Identifying non-nucleoside inhibitors of RNA-dependent RNA-polymerase of SARS-CoV-2 through per-residue energy decomposition-based pharmacophore modeling, molecular docking, and molecular dynamics simulation

Abstract

Background and Objective

The current coronavirus disease-2019 (COVID-19) pandemic has triggered a worldwide health and economic crisis. The severe acute respiratory syndrome coronavirus-2 (SARS-CoV-2) causes the disease and completes its life cycle using the RNA-dependent RNA-polymerase (RdRp) enzyme, a prominent target for antivirals. In this study, we have computationally screened ∼690 million compounds from the ZINC20 database and 11,698 small molecule inhibitors from DrugBank to find existing and novel non-nucleoside inhibitors for SARS-CoV-2 RdRp.

Methods

Herein, a combination of the structure-based pharmacophore modeling and hybrid virtual screening methods, including per-residue energy decomposition-based pharmacophore screening, molecular docking, pharmacokinetics, and toxicity evaluation were employed to retrieve novel as well as existing RdRp non-nucleoside inhibitors from large chemical databases. Besides, molecular dynamics simulation and Molecular Mechanics/Generalized Born Surface Area (MM/GBSA) method were used to investigate the binding stability and calculate the binding free energy of RdRp–inhibitor complexes.

Results

Based on docking scores and significant binding interactions with crucial residues (Lys553, Arg557, Lys623, Cys815, and Ser816) in the RNA binding site of RdRp, three existing drugs, ZINC285540154, ZINC98208626, ZINC28467879, and five compounds from ZINC20 (ZINC739681614, ZINC1166211307, ZINC611516532, ZINC1602963057, and ZINC1398350200) were selected, and the conformational stability of RdRp due to their binding was confirmed through molecular dynamics simulation. The free energy calculations revealed these compounds possess strong binding affinities for RdRp. In addition, these novel inhibitors exhibited drug-like features, good absorption, distribution, metabolism, and excretion profile and were found to be non-toxic.

Conclusion

The compounds identified in the study by multifold computational strategy can be validated in vitro as potential non-nucleoside inhibitors of SARS-CoV-2 RdRp and holds promise for the discovery of novel drugs against COVID-19 in future.

Introduction

Coronaviruses (CoVs) constitute a distinct family of positive-stranded RNA viruses that infect humans and animals, causing respiratory and gastrointestinal diseases [1]. Alpha, Beta, Gamma, and DeltaCoVs are the four genera of CoVs; the first two specifically infect humans. Severe acute respiratory syndrome coronavirus (SARS-CoV) caused an outbreak of SARS (2002–2003) with a 10% death rate, while MERS-CoV triggered a deadly epidemic (2012) with a fatality rate of 37% [2]. Severe acute respiratory syndrome coronavirus-2 (SARS-CoV-2) is a novel human CoV that emerged in the Chinese city of Wuhan in late 2019 [3]. As per the weekly epidemiological update of the World Health Organization (WHO) [4], the virus has infected over 641 million people and resulted in 6.6 million deaths globally as of December 4, 2022. The ongoing coronavirus disease-2019 (COVID-19) pandemic is caused by the SARS-CoV-2 [5]. This virus infection causes symptoms that range from moderate coughing, respiratory illness, and fever to severe pneumonia, breathing difficulty, multi-organ failure, and death [6]. Amidst the preventative measures taken to contain the disease, its spread remains uncontrolled. Researchers across the globe are working actively to stop viral multiplication and spread [7].

SARS-CoV-2 belongs to the BetaCoVs genus and is a positive-strand ss-RNA virus [8]. The viral genome is about 30 kb long and consists of fourteen open reading frames (ORFs) that encode 29 proteins[9], [10]. ORF1ab and ORF1a, respectively, encode the polypeptides pp1ab and pp1a [9]. Genes code for nucleocapsid (N), envelope (E), membrane (M), surface (S), and eight accessory proteins towards the 3′ end of the genome. The non-structural proteins nsp1–nsp16 are produced by the cleavage of two large polyproteins, pp1ab and pp1a, via proteolytic enzyme [9].

RNA-dependent RNA polymerase (RdRp), which mediates the virus RNA production, is one of these non-structural proteins. With the help of co-factors (nsp7 and nsp8), RdRp plays an essential part in the transcription and replication cycle of SARS-CoV-2 [11], [12]. This enzyme is a prime target for drug development owing to its conservation across evolutionarily distant RNA viruses and the lack of a human counterpart [13], [14]. Due to its critical function, RdRp has been the effective target of multiple drugs that are either approved or under evaluation in clinical stages for treating various viral infections, including influenza viruses, respiratory syncytial virus, viral hemorrhagic fever, and hepatitis C virus[15], [16], [17], [18], [19], [20], [21]. The structure of the SARS-CoV-2 RdRp has recently been determined using cryo-electron microscopy (Cryo-EM)[2], [12], revealing the protein right-handed cup architecture where an N terminal β-hairpin [amino acids (a.a) 31–50] and an expanded NiRAN domain (nidovirus RdRp-associated nucleotidyl transferase, 115–250 a.a) are present. The NiRAN domain is connected to the RdRp domain (a.a 366–920) via an interface domain of 251–365 a.a [2], [22]. The polymerase domain is segregated into three subdomains: palm, thumb, and finger. The palm domain (composed of 582–620 and 680–815 a.a) is the largest of the three subdomains. The thumb subdomain (397–581 and 621–679 residues) forms a closed circle with the finger subdomain (819–920 a.a). The interaction of nsp7 and nsp8, as well as an nsp7–nsp8 heterodimer, stabilizes the close conformation [2], [12].

The primary therapeutic strategy for COVID-19 is supportive care [23], supplemented with a combination of wide-ranging antibiotics, antiviral drugs, corticosteroids, and convalescent plasma [24], [25]. Previously, several computer-aided studies reported the potential activity of anti-COVID-19 compounds targeting structural and non-structural proteins of SARS-CoV-2[26], [27], [28], [29]. Besides, specific in silico and in vitro research recounted nucleoside analogs (NAs) activity, including remdesivir, galidesivir, favipiravir, and sofosbuvir against RdRp of SARS-CoV-2 [7]. Among these, remdesivir, the Ebola virus drug, received greater attention [30]. It has also been shown to block the MERS coronavirus's RdRp [31]. Due to encouraging clinical results initially, remdesivir received emergency approval from Food and Drug Administration (FDA) for use against critically ill patients in May 2020 [32]. Nevertheless, the drug was reported ineffective and suspended from use on COVID-19 patients in November 2020 by WHO [33]. A distinctive proof reading function of the CoV makes them challenging to target with NAs due to the presence of an exonuclease domain nsp14 which removes the incorporated NAs, thereby resisting such antivirals [34], [35]. Although novel COVID-19 vaccines were developed, waning humoral responses, especially in vulnerable populations, were reported [36], [37], further highlighting the need to continue screening for compounds that could mitigate acute infections brought about by SARS-CoV-2 and emerging variants.

Identifying non-nucleoside inhibitors against the major therapeutic target, RdRp, could be a promising therapeutic strategy to treat the SARS-CoV-2 infection. Herein, we screened a huge collection of ∼690 million and 11,698 small molecule inhibitors from ZINC20 and DrugBank, respectively, to identify potential inhibitors of SARS-CoV-2 RdRp by targeting the RNA-primer strand binding site. Based on the current computational results, ZINC285540154, ZINC98208626, ZINC28467879, ZINC611516532, ZINC739681614, ZINC1166211307, ZINC1398350200, and ZINC1602963057 were identified as potential non-nucleoside inhibitors of SARS-CoV-2 RdRp and warrants in vitro and in vivo evaluation to establish their therapeutic relevance concerning COVID-19.

