Biophysical Analysis of Potential Inhibitors of SARS-CoV-2 Cell Recognition and Their Effect on Viral Dynamics in Different Cell Types: A Computational Prediction from In Vitro Experimental Data

Abstract

Recent reports have suggested that the susceptibility of cells to SARS-CoV-2 infection can be influenced by various proteins that potentially act as receptors for the virus. To investigate this further, we conducted simulations of viral dynamics using different cellular systems (Vero E6, HeLa, HEK293, and CaLu3) in the presence and absence of drugs (anthelmintic, ARBs, anticoagulant, serine protease inhibitor, antimalarials, and NSAID) that have been shown to impact cellular recognition by the spike protein based on experimental data. Our simulations revealed that the susceptibility of the simulated cell systems to SARS-CoV-2 infection was similar across all tested systems. Notably, CaLu3 cells exhibited the highest susceptibility to SARS-CoV-2 infection, potentially due to the presence of receptors other than ACE2, which may account for a significant portion of the observed susceptibility. Throughout the study, all tested compounds showed thermodynamically favorable and stable binding to the spike protein. Among the tested compounds, the anticoagulant nafamostat demonstrated the most favorable characteristics in terms of thermodynamics, kinetics, theoretical antiviral activity, and potential safety (toxicity) in relation to SARS-CoV-2 spike protein-mediated infections in the tested cell lines. This study provides mathematical and bioinformatic models that can aid in the identification of optimal cell lines for compound evaluation and detection, particularly in studies focused on repurposed drugs and their mechanisms of action. It is important to note that these observations should be experimentally validated, and this research is expected to inspire future quantitative experiments.

1. Introduction

Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) is a single-stranded, positive-sense RNA virus that belongs to the betacoronavirus genus of the Coronaviridae family.¹ The spike protein of SARS-CoV-2 is highly conserved among human coronaviruses, including SARS-CoV-1 and MERS-CoV.¹ The spike protein is important for mediating infection by binding to host cell receptors through the receptor-binding domain (RBD).² The susceptibility of cells to the SARS-CoV-2 infection can be influenced by a variety of factors. These include the presence of proteins capable of binding to the virus, the level of expression of receptors on the host cell, and the ability of proteases to cleave the spike protein.²⁻⁵

For instance, the angiotensin-converting enzyme 2 (ACE2) receptor plays a vital role in the entry of SARS-CoV-2. ACE2 is expressed in various organs, but its expression levels differ across different human tissues. This variability in expression levels may contribute to varying levels of cellular susceptibility to SARS-CoV-2 infection.⁶ Post-translational regulation of ACE2 can occur through another potential receptor called Basigin2 (CD147). Another candidate receptor, Neuropilin 1 (NRP1), has been found to increase cell susceptibility to SARS-CoV-2 infection by enhancing the virus’s infectivity. NRP1 facilitates greater viral entry into cells, rather than simply increasing the binding of the virus to the cell membrane.^2,3 Although other potential receptors for SARS-CoV-2 have been identified, they have not been studied as extensively as ACE2. These include glucose-regulated protein 78 or heat shock protein A5 (HSPA5) angiotensin II receptor type 2 (AGTR2) and the receptor for advanced glycation end products (RAGE).³

After binding to the ACE2 receptor, the spike protein of SARS-CoV-2 must undergo proteolytic cleavage for a successful viral entry. There are several proteases that can activate SARS-CoV-2 entry, including transmembrane serine protease 2 (TMPRSS2), FURIN protease, and cysteine proteases such as L-type cathepsins (Cathepsin L).^3,4,7 The addition of a FURIN cleavage site to the spike protein of SARS-CoV-2 is one of the key differences between this virus and the SARS-CoV-1 virus. The FURIN cleavage site makes it easier for SARS-CoV-2 to enter cells, which is one reason why this virus is more infectious than SARS-CoV-1.⁸ Interestingly, sequential passaging of SARS-CoV-2 in Vero cell lines causes a loss of the FURIN cleavage site. This results in a viral strain that is less infectious, but it is also more fit and can replicate more efficiently in cells.⁹ There has been some research suggesting that cathepsin L inhibitors could be used as therapy for COVID-19. Cathepsin L inhibitors block the activity of cathepsin L, which could prevent the cleavage of the spike protein and the entry of SARS-CoV-2 into cells. However, more research is needed to confirm the effectiveness of this approach.¹⁰

It has been reported that the presence of multiple receptors can further enhance the susceptibility of cells to infection. The SARS-CoV-2 virus can utilize various receptors, such as the ACE2 receptor and the TMPRSS2 protease, to enter cells. When multiple receptors are present, blocking the virus from entering cells becomes more challenging.³ Consequently, inhibiting only one of these proteins may not be sufficient to prevent the entry of SARS-CoV-2 into cells.⁸ Moreover, the differential expression levels of these receptors can contribute to the observed variations in the cell line susceptibility to SARS-CoV-2. For instance, certain cell lines may exhibit higher ACE2 receptor expression levels than others, rendering them more vulnerable to SARS-CoV-2 infection.¹¹⁻¹³ Therefore, the choice of cell line becomes crucial when studying the infectivity of SARS-CoV-2, particularly in drug-related investigations.^14,15

Viral entry is the first step in the SARS-CoV-2 infectious cycle. Blocking viral entry into cells is a promising therapeutic target for COVID-19 treatment.^16,17 Many repurposed compounds have been investigated for their ability to prevent viral entry, including the serine proteases nafamostat¹⁸ and camostat,¹⁶ the nonsteroidal inflammatory drug ibuprofen,^19,20 the antimalarials chloroquine^17,21 and hydroxychloroquine,¹⁷ and the ACE inhibitors captopril and telmisartan.²² The ACE inhibitors captopril and telmisartan were found to decrease ACE2 protein expression in kidney-isolated membranes. This suggests that patients taking these drugs may not need to stop treatment for the COVID-19 protection. Although these compounds act as competitive inhibitors to other host receptors, they have also been implicated in interacting with the ACE2–SARS-CoV-2 spike protein complex. This interaction may be responsible for their antiviral activity.²³

To gain a deeper understanding of the viral particle’s biological cycle, researchers have proposed predictive mathematical models. These models not only describe the dynamics of various viral infections but also simulate the effectiveness of potential repurposed drugs against the virus.²⁴ In this study, we employed well-established and recommended mathematical models to investigate viral dynamics. By utilizing data obtained from in vitro assays involving different cell lines, we calculated parameters associated with the SARS-CoV-2 viral infection. This approach allowed us to model and predict the potential efficacy of different drugs, taking into account the specific cell type.^15,24−27 In our study, we focused on considering the expression values that are deemed more favorable for the most relevant proteins, namely, NRP1, CD147, FURIN, TMPRSS2, and ACE2.^2−4,6,7,28 Additionally, we also considered other less studied proteins, such as HSPA5, AGTR2, RAGE, and Cathepsin-L.^3,8,29 These evaluations were conducted using recommended cell lines for SARS-CoV-2 research, including Vero E6, HEK293T, and Calu-3.^{2,14,15,30−32} As a negative control, we selected the HeLa cell line, which is not recommended for SARS-CoV-2 research.³⁰

The objective of this research is to forecast the vulnerability of different cell lines employed in in vitro assays to SARS-CoV-2. This is accomplished through theoretical modeling of the expression patterns of a cluster of proteins known to serve as potential receptors for the virus.³ Additionally, we simulate the infection of these cells under both drug-treated and untreated conditions, aiming to inhibit the virus’s binding to its receptors.^16,17,22,23 Our primary aim of this work is to contribute to the advancement of suitable cellular models for in vitro replication assays of SARS-CoV-2.

