Probing the Dynamic Distribution of Bound States for Methylcytosine-binding Domains on DNA*

Background: Although highly homologous to MBD2, the functional role of MBD3 remains in question. Results: MBD3 preferentially localizes to methylated and, to a lesser degree, unmethylated CpG dinucleotides. Conclusion: Dynamic distribution between methylated and unmethylated sites modifies the genomic localization of MBD3. Significance: Changes in the dynamic distribution on DNA dictate functional differences between MBD proteins. Although highly homologous to other methylcytosine-binding domain (MBD) proteins, MBD3 does not selectively bind methylated DNA, and thus the functional role of MBD3 remains in question. To explore the structural basis of its binding properties and potential function, we characterized the solution structure and binding distribution of the MBD3 MBD on hydroxymethylated, methylated, and unmethylated DNA. The overall fold of this domain is very similar to other MBDs, yet a key loop involved in DNA binding is more disordered than previously observed. Specific recognition of methylated DNA constrains the structure of this loop and results in large chemical shift changes in NMR spectra. Based on these spectral changes, we show that MBD3 preferentially localizes to methylated and, to a lesser degree, unmethylated cytosine-guanosine dinucleotides (CpGs), yet does not distinguish between hydroxymethylated and unmethylated sites. Measuring residual dipolar couplings for the different bound states clearly shows that the MBD3 structure does not change between methylation-specific and nonspecific binding modes. Furthermore, residual dipolar couplings measured for MBD3 bound to methylated DNA can be described by a linear combination of those for the methylation and nonspecific binding modes, confirming the preferential localization to methylated sites. The highly homologous MBD2 protein shows similar but much stronger localization to methylated as well as unmethylated CpGs. Together, these data establish the structural basis for the relative distribution of MBD2 and MBD3 on genomic DNA and their observed occupancy at active and inactive CpG-rich promoters.

Although highly homologous to other methylcytosine-binding domain (MBD) proteins, MBD3 does not selectively bind methylated DNA, and thus the functional role of MBD3 remains in question. To explore the structural basis of its binding properties and potential function, we characterized the solution structure and binding distribution of the MBD3 MBD on hydroxymethylated, methylated, and unmethylated DNA. The overall fold of this domain is very similar to other MBDs, yet a key loop involved in DNA binding is more disordered than previously observed. Specific recognition of methylated DNA constrains the structure of this loop and results in large chemical shift changes in NMR spectra. Based on these spectral changes, we show that MBD3 preferentially localizes to methylated and, to a lesser degree, unmethylated cytosine-guanosine dinucleotides (CpGs), yet does not distinguish between hydroxymethylated and unmethylated sites. Measuring residual dipolar couplings for the different bound states clearly shows that the MBD3 structure does not change between methylation-specific and nonspecific binding modes. Furthermore, residual dipolar couplings measured for MBD3 bound to methylated DNA can be described by a linear combination of those for the methylation and nonspecific binding modes, confirming the preferential localization to methylated sites. The highly homologous MBD2 protein shows similar but much stronger localization to methylated as well as unmethylated CpGs. Together, these data establish the structural basis for the relative distribution of MBD2 and MBD3 on genomic DNA and their observed occupancy at active and inactive CpG-rich promoters.
The mammalian methylcytosine-binding domain proteins (MeCP2 and MBD1-4) selectively bind symmetrically methylated CpGs through a common methylcytosine-binding domain (MBD) 2 (1) and likely arose from a gene duplication event of a single common ancestral protein (MBD2/3) (2). The preference of MBD2 for methylated DNA has been retained in both invertebrates and vertebrates; (3); however, the highly homologous MBD3 shows little to no preference for methylated DNA as a result of key differences in amino acids critical for DNA contact within the MBD (4,5). Both proteins recruit a nucleosome remodeling and deacetylase (NuRD) complex (6) but in a mutually exclusive manner (7). The MBD2-NuRD complex has been specifically shown to promote methylation-dependent gene silencing and represents a potential therapeutic target for gene reactivation (6, 8 -13), whereas the function of the MBD3-NuRD complex has not been clearly delineated.