Methodology

SARS-CoV-2 RdRp structure

The Cryo-EM structures of SARS-CoV-2 RdRp in apo form (PDB ID: 7BV1) and in complex with suramin (PDB ID: 7D4F) were retrieved from the RCSB Protein Data Bank (https://www.rcsb.org/). The structure preparation tool implemented in Molecular Operating Environment version 2020.0901 (MOE) was employed to refine the structure with all atoms Amber14:EHT [38], [39] (Amber ff14SB combined with EHT) force field. The unclear residues in protein were fixed. The MOE loop builder was used to build the missing loops in protein. The C–N terminals were identified and charged for clarifying the start and termination of the protein. The amino acids were added with the missing hydrogens. All atoms were given the force field parameters, missing atom types, bond stretch parameters, missing angle parameters, and missing van der Waals parameters. For the protein–ligand complex, RESP and AM1-BCC charges were calculated [40].

Molecular dynamics simulation

Molecular dynamics (MD) simulation is used to explore protein behavior at the atomic level upon inhibitor binding. The 3D structure of the SARS-CoV-2 RdRp in complex with suramin was used as a reference structure and subjected to MD simulation in an implicit solvent implementing GPU version, Particle Mesh Ewald Molecular Dynamics (PMEMD)[41] engine embedded in AMBER 20 [42]. Using the residue-specific ff19SB [43] force field, the coordinate and topology file of the systems for protein residue parameters were generated with the LEap module incorporated in AMBER 20 [42]. Monovalent OPC ions [44], including Na+ and Cl^– (∼0.1 M), were added to neutralize the system. General Amber Force Field-2 (GAFF2) [45] with AM1-BCC [45] charges calculation was used to treat the small molecule inhibitors for the structure parametrization. In addition, the Parmchk2 tool of AMBER 20 was used to generate the small molecule’s missing force field parameters, and missing hydrogens were added by LEaP module. All systems were solvated in a truncated octahedral box with an OPC (optimal point charge) water model with 10 Å of the buffer distance. PMEMD engine [41] on GPUs was used to optimize parallel scaling in long-range electrostatics.

Initial minimization of the system was carried out in two steps: First, 20000 steps of steepest descent minimization, then 10000 steps of conjugate gradients minimization. Before the minimization step, steady heating from 0.1 to 300 K was performed in 400 ps time using an NVE (Microcanonical ensemble) and a Langevin thermostat. Using a Langevin thermostat [46] and 2.0 ps⁻¹ collision frequency, the kinetic energy of harmonic oscillators was adjusted for dynamic propagation. Following heating, density adjustment in the 400 ps run was made by following the same procedure. Next, the equilibration of all systems at 300 K for 2000 ps was carried out in an NVE ensemble for 400 ps without restraint and a pressure relaxation time of 2 ps. Employing the isotropic position scaling approach, the pressure was upheld constantly in the equilibration with a pressure relaxation time of 1 ps. In order to constrain all hydrogen bonds, the SHAKE algorithm was applied [47]. To compute long-range electrostatics, the particle-mesh Ewald [48]method was applied with a cut-off set to 8 Å. Finally, 100 ns production MD was continually run for apo (PDB ID: 7BV1), reference (PDB ID: 7D4F), and selected inhibitor systems following the protocol of equilibration stage. The final trajectory in every 10 ps was written for analysis.

Per Residue-free energy decomposition and hydrogen bond analysis

Per residue energy decomposition (PRED) of the total interaction energy was performed using the MMPBSA.py plugin of the AMBER 20 [49]. The final 1000 frames of the MD trajectory were used to perform PRED analysis to determine each residue's contribution to the overall binding energy profile between SARS-CoV-2 RdRp and ligand. The contribution of each residue was determined by the internal van der Waals and electrostatic energy as an overall average of the last 1000 steps selected from the trajectory. The total contribution energy of each residue was reported in kcal/mol and was followed by the backbone and side chine atom’s further decomposition. The hydrogen bonds (H-bonds) formed between the ligand and protein were calculated using the H-bond of the CPPTRAJ module. The 100 ns trajectory was used to analyze the bonding residues during the course of the simulation. The bond cut-off distance was selected at a 3.5 Å for the hydrogen donor and acceptor atoms with an angle of 120 degrees, respectively. The lifetime of each bond between the protein residue and ligand was reported with average distance and angle in angstroms.

Pharmacophore modeling and validation

Based on PRED and H-bond analysis of the RdRp–suramin complex, the pharmacophore model was generated by MOE pharmacophore query editor using EHT pharmacophore scheme, which creates annotation points on the ligand interacting atoms with the protein. Those H-bond donor/acceptor and aromatic features in ligands were identified that interact with critical residues of RdRp. The minimal energy strength for H-bond acceptor and donor features was set at 0.8 kcal/mol and 0.5 kcal/mol, respectively. A test database was created containing 36 inhibitors of RdRp having IC50 value range from 0.001–10 nM, retrieved from the BindingDB database [50] (Table S1). To differentiate between active and inactive compounds more precisely, the DUD-E decoy set (obtained from the DUD-E decoys database) [51] was used to assess the observed active molecules. The active molecules and the chosen DUD-E decoys were converted to an "mdb" format using MOE, and then a ROC curve (receiver operating characteristic curve) was generated. Key metrics, including active hits (AH), decoy compounds (DC), early enrichment factor (EF), overall compounds in the dataset (D), the total number of hits (TH), and quality of hit scoring (rscore) were used to evaluate the model's efficacy [52].

Compounds selection

A validated pharmacophore model was used to virtually screen the ZINC20 drug-like database (https://zinc20.docking.org/) containing ∼0.69 billion compounds. Before performing virtual screening (VS), applying cleaning rules on database compounds is essential. The MMFF94x force field was used to balance the ligands’ protonation states, and the geometry of each molecule was improved by adding missing hydrogens. The quick preparation tool of MOE database was used to compute and adjust the interatomic distances, angles, and dihedrals. The protein–ligand complex was optimized using constrained minimization of 0.01 Kcal/A². The generated pharmacophore features were used to filter the 3D structures from the database, and the extracted compounds were saved in a new database for future analysis.

Structure-based virtual screening

Molecular docking was employed in the structure-based VS of selected compounds via Dock application of MOE to predict the optimum binding mode of retrieved small molecules at the RNA primer strand binding site of RdRp. MOE's docking protocol was validated by re-docking the co-crystallized ligand of RdRp. Initially, the protein file was treated by the MOE-QuickPrep module to add hydrogen atoms and partial charges according to AMBER12: EHT force field. The described procedure was used to re-dock the co-crystallized ligand. Next, the Root Mean Square Deviation (RMSD) was estimated for re-docked conformation and crystal ligand. The obtained screening hits were docked in the active pocket of the RdRp using a validated docking protocol. Protein-Ligand Interaction Fingerprinting (PLIF) was used to examine the interactions of small molecules with the residues of RdRp binding site in the MOE database. Separate PDB files were created for each protein–ligand complex.

Prediction of pharmacokinetics, drug-likeness, physicochemical and toxicity

SwissADME tool [53] was used to study physicochemical properties, drug-likeness, pharmacokinetic and medicinal properties. Moreover, ProTox-II server [54] was used to study organ toxicity, toxicity endpoints, Tox21-Nuclear receptor signaling pathways, and the Tox21-Stress response pathway.

MD simulation of selected compounds

Apo structure of RdRp was used to compare the structural alteration upon inhibitor binding in the selected docked complexes. The protocol mentioned above reported for the RdRp-suramin complex, and RdRp-apo-state was followed to simulate selected inhibitors. The stability of the simulated complexes was analyzed by CPPTRAJ, the prime module of AMBER20, for analyzing output MD trajectory. The Cα atoms of the protein were used to calculate the Root mean square deviation (RMSD) of each system in angstrom [55].

Binding free energy calculation

The binding free energies of protein–ligand complexes were estimated by the Molecular Mechanics/Generalized Born Surface Area (MM/GBSA) approach [56], [57], [58] using the MM/PB (GB) SA script in AMBER 20. For this purpose, 1000 snapshots were selected from the last 10 ns trajectory from the production simulation.