2. Materials and Methods

2.1. Search for Cell Lines and Proteins in Databases

Based on a comprehensive analysis of multiple studies,^{2,14,15,31,32} we carefully selected the following cell lines for our SARS-CoV-2 research. To identify these cell lines, we utilized the “browser by cell line group” feature of the Cellosaurus server, developed by Amos Bairoch of the CALIPHO group at SIB - Swiss Bioinformatics Institute [https://web.expasy.org/cellosaurus/browse_by_group]. First, we chose the Vero E6 cell line, derived from African green monkey kidney epithelial cells, as it is specifically recommended by Cellosaurus for culturing SARS-CoV-2. In addition, we included the HEK293 cell line, derived from human embryonic kidney cells, and the CaLu3 cell line, derived from human epithelial cells found in lung adenocarcinoma. Cellosaurus designates these cell lines as valuable for investigating SARS-CoV-2. Furthermore, we incorporated the CaLu3 cell line due to its high expression levels of most of the receptors considered in this study, as indicated by the Human Protein Atlas database. This added perspective enhances the comprehensiveness of our analysis. Lastly, we included the HeLa cell line, derived from human cervix - cervical adenocarcinoma, as a control. Cellosaurus classifies this cell line as unsuitable for culturing SARS-CoV-2 [https://web.expasy.org/cellosaurus/sars-cov-2.html].³⁰

In order to identify the proteins that could potentially act as receptors for SARS-CoV-2, we conducted an extensive analysis of various studies.²⁻⁵ Additionally, we examined the presence and expression levels of a total of nine proteins associated with SARS-CoV-2 infection using data from the Human Protein Atlas database [https://www.proteinatlas.org/]. Specifically, we focused on five proteins that have been identified as particularly significant in this context: NRP1, CD147, FURIN, TMPRSS2, and ACE2.^2−5,28 Furthermore, we considered four less extensively studied proteins: HSPA5, AGTR2, RAGE, and Cathepsin-L.³ For cell lines where data were not available in the Human Protein Atlas server, we conducted a search in the NCBI/GenBank database to determine the taxonomic group to which the cell line belongs. Subsequently, we performed a protein–protein BLAST search between the potential receptors of the cell line of interest and the receptors expressed in the cell lines present in the Human Protein Atlas, focusing on genomic, proteomic, and functional relationships within the specific tissue/region/cell type.

2.2. Design of Hypothetical Cellular Expression Systems from Experimental Data

We have developed a collection of four hypothetical multiexpression systems to mimic the simultaneous involvement of multiple proteins in four different cell lines: Vero E6, HEK293, HeLa, and CaLu3. These systems incorporate a set of nine proteins, namely, NRP1, CD147, FURIN, TMPRSS2, ACE2, HSPA5, AGTR2, RAGE, and Cathepsin-L (see Table 1).²⁻⁴ Simulating the expression of all of these proteins simultaneously in each cell line is justified by the fact that it is a plausible occurrence. Single-cell studies have revealed significant variations in protein expression among cells, including instances where certain genes are not expressed individually.³³ Consequently, it is essential to consider heterogeneous coexpression patterns. By simulating these group expression scenarios (systems), we can explore how the susceptibility of cells may be influenced by the type, number, and level of expression of proteins that potentially act as receptors.³³

Table 1. Summary of the RNA Expression Levels of the Receptors Studied in Different Cell Lines Analyzed in the Atlas of Human Proteins^a.

protein(s)	Vero6	HEK293	HeLa	CaLu3
NRP1	5.8	1.7	3.9	33.8
CD147	72.9	88.7	72.4	85.6
FURIN	9.7	10.3	15.8	8.8
TMPRSS2	11.9	0.01	0.01	0.01
ACE2	1.0	0.1	0.01	0.01
HSPA5	35.2	14.3	30.1	24.0
AGTR2	0.01	0.01	0.01	0.01
AGER	1.7	4.9	0.1	0.2
Cathepsin-L	75.7	37.8	100	100

^aThe generated RNA sequencing results are reported as normalized Transcription Values Per Million (nTPM). Numerical values and a pie chart representation of the expression levels of angiotensin-converting enzyme 2 (ACE2), transmembrane serine 2 (TMPRSS2), basigin2 (CD147), FURIN protease, neuropilin 1 (NRP1), heat shock protein A5 (HSPA5), angiotensin II receptor type 2 (AGTR2), receptor for advanced glycation end products (RAGE) and Cathepsin-L.

To assess the susceptibility of cell lines to SARS-CoV-2 infection, we employed the expression levels of potential receptors as a baseline. Utilizing transcriptomics data sourced from The Human Protein Atlas, we classified genes based on their expression patterns in specific tissues, individual cell types, brain regions, blood cells, and cell lines. The RNA expression values were obtained by normalizing the transcription values per million (nTPM) for each gene in tissue culture cell lines. The nTPM value represents the highest expression level found in The Human Protein Atlas transcriptomics database. An nTPM value of ≥1 indicates expression in at least one tissue, region, or cell type. Conversely, an nTPM value of <1 suggests a lack of expression.²⁸

To mitigate potential errors stemming from database limitations and detection methods for expression levels, it has been proposed to assign minimum expression values.¹⁵ Following this recommendation, we considered an nTPM value of 0.0 to be equivalent to 0.01. This is because values of 0.0 or less may be attributed to data constraints rather than a lack of protein expression. For a comprehensive understanding of transcriptomic data normalization, classification, and data sources, we encourage referring to The Human Protein Atlas’ essays and annotations section.³⁰

2.3. Simulation of Viral Dynamics from Experimental Data

To depict our findings, we utilized an in vitro data set of SARS-CoV-2 virus infection in cell lines specifically recommended for culturing the virus. In order to assess the virus’s theoretical infectivity based on the susceptibility level mediated by protein expression associated with virus receptors, we held the viral titer constant, with a multiplicity of infection (MOI) set to 1. This approach aligns with previous studies.^14,34−40 The purpose of this MOI value is to ensure a consistent probability of infection and to consider the recommended viral load necessary for inducing cytopathological effects postinfection (pi).³⁵ In addition, we assumed an initial cell concentration of approximately 1 × 10⁵ cell/mL, which is in line with previous research.^2,14,15,38 The infection measurement was scheduled at 24 h postinfection (hpi), a time frame consistent with prior investigations.^{14,37,39−41} Notably, earlier reports suggest that several cell lines can reach their maximum infectious SARS-CoV-2 titer in plaque assays within 24 hpi.^37,42

To calculate the hypothetical concentration of target or susceptible cells based on the type and level of protein expression, the following steps were undertaken:

• 1.The same initial total number of cells/mL (approximately 1 × 10⁵) as previously described for each cell line was assumed.
• 2.Given the low expression levels of certain crucial receptors in specific cell lines, minimum expression values were assigned on a percentage scale. This was done to estimate the hypothetical number of cells and the percentage of expressed receptors, following a suggestion put forth.¹⁵

To determine the effective infection rate (β), a value of 2 was assumed, taking into account the reported basic reproduction number, ₀.¹²

In this sense, each nTPM value was assumed to represent the percentage value of susceptible cells. To do this, we considered the mean of the nTPM values reported for the potential receptors and for each cell line. To address the hypothesis regarding cell susceptibility based on receptor expression levels, a constrained model of target cells was employed. This model relies on available data and aims to estimate infection rates among susceptible cells. Specifically, we utilized a well-established viral infection model known as the limited target cell model.^15,26 This model encompasses three compartments: susceptible cells (U), infected cells (I), and viral titers (V). The applied model is represented by the following set of differential equations:

1

2

The equations’ left-hand side represents the rate of change of the variables over time. The parameters β and δ denote the effective infection rates and the number of deceased infected cells, respectively. In this model, the virus (V) is assumed to infect susceptible cells (U) at a rate of β, while infected cells are eliminated at a rate of δ. Using this framework, we calculated the basic reproduction number, R₀. This value represents the average number of cells that a single infected cell can infect at the beginning of an infection:

3

Virions are actively released from infected cells at a productive rate of p per day. These virions are either cleared from the circulation at a rate of c or lost upon infecting a target cell. For the virus clearance rate, c, we adopted a value of 10 days^–1, as recommended elsewhere.¹⁵ The rate p, which is approximately equal to 22.7 copies days^–1 cell^–1, is suggested to be dependent on the susceptible cell population (U). Therefore, to estimate the actual rate p, it is necessary to calculate the product of p and U, as proposed elsewhere.¹⁵

2.4. Drugs with Potential Antiviral Activity Aimed at Blocking the SARS-CoV-2 Receptor-Binding Domain (RBD) Interactions with Host Receptors

In this study, we considered the crystal structure of the RBD complex (PDB: 6M0J) that facilitates the interaction between SARS-CoV-2 and cellular receptors, such as ACE2. The structure was obtained from the RCSB protein database, as suggested.¹⁹ To gather the pharmacokinetic/pharmacodynamic properties of drugs with reported potential antiviral activity, specifically targeting the SARS-CoV-2 spike protein or viral entry through host receptors, we relied on the literature sources. These sources provided kinetic models and parameter values for drugs like camostat,¹⁶ hydroxychloroquine,¹⁷ nafamostat,¹⁸ ibuprofen,^19,20 chloroquine,^17,21 captopril,^22,23 and telmisartan.²² Additionally, we searched for reported in vitro EC₅₀ values, which can be found in Table 2. Likewise, pyronaridine was included as a positive control for docking, as it is a ligand tested with AutoDock-based algorithms under similar blind docking conditions to those considered in this study and targeted to the RBD-ACE2 complex. Furthermore, pyronaridine also exhibits thermodynamically favorable binding, as confirmed experimentally by enzyme-linked immunosorbent assay and biolayer interferometry.⁴³

