A fragment-based approach to discovery of Receptor for Advanced Glycation End products inhibitors

Abstract

The Receptor for Advanced Glycation End products (RAGE) is a pattern recognition receptor that signals for inflammation via the NF-κB pathway. RAGE has been pursued as a potential target to suppress symptoms of diabetes and is of interest in a number of other diseases associated with chronic inflammation, such as inflammatory bowel disease and bronchopulmonary dysplasia. Screening and optimization have previously produced small molecules that inhibit the activity of RAGE in cell-based assays, but efforts to develop a therapeutically viable direct-binding RAGE inhibitor have yet to be successful. Here, we show that a fragment-based approach can be applied to discover fundamentally new types of RAGE inhibitors that specifically target the ligand-binding surface. A series of systematic assays of structural stability, solubility, and crystallization were performed to select constructs of the RAGE ligand-binding domain and optimize conditions for NMR-based screening and co-crystallization of RAGE with hit fragments. An NMR-based screen of a highly curated ~14 000-member fragment library produced 21 fragment leads. Of these, three were selected for elaboration based on structure-activity relationships generated through cycles of structural analysis by X-ray crystallography, structure-guided design principles, and synthetic chemistry. These results, combined with crystal structures of the first linked fragment compounds, demonstrate the applicability of the fragment-based approach to the discovery of RAGE inhibitors.

1 |. INTRODUCTION

The Receptor for Advanced Glycation End products (RAGE) is an immunoglobulin, pattern-recognition cell surface receptor that plays an important role in the inflammatory response upon activation by damage-associated molecular patterns (DAMPs).^1–5 RAGE is present at low concentration in several tissues, including kidneys, and liver, as well as cells in the cardiovascular, nervous, and immune systems. However, RAGE levels are increased during certain physiological states or pathological conditions, such as chronic inflammation.⁶ Overexpression of RAGE results from activation of many cellular signaling pathways, such as the MAPK, JNK, and NF-κB pathways^6,7 that, in turn, leads to extracellular secretion of additional RAGE binding partners, generating a positive feedback loop.^4,5,8,9

Advanced Glycation End products (AGEs) are reactive species that arise from elevated levels of glucose.⁸ AGE-modified proteins and lipids bind to many cell surface receptors, including RAGE, and are linked to multiple inflammation-related pathologies via NF-κB signaling.⁵ The RAGE-AGE interaction has been shown to accelerate atherosclerosis in both in vitro and in vivo models.¹⁰ RAGE overexpression in diabetic mice contributes to glomerulosclerosis.¹¹ Another RAGE ligand, calprotectin (CP, S100A8/S100A9 heterodimer), is elevated in the colon in inflammatory bowel diseases which, as a result of the positive feedback loop, causes RAGE levels to also increase.^12,13 RAGE expression is crucial for normal lung development, and lethality can occur if its expression is hindered during development by chronic hyperoxia, as seen in neonates with bronchopulmonary dysplasia.¹⁴ Together, these findings demonstrate the breadth in pathogenesis caused from ligand induced RAGE activation.

The structure of RAGE consists of three extracellular Ig domains (V, C1, and C2), a single membrane span, and a small cytosolic tail. The tandem of the V and C1 domains are not structurally independent, but rather form an integrated structural unit independent of the C2 domain.¹⁵ The exact nature of how ligands activate the protein is not fully understood. A proposed model for signaling suggests that RAGE can be assembled into higher order oligomeric states, shifting a pre-existing equilibrium through the binding of oligomeric ligands.^16,17 This supports a mechanism whereby RAGE assembly into higher order oligomers increases the efficiency of signaling and drives an increase in RAGE-mediated inflammation.^5,16 Inhibition of ligand binding to RAGE has therefore been targeted as a means to suppress adverse effects from over-stimulated RAGE-mediated inflammation observed in a range of disorders.¹⁸ Nearly all ligand activators of RAGE are understood to bind to the V domain,¹⁸ so our inhibitor discovery program has been directed to this domain.

Several RAGE inhibitors are currently under study, including some in clinical trials.^11,18 However, structural data to confirm that these molecules bind directly to RAGE is lacking. For example, the small molecule RAGE inhibitor FPS-ZM1 was found, by surface plasmon resonance (SPR), to inhibit the extracellular portion of RAGE from binding protein ligands HMGB1 (K_i = 148 nM) and S100B (K_i = 230 nM).¹⁹ Additionally, the presence of FPS-ZM1 inhibited activation of primary microglia by AGEs, which led to a decrease in RAGE expression, lower oxidative stress levels, and overall lower levels of inflammation.²⁰ Another small molecule, azeliragon, has been reported to inhibit the interaction of RAGE with HMGB1, S100B, and AGEs.²¹ This molecule inhibited the binding of the extracellular portion of RAGE to Aβ_1–42 with an IC₅₀ = 500 nM, as determined by a fluorescence polarization assay.¹⁸ Azeliragon was also shown to reduce Aβ plaque deposition and levels of inflammatory cytokines by reducing neuroinflammation, as well as to improve cerebral blood flow.²² RAGE antagonistic peptides (RAPs) derived either from HMGB1 or S100P have also been designed to act as inhibitors. A RAP derived from the sequence of the COOH-terminal motif of HMGB1 inhibits the binding of HMGB1, as shown via pull-down and cell-based studies.²³ Using Elisa and cell-based assays, a RAP derived from S100P blocked RAGE interactions with HMGB1, S100P, and S100A4.²⁴ Despite the interest in these molecules as RAGE inhibitors, their mechanism of action remains ambiguous; no structural data are available to rigorously establish their binding modes and locations on RAGE, if indeed they directly bind. Given these issues and the uncertainty of success with advancing the limited number of candidates to clinical applications, we have adopted a fragment-based discovery strategy to generate new RAGE inhibitors specifically directed to its ligand-binding domain.