Eq. 1 was used to estimate complexes' binding free energies (G bind) [59].

$$\mathrm{\Delta }\mathrm{Gbind}=\mathrm{\Delta }\mathrm{GR}+\mathrm{L}\mathrm{–}\left(\mathrm{\Delta }\mathrm{GR}+\mathrm{\Delta }\mathrm{GL}\right)$$ (1)

GR + L = energy of the protein–ligand complex, GR = energy of free protein, and GL = ligand’s energy. Each G (free energy) term in the above eq. is estimated using Eq. 2 in the MM/GBSA, and MM/PBSA approaches.

$$\mathrm{G}=\mathrm{Ebond}+\mathrm{Evdw}+\mathrm{Eelec}+\mathrm{GPB}+\mathrm{GSA}\mathrm{–}\mathrm{TSS}$$ (2)

E-bond calculates the bond energy, which includes bond, angle, and dihedral energies; Evdw denotes van der Waals contribution, and Eelec denotes electrostatic energy. GPB and GSA energy of solvation indicate polar and non-polar contributions, respectively; T represents the absolute temperatures, and SS represents solute entropy.

The force field, type of MD simulation, specificity of the protein–inhibitor complex, inner dielectric constant, particle charges utilized for the small molecule, and pose chosen from the docking procedures all influence the free energy calculation model MM/PB(GB)SA. Small inhibitor-attached RdRp systems were subjected to MM/PB(GB)SA calculations, including one reference and eight selected protein–inhibitor complexes. The topology of the systems was optimized by a solvent probe with a radius of 2 Å, and the raddi mbondi2. The results of all energy calculations were given in kcal/mol. To examine the protein structural alterations in free and inhibitor bind states, whole electrostatics energies, including the van der Waals energy, electrostatic energy, bond energy, angle energy, and dihedral energy, were computed for the system from the unit cell parameters. The direct sum of the electrostatic energies was used to determine the system's overall energy.

Protein structural compactness and fluctuation

Each system's Root Mean Square Variations (RMSF) were computed using the below Eq. 3 [55] to detect the atomic positional fluctuations of the given residues in the protein. The RMSF indicates how flexible each protein residue is.

$$\mathit{RMSF}\left(i\right)=\sqrt{⟨{\left(x_i-\left(\left.⟨x_i\right)\right.⟩\right))}^2⟩}$$ (3)

The averages of atom positions over total simulation time frames are denoted by X in Eq. 3.

The radius of gyration (Rg) was used to compute the atomic motion in the protein structure from their shared center of gravity, applying Eq. 4.

$$\mathit{Rg}=\sqrt{\frac{1}{N}\sum _N^{i=0}{\left(r_i-r_m\right)}^2}$$ (4)

In equation6, "ri" and "r_m" represent atom position and mean position in protein, respectively. The Altona and Sundaralingam [40] technique was utilized to compute the Five-membered ring pucker. The standard deviations and averages of the Rgs were computed. Using the periodic torsions, appropriate cyclic averages were computed.

Solvent accessible surface area (SASA) analysis was carried out to investigate the surface properties of proteins such as polar, non-polar, exposed, and buried residues. The atom type and bond information in the MD trajectory was used to determine the area of the protein exposed to the solvent in Ų. The Linear Combination of Pairwise Overlaps (LCPO) [49]algorithm is used by the CPPTRAJ package of AMBER 20 to identify the protein solvent-exposed residues involved in the protein's synthesis and stability.The alteration that occurred in the structure was examined using the 100 ns trajectory data of each sytstem, where the first frame of the simulation trajectory was selected as a reference to the others while calculating the RMSD of each residue and was populated in 30 bins with a range of 1–6 Å.

Protein dominant motions triggered by inhibitor binding

Using the CPPTRAJ module of AMBER 20, principal component analysis (PCA) was performed to extract the protein’s functional and slow motions [55], [60]. The covariance matrix was computed first to perform PCA. In matrix C, element Cij is given as in Eq. 5:

$$\mathit{Cij}=\left.⟨\left(\mathit{xi}\right)-\right)\left(〈\mathit{xi}〉\right)\left.⟨\left(\mathit{xi}\right)-\right)\left(\left(〈\mathit{xi}〉\right)\right.⟩$$ (5)

Where xi and xj represent the cartesian coordinate of Cα atoms at the number i and j; 〈xi〉 and 〈xj〉 are the ith and jth atom's average coordinates throughout the ensemble. The 3D positional coordinates of each system trajectory were used on a time scale with ten motion modes to calculate the coordinate covariance matrix. The eigenvalue and eigenvector for the covariance matrix C estimated over 10000 snapshots from the trajectory of each system were solved and diagonalized to obtain the principal components (PCs). The directions of motions are represented by the eigenvectors, PCs, while the eigenvalues show their magnitudes. Each eigenvector's fraction was expressed as a percentage. The first two principal components, PC1 and PC2, were plotted to investigate their motions.

Physical motions of atoms with inhibitor attachment

Dynamic cross-correlation map (DCCM) was used to uncover the correlation across all Cα atoms in all systems of SARS-CoV-2 RdRp. The cross-correlation coefficient, Sij, between two residues (i and j) during the whole MD trajectory was calculated as in Eq. 6:

$$\mathrm{S}ij=\frac{\mathrm{〈}\mathrm{\Delta }\mathrm{ri}.\mathrm{\Delta }\mathrm{rj}\mathrm{〉}}{(\mathrm{〈}\mathrm{\Delta }\mathrm{ri}2\mathrm{〉}\mathrm{〈}\mathrm{\Delta }\mathrm{rj}2\mathrm{〉})2}$$ (6)

Where Δri or Δrj are the displacement vector and calculated as ith and jth atom's instantaneous position minus their mean positions. The same direction/positive correlated motions between ith and jth atoms are indicated when Sij> 0. At the same time, opposite direction /negative correlated motions are shown when Sij< 0. DCCM analysis was carried out using the CPPTRAJ module [55].

Gibbs free energy distribution

The conformational free energy values of the complexes in both stable and transient states were examined. The CPPTRAJ module of AMBER 20 was used to analyze the systems' Free Energy Landscape (FEL) [55]. Using PC1 and PC2 principle components, the trajectories data were distributed into 100 bins. PC1 and PC2 values show the highest variation reported. Bins with no population were maintained at a population size of 0.5 as an artificial barrier during the free energy calculations. Free energies were calculated at 300 K and reported in kcal/mol.

Hydrogen bonds analysis

The CPPTRAJ module of the AMBER 20 [61] was used to compute all the H-bonds involving amino acid residues and ligand atoms. The number of H-bonds that can form between a given ligand atom and a protein residue is an average of all the calculations made for a 100 ns simulation run. The cut-off distance and angle between the donor–hydrogen–acceptor atoms were fixed at 3.5 Å and 120º, respectively.

Data analysis

MOE 2020.0901 [62], VMD [63], Pymol [64], and Blender [65] were used to create illustrations. The lowest energy structure ensembles of all the systems were extracted by Origin pro [66].

Results

The 3D structure of SARS-CoV-2 RdRp in the apo and in suramin attached state is presented in Fig. 1 . A highly potent non-nucleoside inhibitor ‘suramin’(IC_{50 =} 0.26 µM) [67] is attached at two different sites of RdRp, including RNA-template strand and RNA-primer strand binding sites. At RNA-template strand binding site, suramin occupies the cavity close to the active catalytic site where the sulfonate groups of suramin mediate H-bonds with the side chains of Ala550, Lys551, Arg555, and Arg836. At the same time, Asp865 provides a H-bond to the amino group of suramin. The amide bond in ligand also form a H-bond with the Arg555. Furthermore, Ser549 and Leu862 stabilize the ligand through van der Waals interactions.

Fig. 1.

The 3D structure of SARS-CoV-2 RdRp is presented in a complex with suramin molecules (PDB ID: 7D4F). The suramin binding is shown at the RNA primer strand binding site (A) and RNA template binding site (B).