A recent study showed that MBD3 co-localizes with Tet1 and suggested preferential binding to hydroxymethylated CpGs (hmCpG) (14). Subsequent experiments, however, failed to show a binding affinity preference for hmCpG (15) but instead found that both MBD2 and MBD3 preferentially localize to transcriptional start sites with CGIs (16,17). MBD2 predominates at methylated CGIs, and the associated genes show reduced expression, whereas MBD3 appears to favor transcriptional start sites with unmethylated CGIs and is enriched at active promoters (16).
To help elucidate the structural differences between MBD2 and MBD3 and evaluate the recently proposed hydroxymethylation selectivity (14), we determined the structure of MBD3 bound to DNA containing a single hmCpG dinucleotide. We observed that MBD3 adopts a structure very similar to that of MBD2. A critical loop connecting two anti-parallel ␤ strands is less well defined in MBD3, but otherwise the two structures are nearly identical. Furthermore, we show that MBD3 does not specifically recognize hmCpG, but chemical shift analysis indicates that MBD3 binds differently to mCpG and spends a significant proportion of time on methylated sites. Occupancy depends on the number of unmethylated sites available, and MBD3 demonstrates chemical shift averaging indicative of fast exchange between the methylated and nonspecific binding modes. Residual dipolar coupling (RDC) analysis confirms our findings by showing that MBD3 preferentially localizes to mCpG sites and that MBD3 adopts a very similar structure on mCpG, CpG, and hmCpG DNA. As expected, MBD2 shows a strong preference for mCpG sites, exclusively localizing to the mCpG dinucleotide. We also find that MBD2 localization is influenced by unmethylated CpG density and that MBD2 shows an unanticipated additional weak localization to hmCpG.
Taken together, this information leads to a model in which the methylation specificity and occupancy of an MBD can be titrated by single amino acid substitutions. Importantly, the tendency to localize on a specific site does not necessarily translate into a global binding affinity preference. These data are consistent with recent studies showing that both MBD2 and MBD3 localize to transcription start sites associated with CGIs (16,17). The ability to condense chromatin and silence transcription at or near methylated CGIs reflects MBD2 high affinity and stable occupancy of mCpG sites. Hence, we propose that MBD3 evolved, at least in part, to counterbalance MBD2 on unmethylated CGIs. The presence of MBD3 at unmethylated CGIs could modify the distribution of MBD2 and potentially mitigate strong silencing by the high affinity, more strongly localizing MBD2 protein, thereby preserving bivalency with respect to transcription.

EXPERIMENTAL PROCEDURES
Purification of Proteins and DNA-The methylcytosinebinding domain of MBD3 (amino acid residues 1-70) was cloned, expressed, and purified as described previously for cMBD2 (18). 17-and 27-bp complementary oligonucleotides (Table 1) were purchased (Integrated DNA Technologies), annealed, and purified as described previously (18). The sequences were derived from the p16 INK4a promoter known to be a native target sequence for MBD2 (5).
NMR Spectroscopy-Purified protein was combined with 10% excess purified dsDNA and buffer exchanged into 10 mM NaPO 4 , pH 6.5, 1 mM dithiothreitol, 10% 2 H 2 O, and 0.02% sodium azide and concentrated to 0.2-1 mM. NMR spectra from standard experiments for resonance assignments, distance, and torsional angle restraints were collected on a Bruker Avance III 700-MHz instrument. The data were processed using NMRPipe (19) and analyzed with CcpNmr (20). Residual dipolar couplings were measured for complexes containing 2 H, 13 C, 15 N-labeled protein using standard in-phase/anti-phase experiments and samples aligned by adding ϳ12 mg/ml pf1 bacteriophage (Asla Biotech, Ltd.). For each aligned sample, a one-dimensional 2 H spectrum of 2 H 2 O was collected, and the deuterium quadrupole splitting was measured. When comparing RDC values between samples, the observed RDC values were normalized to an effective deuterium quadrupole splitting of 10 Hz.
