The Methyl-CpG Binding Domain 2 and 3 proteins and formation of the Nucleosome Remodeling and Deacetylase complex

Abstract

The Nucleosome Remodeling and Deacetylase (NuRD) complex uniquely combines both deacetylase and remodeling enzymatic activities in a single macromolecular complex. The Methyl-CpG Binding Domain 2 and 3 (MBD2 and MBD3) proteins provide a critical structural link between the deacetylase and remodeling components while MBD2 endows the complex with the ability to selectively recognize methylated DNA. Hence, NuRD combines three major arms of epigenetic gene regulation. Research over the past few decades has revealed much of the structural basis driving formation of this complex and started to uncover the functional roles of NuRD in epigenetic gene regulation. However, we have yet to fully understand the molecular and biophysical basis for methylation-dependent chromatin remodeling and transcription regulation by NuRD. In this review, we discuss the structural information currently available for the complex, the role MBD2 and MBD3 play in forming and recruiting the complex to methylated DNA, and the biological functions of NuRD.

Introduction

The highly homologous Methyl-CpG Binding Domain proteins 2 and 3 (MBD2 and MBD3) are the ancestral orthologs to the larger MBD family found in all vertebrate species [1–3]. Members of this family share an approximately seventy amino acid domain that selectively recognizes methylated cytosine-guanosine dinucleotides (CpGs). Genetic duplication and diversification generated seven known family members (MBD1–6 and MeCP2), each of which independently associates with different co-regulatory complexes and have unique functional roles [2]. MBD2 and MBD3, which arose by duplication of the MBD2/3 protein found in invertebrates, provide critical structure to the Nucleosome Remodeling and Deacetylase (NuRD) complex [4–8]. NuRD comprises two separable functions consisting of histone deacetylase and chromatin remodeling sub-complexes which interact with MBD2 and MBD3 through distinct interfaces [7, 8].

Consequently, MBD2 and MBD3 physically bridge three major arms of epigenetic gene regulation to form NuRD: histone deacetylation, ATP-dependent nucleosome remodeling, and selective recognition of methylated DNA. Importantly, MBD2 and MBD3 form mutually exclusive NuRD complexes, each recognizing methylated DNA with markedly different affinity and selectivity [2, 9–12]. Therefore, MBD2-NuRD and MBD3-NuRD assemble functionally distinct complexes with high and low methylation selectivity, respectively. In this review, we discuss the formation of the NuRD complex focusing on the structures of the different components, the critical role provided by MBD2 or MBD3, how differences in methylation selectivity impact dynamic interaction with DNA, and we describe a model of NuRD function that attempts to reconcile these contrasting functional and biophysical data.

The NuRD complex

Early work in the Adrian Bird laboratory utilized methylated DNA probes to isolate two nuclear protein fractions that selectively bind methylated DNA, the methyl-CpG binding proteins 1 and 2 (MeCP1 and MeCP2) [13, 14]. MeCP1 was a large complex (400 – 800 kDa) that required multiple mCpGs for binding, while MeCP2 was a single polypeptide (52 kDa) that bound to individual methylated sites. Further characterization of MeCP2 led to the identification of the protein domain responsible for methylation selectivity and, by homology, the MBD family of proteins in mammals [2, 15]. A few years later, work by Paul Wade in the Alan Wolffe laboratory characterized a multi-protein nucleosome remodeling complex with methylation specificity [5, 6]. This complex became known as NuRD and likely represented the original MeCP1 protein complex identified in Bird’s laboratory [4]. At the same time the Danny Reinberg laboratory dissected the proteins comprising the histone deacetylase core complex of NuRD and co-expressed distinct combinations of these proteins to form stable sub-complexes [16, 17].

Together these and subsequent studies have shown that the core NuRD comprises at least seven proteins, each of which has multiple paralogs in humans (Figure 1) [5, 9, 17–20]. These proteins include Metastasis Tumor-Associated 1, 2, or 3 (MTA1, 2, or 3), Histone Deacetylase 1 or 2 (HDAC1 or 2), Retinoblastoma Binding Protein 4 or 7 (RBBP4 or 7), GATAD2A or B, Chromodomain Helicase DNA Binding Protein 3, 4, or 5 (CHD3, 4, or 5), Cyclin-Dependent Kinase 2 Associated Protein 1 (CDK2AP1), and MBD2 or 3 [5, 9, 16, 17].

Figure 1. Overview of the NuRD complex.

(a) Schematic diagrams of the seven major NuRD components highlight the domain architecture and relative size of each protein: CDKAP1, MBD2 or 3, RBBP4 or 7, HDAC 1 or 2, GATAD2 A or B, MTA 1, 2, or 3, and CHD 3, 4, or 5. (b) A model of NuRD depicts how these components interact.

• HDAC1, HDAC2, RBBP4 and RBBP7 proteins each contain a single structured domain and function in a variety of co-regulatory histone-modifying complexes (e.g., Sin3a, NuRF, and CAF-1). HDACs provide histone deacetylase enzymatic activity while RBBPs present protein-protein interaction surfaces involved in the recruitment of the complex to chromatin.

• MTA1, MTA2, and MTA3 proteins contain multiple domains separated by disordered stretches of the polypeptide. Different regions and domains of the MTAs make distinct protein-protein interactions with HDAC and RBBP paralogs [21–26]. Therefore, MTAs provide a critical structural role in forming the histone deacetylase core sub-complex (HDCC) of NuRD comprising the MTA, HDAC, and RBBP proteins.

• GATAD2A and GATAD2B proteins contain two relatively small conserved regions implicated in protein-protein interactions separated by a long, likely unstructured, polypeptide [27]. These two conserved regions of the GATAD2s from critical interactions between the MBD and CHD proteins [8, 23].

