Methotrexate Inhibits the Binding of the Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV-2) Receptor Binding Domain to the Host-Cell Angiotensin-Converting Enzyme-2 (ACE-2) Receptor

Abstract

As the severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) virus mutates, finding effective drugs becomes more challenging. In this study, we use ultrasensitive frequency locked microtoroid optical resonators in combination with in silico screening to search for COVID-19 drugs that can stop the virus from attaching to the human angiotensin-converting enzyme 2 (hACE2) receptor in the lungs. We found 29 promising candidates that could block the binding site and selected four of them that were likely to bind very strongly. We tested three of these candidates using frequency locked optical whispering evanescent resonator (FLOWER), a label-free sensing method based on microtoroid resonators. FLOWER has previously been used for sensing single macromolecules. Here we show, for the first time, that FLOWER can provide accurate binding affinities and sense the inhibition effect of small molecule drug candidates without labels, which can be prohibitive in drug discovery. One of the candidates, methotrexate, showed binding to the spike protein 1.8 million times greater than that to the receptor binding domain (RBD) binding to hACE2, making it difficult for the virus to enter cells. We tested methotrexate against different variants of the SARS-CoV-2 virus and found that it is effective against all four of the tested variants. People taking methotrexate for other conditions have also shown protection against the original SARS-CoV-2 virus. Normally, it is assumed that methotrexate inhibits the replication and release of the virus. However, our findings suggest that it may also block the virus from entering cells. These studies additionally demonstrate the possibility of extracting candidate ligands from large databases, followed by direct receptor–ligand binding experiments on the best candidates using microtoroid resonators, thus creating a workflow that enables the rapid discovery of new drug candidates for a variety of applications.

The coronavirus disease 2019 (COVID-19) pandemic caused by the severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) virus has created both a public health crisis and an economic crisis. This virus forced much of the world to adopt a lockdown mode, causing staggering economic fallout and human suffering.¹⁻³ Several studies have identified new drug candidates for COVID-19. High-throughput screening of ∼3000 drugs tested and validated 23 diverse antiviral drugs in human Huh7.5 cells.⁴ Virtual screening using 6218 approved and clinical trial drugs for COVID-19 were screened against the main protease and RNA-dependent RNA polymerase of SARS-CoV-2, resulting in 15 and 23 potential repurposed drugs, respectively. Among them, seven compounds can inhibit SARS-CoV-2 replication in Vero cells.⁵In silico studies (molecular docking and transcriptomic analyses) have identified a set of potential drug repurposing candidates targeting three viral proteins (3CL viral protease, nonstructural protein (NSP) 15 endoribonuclease, and NSP12 RNA-dependent RNA polymerase), which include rutin, dexamethasone, and vemurafenib.⁶ Most of these drug candidates act on processes that occur after the invasion by the virus into the cell. Our strategy is to prevent the initial insertion of the virus into the cell. We note that the only drug that continues to show efficacy in early COVID-19 is nirmatrelvir, the antiviral contained in Paxlovid.⁷ Remdesivir still shows modest efficacy.⁸ However, our therapeutic options have decreased, not increased, with the loss of monoclonal antibody therapy. Our work offers a rapid path toward filling in this therapeutic void, modeled on each new variant.

We aimed at identifying Food and Drug Administration (FDA) approved drugs⁹ that would bind to SARS-CoV-2 sufficiently strongly to block virus binding to human angiotensin-converting enzyme 2 (hACE2) in Figure 1a. To do this, we first defined a pharmacophore for virtual screening with 2.5 Å resolution based on the binding site of the receptor binding domain (RBD) to hACE2 from the first reported crystal structure of SARS-CoV-2 complexed with hACE2 (PDB ID: 6LZG) (Figure 1b).¹⁰ Another crystal structure with 2.45 Å resolution of the RBD of the spike protein of SARS-CoV-2 bound to the hACE2 (PDB: 6M0J) has also been reported.¹¹

Figure 1.

(a) Overall schematic of how drug candidates are found and validated. (b) X-ray structure of the SARS-CoV-2 receptor-binding domain (RBD) complex (green) bound to human angiotensin-converting enzyme 2 (hACE2) (yellow) (PDB ID: 6LZG). Blue circles indicate the three common hydrogen bonds shared between SARS-CoV and SARS-CoV-2. Additionally, the SARS-CoV-2-RBD complex used for virtual screening exhibits two unique interactions, represented by red circles: a salt bridge between K417 in RBD and D30 in hACE2, and a water-mediated hydrogen bond between E484 in RBD and K31 in hACE2. (c) Overview of experimental steps. As the spike receptor binding domain, or spike receptor binding domain-drug complex binds to receptors on the surface of the toroid, the resonance frequency of the toroid changes. (d) Human ACE2 receptors are bound to the surface of the silica microtoroid using EDC/NHS functionalization.

Based on this pharmacophore, we carried out in silico virtual screening using Phase virtual screening software from Schrödinger¹² to identify 46 candidates (29 unique molecules) out of 1657 in the FDA list. We followed this with DarwinDock¹³ complete sampling calculations on these 46 candidates to optimize binding site and free energy. This finally led to four ligands predicted to bind strongly to the RBD pharmacophore.

We then tested three of these four experimentally using the fast and accurate frequency locked optical whispering evanescent resonator (FLOWER) technique.^14,15 An overview of the experimental steps is shown in Figure 1c. Previously, FLOWER has sensed the individual macromolecules. Here, we show that FLOWER can provide accurate binding affinities and detect the inhibition effect of small molecule analytes without labels (such as fluorescent tags that would require additional cost and might affect accuracy).¹⁴⁻¹⁹ This led to two drugs, methotrexate and diethylenetriamine pentaacetate (DTPA), predicted to bind to the RBD over 1.8 million times stronger than RBD binds to hACE2. We then predicted the binding of all four candidates to the alpha, delta, and omicron variants and found that methotrexate alone binds much stronger to all three RBDs than RBD binds to hACE2. Indeed, it has been reported that methotrexate inhibits SARS-CoV-2 virus replication in vitro(20) and that methotrexate-treated patients can remain healthy despite long and close contact with SARS-CoV-2-infected individuals.²¹ Other in vitro and in vivo studies have also reported that methotrexate is able to inhibit COVID-19 via multiple mechanisms, such as suppressing SARS-CoV-2 entry and replication in all four SARS-CoV-2 variants of concern by targeting the furin enzyme and the host’s dihydrofolate reductase (DHFR), respectively.²²

These studies demonstrate the use of virtual screening combined with DarwinDock to extract candidate ligands from large databases that can be first validated experimentally with FLOWER, with the winners tested experimentally on full human cells. After identifying hits in this way, we can next do R-group screening to optimize new ligands much more selectively for blocking the invasion of SARS-CoV-2 and its mutants. Our studies demonstrate a novel strategy for preventing SARS-CoV-2 infection prior to all of the insidious effects set in motion once the virus is inside the cell.

Methods

We applied a strategy of virtual screening, followed by DarwinDock calculations, to identify potential ligands for blocking the binding of SARS-CoV-2 to hACE2. Three of the four predicted candidates were tested experimentally using FLOWER.

Virtual Screening for SARS-CoV-2 Binding to hACE2

We used the Phase in silico techniques in Maestro from Schrödinger on the 1657 drugs from the ZINC FDA-approved Drug Bank⁹ to predict 46 ligands that bind to the SARS-CoV-2 RBD. To do this, we generated a pharmacophore hypothesis based on the 2.5 Å X-ray structure of SARS-CoV-2 complexed with hACE2.¹⁰

The protein residue hypothesis included the following features: one hydrogen bond (H-bond) acceptor (A6) and donor (D12) at Q498, three negatively charged residues (N15, N16, N17) at Y449, Q493, and K417, and one positively charged residue (P18) at E484. We required matches with at least three out of these six pharmacophore points. The default rejection criteria (alignment score > 1.2, vector score < −1.0, volume score < 0.0, and included volume score < 1.0) and the Phase screen score as a scoring function were used to generate 50 minimized conformers.