Table 2. Estimates of the Number of Susceptible Cells (U), Dead Infected Cells (δ), Productive Release Rate of Virions from Cells (p), and Basic Reproduction Number (R₀).

system	U(cell/mL)	p(copies-cell–1day–1)	δ (cell/mL)	R0
Vero E6	2.4 × 104	5.4 × 105	1.3 × 104	20.5
HEK293	1.8 × 104	4.0 × 105	9.8 × 103	20.3
HeLa	2.5 × 104	5.6 × 105	1.4 × 104	20.6
CaLu3	2.8 × 104	6.4 × 105	1.5 × 104	20.7

2.5. Prediction of the Affinity and Theoretical Stability of Compounds Targeting the SARS-CoV-2 Receptor-Binding Domain (RBD)

Due to the limitations described associated with molecular dockings such as variations in the performance of each method and the inherent limitations of crystal structures to accurately reflect the dynamic nature of protein structures in their biological state, a comparative analysis was performed using various popular molecular docking models to ensure consistent quality for the majority of the targets as suggested.⁴⁴ The complexes were built using the DINC server (https://dinc.kavrakilab.org/), which offers a version for the analysis of structures associated with COVID-19 with the docking algorithms AutoDock Vina (Vina), AutoDock Vina RaDii Optimized (Vinardo), and AutoDock 4 (AD4). Additionally, the DockThor server (https://www.dockthor.lncc.br/) was also used, which is a sophisticated machine learning algorithm that utilizes the refined PDBbind data set for training, similar to Vina. These programs were selected due to their advanced and enhanced features, making them improved versions of molecular docking tools. They are based on the efficient AutoDock algorithm, which has been previously recommended.⁴³ To ensure accuracy, a minimum of 10 runs per program were performed, resulting in approximately 1 × 10⁶ evaluations per run. The default parameters were used for the rest of the settings. As part of the preparation, all water molecules were removed, and the PDB files were separated into two distinct files: one containing the protein and the other containing the ligand structure. During the sampling process, only the three runs with the most favorable binding poses were considered. This approach allowed us to focus on the most feasible and thermodynamically favorable positions in the complexes. Based on this criterion, the selected complexes underwent further analysis, including potential theoretical inhibition and molecular dynamics. The binding constant K, derived from the binding free energy, was calculated using the equation described:^19,20

4

And the inhibition constant for binding of ligand to proteins (K_i) (in units of M) was obtained as

5

In the provided equation, ΔG represents the binding affinity in kcal mol^–1, R denotes the universal gas constant (1.987 cal K^–1 mol^–1), and T represents the absolute temperature (298.15 K). According to the equation, a higher K_i value indicates weaker binding between the inhibitor and the protein. Consequently, the protein–inhibitor complex is more likely to dissociate.⁴⁵⁻⁴⁷ To determine the IC₅₀ values, the Dixon plot was employed, assuming competitive inhibition, as recommended.⁴⁷ This plot offers a reliable estimation of IC₅₀ values for competitive inhibition of a specific substrate.⁴⁸ In this study, we considered a hypothetical substrate-inhibitor relationship (1:1) to avoid concentration-driven preferential associations, assuming structural similarities between the ligand and the substrate. Additionally, the IC₅₀-to-K_i web tool was utilized to predict theoretical inhibition potential following the considerations for competitive inhibition with the following equation:

6

The variables in the equation include P₅₀, which represents the protein concentration at 50% inhibition; K_I, denoting the affinity constant of the ligand to the protein; [PL₅₀], indicating the protein concentration divided by the ligand concentration at 50% inhibition; and [L₅₀], representing the ligand concentration at 50% inhibition. To predict the concentration of the inhibitor required to achieve 50% inhibition (IC₅₀), we utilized the IC₅₀-to-K_i web tool. This tool utilizes the provided values of [P], [L], and K_i (all units in μM) to generate the IC₅₀ value. All calculations were performed using the IC₅₀-to-K_i web tool, accessible at https://bioinfo-abcc.ncifcrf.gov/IC50_Ki_Converter/index.php.^19,46 On the other hand, the experimental inhibitory concentration values were obtained in molar expression, as described later for each of the compounds considered.^43,49−60

Simulations were conducted on docking hits for two main purposes:

• 1.To examine the relative stability of the ligand within the binding pocket.
• 2.To sample the minimum energy conformations and assess the perturbation of thermodynamic and structural stability in the complexes.

For each protein–ligand complex, a series of minimization procedures were performed to relax the MD system. This process involved three phases: relaxation, equilibrium, and sampling, as recommended.^19,61,62 The MD simulation was carried out in an explicit water system. Specifically, the system was solvated in a cubic water box with dimensions of 80 Å for each axis (X, Y, Z). The system consisted of 12424 water molecules, 58 Na⁺ ions, and 34 Cl^– ions, resulting in a molar density of Na/Cl = 2.76 × 10^–3. Each MD system included one copy of the protein system and one copy of the docking ligand. The topology files for the ligands and proteins were automatically generated by using myPresto. The Amber99SB-ILDN force field and the TIP3P water model were employed. The entire system was neutralized, and water molecules were treated as rigid bodies in all models. The simulation time interval was set to 2 fs, and periodic boundary conditions were applied. Temperature and pressure control were achieved using the Berendsen algorithm. After an initial energy minimization phase, which involved steepest descent (5,000 steps) and conjugate gradient (5,000 steps) methods with positional restraints on the solute, a 100 ps simulation was conducted. During this simulation, the positions of the solute atoms were restrained by a force constant of 10 kcal mol^–1 Å^–2. This allowed the water molecules to diffuse around the molecule and to equilibrate the system. The particle mesh Ewald (PME) method was used to calculate the electrostatic contribution to nonbonded interactions with a cutoff distance of 14.0 Å and a time step of 1 fs. The van der Waals interaction cutoff distance was also set to 14.0 Å. Following the equilibration run, a production run was performed at a constant temperature of 300 K using the NVT (particle numbers, volume, and temperature) ensemble. The cell size remained the same, and the SHAKE algorithm was applied to the system. The time step for this run was set to 2 fs. Ten structures were extracted from a 100 ns trajectory, with each structure obtained every 10 ns. These structures served as target structures for further analysis. The root mean square deviation (RMSD) calculations were performed using the following equation:

7

where δ_i is the distance between atom i and either a reference structure or the mean position of the n equivalent atoms. The neutralization of this type of system as well as the final production run of 100 ns with a subsequent molecular mechanics/Poisson–Boltzmann surface area (MM/PBSA) calculation has already been widely proposed by other specialists for the study of ligand-protein systems associated with COVID-19. In addition, diverse MD simulation analyses were incorporated including RMSD, RMSF, and R_g, as well as comparisons with positive controls as suggested.^43,64 For the calculation of the root-mean-square fluctuation (RMSF) the algorithm of Calculations and Analysis of Normal Modes WEBnm@ (http://apps.cbu.uib.no/webnma3/) was used.⁶³ All MD simulations and additional adjustments were carried out using COSGENE/myPresto.^19,61,62 As part of the simulations, binding affinities were predicted based on the force constant relative to the initial frame for each complex during the 100 ns MD simulation. The WebPSN algorithm (accessible at http://webpsn.hpc.unimo.it/wpsn.php) was employed to calculate the binding force using two alternative versions of Elastic Network Models (ENM). These models were utilized to evaluate the cross-correlation of the motion of Cα atoms and the pairwise interactions between Cα atoms. The two versions of ENM used were linear cutoff-ENM and Kovacs-ENM.⁶⁵

To reevaluate the complexes and determine the free energy binding (ΔG_bind) of the MD trajectories, we employed the widely used MM/PBSA method⁶⁶⁻⁶⁹ as a thermodynamic integration approach. The final snapshots of the minimized system were subjected to MM/PBSA rescoring, which was implemented in AMBER16. As suggested,^43,64 2500 frames extracted from the last 10 ns of molecular dynamics were used to compute the solvation free energy and molecular mechanics potential energy for estimation of total binding affinity of the complexes of interest. The formulas for the MM/PBSA binding free energy estimation are shown in eqs 8 and 13:

