Structure-Based Discovery of a Novel Class of Small-Molecule Pure Antagonists of Integrin αVβ3

Abstract

Inhibitors of integrin αVβ3 have therapeutic promise for a variety of diseases. Most αVβ3-targeting small molecules patterned after the RGD motif are partial agonists because they induce a high-affinity, ligand-binding conformation and prime the receptor to bind the ligand without an activating stimulus, in part via a charge–charge interaction between their aspartic acid carboxyl group and the metal ion in the metal-ion-dependent adhesion site (MIDAS). Building upon our previous studies on the related integrin αIIbβ3, we searched for pure αVβ3 antagonists that lack this typical aspartic acid carboxyl group and instead engage through direct binding to one of the coordinating residues of the MIDAS metal ion, specifically β3 E220. By in silico screening of two large chemical libraries for compounds interacting with β3 E220, we indeed discovered a novel molecule that does not contain an acidic carboxyl group and does not induce the high-affinity, ligand-binding state of the receptor. Functional and structural characterization of a chemically optimized version of this compound led to the discovery of a novel small-molecule pure αVβ3 antagonist that (i) does not prime the receptor to bind the ligand and does not induce hybrid domain swing-out or receptor extension as judged by antibody binding and negative-stain electron microscopy, (ii) binds at the RGD-binding site as predicted by metadynamics rescoring of induced-fit docking poses and confirmed by a cryo-electron microscopy structure of the compound-bound integrin, and (iii) coordinates the MIDAS metal ion via a quinoline moiety instead of an acidic carboxyl group.

Graphical Abstract

INTRODUCTION

The integrin receptor αVβ3 is expressed on the surface of a variety of cell types, most notably osteoclasts,¹ and also tumor cells, vascular smooth muscle cells, and neovascular endothelial cells.^2,3 The protein has been associated with a broad range of pathological processes, from osteoporosis¹ to tumor angiogenesis,^3,4 metastasis², sickle cell disease vaso-occlusion,^5-7 dermal and hepatic fibrosis,^8-13 Crohn’s disease strictures,¹⁴ viral invasion,¹⁵ supravalvular aortic stenosis associated with Williams syndrome,¹⁶ acute myelogenous leukemia,¹⁷ T-cell lymphoma,^18,19 ocular neovascular diseases,²⁰ and disruption of glomerular barrier function.²¹ However, to date, no αVβ3 antagonist drugs have been approved for clinical use,²² even though promising data were obtained in a human phase 2 study of osteoporosis.²³

Endogenous integrin ligands contain the Arg–Gly–Asp (RGD) sequence, which strongly binds to the integrins’ RGD-binding pocket, in part, via a charge–charge interaction between the ligand carboxyl group of the aspartate in the RGD sequence and a positively charged bivalent ion in the β3 metalion-dependent adhesion site (MIDAS). In addition to coordinating the MIDAS metal ion, the ligand’s carboxyl oxygens may also interact with the backbone nitrogens of Y122 and/or S123 (the latter in αIIbβ3 only) on the neighboring β1–α1 loop, and this has been proposed to move the loop toward the MIDAS, resulting in reorganization of the coordination of both the MIDAS and adjacent to MIDAS (ADMIDAS) metal ions.^24-26 The reorganization of these protein–metal ion interactions causes the loss of the interaction of the backbone carbonyl of M335 with the ADMIDAS metal ion, freeing the α6-β7 loop to undergo a major conformational change leading to a dramatic swing-out motion of the β3 hybrid domain and adoption of a high-affinity, ligand-binding state of the protein.^24,27-31 Many small-molecule antagonists of integrin αVβ3 have been patterned after the RGD sequence, but because they contain a carboxyl group that interacts with the integrin in the same way as natural ligands, many of them induce the same conformational changes and thus are considered partial agonists. Indeed, under certain experimental conditions, RGD-based molecules have been reported to prime αVβ3 to bind ligand in the absence of an activating stimulus in vitro,^32,33 and to perhaps trigger blood vessel formation in explanted tissues ex vivo and tumor angiogenesis in vivo.³⁴.

Reasoning that a pure αVβ3 antagonist, that is, one that blocks the receptor without inducing the conformational change leading to the high-affinity, ligand-binding state of the receptor might have therapeutic benefits, and inspired by crystallographic evidence that a mutant form of a fibronectin fragment (hFN10) could bind to the receptor without inducing the β3 hybrid domain swing-out motion of αVβ3,³⁵ we developed novel pure αVβ3 antagonists (TDI-4161 and TDI-3761) that do not produce β3 hybrid domain swing-out and thus do not prime the receptor to bind the ligand.³³ They achieve this—despite having a carboxyl group that coordinates the MIDAS metal ion—by forming a π–π interaction with β3 Y122 on the β1–α1 loop, thus preventing the loop from moving and disrupting the coordination of the MIDAS and ADMIDAS metal ions.³³ We confirmed the lack of swing-out motion by antibody binding, negative-stain electron microscopy (EM), and X-ray crystallography. We also showed that the compounds could inhibit bone resorption by osteoclast-like cells in vitro. Unlike the αVβ3 cyclic RGD peptide partial agonist cilengitide, however, they did not enhance aortic sprout angiogenesis at concentrations below their IC₅₀,³³ a feature of cilengitide that in other studies correlated with enhanced tumor growth in vivo.³⁴

We previously also developed novel pure antagonists of the closely related β3 integrin family member, αIIβ3,^36-40 that operate through a different mechanism that involves displacement of the MIDAS metal ion by direct binding to a β3 residue (E220) that helps coordinate the MIDAS metal ion. One of these compounds, RUC-4 (zalunfiban), is now in clinical phase testing as therapy for ST-segment elevation myocardial infarction.^40,41 To assess whether we could identify pure antagonists of αVβ3 that have a similar mechanism of action, we performed an in silico screen in the absence of the MIDAS metal ion, searching for small molecules that interact with residue E220 on the β3 subunit directly. Functional and structural characterization of a chemically optimized version of one of the identified compounds led us to discover a novel αVβ3-targeting small molecule that coordinates the MIDAS metal ion via a quinoline moiety rather than a carboxyl group and does not induce the conformational changes that lead to the high-affinity, ligand-binding state of the protein, thus defining a new class of pure αVβ3 antagonists.

MATERIALS AND METHODS

Virtual Screening Strategy: Chemical Libraries, Molecular Docking, and Filtering.

The workflow of the structure-based virtual screening strategy used to predict novel pure antagonists of integrin αVβ3 is depicted in Figure S1. Two different chemical libraries were utilized for virtual screening: a large one including ~340 million “Wait OK” compounds from the “lead-like” segment [molecular weight (MW) <375 Da and Log P < 4] of the ZINC15 database^42-45 downloaded in ready-to-dock three-dimensional (3D) representations on April 8, 2019 and another one composed of ~4.2 million “in-stock” compounds with MW > 375 Da and Log P < 5 from the ZINC15 database downloaded as ready-to-dock 3D structures on June 24, 2019. Consideration of tautomers and ionization states of these molecules expanded the datasets to ~471 and ~6.5 million compounds, respectively, which were docked to αVβ3 with Dock 3.7⁴³ or Glide High Throughput Virtual Screening (HTVS),^44,45 respectively. These docking algorithms were used to enable rapid, low-cost screening of these large datasets and were subsequently replaced by increasingly discriminating docking procedures implemented in Glide Standard Precision (SP)^44,45 and Glide Extra Precision (XP)⁴⁶ for virtual screening of pruned datasets (Figure S1).