This pharmacophore model served as the 3D query for virtual screening of 1657 drugs from the ZINC FDA-approved Drug Bank.⁹

We selected the best 46 hits based on energy and structural diversity. Then, we reranked these 46 using the much more accurate DarwinDock scoring. This resulted in identifying four structurally diverse hits that bind to wild type RBS. Next, we optimized the side chains of these hits using the side chain rotamer excitation analysis method (SCREAM)²³ to obtain improved binding configurations.

DarwinDock: Calculation of Ligand–Protein Binding Sites and Energies

The identification of optimal binding ligands goes beyond the capabilities of the virtual screening software, which primarily serves to identify candidate ligands. To obtain more accurate energetics, we employed the DarwinDock complete sampling method for predicting ligand binding sites. Considering conformational fluctuations, as proteins and ligands are dynamic structures, we employed ensemble docking, taking into account multiple ligand conformations (45 conformations for versetamide, 94 conformations for DTPA, 49 conformations for methotrexate (MTX), and 12 conformations for pentosan polysulfate (PPS)). Instead of docking ligands to a single protein structure, we utilized SCREAM to generate multiple conformations of the side chains and RBD, representing various conformational states.²³ Once the optimal binding pose was determined from docking, molecular dynamics (MD) simulations were employed to thoroughly explore the conformational space of the system. This helped identify the best binding pathways and potential conformational changes during ligand binding, accounting for the receptor’s flexibility to enhance the accuracy of binding predictions. This methodology, outlined below, enables the comprehensive exploration of ligand–protein interactions.

Initially, we prepared the likely binding region of the RBD by substituting six hydrophobic residues (I, L, V, F, Y, and W) with alanine (A). This alanization step created a sufficient space to accommodate ligand docking.

Subsequently, DOCK4.0 was employed to generate approximately 50,000 poses without energy calculations, effectively spanning the putative binding regions of the alanized protein. To ensure comprehensive sampling, we generated these poses incrementally in batches of 5000 and clustered them into Voronoi families based on root-mean-square deviation (RMSD) until less than 2% of new families were formed. To ensure the complete sampling of the binding site (completeness), our objective was to generate a comprehensive set of poses covering the entire binding region and then select the best ones based on their energy. We achieved this by sampling binding configurations, ensuring that they had no more than “MaxBump” bad contacts, resulting in the generation of approximately 2500 families with a 1.2 Å diversity. Here, MaxBump was defined as Max(2, 0.1N), with N representing the number of atoms in the ligand. We subsequently calculated the Dock score for each of these configurations and retained the top 10% (250) of them. We then generated additional configurations and retained those that overlapped with any one of these 250 families until each family had an average of six children. These 1500 structures were then divided into families with a 0.6 Å diversity, resulting in 750 families.

For each family, we evaluated the energy using the Dreiding force field (FF), which incorporates more accurate 3-point hydrogen bond interactions compared to the more standard CHARMM and Amber FF. We selected the best 150 families (the top 10%) for further consideration. Within each of these 150 families, we evaluated the Dreiding energy for each structure and retained the best one for each family. Subsequently, we employed SCREAM21 to optimize the protein back to its original residue, determining the optimal side chains for each of the 150 ligand poses.

SCREAM uses a library of residue conformations for each amino acid, allowing for a range of diversity from 0.2 to 5.0 Å (we used 1.0 Å, leading to 1478 rotamers), in conjunction with Monte Carlo sampling based on the sum over four energy terms: full valence, hydrogen bond, electrostatic, and van der Waals (vdW). A distinctive feature of SCREAM is the utilization of optimized flat-bottom vdW potentials that reduce the penalty for contacts that are slightly too short while retaining normal attractive interactions at full strength. Accurate hydrogen bond and vdW energies were obtained using Dreiding3.

We then selected the best 50 structures and performed 50 steps of minimization for the unified binding site. Subsequently, we chose a subset of approximately five structures from this group. For this subset, we further minimized the structure for the complete ligand–protein complex and selected one or two structures for subsequent analysis.

DarwinDock allows for the elimination of a percentage of the protein + ligand complexes at each step based on energy criteria, including analytical volume generalized Born (AVGB) continuum solvation and a Delphi-based solvation method. We scored these structures using several energy criteria, some with solvation and some without, some with ligand minimization and some without, some using just the energies within 5 Å, and some for all distances.

We find that combination of the winners using these various criteria is successful at recognizing the best cases, which are then validated with full MD. The specific definitions are as follows:

• Solvation E: the binding energy with solvation but without ligand minimization.

  BE = complex – [protein + (ligand_vac + ligand_solv)]

• Solvation Emin: the binding energy with ligand minimization and with solvation.

  BE = complex – [protein + (ligand_vac + ligand_solv) – (ligand_min + ligand_{min_solv})]

• Ligand strain E: ligand strain energy in the binding site, ligand_com – ligand_min

• Cavity E: the cavity analysis where only residues within the local 5.0 Å binding cavity contribute to the energy.

This step involved the independent optimization of side chains for each of the 100 poses, followed by additional full geometry minimization by 10 steps with a final energy gradient (force) of 0.05 kcal/mol/Å.

Finally, the docked structure exhibiting the most favorable binding energy was selected as the outcome. Notably, when applying this procedure to the antagonist JDTic ligand within the X-ray structure,²⁴ we achieved a structure deviating by only 0.23 Å RMSD from the reference X-ray structure. This robust approach has successfully predicted ligand binding sites for various targets, including CCR5,^11,12 GLP1R,²⁵ CB1,²⁶ AA3,²⁷ TAS2R38,²⁸ DP prostaglandin,²⁹ and TAS1R2/1R3 heterodimer.³⁰ Building upon these predictions, we identified four candidate ligands for subsequent experimental validation.

Prediction of Binding Affinity for SARS-CoV-2 Variants

Certain SARS-CoV-2 variants, such as those carrying the N501Y, K417T, or N/E484 K mutations, exhibit changes within the binding site of the RBD on the spike protein. In light of this, we aimed to predict the binding affinity of our drug candidates against four SARS-CoV-2 variants, including SARS-CoV, which served as a test reference. To achieve this, we first aligned the best pose from the SARS-CoV-2 complex with the corresponding mutants. Subsequently, the side chains within the binding site were optimized using SCREAM, followed by complex minimization. Next, we employed MD annealing to optimize the movable atoms within the binding site. This involved performing 10 cycles of annealing from 50 to 600 K over a duration of 0.1 ps, with a maximum of 1000 steps and a target root-mean-square (rms) value of 0.5 Å. After the MD annealing process, we rescored the binding energy and compared it to the binding energies obtained from the original SARS-CoV-2 complex. This comparison allowed us to assess the changes in binding energy induced by the mutations in the SARS-CoV-2 variants.

Full Solvent MD

To further examine the interactions between the hACE2-RBD complex and potential ligands, we conducted full solvent MD simulations. Although one glycosylation site at N343 out of 17 N-glycosylation sites was reported in the RBD,³¹ we did not include the glycan for simulation. Because N343 is not located in the binding site of either hACE2 or drug, N-glycosylation is important for proper protein folding and priming by host proteases. The hACE2-RBD complex was incorporated into a bilayer lipid membrane along with water at a physiological salt concentration. Using visual MD (VMD), we constructed a system with a total of 163,337 atoms, including 50,275 water molecules, 156 Na⁺ ions, and 131 Cl^– ions in a periodic cell of size (in the XYZ plane) of 100 Å × 120 Å × 150 Å. After the protein was inserted into the water cell, any water molecules overlapping within 5 Å were removed. Periodic boundary conditions were applied to simulate an infinite system. We assigned the protonation state of protein residues at physiological pH = 7.4. For all calculations, the temperature was maintained at 310 K using a velocity-rescale thermostat with a damping constant of 1.0 ps for temperature coupling, and the pressure was controlled at 1 bar using a Parrinello–Rahman barostat algorithm with a 5.0 ps damping constant for the pressure coupling. Semi-isotropic pressure coupling was used during this calculation. The Lennard-Jones cutoff radius was 12 Å, where the interaction was smoothly shifted to 0 from 10 to 12 Å. Periodic boundary conditions were applied in all three directions.