Structure Calculations-The structure of the MBD3 MBD was calculated by simulated annealing as implemented in the Xplor-NIH software package (21) and based on NOE-derived distance constraints, torsion angle restraints, and residual dipolar couplings, as well as a torsion angle database potential of mean force (22) and a quartic van der Waals repulsion term for nonbonded contacts (23). Backbone torsional angle restraints were derived from chemical shifts using the TALOSϩ software (24), and hydrogen bond distance and angle restraints were introduced based on backbone torsional angles and characteristic NOE patterns.
Binding Affinity-Binding affinities were determined by surface plasmon resonance analysis on a Biacore T100 system (GE Healthcare) as described previously (18). The binding affinity was determined from steady state analysis of the SPR relative response at varying concentrations of protein. As previously shown, the maximum steady state response (R max ) in SPR depends on stoichiometry (n) of binding (25). Prior to fitting, the steady state response at each protein concentration (R A ) was normalized (R norm ) to the total DNA immobilized (R l ) and molecular weights of the DNA and protein (MW L and MW A , respectively).
Final data analysis, plotting, and curve fitting were performed with pro Fit software (QuantumSoft).

RESULTS
Solution Structure of the MBD3 MBD Is Nearly Identical to That of MBD2-We determined the solution structure of the MBD from MBD3 (amino acids 1-70) bound to a 17-bp dsDNA with a central hydroxymethylated CpG dinucleotide. The structure was calculated based on 528 NOE-derived distance constraints, 120 dihedral angle restraints, and 53 residual dipolar coupling restraints ( Table 2). The overall protein structure is well defined (Fig. 1, A and B) with average pairwise root mean square deviations (RMSD) of 0.7 Ϯ 0.1 Å (backbone) and 1.2 Ϯ 0.1 Å (all heavy atoms) for ordered regions (residues 6 -23 and  27 27 GAGCTAGAGCGCT(mC)GGCGGCGCCAGGC mCpG 10 10 GGAT(mC)GGCTC 34 -71). As expected, the fold is very similar to that of chicken MBD2 (cMBD2; RMSD ϭ 2.0 Ϯ 0.1 Å) for the same ordered regions ( Fig. 1D) (18). As with all MBD proteins studied to date, the topology of the MBD comprises a four-stranded ␤-sheet followed by a single ␣-helix and a C-terminal loop. The central two strands of the ␤-sheet (␤2 and ␤3) form a long finger-like projection that can extend down and across the major groove of DNA to make base specific contacts. The most notable difference between the cMBD2 and MBD3 structures is that the loop connecting the long fingerlike projection is not as well ordered in MBD3 (residues 24 -33; Fig. 1, A and C) with an RMSD of 1.7 Ϯ 0.5 Å (backbone) as compared with cMBD2 with an RMSD of 0.9 Ϯ 0.3 Å. Residues at the base of this loop form critical DNA specific contacts and a hydrogen bond network that stabilizes the interaction with the methylated CpG dinucleotide (Fig. 1E).
Predicting backbone order parameters (S 2 ) based on chemical shifts using the random coil index method (24,26) confirms that this loop is less well ordered in MBD3 (Fig. 2). The difference in predicted S 2 between complexes shows that residues 24 -33 become progressively more ordered between the MBD3-hmCpG, MBD3-mCpG, and cMBD2-mCpG complexes, respectively. Thus in the absence of a methylation-specific binding mode, the loop connecting the central two ␤-strands is more flexible.
