluripotent, or in the case of PGCs can give rise to cells with pluripotent ability (that is, the capability to differentiate into all cell types of the embryo), it raises the question of whether and how epigenetic reprogramming is connected with developmental capacity.

5Here we compare the synergistic mechanisms that result in epigenetic memory loss in PGCs, the early embryo, and naive ESCs, and we explore the link between pluripotency and loss of epigenetic memory. Importantly, as cells exit pluripotency and begin cell fate commitment, they re-establish epigenetic marks. We propose a model whereby heterogeneous patterns of DNA methylation are generated at this time to allow cell lineage priming prior to commitment.

6Primordial germ cell (PGC) specification results in the single largest change of DNA methylation in the mammalian life cycle. PGCs differentiate from precursor cells in the epiblast, which at E6.5 is as highly methylated as somatic cells, and over the following 7 days or so lose 90% of their global methylation (Figure 1). Significantly, CGI methylation patterns in early PGCs correlate highly with the epiblast and other somatic tissues but not with the oocyte, the ICM, or any other cells of the preimplantation embryo. Although quantitative single-cell analysis will undoubtedly further illuminate this phenomenon, the presence of an epiblast-derived methylation signature in PGCs almost certainly demonstrates that they are reprogrammed from differentiated tissues rather than being rare remnants of the ICM that have escaped de novo methylation.

7Demethylation in PGCs occurs in two phases (Figure 2), the first during their migration to the gonad anlagen, beginning at around E8.0. Careful analysis of migratory PGC demethylation over consecutive cell divisions and hairpin bisulfite analysis has revealed that this process mainly occurs by inactivation of the maintenance methylation system. The underlying mechanism presumably involves transcriptional downregulation of Uhrf1 and exclusion of UHRF1 protein from the nucleus, as well as loss of H3K9me2, which in other systems causes DNMT1 to be mistargeted at replication foci. As such, relatively indiscriminate passive demethylation reduces global methylation levels to 30% by the time PGCs first arrive in the gonad anlagen. In addition, downregulation of the de novo methyltransferases DNMT3A, -3B, and -3L precludes restoration of methylation patterns following replication, contributing to rapid loss during the first wave of demethylation. Despite this widespread reduction, specific sequences remain methylated until the second wave of demethylation occurs around E11.5, including differentially methylated regions (DMRs) in imprinted loci, CGIs on the X chromosome, and germline-specific genes. Demethylation of these sequences appears to require the TET1 and TET2 proteins, which oxidize 5-methylcytosine (5mC) to 5-hydroxymethylcytosine (5hmC). How the TET enzymes are targeted and the ultimate fate of hydroxymethylated DNA is the subject of debate—5hmC may be passively lost by replication (currently there is no known mechanism to maintain 5hmC), or it could be subjected to further oxidation to 5-formyl- and 5-carboxyl-cytosine (5fC and 5caC, respectively), which can be actively removed by decarboxylation or base excision repair. Interestingly, base excision repair (BER) is also implicated in PGC demethylation, and deficiency of the deaminase AID results in 2- to 3-fold-higher levels of methylation in PGCs at E13.5. However, the time and substrate of action, as well as the genomic targets of AID, are unknown. The glycosylase TDG, which can excise 5fC and 5caC as well as deaminated 5mC, is not expressed during the second window of demethylation. Thus, while the BER pathway has been implicated in PGC demethylation, it is not clear precisely which components are involved and to what extent. Elucidating the combinatorial action of various factors during the demethylation process will need detailed studies analyzing where these factors are targeted in the genome and which downstream pathways resolve specific base modifications.

8Despite the magnitude of PGC reprogramming, there are some sequences in the genome that are resistant to demethylation. These recalcitrant sequences include members of the intracisternal A particle (IAP) retrotransposon family and adjacent CGIs, as well as approximately 250 CGIs in the genome, which are variably erased (VECs). This variable erasure may be significant because it provides a potential mechanism for transmission of epigenetic traits acquired in one generation to the next, including metabolic phenotypes recently observed in animal models. The large-scale erasure of epigenetic marks in the germline may explain why transgenerational epigenetic inheritance in mammals is rare. Indeed, it is more common in plants that do not undergo genome-wide epigenetic reprogramming in germ cells.