Similarly, the drug-RBD complex having a cell size in the XYZ plane of 60 Å × 80 Å × 90 Å was solvated in water at physiological salt concentration, resulting in a total of 37,461 atoms, including 11,509 water molecules, 32 Na⁺ ions, and 31 Cl^– ions. Waters within a 5 Å radius of the protein were excluded from the system.

To evaluate electrostatic interactions, the particle mesh Ewald (PME) method was employed. The overall charge of the system was balanced to zero by replacing water molecules with Na⁺ and Cl^– ions. Within the hACE2 protein, three disulfide bridges were constrained: C133–C141, C344–C361, and C530–C542. Additionally, four disulfide bridges within the RBD were constrained: C336–C361, C379–C432, C391–C525, and C480–C488. These constraints ensured the stability of the protein structure during the simulations.

After inserting the complex into the water and ion-containing box, the protein was initially fixed, while the other atoms were energy minimized for 1000 steps to optimize the positions of water and ion atoms. Subsequently, the system was equilibrated using NPT dynamics for 500 ps with 1 fs (fs) time steps, followed by 1000 steps of minimization for the full system while still keeping the protein fixed. This allowed the water and ions to readjust to the presence of the protein.

Following this equilibration of the solvent, a full system minimization was performed for 1000 steps, after which we carried out NPT dynamics at 310 K to equilibrate the full system at 1 atm pressure. MD simulations were conducted for a duration of 50 ns (ns), using NAnoscale MD (NAMD) 2.9 software.³² During the simulations, the system was maintained at a temperature of 310 K and a pressure of 1 atm.

Langevin dynamics were utilized for temperature control with the thermostat set at 310 K. The Nose–Hoover Langevin piston pressure control was employed to manage fluctuations in the barostat, which was set at a pressure of 1 bar. In this context, the periodic cell was constrained to remain orthorhombic but the cell parameters were allowed to vary. A dielectric constant of 1 was used for the electrostatic interactions, which were calculated using the PME method. The grid in the x, y, and z directions for the PME method was set at 55, 75, and 90 points for the lipid–protein complex, respectively. The van der Waals interactions were described using a Lennard-Jones function multiplied by a cubic spline switching function starting at 8 Å and stopping at 12 Å. The cutoff radius for including atoms in the nearest-neighbor list was 13.5 Å. All 1–2 and 1–3 interactions were excluded, and 1–4 interactions were scaled with a predefined factor. The bonded interactions were calculated at every time step, nonbonded interactions were calculated every other time step, and electrostatic interactions were calculated every fourth time step. The nearest-neighbor list was updated every 20 time steps. Every 10 ps, a snapshot was written to the trajectory file for subsequent analysis.

The CHARMM22 force field parameters were used for the protein, the TIP3 model was employed for water molecules,³³ and the CHARMM27 force field parameters³⁴ were utilized for the lipids. These force field choices ensured an accurate representation of the interactions within the system components.

Spike-RBD Affinity and Drug Inhibition Studies Using FLOWER

Immobilization of hACE2 on Microtoroid Optical Resonators

Microtoroids were prepared as previously described.¹⁸ The microtoroid chip was functionalized by treating it with 1% APTES silane in 1 mM acetic acid for 10 min. After rinsing with DI water and ethanol, the chip was dried using nitrogen and then succinylated overnight using 10% (w/v) succinic anhydride in dimethylformamide. Subsequently, the succinylated chip was rinsed with ethanol, dried with nitrogen, and activated using EDC/NHS (0.1/0.05 M) in MES buffer (50 mM, pH 4.5) for 10 min. Following activation, the chip was incubated with 400 nM hACE2 receptor (Invivogen) in HEPES 25 mM, NaCl 160 mM, CHAPS 8 mM, and pH 7.5 for 1 h. To minimize nonspecific binding, the chip was exposed to blocking buffer (ethanolamine 100 mM) for 5 min, followed by rinsing with blocking buffer and sensing buffer. The chip was then stored in the sensing buffer until further use. An overview of these steps is shown in Figure 1d.

Drug Preparation

Stock solutions of methotrexate, diethylene triamine pentaacetic acid (DTPA), and versetamide were prepared. Methotrexate (Sigma-Aldrich, CAS number: 59-05-2) was dissolved in 10 mM NaOH to obtain a stock concentration of 1 mM. DTPA (Sigma-Aldrich, CAS number: 67-43-6), the gadolinium-free component of Magnevist, was also prepared as a stock solution of 1 mM. Versetamide was obtained by removing gadolinium ions from gadoversetamide (Sigma-Aldrich, 1287675) using Dowex resin beads. The gadoversetamide-HCl solution was added to the resin beads, and after shaking for 5 min, the beads were centrifuged and the supernatant containing versetamide was collected. Liquid chromatography–mass spectrometry confirmed the successful removal of gadolinium ions. A stock solution of 197 mM purified versetamide was prepared by solubilizing 30 mg of versetamide in 300 μL of DI water. All of the stock solutions were stored at 4 °C until further use. LC-MS data showing the purification of versetamide from gadoversetamide are shown in Figure S1.

Measurement of SARS-CoV-2 Spike-RBD Affinity to hACE2

Real-time binding measurements were performed using the FLOWER system, as described previously.^14,15,17,19 Microtoroid optical resonators were functionalized with hACE2 antibodies, and the resonance frequency shift was tracked during the binding process. The FLOWER system consisted of a tunable laser (Newport, Velocity 6712), a photodetector (Newport, Nirvana 2007), and a digital laser locking module (Toptica, Digilock 110). A high-precision data acquisition card (National Instrument, PCI-4461) recorded the resonance wavelength shift data throughout the experiments.

A pressure-driven perfusion system (AutoMate Scientific, VALVELINK 8.2) connected to the microfluidic chamber facilitated the controlled introduction of the running buffer and samples. Custom LABVIEW software controlled the sequence of the introduction of the solution into the chamber.

For the experiments, two different running buffers were used: a sodium phosphate buffer (10 mM, EDTA 2 mM, AEBSF 1 mM, Tween 20 0.01%, pH 7) and a tricine buffer (10 mM, EDTA 2 mM, AEBSF 1 mM, Tween 20 0.01%, pH 7). The binding affinity of SARS-CoV-2 Spike-RBD proteins to hACE2 was measured using tricine buffer for all four different variants. The Spike-RBD proteins used in this test included Spike-RBD-His WT (Invivogen, product code his-sars2-rbd), Spike-RBD (N501Y)-His (Alpha) (Sino Biological, 40592-V08H82), Spike-RBD (L452R, T478 K)-His (Delta) (Sino Biological, 40592-V08H90), and Spike-RBD-His (B.1.1.529 Omicron) (Sino Biological, 40592-V08H121).

To initiate the measurement, a blank injection of the binding buffer was performed to establish the sensor baseline. Varying concentrations of the SARS-CoV-2 Spike-RBD protein were diluted in the running buffer and introduced into the fluidic chamber, starting from the lowest to the highest concentration. Between each concentration, rinsing with running buffer was performed. The real-time resonance wavelength shifts corresponding to each concentration were measured.

From the obtained data, the resonance wavelength shift at equilibrium (Δλ_eq) was determined by extracting the fitting parameters using eq 1.³⁵

1

Here, Δλ(t) represents the resonance wavelength shift at time t, Δλ_start is the wavelength shift at time t₀ (start time introducing the sample), Δλ_eq is the wavelength shift at equilibrium, k_obs is the observed rate constant, and t₀ is the starting time of introducing the sample. The association constant (k_on), dissociation constant (k_off), and ligand concentration ([L]) are related to k_obs.

The binding curve was constructed by plotting the extracted wavelength shift at equilibrium (Δλ_eq) as a function of the ligand concentration. A one-site-specific binding model was used to fit the data and determine the binding affinity (K_d). The maximum wavelength shift (B_max) represents the maximum shift when all of the receptors are occupied. The fitting equation used was³⁶

2

Inhibition of Spike-RBD Binding to hACE2 Receptors

In the inhibition study, three predicted drug candidates were examined in a tricine buffer: methotrexate (MTX), versetamide, and diethylenetriamine pentaacetic acid (DTPA). Varying concentrations of each drug were mixed with 10 nM Spike-RBD and incubated for 1 h before the experiment.