Filtered intermolecular NOE spectra did not contain NOE cross-peaks, which is consistent with the observed line broadening for residues at the DNA interface and the overall lower affinity of MBD3 for DNA and suggests nonspecific protein-DNA interaction with dynamic exchange between binding sites. Based on subsequent analyses that indicate MBD3 preferentially localizes to methylated sites, we collected filtered NOE spectra for MBD3 bound to methylated DNA (mCpG). Like-FIGURE 1. Solution structure of MBD3 methyl-binding domain bound to hydroxymethylated DNA. A, stereo ribbon diagram (blue) of the MBD3 solution structure is shown for the ensemble of 20 calculated structures (Protein Data Bank code 2mb7). The loop connecting ␤2 and ␤3 (residues 24 -33) is highlighted in light blue. B, ribbon diagram of the lowest energy solution structure is shown with key contact and chemical shift reporter residues depicted as sticks. C, per residue RMSD for backbone atoms is plotted for the solution structure ensemble of MBD3 (blue) and for the solution structure ensemble of cMBD2 (red) previously reported (Protein Data Bank code 2ky8) (18). D, the best fit protein alignment of the solution structures of cMBD2 (green) and MBD3 (blue) MBD is shown bound to the methylated DNA from the cMBD2-dsDNA solution structure (Protein Data Bank code 2ky8). E, diagram highlighting the cMBD2 hydrogenbonding network while bound to methylated DNA and with key residues depicted as sticks. Structure diagrams were generated using the PyMOL program (Delano Scientific LLC). wise, we did not detect any intermolecular NOE cross-peaks with methylated DNA, indicating that MBD3 exchanges among different binding modes whether on methylated or hydroxymethylated DNA. Given the absence of informative intermolecular NOEs, we did not determine a solution structure of the protein-DNA complex. Importantly, we noted that chemical shifts of MBD3 bound to hmCpG (discussed in detail below) are more similar to MBD3 on unmethylated DNA than methylated DNA, which led us to compare spectra of MBD3 on different DNA molecules to probe both methylation-specific and nonspecific DNA association. MBD3 Spends a Significant Portion of Time on Methylated Sites-As first described for MeCP2 (27) and later described for cMBD2 (18), a pair of highly conserved arginine residues in MBD proteins form bidentate hydrogen bonds with the symmetrically opposed guanine bases of an mCpG dinucleotide (Fig. 1E). The aliphatic portion of each arginine packs against the methyl group of the neighboring methylcytosine. Given that an unmethylated CpG contains symmetrically opposed guanines, we hypothesized that MBDs should still recognize the CpG dinucleotide, but with lower affinity. Even though MBD3 binds DNA with lower overall affinity and shows less selectivity for methylated DNA, the critical arginine residues are conserved and could provide sequence specific recognition of CpG dinucleotides.
In 15 N HSQC spectra, we noted that several cross-peaks that showed unusual chemical shifts for cMBD2-mCpG did not show the same chemical shifts for MBD3-hmCpG (Fig. 3B). In particular, the 1 H⑀ of Arg 24 , which forms a side chain hydrogen bond with Asp 32 and is shifted far downfield (9.5 ppm) in cMBD2-mCpG, is only shifted to 7.5 ppm in MBD3. Likewise, Gly 27 is shifted upfield in 15 N to 102 ppm in cMBD2-mCpG, but not to the same degree in MBD3-hmCpG (105 ppm), and finally Ala 30 is shifted upfield in 1 H to 6.8 ppm in cMBD2-mCpG but only 7.6 ppm in MBD3-hmCpG.
We hypothesized that these large chemical shift changes reflect a difference between methylation-specific and nonspecific binding modes. Arg 24 and Asp 32 are positioned at the N and C termini of the poorly structured loop in MBD3, whereas Ala 30 and Gly 27 also fall within this same loop. Methylationspecific binding stabilizes the Arg 24 -Asp 32 H-bond and the loop containing Gly 27 and Ala 30 . To test this hypothesis, we compared HSQCs for MBD3 bound to a DNA sequence with three CpG dinucleotides in which the central CpG is methylated (mCpG), hydroxymethylated (hmCpG), or unmethylated (CpG(ϫ3)), as well as similar DNA sequences with only one (CpG(ϫ1)) or no CpGs (CpG(ϫ0)) ( Table 1). We found that the chemical shifts of these reporter resonances fall on a line between extrema represented by cMBD2-mCpG and MBD3-CpG(ϫ0). Importantly, the peak position for each of the reporter resonances falls at the same fractional distance between these extrema (Fig. 3B), shifting toward the position in the cMBD2-mCpG complex as the number of unmethylated CpG sites increases and with the addition of a methylated CpG. These observations strongly indicate chemical shift averaging between two binding modes (28) reflective of fast exchange between methylation-specific and nonsequence specific interaction with DNA. Furthermore, the chemical shift changes are consistent with preferential localization at the CpG and mCpG sites.