Specifically, docking calculations for these datasets were performed using the crystal structure of αVβ3 that was determined in complex with TDI-4161 (PDB code: 6MK0³³). Prior to docking, all small molecules and metal ions were removed from the system and the most probable protonation states of ionizable protein residues at pH 7.4 were assigned using PROPKA^47,48 as implemented in Schrödinger’s “Protein Preparation Wizard” tool.⁴⁹ For a first docking-based virtual screening of the large dataset of ~471 million compounds with Dock 3.7,⁴³ CHEMGRID⁵⁰ was used to make a van der Waals grid centered on TDI-4161 using the Amber force field.⁵¹ The docking grid utilized to dock compounds with Glide HTVS was centered on TDI-4161 and extended 10 Å in all three dimensions. Docking with Glide included additional steps using Glide SP^44,45 and Glide XP,⁴⁶ which were also applied to rescore top-ranked Dock 3.7 molecules (<–75 kcal/mol). Docked ligands were sequentially filtered based on (i) the plateau of docking scores (see Figure S2 for the cutoffs applied to the initial Dock 3.7 and Glide HTVS screens), (ii) their ability to form direct interactions (within 3.0 Å) through nitrogen or oxygen atoms with carboxylic groups of β3 E220 and αV D218 or D219, (iii) lowest estimates of binding free energy ΔG(bind) from molecular mechanics/generalized Born surface area (MMGBSA) calculations,^52,53 (iv) a favorable ligand strain energy depending on the number of rotatable bonds in the molecule,⁵⁴ and (v) Tanimoto coefficients (Tc) < 0.4 with respect to other ligands within the group, as well as the 1679 annotated αVβ3 binders in the ChEMBL25 database.⁵⁵ A total of 71 compounds were ultimately purchased for experimental testing (Table S1).

Compounds.

The compounds were purchased from MolPort (Riga, Latvia) and dissolved in either 100% (1 compound) or 10% dimethyl sulfoxide (DMSO) (70 compounds) at 10 mM. ZINC72424116, which was synthesized by ChemBridge (88% purity), produced 96% inhibition in the screening assay and was selected for further study. The chemical syntheses of (R)-2-(4-(quinolin-4-yl)-1-((5,6,7,8-tetrahydro-1,8-naphthyridin-2-yl)methyl)piperazin-2-yl)ethan-1-ol hydrochloride (MSR01R), (S)-2-(4-(quinolin-4-yl)-1-((5,6,7,8-tetrahydro-1,8-naphthyridin-2-yl)methyl)-piperazin-2-yl)ethan-1-ol hydrochloride (MSR01S), 7-((4-(quinolin-4-yl)piperazin-1-yl)methyl)-1,2,3,4-tetrahydro-1,8-naphthyridine (MSR02), and 1-(quinolin-4-yl)-3-(3-(5,6,7,8-tetrahydro-1,8-naphthyridin-2-yl)propyl)urea dihydrochloride (MSR03) are described in detail in the Supporting Information (SI).

Functional αVβ3-Mediated Cell-Adhesion Assay.

Preparation of Cells Expressing Human αVβ3.

HEK-293 cells were transfected with the cDNA for αV using the pEF1/V5-His A vector and the cDNA for β3 using the pcDNA3.1 vector. HEK-293 cells expressing αVβ3 (HEK-αVβ3 cells) were identified and stable cell lines were established by repetitive sorting. HEK-αVβ3 cells for assays were counted and the number adjusted to values appropriate for each assay.

αVβ3-Mediated Cell Adhesion to Fibrinogen.

Polystyrene 96-well microtiter plates (Costar, 3590) were precoated with 5 μg/mL of purified fibrinogen in 100 mM NaCl, 50 mM Tris/HCl, pH 7.4, and then incubated with N-(2-hydroxyethyl)-piperazine-N′-ethanesulfonic acid (HEPES)-buffered modified Tyrode’s solution [HBMT; 128 mM NaCl, 10 mM HEPES, 12 mM NaHCO₃, 0.4 mM NaH₂PO₄, pH 7.4, 2.7 mM KCl, 0.35% bovine serum albumin (Fisher), 0.1% glucose] overnight at 4 °C. The wells were washed with HBMT containing 1 mM MgCl₂ and 2 mM CaCl₂ and then 50 μL of HEK-αVβ3 cells (3000 cells/μL) were added to each well. The cells had been pretreated for 20 min at room temperature with the compound to be tested (for the screening assays, final concentrations were either 50 μM compound, 0.1% DMSO or 100 μM compound, 0.6% DMSO). After ~30 min, the wells were washed three times with HBMT containing 1 mM MgCl₂ and 2 mM CaCl₂. The adherent cells were lysed and the released acid phosphatase activity was measured by adding freshly prepared phosphatase substrate (2 mg/mL) (Sigma #P4744) in 100 mM citrate, 0.1% Triton X-100. After 60 min at room temperature, the optical density was measured in a spectrophotometer at 405 nm. In each assay, 10 mM ethylenediaminetetraacetic acid (EDTA) was used as a positive control, and untreated cells in the appropriate concentration of DMSO was used as a negative control. All assays were performed in triplicate, and the values averaged. The optical density in the presence of EDTA was subtracted from the value in the presence of the compound and then converted into percent inhibition relative to the value in the absence of the compound (Table S2). The Merck αVβ3 antagonist MK-429 was included as a positive control (3 nM producing >90% inhibition) for the assay. The compound that showed greater than 95% inhibition of αVβ3-mediated cell adhesion in the screening assay (ZINC72424116) was analyzed for its IC₅₀, defined as the concentration of the test compound that reduced the adhesion of the HEK-αVβ3 cells by 50%, taking the results with untreated cells as 100% and the results in the presence of EDTA as 0%. Representative dose–response experiments of inhibition of αVβ3-mediated cell adhesion to fibrinogen by this compound, as well as MSR01S, MSR01R, MSR02, and MSR03 are shown in Figure S3A.

Monoclonal Antibody AP5 Binding Assay.