To establish a baseline (B), a solution of 10 μM drug without Spike-RBD was first introduced into the microfluidic chamber. Next, premixed solutions of the drug with 10 nM of Spike-RBD were injected into the chamber, starting from the highest drug concentration. The resonance wavelength shift at equilibrium (Δλ_eq) was determined from eq 1.

After the premixed solution with the lowest drug concentration was introduced, a drug-free sample (10 nM Spike-RBD) was injected for normalization. The normalized responses (Y) and log[L] (logarithm of drug concentration) were used to fit the Hill equation to the inhibition curve, as described in eq 3.³⁵

3

When multiple binding sites are present, a competitive binding model can be employed by summing individual binding equations. For example, in the case of two binding sites, the competitive binding equation (eq 4) is used to describe the normalized responses (Y), log[L], and the fraction of sites corresponding to IC50₁.

4

To determine the inhibition constant (K_i) using the Cheng–Prusoff equation³⁷ (eq 5), the fixed concentration of Spike-RBD (10 nM) and the experimentally obtained dissociation constant (K_d) are used. The IC50 value is extracted from the curve fit using eqs 3 and 4.

5

Results and Discussion

Binding Site Analysis

It is interesting to compare the binding sites of SARS-CoV and SARS-CoV-2 RBD to ACE2 to see what is special about SARS-CoV-2. The sequence alignment of SARS-CoV and SARS-CoV-2 RBD reveals a strong structural homology with a high sequence identity of approximately 73.9%. The overall structure of the SARS-CoV-2 RBD closely resembles that of the SARS-CoV RBD, with a RMSD of 1.2 Å for 174 aligned Cα atoms. However, the SARS-CoV-2 RBD features a twisted five-stranded antiparallel β-sheet, which is stabilized by three disulfide bridges: C336–C361, C379–C432, and C391–C525.

Comparing the binding site of SARS-CoV and SARS-CoV-2 to hACE2, we find several hydrogen bonds at the binding interface. These include the conserved interaction between Y449 in RBD β1′ and D38 and Q42 of hACE2, the interaction between Q493 in RBD β2′ (N in SARS-CoV) and E35 of hACE2, and the interaction between Q498 in RBD (Y in SARS-CoV) and Q42 in hACE2 (shown as a blue circle in Figure 1).

However, the most significant sequence variation occurs at the β1′/β2′ loop of the RBD due to insertions in SARS-CoV-2. The inserted E484 in the RBD β1′/β2′ loop interacts with K31 in hACE2 (shown as a red circle in Figure 1) through an intervening water molecule. Another unique residue in the SARS-CoV-2 RBD is K417, located in RBD α3, which forms a salt bridge interaction with D30 of hACE2 (also shown as a red circle in Figure 1). Notably, K417 in SARS-CoV-2 RBD corresponds to Val in SARS-CoV RBD, leading to a significantly higher binding affinity of SARS-CoV-2 RBD to hACE2 compared to SARS-CoV RBD (4.7 vs 31 nM).

Based on these distinct interactions in the SARS-CoV-2 RBD, as well as conserved polar contacts in SARS-CoV, we focused on generating a pharmacophore for virtual screening.

Virtual Screening of the ZINC FDA-Approved Drug Bank Database

In our study, we conducted virtual screening using a pharmacophore model based on the X-ray structure of the SARS-CoV-2 RBD-ACE2 complex (PDB: 6LZG), as depicted in Figure 1. This pharmacophore model included common hydrogen bonds shared between SARS-CoV and SARS-CoV-2, as well as unique interactions specific to SARS-CoV-2.

Two distinct pharmacophore hypotheses were generated for virtual screening analysis. The first hypothesis focused on protein residues, encompassing specific features, such as acceptor, donor, and charged residues. The second hypothesis targeted the receptor cavity, incorporating acceptor, donor, charged, hydrophobic, aromatic, and excluded volume features. The detailed pharmacophore models are presented in Figure S2.

To perform the virtual screening, we utilized the ZINC FDA-approved Drug Bank database, which consists of 1657 drugs approved by the FDA from clinical trials. Applying our pharmacophore models (Figure S2), we employed Phase virtual screening to identify potential hit molecules. A total of 46 hits were obtained, representing 29 unique molecules, accounting for some duplicates due to multiple ligand conformations.

To refine the scoring and prioritize the hit molecules, we employed an optimization strategy involving side chain adjustments within the binding site using SCREAM and subsequent structure relaxation through minimization. This process provided two scoring energies for each compound: Ucav, which considered a unified cavity interaction energy, and BE-Dock, representing the total binding energy of the ligand to the protein relative to the solvated free ligand plus receptor.

Based on the refined scoring energies, we reassessed all 46 compounds and ranked the top 30 based on their BE-Dock values (in kcal/mol), as summarized in Table S1. Among these top-ranked compounds, two promising candidates emerged from the virtual screening results: (1) pentosan polysulfate (Hit 13-Ucav 1 in Table S1) exhibited favorable interactions, forming hydrogen bonds with specific residues; (2) versetamide (Hit 36-Ucav 6 in Table S1) demonstrated notable stability through multiple hydrogen bond interactions. The poses of two hits from refined virtual screening are shown in Figure S3.

DarwinDock Calculations for the Top 46 Hits from Virtual Screening

Following the virtual screening analysis, we proceeded to perform molecular docking using DarwinDock for all 46 hit molecules. The docking results revealed significantly stronger binding energies and interactions with the spike protein RBD, as presented in Table S1. From this set, we identified four particularly promising candidates for further experimental validation, as depicted in Figure 2.

Figure 2.

Four most promising poses generated by DarwinDock for the 46 virtual screening hits targeting the receptor-binding domain (RBD) of SARS-CoV-2 were applied. These hits were extracted from the X-ray structure of the SARS-CoV-2 receptor-binding domain bound to human angiotensin-converting enzyme 2 (PDB: 6LZG). (A) Versetamide: the docked pose of Hit 36 (Ucav-VS 6 in Table S1) displays salt bridges at R403, R408, and K417, along with hydrogen bonds at Y453 and Y505. (B) Diethylenetriamine pentaacetic acid (DTPA): the docking pose of Hit 2 (Ucav-VS 11 in Table S1) reveals salt bridges at R408 and K417, as well as hydrogen bonds at Y449, Q493, and N501. (C) Methotrexate (MTX): the docking pose of Hit 28 (Ucav-VS 45 in Table S1) exhibits salt bridges at R403, R408, and K417, with hydrogen bonds at G496, N501, and Y505 within the RBD binding site. (D) Pentosan polysulfate (PPS): the docking pose of Hit 13 (Ucav-VS 1 in Table S1) demonstrates salt bridges at R403, R408, and K417, accompanied by hydrogen bonds at Q409, Y453, and Y505 within the RBD binding site.

Versetamide, Hits 32 and 36

Versetamide, which is a derivative of gadoversetamide (marketed as OptiMARK), proved to be a compelling candidate (Figure 2A). Hit 32 and Hit 36 represent different conformations of the same ligand. The docking simulations revealed notable features in the binding site of Hit 32, including salt bridges formed with R403 and K417, as well as a hydrogen bond with Q409. Similarly, Hit 36 displayed three salt bridges with R403, R408, and K417, along with two hydrogen bonds involving Y453 and Y505 within the RBD binding cavity. Hit 41 was excluded from further consideration due to its inferior Ucav value compared to Hit 36 (−123.41 vs −138.17 kcal/mol).

Diethylenetriamine Pentaacetic Acid (DTPA), Hit 2

Derived from gadopentetic acid (commonly known as Magnevist), DTPA is a chelating agent complexed with gadolinium (Figure 2B). Magnevist is utilized to enhance imaging in cranial and spinal magnetic resonance imaging (MRI). Upon removal of gadolinium, the docking pose of DTPA demonstrated three salt bridges (R403, R408, and K417) and three hydrogen bonds (Y449, Q493, and N501) within the binding site.