Using cMBD2 as representative of the methylation-specific binding mode could introduce structural and primary sequence differences that affect observed chemical shifts. Therefore we sought to generate an MBD3 MBD that binds with high selectivity for mCpG, which would allow us to evaluate chemical shift changes for more backbone resonances with fewer confounding sequence variations. Previous studies have established that the lack of mCpG specificity for MBD3 reflects two amino acids (His 30 and Phe 34 ) that differ from other MBD proteins (Lys 32 and Tyr 36 in cMBD2) (4,5). We introduced the H30K,F34Y double mutation into MBD3 (MBD3 KY ), and as expected, this mutant bound with higher affinity and selectivity for mCpG comparable to cMBD2 (Fig. 7 and Table 3). Each of the reporter resonances now show chemical shifts that are very similar to those of the cMBD2-mCpG complex, confirming that the unique chemical shifts do reflect a methylation-specific binding mode.
To further explore the chemical shift changes associated with the different binding modes, we assigned the backbone resonances ( 15 N, 1 H) for MBD3 bound to mCpG or CpG(ϫ3) and MBD3 KY bound to mCpG. In Fig. 3A, the chemical shift distances are plotted for backbone resonances between the different complexes. The largest chemical shift changes are seen for the poorly structured loop (residues 24 -33) when comparing the MBD3 KY -mCpG and MBD3-mCpG complexes with the MBD3-hmCpG complex. In contrast, the 15 N HSQC spectra shading from blue to red reflects low to high). B, the predicted order parameters (S 2 ) are plotted for the MBD3-hmCpG (black), MBD3-mCpG (blue dotted), and cMBD2-mCpG (red) complexes. C, bar plots are shown for the difference in order parameters (⌬S 2 ) between the cMBD2-mCpG complex and MBD3-mCpG (black) and MBD3-hmCpG (gray) complexes. The loop connecting ␤2 and ␤3 (residues 24 -33) is highlighted in light yellow in B and C.
These observations strongly support a binding model in which MBD3 exhibits fast exchange between methylation-specific and nonspecific binding. Thus the observed chemical shift ( obs ) reflects a weighted average of the methylation-specific ( MBD3(KY)-mCpG ) and nonspecific ( MBD3-CpG(ϫ3) ) binding modes, as given by Equation 2, where m is the fraction in the methylation-specific binding mode. Hence, these chemical shifts are direct measures of the average time spent on the mCpG site. Using 11 backbone 1 H, 15 N resonances that show a clear linear relationship between chemical shift and DNA bound, we find that MBD3 spends ϳ43% (Ϯ5%) of the time on the mCpG site (Fig. 3B).
These data further indicate that, despite a lack of a strong global binding affinity preference for mCpG DNA, MBD3 still spends a significant proportion of time on methylated sites. To confirm that this finding was not the result of very weak binding with chemical shift averaging between DNA bound and free MBD3, we compared HSQCs for 600 and 300 M samples of protein on DNA. The peaks for each reporter residue show nearly identical chemical shifts at both concentrations (Fig. 5B), indicating that the observed chemical shift changes of the reporter residues were not the result of exchange between bound and free states but instead represent averaging between different bound states. The observed differences in chemical shift represent changes in the binding distribution on DNA, not changes in the distribution between bound and free states.
We developed a statistical-mechanical model to describe the distribution of MBD3 on methylated DNA (Fig. 3C) in which the partition function comprises a sum of Boltzmann factors (e ϪE(i)/kT ) for methylation-specific and nonspecific binding modes, ⌬E is the difference in energy between binding modes, and N is the number of nonspecific sites. The additional methylation-specific interactions formed by MBD2 lead to a larger  ⌬E, thereby increasing m and the relative selectivity for mCpG.