Fragment-based discovery (FBD) can provide a substantial increase in efficiency over high throughput screening for the generation of inhibitors of protein-protein interactions due to the ability of fragments to find multiple small sub-pockets across large binding interfaces. This method also provides access to a much higher degree of chemical diversity, lead molecules that are easier to elaborate, and can identify allosteric effectors outside of the target binding site. The generation of high affinity inhibitors by covalent attachment of one or more fragments binding to different nearby sites leverages the “linkage effect”,²⁵ ie, linking produces compounds with a theoretical affinity that is the product of the affinity of the constituent fragments (eg, linking two fragments with K_d values of 1 mM can produce a 1 μM inhibitor).^26,27

Here we have employed the “SAR by NMR” FBD approach,²⁸ which has been extensively used for the design of inhibitors of protein interactions (eg, ²⁹). In this technique, chemical shift perturbations (CSPs) of the amide backbone peaks in the NMR spectra of the protein target detect binding of small (<300 Da), weakly binding “fragment” molecules. This can enable specific identification of hits at the binding interface from libraries of chemically diverse molecular fragments.³⁰ Multiple rounds of NMR screens are used to identify fragments that bind to different sites within the target-binding interface, although some screens directly reveal multiple fragment binding sites. Iterative cycles of chemical elaboration based on structure-activity relationships (SAR), design of linkers between fragments, and structural analysis of protein-ligand complexes are then undertaken to generate high affinity inhibitor compounds.

This report demonstrates the applicability of FBD to RAGE. We describe the properties of various constructs of RAGE assessed for screening and structural studies. The V domain alone was used for the NMR screens of a curated library of ~14 000 fragments. A tandem construct with both the V and C1 domains was used for determining X-ray crystal structures of lead fragments bound to RAGE to establish their location and orientation on the RAGE ligand-binding surface. After elaboration of select lead fragments, structures of two linked compounds were determined. These data demonstrate the feasibility of linking lead molecules based on structures of VC1-fragment complexes and revealed an additional fragment-binding site in the interaction interface. Our results show that a FBD approach can be applied for the discovery of new RAGE inhibitors.

2 |. EXPERIMENTAL METHODS

2.1 |. Molecular biology

DNA fragments were PCR amplified with primers containing 5′ NdeI and 3′ XhoI restriction sites and the sequences for RAGE V_23–119, V_23–132, VC1_23–233, and VC1_23–243. The PCR products were subcloned into pET15b vectors (Novagen), which includes a N-terminal 6-Histidine tag and a thrombin cleavage site. Site-directed PCR mutagenesis was used to produce VC1_23–243 mutants E132A, W230A, W230Y, and E132A/W230A. The DNA for VC1_23–231 was PCR amplified with primers containing 5′ BamHI and 3′ XhoI restriction sites and subcloned into a pSV281 vector, which contains a N-terminal 6-Histidine tag and a tobacco etch virus protease (TEV) cleavage site.

2.2 |. Protein expression and purification

RAGE V_23–132, VC1_23–233, VC1_23–243 and mutants were overexpressed in E. coli strain OrigamiB (DE3) (Novagen) cells grown at 37 °C to OD₆₀₀ 0.6–0.8, after which the cells were moved to 20 °C for ~30 minutes. Protein expression was induced with 0.5 mM IPTG and expression was monitored overnight at 20 °C. Cells were harvested at 4 °C and 6000 rpm for 20 minutes. Pellets were lysed using sonication in 20 mM Tris-HCl (pH 8), 20 mM imidazole, 300 mM NaCl (Buffer A) at 4 °C in the presence of Roche cOmplete EDTA-free protease inhibitor cocktail tablet. Clarified lysate was purified using a HisTrap column (GE Healthcare) equilibrated with Buffer A. Protein was eluted with a 4 CV linear gradient to 20 mM Tris-HCl (pH 8), 500 mM imidazole, 300 mM NaCl (Buffer B) and then dialyzed in 20 mM MES (pH 6), 300 mM NaCl (Buffer C) at 4 °C overnight. Following dialysis, the 6-His tag was removed with thrombin (1–2 units/mg protein) incubated at room temperature 1–2 hours. Protein was further purified over a SP sepharose (GE Healthcare) column equilibrated with Buffer C and eluted with 18 CV of 20 mM MES (pH 6) and 1 M NaCl (Buffer D).

