Transcription factors as readers and effectors of DNA methylation

Tandem mass spectrometry

One systematic approach for the discovery of mCpG-binding proteins is based on tandem mass spectrometry (MS/MS)^50,51. A recent study used a generic DNA sequence harbouring an mCpG site to pull down interacting proteins from nuclear extracts of cultured cells⁵¹. Proteins bound to methylated DNA sequences were then identified by MS/MS. Based on this approach, 19 proteins were identified that interact preferentially with the methylated DNA probe rather than the non-methylated counterpart in mouse embryonic stem (ES) cell nuclear extracts. Besides the known MBD proteins (such as MeCP2, MBD1 and MBD4), many TFs (such as MHC class II regulatory factor RFX1, zinc-finger homeobox 3 (ZFHX3), lysine-specific histone demethylase 1A (LSD1), zinc-finger and BTB domain-containing protein 44 (ZBT44) and thymocyte nuclear protein 1 (THYN1; also known as THY28), and the Krüppel-like factors (for example, KLF2, KLF4 and KLF5)), were identified as new mCpG-binding proteins (TABLE 2). The authors applied the same approach to neuronal progenitor cells and found that a large and distinct set of proteins showed preferential binding to mCpG sites, suggesting that the interaction with mCpG is dynamic and thus varies under different physiological conditions. A similar approach was also used to identify nucleosome-interacting proteins that are affected by DNA methylation⁵⁰. Although proteins from cell extracts are in a more native state in the MS/MS-based approach, the DNA probes used in this approach are typically generic and, therefore, sequence specificity of observed interactions remains elusive.

Functional protein microarray

Functional protein microarrays have been used as a powerful tool to profile protein–DNA interactions in the past⁵². A comprehensive examination of sequence-specific mCpG-binding activities was conducted by sequentially probing a human protein microarray containing 1,321 TFs and 210 cofactors with 154 DNA motifs that each carried at least one mCpG site⁵³. To identify human TFs that preferentially bind to methylated DNA motifs, each methylated motif was mixed with its unlabelled and unmethylated counterpart in tenfold excess in the binding assays. This competition assay ensures that the identified interactions are indeed methylation-dependent, rather than due to CpG-flanking sequences. Of the 154 methylated motifs examined, 150 showed strong binding signals to at least one protein on the microarray. In total, 41 TFs and 6 cofactors were found to bind to at least one methylated sequence. Most of these factors were found to bind to only a few methylated sequences, suggesting that the interactions are not only methylation-dependent but also sequence-specific. Interestingly, the factors that showed binding activity to methylated sequences were widespread among various TF subfamilies, such as zf-C2H2, homeobox, bHLH (basic helix–loop–helix), forkhead, bZIP and HMG (high-mobility group) box. Many of these factors are known to be involved in tissue development or have been associated with cancer. A subsequent validation assay showed that some of these TFs indeed bind to methylated DNA in vivo and regulate gene expression⁵³.

DNA microarray

DNA (or protein-binding) microarray technology has been used to determine the binding specificity of TFs^54,55. A double-stranded DNA microarray, typically comprising 40,000 unique DNA sequences that cover all possible combinations of 8–10-nucleotide-long DNA sequences that could constitute a binding motif, is incubated with a purified TF so that its binding preference can be accurately determined. In a recent study, the bacterial DNMT SssI was used to methylate the CpG sites of the sequences on the array⁵⁶, followed by individual probing with eight purified proteins containing bZIP domains. By comparing the binding profiles of each protein obtained on the methylated and unmethylated microarrays, proteins that preferentially bind to specific sequences were determined. Among the eight bZIP proteins, CEBPα and CEBPβ were found to specifically bind to a methylated sequence⁵⁶. This approach enables a large amount of DNA sequences to be surveyed for protein–DNA interactions; accurate sequence specificity can, therefore, be determined for a given protein. However, prior knowledge of a candidate TF is required because it can be cumbersome to survey an entire TF family. Therefore, this approach is ideally used for fine-mapping sequence specificity of a previously identified mCpG-binding protein.

