Structural Basis for Ligand Recognition and Activation of RAGE

VC1 Forms an Integrated Structural Unit

The tandem RAGE VC1 domains were crystallized and the X-ray structure was refined at 1.85Å resolution to a R_cryst of 20.9% and R_free of 24.0% (Table 1). VC1 is found to form a bent elongated structure with an angle of 145° between the two Ig domains (Figure 1A). A specific interface between the domains is observed, which includes several interdomain hydrogen bonds and hydrophobic interactions (Figure 1A). Among these H-bonds, Gln119 in V engages in a pseudo β sheet hydrogen-bonding pattern with Tyr150 in C1, and the carboxylate oxygen atom of Glu94 in V forms an H-bond with the hydroxyl hydrogen of Tyr150 in C1. In addition, a water molecule bridges the amide nitrogen of Ile120 in C1 domain with the side chains of Gln119 and Arg29 in V. Hydrophobic contacts at the interface are mediated between side chain atoms of Pro215 from the C1 F-G loop and the side chain of Tyr118, as well as between the side chain of Tyr150 from the C1 B-C loop and Ile91. Notably, all residues at the VC1 interface are strictly conserved (see Figure S1 available online). These characteristics confirm that the two domains form an integrated structural unit (Dattilo et al., 2007).

On the basis of sequence alignments, the N-terminal domain of RAGE (V) had been assigned to the V-set type of Ig molecules. However, once the structure was determined, it revealed features typical of I-set topology (Harpaz and Chothia, 1994), which is characteristic for cell adhesion molecules (Casasnovas et al., 1997; Feinberg et al., 2001; Freigang et al., 2000; Kasper et al., 2000; Su et al., 1998) or muscle proteins (Holden et al., 1992; Zou et al., 2006). Strands A, C, C’, F, and G form one sheet, whereas strands B, D, and E form the second sheet (Figure 1). The region in the V domain corresponding to C” strand in V-set Ig domains instead forms an extended loop encompassing three glycine residues (Gly68, Gly69, and Gly70) and two proline residues (Pro66 and Pro71). The frequency of glycine and proline residues is inconsistent with formation of a β strand. High B-factors in this region suggest that the loop might be adopting variable conformations, which is not unusual for a short Gly and Pro-rich sequence. Although more correctly classified as an I domain, the V nomenclature will be retained here in order to maintain consistency with previous literature.

The C1-domain fits to the C1-set of Ig molecules, with strands A, B, D, and E forming the back β sheet and strands C, C’, F, and G forming the front sheet of the β sandwich. However, the RAGE C1-domain has a unique topology distinct from other Ig fold sets: it contains two additional β strands, A’ and G’, which belong neither to the front nor to the back sheet. Both strands form an additional parallel β sheet stabilizing the C terminus of the domain (Figures 1A and 1B). Further variation is observed for strands D and E. In comparison to most C1-type Ig domains, strand D is elongated by two residues encompassing Arg178 and Arg179, as well as the D-E loop, which is slightly longer through the inclusion of Pro181-Gly184. The E-F loop, which includes Thr195-Thr205, is also elongated by residues Arg198-Asp201. This loop has a unique interaction that stabilizes the C terminus of C1 involving a hydrogen bond between the Arg198 side chain and the Pro232 main chain carbonyl oxygen.

A well-ordered 10-residue C-terminal extension from C1 (Glu231-Leu240) is observed in the high-resolution structure (Figures 1A and 1B). This extension constitutes at least in part the linker to the C2 domains, which NMR studies have shown is flexible (Dattilo et al., 2007). A solution NMR structure of the isolated C2 domain has been determined (PDB 2ENS), but the N terminus is not well defined. In our structure, the extension is stabilized by a contact between Leu240 and the C1-domain of a neighboring molecule in the crystal. Hence, some uncertainty remains over the exact length of the linker between the C1 and C2 domains.

VC1 Contains a Highly Basic Surface

One of the most unusual properties of RAGE is its ability to engage very diverse ligands of different size, fold, or symmetry. The structure of the VC1 ligand-binding domain provides a means to assess what specific surface properties enable this unique characteristic. Our analysis revealed large positively charged patches on the V domain (Figure 2), which are formed by highly conserved Arg and Lys residues. A patch of positively charged electrostatic surface was also identified on the C1 domain. Thus, the entire V domain and a major part of C1 form a large area of electropositive surface (Figure 2) that fits well to the acidic (negative) character of the diverse RAGE ligands, including AGE-modified proteins, amyloid-β, and S100 proteins. Comparison of molecular surfaces displaying the isoelectrostatic field potential of structural homologs neural adhesion molecules TAG-1/axonin-1 (Freigang et al., 2000) and N-CAM (Kasper et al., 2000), as well as a structural model of the closest homolog ALCAM, reveals that the electrostatic potential of VC1 is significantly stronger (Figure 2), which suggests RAGE contains an electrostatic trap for negatively charged ligands.