RAGE V_23–119 was expressed as noted above and purified with a modified protocol. After cell lysis with sonication in 20 mM Tris-HCl (pH 8), 20 mM imidazole, 300 mM NaCl (Buffer A) at 4 °C in the presence of protease inhibitor cocktail tablet, soluble protein was separated from the inclusion bodies with centrifugation at 20000 rpm at 4 °C. Guanidine hydrochloride (4 M GndHCl) and sonication was used to extract the insoluble portion of RAGE several times from the inclusion bodies and refolding of the protein was performed in Buffer A by dialyzing away the denaturant. The soluble and the refolded portions were pooled together and run over a HisPrep column (GE Healthcare), eluting with a 9 CV linear gradient to 20 mM Tris-HCl (pH 8), 750 mM imidazole, 300 mM NaCl and then dialyzed twice in 50 mM Tris (pH 7.5), 50 mM NaCl (Buffer E) for 4–6 hours at 4 °C. Following dialysis, the 6-His tag was removed with thrombin (1–2 units/mg protein) incubated at room temperature 1–2 hours. The protein was further purified over a SourceS column (GE Healthcare) equilibrated with Buffer E and eluted with 18 CV of 50 mM Tris (pH 7.5) and 750 mM NaCl (Buffer F). Unlabeled RAGE VC1_23–243 was dialyzed into crystallization buffer containing 10 mM sodium acetate (pH 5.2) and the other proteins were dialyzed into 25 mM HEPES (pH 6.8) and 75–100 mM NaCl.

The RAGE VC1_23–231 construct was a kind gift from Professor Gregers R. Andersen (Aarhus University, Denmark).³¹ It was overexpressed in Escherichia coli T7 SHuffle cells grown at 37 °C to OD₆₀₀ 0.6–0.8, after which cells were moved to 22 °C for ~30 minutes. Protein expression was induced with 1 mM IPTG and expression was monitored overnight at 22 °C. Cells were harvested at 4 °C, centrifuged at 6000 rpm for 20 minutes and subsequently lysed with sonication in 50 mM HEPES (pH 7.5), 30 mM imidazole, 300 mM NaCl, 1 mM PMSF, 1 protease inhibitor tablet. After lysis and centrifugation at 20000 rpm for 20 minutes, lysate was loaded onto a HisPrep column (GE Healthcare), washed with 50 mM HEPES (pH 7.5), 30 mM imidazole, 1 M NaCl, and 1 mM PMSF, then eluted with a seven CV linear gradient to 50 mM HEPES (pH 7.5), 500 mM imidazole, 300 mM NaCl, 1 mM PMSF. Protein fractions were dialyzed into 50 mM HEPES (pH 7.5), 250 mM NaCl overnight in the presence of TEV protease (1:30 TEV: protein). The protein was then passed over a SourceS column and eluted with a 18-CV linear gradient to 50 mM HEPES (pH 7.5), 650 mM NaCl. The protein was dialyzed into crystallization buffer containing 25 mM HEPES (pH 7.5) and 250 mM NaCl. ¹⁵N-enriched V and VC1 proteins were expressed and purified as described with the exception that cells were grown in minimal media with ¹⁵N-ammonium chloride as the sole nitrogen source.

2.3 |. Chemical synthesis and characterization

Fragment analogs 1–23 and 28–29 were purchased from commercial vendors and used as delivered. Fragment analogs 24–27 and 30–31 were prepared as described in the Supporting Information.

2.4 |. Protein NMR

¹⁵N-enriched proteins were dialyzed overnight in NMR buffer: 25 mM HEPES (pH 6.8), 75–100 mM NaCl and 5% D₂O. ¹⁵N-¹H HSQC spectra were acquired at 25 °C on a Bruker Avance 900 equipped with a triple resonance cryoprobe. Acquisition parameters were 2048 points in the direct and 128 points in the indirect dimension with 32 scans. For NMR screening, ¹⁵N-enriched V_23–119 was used and the buffer contained ~4.5% DMSO-d₆ instead of D₂O. ¹⁵N-¹H SOFAST-HMQC³² spectra were acquired at 25 °C on Bruker Avance 600 equipped with a triple resonance cryoprobe. Typical acquisition parameters were 1024 points in the direct and 96 points in the indirect dimension with 28 scans. Spectra were acquired at protein concentrations of 36 to 50 μM with small molecule concentrations of 200 to 800 μM. FPS-ZM1 (Sigma-Aldrich) and azeliragon (VTV Therapeutics) were obtained as kind gifts from Prof. Berislav Zlokovic laboratory (USC Keck School of Medicine).