ChIP–BS-seq

To determine methylation-dependent protein-DNA interactions in vivo, chromatin immunoprecipitation followed by bisulfite sequencing (ChIP–BS-seq) is an ideal approach. ChIP is first performed to obtain the DNA sequences that are bound by a protein of interest, and then the methylation level is sequentially determined using BS-seq. This approach was developed recently to determine the crosstalk between histone modifications and DNA methylation^57–59. However, it requires prior knowledge of mCpG-binding proteins and the availability of antibodies that are directed against the proteins of interest. For example, after KLF4 was determined to bind to mCpG sites, ChIP–bisulfite conversion followed by PCR was used to validate the methylated DNA–protein interactions in vivo⁵³. It is important to note that ChIP–BS-seq is the only approach that does not use naked DNA fragments to identify the TFs that bind to methylated DNA. Therefore, TFs identified using the other methods may not necessarily recognize mCpGs in vivo, and further studies are needed to dissect the functionality of the interactions.

Protein domains that interact with methylated DNA

Identification of the protein domains that recognize mCpG sites is important to characterize mCpG-dependent protein–DNA interactions. Such knowledge will enable the mutation of critical residues within these domains that abolish the mCpG-dependent binding activity of these proteins, while maintaining their ability to bind non-methylated DNA. Therefore, mutated proteins can be useful tools to dissect the physiological roles of mCpG-dependent protein–DNA interactions.

Besides the well-known MBDs, other protein domains seem to interact with mCpG sites. For example, the recent crystal structure of mouse ZFP57 in complex with a methylated DNA sequence demonstrated that its two zinc-fingers interact with methylated DNA, and that an arginine (Arg178), which is involved in hydrophobic interactions, plays a crucial part in mCpG binding⁴⁴. A separate study suggested that an arginine and glutamate pair in KLF4 recognizes the mCpG site⁶⁸. A structural comparison of MeCP2 and KLF4 indeed showed a common structural feature involving one arginine and one asparagine⁵³. A global survey of methylated-DNA-binding proteins suggests that many other protein domains might also be able to interact with mCpG sites, including homeobox, HLH and E2F domains⁵³.

There are currently no general rules of evolutionary conservation for the domains that interact with methylated sites, owing to a lack of data from multiple species. However, in a comparison of mCpG-binding proteins between humans and mice^51,53, a few proteins such as KLF4 and homeobox A5 (HOXA5) were shown to bind mCpG sites in both species, which is indicative of the functional importance of methylation-dependent protein–DNA interactions.

Sequence specificity

Notably, many proteins can bind both non-methylated and methylated sequences in a different sequence context. For example, CEBPα is known to bind a particular sequence element, 5′-TGACGTCA⁴². However, when the CpG is methylated, CEBPα can effectively recognize half of the motif: 5′-mCGTCA⁴². Similarly, although KLF4 recognizes a non-methylated canonical motif of 5′-TTTACGCC, it has been demonstrated that KLF4 specifically recognizes a 5′-TCCmCGCCC motif only when the CpG is methylated⁵³. If the methylation status of these two sequences is exchanged, KLF4 loses the ability to bind to either sequence⁵³. Indeed, for many newly discovered methylated DNA-binding proteins, the methylated motifs differ from the non-methylated motifs⁵³. Therefore, it is reasonable to speculate that 5mC might represent the fifth nucleotide that further fine-tunes the specificity of protein–DNA interactions; that is, 5mC acts as an additional regulatory layer to remove, create and/or change TF binding sites (FIG. 1).

Several recent studies have started to provide the structural basis for the altered sequence specificity due to DNA methylation. Both in vitro DNase I digestion assays and structural studies indicate that methylation has a profound impact on DNA structure and shape (for an in-depth review see REF. 69). Adding a methyl group to the cytosine could affect the local DNA shape, as evidenced by the altered DNase I digestion rate and patterns^70,71. Similarly, based on a few reported crystal structures of double-stranded DNA fragments with 5mC bases^68,72,73, the presence of a bulky methyl group in the major groove leads to a subtle widening of the major groove and a subtle narrowing of the minor groove. Consequently, 5mC can affect the access of a given TF to the affected motifs in both major and minor grooves in genomic DNA and thus change the sequence specificity of protein–DNA interactions.