9Another major wave of demethylation in the mammalian germline occurs in the early embryo, and while the magnitude of methylation loss might not be as large as in PGCs, it is perhaps more dramatic due to the precipitous loss of DNA methylation specifically in the male pronucleus. Here, highly methylated sperm-derived DNA is rapidly and actively demethylated before and during the first S phase in the zygote, making it a definitive example of large-scale active demethylation (Figure 2). A number of factors have been implicated in this rapid demethylation, most prominently hydroxylation by TET3. Unlike TET1 and TET2, TET3 is highly expressed in the zygote and is required for most of the 5hmC that appears in the paternal genome upon fertilization, and for a proportion of its demethylation. Until recently, it was thought that only the paternal genome undergoes active demethylation in the early embryo; however, the discovery of considerable amounts of 5hmC and 5fC in both pronuclei suggests that it may also affect the maternal genome to some extent. As TDG is apparently not involved in removing these modifications, it is thought that they are lost during DNA replication. Recent work has shown that TET3-mediated hydroxylation and demethylation is confined to S phase chromatin, while surprisingly the bulk of the demethylation occurs prior to DNA synthesis (Figure 2, stages 1 and 2). AID was also shown to be required for paternal demethylation, and initial deamination may be resolved by the UNG2 glycosylase rather than TDG, opening up the intriguing possibility that cytosine deamination and long-patch BER are involved in demethylation. Unravelling the targets of these active demethylation pathways in the zygote and to what extent they synergize or operate in parallel will require detailed genome-wide profiling coupled with disruption of specific pathways.

10As in PGCs, inactivation of the maintenance methyltransferase and de novo methylation system is thought to play a role in demethylation in preimplantation embryos following the zygote stage. UHRF1 protein is highly abundant in the zygote but is excluded from the nucleus together with DNMT1, resulting in passive demethylation of the maternal (and paternal) genome in the early embryo such that by E3.5 methylation in the blastocyst reaches a global methylation level of ∼25% (Figure 1). From basal levels in the early blastocyst, methylation slightly increases in the ICM (E3.5) and the placenta. As in migratory PGCs, some regions such as IAPs, and DMRs in imprinted genes, are protected from both active and passive demethylation in the early embryo. The mechanism by which DMRs are protected is thought to involve binding by proteins such as STELLA, ZFP57, and KAP1 and recruitment of the low level of nuclear DNMTs and UHRF1.

11An outstanding question in the field is the extent to which DNA demethylation contributes to the activation of the early embryo transcriptional network and acquisition of totipotency. A large number of retrotransposon sequences are expressed in the early embryo. Expression of some of these repetitive sequences may be required for zygotic genome activation—inhibition of both endogenous retroviral elements (ERVs) and LINE1 elements can impair developmental competence of embryos. Moreover, around 300 early embryo genes are activated by the LTRs of the MERV-L class of endogenous retroviruses through the production of chimeric transcripts, directly implicating activation of ERVs and establishment of the early embryo transcriptional network. Along with the rest of the genome, these elements are subjected to rapid DNA demethylation in the early embryo that coincides with their activation. Recently, a rare population of ESCs that express markers of the early embryo was identified, and as these cells can contribute to the extraembryonic lineages, they are considered to have totipotent features. Relative to other ESCs, this subpopulation has reduced methylation at MERV-L repetitive sequences, as in the early embryo. Despite this, complete removal of DNA methylation from mouse ESCs apparently does not stimulate expression of ERV elements or the transcriptional network of the early embryo. Thus, the extent to which the totipotent state is dependent upon DNA demethylation, either in the context of ERVs or the transcriptional network of the early embryo in general, is currently unclear.

12While the link between pluripotent cell types and genome-wide methylation is apparent in vivo (as witnessed by strikingly similar methylation levels of E9.5 PGCs and ICM), it was initially puzzling that ESCs conventionally grown in serum containing media exhibited high CpG methylation levels (∼80%) characteristic of somatic cells, in contrast to the hypomethylated ICM from which they are derived (Figure 1). This paradox was recently resolved; by blocking the prodifferentiation signal that serum-grown ESCs receive (by inhibiting ERK1/2 and GSK3β signaling with two small molecule inhibitors), genome-wide demethylation was induced to the same extent as in preimplantation embryos (∼30%). In addition to achieving striking demethylation, use of these two inhibitors (commonly known as “2i”) enabled derivation of bona fide naive ESCs from previously recalcitrant mouse strains, rats, and even humans. Accordingly, 2i ESCs are now considered the best cell culture representation of in vivo pluripotent stem cells.