Monoclonal antibody AP5, which binds to a ligand-induced binding site on the PSI domain of β3,^33,56 was a kind gift of Dr. Peter Newman, Versiti BloodCenter of Wisconsin. HEK-αVβ3 cells were harvested, washed with HBMT once, and resuspended in HBMT containing 1 mM MgCl₂ and 2 mM CaCl₂; 5 × 10⁵ cells were incubated with the compound to be tested or cilengitide at the indicated concentrations for 20 min at room temperature and then the cells were incubated with fluorescently labeled AP5 (10 μg/mL) for 30 min at 37 °C. The cells were then washed and analyzed by flow cytometry (FACSCalibur; Becton Dickinson). The geometric mean fluorescence intensities (GMFIs) of AP5 binding to untreated cells and cells treated with 1 μM cilengitide were taken as 0 and 100% exposure of the AP5 epitope. The concentration of the test compound required to induce half-maximal exposure of the AP5 epitope as judged by exposure produced by 1 μM cilengitide was calculated and defined as EC₅₀. In cases in which even the highest concentration of the test compound did not induce 50% exposure of the AP5 epitope, the results are reported as greater than the highest concentration tested and the AP5 binding produced by that concentration reported as a percentage of the value obtained with cilengitide.

Priming Assay.

Partial agonists of αVβ3 prime the receptor to bind the fibrinogen by inducing the receptor to adopt a high-affinity, ligand-binding conformation, whereas pure antagonists do not.³³ HEK-αVβ3 cells were washed, resuspended in HBMT containing 1 mM MgCl₂ and 2 mM CaCl₂ at 2 × 10⁶ mL, and either left untreated (control) or incubated with 100 μM RGDS, or the test compound at 10 μM for 20 min at room temperature. The samples were then fixed with 4% paraformaldehyde in Dulbecco’s phosphate-buffered saline (PBS) for 40 min at room temperature, followed by quenching of the reaction with 5 mM glycine for 5 min at room temperature. After the cells were washed with HBMT, they were resuspended in HBMT containing 1 mM MgCl₂ and 2 mM CaCl₂. Alexa⁴⁸⁸-conjugated fibrinogen was then added (Invitrogen), and the mixture was incubated for 30 min at 37 °C. The cells were then washed and analyzed by flow cytometry (FACSCalibur; Becton Dickinson).

Structure-Based Ligand Optimization.

A comparison between the ZINC72424116-bound αVβ3 model and the crystal structure of αVβ3 in complex with TDI-4161 (Figure S4) suggested an opportunity to improve ZINC72424116 binding to the αV chain by replacing its p-ethoxytoluene group with the tetrahydronaphthyridine group of TDI-4161. The resulting compound, which we termed MSR01, was then made more synthetically tractable for more convenient structure–activity relationships (SAR) exploration by removing first the 2-hydroxyethyl group, which led to the design of MSR02, and then the piperazine ring, which led to the design of MSR03 (Figure S5).

Induced-Fit Docking and Metadynamics Rescoring of MSR03.

Induced-fit docking calculations of MSR03 were carried out in the presence or absence of the MIDAS metal ion, using the crystal structure of αVβ3 that was determined in complex with TDI-4161 (PDB code: 6MK0³³) and Glide XP⁴⁶ within the Schrödinger-2020-4 suite. Prior to docking, all small molecules and metal ions were removed from the system except for the MIDAS metal ion (in the set of simulations that included it), as well as the ADMIDAS and synergistic metalion-binding site (SyMBS) ions in the β3 chain and the crystallographic metal ions in the α chain. The most probable protonation states of ionizable protein residues at pH 7.4 were assigned using PROPKA^47,48 as implemented in Schrödinger’s “Protein Preparation Wizard” tool.⁴⁹ MSR03 was simulated with a positively charged tetrahydronaphthyridin moiety and a neutral quinoline group ((t) and (q), respectively in Figure S5A) according to their most probable protonation states predicted by the Schrödinger’s Epik program⁵⁷ at pH 7.4, in the presence or absence of the MIDAS metal ion. To assess whether MSR03 would form a stable direct interaction with β3 E220 in the absence of the MIDAS metal ion, the molecule was also simulated with a positively charged quinoline moiety (Figure S5B) notwithstanding a less probable protonated state for this chemical moiety at pH 7.4, according not only to Epik, but also to a reported pK_a value of 4.85.⁵⁸ The five Glide XP top-scoring binding poses of MSR03 with a positive or neutral quinoline group at the αVβ3 crystallographic ligand-binding pocket in the absence of the MIDAS metal ion, as well as the top-scoring binding poses of MSR03 with a neutral quinolone group and in the presence of the MIDAS metal ion were selected for metadynamics rescoring.

Ten independent metadynamics simulations were carried out for each of these systems using the Binding Pose Metadynamics module^59,60 within the Schrödinger suite 2020-4. MSR03 was parameterized with a charged tetrahydronaphthyridin moiety and a positive or neutral quinoline group using the OPLS3e force field⁶⁰ and the Force Field Builder tool of the Schrödinger suite for automatic generation of torsional parameters that are not present in standard OPLS3e. The MSR03-bound αVβ3 systems were solvated with SPC water molecules. Each metadynamics simulation was performed for 10 ns using Desmond 6.4⁶¹ with a bias factor of 4.0, Gaussian hills with a height of 0.05 kcal/mol and a width of 0.02 Å, and a deposition rate of 1 ps. The collective variable used for these metadynamics simulations was MSR03’s heavy atom root-mean-square deviation (RMSD) with respect to the docked pose after aligning the Cα atoms of the protein residues within 3 Å of the docked MSR03. The stability of a binding pose was assessed by calculating a so-called Comp-Score,⁶² which is the linear combination of a “PoseScore” (i.e., the average RMSD of MSR03 from its starting pose) and a “PersScore” (i.e., the fraction of frames that retain the same hydrogen bonds as the input structure in the last 2 ns of simulation, averaged over all 10 independent simulations). Specifically

$$\begin{gathered} \text{CompScore}=\text{PoseScore}-5\times \text{PersScore} \\ =\int \text{RMSD}\frac{{\mathrm{e}}^{-F(\text{RMSD})∕k_{\mathrm{b}T}}}{\int {\mathrm{e}}^{-F(\text{RMSD})∕k_{\mathrm{b}}T}}-5H \end{gathered}$$

where F(RMSD) is the free energy profile of MSR03 as a function of RMSD, k_b is Boltzmann’s constant, T is the temperature, and H is the measure of the aforementioned hydrogen-bond persistence. CompScore values were averaged over the 10 metadynamics simulations and are reported here as average values with error bars illustrating the 25th and 75th percentiles.

Equilibrium Molecular Dynamics (MD) Simulations of the Most Stable MSR03 Poses as per Metadynamics Rescoring.