Methotrexate (MTX), Hit 28

Methotrexate, a folate derivative known for its anti-inflammatory properties and cell division inhibition, emerged as another compelling hit (Figure 2C). It has been employed in the treatment of arthritis-induced inflammation and neoplastic diseases such as breast cancer and non-Hodgkin’s lymphoma.³⁸ Docking simulations unveiled three salt bridges (R403, R408, and K417) and hydrogen bonds involving the backbone of G496, as well as the side chains of N501 and Y505 within the RBD binding site.

Pentosan Polysulfate (PPS), Hit 13

PPS, commercially known as Elmiron, possesses heparin-like properties with anticoagulant and fibrinolytic effects (Figure 2D). The docking analysis of Hit 13 (Ucav-VS 1 in Table S1) revealed salt bridges with R403, R408, and K417, along with three additional hydrogens bond involving Q409, Y453, and Y505.

To assess the binding affinity of versetamide to both SARS-CoV and SARS-CoV-2 proteins, we introduced six missing residues (DLCFSN) into the loop of SARS-CoV (PDB: 2AJF) and performed energy minimization on the protein structure (Figure S4). Table S2 demonstrates a binding affinity that is 33 kcal/mol weaker for versetamide (Hit 36) docked at SARS-CoV compared to the SARS-CoV-2 protein, as anticipated. The binding of versetamide 36 at SARS-CoV-2 involves an additional salt bridge at K417 and a hydrogen bond at R403, depicted in red in Figure S5. Notably, these corresponding amino acids are replaced by V404 and K390 in the SARS-CoV protein, respectively. The absence of a salt bridge and a hydrogen bond in the binding site of SARS-CoV explains the weakened interaction observed between versetamide 36 and the SARS-CoV protein.

Considering the global circulation of multiple SARS-CoV-2 variants, it is crucial to investigate their effect on binding affinity. Several prominent variants, including Alpha, Delta, and Omicron, have been identified according to the World Health Organization.³ Each variant exhibits distinct mutations in the spike protein’s binding cavity, with N501Y, K417N, K417T, and E484 K being the focus of our study.

• 1)Alpha (B.1.1.7): initially detected in the U.K., the Alpha variant (also known as 20B/501Y.V1 or B.1.1.7 lineage) stands out due to an unusually high number of mutations. This variant has spread to numerous countries worldwide, including the United States and Canada. Notably, the N501Y mutation occurs within the binding cavity and is accompanied by a total of 17 amino acid changes in the SARS-CoV-2 proteins, eight of which are found in the spike protein.
• 2)Delta (B.1.617.2): first detected in India in October 2020, the Delta variant features the L452R and T478 K mutations within the binding cavity.
• 3)Omicron (B.1.1.529): reported to the World Health Organization (WHO) from South Africa, the Omicron variant possesses a total of 50 mutations, with 32 of them occurring on the spike protein and 15 within the RBD. Notable mutations within the binding cavity include G339D, S371L, S373P, S375F, K417N, N440K, G446S, S477N, T478K, E484A, Q493R, G496S, Q498R, N501Y, and Y505H.

To explore the binding affinity of versetamide for these variants, we employed DarwinDock to predict their new binding sites and compared the resulting binding affinities to those of versetamide, as shown in Table S3. While the E484 K mutation exhibited improved binding energy, the K417T and K417N mutants displayed considerably worse Ucav binding energies. This discrepancy can be attributed to the loss of the salt bridge with versetamide as observed in the docking pose. It is worth noting that N501 and E484 do not directly interact with the ligand, resulting in a minimal impact on the binding energy.

MD Simulations

MD Simulations on the hACE2-RBD Complex

MD simulations were performed for hACE2-RBD complexes of four variants (wild type (WT), Alpha, Delta, and Omicron) over 50 ns. The WT variant displayed the best interaction energies, with stable salt bridges and hydrogen bonds between ACE2 and RBD. The Alpha variant showed a new hydrogen bond between hACE2 K353-NZ and RBD Y501-OH due to the N501Y mutation. The Omicron variant exhibited three new hydrogen bonds arising from mutations in the RBD (Q498R, GS446, and N501Y). However, other mutations (K417N, Q493R, and Y505H) resulted in the loss of hydrogen bonds compared with the WT.

The simulations were divided into four phases: (i) 0–10 ns with all backbone constraints, (ii) 10–20 ns with helix and helix backbone constraints, (iii) 20–30 ns with helix constraints, and (iv) 30–50 ns without any constraints. The root-mean-square deviations (RMSDs) were calculated and averaged over the trajectory (Figure S6). Additionally, the average interaction energies between ACE2 and RBD were computed (Figure S7).

Among the variants, the WT exhibited the most favorable interaction energies between ACE2 and RBD. The stable salt bridges or hydrogen bonds observed in WT, Alpha, Delta, and Omicron (with average distances less than 5 Å) were 8, 7, 7, and 5, respectively (indicated in red, Figure S8). In the Alpha variant, a novel hydrogen bond formed between hACE2 K353-NZ and RBD Y501-OH due to the N501Y mutation in RBD. Although N501Y is in a loop, it did not induce significant conformational changes. Notably, the RBD Y501-OH formed a hydrogen bond with hACE2 K353-NZ. In the Omicron variant, three new hydrogen bonds were observed: between hACE2 Y41-OH and RBD R498-CZ, between hACE2 Q42-CD and RBD S446-OG, and between hACE2 K353-NZ and RBD Y501-OH. These hydrogen bonds arose from the mutations in the Omicron RBD (Q498R, GS446, and N501Y). However, other mutations, namely, K417N, Q493R, and Y505H, resulted in the loss of certain hydrogen bonds compared to the WT.

MD Simulations of Drug-RBD Complexes

We initiated two independent MD simulations to investigate the behavior of MTX-RBD variants (WT, Alpha, Delta, Omicron) over a duration of 100 ns. The simulations were performed with different restraints: 0–10 ns with backbone and salt bridge constraints, 10–50 ns with helix, helix backbone, and salt bridge constraints, and 50–100 ns with helix and helix backbone constraints. The average RMSD values for the protein backbone of WT, Alpha, Delta, and Omicron bound with MTX from two independent MD runs were found to be 1.61 and 1.36, 2.05 and 1.37, 1.35 and 2.51, and 1.45 and 1.93 Å, respectively. Similarly, the average RMSD values for MTX ligand were 28.36 and 5.79, 27.88 and 22.56, 82.53 and 70.20, and 28.83 and 23.87 Å for WT, Alpha, Delta, and Omicron, respectively. Compared to standard protein RMSDs, the ligand RMSD is higher over the dynamics, displaying the dramatic change of the binding sites. Notably, the Delta RBD variant exhibited significantly higher RMSD values for MTX, indicating a greater degree of flexibility and the presence of multiple binding sites. This observation was further supported by the analysis of average total interaction energies, where MTX demonstrated −82.06 and −114.64, −97.22 and −84.31, −113.77 and −94.14, and −72.66 and −100.72 kcal/mol for WT, Alpha, Delta, and Omicron, respectively. Interestingly, Delta exhibited the most favorable interaction energy with MTX (−103.95 kcal/mol) in average from two MDs, while Omicron displayed the weakest interaction energy (−86.09 kcal/mol).

Subsequently, we conducted MD simulations for DTPA-RBD variants (WT, Alpha, Delta, Omicron) without any restraints over a duration of 100 ns. The RMSD analysis revealed average values of 1.50 and 2.02, 1.26 and 1.29, 1.44 and 1.45, and 1.41 and 1.31 Å for the protein backbone of WT, Alpha, Delta, and Omicron, respectively. For DTPA, the average RMSD values of the ligands were 28.36 and 63.75, 30.54 and 8.45, 37.31 and 42.28, and 42.41 and 34.67 Å for WT, Alpha, Delta, and Omicron, respectively. As MTX shows, the Delta RBD variant exhibited the highest RMSD for DTPA. The average total interaction energies between DTPA and RBD were determined as −114.81 and −121.89, −116.81 and −132.51, −120.22 and −121.15, and −166.05 and −165.27 kcal/mol for WT, Alpha, Delta, and Omicron, respectively. Remarkably, DTPA demonstrated the most favorable interaction energy with Omicron (−165.66 kcal/mol) in average from two MD simulations, whereas WT exhibited the weakest interaction energy (−118.35 kcal/mol).