8

9

10

11

12

13

The total binding free energy (ΔG_bind) represents the difference in free energy between the bound state (G_complex) and the free state (G_receptor + G_ligand). It can also be expressed as the sum of the enthalpy (ΔH) and entropy part (−TΔS). In this study, the enthalpy changes were computed using the MM/PBSA approaches, while the entropy changes were neglected due to their computational cost and lower accuracy. The enthalpy part can be further divided into the molecular mechanical energy (ΔE_MM) and the solvation free energy (ΔG_solv). The ΔE_MM term includes the intramolecular energy (ΔE_bonded), the electrostatic energy (ΔE_ele), and the van der Waals energy (ΔE_vdW). The ΔG_solv term encompasses both the polar contribution (ΔG_polar) and the nonpolar contribution (ΔG_nonpolar). The polar contributions are accounted for using the Poisson–Boltzmann (PB) model, while the nonpolar contributions are assumed to be proportional to the solvent-accessible surface area (SASA).⁷⁰ To calculate the energetic components and predict electrostatic solvation and free binding energies, we utilized the size-modified Poisson–Boltzmann equation (SMPBE) available at https://web.uwm.edu/smpbs/ and the APBS program accessible at https://server.poissonboltzmann.org/.⁶⁶ The ΔG_bind of the complexes was determined based on frames extracted after a 100 ns cycle. For visualization and analysis, we employed Molegro Molecular Viewer, version 7.0 (MMV_7.0). Detailed parameters for the MM/PBSA calculation procedures can be found in the Supporting Information Table S1.

2.6. Prediction of Theoretical Antiviral Effectiveness of Ligand from Experimental Data

As mentioned in previous studies,^15,72,73 we assume that compounds exhibiting antiviral properties with a constant effectiveness ε can reduce R₀ by a factor of (1 – ε). In eq 14, ε can take values ranging from 50% to 99%:

14

In the equation, C represents the concentration of the ligand in molar units, ε denotes the effect caused by the ligand at concentration C, and K_d stands for the equilibrium dissociation constant. In classical dose–response curves, K_d is referred to as EC₅₀ (dose for 50% effect), while in inhibition screens, it is known as IC₅₀ (50% inhibition). The Hill coefficient, denoted as n, is a measure of the ligand’s response or efficacy ε, which can range from 0 (no drug effect) to 1 (maximum effect). The Hill coefficient can be greater or less than 1, with values of n < 1, indicating negative cooperativity, and n > 1, indicating positive cooperativity. Typically, small molecules that bind to a target with a single drug-binding site exhibit an uncooperative dose–response curve with a slope parameter of n = 1, representing the fraction of bound ligand. For the purposes of this study, we assumed this value.^15,71 The Hill equation describes the relationship between the fraction of bound ligands and the ligand concentration. The dimension of K_d (IC₅₀ or EC₅₀) in this inhibition approximation of the Hill equation is M (molar), as suggested.^72,73 Additionally, due to limitations in determining the concentrations of active pharmaceutical ingredients (APIs) within the cellular medium of the compounds considered in this study, as they are not approved drugs for the inhibition of SARS-CoV-2, and given the incompatibility in the Hill equations to establish the relationship between the administered drug doses and the API concentrations at the cytoplasmic level, it was theoretically assumed that the reference values for ligand concentration corresponded to the experimental molar inhibitory concentration values reported for each of the compounds,^43,49−60 in order to illustrate the potential application of the Hill mathematical model. In this regard, the predictions of ε made in this study are only demonstrative, and further studies considering precise API values are recommended. On the other hand, it has been described that antiviral drugs that reduce viral infectivity β by a factor (1 – ε_β) reduce the basic reproductive number, R₀, by a factor:

15

If (1 – f(ε_β))R₀ is less than 1, the virus will almost surely go extinct.⁷⁴ The relationship (1 – f(ε_β))R₀ to describe the theoretical antiviral efficacy after treatment was designated in this work with the letter “ϕ” for illustrative purposes.

2.7. Prediction of Theoretical Drug Toxicity

Designed to predict potential toxicity, eMolTox is a web server (http://xundrug.cn/moltox) that can assess the likely toxic properties of a given molecule. By leveraging advanced machine learning methodologies, eMolTox not only forecasts but also presents comprehensive information about the potential toxic characteristics of a molecule. It also provides comparative data on known toxic compounds, thereby aiding safety analysis in the process of drug development.⁷⁵ Furthermore, the obtained data was cross-referenced with ProTox-II, a server available at https://tox-new.charite.de/protox_II/index.php?site=home. ProTox-II utilizes molecular similarity, fragment propensities, and machine learning techniques to predict a range of toxicity end points. These end points include acute toxicity, hepatotoxicity, cytotoxicity, carcinogenicity, mutagenicity, immunotoxicity, adverse outcome pathways, and toxicity targets.⁷⁶ It is important to note that these comparative analyses were conducted despite the fact that most of the compounds considered in this study are FDA-approved, indicating their safety for humans (https://pubchem.ncbi.nlm.nih.gov/). Therefore, they far exceed the preclinical judgment provided by computational tools. The purpose of this is to demonstrate their potential toxicity at the cellular level due to limitations in determining the concentrations of the Active Pharmaceutical Ingredients (APIs) within the cellular medium of the compounds considered in this study. This is particularly because they are not FDA-approved drugs for the inhibition of SARS-CoV-2, and it was theoretically assumed that the experimental inhibitory concentration values serve as the theoretical concentrations of the APIs at the cytoplasmic level.

3. Results and Discussion

3.1. Theoretical Susceptibility to SARS-CoV-2 Infection Depending on the Type and Expression Level of Potential Receptors in Cells of Interest from Experimental Data

A statistically significant difference was observed between the consensus normalized expression levels (nTPM) of proteins associated with SARS-CoV-2 infection in the cell lines studied (F_(8, 27) = 29.524, p < 0.00001, α = 0.01). No correlation was found between the nTPM levels of any of the proteins studied, regardless of the cell line type. The only exception was a moderate negative correlation between the CD147 protein, and all potential receptors tested (r ≈ −0.50). A similar correlation was observed between NRP1 and AGTR2, Cathepsin-L, and HSPA5 (r ≈ −0.50 and r ≈ −0.60, respectively). As expected, a perfect and positive correlation was observed between the ACE2 receptor and the TMPRSS2 protease (r = 0.99). On the other hand, very high and negative correlations were observed between the AGTR2 protein and Cathepsin-L (r = −0.99) and between CD147 and HSPA5 (r = −0.92). The results of the principal components analysis (PCA) showed that the ACE2 and TMPRSS2 receptors, together with the possible CD147 and HSPA5 receptors, are responsible for most of the observed variability (Cp1 ≈46%). The positive correlation between the ACE2 receptor and the TMPRSS2 protease as well as its relevance as they are expressed simultaneously has been previously reported.⁷⁷

Under the conditions of this study, all multiple theoretical expression systems were predicted to have a similar susceptibility (U) of up to ≈30% cells/mL, with a mean of ≈2.4 × 10⁴ and range of 1.8 × 10⁴–2.8 × 10⁴ potentially susceptible cells/mL. Specifically, the CaLu3 system was predicted to have the highest theoretical susceptibility, mediating a susceptibility of ≈2.8 × 10⁴ cells/mL, followed by the Vero E6 and HeLa systems, with a theoretical susceptibility of ≈2.4 × 10⁴ cells/mL. The HEK293 system was the least susceptible to infection under the conditions of this study (≈1.8 × 10⁴ cells/mL). Although the susceptibility values are very close, the difference in terms of percentage susceptibility between the CaLu3 system and the rest of the systems was between 11 and 37% (see Table 2). These results are consistent with what has been reported because, although all the cell lines tested in this study have shown susceptibility to SARS-CoV-2, it has been reported that SARS-CoV-2 pseudovirions can significantly increase in CaLu3 cells as well as in Vero E6 cells (the cell line most used to replicate and isolate SARS-CoV-2). Meanwhile, HEK293 cells have been characterized as showing modest viral replication.⁷⁸

According to the primary variables considered in this study, the difference in susceptibility is due to the simulated contribution of the possible receptors in each cell line. In the CaLu3 system, CD147 and Cathepsin-L receptors could contribute more than 90% of the observed susceptibility. HeLa and Vero E6 followed, with contributions of more than 80% and 70% of the observed susceptibility by the same potential receptors, respectively. It is important to note that the receptors with the highest correlation in terms of nTPM (ACE2 and TMPRSS2) had a significant contribution in the Vero E6 model (≈6%), while in the rest of the cellular models, their average contribution was low (≈0.05%). These results allow us to infer that the higher propensity to infection in CaLu3-type systems could also be associated at a statistical level with the contribution of other potential receptors other than ACE2 and TMPRSS2, as has been reported.²