2.5 |. X-ray crystallography

For crystallization of RAGE VC1_23–243 and mutants, crystals were grown by hanging drop vapor diffusion at 25 °C using the published protocol from Koch et al.¹⁶ Briefly, 2 μL of protein was mixed with 2 μL of 0.1 M Na cacodylate buffer, 0.2 M ZnOAc (pH 6.5), and 11% PEG 8000. Crystals were soaked in 12% glycerol for 1 minute prior to flash freezing. Crystallization of RAGE VC1_23–231, was based on the published protocol.³¹ Crystals were grown by hanging drop vapor diffusion at 4 °C by mixing 2 μL of protein with 2 μL of 2.5 M NaOAc (pH 7.5). Rhombus shaped crystals appeared after 4–6 days. Crystals were soaked either overnight in 2 μL of 20 mM small molecule + 2.48 M NaOAc + 25% sucrose or 1 min in 2 μL of 2.48 M NaOAc + 25% sucrose prior to flash freezing them in liquid nitrogen. The data sets were collected at the Life Sciences Collaborative Access Team beamline 21-ID-G or 21-ID-F at the Advanced Photon Source, Argonne National Laboratory or at the Advanced Light Source (ALS) at the Lawrence Berkeley National Laboratory SIBYLS beamline. Collected data were reduced and scaled with HKL2000.³³ Phase calculation was by molecular replacement using Phaser³⁴ and refinement was performed with Phenix refine.³⁵ The final model was built in COOT³⁶ and visualized using Pymol (version 2.0 Schrodinger, LLC). Crystals of RAGE VC1_23–243 belonged to the P62₁2₁2₁ orthorhombic space group and crystals of RAGE VC1_23–231 belonged to the P6₂ hexagonal space group. Data collection and refinement statistics can be found in Table S1.

3 |. RESULTS AND DISCUSSION

3.1 |. Selection of RAGE constructs for NMR screens

Two different V and two different VC1 constructs of RAGE have been prepared and studied.^15,31 Since fragment-based approaches require large quantities of protein, a systematic analysis of these four constructs was performed to minimize protein production while obtaining high quality results for analyses. After optimization of expression and purification protocols of soluble protein for each construct, we obtained higher yields for both VC1 constructs (6 mg/L) than for the V constructs (1.5–3 mg/L). However, during the purification of the V constructs, it was evident that most of the V domain was insoluble and remained in the inclusion bodies, so we developed a refolding protocol.

After unfolding in 4 M GndHCl, a rapid one-time dialysis step was successful in producing folded protein, although a significant portion of protein precipitate was observed. To circumvent this problem, several changes to the dialysis step were tested. However, the rapid one-time dialysis worked best for producing the highest overall yields. ¹⁵N-¹H HSQC NMR spectra were acquired to confirm that the refolded protein was structurally similar to the soluble protein; the peaks in the two spectra overlaid nicely, confirming that the two sources of folded V domain could be combined. This raised the yield to 6 mg/L of culture of V_23–119 and 4 mg/L of culture of V_23–132.

With these improvements, the yield of various constructs was comparable, so the selection of which construct to use for the NMR screening could be based on the quality of 2D ¹⁵N-¹H NMR spectra (Figure 1). The spectra of all four constructs contain well-dispersed peaks indicative of folded globular proteins. Given the similarities in yield, spectral quality, and the fact that known protein ligands bind primarily to the V domain, we decided to proceed with the V construct vs VC1. In addition, this construct has approximately half the number of peaks in its NMR spectra, which aids the identification CSPs caused by ligand binding. Since V_23–119 has a 67% higher yield than V_32–132, all NMR fragment screens were carried out with this shorter construct.

FIGURE 1.

Comparison of the 600 MHz ¹⁵N-¹H HSQC NMR spectra of RAGE V_23–119, V_23–132, VC1_23–233, and VC1_23–243

3.2 |. NMR screen of the fragment library

An NMR screen of V_23–119 was carried out using a library of ~14 000 molecular fragments acquired from various vendors and curated to maximize chemical diversity while removing frequent false positive binders.³⁷ The fragments are arrayed into 12 plates of 96 wells, each with a mixture of 12 fragments. ¹⁵N-¹H SOFAST-HMQC spectra were acquired for each pool over the course of 2 days. The data were processed using an in-house automated pipeline and rapidly inspected to identify pools that caused perturbations of the backbone amide chemical shifts (CSPs).³⁷ The pools that gave rise to CSPs were filtered for residues in the target-binding surface, based on the resonance assignments for the V domain³⁸ and the V-binding sites reported for AGEs³⁸ and S100B.¹⁶ In all, on-target CSPs were observed for 47 out of 1152 pools of fragments in this library. The hit pools were deconvoluted by acquiring individual spectra for each member of the pool to identify which fragment gave rise to the observed CSP. Figure 2 shows an example of the similarity of the CSPs observed for a fragment pool and one of the fragments (1) in the pool. Notably, the peaks exhibiting CSPs correspond to key residues in the binding surface, including K52, Q100, A101, and K110 (Figure 2, inset). Importantly, some pools of fragments also caused CSPs of several key ligand-binding residues, but the overall shift patterns were different. For example, CSPs of G70 and R77 were observed with some fragment pools, indicating they likely contained a fragment that bound at different locations from fragment 1 (-Figure S1A,C).

FIGURE 2.