For the versetamide-RBD variants (WT, Alpha, Delta, and Omicron) complex, MD simulations were performed for 100 ns. The average RMSD values for the protein backbone were found to be 1.80 and 1.51, 1.60 and 1.38, 1.39 and 1.59, and 1.55 and 1.32 Å for WT, Alpha, Delta, and Omicron, respectively. In contrast, the average RMSD values for the versetamide ligand were 51.04 and 74.60, 17.15 and 30.04, 63.50 and 65.14, and 66.87 and 47.52 Å for WT, Alpha, Delta, and Omicron, respectively. Notably, the WT, Delta, and Omicron RBD variants displayed high RMSD values for versetamide. Overall, versetamide exhibited significant fluctuations compared to the other drugs. The average total interaction energies between versetamide and RBD were −88.30 and −100.45, −105.41 and −58.58, −110.63 and −107.33, and −82.93 and −99.14 kcal/mol for WT, Alpha, Delta, and Omicron, respectively. The Delta variant showed the most favorable interaction energy with versetamide (−108.98 kcal/mol) in average from two MDs, whereas the Alpha RBD exhibited the weakest interaction energy (−81.99 kcal/mol).

In Table S4, we summarized the average interaction energy of total (TOT), electrostatic (ELECT), and van der Waals (VDW) interactions between the drugs and RBD from two independent 100 ns MD simulations. Interaction energy trajectory analyses are shown in Figures S9–S11. Among the drugs, DTPA displayed the most favorable interaction energy of −118.35 kcal/mol on average (Figure S12). MTX and versetamide show similar interaction energies, with averages of −98.35 and −94.37 kcal/mol. In binding experiments, MTX has the best binding affinity (K_i) of 82.6 pM while DTPA is the second best with a K_i of 420.9 nM. Versetamide has a much lower binding affinity of 11.3 nM. However, except for DTPA, the Omicron variants showed weakened interaction energies, which is consistent with the experiments (55.0 M for MTX and no inhibition for versetamide). On the other hand, the Delta variants exhibited increased interaction energies for all three drugs. It is important to note that the disparities between the calculated interaction energies and the experimental binding affinity may arise due to the presence of multiple binding sites for the drugs and the inherent flexibility of the isolated RBD, which is normally stabilized through its trimeric association within the spike protein.

To further analyze the methotrexate-RBD complex, we compared two conformations with the strongest binding energies to the initial conformations at 0 ns MD (see Figure S13). At 0 ns, in the WT RBD, MTX forms three salt-bridges at R403, R408, and K417, as well as one H-bond with Y505. After 58.6 ns, the binding site of MTX shifts toward the top, resulting in the formation of three H-bonds with Y453, Q493, and Y505. Subsequently, at 82.1 ns, the binding site of MTX moves toward the left, leading to the formation of alternative H-bonds with Y453 and Y473.

In the RBD variant, at 0 ns, MTX establishes two salt-bridges at R403 and R408, in addition to two hydrogen bonds at Q409 and N417. After 66.4 ns, the binding site of MTX shifts toward the top, forming a salt bridge with R493 and two H-bonds with the backbone of S496 and H505. Finally, at 87.2 ns, the binding site of MTX remains at the top, forming an alternative salt bridge with R493 and H-bonds with S494 and the backbone of R498. In both cases, compared to the WT RBD, methotrexate exhibited multiple binding sites on the RBD surface, likely due to the Q493R mutation (indicated by the arrow) in the Omicron RBD variant.

Experimental Validation Using FLOWER

To validate the candidate ligands for drug development, we conducted two sets of experiments: (1) measurement of Spike-RBD affinity to hACE2 receptors and (2) inhibition of Spike-RBD binding to hACE2. For these experiments, we utilized FLOWER, an ultrasensitive label-free optical biosensing technique based on whispering gallery mode optical resonator technology.^15,39 FLOWER was chosen due to its high sensitivity and the ability to perform experiments on drug candidates without the need for tagging.

Spike-RBD Affinity to hACE2 Receptors

We validated FLOWER’s performance by first measuring Spike-RBD affinity to hACE2 receptors and comparing these results to published reports. Figure 3a provides an overview of the Spike-RBD affinity for hACE2 experiment. Various concentrations of Spike-RBD protein were introduced into a microfluidic chamber, and buffer rinsing (PBS or tricine) was performed between each concentration. Figure 3b shows the resulting wavelength shift observed when introducing different concentrations of Spike-RBD-His wildtype (WT). The binding curve was constructed using eq 1 from the Methods section, and the equilibrium wavelength shift (Δλ_eq) was determined for each concentration (Figure S14a). The binding curve was then fitted to eq 2 to extract K_d and B_max values. The percentage binding (Figure 3c,d) was calculated by dividing the wavelength shift by B_max. A comparison of K_d values in PBS and tricine buffer is shown in Figure 3c for Spike-RBD-His WT and in Figure 3d for Spike-RBD (L452R and T478 K)-His (Delta). Table 1 summarizes our measured K_d values, indicating that Spike-RBD exhibits stronger binding to hACE2 in PBS compared to tricine. K_d values obtained in PBS (Table 1) are consistent with other literature reports which range from approximately 5–400 nM.^11,40−44 No other studies report K_d values in tricine.

Figure 3.

(a) Experimental setup for binding of Spike-RBD to hACE2. A microtoroid functionalized with hACE2 receptors is exposed to a series of Spike-RBD solutions at varying concentrations. (b) Sensorgram illustrating the binding of Spike-RBD-His WT to hACE2 in tricine. The blue shaded areas represent the duration of exposure to Spike-RBD solutions, while the white areas indicate running buffer rinsing. (c) Binding curve showing the interaction between Spike-RBD-His WT and hACE2 in PBS (dashed line) and tricine (solid line). (d) Binding curve depicting the interaction between Spike-RBD (L452R, T478 K)-His (Delta) and hACE2 in both PBS and tricine.

Table 1. Summary of Binding Constants to hACE2 for Different Spike-RBD Variants in Tricine and PBS Buffers^a.

buffer	SARS-CoV-2 spike-RBD type
	wild type	Alpha (N501Y)	Delta (L452R, T478 K)	B.1.1.529 Omicron
tricine: Kd (nM)	366.4 [308.1, 424.6]	146.2 [98.2, 194.2]	30.2 [15.2, 45.3]	31.7 [20.7, 42.8]
PBS: Kd (nM)	56.5 [26.4, 86.6]		5.0 [0, 10.8]

^aThe numbers in the square brackets represent the range of uncertainty.

Figure 4 presents the binding curves for Spike-RBD wildtype and its three variants to hACE2 in tricine. K_d values were extracted by fitting the curves using eq 2, as listed in Table 1. The lower K_d values for the Spike-RBD variants compared to wildtype indicate stronger binding, in quantitative agreement with previous studies.^40,43,45

Figure 4.

Binding curves of the Spike-RBD wild type and variants to hACE2 in tricine buffer. The lines represent the fitting results using eq 2 to model the binding interaction.

Drug Inhibition Experiments

We further characterized three drug candidates (methotrexate, versetamide, and DTPA) in tricine. PPS could not be tested due to the purchased samples being polymerized. The drug concentrations were mixed with 10 nM Spike-RBD and incubated for 1 h before measurements (Figure 5a). Similar to the Spike-RBD affinity measurement, each concentration was fitted using eq 1. The equilibrium wavelength shift (Δλ_eq) obtained from the fitting was used to construct the inhibition binding curves (Figure 5). Pure drug and Spike-RBD were added at the beginning and end of the experiment to determine T and B values in eq 3 for response normalization (Figure 5b-d). Repeated experiments for DTPA are shown in Figure S15. Additional sensorgrams and inhibition curves are shown in Figures S16 and S17. For the methotrexate experiment, one concentration mixed with 10 nM Spike-RBD (N501Y)-His (Alpha) was repeated three times before moving to the next concentration, as indicated by the hollow red circles in Figure 5. The deviation is also shown in Figure S18.

Figure 5.