• The largest proteins in the complex, CHD3, CHD4, and CHD5 contain an Snf2-type helicase/ATPase domain that physically moves nucleosomes in an ATP dependent manner [5, 16, 28]. The CHDs have two chromodomains and two plant homeodomains (PHD) that interact with DNA and histones, respectively [29, 30]. CHDs also have an uncharacterized C-terminal domain and an N-terminal High Mobility Group (HMG) Box-like domain that binds poly-ADP ribose [31].

• MBD2 and MBD3 proteins contain an N-terminal MBD and C-terminal coiled-coil domain separated by an intrinsically disordered region implicated in binding to the HDCC [4, 7, 8, 27]. The coiled-coil domain of MBD2 and MBD3 binds a GATAD2 protein, thereby linking the HDCC with the CHD proteins.

• CDK2AP1 appears to be a core component of NuRD but often overlooked because of its small size (~12 kDa). Studies indicate that CDK2AP1 negatively regulates CDK2 activity, inhibits tumor growth, and acts as a tumor suppressor in oral cancer [9, 18, 32–34]. While its function within the NuRD complex has not been clearly defined, a recent study revealed that CDK2AP1 promotes NuRD recruitment to specific loci where it displaces SWI/SNF and represses transcription [35].

The presence of multiple paralogs for each core component leads to a sizeable combinatorial array of potentially distinct NuRD complexes. Experiments indicate that at least some of the paralogs interact with NuRD in a mutually exclusive manner and are expressed in a cell and tissue-specific manner with unique functional roles [9, 11, 36–39]. For example, the MTA3 paralog is expressed in a specific subset of B lymphocytes and promotes germinal center differentiation [40, 41]. The CHD3, CHD4, and CHD5 paralogs are expressed at different stages during cortical brain development in mouse and regulate distinct functions [42]. The GATAD2A paralog contains a Zinc Finger MYND-Type Containing 8 (ZMYND8) binding motif such that GATAD2A-NuRD localizes to sites of DNA damage [39].

In addition to the core components, other proteins bind to the complex at sub-stoichiometric ratios, such as spalt-like transcription factor 4 (SALL4), friend of GATA (FOG)- 1 and 2, B-cell lymphoma/leukemia 11a (BCL11a), ZMYND8, and several zinc finger proteins (ZNF-512B, −532, −592, and −687) [19, 20, 39, 43]. Several of these components interact with NuRD through a small (~12 amino acid) N-terminal peptide that binds to RBBP4/7 and was first identified in the FOG-1 protein [44–48]. This peptide can be used to isolate relatively pure NuRD complexes from cell extracts in a single step [45, 49], indicating that the interaction interface is uniquely accessible when the RBBPs are bound to NuRD. The sub-stoichiometric components can localize NuRD to different regions of the genome for distinct functions [39, 44, 47, 50–53]. Therefore, NuRD appears to act as a co-regulatory complex that can be targeted to different regions of the genome through either its intrinsic binding affinity for methylated DNA or co-recruitment by different transcription factors and chromatin-associated proteins.

Structural Analyses of NuRD

The variety of distinct NuRD complexes containing different isoforms with long disordered regions have challenged investigations of complex formation. As a result, structural studies have focused on well-defined sub-complexes between separate components. We determined one of the first structures of a NuRD sub-complex formed between the coiled-coil domains of GATAD2A (also known as p66α) and MBD2 [7, 54]. This complex involves two small helical peptides that form a very stable high-affinity interaction critical to NuRD function. Each of the coiled-coil domains behaves as monomeric helices in isolation, which is unusual for coiled-coil domains. The charge distribution of the domains inhibits homodimerization and stabilizes the heterodimeric interaction (Figure 2a,b) [54]. Importantly, we showed that enforced expression of the isolated coiled-coil domain from GATAD2A inhibits methylation-dependent gene silencing by NuRD. This same peptide can immuno-precipitate all native components of NuRD except CHD4 and GATAD2A. Hence, the coiled-coil interaction forms a necessary and sufficient link between the histone deacetylase and chromatin remodeling portions of NuRD. Of note, the coiled-coil domains of the MBD2/3 and GATAD2 proteins are conserved throughout Metazoan species, independent of the presence of DNA methylation. Therefore, this coiled-coil interaction and the NuRD complex emerged with the earliest multicellular organisms and has been maintained throughout animal evolution [1]. More recent work has shown that the second highly conserved region of the GATAD2 proteins interacts with the C-terminal portion of CHD4, thus delineating the minimal protein regions necessary for the structural link between NuRD and CHD3/4 [23, 55].

Figure 2. Structural analyses of the MBD2:GATAD2A and MTA1:HDAC1 complexes.

(a) A cartoon diagram depicts the solution structure of the coiled-coil complex between MBD2 (cyan) and GATAD2A (blue) (PBD ID: 2L2L). Amino acids at the protein-protein interface are shown as sticks. (b) A surface rendition of MBD2-CC and GATAD2A colored by electrostatic potential shows alternating patches of positive (blue) and negative (red) regions that promote heterodimerization and minimize homodimerization. (c) A cartoon diagram depicts the crystal structure of the complex between MTA1 (red and light red), HDAC1 (brown and light brown), and a peptide inhibitor (magenta). MTA1 forms a dimer and wraps around HDAC1. An inositol tetraphosphate molecule (space-filled, yellow) binds near the HDAC1 active site. (d) Surface rendering of HDAC1 and MTA1 colored by electrostatic potential shows that IP₄ binds to positively charged regions (blue) of each, thereby bridging between surfaces that otherwise would not interact favorably (PDB ID: 5ICN). All structure figures were generated using PyMOL [130] and the electrostatic surface potential using the APBS plugin [131].