Binding affinity

One important question is whether methylated-DNA–protein interactions have a similar binding affinity to the interactions between the same protein and a non-methylated DNA, or whether they are just labile interactions. Using an in vitro pulldown-coupled MS/MS approach, the relative affinity of protein–mCpG interactions can be estimated⁵¹. For example, for a particular sequence (5′-GGGCGTG), which was determined on the basis of the KLF4 ChIP– seq data sets⁷⁴, KLF4 showed higher affinity when the cytosine in the motif was methylated compared with the unmethylated sequence⁵¹. The protein microarray approach⁵³, which uses the concept of relative affinity to identify proteins that preferentially bind to methylated DNA, revealed proteins with strong fluorescent signals, which are expected to bind tighter to the methylated motif than to the unmethylated counterpart. The absolute binding affinity (that is, K_d values) can be measured by applying the oblique incidence reflectivity difference (OIRD), which is a real-time, label-free method to measure the kinetics of a binding event^53,75. Three proteins (ZMYM3, AP2α and KLF4) were selected to determine the K_on and K_off values with their corresponding motifs in methylated forms. The deduced K_d values of ZMYM3, AP2α and KLF4 were determined as 460 nM, 399 nM and 479 nM, respectively. Importantly, no obvious affinity could be detected when these tested motifs were unmethylated. As a comparison, the K_d values of the short isoform of MBD2, MBD2b, for the same motifs ranged from 97 nM to 197 nM, suggesting that MBD-lacking TFs bind to methylated DNA motifs nearly as strongly as MBD2b.

Cis-regulatory elements

To better understand the physiological role of mCpG-binding proteins, it is important to determine which methylated regions in the genome can be specifically recognized by these proteins. Although MBD family proteins tend to bind to regions with a high methylation density (that is, high methylation level and high CpG density)⁷⁶, it is interesting to examine whether the same is true for sequence-specific mCpG-binding proteins.

As protein-DNA interactions are dynamic, differentially methylated regions might be possible candidates for methylation-dependent interactions. Analysis of the methylomes obtained from 17 adult mouse tissues at single base-pair resolution showed that approximately 6.7% of the mouse genome is differentially methylated, mostly at distal cis-regulatory regions⁷⁷. Another study discovered that regions with a low level of methylation, ranging from 10% to 50%, often occur at distal regulatory regions⁷⁸; that is, regions that are enriched for enhancer marks, including high levels of histone H3 lysine 4 monomethylation (H3K4me1) as well as binding sites for p300 histone acetyltransferase and other regulatory factors. Similarly, extensive DNA methylation was found to coexist with active H3K27 acetylation (H3K27ac) marks in a large number of enhancers⁷⁹. More importantly, the reduction of DNA methylation led to a decrease in H3K27ac marks, suggesting an active role of DNA methylation in regulating enhancer activity. Based on the analysis of KLF4 binding in ES cells, we also found that KLF4 binds to methylated enhancer regions⁵³, which may suggest that sequence-specific mCpG-binding proteins interact preferentially with distal enhancer regions.

Protein binding affects the DNA methylation status

The binding of methyltransferases or methylcytosine dioxygenases (for example, DNMTs and TETs (ten-eleven translocation proteins)) affects the status of DNA methylation, but recent studies suggest that many non-enzymatic proteins, such as TFs, could regulate the establishment and maintenance of the local DNA methylation levels in a sequence-specific fashion. One such regulator is the transcriptional repressor CTCF, which is known to have an essential role in imprinting control; that is, to achieve allele-specific gene regulation⁸⁰. CTCF binds to the unmethylated ICRs in maternal alleles, which prevents distal enhancers from activating downstream genes⁸¹. By contrast, when the paternal ICR is methylated, CTCF cannot bind to the ICR, thus allowing the activation of downstream genes by distal enhancers. One study suggests that CTCF itself contributes to the maintenance of the non-methylated status of maternal ICRs, as maternally transmitted mutant ICRs in neonatal mice that harbour point mutations in CTCF binding sites acquire a heterogeneous degree of methylation⁸². Although the traditional view of imprinting control is that differential methylation leads to differential binding of CTCF, and thus yields allele-specific gene regulation, this study suggests that CTCF binding itself is necessary to maintain differential methylation of ICRs.