Binding of hit fragment 1 to V. A, Overlay of the 600 MHz ¹⁵N-¹H SOFAST HMQC spectrum of V_23–119 with the pool of 12 fragments containing 1. B, Overlay of the 600 MHz ¹⁵N-¹H SOFAST HMQC spectrum of V_23–119 with fragment 1 showing it has similar CSPs. The inset shows an expanded region of the spectrum containing four of the critical ligand-binding residues (K52, Q100, A101, K110)

Some pools of fragments that showed CSPs gave rise to substantially reduced CSPs (or none) upon deconvolution. This phenomenon can be attributed to interactions between fragments promoting interactions with the protein. For other pools, general peak broadening was observed, which might be attributed to either intermediate exchange or fragment-induced protein aggregation (Figure S1B). Hence, after deconvolution of the 12-fragment pools and validation of binding by NMR titrations, 21 different fragments were deemed “hits” of sufficient promise for the subsequent steps (Table 1). While the hit fragments are relatively diverse, the majority of the structures are heteroaromatic ring systems, many of them bicyclic. As a group, the hits trend toward relatively high lipophilicity, particularly when considering their modest fragment-like size. However, no clear trend associating higher lipophilicity with stronger shifts was immediately evident. Despite the low hit rate (0.16%), the screen confirms the ligandability of RAGE. Notably, molecules with a 5,6 bicyclic group (including the carboxylic acid-containing indole acid and benzofuran groupings) caused some of the largest CSPs and are structurally distinct from existing RAGE inhibitors such as FPS-ZMT1 and azeliragon.

TABLE 1.

Data for two commercially available RAGE inhibitors and hits from the curated fragment library identified in the NMR screen

Molecule #	Structure	Chemical class	Chemical shift magnitude	AlogPa
Azeliragon			No shift	6.78
FPS-ZM1			No shift	5.31
1		Indole acid	Strong	4.05
2		Indole acid	Weak	3.00
3		Benzofuran	Strong	3.56
4		Benzofuran	Strong	2.77
5		Benzofuran	Weak	3.90
6		5,6-bicyclic	Weak	1.72
7		5,6-bicyclic	Weak	1.75
8		(Heterocyclic) Biphenyl	Weak	2.01
9		(Heterocyclic) Biphenyl	Weak	4.07
10		(Heterocyclic) Biphenyl	Weak	2.67
11		(Heterocyclic) Biphenyl	Weak	3.57
12		Phenyl Furan	Weak	2.33
13		Phenyl Furan	Weak	3.82
14		Phenyl Furan	Weak	3.53
15		Other	Strong	2.86
16		Other	Weak	1.69
17		Other	Weak	3.15
18		Other	Peak disappearance	1.87
19		Other	Peak disappearance	3.43
20		Other	Peak disappearance	4.08
21		Other	Peak disappearance	2.27

^aAlogP calculated using the Adriana code from Molecular Networks.

We also performed NMR titrations with known RAGE inhibitors, FPS-ZM1, and azeliragon, using the same-labeled V domain construct used to discover the fragment hits. Interestingly, no CSPs were observed in the NMR spectra with either molecule at similar concentrations to our hits.

3.3 |. Optimization of RAGE crystallization for determining structures of fragment complexes

After completing the fragment screen, the next objective was to obtain crystal structures of protein-fragment complexes. These experiments were performed with VC1 because the V domain is only half of the VC1 structural unit and is meta-stable, and VC1 had been previously crystallized. In following our previously published protocol for crystallizing V_23–243,¹⁶ we found that crystals appeared slowly, were small in size, and did not consistently diffract to atomic resolution (ie, <2.4 Å). Several approaches were attempted to try to improve the crystals, including decreasing temperature, varying pH values, microseeding, and macroseeding. Microseeding produced relatively small crystals more quickly (3–5 days) but these diffracted over a broad range, from 1.8 to 3 Å. However, crystals allowed to grow an additional 1 to 2 weeks were significantly larger and generated highly reproducible diffraction in the range of 2 Å.

Toward obtaining crystal structures in complex with the hit fragments and future analogs, we carefully examined the crystal lattice in the structure of VC1_23–243. This analysis revealed contacts between molecules in the unit cell, including a salt bridge between R98 and E132/W230, with R98 3.5 Å from E132 and 5.3 Å from W230 (Figure 3A–C). This raised concern because R98 is well within the ligand-binding surface previously mapped for AGEs. To ensure that this part of the VC1 surface was not occluded in the crystal, several mutants were constructed with the goal of altering crystal packing, including E132A, W230A, W230Y, E132A/W230A, and E132D/W230Y. However, all efforts to crystallize these mutants proved unsuccessful, even after microseeding with wild-type crystals.

FIGURE 3.