The lowest CompScore poses of MSR03 with a neutral quinoline group in the presence or absence of the MIDAS metal ion, or a positively charged quinoline group in the absence of the MIDAS metal ion, were selected for subsequent 1 μs equilibrium molecular dynamics (MD) simulations for each system. For these simulations, only the αVβ3 head domain (residues 1–437 of αV and residues 109–352 of β3) along with the ligand and crystallographic SyMBS and ADMIDAS cations, in the absence or presence of the MIDAS metal ion, were used. The systems were solvated with explicit SPC water and ionized by a 150 mM NaCl solution. The solution ions were placed at least 10 Å away from crystallographic cations. Brownian dynamics and classical MD simulations were performed using Desmond 6.4⁶¹ and the OPLS3e force field.⁶⁰ The simulation strategy for each system consisted of the following steps: (1) 100 ps of Brownian dynamics, NVT (constant number (N), volume (V), and temperature (T)), T = 10 K, timestep = 1 fs, and all of the solute heavy atoms restrained; (2) 12 ps of MD using the Berendsen thermostat, NVT, T = 10 K, timestep = 1 fs, and all of the solute heavy atoms restrained; (3) 12 ps of MD using the Berendsen thermostat and barostat, NPT (constant number (N), pressure (P), and temperature (T)), T = 10 K, timestep = 1 fs, P = 1 atm, and all of the solute heavy atoms restrained; (4) 12 ps of MD using Berendsen thermostat and barostat, NPT, T = 300 K, timestep = 1 fs, P = 1 atm, and all of the solute heavy atoms restrained; (5) 24 ps of MD using the Berendsen thermostat and barostat, NPT, T = 300 K, timestep = 1 fs, and P = 1 atm; and (6) four MD replicas of 250 ns each using the Nose–Hoover thermostat and Martyna–Tobias–Klein (MTK) extended Lagrangian barostat, NPT, T = 300 K, timestep = 2 fs, and P = 1 atm, yielding a 1 μs simulation trajectory for each system.

The 1 μs simulation trajectories (4000 frames) of the systems with MSR03 containing a protonated quinoline group in the absence of the MIDAS or a neutral quinoline group in the presence of the MIDAS metal ion (the ligand with a neutral quinoline group simulated in the absence of the MIDAS metal ion did not remain in the pocket for any of the simulation replicas) were further analyzed by Structural Interaction Fingerprint (SIFt) analysis⁶³ with 9-bit representations using an in-house python script. The 9-bit representation is based on the following nine types of interactions: apolar interactions (carbon–carbon atoms in contact), face-to-face (Aro_F2F) and edge-to-face (Aro_E2F) aromatic interactions, electrostatic interactions with positively (Elec_ProP) and negatively (Elec_ProN) charged protein residues, hydrogen-bond interactions with the protein as the hydrogen-bond donor (Hbond_proD) or hydrogen-bond acceptor (Hbond_proA), one-water-mediated H-bond (Hbond_1wat), and two-water-mediated H-bond (Hbond_2wat) interactions. A distance cutoff of 4.5 Å was used to define apolar interactions, while a cutoff of 4 Å was used to describe aromatic and electrostatic interactions. For the H-bonds, we used a cutoff of 3.5 Å. Interactions were calculated only for protein side chain atoms. The probability of each interaction was estimated using a two-state Markov model, sampling the transition matrix posterior distribution using standard Dirichlet priors for the transition probabilities as described by Noé et al.⁶⁴

Negative-Stain Electron Microscopy.

Recombinant αVβ3 ectodomain was produced and purified as previously reported⁶⁵ except that stably expressing HEK-293S GnT1⁻cells⁶⁶ were used instead of CHO-lec cells. The ectodomain at 0.89 mg/mL in HEPES-buffered NaCl, pH 7.2, 1 mM CaCl₂, 1 mM MgCl₂ was mixed with 20 mM HEPES, pH 7.4, 150 mM NaCl, 1 mM MgCl₂, 2 mM CaCl₂ and MSR03 in 1 M NaCl to give final concentrations of 0.02 mg/mL protein, 312 μM MSR03, 17.5 mM HEPES, pH 7.4, 275 mM NaCl, 1.75 mM CaCl₂, 0.87 mM MgCl₂. As a control, 1 M NaCl without MSR03 was used. Similarly, cilengitide at 10 mM in DMSO was mixed with the protein solution to give final concentrations of 0.02 mg/mL protein, 10 μM cilengitide, 0.1% DMSO, 20 mM HEPES, pH 7.4, 150 mM NaCl, 2 mM CaCl₂, 1 mM MgCl₂; a DMSO control was also prepared. A 5 μL aliquot of integrin solution (~0.007 mg/mL) was applied to a glow-discharged thin carbon film that had been evaporated onto a plastic-covered copper grid. After 30 s, the grid was blotted, washed twice with deionized water, and stained with 0.7% (w/v) uranyl formate as described.⁶⁷ Grids were imaged with a Philips CM10 electron microscope operated at an acceleration voltage of 80 kV using a defocus of about −1.5 μm and a magnification of 52,000×, yielding a calibrated pixel size of 2.7 Å at the specimen level. Images (Figure S6) were collected with an AMT 3k × 5k ActiveVu CCD camera (53 images for αVβ3 (control), 56 images for αVβ3 + MSR03, and 53 images for αVβ3 + cilengitide). Particles were picked manually using EMAN2,⁶⁸ yielding datasets of 12,194 particles for αVβ3 (control), 10,786 particles for αVβ3 + MSR03, and 15,735 particles for αVβ3 + cilengitide. The particles were extracted into 192 × 192 pixel images, centered, and normalized in EMAN2. The particles were classified into 100 groups using K-means classification procedures implemented in SPIDER.⁶⁹ Class averages with clear structural features were manually assigned to represent αVβ3 in a compact-closed, extended-closed, or extended-open conformation; the remaining averages were not assigned (Figure S7). The criteria for differentiating the extended-closed from the extended-open conformation included whether the legs are crossed and whether the β3 hybrid domain appears to face more toward the headpiece rather than away from the headpiece.

Cryo-EM Sample Preparation, Data Collection, and Image Processing.

Vitrified grids of MSR03-bound αVβ3 (~0.22 mg/mL) were prepared with a Vitrobot Mark IV (Thermo Fisher Scientific) set at 4 °C and 100% humidity. The 3.5 μL aliquots were applied to glow-discharged Quantifoil 300 mesh R 1.2/1.3 copper grids, which were blotted for 5 s with a blot force setting of −1 and then plunged into liquid nitrogen-cooled ethane.

Image stacks were collected with a Titan Krios electron microscope (Thermo Fisher Scientific) in the Cryo-EM Resource Center at the Rockefeller University operated at an acceleration voltage of 300 kV. Data were collected with a K3 direct detection camera (Gatan) in super-resolution counting mode using SerialEM.⁷⁰ The nominal magnification was 81,000×, corresponding to a calibrated pixel size of 0.86 Å at the specimen level, and the defocus range was −1.5 to −2.5 μm. Exposures of 2 s were dose-fractionated into 40 frames (50 ms per frame), with a dose rate of 20 electrons/pixel/s, resulting in a total dose per frame of 1.35 electrons/Ų.

Image stacks were motion-corrected, dose-weighted, summed, and binned over 2 × 2 pixels in MotionCor2.⁷¹ The contrast transfer function (CTF) parameters were determined with cryoSPARC,⁷² which was used for all subsequent image processing steps. Blob picking was used to autopick 5351 micrographs without templates. The 1,588,825 picked particles were extracted into 400 × 400 pixel images, which were binned over 5 × 5 pixels, and subjected to two-dimensional (2D) classification into 100 classes. Twenty-six of the resulting classes were used to generate an ab initio model and 10 class averages were used as templates to repick the images, yielding a total of 5,007,268 particles. These particles were reextracted, binned over 5 × 5 pixels, and subjected to three rounds of 2D classification into 100 classes to remove junk particles.