Without these interactions, MBD3 shows a smaller ⌬E, m is reduced but still not 0, and MBD3 distributes unevenly between mCpGs and unmethylated sites. Fig. 3C depicts a simplified structural model of this distribution in which the MBD3 structure was docked onto the 17-bp DNA fragment at the centrally methylated mCpG (red) as well as four "nonspecific" binding sites (blue). The statistical model further indicates that localizing to an mCpG depends on the number of nonspecific sites available (N) and therefore the length of DNA. We tested this latter prediction by comparing chemical shifts for the reporter residues of MBD3 on 10-, 17-, and 27-bp DNA with a single mCpG (Fig. 3D). Indeed, we found that increasing DNA length results in chemical shift changes for Gly 27 and Ala 30 toward the unmethylated binding mode.

Residual Dipolar Couplings Confirm MBD3 Localizes to Methylated Sites without Significant Conformational Change-
As an alternative method to assess the ensemble of binding modes, we measured residual dipolar couplings ( 1 D NH ) for MBD3 bound to methylated and unmethylated DNA as well as for MBD3 KY bound to methylated DNA. As can be seen in Fig.  5, the observed 1 D NH are similar but not identical between the different complexes. When plotting 1 D NH from the unmethylated complex against those from wild type MBD3 or MBD3 KY methylated complexes, the data fall off of the line of identity (y ϭ x), as highlighted by the red dotted ovals in Fig. 5 (A and B). Because the observed residual dipolar couplings reflect a weighted average of the different binding modes (29), the 1 D NH for each residue (n) of MBD3 bound to mCpG is a linear combination of 1 D NH for MBD3 bound to CpG(ϫ3) (nonspecific binding mode) and 1 D NH for MBD3 KY bound to mCpG (methylation-specific binding mode). In Equation 3, m is the fraction in the mCpG specific binding mode and 1 D NH (n) are the residual dipolar couplings for each residue in the MBD3-mCpG, MBD3-CpG, and MBD3 KY -mCpG complexes. Fitting the observed 1 D NH to Equation 3 as a function of m provides the fraction of MBD3 in the methylation-specific binding mode. As shown in Fig. 5C, the sum of  squared residuals between predicted and observed 1 D NH is minimized when m is 0.37. Plotting observed 1 D NH,MBD3-mCpG versus predicted at m ϭ 0.37 shows good agreement with tight clustering around y ϭ x (Fig. 5D). Therefore, using RDCs as an independent measure of methylation selectivity, we find that MBD3 spends ϳ37% of the time on the methylated binding site, which agrees within experimental error with the results from chemical shift analysis.
The residual dipolar couplings for each complex were fit to the solution structure of MBD3 bound to hmCpG using singular value decomposition as implemented by PALES software (30). Despite the differences in RDCs between complexes, each data set fit quite well to the MBD3 structure with Q factors of 19.9, 23.1, and 25.9% for the MBD3-mCpG, MBD3-CpG(ϫ3), and MBD3 KY -mCpG complexes, respectively (Fig. 5E). Therefore, the change between methylated and nonspecific binding modes does not involve backbone structural changes. The observed chemical shift changes likely reflect a stabilization of the dynamic loop between ␤2 and ␤3 but without significant structural rearrangements.
MBD2 and MBD3 Distribution Is Influenced by DNA Methylation Status and CpG Density-In contrast to MBD3, MBD2 appears to spend most of its time on the methylated site (Fig.  6A). The chemical shifts for reporter residues in MBD2 represent the extrema for the complexes studied. As shown in Fig.  6B, these peaks do not change with increasing DNA length, indicating that a large ⌬E dominates the fraction bound to the methylated site and m Ϸ 1 (Fig. 3C). The higher affinity and methylation selectivity of MBD2 results in exclusive occupancy of mCpG.