Crystal lattice and packing in VC1_23–243 and VC1_23–231 crystals. VC1_23–243 is shown in (A-C) and VC1_23–231 in (D-F). (A, D) Layout of the V and C1 domains in the corresponding lattices. (B, E) Ribbon diagram representation of the lattices color-coded to identify the V and C1 domains. (C, F) Close-up view of the salt bridge between R98 and E132/W230 in VC1_23–243 crystals and the exposure of R98 in VC1_23–231 crystals

With these concerns in mind, we next turned to the shorter VC1_23–231 construct to determine if the crystals were more promising, particularly since there are clear differences in the crystal packing lattice contacts.³¹ For example, R98 is exposed in a solvent channel, distal from E132 and W230 (Figure 3D–F). Using similar, but not identical, conditions to those published, crystals were apparent after 2 days. These crystals consistently diffracted to better than 2.1 Å resolution, even higher than the published structure, presumably the result of a slower growth rate than the published conditions, under which crystals appeared overnight.

To obtain structures of RAGE-fragment complexes, both co-crystallization and ligand soaking into already formed crystals were investigated. DMSO was required to solubilize fragments and NMR was used to confirm that the structure of the protein was not affected by the addition of up to 25% v/v of this co-solvent. Conditions for co-crystallization were varied, including ratios of fragment to protein (10:1, 5:1, 3:1, and 2:1), incubation times, concentration of DMSO, and seeding into drops. However, none of these co-crystallization trials was successful, so we attempted ligand soaking instead. To prevent crystals from cracking during the soaking period, we tested the addition of various types of polyethylene glycol polymers (PEG) (300–8000) to the crystal prior to addition of DMSO. From the different PEG molecules, it was found that a crystal soaked for 7 hours in PEG 3350 or 8000 was protected from ~22% DMSO shock (35 mM fragment). A cryoprotectant screen revealed that crystals transferred to new drops with sucrose could tolerate the same amount of DMSO. Soaking times were also varied; however, incubation longer than 1 day either cracked or dissolved crystals. The final protocol adopted involved transferring the crystal to a new drop with sucrose and then adding the fragment in DMSO to the drop overnight. High fragment concentrations were used to maximize the occupancy of the fragment in the ligand-binding site. Since the 22% DMSO required for a fragment concentration of 35 mM would occasionally lead to crystal dissolution, the fragment concentrations were lowered to 20 mM, which enabled lowering DMSO to 12.5%.

3.4 |. Structural analysis and elaboration of hit fragments bound to VC1

Using this soaking technique, a high-resolution structure was obtained with fragment 1 bound to RAGE. We named this binding location “site 1.” Figure 4 shows the co-crystal structure of 1 (the fragment with the largest CSPs in the NMR screen) bound to VC1; both 2Fo-Fc and polder omit maps³⁹ (Figure S2) were generated to confirm the location and orientation of the ligand. Key ligand interaction residues E50, K52, R98, Q100, K110, and N112 are highlighted. Interesting intermolecular contacts include a charge-charge interaction between the carboxylic acid of 1 and VC1 K52 as well as interactions between aromatic pi orbitals of the fragment and charged residues of the protein. Of note, the binding pocket is only weakly hydrophobic and no significant hydrophobic interactions were identified.

FIGURE 4.

Co-crystal structure of VC1 with fragment 1. Side chains are rendered and labeled for key contact residues in VC1. The 2Fo-Fc electron density map for the ligand (red) is rendered as blue mesh

Using this structural information, we designed and synthesized analogs of 1 (Table 2, 22–27). NMR CSPs showed that each of these molecules binds at the same site as 1. Notably, several of these analogs were found to cause significantly larger CSPs compared to the parent fragment with 24 and 25 causing the largest CSPs (Figure 5). Soaking experiments were performed and high-resolution crystal structures were obtained for analogs 22, 23, and 24. All three were bound in site 1 with functional groups making interactions with the protein similar to the parent molecule 1, as shown in Figure S3 for 23 and 24. Together, these data on the analogs suggest that in certain instances the N-methyl indole may better anchor the molecule, and confirms that replacement of the 3-phenyl ring with a less lipophilic pyrazole is tolerated, as is exchanging the 7-methyl group for a 6-methoxy.

TABLE 2.

Molecular structures of selected fragments and analogs. Molecules 22 and 23 are analogs of parent fragment 2 that bind in both site 1 and site 2. Molecules 24–27 are analogs of fragment 1 that have larger CSPs than the parent. Fragments 28 and 29 are analogs of parent fragment 17 that bind in site 2 and have larger CSPs than the parent

Molecule #	Structure	Chemical class	Chemical shift magnitude	AlogPa
22		Indole acid	Weak	3.74
23		Indole acid	Weak	3.69
24		Indole acid	Very strong	2.49
25		Indole acid	Strong	2.48
26		Indole acid	Strong	3.75
27		Indole acid	Very strong	4.26
28		Other	Weak	3.02
29		Other	Strong	2.77

^aAlogP calculated using the Adriana code from Molecular Networks.

FIGURE 5.