A breakthrough in structural analysis of NuRD came from studies of the histone deacetylase core by John Schwabe’s group. They determined high-resolution structures of a complex between the ELM2-SANT domains of MTA1 and HDAC1 [24, 25, 56]. This work built on prior research from the same group investigating the complex between homologous domains from the SMRT co-repressor and HDAC3 [57]. The MTA1-HDAC1 interaction involves an extensive protein-protein interface formed by ELM2 and SANT domains of MTA1 as well as an extended region between these two domains that wraps around the surface of HDAC1 (Figure 2c). Their structural analyses led to two interesting observations. First, the SANT domain of MTA1 binds to HDAC1 through inositol tetraphosphate (Ins(1,4,5,6)P₄). The phosphate groups from Ins(1,4,5,6)P₄ favorably interact with positively charged surfaces at the interface, bridging an otherwise repulsive interaction (Figure. 2d). The contact surface on HDAC1 is near the enzymatic cleft such that the presence of Ins(1,4,5,6)P₄ augments deacetylase activity [56]. Also, previous work has suggested that Ins(1,4,5,6)P₄ levels fluctuate during the cell cycle [58]. This observation raises the interesting possibility that Ins(1,4,5,6)P₄ regulates histone deacetylation activity in a cell cycle-dependent manner. Second, the ELM2 domain forms a homodimer within the crystal lattice leading to a 2:2 molar ratio for the histone deacetylase core. Dimerization through the ELM2 domain helps clarify stoichiometric analyses of the full NuRD complex. While not wholly consistent, quantitative mass spectrometry and protein cross-linking experiments have found that a single CHD interacts with 1–2 copies of the MBD, GATAD2, and HDAC proteins, 2–3 copies of the MTA proteins, and at least four copies of the RBBP proteins [19, 20, 59]. These results, along with the structural studies, suggest an overall stoichiometry of 1:2:4 between the CHD/GATAD2/MBD/CDKAP1:MTA/HDAC:RBBP components.

Several studies have delineated the specific protein interactions involving the RBBP proteins. These proteins comprise a single WD-repeat domain, which forms a seven-bladed β propeller fold, with an extended N-terminal α-helix. This 3-dimensional fold generates a large surface area which allows RBBP4 and RBBP7 proteins to interact with different chromatin-associated proteins through independent interaction surfaces. Joel Mackay’s group mapped and solved the first structure of a complex between MTA1 and RBBP4 (also known as RbAp48) [21]. This interaction involves a conserved positively charged motif (KRAARR) from MTA1 (amino acids 678–683)^* that binds in a groove formed between non-canonical structural elements from the N-terminal α-helix, an extended loop, and a C-terminal helical turn of RBBP4 (Figure 3). The MTA1 peptide occupies the same binding site as a homologous sequence from Histone H4 [60], so that interaction with MTA1 precludes binding to Histone H4. In subsequent work, the Mackay laboratory showed that a separate region in MTA1 (amino acids 440–550), contains a similar motif and binds RBBP4 [22]. At the same time, the Schwabe group determined the structure of this second interaction showing that it occupied the same binding site of RBBP4 as the more C-terminal peptide, but with a more extensive interface [26]. Unlike the more C-terminal peptide, this central region of MTA1 folds into three α-helices and a β-strand that wrap around RBBP4.

Figure 3. Binding of MTA1 and FOG1 to RBBP4.

A cartoon diagram depicts an overlay of crystal structures showing RBBP4 bound to four different peptides. Two different peptides from MTA1 (amino acids 462–546 in yellow and 670–695 in cyan; PDB ID: 5FXY and 4PBZ, respectively) bind RBBP4 (green) through a common interface. Likewise, the N-terminal helix from histone H4 (amino acids 24–41 in blue; PDB ID: 3CFV) binds to the same region of RBBP4. In contrast, the N-terminal FOG1 peptide (amino acids 1–15 in magenta; PDB ID: 2XU7) binds to a distinct region of RBBP4 that does not overlap with the binding surfaces for MTA1 or histone H4. Note that MTA1 residues not observed in the crystal structure (519–528) are indicated with a yellow dashed line.

To extend the crystallographic analyses of the RBBP-MTA peptide interaction, several groups have investigated larger complexes involving the MTA, HDAC, and RBBP proteins. First, the Schwabe group used a combination of small-angle x-ray scattering and single-particle negative-stain electron microscopy to study a complex formed between HDAC1, MTA1 (residues 162–546), and RBBP4 [26]. The MTA1 construct includes the ELM2-SANT region, adjacent zinc finger, and one of the two regions that bind RBBPs. This work showed that the ELM2-SANT region forms a dimer at the center of the complex with the associated HDAC1 proteins immediately adjacent to that center. The two RBBP proteins extend further away on opposite sides, while making contact with the HDAC1 proteins. In this model, the HDAC1 enzymatic cleft remains exposed to solvent and available for interaction with an acetylated histone tail. The MacKay group studied the complex between MTA1 (residues 449–715) and RBBP4 using chemical crosslinking analyses and single-particle negative-stain electron microscopy [22]. This MTA1 construct starts just after the ELM2-SANT domains and includes both of the RBPP binding motifs. The results confirmed that two copies of RBBP4 could bind MTA1 simultaneously. More recently, Brasen et al. [61] purified a complex between full-length MTA2 and RBBP7 for analysis by negative-stain electron microscopy. Two RBBP7 molecules extend from either side of the centrally located ELM2-SANT dimer to form an extended structure containing four RBBPs and two MTAs. The BAH domains protrude from the center of the complex adjacent to the ELM2-SANT domains. Together, these studies have shown how each MTA can homodimerize and recruit two independent copies of the RBBP proteins, for a total of four RBBP proteins per NuRD complex. This multiplicity also agrees with stoichiometric analyses [19, 20] and allows the RBBPs in NuRD to bridge multiple histones and simultaneously interact with transcription factors and other co-regulatory proteins.