At RNA-primer strand binding site, the other molecule of suramin is fitted in the conserved motifs, namely motif B (N-terminus) and motif G (C-terminus), and follow a similar binding pattern as in RNA-template strand binding site. The side chains of Asn496, Asn497, Lys500, and Arg569 establish H-bonds with two sulfonate groups of suramin, while Asn497 form H-bonds with two sulfonates of this molecule. Apart from interacting with the solvent, another sulfonate group creat an H-bond with the Asn496. In addition, Lys577 and Gly590 form H-bonds with the naphthalene ring and the amide bond. Furthermore, Ser549 and Arg553 also provide van der Waals interactions with the ligand.

Per-residue free energy decomposition analysis

The SARS-CoV-2 RdRp in complex with suramin (bound at RNA-primer strand binding site) was further subjected to MM/GBSA free energy decomposition analysis to determine the contribution of each residue to the total energy of the complex. Significant contributions to the total binding free energy of the complex were −6.02 kcal/mol, −5.60 kcal/mol, −3.43 kcal/mol, and −2.58 kcal/mol, respectively, from residues Lys553, Arg557, Lys623, and Arg555. The van der Waals interactions energy contributed more to the overall binding energy of these residues with the ligand than the electrostatic component ( Table 1). Whereas Cys815 (–0.60 kcal/mol), Ser551 (–0.595 kcal/mol), Ser816 (–0.758 kcal/mol), and Arg838 (–0.603 kcal/mol) also made moderate contributions to the total energy.

Table 1.

Molecular Mechanics/Generalized Born Surface Area (MM/GBSA) per-residue energy decomposition analysis of SARS-CoV-2 RdRp–suramin complex. ΔG Total, total binding free energy.

Residues	MM/GBSA Per-Residue Energy Decomposition Analysis (kcal/mol)
	Van der Waals	Electrostatics	Polar Solvation	Non-Polar Solvation	ΔG Total
LYS547	–0.25 ± 0.0034	–53.98 ± 0.164	54.35 ± 0.1639	–0.082 ± 0.0027	0.035 ± 0.0088
SER551	–0.45 ± 0.0088	–4.48 ± 0.0467	4.43 ± 0.0324	–0.09 ± 0.0012	–0.59 ± 0.0215
LYS553	–3.09 ± 0.0215	–193.79 ± 0.1605	94.65 ± 0.1465	–0.68 ± 0.0020	–6.02 ± 0.0351
ARG555	–1.248 ± 0.017	–62.39 ± 0.200	61.23 ± 0.17108	–0.18 ± 0.00159	–2.58 ± 0.0406
ARG557	–0.93 ± 0.0222	–82.39 ± 0.123	78.10 ± 0.1041	–0.37 ± 0.0018	–5.60 ± 0.0326
LYS623	–0.33 ± 0.0199	–86.59 ± 0.4146	83.78 ± 0.3652	–0.29 ± 0.0034	–3.43 ± 0.0569
LYS800	–0.11 ± 0.0033	–52.63 ± 0.1959	52.44 ± 0.1892	–0.04 ± 0.0015	–0.36 ± 0.01277
CYS815	–0.65 ± 0.0134	–1.03 ± 0.0350	1.21 ± 0.024	–0.13 ± 0.0023	–0.60 ± 0.0150
SER816	–1.86 ± 0.0156	0.17 ± 0.0738	1.39 ± 0.0730	–0.46 ± 0.0029	–0.75 ± 0.0171
GLN817	–0.49 ± 0.0116	–0.78 ± 0.04182	0.98 ± 0.0375	–0.05 ± 0.0024	–0.35 ± 0.0159
ARG838	–1.17 ± 0.0203	–24.37 ± 0.13152	24.40 ± 0.1364	–0.28 ± 0.0023	–0.60 ± 0.0182

Attribution of critical residues of RdRp

H-bond analysis of the RdRp-suramin complex unveiled significant occupancy time of Arg838 (50%), Ser816 (∼41%), Arg867 (34%), Lys553 (∼29%), Lys547 (∼28%), Arg557 (∼28%), Ser551 (12%), Lys623 (∼6%), and Cys815 (∼3%) with suramin (Table S1). The rest of the H-bonds were short-lived and didn't contribute significantly to protein–ligand stability. Based on the higher free energy decomposition value and H-bond occupancy, five residues (Lys553, Arg557, Lys623, Cys815, and Ser816) were found critical for suramin binding at RdRp binding site. Therefore, it is expected that those inhibitors that interact with these specific residues can disrupt RNA binding to RdRp for replication, thus, halting the viral replication.

Ligands steric and electronic features

The generated pharmacophore model revealed one steric and six electronic features on ligand atoms that strongly interact with the critical residues (Arg553, Arg557, Lys623, Cys815, and Ser816). Two oxygen atoms of ligand were taken as hydrogen bond acceptors (HBA). In comparison, three interacting nitrogen atoms were selected as hydrogen bond donors (HBD) ( Fig. 2), and one aromatic (ARO) feature of the ligand was selected, while the protein's pocket volume was omitted. The pharmacophore model was verified with the help of the active compounds and their associated decoy sets. In the first stage of the pharmacophore model-based screening procedure, a ROC curve was produced to evaluate the model's capacity to distinguish between active and inactive compounds. The resulting ROC curve showed a combination of specificity and sensitivity. EF and area under the curve (AUC) values demonstrated the curve's quality. AUC values between 0 and 0.5 indicate a possibility of discriminating; values between 0.51 and 0.70 indicate good performance, and values between 0.71 and 0.80 indicate an excellent model. The ROC curve presented in Fig. S1 showed good results, with an AUC value of 0.958. The pharmacophore model was validated on the test database, which comprised 36 active (Table S2) and 360 inactive compounds. The pharmacophore model efficiently picked all the active molecules in the test set, indicating the accuracy of the generated model in differentiating between active and inactive compounds. Fig. S1 presents a good ROC curve with an AUC of 0.958, which signifies the pharmacophore quality.

Fig. 2.

The complex-based Pharmacophore model is shown. The Pharmacophore features were constructed on the ligand atoms that bind with the most important residues. H-bond acceptor and donors are shown in the green and purple spheres, respectively, and the aromatic feature is presented in the orange sphere. ACC, acceptor; ARO, aromatic; DON, donor.

Pharmacophore-based filtration of ZINC20 database

A set of ∼690 million small molecules from the ZINC20 drug-like database was virtually filtered through the electrostatic/steric features constructed on the ligand atoms. Subsequently, 3542 compounds were matched with the constructed pharmacophore model.

Structure-based virtual screening

The hits retrieved from the ZINC20 database (3542) and small molecules from DrugBank [segregated into FDA-approved (2701), drugs in investigational phase (4094), and experimental group (6310) categories] were further filtered through structure-based virtual screening protocol. Prior to SBVS of the huge database, the docking protocol was validated by re-docking of co-crystallized conformation of suramin at the RNA-primer strand binding site. The suramin molecule was re-docked perfectly with an RMSD value of 0.97 Å and a docking score of −6.66 kcal/mol (Fig. S2). Subsequently, a total of 15,240 compounds were docked at the RNA-primer strand binding site. Later, the docked database was sorted according to the docking score, and 14 compounds from ZINC20 and 10 molecules from DrugBank were selected based on good docking score and interaction of ligands with the crucial residues of RdRp. The docking results are given in Table S3–S4 .