Relative to the number of virions that could theoretically be released from productively infected susceptible cells, the systems tested were predicted to have a mean (p) release rate of 5.3 × 10⁵ copies/cell/day. Specifically, the CaLu3 system was predicted to have the most favorable release rate (p) with a value of 6.4 × 10⁵ copies/cell/day. The least favorable release rate (p) was for HEK293 with 4.0 × 10⁵ copies/cell/day. The theoretical number of dead infected cells (δ) was also predicted, and it was calculated that the tested systems could generate dead cells with a mean of 1.3 × 10⁴ cells/mL. The minimum and maximum values were 9.8 × 10³ and 1.5 × 10⁴ cells/mL, respectively, in the HEK293 and CaLu3 cell lines (see Table 2). This behavior in terms of the virions that could be released from infected cells has already been reported.⁷⁸

In order to validate the susceptibility of the cell lines used in this study, the basic reproduction number (R₀) was calculated. R₀ represents the average number of susceptible cells that become infected from a single infected cell at the beginning of the infection. A mean R₀ ≈ 20 was predicted in all systems. In all cell lines, a R₀ ≥ 20 was predicted, with a minimum of R₀ = 20.3 and a maximum of R₀ = 20.7, for the HEK293 and CaLu3 systems, respectively. When comparing the predicted values of R0 for the group of renal cell lines (Vero E6 and HEK293) against the group of nonrenal cell lines (HeLa and CaLu3), it was found that the difference is not significant (t = −1.543, p = 0.183, α = 0.01) under the conditions of this study, assuming equal variances with a normal distribution of the basic reproduction number. This suggests that the considered cell lines show similar viral kinetic behavior (see Table 2). The absence of difference between the values of R₀ for the studied cell groups is important because it corresponds to the spectrum of cell lines susceptible to viral infection by coronavirus, which has been reported as similar for the spike protein of SARS-CoV-2, with confirmed entry in CaLu3, HEK293, and Vero E6 cells.⁷⁹

Viruses are considered to infect susceptible cells (U) at an effective infection rate (β) according to eqs 1 and 2 (see section materials and methods). It was assumed that the normalized transcriptional values per million (nTPM) represent the percentage value of U. To do this, the average of the nTPM values reported for potential receptors and each cell line was considered. To calculate the basic reproduction number (R₀),eq 3 was used, where virions are released from infected cells productively at a rate per day (p), generating a number of dead infected cells (δ).

However, under the conditions of this study, the predictions made differ from those of other reports. These reports indicate that virus production tends to be higher in Vero cells than in CaLu3 cells after SARS-CoV-2 infection. These reports suggest that in Vero cells, SARS-CoV-2 entry is mediated by cathepsin L rather than TMPRSS2 due to low or rare expression of TMPRSS2.⁸⁰ However, RNA expression analysis from normalized cell line transcript values reported in The Human Protein Atlas differs from these statements and points to higher TMPRSS2 expression in Vero cells than in CaLu3 cells.²⁸ These discrepancies could affect predictions in terms of cell susceptibility and hence estimates of viral replication.

3.2. Theoretical Affinity of Targeted Drugs for Disruption of SARS-CoV-2 Receptor-Binding Domain (RBD)

Table 4 shows the results of the molecular docking methods and scoring functions used to predict the relative binding energies of the considered drugs to the SARS-CoV-2 spike system. All drugs showed thermodynamically favorable binding, with a mean of −8.30 kcal mol^–1. Telmisartan showed the most favorable thermodynamic mean, with an energy of ≈−9.45 kcal mol^–1. Followed by the anticoagulant nafamostat (≈−9.15 kcal mol^–1) (see Figure 1). The rest of the analyzed compounds showed a thermodynamically favorable binding energy of ≈−7 kcal mol^–1. In all cases, captopril was the compound with the least favorable binding energy (≈−5.56 kcal mol^–1), very similar to the control drug used (see 3).

Table 4. Results of the Predicted Kinetics and Inhibitory Potency of the Drugs Considered in This Study in Terms of IC₅₀ and pIC₅₀ on the SARS-CoV-2 Receptor-Binding Domain (RBD).

druga	EC50(M)b	pEC50b	IC50(M)c	pIC50c	IC50(M)d	pIC50d
telmisartan (CID_65999)	1.0 × 10–7	5.0	1.9 × 10–8	7.7	6.9 × 10–8	7.2
nafamostat (CID_4413)	1.0 × 10–7	7.0	3.5 × 10–8	7.5	1.0 × 10–7	7.0
camostat (CID_2536)	1.0 × 10–6	6.0	3.0 × 10–7	6.5	3.1 × 10–7	6.5
hydroxychloroquine (CID_3652)	4.2 × 10–6	5.4	4.2 × 10–7	6.4	3.4 × 10–7	6.5
chloroquine (CID_2719)	4.4 × 10–6	5.4	4.5 × 10–7	6.3	3.5 × 10–7	6.5
ibuprofen (CID_3672)	2.5 × 10–5	4.6	7.3 × 10–7	6.1	3.9 × 10–7	6.4
captopril (CID_44093)	1.0 × 10–5	5.0	1.3 × 10–5	4.9	4.9 × 10–7	6.3
pyronaridine (CID_107771) (control)			4.5 × 10–5e	4.3e	1.0 × 10–8d	8.0d

^aAll compounds were obtained from the PubChem database (https://pubchem.ncbi.nlm.nih.gov/).

^bValues calculated in molar expression from the experimental data (see section materials and methods).

^cValues calculated in molar expression from theoretical binding kinetics and relative energies using the Dixon plot method.

^dValues calculated in molar expression from theoretical binding kinetics and relative energies using the IC50-to-K_i web tool (https://bioinfo-abcc.ncifcrf.gov/IC50_Ki_Converter/index.php). EC₅₀, half maximal effective concentration; pEC₅₀, negative logarithm of the EC₅₀ value; IC₅₀, concentration at which the drug is capable of inhibiting activity by 50%; pIC₅₀, negative logarithm of the IC₅₀ value

^eExperimental values (see section materials and methods).

Figure 1.

To illustrate, we present the most stable conformation of two compounds: (a) nafamostat (PubChem CID_4413) and (b) pyronaridine (control) (PubChem CID_107771) within the binding pocket of the interface connecting the ACE2 receptor and the receptor-binding domain (RBD). The position and orientation of the nafamostat and pyronaridine structures are indicated within a circle, while the closest residues are displayed in the lower right corner. These interactions were predicted by using the BIOVIA Discovery Studio Visualizer tool. In the context of this study, nafamostat exhibited the most favorable pose within the interface formed between the RBD of SARS-CoV-2 and ACE2.

Table 3. Comparative Analysis of Various Affinity Parameters of the Drugs and Controls Considered in This Study against the Receptor-Binding Domain (RBD) of SARS-CoV-2.

		kcal mol–1
druga	types	DockTb	Vinac	Vinardoc	AD4c	Ki (uM)d	Binding forcee (kcal.mol–1 Å–2)	MM/PBSAe (kcal.mol–1)	Interface RBDg
telmisartan (CID_65999)	ARBs	–8.71	–9.65	–9.70	–9.72	0.13	8.96	–4.03	–
nafamostat (CID_4413)	anticoagulant	–7.47	–9.75	–9.94	–9.45	0.23	9.18	–10.38	+
camostat (CID_2536)	serine protease inhibitor	–7.38	–8.18	–8.00	–7.72	1.99	9.31	–15.24	–
hydroxychloroquine (CID_3652)	antimalarials	–7.97	–6.84	–7.84	–7.84	2.79	9.78	–8.72	–
chloroquine (CID_2719)	antimalarials	–7.76	–7.04	–7.77	–7.77	2.98	9.20	–13.79	–
ibuprofen (CID_3672)	NSAID	–7.24	–7.35	–7.93	–6.65	4.85	9.34	–9.75	–
captopril (CID_44093)	ARBs	–6.40	–5.95	–4.77	–5.13	88.55	9.26	–7.61	–
pyronaridine (CID_107771) (control)	antimalarials	–7.60	–5.31	–5.09	–4.22	0.36f*	7.78	–3.27	+

^aAll compounds were obtained from the PubChem database (https://pubchem.ncbi.nlm.nih.gov/).