13In contrast to serum-grown ESCs, naive ESCs grown in 2i have low levels of DNMT3A/B and their targeting factor DNMT3L, suggesting a mechanistic explanation for their low methylation levels. While knockdown of the de novo methyltransferases in serum-cultured ESCs does not immediately affect many regions thus far analyzed, LINE1 elements undergo demethylation with similar kinetics to 2i treatment following DNMT3A/B ablation. These elements (and some single-copy loci) were previously shown to require continuous de novo methylation in order to maintain high levels of methylation in serum-grown ESCs, indicating that DNA methylation maintenance is inefficient at these loci. Thus, DNMT3A/B repression in 2i is sufficient to cause demethylation of certain genomic regions in these cells.

14Demethylation in 2i occurs with similar kinetics to that of PGCs and early embryos (Figure 2), but surprisingly without global alteration of UHRF1 and DNMT1 expression. TET2 is upregulated by 2i treatment and hydroxylation is induced. TET1 is targeted preferentially to CGIs while TET2 is enriched at gene bodies, where more substantial demethylation is observed in 2i. The extent of demethylation can be enhanced by the addition of Vitamin C to 2i culture medium, which is a cofactor for activity of the TET enzymes. Interestingly, this demethylation is focused on specific sequences including germline-specific genes. Hence a significant mode of demethylation may be erosion of methylation patterns by hydroxylation and failure to repair this erosion by de novo methylation after replication.

15The similarity of demethylation mechanisms between migratory PGCs, preimplantation embryos, and 2i ESCs extends to resistance to methylation erasure of IAP retrotransposons and DMRs in imprinted genes. However, with prolonged culture of ESCs in 2i, some erosion of DMR methylation does occur, perhaps in part because of the consistently elevated levels of PRDM14. In that respect, prolonged 2i treatment may mimic, to some extent, the transition of migratory to gonadal PGCs.

16In summary, genome-wide demethylation is achieved in naive ESCs by a combination of disabling the de novo methylation machinery and increased hydroxylation, which may occur in the context of a partially impaired maintenance methylation system (Figure 2). Further study is required to elucidate the exact contribution of these different mechanisms at distinct genomic sites and how this relates to what is known about the removal of epigenetic memory in in vivo systems.

17Lack of epigenetic memory is a common characteristic of pluripotent cell types and their precursors, including PGCs, ESCs, and induced pluripotent stem cells (iPSCs). This is thought to ensure that future differentiation decisions are not affected by events in the past. In support of this idea, inefficient methylation erasure at imprinted loci has been associated with a spectrum of developmental abnormalities in the progeny of Tet1-deleted male mice. Insights into the mechanistic links between reprogramming and pluripotency are now beginning to emerge.

18All experimental reprogramming techniques, including somatic cell nuclear transfer (SCNT), cell fusion, and iPSC reprogramming, involve demethylation of the genome that appears to be crucial for successfully achieving pluripotency ( this issue of Cell Stem Cell). Furthermore, ESCs that lack all three methyltransferases (and are devoid of virtually all methylation) are highly resistant to differentiation, spontaneously reverting to pluripotency. Full expression of the pluripotency network is observed in both migratory PGCs and ICM cells, showing that a genomic methylation level of 30% or less is characteristically associated with the naive pluripotent state (Figure 1). This appears to be true for naive ESCs derived from mice and may be useful as a diagnostic marker for naive cells from other species. Expression of key pluripotency factors is associated with demethylation of these loci in PGCs, in early embryos, and during experimental reprogramming. However, global demethylation in PGCs is not associated with promiscuous transcription, nor is promoter demethylation in 2i associated with transcriptional activation of demethylated genes. Thus, while global DNA demethylation is required for activation of the pluripotency network, pluripotent cells exhibit an uncoupling of DNA methylation and transcriptional regulation. Interestingly, expression of the pluripotency network declines in gonadal PGCs (E13.5), which show further demethylation to less than 10%. Perhaps extreme demethylation is detrimental since it may lead to genome instability especially during cell division (methylation of major satellites, for example, appears to be needed for proper chromosome segregation).