The optimal docked poses of selected hits are shown in Fig. 3. ZINC98208626, an experimental stage drug, showed H-bond interactions with the side-chains of Arg557, Lys623, Glu813, Cys815, and backbone atoms of Cys801. In addition, ligand-mediated π-alkyl interactions with Ser761. Similarly, nitrogen atoms and OH-moiety of ZINC28467879, an investigational phase inhibitor, mediated hydrogen bonding with Arg555 and Glu813. Besides, oxygen and sulfur atom of ZINC28467879 mediated H-bonds, respectively, with Arg557 (2x), Ser816, and Lys623. While Arg555 formed π-cation interaction with the benzene ring of ligand. The best-docked view of ZINC285540154 revealed several H-bonds; its phosphate moiety formed multiple H-bonds with Lys553, Arg555, Arg557, and Lys623 and the carbonyl group of ligand was H-bonded with Ser816. In addition, OH-moieties and nitrogen atoms of the ligand showed H-bonds with Asp620, Tyr621, and Pro622, and Asp763. Also, Tyr460 and Lys623 provided π-cation interactions with the inhibitor. The stability of ZINC611516532 in the binding pocket of RdRp was maintained by multiple H-bonds with Asp620. The carbonyl oxygen and OH-moiety of the ligand mediated H-bond with Lys623 and Glu813, while the terminal nitrogen of ligand was bound with Ser551 via H-bond. Furthermore, Arg557 made π-alkyl interaction with the ligand benzene ring.

Fig. 3.

The docked poses of selected inhibitors of RdRp are shown. The binding interactions of ligands with protein are presented in 2D form (A) and 3D format (B).

The ZINC739681614 established H-bond interactions with the side chains of Asp454, Lys547, and Arg557. The stability of ZINC1166211307 in the protein's binding pocket was mainly contributed by hydrogen bonding with Asp620, Lys623, and Asp762. Also, Asp620 demonstrated π-alkyl interaction with the ligand. The OH-group of ZINC1398350200 formed H-bond with the side chain of Trp802 and the backbone of Glu813, and the carbonyl oxygen of the ligand mediated H-bond with Arg557 and Lys623. In addition, Asp620 and Ser816 also mediated H-bond with the ligand. Moreover, Glu813 showed ionic interactions with the ligand. The best-docked orientation of ZINC1602963057 showed that the carbonyl and amino groups of this molecule mediated H-bond with the side chains of Lys553 and Glu813. The side chain of Lys547 mediated two H-bonds with the ligand nitrogen atoms. In addition, Arg557 showed ionic interactions and His818 showed π-alkyl interaction with the ligand.

Prediction of physicochemical, pharmacokinetic, and toxicity parameters

The physicochemical characteristics of selected hits revealed their molecular weight (MWT) range from 305 to 324 g/mol. All the compounds (except ZINC611516532) possess 6–10 rotatable bonds (RBs) in their structure. These compounds have 3–8 HBAs, and 2–4 HBDs and their topological polar surface area (TPSA) range from 65.46 to 147.13 Ų, while the molar refractivity (MR) of compounds vary from 70.99 to 96.86. These findings were compared to suramin's physicochemical characteristics, which possess MWT of 1297.28 g/mol, TPSA of 534.03 Ų, RBs = 22, MR = 306.12, HBA= 23, and HBD = 12.

The drug-likeness of the selected hits was assessed by Lipinski's rule of five [68], Veber's rule [69], and Muegge's rule [70]. Overall, all the compounds obeyed these rules; however, ZINC2098490495 violated one of Veber's rule (TPSA>140) and two rules of Muegge's rule (XLOGP3 <–2, TPSA>150). Similarly, ZINC1602963057 also violated one Veber's rule (TPSA>140). In comparison, suramin showed three violations of Lipinski's rule of five (MW>500, N or O>10, NH or OH>5), two violations of Veber's rule (Rotors>10, TPSA>140), and six violations of Muegge's rule (MW>600, TPSA>150, No. of rings>7, Rotors>15, HBA>10, HBD>5). These filters (Lipinski, Veber, and Muegge's rules) are used as critical guiding principles in computational drug designing methods that offer primary generalization about the approved form of a drug [71]. The Log S scale predicted the solubility level of 14 selected compounds, which indicates four compounds are very soluble, one compound is highly soluble, and the rest are soluble in an aqueous solvent. The predicted partition coefficient (LogP octanol/water) of the chosen hits are in the range of + 0.10–2.43, indicating their solubility in a hydrophobic media. In contrast, Log Po/w of suramin is 2.59 and predicted soluble in the aqueous media.

The predicted pharmacokinetic characteristics of chosen inhibitors also aided in selecting a more suitable drug candidates. According to SwissADME prediction [53], 12 compounds are estimated to have high human gastrointestinal absorption (except ZINC1602963057 and ZINC2098490495). Similarly, all the compounds (except ZINC1093006744 and ZINC1092867333) demonstrated no blood-brain barrier penetration, and most of the compounds are not expected to inhibit cytochrome p450 enzymes (CYP3A4, CYP2D6, CYP2C9, CYP2C19, and CYP1A2) and except ZINC1317807769, ZINC1717686152, ZINC970738350, all compounds displayed non-substrate-like properties for P-glycoprotein (P-gp). The ligands' skin permeability (LogKp) is in the range of –6.15 to −11.34 cm/s, indicating these compounds are not permeable through the skin.

The bioavailability score of ZINC1602963057 is 0.56, while the rest of the compounds possess a 0.56 score indicating their good bioavailability. Besides, none of the compounds showed PAINS alert. The estimated synthetic accessibility of selected compounds ranges from 2.34 to 3.43, implying that these compounds are synthesizable. In addition, compounds ZINC739681614, ZINC738028786, ZINC1166211307, ZINC2098490495, ZINC1398350200, and ZINC81523997 showed one violation (Rotors>7), while the rest of the compounds passed the lead-likeness criteria. In contrast, the medicinal properties of suramin revealed that the selected compounds have comparable scores. Bioavailability and synthetic accessibility scores for suramin are 0.11 and 6.41, respectively. Also, suramin did not show any PAINS alert. However, it showed two violations (MWT>350, Rotors>7) of lead likeness. The results are tabulated in Tables S5–S7.

Using the ProTox-II server [54], the toxicity and adverse effects of the selected ZINC20 compounds were determined. The median lethal dose (LD50) of our hits are in range of 593 mg/kg (class 3) to 3500 mg/kg (class 5). Only ZINC739681614 indicated immunotoxicity and cytotoxicity, and ZINC1602963057 and ZINC970738350 showed carcinogenic and immunotoxic properties, while the rest of the compounds were found non-toxic and safe. The toxicity reports of the selected compounds are tabulated in Table S8.

Stability of the simulated complexes

It is vital to understand the inhibitory mechanism of SARS-CoV-2 RdRp by identifying the protein's key structural characteristics. MD simulation of SARS-CoV-2 RdRp was performed in the apo and inhibited states, using the suramin molecule as a reference and the most suitable selected inhibitors. A comparative analysis in both (apo- and ligand-bound) states of RdRp was performed to investigate structural stability during MD simulation. After 100 ns, the RMSD of the output trajectories was determined.

The apo-RdRp (PDB ID: 7BV1) experienced a free moment, and the RMSD was increased to 30 ns ( Fig. 4). The protein fluctuated between 2.2 Å and 3.4 Å until 85 ns, suggesting its instability. However, the apo-state was stabilized at the end of the simulation. The RMSD of the suramin-bound complex (7D4F) was raised initially until 20 ns. However, the system got stabilized at the end of the simulation with an average value of 2.5 Å. The ZINC98208626–RdRp complex reflected steady behavior during 100 ns simulation run (except for the initial RMSD shift till 20 ns).

Fig. 4.

Root mean square deviation (RMSD) of apo-RdRp (PDB ID: 7BV1), suramin-bound RdRp (PDB ID: 7D4F), and RdRp in complex with chosen eight inhibitors molecules. ns, nanosecond.