^bScore calculated from DockThor (https://dockthor.lncc.br/v2/).

^cScore calculated from DINC (http://dinc.kavrakilab.org/).

^dTheoretical inhibitory kinetics of each ligand-protein complex calculated from the mean energy.

^eForce constant and Molecular Mechanics/Poisson–Boltzmann surface area (MM/PBSA) calculations from the minimum energy structures of the complexes after an MD of 100 ns. ARBs, angiotensin II receptor blockers; NSAID, nonsteroidal anti-inflammatory drugs.

^fExperimental values (see section materials and methods).

^gThe negative sign (−) is used to designate the docking outside the interface established by the RBD with receptors like ACE2, and the positive sign (+) to indicate the docking that occurs within the interface.

These results correspond to reports that compounds such as nafamostat and camostat have very favorable binding energy values against the coronavirus spike glycoprotein. Indeed, nafamostat has shown the most favorable interactions and inhibitory power in previous experiments.⁸¹ Moreover, the results garnered with Vinardo carry a significant weight. This is because Vinardo is an algorithm that boasts superior docking, scoring, classification, and virtual evaluation capabilities, consistently outperforming both Vina and AD4. It can classify active compounds over inactive ones. Vinardo (Vina RaDii Optimized) is a scoring function that incorporates component terms similar to those of the Vina scoring function. These component terms include steric interactions, hydrophobic interactions, and nondirectional hydrogen bonds. However, Vinardo distinguishes itself from Vina through several modifications. It features a modified steric interaction term, utilizes new atomic radii, and employs simplified interactions.⁸²

In terms of the theoretical inhibitory kinetics of each ligand–protein complex, which was calculated from the respective relative energies of binding, a mean K_i = 14.50 μM was predicted, with a minimum and maximum inhibition constant for the compounds telmisartan (K_i = 0.13 μM) and captopril (K_i = 88.55 μM), respectively. Approximately 70% (5/7) of the tested compounds (targets of this study) showed a theoretical inhibition constant K_i < 3 μM. Only a value close to 28% (2/7) of the compounds had a K_i < 0.5 μM (telmisartan and nafamostat), similar to what was described for the control drug (K_i ≈ 0.4 μM) (see Table 4). Additionally, binding affinities based on the relative force constant and free energy binding calculation (ΔG_bind) considering the MM/PBSA method on the final lowest energy structures after MD predicted thermodynamically favorable docking. In the case of the force constant calculation, a mean of ≈9 kcal mol^–1 Å^–2 was predicted, with a minimum and maximum of 8.96 and 9.78 kcal mol^–1 Å^–2 for the compounds telmisartan and hydroxychloroquine, respectively. While with MM/PBSA, a mean of ≈ −9 kcal mol^–1 was predicted, with a minimum and maximum of ≈ −15 kcal mol^–1 Å^–2 and ≈ −4 kcal mol^–1 for the camostat and telmisartan compounds, respectively (see Table 3).

As the relevance of the ACE2 receptor in mediating infection has been confirmed, its interaction with the SARS-CoV-2 receptor-binding domain (RBD) was studied.⁸³ The interface formed between the RBD and ACE2 was explored for illustrative purposes only. Of the compounds tested and under the conditions of this study, only nafamostat showed favorable poses in the interface formed between the receptor-binding domain (RBD) of SARS-CoV-2 and ACE2 after 24 runs with a total of 10⁶ evaluations per run, similar to what was described for the pyronaridine control (see Figures 1–3 and Table 3). The poses presented by nafamostat turned out to be the most statistically likely to be among the top three thermodynamically most favored scores out of a total of 2.4 × 10⁷ possible positions, which were guided by blind docking. In all cases, the interactions of the compounds with the spike were dominated by hydrophobic interactions, with the greatest diversity of types of interactions observed for the compound camostat. All compounds mediated interactions with a relative binding strength of ≈9 kcal mol^–1 Å^–2.

Figure 3.

MD simulation of compounds with the most favorable pose in the interface formed between the RBD of SARS-CoV-2 and cellular receptors like ACE2 under the conditions of this study (total time of 100 ns). (a) Root mean squared deviations (RMSDs) of Cα during each 10 ns, (b) root means square fluctuation (RMSF), (c) radius of gyration-guided motions (R_g), (d) number and type of interactions of nafamostat, and (e) pyronaridine. RBD, receptor-binding domain.

In relation to the inhibitory potency calculated from the reported experimental EC₅₀ values, a mean potency was predicted in terms of pEC₅₀ = 5.5. Specifically, nafamostat was predicted to be the compound with the highest theoretical inhibitory potency calculated from the reported experimental EC₅₀ values, with a value of pEC₅₀ = 7.0. While ibuprofen was the compound with the lowest inhibitory potency calculated from the data reported for NSAIDs, with a pEC₅₀ = 4.6 (see Table 4).

Additionally, the predictions based on EC₅₀ were validated applying the Dixon method and the IC50-to-K_i web tool using the theoretical values of IC₅₀. In all cases, the inhibitory potency calculated in terms of pIC₅₀ was reproducible regardless of the compound and predictive method used (Dixon plot and IC50-to-K_i web tool) for the calculation of IC₅₀, showing an overall difference between methods with a mean of ≈0.1, and a difference between the mean of pEC₅₀ (experimental data reported) and pIC₅₀ ≈ 1 (see Table 4).

Studies have shown that nafamostat exhibits strong inhibitory effects on coronavirus-mediated membrane fusion in HEK293 and CaLu3 cells.¹⁸ Compared to other compounds such as camostat, nafamostat has demonstrated more favorable inhibitory activity. This superiority has been attributed to nafamostat’s ability to readily form a stable covalent enzyme–substrate intermediate, which effectively explains its high potency. These findings align with in vitro experiments and are consistent with previous virus cell entry assays.⁸⁴

Molecular dynamics simulations showed that upon ligand binding, all complexes exhibited conformational fluctuations with an RMSD of ≤3.2 Å. These results in terms of RMSD indicate that the junctions are stable throughout the simulation period of 100 ns. This is in addition to being thermodynamically favorable (as previously predicted). These observations together with the predictions in terms of RMSF and R_g allow us to infer that all the tested compounds form stable complexes (see Figure 2). It is important to note that both nafamostat and the pyronaridine control were compound-guided by blind docking and showed interactions with the RBD interface that were predicted to be stable over time according to the molecular dynamics (MD) simulation analysis at 100 ns and in terms of RMSD, RMSF, R_g, and the number and type of interactions (see Figure 3). Despite these standard theoretical approaches showing promising favorable and stable interactions over time for the complexes considered in this study, it is recommended to validate the predictions of such systems through lead compound similarity analysis in order to better understand the behavior of each selected drug against the target protein, as suggested.^85,86

Figure 2.

MD simulation for each complex (total time 100 ns). (a) Root mean squared deviations (RMSD) of Cα during each 10 ns, (b) root means square fluctuation (RMSF), and (c) radius of gyration-guided motions (R_g). RBD, receptor-binding domain; HCQ, hydroxychloroquine.

3.3. Prediction of Theoretical Antiviral Effectiveness of Drugs from Experimental Data

Table 5 shows the prediction of the theoretical antiviral effectiveness of each of the compounds considered based on the susceptibility of the cell type. Predictions made from reported experimental EC₅₀ (μM) and IC₅₀ (μM) values calculated in this study from relative binding energies and theoretical inhibitory kinetics are presented. In terms of antiviral effectiveness calculated from the EC₅₀, similar mean drug efficacy was predicted as a function of cell type susceptibility (F_(3,24) = 0.002, p = 0.999, α = 0.01) with an effect in percentage terms of ≈68%.

Table 5. Results of the Predicted Antiviral Efficacy (ε) of the Drugs Considered in This Study from the EC₅₀ and IC₅₀ Values and in Relation to the Types of Cell Lines Studied.

drug(s)a	EC50(μM)b	antiviral efficacy (%)e
		Vero E6	HeLa	HEK293	CaLu3
nafamostat	0.1	99	99	99	99
camostat	1	86	86	86	86
chloroquine	4.4	80	80	80	80
hydroxychloroquine	4.2	72	72	72	72
telmisartan	10	57	57	59	56
captopril	10	57	57	59	56
ibuprofen	25	28	28	29	28

drug(s)a	IC50(μM)c	antiviral efficacy (%)
		Vero E6	HeLa	HEK293	CaLu3
nafamostat	0.03	99	99	99	99
camostat	0.3	95	95	95	95
chloroquine	0.45	97	97	97	97
hydroxychloroquine	0.42	97	97	97	97
telmisartan	0.02	99	99	99	99
captopril	13.28	89	89	91	88
ibuprofen	0.73	94	94	94	94

drug(s)a	IC50(μM)d	antiviral efficacy (%)
		Vero E6	HeLa	HEK293	CaLu3
mafamostat	0.1	99	99	99	99
camostat	0.31	95	95	95	95
chloroquine	0.35	98	98	98	98
hydroxychloroquine	0.34	97	97	97	97
telmisartan	0.07	98	98	98	98
captopril	0.49	88	88	88	88
ibuprofen	0.39	97	97	97	97

^aAll compounds were obtained from the PubChem database (https://pubchem.ncbi.nlm.nih.gov/).