As discussed previously, RBBPs bind to a small N-terminal peptide found in different transcription factors that recruit NuRD. Crystal structures have been determined for a number of these peptides bound to the RBBPs (BCL11A [46], SALL4 [48], AEBP2 [62], PHF6 [63] and others), all of which show a similar interaction as between FOG-1 and RBBP4 [45]. As shown in Figure 3, the FOG-1 peptide (amino acids 1–15) binds across one of the flat axial surfaces of the RBBP4 β-propeller, making multiple specific interactions. The conserved RRK motif from FOG-1 interacts with negatively charged glutamate residues at the surface, with the middle arginine protruding into the central cavity of RBBP4. Importantly, the peptide binding surface on RBBP4 does not overlap with that of MTA1, which should permit the RBBPs to bind both MTA and a transcription factor concurrently.

In addition, structural work by Schmitges et al. [64] showed that a similar peptide sequence from the N-terminal tail of histone H3 binds to the same region of Nurf55, the RBBP homolog from D. melanogaster. They studied this interaction in the context of the polycomb repressive complex 2 (PRC2) which contains an RBBP protein as a core component. Importantly, they found that lysine 4 of histone 3 (H3K4) makes a critical intermolecular contact, such that H3K4 methylation blocks interaction with Nurf55 and inhibits methylation of lysine 27 of histone 3 (H3K27) by PRC2. Together, these studies show that the RBBPs function as a handle for recruitment of NuRD by transcription factors and chromatin-interacting proteins and bind to nucleosomes that lack H3K4 methylation.

The CHD3, 4, and 5 proteins are the largest components of the NuRD complex (~110 kDa – Figure 1); however, only a few published studies have investigated the structure of these proteins. Solution structures of the PHD domains in isolation and bound to histone peptides reveal a canonical PHD fold in which the histone peptide binds to form a third β-strand on the central sheet of the domain [29, 65]. Affinity analyses show that the second PHD domain preferentially binds trimethylated lysine nine in the histone H3 peptide (H3K9me3) while the first PHD domain does not. However, both PHD domains favorably bind peptides with unmodified lysine four (H3K4). Work by Tatiana Kutateladze’s group showed that the separation between the dual PHD domains permits the CHDs to bridge neighboring nucleosomes in chromatin [30]. They found that gene silencing by NuRD requires this bivalent interaction with chromatin. This work provides tantalizing insight into how the CHD proteins from NuRD could establish an orderly array of nucleosomes, where bridging by the PHD domains could help establish compact spacing between nucleosomes. The Mackay group recently determined the structure of an N-terminal HMG-like domain from CHD4 [31]. They found that this domain binds poly-ADP ribose with high affinity, yet the role of this interaction in chromatin remodeling or DNA damage repair [39] remains unclear. At this point, the published data describing the structure of the CHDs and their interaction with nucleosomes has been limited to these few studies, which precludes a more detailed understanding of how the protein interacts with chromatin and an accurate model of CHD function.

Based on the above structural studies, a picture of the overall NuRD complex emerges. The MTAs form the structural backbone of the HDCC and provide a handle for recruitment to chromatin through binding of multiple RBBPs. The GATAD2 proteins bridge the CHDs and MBDs through two small conserved domains separated by a long intrinsically disordered region. The missing piece to this puzzle is how the MBD/GATAD2/CHD remodeling portion interacts with the MTA/HDAC/RBBP deacetylase portion of the complex. Based on our observation that the isolated coiled-coil domain from GATAD2A interacts with a sub-complex consisting of MBD, HDAC, MTA, and RBBP proteins, excluding native GATAD2A and CHD4 [7], we hypothesized that the MBD proteins would provide this missing connection. Importantly, NuRD appears to be present across all animal species; yet, in invertebrate species, the MBD2/3 proteins do not always contain an MBD domain. This latter observation implies that the MBD domain is not necessary for binding to other components to form NuRD.

Therefore, we anticipated that the region between the N-terminal MBD and the C-terminal coiled-coil likely binds to the HDCC comprising the MTA, HDAC, and RBBP proteins. Based on NMR and circular dichroism studies, we found that this central portion of MBD2 (amino acids 238–356) behaves as an intrinsically disordered region (IDR) in isolation (Figure 4a–c). Immune-precipitation studies showed that the IDR was necessary and sufficient to bind the HDCC [8]. Mutating highly conserved residues across the MBD2-IDR revealed that modifying two consecutive residues (R286E/L287A) disrupts binding to the HDCC and blocks methylation-dependent gene silencing by NuRD [66]. These observations lead to the current model depicted in Figure 1 which the MBDs play a critical structural role linking the GATAD2/CHD4 and HDCC subcomplexes.

Figure 4. The intrinsically disordered region and DNA binding domains of MBD2.

(a) A schematic diagram of MBD2b (amino acids 214–361) highlights the large intrinsically disordered region located between the N-terminal MBD and C-terminal coiled-coil domain. (b) The intrinsically disordered region (IDR) binds to the HDCC (brown), while mutating two consecutive residues (R286E/L287A, red) disrupts this interaction. (c) An alignment of a region from PWWP2A that binds to the HDCC shows that this peptide shares homology with the MBD2 IDR. This region includes the double-mutation (red) that disrupts MBD2 binding to to the HDCC. (d) A mixed diagram depicts the solution structure of MBD2 (cyan) bound to an mCpG dinucleotide (sticks) (PDB ID: 2KY8). MBD2 makes three key interactions with DNA through a conserved RRY motif. (e) A closeup view shows that arginine residues 24 and 46 each hydrogen bond to symmetrically related guanine bases and pack against methyl groups (orange spheres) of the adjacent methyl-cytosine bases. While tyrosine 36 positions its hydroxyl group towards methyl-cytosine and likely binds to structured water surrounding the methyl group.