The RMSD of ZINC28467879–RdRp complex steadily increased till 35 ns, then remained stable till 70 ns and steadily rose afterward before it finally reached equilibrium at the end, indicating instability of the protein during simulation. Similarly, the ZINC85540154–RdRp complex system showed considerable increase in first 20 ns, and continued to steadily rise until the end, with an average RMSD value of 2 Å, indicating unstable behavior of protein during the simulation. The ZINC611516532–RdRp complex showed an enhanced stability shift from 0 to 25 ns to 40–70 ns, impacting the protein's stability. After 70 ns, the protein reached equilibrium. Although RMSD increased in 0–5 ns and 20–60 ns intervals, the ZINC1602963057–RdRp complex was stabilized after the 65 ns. The ZINC1398350200–RdRp complex showed a drastic stability change during 0–15 ns, 20–30 ns, and 55–60 ns intervals. During the simulation, the RdRp–ZINC1166211307 complex also displayed signs of instability. In the 100 ns simulation timeline, this complex showed a steady increase in RMSD. The RMSD value of the ZINC739681614–complex increased from 0 to 15 ns to 20–35 ns intervals; however it reached equilibrium after 35 ns. The results showed that no inhibitor-bound complex fluctuated more than the free state RdRp.

Binding free energy calculations

The binding free energy was computed by MM/GBSA method to determine the binding affinity between SARS-CoV-2 RdRp–inhibitor complexes shown in Table 2. In addition, different free energy components that contribute to binding were estimated. One thousand frames from the 100 ns MD trajectory of the binding free energy calculations were used to calculate binding free energy. The MM/GBSA analysis revealed energy differences among suramin and eight small molecules. Suramin exhibited van der Waals free energy (ΔE_VDW −29.61 kcal/mol), electrostatic free energy (ΔE_elec 40.38 kcal/mol), gas-phase free energy (ΔG_gas 10.77 kcal/mol), solvent-accessible surface area free energy (SASA −4.94 kcal/mol) with a total binding free energy (ΔG Total) of −9.94 kcal/mol.

Table 2.

Molecular Mechanics/Generalized Born Surface Area (MM/GBSA) binding free energy calculation and the individual energy components of the RdRp in complex with ligands.

Complex Name	MM/GBSA (kcal/mol)
	ΔEvdW	ΔEelec	ΔGgas	SASA	ΔG Total
RdRp–suramin	–29.61 ± 3.55	40.38 ± 26.33	10.77 ± 24.02	–4.94 ± 0.42	–9.94 ± 5.34
RdRp–ZINC285540154	–53.84 ± 2.89	–	–53.84 ± 2.89	–7.60 ± 0.30	–37.74 ± 2.90
RdRp–ZINC28467879	–42.84 ± 4.22	–	–42.84 ± 4.22	–5.58 ± 0.40	–26.36 ± 3.76
RdRp–ZINC611516532	–25.02 ± 2.58	–193.2 ± 12.92	–218.25 ± 12.35	–3.55 ± 0.31	–22.03 ± 4.03
RdRp–ZINC1398350200	–28.22 ± 2.93	–244.77 ± 16.70	–273.00 ± 16.88	–4.95 ± 0.27	–20.82 ± 2.90
RdRp–ZINC739681614	–25.15 ± 3.88	–14.65 ± 13.02	–39.80 ± 13.15	–3.43 ± 0.78	–16.04 ± 4.72
RdRp–ZINC1602963057	–27.82 ± 2.44	–53.40 ± 9.58	–81.22 ± 8.83	–4.035 ± 0.21	–12.87 ± 2.74
RdRp–ZINC1166211307	–13.82 ± 7.47	–143.16 ± 11.29	–156.99 ± 13.39	–1.76 ± 0.95	–7.83 ± 5.54
RdRp–ZINC98208626	–43.32 ± 4.11	–145.05 ± 16.99	–188.37 ± 16.64	–6.55 ± 0.21	–6.50 ± 7.6

ΔE_vdw, van der Waals free energy: ΔE_elec, electrostatic free energy: ΔG_gas, gas-phase free energy: SASA: solvent accessible surface area free energy: ΔG Total, total binding free energy.

Among all the compounds, ZINC285540154 exhibited the highest free energy (i.e., −37.74 kcal/mol) in MM/GBSA calculation. The higher binding free energy of RdRp–ZINC285540154 complex than the RdRp–suramin complex was attributed to its higher ΔE_VDW (–53.84 kcal/mol), SASA (–7.60 kcal/mol), and ΔG_gas free energy (–53.84 kcal/mol). Afterward, ZINC28467879 also demonstrated higher free energy (i.e. −26.36 kcal/mol) than the reference complex due to its larger ΔE_VDW (–42.84 kcal/mol), SASA (–5.58 kcal/mol), and ΔG_gas (–42.84 kcal/mol) free energy components. The ΔE_elec and ΔG_gas free energy considerably contributed to the high binding energy of ZINC611516532 (–22.03 kcal/mol). Moreover, ZINC1398350200 also exhibited higher binding energy (–20.82 kcal/mol) than the reference complex, owing to its high ΔG_gas (–273 kcal/mol) and ΔE_elec (–244.77 kcal/mol) free energy components among all complexes. Furthermore, the binding free energy of RdRp–ZINC739681614 complex (–16.04 kcal/mol) was also found greater than the reference complex because of its higher ΔG_gas free energy (–39.80 kcal/mol) and ΔE_VDW free energy (–25.15 kcal/mol). The estimated binding free energy of the ZINC1602963057 (–12.87 kcal/mol) was also greater than the reference complex, which was dominated by ΔG_gas free energy. These compounds showed a high affinity for the SARS-CoV-2 RdRp binding site, suggesting these molecules as effective non-nucleoside inhibitors of the novel virus. The binding free energies of ZINC1166211307 and ZINC98208626 were −7.83 kcal/mol and −6.50 kcal/mol, respectively. These compounds also bind at the RdRp primer strand binding site with appropriate affinity mainly dominated by ΔG_gas free energy component.

Electrostatic energies of complexes

The structural fluctuations can affect the complexes' total electrostatic free energies; therefore, changes in each complex's total free energy were also computed (Fig. S3). The overall energy of RdRp in the apo state (7BV1) drastically increased from –16200 to −17100 kcal/mol during the simulation timescale. In contrast, the suramin-bound reference complex experienced a decrease in total energy to −15450 kcal/mol. Following the pattern of reference inhibited state, a decline in total energy was noted for ZINC98208626 (–17300 kcal/mol), ZINC28467879 (–16200 kcal/mol), and ZINC85540154 (–15900 kcal/mol) bound RdRp. Apart from the initial drastic energy variation, the total energy of the RdRp–ZINC611516532 complex was retained with slight changes. Similarly, the whole energy of RdRp was decreased after binding with ZINC739681614 from –17140 to −16300 kcal/mol up to 80 ns, followed by a slight variation in the overall energy till the end. During 10–30 ns interval, a decline in the total energy (–16000 kcal/mol) was observed in RdRp–ZINC1166211307; however, during the 30–60 ns interval, the whole energy of this complex was gradually deteriorated and sustained again with a slight increase towards the end. At the same time, the overall energies of RdRp–ZINC1398350200 and RdRp–ZINC1602963057 complexes decreased from –17420 to −16000 kcal/mol and − 17200 to −16100 kcal/mol, respectively, with slight alteration during the simulation timescale.

Effect on compactness of protein structure and fluctuation

Root mean square fluctuation

RMSF was computed to study the fluctuation of individual residues of RdRp in 7BV1, 7D4F, and eight selected inhibitor-bound states. The ligand attachment caused alteration in the total protein flexibility of RdRp as compared to the apo-state, which exhibited the highest variation with a maximum RMSF value of 7.8 Å. In the apo-RdRp, residues 20–50 (corresponds to β-sheets and loop region of N-terminal β-hairpin), 70–90 (α-helix and loop region), 225–245 (α- helix and loop region of the NiRAN domain), 465–490(α-helix and loop region of the finger domain), and 865–885 (α-helix of the thumb domain) showed a greater extent of RMSF fluctuation. Unlike ligand-bound complexes, residues 20–50 and 70–90 indicated higher fluctuations in the apo-RdRp.