(a) Drug inhibition of Spike-RBD at various drug concentrations with a fixed Spike-RBD concentration of 10 nM. (b–d) Inhibition curves for methotrexate, versetamide, and DTPA against Spike-RBD wild type and variants in tricine buffer, plotted as a function of drug concentration. The solid lines represent the fitting results obtained using eq 3. Dashed lines connect data points for the cases with no inhibition. The data from two separate runs of methotrexate against Spike-RBD (N501Y)-His (Alpha) are presented as solid and hollow circles in (b).

The inhibition binding curves, fitted with eq 3, are displayed as solid lines in Figure 5b–d. K_i values were calculated using the IC₅₀ obtained from eq 5. The characterized parameters (IC₅₀, K_i, and Hill slope) are summarized in Table 2. Methotrexate exhibits a significant inhibition against all four variants. Notably, increasing the DTPA concentration does not reduce Spike-RBD inhibition. The normalized responses of different versetamide concentrations remain close to 1.0 for Spike-RBD (N501Y)-His (Alpha), Spike-RBD (L452R, T478 K)-His (Delta), and Spike-RBD-His (B.1.1.529 Omicron), indicating the absence of inhibition by versetamide.

Table 2. Summary of Binding Characteristics of Three Drug Candidates against the Spike-RBD Wild Type and Three Variants Using the Hill Equation (Eq 3 from the Methods Section).

variants	binding parameters	MTX	DTPA	versetamide
wild type	IC50	84.9 pM [36.6, 193.2 pM]	432.4 pM [128.1, 28.9 nM]	11.6 nM [695.1, 193.2 nM]
	Ki	82.6 pM [35.6, 188.1 pM]	420.9 pM [124.7, 28.1 nM]	11.3 nM [676.6, 188.1 nM]
	Hill slope	0.47 [0.32, 0.62]	0.18 [0.08, 0.27]	0.30 [0.03, 0.57]
Alpha (N501Y)	IC50	4.2 nM [1.3, 13.4 nM]	12.3 nM [4.2 nM, 36.6 nM]	no inhibition
	Ki	3.9 nM [1.2 nM, 12.5 nM]	11.5 nM [3.9 nM, 34.3 nM]
	Hill slope	0.19 [0.15, 0.23]	0.11 [0.12, 0.18]
Delta (L452R, T478 K)	IC50	167.7 pM [67.0 pM, 419.3 pM]	no inhibition	no inhibition
	Ki	126.0 pM [50.3 pM, 315.0 pM]
	Hill slope	0.19 [0.16, 0.22]
B.1.1.529 Omicron	IC50	72.3 M [252.4 nM, 20.8 mM]	no inhibition	no inhibition
	Ki	55.0 M [191.9 nM, 15.8 mM]
	Hill slope	0.07 [0.04, 0.11]

Comparatively, the IC₅₀ values for these three drug candidates against Spike-RBD-His WT reveal that a lower concentration of methotrexate is required to achieve a 50% reduction in the level of binding of Spike-RBD-His WT to hACE2, in contrast to the other variants (Figure 5). Notably, the inhibition experiment for Spike-RBD (N501Y)-His (Alpha) using DTPA was repeated twice (Figure S15). The results presented in the main text correspond to the second experiment, which involved refined sample delivery protocols. A comparison between the two sets of experiments is shown in Figure S14.

It is important to acknowledge that the validity of the Hill equation is contingent upon the ligand (here, the SARS-coV-2 RBD) being fully folded and the binding site remaining unchanged. Table 2 demonstrates negative cooperativity, with Hill slopes less than 1.0, which could arise from several factors, including:

• 1.Possibility of two or more drug molecules binding to Spike-RBD.⁴⁶
• 2.Existence of multiple conformations or orientations in the binding domain.^46,47
• 3.Ligand binding to one site influencing its affinity for the remaining site(s).⁴⁸

The possibility of the RBD having multiple binding sites can arise because multivalent binding allows for increased affinity and stability of the interaction between the virus and receptor. To explore the potential presence of multiple binding sites, we applied eq 4 and refitted the curves, allowing for competitive binding at two sites. We fit the minimum number of binding sites for which we could achieve a good fit. The fitting results are presented in Table 3 and visualized in Figure 6. Only cases that exhibited an inhibitory effect are plotted in Figure 6.

Table 3. Summary of Competitive Binding Characteristics of Three Drug Candidates against Spike-RBD Wild Type and Variants Analyzed Using a Two-Site Model (Eq 4)^a.

variants		MTX	DTPA	versetamide
wild type	T1	0.76 [0.54, 0.98]	0.58 [0.30, 0.85]	0.53 [0.23, 0.84]
	IC501	405.0 pM [106.3 pM, 1.5 nM]	88.7 nM [7.4 nM, 1.1 μM]	721.1 nM [60.1 nM, 8.7 μM]
	IC502	0.2 pM [3.5 fM, 13.5 pM]	0.2 pM [5.5 fM, 5.9 pM]	134.6 pM [136.1 fM, 133.1 nM]
	Ki2	0.2 pM [3.4 fM, 13.1 pM]	0.2 pM [5.4 fM, 5.6 pM]	131.0 pM [132.5 fM, 129.6 nM]
Alpha (N501Y)	T1	0.64 [0.45, 0.83]	0.61 [0.37, 0.85]	no inhibition
	IC501	272.9 nM [66.1 nM, 1.1 μM]	1.6 μM [119.4 nM, 20.6 μM]
	IC502	3.0 pM [0.2 pM, 44.9 pM]	0.1 pM [1.6 fM, 6.8 pM]
	Ki2	2.8 pM [0.2 pM, 42.0 pM]	0.1 pM [1.5 fM, 6.4 pM]
Delta (L452R, T478 K)	T1	0.51 [0.34, 0.69]	no inhibition	no inhibition
	IC501	118.6 nM [17.4 nM, 809.1 nM]
	IC502	0.2 pM [17.4 fM, 1.3 pM]
	Ki2	0.2 pM [13.1 fM, 1.0 pM]
B.1.1.529 Omicron	T1	0.73 [0.62, 0.83]	no inhibition	no inhibition
	IC501	2.3 μM [731.1 nM, 7.0 μM]
	IC502	2.1 fM [0.1 fM, 65.2 fM]
	Ki2	1.6 fM [0.1 fM, 49.6 fM]

^aT1 represents the fraction of binding site 1 (for IC50₁). The numbers in the square brackets indicate the range of uncertainty.

Figure 6.

Inhibition of Spike-RBD binding with varying drug concentrations, while keeping the Spike-RBD concentration fixed at 10 nM. (a–c) Inhibition curves for methotrexate (MTX), versetamide, and diethylenetriamine pentaacetic acid (DTPA) against Spike-RBD wild type and variants in tricine buffer, plotted as a function of drug concentration. The solid lines represent the fitting results obtained using the competitive binding for a two-site equation (eq 4). Only the cases showing inhibition are included in the plot. The combined data from two experiments of methotrexate against Spike-RBD (N501Y)-His (Alpha) are represented as solid and hollow circles in (a).

To demonstrate the obstruction of Spike-RBD binding to hACE2 by drug molecules, we conducted an experiment (Figure S18) in which 10 μM of methotrexate was mixed with serial dilutions of Spike-RBD (L452R, T478 K)-His (Delta). The premixed solutions were incubated for an hour before measurement, and each concentration was introduced three times, with buffer rinsing between injections. Pure MTX was injected initially as a baseline. Figure S18c illustrates the binding curve for Spike-RBD (L452R, T478 K)-His (Delta) to hACE2 with and without 10 μM of methotrexate. The presence of methotrexate shifts the entire curve to the right, indicating that a higher concentration of Spike-RBD is required to achieve the same level of response.

Additionally, we found experimental evidence supporting the role of methotrexate in inhibiting SARS-CoV-2 virus infection. Schälter et al.²¹ provided compelling evidence that individuals taking methotrexate were protected against SARS-CoV-2 infection. Furthermore, treatment of mice with methotrexate resulted in a significant decrease in ACE2 expression in the lung, intestinal epithelium, and intestinal organoids.²¹ Given these findings, it is crucial to quantify the minimum effective methotrexate doses required to prevent infection. Since methotrexate is FDA-approved, setting up these trials should be a straightforward process.