The averages of 91 classes showed well-defined integrins and the 3,547,477 particles from these classes were combined and subjected to 3D classification into four classes using the ab initio model and three noise maps as references. The particles in the class resulting from the ab initio model as reference (1,515,783 particles) were subjected to a second 3D classification, this time using the resulting density map and two noise maps as references. After removing the particles that were assigned to the two noise maps, the remaining 1,251,376 particles were recentered, reextracted into 400 × 400 pixel images, and subjected to nonuniform refinement, yielding the final map at 2.5 Å resolution. The local-resolution map was calculated in cryoSPARC. Cryo-EM data collection statistics are provided in Table S3.

RESULTS

Virtual Screening and Compound Selection for Experimental Testing.

We docked a large chemical library of ~340 million compounds with MW < 375 Da and Log P < 4 using Dock 3.7,⁴³ as well as another library of ~4.2 million compounds with MW > 375 Da and Log P < 5 using Glide HTVS,^44,45 to the crystal structure of αVβ3 that was determined in complex with TDI-4161 (PDB code: 6MK0³³). This was done after removal of TDI-4161, and according to the virtual screening procedure illustrated in Figure S1 and described in detail in the Materials and Methods section. A first round of filtering was then performed by removing the ligands with a Dock 3.7 docking score greater than −75 kcal/mol or a Glide HTVS docking score greater than −5 kcal/mol (Figure S2). Considering only compounds with nitrogen or oxygen atoms within 3 Å from the carboxyl groups of β3 E220 and αV D218 or D219, a total of 200,000 compounds were selected from the combined sets. These compounds were then redocked using the more accurate Glide SP scoring function,^44,45 and a total of 1007 compounds with MW < 375 Da and 1142 with MW > 375 Da were selected, based on their negative docking scores, their forming interactions with β3 E220 and αV D218 or D219, and their availability for purchase. These compounds were then docked one more time using the higher accuracy Glide XP procedure and filtered to keep only compounds with negative docking scores and direct interactions with β3 E220 and αV D218 or D219. This filtering procedure resulted in the selection of 190 compounds with MW < 375 Da and 293 compounds with MW > 375 Da. Estimates of binding energy (ΔG(bind)) for these compounds were calculated using the MMGBSA method. Compounds from the two sets were merged and their total number was further reduced to 86 by applying the following criteria: (i) keeping only compounds that interact with β3 E220 when their nitrogen atom is positively charged, (ii) removing compounds with calculated positive ΔG(bind) values, (iii) removing compounds with large strain energy, (iv) removing similar compounds within the group based on a Tc > 0.4 calculated using radial fingerprints, keeping the ones with the lowest ΔG(bind) values, and (v) removing compounds that have any resemblance with the reported 1679 αVβ3 binders in ChEMBL using the same Tc > 0.4 cutoff. Of the 86 remaining compounds, 71 were available from vendors and were purchased for experimental testing (Table S1).

Experimental Validation and Confirmation of Pure αVβ3 Antagonists.

Of the 71 compounds, two were not soluble in 10% DMSO at 50 μM. The remaining compounds variably inhibited αVβ3-mediated cell adhesion in the screening assay (Table S2). The racemic compound ZINC72424116 was the most potent, producing 95.6% (n = 3) inhibition of adhesion at 50 μM and thus was selected for further study. Its IC₅₀ was determined using triplicate measurements on four occasions (the first with the first sample obtained from the manufacturer, the second and third with the second sample obtained, and the fourth with the third sample obtained) and found to be 13.7, 2.0, 3.8, and 3.0 μM, respectively (Table 1 and Figure S3A).

Table 1.

IC₅₀^a and EC₅₀^b Values for the Original Zinc Compound and the MSR Derivative Compounds

compound	IC50 (μM)a	EC50 (μM)b
ZINC72424116	5.6 ± 0.38 (n = 4)c	>50 (n = 3)
MSR01R	1.3 ± 0.33 (n = 5)	>10 (n = 4)
MSR01S	1.0 ± 0.30 (n = 5)	>10 (n = 4)
MSR02	2.0 ± 1.73 (n = 9)	>10 (n = 5)
MSR03	1.6 ± 0.58 (n = 6)	>10 (n = 5)d

^aThe concentration of the test compound required to inhibit half-maximal adhesion of cells expressing human αVβ3 to immobilized fibrinogen.

^bThe concentration of the test compound required to induce half-maximal exposure of the AP5 epitope.

^cAverage value from five triplicate determinations using three different samples of compounds.

^dStudies performed with MSR03 prepared in DMSO. Repeat study of MSR03 dihydrochloride prepared in 1 M NaCl indicated that at 300 μM MSR03 did not increase AP5 binding above baseline and so its EC₅₀ is >300 μM (n = 3).

Antibody AP5 binds to an epitope on αVβ3 that only becomes accessible upon the conformational change in αVβ3 that is induced by ligand binding. Despite its inhibition of αVβ3-mediated cell adhesion to fibrinogen, ZINC72424116 did not expose the AP5 epitope, with its EC₅₀ > 50 μM, the highest concentration tested, and ~10-fold higher than its average IC₅₀ (Table 1 and Figure S3B). Figure S4 shows the selected docking pose of ZINC72424116 from virtual screening using Glide XP and the filtering criteria used to prioritize the purchase of molecules for experimental testing (see Materials and Methods section and Figure S1). According to this model, the ZINC72424116 S-enantiomer forms (i) hydrogen bonds between its 2-hydroxyethyl group and both the carboxyl group of αV D219 and the backbone amine nitrogen of β3 A218, (ii) a cation–π interaction between the charged nitrogen atom of its piperazine ring and the aromatic ring of αV Y178, and (iii) a charged interaction between its positively charged quinoline ring and the carboxyl group of β3 E220.

Structure-Based Design of the Novel MSR Chemical Series.

A structural comparison between the ZINC72424116-bound αVβ3 model from Glide XP docking and the crystal structure of αVβ3 in complex with TDI-4161 (Figure S4) revealed a much weaker interaction of ZINC72424116 with the αV chain compared to TDI-4161. Reasoning that this interaction could be improved by replacing the p-ethoxytoluene group of ZINC72424116, which is not involved in a direct polar interaction with the protein in our predicted model, with the tetrahydronaphthyridine moiety of MK-429 and TDI-4161,³³ which forms a hydrogen-bond and charge–charge interaction with αV D218, as well as a π–π stacking interaction with αV Y178, we designed MSR01 (see the Supporting Information for the synthesis and characterization). Both enantiomers of MSR01 were synthesized and determined to have IC₅₀ values of 1.3 ± 0.33 (mean ± SD; n = 5) and 1.0 ± 0.30 μM (n = 5) (MSR01R and MSR01S, respectively; Table 1 and Figure S3A). The lower IC₅₀ values for both compounds compared to that of ZINC72424116 are consistent with the prediction that MSR01 would form more and tighter ligand–receptor interactions. Both enantiomers had EC₅₀ > 10 μM (Table 1).