^bValues calculated in molar expression from the experimental data (see section materials and methods).

^cValues calculated in molar expression from theoretical binding kinetics and relative energies using the Dixon plot method.

^dValues calculated in molar expression from theoretical binding kinetics and relative energies using the IC50-to-K_i web tool (https://bioinfo-abcc.ncifcrf.gov/IC50_Ki_Converter/index.php). EC₅₀, half maximal effective concentration; pEC₅₀, negative logarithm of the EC₅₀ value; IC₅₀, concentration at which the drug is capable of inhibiting activity by 50%; pIC₅₀, negative logarithm of the IC₅₀ value.

^ePercentage expression of the reduction in viral infectivity (ε) in terms of the effective radius of infection (β) by the factor (1 – ε_β).

However, a statistically significant difference was predicted between the individual antiviral effectiveness of each of the compounds tested (F_(6,21) = 4088.756, p < 0.00001, α = 0.01). From the EC₅₀ values, it was predicted that ≈42% (3/7) of the considered drugs may show a theoretical antiviral effectiveness ≥80%. Specifically, the compound nafamostat was predicted to have the highest theoretical antiviral effectiveness, regardless of the type of simulated cell susceptibility, with ≈99% effectiveness, followed by camostat (≈86%) and chloroquine (≈80%).

While ibuprofen presented the lowest predicted theoretical antiviral effectiveness (≈28%). In the HEK293 cell line, the theoretical antiviral effect of approximately 42% (3/7) of the tested compounds (telmisartan, captopril, and ibuprofen) was slightly less favorable followed by a similar behavior in the Vero E6 and HeLa lines. On the other hand, and in a strict (but not significant) sense, the lowest theoretical antiviral response was predicted in the CaLu3 system under the conditions of this study.

For illustrative purposes, we sought to predict and compare the theoretical antiviral effectiveness of the compounds from the IC₅₀ values calculated in this study. The predictions with the Dixon plot model reproduced the trend calculated from the EC₅₀ values (although with a greater magnitude of the values) showing a similar mean response of the cells (F_(3,24) = 0.020, p = 0.996, α = 0.01) and also a statistically significant difference between the antiviral effectiveness of each of the compounds tested (F_(6,21) = 196.708, p < 0.00001, α = 0.01). It is important to note that, just like the values of binding free energy (ΔG) and K_i, the values in terms of IC₅₀ and pIC₅₀ presented by the control are also within the predicted values for each compound considered in this study. Therefore, a similar behavior would be expected in terms of theoretical antiviral activity (see Table 4).

It is important to point out that after applying the Dixon method and the IC50-to-K_i web tool to estimate the IC₅₀ values, an increase in the theoretical antiviral effectiveness of all of the compounds was observed. This decreased the discriminatory capacity of the model proposed here by using the EC₅₀ experimental values. However, regardless of the method applied to calculate the IC₅₀, nafamostat presented the highest theoretical antiviral effectiveness in each case (see Table 5). It is important to note that one of the limitations of this study is that the predictions of antiviral efficacy, made from the kinetic/dynamic properties reported for the drugs studied, only consider a single initial dose of each drug for 24 h. This is consistent with the findings of previous studies of other drugs, such as hydroxychloroquine,¹⁷ nafamostat,^51,52 camostat,⁵³ telmisartan,⁵⁴ chloroquine,⁵⁵ ibuprofen,⁵⁶ and captopril.^57,58 With this in mind, we suggest conducting an analysis under the same conditions as this study but with variations in concentration taken into account as well as extending the exposure time to the drugs under investigation. This approach has been recommended in other studies.¹⁵

The findings of this study regarding the antiviral activity and the impact of drugs on the viral kinetics of SARS-CoV-2 have previously been reported in cell lines such as CaLu3 and HEK293 with other coronaviruses. Nafamostat has been shown to possess superior antiviral activity compared to camostat against MERS-CoV.^17,18 Conversely, the camostat has demonstrated activity against the TMPRSS2 receptor, partially inhibiting spike-driven entry into Caco-2, Vero E6, and CaLu3 cells against MERS-CoV, SARS-CoV-1, and SARS-CoV-2. However, it has been suggested that the spread of SARS-CoV-2 relies on the activity of multiple receptors, which could explain the partial inhibition of viral activity in CaLu3 cells when exposed to camostat.^16,17

While chloroquine and hydroxychloroquine are currently used for the treatment of malaria and amebiasis, studies have reported their potential to block the interaction between the SARS-CoV spike protein’s receptor-binding dDomain (RBD) and ACE2 under cell culture conditions in Vero E6 cells. Both chloroquine and hydroxychloroquine have the ability to increase the endosomal pH, which is necessary for the conformational changes in the spike protein during virus-cell fusion. Additionally, these compounds can affect the terminal glycosylation of the cellular receptor ACE2, thereby reducing the affinity of SARS-CoV/SARS-CoV-2 for ACE2.^17,21

Although there is affinity in energetic terms, a generalized effect of inhibitory activity of captopril and telmisartan was not observed. This could be because these compounds have more affinity for the ACE2 cell receptor than for viral spike regions without compromising the activity of ACE2. This has been reported previously.^22,23 Although the predicted results with ibuprofen correspond to a potential ability to interact favorably with the spike protein, as has been reported,^19,20,59 no significant inhibitory activity was observed under the conditions of this study. Ibuprofen has been extensively studied in cell lines, including CaLu3. It has been shown that this drug has no effect on ACE2 modulation, despite hypotheses earlier in the pandemic. This has led to the conclusion that there are no risks associated with the use of nonsteroidal anti-inflammatory drugs (NSAIDs) like ibuprofen during a SARS-CoV-2 infection.⁵⁹

The theoretical susceptibility previously predicted for each of the cell lines considered was used as a control of the tests to evaluate the theoretical antiviral activity of the drugs (Table 2 and Figure 3). For illustrative purposes, the effect of drugs on the number of susceptible cells and the effect of their potential antiviral activity on the number of cells to be infected were simulated. Specifically, antiviral drugs that reduce viral infectivity in terms of the effective radius of infection (β) by the factor (1 – ε_β) reduce the basic reproductive number, R₀, by a factor (1 – f(ε_β)). If (1 – f(ε_β))R₀ is <1, the virus will almost certainly become extinct.

In this sense, the relationship (1 – f(ε_β))R₀ to describe the theoretical antiviral efficacy of drugs after in vitro treatment for 24 h was designated as ϕ. In all the cell lines tested, a mean of the relationship ϕ ≈ 0.065 was predicted, indicating that in all the cell models considered, a decrease in total viral particles is predicted. This effect was simulated for a period of 24 h exposure to drugs, with the HEK293 system being the cell line where the greatest inhibition effect was predicted, but without a statistically significant difference (F_(3,24) = 0.003, p = 0.999, α = 0.01) (see Table 6).

Table 6. Predicted Antiviral Efficacy of the Drugs Considered in This Study in Terms of the ϕ Relationship Calculated from the Reduction in the Effective Radius of Infection (β) and the Reduction in the Basic Reproductive Number, R₀ with Respect to the Cell Lines Studied.

	antiviral efficacy (ϕ = (1 – f(εβ))R0)b
drug(s)a	Vero E6	HeLa	HEK293	CaLu3
nafamostat	0.002	0.002	0.002	0.002
camostat	0.029	0.028	0.029	0.029
chloroquine	0.041	0.041	0.041	0.041
hydroxychloroquine	0.057	0.057	0.058	0.058
telmisartan	0.088	0.087	0.084	0.091
captopril	0.088	0.087	0.084	0.091
ibuprofen	0.148	0.146	0.146	0.149

^aAll compounds were obtained from the PubChem database (https://pubchem.ncbi.nlm.nih.gov/).