Overall, only methotrexate demonstrated measurable inhibition against all four variants in our experiments, emphasizing the importance of further investigations to determine the optimal MTX dosage for infection prevention.

Conclusions

Our aim in this study was to identify FDA-approved drugs that could bind specifically to SARS-CoV-2 and its variants at the site responsible for binding (and subsequent insertion) into ACE2. To achieve this, we employed in silico virtual screening methods and analyzed the ZINC FDA-approved Drug Bank database, which contains 1657 drugs. Using Phase virtual screening software, we identified 46 hits, of which 29 were unique molecules.

To obtain more accurate binding sites and binding energies, we utilized DarwinDock to analyze these 46 hits. This analysis led us to identify four molecules that were predicted to selectively bind to the target binding site of the RBD for wild type. These molecules were versetamide, diethylenetriamine pentaacetic acid (DTPA) (Magnevist without gadolinium), methotrexate, and pentosan polysulfate (PPS) (Elmiron). However, we encountered difficulties in conducting binding experiments with Elmiron due to its high polymerization.

To experimentally measure the binding of the first three molecules (methotrexate, DTPA, and versetamide), we employed the highly sensitive FLOWER technique that enabled us to measure binding constants down to attomolar levels without modifying the ligands. Our results revealed that methotrexate and DTPA exhibit a binding affinity (inhibitory constant, K_i) of 0.2 pM to the WT RBD in tricine buffer. This represents a remarkable 1.8 million-fold increase in binding strength compared to that of hACE2 (K_d: 366.4 nM). Importantly, methotrexate demonstrated strong binding not only to the WT RBD but also to the Alpha, Delta, and Omicron variants.

Methotrexate, a widely used drug for treating inflammation associated with arthritis and controlling cell division in neoplastic diseases like breast cancer and non-Hodgkin’s lymphoma, emerged as an excellent candidate for blocking SARS-CoV-2 from invading cells. In vitro studies have shown that methotrexate inhibits SARS-CoV-2 virus replication, and individuals treated with methotrexate have remained healthy despite close contact with SARS-CoV-2-infected individuals.²¹ Supporting this, in vitro and in vivo studies also report that methotrexate is able to inhibit COVID-19 via multiple mechanisms, such as the suppression of SARS-CoV-2 entry and replication by targeting the host’s furin and DHFR.²²

Although methotrexate is FDA-approved and commonly used for rheumatoid arthritis therapy, it can potentially cause gastrointestinal, bone marrow, lung, and liver toxicity. Hence, long-term administration for infection prophylaxis may present challenges. However, given the critical need to reduce infections, it is worth considering that the first step of successful cell entry by the virus involves the production of NSP1, which blocks innate immune functions by binding to the human 40S subunit in ribosomal complexes, thereby shutting down normal cell functions. Consequently, methotrexate’s ability to block cell entry offers the possibility of blocking the virus prior to NSP1 production while dramatically reducing the dosage to minimize its usual side effects.

Considering that methotrexate delivery through a metered dose inhaler has been demonstrated,²⁰⁰ we suggest utilizing inhalers to deliver methotrexate directly to the lungs as a preventive measure against SARS-CoV-2 infection. This approach would enable minimizing the dosage and potential side effects. Previous studies have administered methotrexate in normal ways so that it is transported to the ACE2 cells in the lungs by normal processes.²⁰¹ Our studies suggest a novel delivery strategy using inhalers to concentrate the methotrexate in the cells, likely to lead to the initial invasion. Therefore, we recommend immediate trials to determine the minimum methotrexate dose required for inhalers to effectively prevent the SARS-CoV-2 invasion. Controlled tests should be carried out prior to authorizing doctors to prescribe methotrexate for combating COVID-19.

The successful validation of three hits from the FDA-approved database (out of 1637 molecules) using the FLOWER experiments validates the effectiveness of combining virtual screening with DarwinDock for identifying candidate ligands with strong binding to the RBD. This approach can be further applied to much larger databases, such as PubChem with 164 million compounds, to identify new ligands with increased binding to the target RBD binding site. Subsequently, these ligands can be tested using FLOWER to identify potential drug candidates for further validation against the real virus. Additionally, we suggest conducting R-group screening to optimize new ligands that are highly selective for blocking the invasion of SARS-CoV-2 and its mutants while minimizing binding to other cells that could lead to side effects.

In the future, virtual screening and post virtual screening methods can be automated. Docking can also be automated, but analysis of docking and selecting the best pose for each ligand needs user interpretation as well as intervention.

In conclusion, our study demonstrates the feasibility of using FLOWER in combination with in silico virtual screening methods to identify FDA-approved drugs with specific binding capabilities to the SARS-CoV-2 RBD. FLOWER confirmed the high binding affinity of methotrexate and diethylenetriamine pentaacetic acid (DTPA) to the RBD, making them promising candidates for blocking virus entry. The identification of methotrexate as a potential preventive measure through inhaler delivery provides an avenue for further research and immediate trials to determine the optimal dosage for SARS-CoV-2 invasion prevention. Overall, our findings highlight the potential of FLOWER in combination with computational strategies in drug discovery and suggest avenues for the future optimization and expansion of candidate ligands.

Supporting Information Available

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acsptsci.3c00197.

• Table S1: rankings from virtual screening (VS) and DarwinDock calculations; Table S2: unified cavity energy (Ucav, kcal/mol) and binding energy (BE, kcal/mol) of versetamide at SARS-CoV-2 and SARS-CoV proteins; Table S3: predicted binding constants for versetamide to single mutants of SARS-CoV-2 protein; Table S4: average interaction energy (kcal/mol) of total (TOT), electrostatic (ELECT), and van der Waals (VDW) between the drug and RBD during 100 ns MD simulations; Figure S1: LC-MS data showing the purification of versetamide from gadoversetamide; Figure S2: pharmacophore hypothesis for virtual screening; Figure S3: poses of two hits from refined virtual screening; Figure S4: X-ray structures of SARS-CoV-2 and SARS-CoV bound to human angiotensin-converting enzyme 2; Figure S5: best docking pose of versetamide at SARS-CoV and SARS-CoV-2; Figure S6: RMSD trajectory analysis of ACE2-RBD protein variants; Figure S7: interaction energy analysis along the trajectory of ACE2-RBD variants during MD; Figure S8: analysis of salt bridges and hydrogen bonds from the MD trajectory of ACE2-RBD variants during MD; Figure S9: interaction energy trajectory analysis of methotrexate-RBD variants during MD; Figure S10: interaction energy trajectory analysis of DTPA-RBD variants during MD; Figure S11: interaction energy trajectory analysis of versetamide-RBD variants; Figure S12: average total interaction energy trajectory analysis for versetamide, DTPA, and MTX at four RBD variant complexes from two separate 100 ns MD simulations; Figure S13: binding site change of methotrexate-RBD WT and Omicron; Figure S14: Spike-RBD-His (B.1.1.529 Omicron) affinity to hACE2 using tricine buffer; Figure S15: comparison of DTPA inhibition against Spike-RBD (N501Y)-His (Alpha) from two experiments; Figure S16: inhibition binding of Spike-RBD (L452R, T478 K)-His (Delta) and methotrexate (MTX); Figure S17: inhibition curves of drug candidates against Spike-RBD-His WT; and Figure S18: 10 μM of MTX mixed with varying Spike-RBD (L452R, T478 K) His (Delta) concentrations and incubated at room temperature for an hour (PDF)

Author Contributions

S.-K.K. and W.A.G. conceived the study. S.-K.K. performed the docking and virtual screening. S.S. performed the experiments. A.G. assisted with the optical sensing experiments. P.-D.N. performed early experiments and removed the gadolinium ions from gadoversetamide. Y.T. assisted with surface functionalization of the sensor in early experiments. J.S. supervised the experiments. S.S., J.S., and W.A.G. analyzed and interpreted the experiments. S.-K.K. and S.S. wrote the manuscript with inputs from others. Every author has read, edited, and approved the final manuscript. S.-K.K. and S.S. contributed equally to this submission.

The authors declare the following competing financial interest(s): J.S. owns a financial stake in Femtorays Technologies which develops label-free molecular sensors.