To simplify the synthesis of MSR01 for more convenient SAR exploration, the chiral center of the compound was eliminated by removing the 2-hydroxyethyl group from the piperazine ring, producing MSR02. Removal of the hydrogen bond between the backbone carboxylic oxygen of αV A218 and the 2-hydroxylethyl group of MSR01 did not significantly affect ligand binding as demonstrated by a comparable IC₅₀ of 2.0 ± 1.73 μM (n = 9; Table 1 and Figure S3A). It also did not affect the ability to expose the AP5 epitope, with an EC₅₀ >10 μM (Figure S3B). To further simplify chemical synthesis, the piperazine ring of MSR02 was also removed, yielding the design of MSR03, which was determined to have an IC₅₀ of 1.6 ± 0.58 μM (n = 6; Table 1 and Figure S3A). MSR03 also did not expose the AP5 epitope, with an EC₅₀ >10 μM tested with MSR03 prepared in DMSO (Figure S3B) and an EC₅₀ > 300 μM tested with MSR03 dihydrochloride prepared in 1 M NaCl (n = 3).

Neither MSR02 nor MSR03 Prime αVβ3 to Bind Fibrinogen.

The partial αVβ3 agonist peptide RGDS (100 μM)-primed αVβ3 expressed on HEK-293 cells to bind fibrinogen, but neither MSR02 nor MSR03 enhanced fibrinogen binding above the control level (Figure 1).

Figure 1.

Neither MSR02 nor MSR03 prime αVβ3 to bind fibrinogen. HEK-αVβ3 cells were washed and either left untreated (control) or incubated with 100 μM RGDS, 10 μM MSR02, or 10 μM MSR03 for 20 min at room temperature and then fixed with 4% paraformaldehyde for 40 min at room temperature. After the cells were washed, they were resuspended in a buffer containing 1 mM Mg²⁺ and 2 mM Ca²⁺ and Alexa⁴⁸⁸-conjugated fibrinogen added for 30 min at 37 °C. The cells were then washed and analyzed by flow cytometry. Data are displayed as mean ± SD with n = 6 for all values. GMFI, geometric mean fluorescence intensity.

Unlike Cilengitide, MSR03 Does Not Induce the Extended-Open Conformation of the αVβ3 Ectodomain.

Negative-stain EM of the αVβ3 ectodomain showed that in the presence of 275 mM NaCl, the vehicle for MSR03, it adopts mostly the compact-closed and only rarely the extended-open conformation (Figure S6A). Similar results were found with 0.1% DMSO, the vehicle for cilengitide, as we previously reported,³³ and so only the results with 275 mM NaCl are reported in this paper. Addition of MSR03 at 312 μM did not appear to have a discernable effect on the αVβ3 conformation (Figure S6B), whereas cilengitide at 10 μM had a strongly activating effect, with the majority of αVβ3 now adopting the extended-open conformation (Figure S6C). Quantification (Figures S7 and 2) confirmed the visual impressions, showing that only 5% of αVβ3 in the resting state adopted the active, extended-open conformation and that addition of MSR03 did not increase this percentage. In sharp contrast, and in accord with our previous report,³³ addition of cilengitide increased the percentage of receptors in the active, extended-open conformation to 82% (Figures S7 and 2).

Figure 2.

Binding of MSR03, unlike cilengitide, does not affect the conformation of αVβ3. Top: Typical averages of classes categorized as being in the compact-closed (red border), extended-closed (blue border), or extended-open (green border) conformation. Side length of individual averages is 51.8 nm. Below: Percentage of molecules in each of the conformational states in untreated αVβ3 (control) and in the presence of MSR03 (312 μM) or cilengitide (10 μM). Percentages were calculated as the fraction of integrins in a particular conformation with respect to all of the assigned particles.

Predicted Binding Mode of MSR03 to αVβ3.

For a more reliable prediction of the mode of binding of MSR03 to αVβ3, we used induced-fit docking and metadynamics rescoring of the ligand, which has been shown to yield more accurate predictions of the structures of protein–ligand complexes.⁶² The virtual screen was based upon the quinoline moiety in ZINC72424116 being protonated, but the group has a pK_a value of 4.85,⁵⁸ and therefore a very large fraction of MSR03 molecules are likely to have a neutral quinoline at physiological pH. Accordingly, we conducted the induced-fit docking and metadynamics rescoring of MSR03 molecules encompassing both the most probable neutral quinoline moiety (Figure S5A) and a positively charged quinoline moiety (Figure S5B). Specifically, computations of the former were carried out in the presence or absence of the MIDAS metal ion, whereas the latter was simulated in the absence of the MIDAS metal ion to mimic the possible displacement of the MIDAS metal ion in exchange for a direct interaction with its coordinating β3 E220 residue. The five top-scoring docking poses of MSR03 that were rescored by metadynamics with a neutral quinoline moiety in the presence or absence of the MIDAS metal ion, or a positively charged quinoline group in the absence of the MIDAS metal ion, are shown in Figures 3, 4, and 5, respectively, along with their docking scores and CompScore values averaged over 10 metadynamics simulations.

Figure 3.

Top-scoring docking poses and scores (in kcal/mol) obtained for MSR03 with a neutral quinoline group in integrin αVβ3 in the presence of the MIDAS metal ion. The top-ranked metadynamics-rescored poses (lowest CompScore values) selected for equilibrium molecular dynamics simulations (poses #2 and #5) are indicated by red boxes. The integrin αV and β3 chains are shown in orange and green cartoon representations, respectively. The αV D218 side chain forming polar interactions with the ligand is shown as sticks, as are the MSR03, E220, and Y122 residues, the latter two shown as points of reference. SyMBS, MIDAS, and ADMIDAS metal ions are shown as purple spheres from left to right. CompScore values calculated for each of the five top-scoring complexes obtained from induced-fit docking over 10 independent metadynamics runs are shown as box plots with error bars denoting the 25th and 75th percentiles.

Figure 4.

Top-scoring docking poses and scores (in kcal/mol) obtained for MSR03 with a neutral quinoline group in integrin αVβ3 in the absence of the MIDAS metal ion. The top-ranked metadynamics-rescored pose (lowest CompScore value) selected for equilibrium molecular dynamics simulations (pose #3) is indicated by a red box. Integrin αV and β3 chains are shown in orange and green cartoon representations, respectively. Residues of the αV and β3 chains forming polar interactions with MSR03 (D218 and R214, respectively) are shown as sticks, as are MSR03, E220, and Y122, the latter two residues shown as points of reference. SyMBS and ADMIDAS metal ions are shown as purple spheres. CompScore values calculated for each of the five top-scoring complexes obtained from induced-fit docking over 10 independent metadynamics runs are shown as box plots with error bars denoting the 25th and 75th percentiles.