Self-Association of RAGE

Recently, it was shown that RAGE forms oligomers on the plasma membrane (Xie et al., 2008), and it was suggested that oligomerization is mediated by the C1 domain (Xie et al., 2007). Indeed, in the crystal structure, there are side-by-side contacts between VC1 molecules mainly involving residues in the C1 domain. The solvent accessible surface area buried by this interface is 980 Ų (Figure 3A). The contact surface contains several charged residues and is connected by a tight network of hydrogen bonds and polar interactions (Figures 3B and 3C). The guanidine of Arg178 side chain forms a salt bridge with the carboxylate group of Glu168′ and a hydrogen bond to the carbonyl oxygen of Pro163′ of the neighboring molecule (Figure 3C). The Gln176 side chain oxygen forms two hydrogen bonds to the Nε of Arg178, which positions the guanidine group for optimal interaction with the neighboring molecule (Figure 3C; Figure S2). These interactions are reminiscent of the critical salt bridge formed between Arg and Glu/Asp residues at the interface between Ig domains in the active dimer of the receptor tyrosine kinase KIT (Yuzawa et al., 2007) or between the membrane-proximal Ig-domains of the VEGF receptor (Yang et al., 2010). The extensive set of stabilizing interactions between C1 domains in adjacent molecules in the crystal structure is consistent with the proposal of RAGE aggregation in the plasma membrane. In addition to H-bonds and salt bridges, a Zn²⁺ ion is found at the intermolecular interface. The Zn²⁺ is coordinated by residues His180, Glu182, and His158′ from the adjacent VC1 molecule (Figure 3B). These residues, like most involved in at the C1-C1 interface, are conserved across species (Figure S1), which implies a role for Zn²⁺ in receptor multimerization. We note that Zn²⁺ can induce dimerization of the class II major histocompatibility complex, leading to cooperative binding of a superantigen ligand (Li et al., 2007). Moreover, homophilic interactions of receptor Ig-type ectodomains play a central role in ligand-binding and cellular activation (Stuttfeld and Ballmer-Hofer, 2009; Xu and Jin, 2010; Yang et al., 2010).

In order to further investigate self-association of VC1 and a potential role for Zn²⁺, we performed dynamic light scattering experiments. The data show that VC1 is monomeric in the absence of Zn²⁺ and at low pH. The tendency for VC1 to self-associate is heavily influenced by Zn²⁺ and increased pH into the physiological range (Figures 4A–4C). The degree of self-association of VC1 is concentration dependent (Figure 4D), which is consistent with self-association being a specific binding phenomenon. The concentration dependence of self-association is promoted by the presence of Zn²⁺ and shows a saturation behavior that could be fit by a hyperbolic function yielding a K_{d app} of 3.8 ± 1.3 μM.

An NMR-Based Model for the Complex of VC1 with the RAGE Ligand S100B

Structural details about ligand binding to RAGE have been limited to a single study that highlighted severalr esidues in the V domain that are important for binding of AGEs (Matsumoto et al., 2008). We have investigated the structural basis for binding of the ligand S100B, combining two-dimensional heteronuclear NMR experiments to identify the binding interface on both VC1 and S100B and HADDOCK-based docking calculations to generate a model of the complex.

To map the S100B-binding site on RAGE, a series of 2D ¹⁵N-¹H HSQC NMR experiments were acquired as Ca²⁺-S100B was titrated into a solution of ¹⁵N-enriched V domain (Figure 5B). The isolated V domain was used in NMR experiments rather than sRAGE or VC1 because, first, the majority of the S100B binding affinity has been shown to be localized to this domain (Dattilo et al., 2007), second, S100B is a homodimer and working with the smallest complex was important to facilitate the NMR analysis, and third, this enables direct comparison to a published study with AGE ligands (Matsumoto et al., 2008). Broadening of all NMR signals over the course of the titration was observed as expected for formation of the V-S100B complex. However, careful analysis of the data revealed differential line broadening effects. As demonstrated elsewhere (Takeuchi and Wagner, 2006), resonances exhibiting enhanced line broadening correspond to residues experiencing the largest perturbation upon binding of the ligand, and these can be correlated to the ligand-binding sites. The binding site was mapped by tracking 52 resonances distributed throughout the V domain that could be directly assigned from the published chemical shift assignments (Matsumoto et al., 2008). Quantification of normalized intensities for these resonances identified a subset of residues that are selectively broadened upon addition of Ca²⁺-S100B (Figure 5B).