Moreover, a recent report confirmed that the binding of some proteins (for example, CTCF and RE1-silencing transcription factor (REST)) can affect local methylation patterns⁷⁸. The authors first created a reporter construct with a CTCF-binding motif that was inserted into a genomic locus in mouse ES cells. Insertion of the binding site induced CTCF binding and resulted in a reduced methylation level in local genomic regions. A single-nucleotide mutation in the CTCF binding motif had no effect on the DNA methylation level. To test the effect in an endogenous setting, the authors generated a Rest^−/− mouse ES cell line and, as expected, observed that the REST binding regions were highly methylated. Most importantly, they found that the methylation levels at these sites were much reduced after reintroduction of wild-type Rest into the cells. Altogether, these results support a model whereby the binding of certain proteins can directly affect DNA methylation levels (FIG. 3).

Another study examined the methylation levels of hundreds of sequences that were individually inserted at the same genomic site in mouse ES cells⁸³. Using this approach, the contribution of various sequence motifs to methylation levels could be quantified. They found that CpG density showed a negative correlation with methylation level, which is consistent with the previously established view that CpG islands are generally unmethylated. Interestingly, when the sequences of binding motifs were altered, overall methylation levels decreased⁸³. This work suggests that protein binding has a general role in reducing DNA methylation levels, perhaps by preventing DNMT enzymes from gaining access to these sites, which is consistent with previous findings⁸⁴.

As TFs have no enzymatic activity to methylate or demethylate a CpG dinucleotide, a possible model would be that these proteins provide sequence-specific guidance and recruit methyltransferases or methylcytosine dioxygenases to these specific sites (FIG. 3). A recent study showed that the nuclear receptor PPARγ (peroxisome proliferator-activated receptor-γ) recruits TET1, resulting in a reduced methylation level around its binding sites through the interaction with TET1 (REF. 85). The reverse can also happen. DNMTs have been found to form protein complexes with various TFs or chromatin modification enzymes. For instance, Sato et al.⁸⁶ demonstrated that DNMT3A and DNMT3B interact with an orphan nuclear receptor, NR6A1 (nuclear receptor subfamily 6 group A member 1), and that this interaction induced the methylation of the OCT4 (also known as OCT3 and POU5F1) promoter carrying the NR6A1 binding site.

DNA methylation dictates protein–DNA interactions

It is well known that DNA methylation can affect the binding of some TFs⁸⁷. The manipulation of the methylation status of DNA sequences has been shown (mostly in in vitro studies) to result in the differential binding of TFs, including E2F, AP2α, MYC and MYN^88–95. Specifically, hypermethylation is often associated with a depletion of TF binding. Recently, a few studies examined the effect of DNA methylation on TF binding in vivo^96,97.

Using a gene-editing approach, Domcke et al.⁹⁷ generated a genetic deletion of three methyltransferases (Dnmt3a, Dnmt3b and Dnmt1) in a mouse ES cell line. A large number of novel binding sites for the TF NRF1 were created as a result of the triple knockout (TKO). These binding sites often correlated with novel DHSs in the TKO cells, which exhibited predominantly low methylation levels due to their generation in cells lacking DNMTs. Interestingly, novel NRF1 binding sites were hypermethylated in the wild-type cell line⁹⁷, suggesting that the removal of DNA methylation in TKO cells generated new binding sites for NRF1. These new binding sites had poor sequence conservation, indicating that these sites are non-functional in the wild-type background. In an earlier study, the same group showed that CTCF was able to reduce the DNA methylation level near its binding sites⁷⁸. In this work, the authors tested whether CTCF binding could affect NRF1 binding by reducing the methylation level of NRF1 binding sites. Reporter constructs harbouring an NRF1 binding motif and a CTCF motif were introduced into the ES cell line. Deletion of CTCF motifs within the construct led to hypermethylation and thus decreased NRF1 binding, suggesting that NRF1 binding in vivo depends on both methylation levels and co-occurring TFs, such as CTCF⁹⁷.