19Just as DNA methylation levels potentially influence the expression of the pluripotency network, the pluripotency network can direct the machinery of epigenetic reprogramming. TET1 and TET2 are directly linked to the pluripotency network by physical interactions with NANOG and PRDM14. TET1 and NANOG have overlapping patterns of DNA binding, and NANOG is required to recruit TET1 to a subset of these common loci. Furthermore, NANOG and TET1 act synergistically during the reprogramming of neural stem cells to iPSCs, and TET1 can replace OCT4 in conventional fibroblast reprogramming. Similarly, PRDM14 recruits TET1/2 to target loci and TET1/2 enhance PRDM14-induced DNA demethylation in ESCs. DNMT3A/B/L are negatively regulated by PRDM14, in part via binding of PRDM14 to an upstream enhancer of Dnmt3b. NANOG also binds this locus and may have similar effects on Dnmt3b expression. As well as the de novo methyltransferases, UHRF1 appears to be suppressed by PRDM14 in PGCs. These molecular interactions conspire to ensure that DNA methylation is maintained at low levels in pluripotent cells.

20In summary, the pluripotent ground state is maintained by direct interactions between the transcriptional network and the DNA methylation machinery. Cross-regulatory mechanisms ensure that robust expression of pluripotency factors is accompanied by stable hypomethylation of the genome, yielding a pluripotent ground state with little epigenetic memory.

21As cells exit pluripotency and begin to differentiate, new epigenetic memories are formed to define cellular identity and restrict lineage choices. The exit from pluripotency is characterized by a steep decline in TET1/2 levels and an increase in DNMT3A/B enzymes together with re-engagement of DNMT1 at replication foci. During this transition, cells appear to pass through an intermediate state of epigenetic priming that is characterized by high levels of both de novo methyltransferases and TET enzymes. For example, cells expressing both DNMT3B and TET1 are present in the ICM of late blastocysts, and the E6.5 epiblast displays high expression of DNMT3A/B and of TET1. Interestingly, serum ESCs are similar to the E6.5 epiblast in the pattern and extent of DNA methylation (Figure 1), and ESCs grown in serum conditions also express high levels of DNMT3A/B/L and of TET1/2, making them a useful model for the relatively transient primed cell population in vivo. These cells also display remarkable methylome plasticity that is highly sensitive to growth conditions. For example, the methylation status of CGIs even varies with cell passage number. We speculate that rapid turnover of DNA methylation may generate epigenetic heterogeneity in primed ESCs and in the epiblast and that this heterogeneity may contribute to cell fate decisions by allowing diversification prior to lineage commitment. Consistent with this hypothesis, the number of low-methylated regions (LMRs, an indirect indicator of heterogeneous methylation across a cell population) decreases as mouse ESCs differentiate to neural progenitor cells. However, a decrease in the number of LMRs was not observed in a similar study of human ESC differentiation. Further studies utilizing single-cell approaches are necessary to resolve these discrepancies by directly analyzing cell-to-cell variation in DNA methylation.

22Heterogeneity at the level of transcription has already been linked to the propensity of serum ESCs to differentiate. Reporter cell lines have shown that key pluripotency factors such as NANOG, STELLA, and REX1 are heterogeneously and dynamically expressed in these cells. For example, while NANOG-low and -high cells are interchangeable, the NANOG-low population has an increased propensity to differentiate and elevated expression of differentiation markers. More recent studies have employed computational approaches, single-cell analyses, and live-cell imaging to demonstrate that serum ESCs are a metastable population and that cells switch between transcriptional states in a stochastic manner.

23Emerging evidence indicates that epigenetic mechanisms may contribute to this metastable state. One study has found that STELLA-low and -high cells have different chromatin modifications at the Stella promoter, and a slight increase in DNA methylation was observed in a subset of STELLA-low cells. We have also reported differences in 5mC and 5hmC between NANOG-low and -high cells, with increased expression of TET1/2 being associated with increased 5hmC in NANOG-high cells. In line with these findings, it has recently been shown that transcription patterns between daughter cells differ depending on the methylation state. The hypomethylated state of ESCs grown in 2i is associated with greater transcriptional similarity between daughter cells than for ESCs grown in serum. Furthermore, cells that are deficient for the methyltransferases (TKO cells) have daughters with highly similar transcriptional patterns despite being in serum culture conditions. Generally, DNA methylation is thought to act downstream of transcriptional changes during cell fate decisions, but this fascinating result suggests that DNA methylation heterogeneity may also be able to generate transcriptional diversity in certain contexts. In support of this idea, a study of DNMT3A null hematopoietic stem cells (HSCs) has suggested that DNA methylation dynamics can direct transcriptional changes and lineage choice in response to differentiation stimuli. Ultimately, transcriptional and epigenetic heterogeneity are likely to be tightly linked and cross-regulatory, such that stochastic differences at either level can generate cell diversity.