Comparing spectra of MBD2 and MBD3 when bound to the unmethylated oligonucleotides CpG(ϫ0), CpG(ϫ1), and CpG(ϫ3) reveals that the reporter resonances shift toward the methylation-specific state with increasing numbers of CpGs ( Fig. 3B; see also Fig. 7A). This finding indicates that both MBD2 and MBD3 localize to sites of increased CpG density. MBD3 shows relatively small chemical shift changes when bound to CpG(ϫ3) as compared with CpG(ϫ0), consistent with localizing to a CpG dinucleotide ϳ9% (Ϯ3%) of the time. MBD2 shows a pronounced difference between CpG(ϫ0) and CpG(ϫ3) (Fig. 6A), showing that MBD2 preferentially localizes to CpG dinucleotides ϳ33% (Ϯ11%) of the time. The reporter resonances also indicate that MBD2, but not MBD3, tends to localize to hmCpG and exhibits higher affinity for these sites ( Fig. 6A and Table 3). To test whether MBD3 could influence how MBD2 distributes on unmethylated CpGs, we added equimolar MBD3 to 15 N-MBD2 bound to CpG(ϫ3) and found that the reporter peaks shifted toward those of the nonspecific binding mode (Fig. 6A). This finding shows that despite relatively weak overall binding affinity, MBD3 can modulate the distribution of MBD2 on CpG sites, shifting MBD2 toward the nonspecific binding mode.
Global Binding Affinity Does Not Reflect Localization Preferences on DNA-Binding affinities for each of the methylated and unmethylated sequences investigated were determined by steady state analysis of surface plasmon resonance data, as described previously (18). To allow for direct comparison of binding stoichiometry, we normalized the relative steady state response to the amount of DNA coupled to the sensor chip such that the maximum steady state response reflects the number of binding sites on the DNA (25). As can be seen in Fig. 7 and Table  3, both cMBD2 and MBD3 KY bind mCpG DNA with high affinity and stoichiometry of approximately one (K D ϭ 105 Ϯ 7 nM and 113 Ϯ 11 nM and R max ϭ 0.83 Ϯ 0.01 and 0.98 Ϯ 0.02, respectively). Both bind unmethylated DNA with much lower affinity and high stoichiometry (K D ϭ 74 Ϯ 4 M and 17 Ϯ 2 M and R max ϭ 4.8 Ϯ 0.1 and 2.1 Ϯ 0.1, respectively) indicative of a high degree of methylation selectivity. Although NMR analyses indicate MBD3 binds DNA with sufficient affinity to be fully bound at 300 M concentration (Fig. 4B), solubility limits of the isolated protein preclude accurate determination of binding constants by surface plasmon resonance. The qualitative results of these studies, however, are very similar to those reported previously by Hashimoto et al. (15), who determined DNA binding affinity for full-length MBD3 by fluorescence anisotropy. In those previous studies, MBD3 bound with similar low affinity to unmethylated and hydroxymethylated DNA and with a small but weak preference (ϳ5-fold) for methylated DNA. The current SPR data indicate a small preference for mCpG as well (Fig. 7).
Likewise, cMBD2 shows similar low affinity and high stoichiometry when binding to hydroxymethylated and unmethylated DNA with 0 -3 CpG dinucleotides ( Fig. 7 and Table 3). Therefore, global binding analysis reveals a marked preference for mCpG by cMBD2 and MBD3 KY but only weak and nonspecific binding for all other complexes. These findings agree with several previous DNA binding analyses of MBD proteins (4,5,15,18).

DISCUSSION
Although it has been established that MBD3 binds DNA with lower affinity and much less specificity for mCpG dinucleotides than MBD2, the functional role of MBD3 has not been well defined. Based on NMR structural, chemical shift, and residual dipolar coupling analyses, we have demonstrated that MBD3 binding to methylated DNA can be described by an ensemble of FIGURE 6. MBD2 distribution is influenced by DNA methylation status and CpG density. A, an overlay of 15 N HSQC spectra for key reporter residues of MBD2 bound to CpG(ϫ0), CpG(ϫ1), CpG(ϫ3) Ϯ MBD3, mCpG, and hmCpG shows that cMBD2 preferentially localizes to DNA with mCpG, hmCpG, and multiple CpG sites and that localization is modified by the presence of equimolar MBD3. B, an overlay of 15 N HSQC spectra cMBD2 bound to mCpG DNA of varying lengths confirms that cMBD2 strongly prefers mCpG sites.