Overlay of ¹⁵N-¹H SOFAST-HMQC NMR spectra of ¹⁵N-enriched V alone (black) and in the presence of fragment 1 (red) and either analog 24 (blue, at left) or 25 (blue, at right). Both analogs show larger CSPs compared to the parent

The majority of hit fragments exhibited CSPs consistent with binding in site 1. However, a few fragments had CSP patterns that differed. Some, such as 17, exhibited CSPs of R48 and Q100, suggesting the presence of a second site. The structure of the complex of 17 with VC1 had clear electron density engaging R48. However, the electron density around the rest of the fragment was relatively poor, consistent with relatively weak binding. The weakness of the electron density also led to some difficulty in determining the orientation of the fragment in the binding site. Careful analysis of the 2Fo-Fc and polder omit maps (Figure S2) was critical to overcoming this challenge.

Targeted commercial purchase and synthetic chemistry were undertaken to improve affinities and better anchor the fragment in site 2. Replacement of the 3-hydroxyl group with a carboxylic acid group in the 4 position (Table 2, analog 28) caused a slight increase in the CSPs of key residues, as did combining the 3-hydroxyl with the 4′-carboxylic acid (Table 2, analog 29). Although the CSPs for these analogs were larger than those of the parent (Figure 6), they are still rather modest overall. Further optimization efforts will be required to produce better site 2 binders.

FIGURE 6.

Overlay of ¹⁵N-¹H SOFAST-HMQC NMR spectra of RAGE alone (black) with RAGE in the presence of fragment 17 (red) and either analog 28 or 29 (blue). Both analogs exhibit larger CSPs than the parent

3.5 |. First steps to generating linked compound inhibitors

The co-crystal structures of several site 1 fragments showed electron density for a second molecule in site 2. However, further investigation and refinement of these structures revealed the molecule in site 2 was the cryoprotectant sucrose (Figure S4). The sucrose molecule was stabilized by interactions with Q100 and R48, as with the fragments described above that bind in site 2. An NMR titration with sucrose was performed to test if VC1 binds sucrose with appreciable affinity, but the spectrum revealed no significant CSPs. Hence, we attributed sucrose binding in site 2 to mass action due to its high concentration as the cryoprotectant. Concerned that sucrose may inhibit fragment molecules from binding at site 2, we sought alternative cryoprotectants. However, out of all of the cryoprotectants screened, sucrose best protected crystals from dissolution over long periods of time and therefore it was kept for subsequent soaking experiments. Given the low affinity and the fact that we did observe fragments binding in site 2, we concluded that using sucrose as cryoprotectant was not a substantial problem.

In fact, some fragment complexes, such as 22, an analog of 2 that was identified using targeted commercial purchase to expand SAR around 1, were found to occupy both site 1 and site 2 (Figure S5). The electron density around the analog is weak in both sites, which makes the precise mode of binding difficult to establish. However, the electron density is sufficient to confirm two different molecules are bound and that these are not solvent molecules. Unlike 1, which makes a hydrogen bond contact with E50, 22 uses salt bridges to make contacts in sites 1 and 2. Given that it appears in both sites 1 and 2 with a short distance between the two fragments in the crystal structure (~8 Å), 22 might be considered a promising candidate to link with itself. However, the potential channel between the two sites is obstructed by Q100 (Figure S5). To better enable fragment linking, we sought additional analogs whose binding might maintain an open channel between the sites.

Rather than searching for new molecules that bind to both sites, we turned to soaking compounds already identified as site 1 and site 2 binders simultaneously into VC1 crystals. This strategy proved successful and a structure of a ternary complex with fragments 1 and 17 bound was determined. The resolution attained was 1.82 Å, with occupancies greater than 0.7 in both sites. In the structure, the two fragments are only ~4 Å from one another, a distance that should be easily amenable to linkage. Although the binding mode of fragment 1 is consistent with earlier results, fragment 17 was found to occupy a reversed orientation and was shifted within the binding site relative to what was observed for 17 alone (Figure 7). Analog 29 has slightly larger CSPs than fragment 17 and is more soluble. Thus, we also crystallized it in a ternary complex with 1 (Figure 8A). In this structure, the electron density for 29 is stronger compared to the ternary complex with 17, consistent with tighter binding. The carboxylic acid group interacts more effectively with R48 relative to the hydroxyl group in fragment 17 and the hydroxyl group at the 3-position of 29 makes an additional contact with Q100. Moreover, this hydroxyl group of 29 extends toward 1 and is therefore a promising lead for developing a linker between the two fragments.

FIGURE 7.

Co-crystal structures of VC1 with fragment 17 and with fragments 1 and 17. A, The chemical structure of 17 (left) and the crystal structure of the complex with VC1, showing residues contacting the fragment (center) and the fit of the molecule in the site 2 (right). Side chains are rendered and labeled for the contact residues in VC1. The 2Fo-Fc electron density maps for the ligand (red) are rendered as blue mesh. B, Crystal structure of the ternary complex showing the fit of the ligands in site 1 and site 2

FIGURE 8.