24DNA methylation heterogeneity may be generated by several mechanisms (Figure 3). Stochastic fluctuations in the expression of TET or DNMT3 enzymes could generate differences in DNA methylation between sister cells. Differential targeting of these enzymes could also produce cells with variable epigenetic marks at the same sequence. Alternatively, strand-specific effects of DNMT3 or TET enzymes could introduce asymmetric DNA modifications that would yield daughter cells with distinct methylation patterns. Differences in the efficiency of methylation maintenance may also lead to DNA methylation heterogeneity within a cell population, and incomplete erasure of inherited methylation patterns (e.g., oocyte-derived methylated regions) might be an additional source of epigenetic heterogeneity. The precise patterns and kinetics of such events across cell populations will depend on the balance between expression of DNMT3 and TET enzymes, as these genes are themselves heterogeneously expressed in ESCs. Rapid turnover of DNA methylation through 5hmC is also likely to contribute to epigenetic heterogeneity as CGIs with variable methylation in ESCs are also enriched for 5hmC. Loss of DNA methylation through 5hmC will occur via hemi-methylated intermediates since there is no known mechanism for 5hmC maintenance (Figure 3). Consistently, 5hmC has been linked to hemi-methylation at repetitive elements, with LINE1 elements having increased 5hmC, increased hemi-methylation, and increased hemi-5hmC relative to IAPs. LMRs are also enriched for 5hmC and are bound by TET1. The presence of characteristic histone modifications and transcription factor binding at LMRs predicts that these loci are distal regulatory elements. Thus, DNA methylation turnover through 5hmC and hemi-methylation may generate heterogeneous methylation at enhancers and other regulatory elements. This could modulate their transcriptional output (Figure 4), since many transcription factors display methylation-sensitive sequence binding with the majority preferring an unmethylated substrate. In support of this model, methylation has been shown to inhibit the activity of lineage-specific enhancers in the context of T cell differentiation.

25The epigenetically primed diversity seen in ESCs may also exist in other contexts such as hematopoietic stem cells (HSCs). In a manner analogous to ESCs, HSCs display transcriptional heterogeneity corresponding to their differentiation potential. These cells can be fractionated based on the expression of the stem cell marker SCA-1; SCA-1-high and -low subpopulations are interchangeable and are predisposed to adopt the myeloid and erythroid lineages, respectively. Loci that are hypomethylated specifically in either myeloid or lymphoid cells have intermediate levels of methylation in HSCs, and a myeloid specific locus displays stochastic DNA methylation in these cells. These findings suggest that DNA methylation heterogeneity in HSCs underscores transcriptional heterogeneity and precedes lineage commitment.

26In general terms, creating heterogeneity of gene expression at critical times in development is expected to help with symmetry breaking during cell fate decisions. By permitting gene expression and epigenetic heterogeneity, primed cells are able to diversify prior to lineage commitment. This diversification could allow cells to respond differently to uniform differentiation stimuli, such that multiple lineages may be initiated from the same pool of stem cells (Figure 4B). Our model thus predicts that the transitional fine-tuned overlap between methylation and demethylation systems is critical for cell fate decisions during gastrulation. This model needs to be interrogated using emerging single-cell techniques, which aim to decipher the complex relationships between DNA methylation and other sources of epigenetic and transcriptional heterogeneity in the same cell.

27Recent studies have elucidated the synergistic mechanisms that orchestrate genome-wide epigenetic reprogramming in germ cells and early embryos. Global demethylation appears to be predominantly the result of disabling the maintenance and de novo methyltransferases, while modifications of cytosine and DNA repair may be needed for more targeted demethylation events. Global hypomethylation of the genome is inextricably connected with pluripotency through reciprocal links between the DNA methylation machinery and the pluripotency transcription factor network. This ensures that naive pluripotency is essentially devoid of epigenetic memory, so that events of the past do not influence future differentiation decisions. Epigenetic memory is re-engaged at the exit from pluripotency with robust expression of the de novo and maintenance methylation systems. We propose that differentiating cells pass through a transient epigenetically primed state characterized by high levels of TET and DNMT enzymes, and heterogeneous patterns of DNA methylation. Notably, primed ESCs (grown in serum) express high levels of methyltransferases and of TET1 and 2, and similar cells appear to exist in the ICM in vivo. Epigenetic heterogeneity in the ICM and epiblast may aid cell fate decisions by allowing diversification prior to lineage commitment.

28H.J.L., T.A.H., and W.R. wrote the manuscript and prepared figures.