methylation-specific and nonspecific binding modes and that MBD3 preferentially localizes, albeit weakly, to methylated sites. In contrast, MBD3 does not distinguish between hydroxymethylated and unmethylated DNA but does show a slight preference for DNA with multiple CpG dinucleotides. A similar chemical shift analysis of cMBD2 showed that cMBD2 localizes almost exclusively to the methylated sites when present. Even on unmethylated DNA, cMBD2 preferentially localizes to CpG dinucleotides, especially as the number of available CpGs increases. Surprisingly, cMBD2 more avidly localized to a hydroxymethylated than an unmethylated site. Taken together, these findings lead to a model in which both MBD2 and MBD3 preferentially localize to DNA with multiple CpGs, whereas MBD2 more exclusively localizes to mCpGs.
Our studies lead to several novel observations and hypotheses. First, the differences in DNA binding between MBD3 and MBD2 reflect a change in the degree of selectivity for mCpG, not absolute differences in binding specificity. As we hypothesized, both MBD2 and MBD3 show a weak preference for CpG dinucleotides even the absence of methylation; however, MBD2 shows a much greater selectivity for mCpG than MBD3. These preferences are not necessarily apparent by global binding analyses but instead are reflected in the partitioning between methylation-specific and nonspecific binding modes on DNA. The observed changes in the NMR spectra arise from changes in the distribution of bound states not from changes in the distribution between bound and free states. In addition, the preference, although weak, for multiple CpG sites within the relatively small 17-base pair oligonucleotide used in these studies correlates with the whole genome analyses that show MBD2 and MBD3 localize to CGIs. Given that a CGI contains 100 -1000 CpG dinucleotides, even a relatively small preference would lead to fairly strong localization at such sites. Furthermore, these findings correlate with the observations that both MBD2 and MBD3 are found at unmethylated CGIs, whereas MBD2 binds with much greater affinity and likely excludes MBD3 from methylated CGIs. Finally, these studies clearly indicate that MBD3 does not exhibit a binding preference for or a structural recognition of hmCpG DNA. From the standpoint of MBD3, hydroxymethylation is functionally equivalent to demethylation.
Genomes containing both MBD2 and MBD3 proteins emerge at the same time as the vertebrate methylation pattern, which includes largely unmethylated CGIs. This concurrence along with the preceding characterization of DNA binding by MBD3 leads us to speculate that MBD3 plays an important role in regulating genes with unmethylated CGIs. One possibility these studies raise is that MBD3 helps counterbalance the tendency of MBD2 to preferentially localize to CpG dinucleotides by competing with MBD2 at unmethylated CGIs. We found that MBD3 does modify the distribution of MBD2 on DNA such that MBD2 spends less time on CpG sites. Thus MBD3 could help prevent gene silencing by MBD2 at unmethylated gene promoters and enhancers depending on the relative concentration of the two proteins and cellular context. Indeed, recent studies have shown that knockdown of MBD3 can lead to decreased gene expression (16,17).
Mammalian cells express multiple MBD proteins as well as different isoforms of individual MBDs. Different splice variants as well as distinct genes encode for MBD2 and MBD3 proteins, some of which lack the DNA binding domain itself. Here we have studied the solution structure and DNA binding of MBD3 by NMR. Chemical shift analyses indicate that MBD3 recognizes and preferentially localizes to both mCpG and CpG sites but not to the same extent as MBD2. Single amino acid differences dictate the degree to which these proteins localize on mCpG sites. Importantly these binding characteristics do not necessarily lead to changes in global binding affinities but rather correlate with localization of MBD2 and MBD3 to CGIs in whole cells. Hence a number of different NuRD complexes can be formed that show varying degrees of DNA methylation selectivity and provide distinct functional roles. For MBD3, these functional differences appear to reflect subtle distinctions in the behavior of the MBD when bound to methylated and unmethylated DNA. Therefore, these data establish a structural basis for the relative distribution of MBD2 and MBD3 on genomic DNA and help explain their observed occupancy at CpG-rich promoters.