More recent work by Zhang et al. [67] and Link et al. [68] found that the PWWP2A protein binds to the HDCC sub-complex and excludes the MBD, GATAD2, and CHD proteins of NuRD. The two groups mapped this interaction to slightly different but contiguous regions of PWWP2A. Interestingly, Zhang et al. [67] found that a small peptide from PWWP2A (amino acids 311–331) bound to a purified MTA1-HDAC1 complex. This peptide is homologous to the MBD2-IDR region containing R286-L287 (Figure 4c). Hence, PWWP2 and MBD2 competitively bind to the HDCC and appear to use a similar peptide motif in this interaction. Recently, we have shown that this same region of MBD2 shows intrinsic helical propensity and that the R286E/L287A mutation reduces this propensity [66]. These data suggest that the MBD2-IDR folds into α-helices upon binding to the HDCC. However, the structure of MBD2 or PWWP2 bound to the HDCC has yet to be determined to verify this model of interaction.

Ernest Laue’s group studied the formation of NuRD from D. melanogaster (dNuRD) using both pull-down assays of endogenous complexes and recombinant expression of sub-complexes in insect cells [69]. They demonstrated that the Drosophila homologs of the RBBP, MTA, HDAC, and MBD proteins formed a stable deacetylase sub-complex with a 5:2:2:1 stoichiometric ratio. This sub-complex could be purified with or without the MBD component. In contrast, the CHD4, GATAD2A, and DOC1 homologs were present in the endogenous dNuRD at sub-stoichiometric ratios and appeared to only weakly associate with histone deacetylase sub-complex. Using single-particle tracking, they showed that association of CHD4-like with dNuRD required the GATAD2A homolog. These observations further support that NuRD can be divided, both functionally and structurally, into chromatin remodeling and histone deacetylase sub-complexes. The instability of the full dNuRD may reflect the relatively low affinity of the coiled-coil interaction between MBD2 and GATAD2A homologs in Drosophila (K_D ~ 7 μM) as compared to the human complex (K_D ~ 50 nM) [1].

Methylation-specific targeting of NuRD

The NuRD complex was first described as selectively binding methylated DNA, and biochemical and cellular studies have supported that observation. However, methylation selectivity of the complex depends on which MBD it contains. Initially, MBD3 was considered a core component and MBD2 secondary. However, work by the Hendrik Stunnenberg laboratory showed that MBD2 and MBD3 form distinct and mutually exclusive complexes [9]. This finding has been well supported by subsequent quantitative mass spectrometry analyses [19, 20] and by genomic localization studies [70–72]. Also, mammalian cells express multiple isoforms of MBD2 and MBD3 [2]. MBD2a, the most abundant isoform of MBD2, contains a long (approximately 150 amino acid) N-terminal region only found in mammalian species. This region contains a positively charged glycine-arginine (GR) repeat that likely contributes to DNA binding and recruits additional cofactors (PRMT5 and MEP50) to NuRD [9]. However, the GR region is not required for methylation-dependent gene silencing by MBD2. The MBD2b isoform, which lacks the N-terminal extension, shares 76% identity with MBD3a isoform. Both isoforms contain an N-terminal MBD (77% identity) and C-terminal coiled-coil (85% identity) domains separated by the intrinsically disordered region (76% identity).

Embryonic stem cells express additional unique isoforms of MBD2 and MBD3. The MBD2c isoform is a splice variant that lacks the IDR and coiled-coil domains. MBD2c binds methylated DNA but does not recruit the NuRD complex, acting as a dominant-negative inhibitor of methylation-dependent gene silencing by NuRD [73]. The MBD3b isoform lacks a full MBD yet retains the IDR and coiled-coil domains [74]. Therefore, the MBD2c and MBD3b isoforms found in embryonic stem cells appear to disrupt the association of NuRD with methylated DNA. The precise role of the two proteins during embryonic development and pluripotency, though, remains an active area of investigation and hotly debated [74–78]. Early studies indicated that MBD3 promotes lineage commitment [74], while more recent research has indicated that MBD3 either promotes or blocks pluripotency and reprogramming [76, 77]. These confounding results likely reflect, at least in part, lack of a detailed molecular model for gene regulation by NuRD complexes.

After differentiation and development, most adult tissues express MBD2a, whereas only testis and brain tissues continue to express MBD3 [79]. This pattern of expression implies that methylation-independent functions of NuRD dominate in undifferentiated tissues while methylation-dependent functions play an essential role during cellular differentiation. However, until we understand the physical connection between NuRD recruitment and changes in gene expression, the biological phenotypes arising from manipulating NuRD expression are difficult to explain fully.

The MBD selectively recognizes symmetric mCpG dinucleotides

The structure of the DNA binding domain of MBDs bound to a variety of DNA substrates have been solved by x-ray crystallography and nuclear magnetic resonance spectroscopy [12, 80–88]. This domain comprises a 3–4 stranded β-sheet, a single α-helix, and a C-terminal loop. The sheet forms the base of the protein with a finger-like projection of two central β-strands that extends down the major groove of DNA (Figure 4d). The helix and C-terminal loop pack against this sheet to form the hydrophobic core of the domain. A conserved structural motif formed by two arginine and tyrosine residues protrudes from the base of the domain to interact with DNA. As depicted in Figure 4e, the two arginine sidechains form bidentate hydrogen bonds with symmetrically related guanosine bases of the CpG. This arrangement allows the arginine sidechain to form a stair motif cation-π interaction with the adjacent methyl-cytosine base [89]. The aliphatic portion of the arginine sidechains packs against the methyl groups of the symmetrically related methyl-cytosines (mC), which together with the cation-π interaction provides approximately 5-fold selectivity for mCpG [10, 12, 86]. The hydroxyl group of the conserved tyrosine sidechain points towards the methyl group of one mC and hydrogen bonds with structured water that surrounds this methyl group. The tyrosine hydroxyl provides the bulk of methylation selectivity, contributing an additional 10–20-fold binding preference for mCpG over CpG [10, 12, 85, 86]. Hence, the combination of these favorable interactions involving both the arginine and tyrosine residues leads to a net 100-fold selectivity for symmetrically methylated mCpGs.