Co-crystal structures of VC1 complexes with fragments 1 and 29 and linked compounds 30 and 31. A, Ternary complex with 1 and 29 showing the fragments have the same orientation as the complex with hit fragments 1 and 17. The electron density for 29 is stronger than the density for 17, consistent with its larger CSPs and suggesting tighter binding. B, Linked compound 30 with a three atom linker between 1 and 29. Note that the linked compound binds in the same orientation as the two isolated fragments in panel A. C, Linked compound 31 with a two-atom linker between fragments 1 and 29. Note that 31 bind in the opposite orientation relative to 30 in (B) and occupies a new site 3. D, Overlay of complexes with 30 and 31 showing differences in their binding sites. E, Surface representation of VC1 with the three fragment binding sites colored

The open channel and short distance between the fragments observed in both ternary structures support the design of a linked RAGE inhibitor, despite the weak binding of the isolated fragments. To establish the feasibility of linking, we synthesized molecules (30, 31) connecting fragments 1 and 29 with ether linkages. A crystal structure of VC1 with linked compound 30 is shown in Figure 8B. Comparing this structure to the corresponding ternary complex reveals a remarkable degree of similarity in the location and orientation of both fragments in the VC1 ligand-binding interface. The clarity of the electron density throughout the whole compound is notable.

Linked compound 31 has a shorter, two-atom (CO) linker, and in the structure of the complex, we found the site 1 moiety was flipped by 180°. As a consequence, the site 2 moiety occupies a completely different site (Figure 8C,D). Close inspection of this structure provided details into the contacts in this new “site 3,” including a strong interaction with K39 and an additional contact with Y113. These two residues are particularly interesting, as a few of the initial hit fragments exhibited CSPs of these residues, and therefore may serve as logical starting points for incorporating the new site 3 into further inhibitor designs (Figure 8E). Overall, these results provide great promise that the fragment-based approach can be used to develop high affinity, direct binding RAGE inhibitors.

4 |. CONCLUDING REMARKS

We have successfully optimized a protocol for a high level of production of the isolated V and VC1 domain constructs for NMR screening of fragments and determining the X-ray crystal structures of the corresponding complexes. From a library of ~14 000 curated fragments, we found 21 hits that bind in the ligand-binding interface, as defined by NMR CSPs. A number of hit fragments have been co-crystallized with VC1, which led to the characterization of two binding sites in close proximity to each other. Elaboration provided analogs with larger NMR CSPs and stronger ligand density in crystal structures. Linking fragments that bound to the two different sites with a three-atom linker afforded a compound that bound in the same orientation as two constituent fragments, demonstrating the viability of linking fragments in this system. Remarkably, an analog of this linked compound with a two-atom linker bound in the reversed orientation and revealed an alternative, third site for fragment binding. Interestingly, this new site extends beyond the surface where the natural ligand S100B binds and thus represents a unique opportunity for further elaboration. Our ability to identify three different fragment binding sites (Figure 8E) using a combination of NMR screening and X-ray crystallography represents a considerable savings over the standard NMR-based fragment discovery approach, which requires multiple rounds of screening to identify different sites.

One limitation in our study is that, at this early stage, we have not directly shown that our molecules inhibit RAGE signaling. Given the very weak binding affinities, demonstrating inhibition of RAGE activity in cellular assays would require such high concentrations that it would not be possible to differentiate RAGE inhibition from general cell toxicity. Experiments have been performed in an effort to demonstrate that a RAGE binder could outcompete a native ligand for RAGE. It was clear from the outset that this endeavor would be an uphill battle, most importantly due to the vast difference in affinity of a natural ligand vs the small molecule RAGE binders, which moreover have yet to be optimized for aqueous solubility. Nevertheless, we performed a large number of experiments using our best binding fragment to challenge the interaction of RAGE-V or RAGE-VC1 with S100B, S100A8/S100A9 or S100A12. Unfortunately, in all of these experiments, we were stymied by the very high fragment concentration required to outcompete the ligand and finding that the fragment was also capable of binding to the ligand. We were thus unable to obtain convincing data.

The results reported here represent the essential first step toward the development of RAGE inhibitors with unique scaffolds compared to the two molecules FPS-ZM1 and azeliragon currently being evaluated as potential drugs. Importantly, the same NMR titrations used to identify our novel RAGE binders revealed no CSPs caused by FPS-ZM1 and azeliragon. This leads us to conclude that the previously known compounds either do not interact with RAGE or do not interact in the same manner as the fragments and linked compounds reported here. Given the slow and uncertain progression of existing RAGE inhibitors toward clinical application, there is value in the pursuit of new chemical scaffolds that interact with the target in a fundamentally different manner than existing candidates. Having established the existence of multiple fragment binding sites in and adjacent to the ligand interaction surface and that linked compounds can be generated, the next steps will be to develop higher affinity, more soluble linked compounds, and measure their ability to block physical interaction of natural RAGE ligands and activation of receptor signaling pathways toward validating the action of direct-binding RAGE inhibitors.

DATA AVAILABILITY STATEMENT

Crystallography data that support the findings of these studies are openly available in the Protein Data Bank (PDB) at https://www.rcsb.org using accession codes 6XQ1, 6XQ3, 6XQ5, 6XQ6, 6XQ7, 6XQ8, 6XQ9, 7LM7, 7LMW. The NMR data and protein constructs that support these studies are available upon request from corresponding author.