The examples above represent the two major mechanisms by which protein–DNA interactions and DNA methylation influence each other (FIG. 3). Of note, these two mechanisms are not mutually exclusive. In some cases, the two mechanisms have been found to coexist for the same TFs. For example, although CTCF is known to change the local methylation status⁷⁸, it has been shown that the binding of CTCF is also methylation sensitive⁹⁶. Finally, it is worth noting that the crosstalk between TF– DNA interaction and DNA methylation is not restricted to the TFs whose binding motifs contain CpGs. Changes in DNA methylation are often associated with chromatin status, resulting in increased or decreased DNA accessibility ^96,97. Differences in chromatin states will either create or eliminate TF binding sites and thus lead to differential TF binding. Although only approximately 25% of known TF binding motifs contain at least one CpG site⁹⁸, through such indirect crosstalk mechanisms, the binding of TFs without CpGs in their binding motifs could also be influenced by DNA methylation.

Activation or repression

Methylation in promoters is often considered the hallmark for gene repression⁹⁹. However, large-scale analyses of gene expression profiles and DNA methylomes have revealed that a substantial proportion of DNA methylation sites is positively correlated with gene expression³¹. This analysis was performed on methylation sites located within 300 bp upstream from transcriptional start sites, which raises the possibility that methylation in promoters could also be positively correlated with increased transcription of a target gene³¹ (FIG. 4). Of course, whether these methylation sites fall exactly within the regulatory elements and whether they are recognized by TFs remains to be tested. Single-gene studies have also demonstrated that DNA methylation can activate gene expression^47–49. For example, the sequence-specific DNA-binding protein RFX activates a methylated promoter⁴⁹. Interestingly, this protein was previously shown to bind to methylated DNA^47,51. Moreover, it was found that methylation at the 3′ end of the CpG island confers tissue-specific transcriptional activation during human ES cell differentiation¹⁰⁰.

A recent comparative study of mouse retina and brain explicitly explored the possible role of methylation sites whose methylation levels were positively correlated with gene expression¹⁰¹. Among the differentially methylated regions located within 4 kb upstream of transcriptional start sites, approximately 47% showed a positive correlation with the expression of their putative target genes. These methylation regions are overrepresented in DHSs and are evolutionarily conserved, suggesting that these sites are likely to be functional¹⁰¹. More importantly, a distinct set of sequence motifs was discovered in these regions, suggesting that some TFs bind preferentially to these regions¹⁰¹.

Pioneer TFs

The human genome is not made of linear, naked DNA strands. Instead, it is mainly organized into two forms. One is heterochromatin (or condensed chromatin), in which DNA sequences and histones are highly condensed, and genes in these regions are inactive. The other form is euchromatin (or open chromatin), in which DNA sequences are largely accessible to TFs, and genes in these regions can be activated^102,103. Chromatin organization is dynamic, and the different types of chromatin can change from one form to another during development or differentiation¹⁰⁴ (FIG. 4).

Pioneer TFs are a unique subset of TFs that drive these chromatin changes. A typical characteristic of pioneer TFs is their ability to bind directly to heterochromatic DNA and recruit other factors to change the status to euchromatin to initiate transcription^105,106. As DNA in heterochromatin is wrapped tightly around the nucleosomes and is often methylated, it is inaccessible to most TFs; pioneer TFs must possess special features to enable protein–DNA interactions. For example, a handful pioneer TFs (such as OCT4, SOX2 and KLF4) were shown to bind only partial motifs displayed on the nucleosome surface¹⁰⁷.