When these changes are mapped onto the VC1 structure, they reveal that binding of Ca²⁺-S100B occurs to one face of the molecule (Figure 5C). Interestingly, these residues map to the highly basic surface of VC1 (Figure 5A), which is complementary to the acidic nature of S100B (Figure 5E). Notably, electrostatic forces have been shown to modulate ligand interactions with other Ig-like domains, such as in the CD2-CD58 protein complex (Wang et al., 1999). Moreover, the S100B-binding site appears to correspond to the site used by RAGE for binding AGE-BSA (Matsumoto et al., 2008). Key residues in AGE binding include Arg48, Arg98, and a patch around Arg104, all of which are perturbed upon binding of Ca²⁺-S100B. Additional residues include Val35, Lys39, Val89, Glu94, Phe97, Ala101, Asn105, Glu108, and Thr109 (Figure 5C). Mutagenesis experiments on V residues serve to validate this as the ligand-binding surface (Matsumoto et al., 2008). The similarity of binding sites for different RAGE ligands has been inferred previously (Matsumoto et al., 2008), but this work combined with the previous NMR studies on binding of AGE ligands, provides the first direct evidence in support of this hypothesis.

In order to map the VC1 binding surface on S100B, a series of 2D ¹⁵N-¹H HSQC NMR experiments were acquired over the course of a reciprocal titration of RAGE VC1 into a solution of ¹⁵N-enriched Ca²⁺-S100B. The effects observed in these spectra were similar to those seen for the S100B titration into the V domain (i.e., differential line broadening for a specific set of S100B resonances). The residues affected include His42, Val52, Asn62, Asp69, Phe70, and Ala78, which cluster to a specific surface on S100B. We note that this surface, and in particular His42, Val52, and Ala78, were shown previously to be important for interactions with other S100B-specific ligands such as NDR kinase (Bhattacharya et al., 2003).

Because residues that are selectively perturbed upon formation of the complex could be identified for both RAGE-V and S100B, we proceeded to generate a model of the complex. The program HADDOCK (de Vries et al., 2007; Dominguez et al., 2003) was used for these calculations. A representative structure was selected on the basis of the lowest HADDOCK score and is shown in Figures 5D–5F. The models show S100B uses a concave, negatively charged surface to clamp around a complementary positively charged face of the RAGE V domain (Figure 5E). A surface area greater than 1400 Ų is buried by formation of the complex. Analysis of the model revealed a number of possible polar hydrogen bonds and salt bridges in the interface. Thus, the model suggests that clustering of basic residues on the binding surface of VC1 and highly acidic surface of S100B are the major driving forces in formation of the complex. In order to validate the result obtained by HADDOCK, we performed a second set of calculations without NMR restraints using the computational docking program HEX (Ritchie and Kemp, 2000; Ritchie et al., 2008). The models obtained in this manner were remarkably similar to the models obtained using the NMR restraints in HADDOCK (Figure S3). Analysis of these models of the VC1-S100B complex reveals a strong electrostatic component to binding, consistent with the proposed electrostatic trap contributing significantly to RAGE ligand binding.