The mC-arginine-guanosine interaction motif has been identified across a much larger group of DNA binding domains that recognize methyl-cytosines in one strand of DNA [90]. Notably, though, thymine has a methyl group in the major groove of DNA analogous to that of mC. Hence, the mC-arginine-guanosine triad can be replaced by thymine-arginine-guanosine, especially for those transcription factors that recognize the mC in only one strand of DNA [90]. In contrast, the symmetry of the two arginine residues in an MBD drives selectivity for symmetrically related guanosines, which, because of base-pairing, are found only in cytosine-guanosine dinucleotides. While other DNA binding domain can recognize one or more methyl-cytosines, the MBDs uniquely recognize the symmetrically related methyl-cytosines in mCpG dinucleotides.

Studies have shown that MBDs can adapt to alternative binding sequences that include a thymine-guanosine dinucleotide. For example, MeCP2 can bind non-CpG sequences such as mCAC which accumulate in neuronal tissues [91, 92]. These alternative binding sequences replace the second guanosine with adenosine, which cannot form a bidentate hydrogen bond with arginine. Therefore, one arginine residue must interact with the phosphate backbone or adjacent DNA bases to accommodate these alternative binding sequences. Nonetheless, MBDs generally show the highest binding affinity for mCpG dinucleotides and cellular studies have found that MBDs localize to methylated CpG islands [10, 71, 91]. Whether mCs in a non-CpG context are functional targets for MBDs, MeCP2 in particular, remains an exciting and active area of investigation.

MBD2 vs. MBD3

The mCpG DNA binding motif (RRY) is highly conserved across the MBD family of proteins, modifications of which predict a lack of methylation specificity. Unlike vertebrates which have both an MBD2 and MBD3, invertebrate species contain only a single MBD2/3 isoform. In species that show evidence of DNA methylation, the MBD2/3 retains the critical RRY motif and selectively binds mCpGs [1, 93–98]. In contrast, those species that lack DNA methylation, such as D. melanogaster, typically have MBD2/3 proteins with modifications that either eliminate or modify this motif [1, 99–101]. In either case, MBD2/3 retains both the IDR and coiled-coil domains critical to forming NuRD, which suggests that NuRD has a critical role in all multicellular organisms that is independent of DNA methylation.

The duplication event that generated MBD3 led to the replacement of the conserved tyrosine by phenylalanine in almost all vertebrate species studied to date (with the notable exception of X. laevis) [2, 3, 6]. This modification contributes to a reduction in overall DNA binding affinity and substantial loss of methylation specificity, which suggests sub-specialization of the complex in vertebrate biology. This sub-specialization correlates with a significant change in the pattern and extent of CpG methylation, and with the development CpG islands associated with gene promoters [93, 102–104].

In studying the MBDs bound to DNA by NMR, we have noted that several backbone and sidechain resonances show substantial chemical-shift differences between methylated and unmethylated DNA substrates. These chemical-shift differences are consistent across all MBDs we have studied to date, including MeCP2, MBD2, MBD4, and an invertebrate MBD2/3 from the freshwater sponge Ephydatia muelleri [1, 12, 85, 86, 92]. Importantly, we found that even MBD3 shows chemical-shift differences for these same resonances, moving in the same direction but not to the same extent when bound to methylated DNA [12]. We demonstrated that these shifts in MBD3 reflect rapid exchange between methylation-specific and non-specific binding modes. Consequently, peak positions can be used to calculate the fraction of MBD3 localized to the mCpG site. Using this information, we showed that the fraction bound to the methylated site depends on the density and relative affinity for mCpGs. Therefore, MBD3 shows structural recognition and preference for a methylated site even though it binds methylated DNA with significantly decreased overall affinity. In addition, we found that MBD3 did not preferentially bind hydroxymethylated DNA, showing identical chemical shifts for both hydroxymethylated and un-methylated substrates. This latter observation, along with binding affinity analyses by several groups [10, 12], provide strong structural evidence that MBD3 does not localize the NuRD complex to hydroxymethylated DNA through its MBD, although previous research has suggested otherwise [105].

These observations suggest to us that the behavior of an MBD when bound to DNA is more relevant to function than the absolute binding affinity and selectivity. Both MBD2 and MBD3 function to direct NuRD complexes to specific regions in the genome to modify chromatin compaction through nucleosome remodeling and deacetylation. If the MBD restricts the mobility of NuRD due to dense methylation of a CpG island, it should modify the ultimate positioning and density of the remodeled nucleosomes. To address this hypothesis, we have been studying the behavior of MBDs on methylated and unmethylated substrates using single-molecule approaches.

MBD2 sliding and bending DNA

In recent work, we developed an in vitro substrate to study MBD2 and MBD3 binding to DNA by atomic force microscopy and single-molecule fluorescence [106]. We inserted a large fragment of the DAPK1 CGI, a known target for MBD2 [107], into a vector completely devoid of CpGs. A surprising result came from atomic force microscopy analysis of MBD2 bound to this substrate. MBD2 induced an acute bend of DNA, especially for methylated CpG rich substrates. This bending has not been observed by NMR or x-ray crystallography studies of MBD2 bound to small DNA fragments, which suggests that it requires a more extended substrate. While the biological significance of this bending is not clear, the correlation between bending angle and methylation of DNA raises the possibility that it contributes to chromatin compaction and gene silencing.