^bThe relationship (1 – f(ε_β))R₀ designated in this study as ϕ describes the theoretical antiviral efficacy of drugs after in vitro treatment for 24 h. Antiviral drugs that reduce viral infectivity in terms of the effective radius of infection (β) by the factor (1 – ε_β) reduce the basic reproductive number, R₀, by a factor (1 – f(ε_β)). If ϕ = (1 – f(ε_β))R₀ is <1, the virus will almost certainly become extinct (see section materials and methods).

In relation to the individual activity of the drugs, it was predicted that under the conditions of this study, all the compounds could theoretically have the capacity to reduce viral infectivity by different magnitudes in terms of the effective radius of infection (β) and reduce the basic reproductive number, R₀, (ϕ ratio) in all cell lines tested. Specifically, in relation to the individual drug activity, nafamostat was predicted to be the compound with the highest ability to reduce β-viral infectivity and to reduce the R₀ number (ϕ = 0.002) in all considered cell lines. It was followed by camostat (ϕ = 0.029) and chloroquine (ϕ = 0.041), with ibuprofen being the least favorable compound (ϕ = 0.147) (Table 6).

Given that renal and pulmonary cell systems were simulated, it is important to note that all the compounds considered in this study are widely known drugs approved by the FDA with known kinetic and dynamic characteristics (all compounds were obtained from the PubChem database: https://pubchem.ncbi.nlm.nih.gov/). Therefore, the prediction of potential toxicities was also carried out using the latest generation of machine learning analysis applied to experimental data in vitro and in vivo and related to toxicology and the analysis of the toxic substructure. Around 77% (7/9) of the tested compounds may be potentially toxic in at least one type of organ or tissue, with a statistical significance of more than 99%. Specifically, ≈55% (5/9) of the compounds may be potentially toxic at the renal level, while ≈22% (2/9) may be potentially toxic at the respiratory level. It is important to note that only nafamostat and camostat did not demonstrate any toxic action from data-driven models (https://tox-new.charite.de/protox_II/index.php?site=home). These predictions hold significance due to the utilization of the eMolTox method (http://xundrug.cn/moltox), which incorporates a comprehensive set of 174 in vitro/in vivo experimental data sets pertaining to toxicology for model development (Supporting Information Tables S2–S8). The method further employs Mondrian’s conformal prediction (MCPs) to estimate the confidence level associated with the generated predictions.⁷⁵

As shown in Figures 4 and 5, the most favorable drugs such as nafamostat and camostat mediated a theoretical decrease in the number of infected cells (I) and the number of dead infected cells (δ) despite the mock susceptibility (U). The theoretical activity presented by compounds like nafamostat can be attributed to the impact of the reduction in the effective radius of infection (β) by the factor (1 – ε_β), which in turn translates into a reduction in the basic reproductive number, R₀, by a factor (1 – f(ε_β)). Specifically, the effective infection ratio was calculated as β ≈ 0.31, which represents an overall theoretical reduction of the simulated daily rate of infection by ≈15% with respect to reported values used as control (β = 2).¹² Consistently, nafamostat was the compound predicted to reduce the daily effective infection rate the most (β = 0.01) in all simulated cell models (see Figure 4B, Supporting Information Figures S1–S6). Followed by camostat in HEK293 (β = 0.14) (see Figure 5B) and chloroquine in CaLu3 (β = 0.20), with ibuprofen being the compound with the least effect on viral kinetics (see Supporting Information Figures S1–S6). These predictions based on theoretical models of cell lines correspond with the reports of the usefulness of nafamostat administered for the prevention of SARS-CoV-2 infection, above camostat.⁶⁰

Figure 4.

Mechanistic model is employed to estimate the rates of uninfected cells (U), infected cells (I), and death of infected cells (δ). For illustrative purposes, a representative graph of the expression systems mentioned in the text is provided for each cell line studied using the limited target cell model. Additional graphics can be found in Supporting Information Figures S1–S6. (A) Vero E6 cell line and (B) Vero E6 + nafamostat.

Figure 5.

Mechanistic model is utilized to provide estimates of the rates of uninfected cells (U), infected cells (I), and death of infected cells (δ). To enhance our understanding, a representative graph depicting the expression systems mentioned in the text is included for each cell line studied using the limited target cell model. Supporting Information Figures S1–S6 contain additional graphics. (A) HEK293 cell line, (B) HEK293 + camostat, (C) CaLu3 cell line, and (D) CaLu3 + nafamostat.

The predictions made in this study indicate that at a theoretical level, it is likely that the infection of cell lines such as CaLu3 may be slightly more favored than in Vero E6 if it is mediated by the contribution of receptors other than ACE2 with a higher level of expression. In a previous work,⁷⁸ in which a smaller number of potential receptors was considered, and without studying the antiviral activity of drugs, our predictions also found that the ACE2 receptor probably does not contribute by itself to explain the theoretical mechanism of infection in CaLu3. However, the limitations reported when measuring the active replication of the virus at a significant level in diverse samples associated with the respiratory tract⁸⁷ do not correspond to the hypothesis of an exclusive participation of potential receptors that exhibit a high level of expression in cells such as CaLu3. This could suggest the involvement of other ACE2-independent cooperative or coupled effects that may mediate theoretical CaLu3 infection, as has been suggested.^88,89

On the other hand, the differential behavior of drugs with possible antiviral properties against SARS-CoV-2 was observed depending on the type of simulated cell line, as has been reported.⁹⁰ In fact, it has been pointed out that CaLu3 cells tend to be more sensitive to inhibitors than HEK293 cells, even with the use of compounds such as nafamostat and camostat.¹⁸ As an instance, camostat, which functions as an inhibitor of the plasma membrane protease TMPRSS2, has exhibited activity in CaLu3 cells while not displaying the same effect in Vero E6 cells, as anticipated in this study. This reinforces the significance of comprehending the translatability of a cellular model of infection,⁸⁰ especially as camostat has been reported to have no antiviral effects in vivo in animal models compared to nafamostat.⁶⁰

Indeed, studies of this nature are crucial, as they provide theoretical backing for the impact of the intricate multiple entry mechanisms associated with SARS-CoV-2. This underscores the necessity to select optimal cell lines for compound evaluation and screening, particularly in studies aimed at uncovering broad antiviral mechanisms.⁸⁵ Finally, it is important to point out that these cell line-based predictions contribute to the tracing of possible routes of infection, the prediction of potentially susceptible organs, and the investigation of interventions for the prevention, control, and treatment of infection by theoretically proposing ideal cell models. These are relevant aspects considering that it has been found that cell lines grown under standard laboratory conditions have difficulties in reproducing the behavior of cells that live in patients.⁹¹

4. Conclusions

Theoretical cellular systems considered and simulated, which express multiple receptors simultaneously, have demonstrated susceptibility similar to SARS-CoV-2. Additionally, the anticipated viral kinetic behavior in the simulated cell lines aligns with the conditions of this study. Among the cell lines tested, the CaLu3 system is projected to be the most susceptible to SARS-CoV-2 infection. This susceptibility can be attributed statistically to the presence of receptors such as CD147 and Cathepsin-L, which likely contribute significantly to the observed susceptibility. Nafamostat, among the compounds tested, exhibits the most favorable characteristics in terms of thermodynamics, kinetics, theoretical antiviral activity, and potential safety (toxicity) associated with SARS-CoV-2 spike protein-mediated infections in the tested cell lines. This study provides mathematical and bioinformatic models for identifying optimal cell lines for evaluating and detecting compounds, particularly in studies focused on the antiviral mechanisms of drugs for potential reuse. It is important to note that these observations must be experimentally validated, and this research is expected to stimulate future quantitative experiments.

Supporting Information Available

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acsomega.3c06968.

• Mechanistic model of the estimates of the rates of uninfected cells (U), infected cells (I), and death of infected cells (δ) (Figures S1–S6); detailed parameters for the MM/PBSA calculation procedures (Table S1) (PDF)

• Supporting information on compound toxicity (Tables S2–S8) (XLS)

Author Contributions

L.G.P.: Conceptualization, Methodology, Investigation, Writing – Reviewing and Editing. C.L. and M.L.H.-L.: Investigation. J.V.-V.: Reviewing. J.L.P.: Writing – Reviewing, Original draft preparation. Y.M.-P.: Writing – Reviewing, Original Draft Preparation. F.M.-R.: Reviewing. Y.J.A.: Writing – Reviewing and Editing, Conceptualization.

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sector.

The authors declare no competing financial interest.