It could be speculated that the ability to bind mCpG sites might prove a useful property for pioneer TFs. If a pioneer TF can interact with an mCpG site, such an interaction would provide an anchor point for the pioneer TF to open up the closed chromatin. Indeed, we observed a large overlap between known pioneer TFs and proteins that bind to methylated DNA (for example, forkhead box protein A (FOXA) and GATA families, which are the best-studied pioneer factors)^{106,108–110}. Interestingly, several of their members showed the ability to bind to methylated DNA, including HOXA5, HOXA9, GATA3 and GATA4 (REF. 53). The mCpG-binding protein KLF4 was also shown to be a pioneer factor^105,111. Although there is no simple assay to identify pioneer TFs, evidence that TFs are able to bind methylated DNA would provide a short list of candidate pioneer TFs for future tests. Notably, as methylation-binding proteins participate in multiple biological processes, including gene regulation and splicing regulation, not all methylation-binding proteins are likely to be pioneer factors. Yet, binding to an mCpG site is just one approach for a pioneer TF to access condensed chromatin. Other pioneer TFs might have alternative approaches such as binding to partial motifs.

Splicing regulation

Historically, RNA splicing was considered to be regulated only at the post-transcriptional level. On the basis of this idea, DNA methylation was not expected to have any substantial role in splicing regulation. However, it is now well-established that splicing occurs co-transcriptionally, which means that DNA modification could influence RNA splicing. In one study, the authors observed that the binding of CTCF in an exon region created a roadblock for RNA polymerase II elongation and thus promoted the inclusion of the exon¹¹². Importantly, the binding of CTCF to the exon or intron was dependent on DNA methylation, suggesting that the methylation status surrounding the spliced exons could affect the inclusion level of these exons. Similarly, the mCpG-binding protein MeCP2 was found to play a part in regulating exon splicing¹¹³. In this case, a high methylation level led to MeCP2 binding to alternatively spliced exons, which resulted in exon inclusion (FIG. 4). The same trend was observed in a study of the brain methylome of honeybees¹¹⁴. A comparison of methylation levels between queen and worker bees revealed that intron-containing histone genes were highly methylated, whereas intronless histone genes were not methylated, suggesting that mCpG-binding proteins might play a part in splicing regulation. This observation is consistent with a global correlation analysis of DNA methylation and differential splicing events between the brain and the retina¹¹⁵. Although CTCF motifs were significantly enriched in differentially methylated regions associated with alternative splicing, other motifs were also enriched, suggesting that other TFs might also participate in splicing regulation. Interestingly, the methylation levels in some of the regions were positively associated with the inclusion level of the spliced exons, indicating that other mCpG-binding proteins are involved in regulating the splicing process.

Human diseases

Many studies have shown that aberrant DNA methylation is associated with various human diseases, including some types of cancer^19–21. For example, profiling the DNA methylation status in the promoters of 272 glioblastoma tumours showed that a distinct subset of samples displayed hypermethylation at a large number of loci, a phenotype termed ‘CpG island methylator phenotype’ (REF. 116). However, the mechanism by which the altered epigenetic state causes disease remains elusive. A recent study analysed the effect of methylation-dependent protein–DNA interactions on gliomas¹¹⁷. The IDH genes (IDH1 and IDH2) encode isocitrate dehydrogenases, and mutations in these genes are among the most frequent found in diffuse gliomas^118,119. Mutant IDH protein is a competitive inhibitor of hydroxylases, including the TET family of 5mC hydroxylases^120–122. As a result, the IDH mutation leads to a remodelling of DNA methylation profiles. Specifically, owing to the interference with TET family proteins, the mutation causes the CpG island methylator phenotype^116,123. IDH mutant gliomas have been shown to exhibit hypermethylation at CTCF binding sites, which leads to a reduction in CTCF binding; loss of CTCF in topologically associated domains removed the domain boundary and caused aberrant gene activation¹¹⁷.