Studies of MBD2 bound to DNA tightropes constructed from the same substrate show that MBD2 rapidly diffuses across unmethylated DAPK1 CGI, but when methylated MBD2 binds statically for a remarkably long time (1–2 minutes). Hence, methylation greatly restricts the dynamic mobility of MBD2 on a large scale, even though NMR studies indicate relatively rapid exchange between methylated sites separated by ~10 base pairs [86]. Ongoing studies with MBD3 show that it diffuses rapidly across CpG rich DNA even when fully methylated. The latter correlates with the rapid exchange between methylated and unmethylated binding modes we observed by NMR studies on small DNA substrates [12]. Together, these results fit with cellular studies which have shown that both MBD2 and MBD3 localize to unmethylated CpG rich regions within the genome while MBD2 more exclusively localizes to methylated CpG rich regions [37, 70, 71]. Knockdown of MBD2 leads to increased expression of associated genes while knockdown of MBD3 correlates with lower expression levels. Therefore, MBD2-NuRD is strongly associated with methylated DNA and silencing of gene expression, while MBD3-NuRD is associated with unmethylated CpG islands and can be associated with active transcription.

The biological function of NuRD

Despite the extensive structural and biophysical analyses over the past decade, the biological role of NuRD remains enigmatic. Early studies showing that MBD2-NuRD selectively binds methylated DNA, contains histone deacetylase activity and is associated with gene silencing lead to the conclusion that NuRD functions solely as a transcriptional repressor. However, more recent studies have begun to question that conclusion.

Several groups have determined the localization of MBD2- and MBD3-NuRD complexes and correlated their localization with gene expression. In studies by Rainer Renkawitz’s group, a green fluorescent protein (GFP) fused MBD2 localized to highly methylated sequences and was associated with repressed genes [37]. In contrast, GFP-MBD3 localized to unmethylated CpG rich regions and was associated with expressed genes, although the number of MBD3 binding sites was low (490 in total) as compared to MBD2 (8200 sites). Also, they found that LacI fused MBD2 dramatically compacted a LacO array in living cells, whereas LacI-MBD3 did not. Paul Wade’s group identified a much larger number of MBD3 binding sites in breast cancer cell lines (9310 sites in MCF-7 cells) using ChIP-seq of endogenous MBD3 [70]. Similarly, they found that MBD3 localized to active CpG rich promoters and was associated with a reduction in nucleosome occupancy near the transcriptional start site.

Finally, Stunnenberg’s group stably expressed a tagged MBD2 to study localization by CHIP-seq in the MCF-7 breast cancer cell line [71]. They found that MBD2 localized to heavily methylated and CpG rich loci, the majority of which were within promoters and exons. Interestingly, they identified a subset in which MBD2 localized downstream of active promoters.

More recently, Brian Hendrich’s laboratory developed a novel inducible MBD3 system in embryonic stem cells to follow MBD3 localization and chromatin remodeling over time [108]. They found that MBD3 rapidly induces nucleosome remodeling within a few hours of induction, which altered local transcription factor binding and gene expression. They could show by micrococcal nuclease digestion that MBD3 locally repositioned nucleosomes. Importantly, whether MBD3 induced remodeling increased or reduced transcription depended on the genetic context. The gene responses were relatively modest (2–3 fold), indicating that MBD3-NuRD functions to dampen or augment expression. Together, these studies show that MBD2-NuRD and MBD3-NuRD both localize to CpG rich regions of the genome, but that MBD2 dominates at methylated regions. The NuRD complex primarily functions by altering nucleosome positioning, while the biological effects depend on the local context and method of recruitment [53, 109–111].

NuRD complex and fetal hemoglobin regulation

The role of DNA methylation in the transcriptional switch from fetal to adult hemoglobin expression has been studied for more than 30 years [112–118]. Inhibition of DNA methylation increases the expression of fetal hemoglobin in adult erythroid cells, suggesting that methylation contributes to gene silencing. However, the γ-globin locus in humans does not contain a CpG island, indicating that silencing involves methylation of only a few CpGs or that methylation indirectly regulates γ-globin expression. In studying methylation-dependent gene silencing of globin, the laboratory of Gordon Ginder isolated an MBD2-NuRD complex that bound to the embryonic globin locus in primary chicken erythrocytes and contributed to methylation-dependent gene silencing [117, 119]. Importantly, disruption of MBD2-NuRD in mouse models of human β-globin gene regulation [117] as well as primary human cells and cell lines [66, 120] leads to a dramatic increase in the expression of fetal hemoglobin. This gene response exceeds the typical 2–3-fold changes associated with genetic disruption of NuRD, suggesting that the NuRD complex plays a particularly important role in globin regulation.

Furthermore, work by Stuart Orkin’s group has shown that Bcl11A binds directly to the γ-globin locus and recruits NuRD to silence gene expression [121–125]. Similarly, Takahiro Maeda’s group found that the ZBTB7A transcription factor (also known as LRF) binds to the γ-globin locus and recruits NuRD to silence gene expression [50]. Therefore, disrupting NuRD formation appears to be a potential strategy to restore fetal hemoglobin expression for therapy of sickle cell anemia and other β-hemoglobinopathies.

Recently, Daniel Bauer’s group used CRISPR screen to identify potential targets within the NuRD complex [55]. The results of those studies showed that the proteins and domains involved in forming NuRD are critical for fetal hemoglobin silencing. Both histone deacetylase and chromatin remodeling functions of the complex are necessary for gene silencing. Also, recent work by Ginder’s group showed that mutations of MBD2 that disrupt recruitment of either the HDCC or CHD4 portions inhibited silencing of fetal hemoglobin [66]. Therefore, fetal hemoglobin silencing requires a functional and complete NuRD complex.