RMSF fluctuation of the reference complex (7D4F) was lower than the apo-RdRp, while residues 400–440 (α-helixand loop region of the finger domain), 495–515 (α-helix and loop region of thefinger domain), and 600–615 (α-helix of the palm domain) demonstrated more movements than the apo-protein. The fluctuation pattern in the selected ligands-inhibited–RdRp was similar to the suramin-bound–RdRp. ZINC98208626, ZINC2846789, and ZINC85540154 showed RMSF value< 3 Å similar to the reference system, except for the α-helix of the Palm domain, which fluctuated less in these systems. Besides, ZINC85540154 indicated slightly more motion in the Thumb domain's α-helix and loop region than the enzyme apo-state. Similar to the ZINC85540154 bound RdRp, ZINC611516532, ZINC1602963057, ZINC1398350200, ZINC1166211307, and ZINC739681614 depicted lower amino acid flexibility. RMSF value of all ligand-bound systems were< 5 Å. Fig. 5 illustrates the RMSF fluctuation of selected compound systems compared to the apo-state (7BV1) and a reference [7], [4]F system.

Fig. 5.

Graphical presentation of residual fluctuations of the RdRp via Root mean square fluctuation (RMSF) in apo (PDB ID: 7BV1), suramin-bound–RdRp (PDB ID: 7D4F) and selected ligand-bound states. Å, Angstrom.

Radius of gyration

The time evolution of Rg of the complete MD trajectory was computed to validate further the protein compactness of apo and inhibited systems ( Fig. 6). The Rg of apo-state increased initially, with the maximum value reaching 32.6 Å. Besides, the average Rg value of the apo-state was around 32.1 Å during the 100 ns simulation run, indicating free motions in the proteins' domains and the unstable state of the protein. Alternatively, the reference system (7D4F) Rg fluctuation was lower than the apo system, with maximum Rg value reaching 31.6 Å and an average value of 31.2 Å throughout the simulation timescale. This indicates more tight packing of reference protein domains, thus confirming its stable position. Among the selected system, the Rg fluctuation of the ZINC611516532 system was highest, particularly in the 60–90 ns range, where Rg value varied from 32.1 Å to 32.5 Å. Besides, the average Rg value was 32.1 Å during the 100 ns simulation timescale for this compound, indicating an unstable behavior of the protein. The average Rg value for the ZINC28467879 system was 32.1 Å, while the average Rg value of 32 Å was estimated for ZINC1398350200 and ZINC739681614. At the same time, the average Rg value of ZINC98208626, ZINC85540154, ZINC1166211307, and ZINC1602963057 systems were less than 32 Å, demonstrating a stable behaviour and lower motion flexibility of these inhibited protein systems throughout the simulation timescale.

Fig. 6.

The radius of gyration graphs of apo (PDB ID: 7BV1), suramin-bound (PDB ID: 7D4F), and eight selected inhibitors-bound SARS-CoV-2 RdRp. Colors are used to indicate each system. Å, Angstrom; ns, nanosecond.

Protein surface conformation

The SASA profile of each system was used to estimate the structural alteration in protein accessibility to solvent. The structural alteration which is caused by the attachment of small molecules affects SASA. The SASA of each complex with respect to time is displayed in Fig. S4 . In the apo-state (7BV1), SASA value increased from 42320 Ų to 45440 Ų during the first 40 ns of simulation, then gradually declined to 40600 Ų towards the end of the simulation.

However, the simulation continuously increased the SASA value of the suramin-bound RdRp complex from 40490Å2 to 44830Å2 during the simulation, which confirms that ligand binding can cause folding and unfolding alteration in the protein structure. The ZINC98208626–RdRp complex showed a drastic increase in SASA value in the first 20 ns up to 42340 Ų. Nevertheless, SASA underwent slight change till the end of the simulation. The ZINC28467879 attached RdRp maintained the SASA value with slight alterations, whereas ZINC85540154 RdRp complex reflected steadily raised SASA value (38700–42700 Ų) during the 100 ns simulation. Similarly, the SASA value of ZINC611516532-bound–RdRp was observed with a persistent increase during the simulation, including a sharp rise during the 70–90 ns interval with a maximum value of 43210 Ų.

In contrast, ZINC739681614-bound–RdRp sustained the SASA for most of the time. At the same time, the SASA value of ZINC1166211307 and ZINC1398350200 attached complexes steadily increased from 39600 Ų to 42700 Ų and 39480–42820 Ų, respectively. The SASA value of ZINC1602963057–RdRp complex was increased initially, then remained with minor alterations till the end of the simulation.

Structural variation in RMSD

RMS-histogram analysis was used to further study the structural alteration during the simulation time ( Fig. 7). The histogram distribution of the residues showed that the apo-RdRp lies between 2.5 Å and 3.5 Å with a smooth slope. In contrast, the reference complex (7D4F) moved towards the left side of the histogram between 1.5 Å and 2.8 Å with a distortion in the peak. Comparing the selected inhibitors complexes with the apo-RdRp, all the inhibited systems are reported in a left shift in the histogram during the normalization of the RMSD. Each inhibited system showed movements of distortion in the peak formation. These roughness confirms the protein folding changes during the simulation time because of the inhibitor attachment. All the inhibited systems’ histograms compared to apo showed the same behavior of the RMSD left shift in the histogram, and structural deviation reported in the peak curve confirms that the selected inhibitors have a significant effect on the protein structure while attaching to the active pocket.

Fig. 7.

RMS-histogram analysis of apo (PDB ID: 7BV1), suramin-bound (PDB ID: 7D4F), and in selected inhibitors bound states of RdRp during the 100 ns MD simulation timescale. RMSD, Root mean square deviation; Å, Angstrom.

Inhibitor-generated dominant motions in protein structure

PCA, based on Cα atoms was used to determine the structural change in the RdRp upon ligand binding.The total combined motions of the protein Cα atoms, represented by eigenvectors, were defined using this approach. The associated eigenvalues showed the amplitude of eigenvectors. Fig. S5 depicts the effect of inhibitors on protein structural dynamics. The first three eigenvectors reflect major dominating motions, whereas the rest depict localized fluctuations. The first three eigenvectors of apo-RdRp (7BV1) accounted for 69% variance, while the inhibited states of RdRp showed different patterns of motion. The first three eigenvectors of the suramin-bound–RdRp (7D4F) contributed 66% motion. Similarly, the first three eigenvectors of ZINC98208626, ZINC28467879, and ZINC285540154 bound RdRp complexes accounted for 58%, 65%, and 63% of the total variance, respectively. At the same time, the first three eigenvectors of ZINC611516532, ZINC739681614, and ZINC1398350200 complexes contributed 68%, 65%, and 63% variance in the entire motion, respectively. Besides, the first three eigenvectors of ZINC1166211307 and ZINC1602963057 were responsible for 67% of the variance in motion. The protein in apo and complex states differed because of structural changes caused by inhibitor binding.

The first eigenvector versus the second eigenvector was plotted to yield possible attributed movements of the RdRp in apo and ligand-bound states. The shifting of conformation with simulation time is shown by the continuous depiction of red to blue color. Each dot in Fig. 8 indicates a single trajectory frame that begins in red and ends in blue colors. Near convergence, the energetically unstable state (red) can be easily separated, resulting in a stable conformational state (blue). The histograms of both PC1 and PC2 are depicted opposite to the X-axis, and Y-axis in Fig. 8 to represent the variance in the data of each PC. Thus, several periodic leaps are required to transition between distinct conformations in the RdRp protein's inhibited states.

Fig. 8.

The principal component analysis of RdRp in apo-form (PDB ID: 7BV1), suramin-bound form (PDB ID: 7D4F) and selected inhibitor-bound states. The first principal component (PC1), and the second principal component (PC2) are plotted at X-axis and Y-axis, respectively.

The dominant motions caused by ligand binding with protein reported from the first 3 PCs are mentioned in Fig. 9. The spikes show the direction of motion, while the length shows the strength of the motion that occurred in a region or domain. In the apo-protein, the RdRp domain acquired inward motion in the active site region, while the NiRAN domain also adapted an inward motion towards the ADP attachment pocket. The active site and ADP attachment region in NiRAN domain acquire outward motion in suramin-inhibited–RdRp (7D4F) and in the selected inhibitors complexes. The outward motion indicates that these molecules force the protein to attain an open conformational state with structural and conformation changes in the protein.