Figure 5.

Top-scoring docking poses and scores (in kcal/mol) obtained for MSR03 with a positively charged quinoline group at integrin αVβ3 in the absence of the MIDAS metal ion. The top-ranked metadynamics-rescored pose (lowest CompScore value) selected for equilibrium molecular dynamics simulations (pose #1) is indicated by a red box. The integrin αV and β3 chains are shown in orange and green cartoon representations, respectively. Residues of the β3 and αV chains forming polar interactions with the ligand (E220 and D218, respectively) are shown as sticks, as are the MSR03 and Y122, the latter shown as a point of reference. SyMBS and ADMIDAS metal ions are shown as purple spheres. CompScore values calculated for each of the five top-scoring complexes obtained from induced-fit docking over 10 independent metadynamics runs are shown as box plots with error bars denoting the 25th and 75th percentiles.

Based on the assumption that the lower the CompScore value, the more stable a binding pose is, we selected poses #2 and #5 of MSR03 with a neutral quinoline group in the presence of the MIDAS metal ion, as well as pose #3 of MSR03 with a neutral quinoline group and pose #1 of MSR03 with a positively charged quinoline group, both simulated in the absence of the MIDAS metal ion, for further evaluation by equilibrium MD simulations (see Materials and Methods section for details). While the MSR03 with a neutral quinoline group simulated in the absence of the MIDAS metal ion (the aforementioned pose #3) either left the pocket completely or lost all interactions with the β3 chain in all simulation replicas, the ligand with either the neutral quinoline group simulated in the presence of the MIDAS metal ion (the aforementioned poses #2 and #5) or the quinoline group in its protonated form simulated in the absence of the MIDAS metal ion (the aforementioned pose #1) showed tighter binding, although they still left the pocket in more than 20% of simulation frames.

Figure 6 shows differences in ligand–receptor interactions that are formed with more than 10% probability during these simulations of MSR03-bound αVβ3, starting from MSR03 binding poses #2 and #5 with the neutral (Figure 6A,B) or protonated (Figure 6C) quinoline group and calculated as averages over four simulation replicas.

Figure 6.

Probabilities of ligand–protein interactions formed by MSR03 with a neutral quinoline group in the presence of the MIDAS metal ion or a positively charged quinoline group in the absence of the MIDAS metal ion during MD simulations. Interaction types formed by MSR03 with (A, B) neutral quinoline group in the presence of the MIDAS metal ion (induced-fit metadynamics-rescored poses #2 and #5, respectively) or (C) positively charged quinoline group in the absence of the MIDAS metal ion, with protein side chains during simulations and calculated as averages over the four replicas amounting to 1 μs of simulation time are: carbon–carbon atomic interactions (Apolar, pink), aromatic edge-to-face (Aro_E2F, light green) and face-to-face (Aro_F2F, dark green) interactions, electrostatic interactions with the protein negatively charged (Elec_ProN, purple), one-water-mediated and two-water-mediated hydrogen-bond interactions (Hbond_1Wat and Hbond_2Wat, light and dark blue, respectively), and hydrogen bond with the protein as the hydrogen-bond acceptor (Hbond_ProA, light orange). Only interactions with an average probability above 10% are displayed. Error bars refer to the 5th and 95th percentile of the probabilities calculated using a two-state Markov model.

The predicted most stable poses #2 and #5 of MSR03 with a neutral quinoline group in the presence of the MIDAS metal ion (Figure 6A,B, respectively) formed similar interactions with probability greater than 50% with αV residues D150, Y178, Q180, and D218, as well as β3 residues S121, Y122, S123, A218, and E220. While β3 residues were not only involved in apolar interactions, the molecule mostly formed apolar interactions with αV D150, but also in electrostatic interactions via the protein negatively charged residue, a hydrogen bond with the protein serving as hydrogen-bond acceptor, and water-mediated interactions, albeit with much lower probability. Interactions with αV Y178 and Q180 were mostly apolar although Y178 was also involved in aromatic face-to-face interactions, as well as aromatic face-to-edge interactions with lower probability. Finally, the molecule involved αV D218 in apolar and electrostatic interactions via the protein negatively charged residue and also hydrogen-bond interactions with the protein serving as the hydrogen-bond acceptor, the latter with slightly lower probability in pose #2. Most importantly, the MSR03 quinoline moiety appeared to directly coordinate the MIDAS metal ion via its nitrogen lone pairs of electrons.

While similar interactions with αV D150, αV Y178, αV Q180, αV D218, and β3 A218 were recorded with a probability larger than 50% for the predicted most stable pose #1 of MSR03 with a positively charged quinoline group in the absence of the MIDAS metal ion, interactions with β3 residues S121, Y122, and S123 had much lower probabilities compared to MSR03 with a neutral quinoline group, and an additional, unique apolar interaction with β3 A252 was recorded for this pose (Figure 6C). Furthermore, as expected, the interaction between the positively charged quinoline group of MSR03 and β3 E220 was strengthened by both electrostatic and hydrogen-bond interactions with the protein in the absence of the MIDAS metal ion.

To experimentally assess the results from computational docking, we produced a single-particle cryo-EM map of MSR03-bound αVβ3 at 2.5 Å resolution (Figure S8). Rigid-body fitting of the crystal structure of TDI-4161-bound αVβ3 (PDB code: 6MK0³³), after removing the TDI-4161 molecule, into this map revealed a clear extra density at the interface between the integrin αV and β3 subunits, indicating the presence of a ligand (see gray mesh in Figure 7).

Figure 7.

Fit of the predicted most stable binding poses of MSR03 with a neutral or positively charged quinoline moiety in αVβ3 from induced-fit docking-metadynamics rescoring into the cryo-EM density map of MSR03-bound αVβ3. Cartoon representations of the αV and β3 chains of integrin αVβ3 shown in transparent orange and green colors, respectively. MSR03 is shown in cyan color for the predicted poses of the molecule with a neutral quinoline group simulated in the presence of the MIDAS metal ion (panels (A) and (B); same poses as poses #2 and #5 in Figure 3, respectively) or a positively charged quinoline group simulated in the absence of the MIDAS metal ion (panel (C), same pose as pose #1 in Figure 5). Cryo-EM density contoured at σ = 0.32 within 1.5 Å of the ligand is shown as a gray mesh and the density attributed to the MIDAS metal ion contoured at σ = 1.03 is colored in magenta. Shown in sticks are the side chains of the αV D218, β3 Y122, and β3 E220 residues. Distances between the tetrahydronaphthyridine group of MSR03 and the carboxyl oxygens of αV D218, as well as the quinoline group of MSR03 and the MIDAS metal ion (when the group is neutral) or a carboxyl oxygen of β E220 (when the group is positively charged), are indicated with dotted lines.