A model of MBD2 and MBD3 function within NuRD

The structural and biological experiments described above raise the possibility that the functional differences between MBD2-NuRD and MBD3-NuRD reflect the behavior of the MBDs on densely methylated and unmethylated CpG islands [106]. As depicted in Figure 5, both complexes freely remodel nucleosomes over unmethylated CpGs. In contrast, MBD2-NuRD restricts mobility and hence drives accumulation of nucleosomes on methylated CpG islands, while MBD3-NuRD does not. Therefore, the two complexes have opposing actions on methylated promoters and enhancers. This model fits with genomic localization studies [37, 70, 71] as well as the expression patters of the two proteins [79]. Embryonic stem cells express MBD3 where it can remodel chromatin independent of DNA methylation status [79, 126, 127]. During the differentiation of embryonic stem cells, a subset of tissue-specific genes become methylated and silenced [128, 129]. Hence, MBD2 expression correlates with methylation and silencing of these genes. Of note, this model only addresses NuRD function when localized to DNA through the MBD components and does not necessarily apply when transcription factors recruit NuRD through the associated RBBP proteins [38]. In the latter case, the function likely depends on the location of the transcription factor binding site, the stability of the interaction with DNA, and whether the complex can interact with nearby methylated CpGs or other chromatin-associated factors.

Figure 5. A model of nucleosome remodeling by the NuRD complex.

A schematic diagram shows a simple model of NuRD function on CpG rich regions of the genome. Given the propensity for MBD2 to bind stably and with high affinity to methylated CpG rich DNA, MBD2-NuRD should reposition nucleosomes away from unmethylated (gray and open circles) and towards densely methylated (black and filled circles) CpG rich regions of the genome. In contrast, MBD3-NuRD would not show the same tendency.

Future Directions

Our understanding of NuRD structure and function has expanded considerably over the past twenty years. Crystal structures of the MTA/HDAC and RBBP/MTA have given us a firm understanding of the histone deacetylase sub-complex structure. However, the molecular details of how the deacetylase complex associates with chromatin and interacts with histone tails remains mostly unknown. The HDACs bind the MTA protein immediately adjacent to its N-terminal BAH domain. Homologous BAH domains have been shown to bind histone tails, and this arrangement suggests that the BAH domain of MTAs could bind and present a histone tail to HDACs for deacetylation. However, no direct evidence has shown whether the BAH domain directly binds to the histones or contributes to deacetylation. Alternatively, this same domain could be involved in protein-protein interactions with the MBDs. In support of the latter, crosslinking studies have indicated that the MBD and BAH domains are close [20, 23].

Furthermore, even though we have shown that the MBD2-IDR binds HDAC/MTA, we still do not know whether MBDs directly bind the MTAs, HDACs, or both. Therefore, structural analyses of the histone core complex, including one of the MBDs, bound to nucleosome core particles would address many of these open questions and provide mechanistic insight into histone deacetylation by NuRD. With the rapid development of cryo-electron microscopy over the past decade, we anticipate that an atomic resolution structure of this complex should be attainable soon.

We know much less about the interaction between the GATAD2 and CHD proteins. The structure of the C-terminal region of CHD4 remains unknown, in part because that region has been difficult to purify in isolation. Also, the GATA-type Zn-finger in the GATAD2 proteins (CR2) has not been studied in detail. In unpublished NMR studies, we found that this Zn-finger remains poorly structured in isolation. Therefore, we anticipate that detailed structural analysis of this region requires co-expression of the CHD and GATAD2 domains as a complex. Studies by Joel Mackay’s group have shown that the CHDs loosely associate with the rest of NuRD, which indicates that the CHD-GATAD2 interaction may be difficult to purify and study in isolation [59]. Nonetheless, recruitment of CHDs to NuRD is critical for function, and determining the structure of this complex should provide needed insight into the molecular mechanism of chromatin remodeling.

Ultimately, structural analyses of the NuRD complex strive to provide functional insight into its role in gene regulation. To accomplish this goal requires a better understanding of how the full complex remodels nucleosomes and how this remodeling depends on DNA methylation and co-recruitment by transcription factors. Dynamic flexibility of several of the core components likely contributes to this remodeling function but also impedes high-resolution structural analysis of the entire complex. Nonetheless, studies by both the Mackay group and Schwabe group have used a combination of electron microscopy and chemical crosslinking to build models of large portions of the complex [23, 26]. These studies suggest a pathway forward to solving the intact, full NuRD structure using a combination of techniques and careful modeling based on the extensive datasets available in the literature.

However, these types of models do not directly address how the NuRD complex functions to remodel chromatin. As discussed above, we do not understand how methylation impacts remodeling, nor do we understand how the remodeling and deacetylation activities function together to modify chromatin compaction and gene expression. Does methylation of specific CpG sites dictate the positioning of nucleosomes, or does it merely function as a barrier to remodeling leading to more condensed nucleosomes arrays? Also, NuRD can be recruited to chromatin through a variety of co-factors which may or may not impact how the complex repositions nucleosomes. Does the transcription factor binding site dictate where NuRD localizes a nearby nucleosome? Alternatively, does the inherent flexibility of NuRD allow it to remodel over a much larger distance even while tethered to specific transcription factor? Solving these questions will provide needed insight into the function of NuRD and lead to a better understanding of epigenetics and transcription regulation.

Abbreviations:

MBD2(3)

methyl-CpG binding domain protein 2(3)

CpGs

cytosine-guanosine dinucleotides

NuRD

Nucleosome Remodeling and Deacetylase

MeCP1(2)

methyl-CpG binding protein 1(2)

MTA1(2,3)

metastasis tumor-associated protein 1(2,3)

HDAC1(2)

histone deacetylase 1(2)

RBBP4(7)

retinoblastoma binding protein 4(7)

GATAD2A(B)

GATA zinc finger Domain 2A(B)

CHD3(4,5)

Chromodomain Helicase DNA Binding Protein 3(4,5)

CDK2AP1

cyclin-dependent kinase 2 associated protein 1

HDCC

histone deacetylase core sub-complex

PHD

plant homeodomain

HMG

high mobility group

SALL4

spalt-like transcription factor 4

FOG1

friend of GATA protein 1

BCL11a

B-cell lymphoma/leukemia 11a

ZMYND8

zinc finger MYND-type containing 8

Ins(1,4,5,6)P4

inositol tetraphosphate

GR

glycine-arginine