Crystallization, Data Collection, and Structure Determination

Crystals were grown by hanging drop vapor diffusion at 298 K by mixing 2 μl of protein with 2 μl of 0.1 M Na cacodylate, 0.2 M Zn acetate (pH 6.5), and 11% PEG 8000. Crystals of a size of 200 μm 3 100 μm 200 μm appeared after 10 to 20 days. The crystals were soaked for 1 min in three consecutive steps in mother liquor containing 4%, 8%, and 12% glycerol and were flash-frozen in liquid nitrogen. Crystals diffracted to a resolution of 1.85 Å. Because Zn²⁺ turned out to be essential for crystallization, we assumed that Zn²⁺ specifically interacts with RAGE VC1 and could, therefore, be used for experimental phasing. Multiple anomalous dispersion data were collected at a wavelength of 1.28 Å at the Swiss Light Source (Villigen, Switzerland) beamline X06SA at 100 K. An additional data set collected at a wavelength of 1.0 Å was used for refinement. The data sets were reduced and scaled with the XDS package (Kabsch, 1993). Crystals belonged to the space group P2₁2₁2 with unit cell parameters of 74.85, 119.96, and 28.89, as determined by XPREP (2005 version, Bruker-AXS). The zinc substructure was determined using SHELXD (Schneider and Sheldrick, 2002). Refinement of the zinc sites and phase calculation was done with SHARP (Fortelle and Bricogne, 1997) followed by solvent flattening and initial fragment building using RESOLVE (Terwilliger, 2000). The final model was built in COOT (Emsley and Cowtan, 2004) and was refined using REFMAC5 (Murshudov et al., 1997; Winn et al., 2001) with final R_work and R_free of 20.9% and 24.0%, respectively (Table 1). Ninety-nine percent of the residues were located in the allowed regions, and 1% were located in the generously allowed regions of the Ramachandran plot (Table 1).

Solution NMR Mapping Studies with Ca²⁺-Loaded S100B

Substoichiometric additions of Ca²⁺-loaded S100B were made into uniformly labeled ¹⁵N-V in 10 mM Imidazole-d₄ (pH 6.5), 100 mM NaCl, and 5 mM CaCl₂. Similar additions of VC1 were made into a uniform solution of ¹⁵N-labeled S100B in 20 mM KCl (pH 7.1), 2.5 mM Ca²⁺, and 5% D₂O. Perturbations in the NMR signals were monitored by 2D ¹⁵N-¹H HSQC spectra using a Bruker DRX500 spectrometer. Data were processed using Topspin 2.0b (Bruker) and analyzed using NMRViewJ. The intensities (I) were extrapolated from peak heights and plotted as a ratio (I/I_o) versus the initial intensity (I_o).

Protein Oligomerization Assay

Protein oligomerization assays of VC1 were performed at 298 K. A stock solution of VC1 with 10 mg ml⁻¹ in 10 mM sodium acetate buffer (pH 5.0) was diluted into 20 mM Tris and 150 mM NaCl (pH 7.6). Large oligomers formed at increasing VC1 concentrations and in the presence of Zn²⁺. The oligomers were pelleted by centrifugation (10 min at 12,000 g). The amount of oligomer formed was calculated by subtracting the amount of protein remaining in the supernatant from total protein. VC1 concentration in the supernatant was determined by the UV absorbance using an extinction coefficient of ε_278nm = 32500 M⁻¹ cm⁻¹. In experiments monitoring the effect of Zn²⁺ ions, 200 μM ZnCl₂ was present in the buffer.

Docking Calculations

Residues exhibiting significant NMR chemical shift perturbations were used as input for HADDOCK calculations (Dominguez et al., 2003) to generate an experimentally derived model of the S100B-VC1 complex. Specifically, those residues in S100B and in the V-domain of RAGE whose intensities were reduced by more than one standard deviation from the mean reduction in intensity were selected. The active and passive categories of restraint were assigned based on criteria outlined in HADDOCK2 (de Vries et al., 2007).

The calculation proceeded in a similar manner to that reported by Nordquist et al. (2010). To improve the conformation sampling, twenty structures each were generated from the X-ray crystal structures of S100B (PDB accession number 2H61) and VC1 using room temperature MD simulations in AMBER 9 (Pearlman and Connelly, 1995). Using these starting structures, 1000 complexes were generated from rigid-body docking, of which 200 were selected for further refinement in explicit water. One hundred and four of the 200 structures fell into 11 clusters, of which the two with the lowest target energies were the most populated (cluster 1,25; cluster 2,16). Cluster 1 had the lowest average Haddockscore and highest average buried surface area, so the 20 structures with the lowest energies were selected as the representative ensemble for further analysis.

Free docking calculations of Ca²⁺-loaded S100B with VC1 were performed using HEX 5.0 (Mustard and Ritchie, 2005; Ritchie and Kemp, 2000) on a multiprocessor Linux workstation. HEX uses 3D parametric functions, which are used to encode both surface shape and electrostatic charge and potential distributions for docking calculations.

Modeling and Analysis

A homology model of the activated leukocyte adhesion molecule (ALCAM) Ig domains 1 and 2 was constructed using VC1 as a template in MODELLER (Eswar et al., 2007; Sali et al., 1995). Electrostatic potentials were calculated with APBS (Baker et al., 2001). All molecular figures were generated using PyMol (DeLano Scientific).