Despite the relatively high resolution of the map, this extra density was poorly defined and insufficient to determine the MSR03 binding pose unambiguously with cryo-EM refinement tools, especially in its interaction with the β3 subunit, whose interactions had shown differences during MD simulations of the molecule with a neutral or charged quinoline group (compare Figure 6A-C). Thus, we fitted our models of αVβ3 in complex with MSR03 with a neutral (Figure 3, poses 2 and 5) or positively charged (Figure 5, pose 1) quinoline moiety from induced-fit metadynamics rescoring to the cryo-EM density map of MSR03-bound αVβ3 (Figure 7A-C, respectively). As shown in Figure 7, the density attributed to MSR03 was reasonably consistent with all of these predicted stable poses. Although unable to distinguish between them, the map supported several conclusions from the molecular dynamics simulations of the three predicted stable poses of MSR03, including a much higher flexibility of the quinoline group compared to the tetrahydronaphthyridine moiety in the pocket, as well as the formation of stable electrostatic and polar interactions between the tetrahydronaphthyridine group and the carboxyl oxygens of αV D218 (see distance values in Figure 7). Importantly, the density map showed a very clear and strong density for the MIDAS metal ion, supporting both the prediction that the neutral form of MSR03 dominates and that the quinoline directly coordinates the MIDAS metal ion. Moreover, the cryo-EM structure showed that there was no reorganization of the MIDAS or ADMIDAS metal-ion coordination, no shift in the β1–α1 loop, and no swing-out of the hybrid domain, all of which are characteristic of the conformational change induced by partial agonists. The poses of MSR03 with a neutral quinoline group in the presence of the MIDAS metal ion showed apolar interactions with β3 Y122, but not the stronger π–π stacking interaction seen in the TDI-4161-bound αVβ3 crystal structure (Figures 6 and S9).³³ However, they also showed apolar interactions with S121 and S123, the latter being absent in the TDI-4161-bound αVβ3 crystal structure.³³ Since we cannot determine whether every αVβ3 molecule in complex with MSR03 contains a MIDAS metal ion, we cannot exclude the possibility that a small percentage of MSR03 molecules with the quinoline group in the charged state displace the MIDAS metal ion and directly coordinate E220, considering that the quinoline nitrogen is predicted to be within 2.7 Å of one of the E220 carboxyl oxygens (Figure 7). However, it must be noted that this interaction is only seen in 50% of simulation frames (Figure 6), denoting a high flexibility of the molecule even in its charged form.

DISCUSSION

Despite the large number of human disorders in which targeting αVβ3 has been proposed as a therapeutic intervention,⁷³ and the promising in vitro, in vivo, and clinical data on the efficacy of a partial agonist αVβ3 antagonist in osteoporosis,²³ no αVβ3 antagonists have been approved for human use. One theoretical consideration for the difficulty in developing an effective αVβ3 antagonist is that the small-molecule antagonists that have been tested, including cilengitide and MK-429, are partial agonists that prime the receptor to bind ligand, and thus may not prevent the downstream signaling that accompanies ligand engagement.^33,34 We previously were able to use structure-guided design to develop pure αVβ3 antagonists that do not activate the receptor by establishing π–π stacking interaction with β3 Y122 and thus preventing the movement of the β1–α1 loop and disruption of the MIDAS and ADMIDAS metal-ion coordinations.³³ We have also developed pure antagonists of the closely related αIIbβ3 receptor that have advanced to clinical testing.^36-41 These lock the receptor in the inactive state by a different mechanism, namely, displacement of the MIDAS MG²⁺ ion by direct interaction with β3 E220. To broaden the repertoire of potential agents for testing the hypothesis that integrin pure antagonists may have therapeutic advantages compared to partial agonists, we sought to make additional pure αVβ3 antagonists.

In silico screening of large libraries of compounds against high-resolution crystal structures of integrin αVβ3 stabilized by pure antagonists make it possible to try to identify novel pure antagonists of integrin αVβ3 that operate via specific mechanisms. Thus, by filtering in silico screened compounds in the absence of the MIDAS metal ion based on their predicted interaction with residue E220 in the β3 subunit, we aimed at biasing the search for compounds that would recapitulate the distinctive mechanism of pure antagonism that we previously described for the inhibition of αIIbβ3 by compounds RUC-2 and RUC-4.^36-39 We were able to identify compound ZINC72424116 from this screen and show that it inhibited the adhesion of cells expressing αVβ3 to fibrinogen, an αVβ3 ligand, with an IC₅₀ of 5.6 μM. To increase ligand-binding affinity, we substituted a tetrahydronaphthyridine group, which has a high affinity for the αV headpiece region of the RGD-binding pocket^33,74 for the p-ethoxytoluene group of ZINC72424116, achieving a 3.5-fold increase in potency in MSR03. We demonstrated that MSR03 fulfills the criteria for a pure antagonist since it does not expose the epitope for the ligand-induced binding site of antibody AP5, does not prime αVβ3 to bind the fibrinogen, and does not induce receptor extension and swing-out when analyzed by negative-stain EM. In contrast, cilengitide exposes the AP5 epitope, primes αVβ3 to bind the fibrinogen, and induces receptor extension as judged by negative-stain EM.³³

Results from induced-fit docking followed by metadynamics rescoring led to a model of interaction between MSR03 and αVβ3 that depended on the protonation state of the quinoline moiety of the small molecule. Specifically, this moiety would interact directly with β3 E220 in the absence of the MIDAS metal ion but only if positively charged. On the other hand, the same molecule with a much more probable neutral state of the quinoline group at pH 7.4 would stably bind the αVβ3 protein in the presence of the MIDAS metal ion by directly coordinating the metal ion. This latter model was supported by a 2.5 Å resolution cryo-EM density map, which revealed a strong density at the expected location of the MIDAS metal ion. Poses of MSR03 with a neutral quinoline group in the presence of the MIDAS metal ion formed apolar interactions with β3 residue Y122, as well as MIDAS-coordinating residues S121 (through side chain oxygen) and S123 (via water-mediated interaction with its side chain), suggesting the possibility of a molecular mechanism of pure antagonism concertedly mediated through these residues, although it did not form the stronger π–π stacking interaction seen in the TDI-4161-bound αVβ3 crystal structure with Y122.³³ Taken together, we consider it most probable that the key factor in MSR03 acting as a pure antagonist is its lack of an aspartic acid carboxyl group with an oxygen interacting with the backbone nitrogen of Y122 on the β1–α1 loop, as well as the establishment of concerted apolar interactions with residues involved in the coordination of the MIDAS metal ion, which are sufficient to keep the β1–α1 loop from moving toward the MIDAS, with no need for strong π–π stacking interactions with Y122 to prevent the conformational change.

MSR03 thus provides a novel way of producing pure integrin antagonism by coordinating the MIDAS metal ion with a quinoline rather than a carboxyl moiety that potentially may be applicable to a wide range of integrin receptors. In addition, our results establish the proof of concept for employing in silico screening against high-resolution αVβ3 experimental structures stabilized by pure antagonists and structure-guided optimization to identify additional integrin pure antagonists that operate through novel mechanisms. Such compounds can serve as valuable starting points for medicinal chemistry modifications to enhance potency and drug-like properties.