Fig. 9.

Principal component analysis of SARS-CoV-2 RdRp in apo-form (PDB ID: 7BV1), suramin-inhibited form (PDB ID: 7D4F), and in complex with eight inhibitors. The arrows show the direction of motion for respective domain.

Physical motions of atoms via inhibitor attachment

DCCM was constructed to analyze the interacting atoms of the protein experiencing functional displacements. The apo-RdRp (7BV1) demonstrated higher positive correlation motions than the inhibited states throughout 100 ns. The apo-state was compared with the inhibitor-bound RdRp, where fewer atoms displayed positive correlation movements; the inhibited RdRp demonstrated a difference in correlated motion ( Fig. 10). High positive correlation motions were detected in suramin-inhibited RdRp (7D4F) between residues 830–860 (α-helix; Thumb domain) and 415–435 (α-helix, β-sheet, Loop; Finger domain), and between residues 560–580 (α-helix; Palm domain) and 480–500 (α-helix, Loop; Finger domain) and anti-correlation movements were also seen in residues 493–535 (α-helix, Loop; Finger domain) and 894–923 (α-helix, Loop; Thumb domain). Residues 830–865 (α-helix; Thumb domain) and 415–435 (α-helix, β-sheet, Loop; Finger domain) moved in the same direction in the ZINC98208626, ZINC28467879, and ZINC285540154 bound RdRp complexes. In these complexes, the localized positive correlation motion was also observed between residues 1–25 (α-helix, Loop) and 60–80 (α-helix, Loop). Whereas moderate anti-correlation motions were observed between residues 885–895 (α-helix; Thumb domain), 695–705 (α-helix; Palm domain), 880–900 (α-helix; Thumb domain), 745–765 (α-helix; Palm domain), 835–845 (α-helix; Thumb domain) and 445–455 (α-helix; Finger domain) in those complexes.

Fig. 10.

Dynamic cross-correlation matrix (DCCM) plot of apo-RdRp (PDB ID: 7BV1), suramin-inhibited (PDB ID: 7D4F), and RdRp in complex with eight selected inhibitors is shown. Green, yellow, and red represent positively correlated motions, whereas dark blue, light blue, and cyan represent negatively correlated motions. The gradient shows the decrease in correlated motion from one color to other.

Similarly, the ZINC611516532, ZINC739681614, ZINC1166211307, ZINC1398350200, and ZINC1602963057 bound RdRp demonstrated significant positive correlation motion in the 830–860 (α-helix; Thumb domain) and 415–435 (α-helix, β-sheet, Loop; Finger domain) regions, as well as localized motion in same-direction between residues 1–25 (α-helix, Loop) and 60–80. The residues 470–490 (α-helix; Finger domain) and 140–160 (α-helix; NiRAN domain) in the ZINC611516532–RdRp complex also displayed motion in the same direction, while residues 740–770 (α-helix, Loop; Palm domain) and 60–85 (α-helix, Loop) regions displayed movement in opposite direction while in ZINC1166211307–RdRp, residues 780–790 (α-helix; Palm domain) and 160–170 (Loop; NiRAN domain) showed a significant positive correlated motion. The ZINC1398350200–RdRp revealed a positive correlation between 505 and 535 (α-helix; Finger domain) and 365–375 (α-helix; Finger domain). On the other hand, the ZINC739681614–RdRp complex acquired moderate negative correlated motions between residues 490–540 (α-helix, Loop; Finger domain) and 845–865 (α-helix, Loop; Thumb domain).

Gibbs free energy distribution

FEL analysis was used to depict the transition stages of each studied system. The first two eigenvectors of the apo-RdRP and inhibited-RdRp were used to generate the trajectories' FEL to examine the evolution from initial positions to metastable states ( Fig. 11). The low energy states in each complex were presented to understand the structural alterations in RdRp after binding with inhibitor. The Gibbs free energy distribution, depths in the confirmational state, and population show the structural variability in apo-RdRp (7BV1) and inhibited systems. The high energy state in FEL plot is shown in red color, while intermediate energy levels are presented in yellow color, and stable energy states are colored blue. Despite stabilizing all energy barriers, apo-RdRp showed three energy barriers with no additional energy distribution constraints. Because the free state of RdRp stayed at one energy level for most of the simulation time, therefore, a considerable difference in the energy profile was observed in the inhibited complexes. In the inhibitor–RdRp complexes, red color zone was more prominent than in apo-RdRp, which indicates that RdRp was unstable due to inhibitor binding. The maximum transitions are reflected in suramin-bound–RdRp (7D4F), illustrating the influence of suramin-binding on RdRp confirmation. Whereas ZINC98208626 and ZINC28467879 bound RdRp complexes showed a great distribution of energy barriers, and these complexes were retained in the high energy state for most of the simulation time.

Fig. 11.

Free energy landscapes (FELs) of RdRp are shown in apo (PDB ID: 7BV1), suramin-bound (PDB ID: 7D4F) and selected inhibitors-bound states. High (shown in red color), intermediate (yellow and green), and low/stable energy (light to dark blue color) levels are shown in graph. PC1, first principal component; PC2, second principal component.

On the other hand, the ZINC85540154–RdRp complex dominantly acquired intermediate (yellow) and high energy states (red). The energy barriers in ZINC611516532, ZINC1602963057, ZINC1398350200, ZINC1166211307, and ZINC739681614 RdRp complexes revealed high Gibbs free energy distribution, indicating the high energy state of RdRp.

Hydrogen bonds analysis

Inter-protein H-bonds were examined to confirm the structural variability and conformational changes in RdRp upon inhibitors binding during the 100 ns simulation time. The apo-RdRp (7BV1) experienced a slight decrease in the number of H-bonds until halfway of the simulation; however, it retained such bonds constantly. Overall, the average count of inter-protein H-bonds held between 410 and 430 throughout the simulation. Compared to apo-RdRp, inter-protein H-bonds were slightly higher (∼430–450) in suramin-inhibited–RdRp (7D4F) with small periodic jumps. Among all the selected inhibitors, ZINC85540154 showed highest decline in H-bonds from 485 to 445, while ZINC611516532, ZINC1398350200, and ZINC1166211307 also reflected a decline in the number of H-bonds from 465 to 435. Moreover, the number of inter-protein H-bonds was constant in ZINC1602963057, ZINC28467879, ZINC98208626, and ZINC739681614 (total count 440–460).

Furthermore, backbone H-bonds are also crucial for the structural stability of the protein; thus, backbone H-bonds were also estimated. The selected inhibitor-bound–RdRp demonstrated a slightly more steady number of backbone H-bonds (220−230) which suggest that these compounds stabilized RdRp more than the 7BV1 and 7D4F (Fig. S6). The apo-RdRp maintained backbone H-bonds between 200 and 210, whereas the suramin-inhibited–RdRp showed constant backbone H-bonds between 210 and 220. The differences and shifts in patterns between the apo and inhibited states suggest structural alterations in RdRp due to the binding of small molecule inhibitors.The analysis of protein–solvent interaction is vital to demonstrate protein structural stability during MD simulation. The number of protein-solvent H-bond interaction was increased during MD (Fig. S7). The number of protein–solvent H-bonds in the apo-RdRp increased drastically from 1300 to 1370 in the first 20 ns; then, the total count was 1370 on average, however; in RdRp–suramin complex (7D4F), the number of protein solvent H-bonds raised with precise periodic jumps from 1320 to 1400. The number of protein solvent H-bonds in the inhibited systems varied from the apo-RdRp and 7D4F, with numerous periodic jumps. The compounds ZINC98208626, ZINC85540154, ZINC1166211307, ZINC1398350200 showed a significant increase in protein solvent H-bonds with different periodic jumps from 1320 to 1430, 1240–1380, 1300–1390, and 1290–1400 respectively. While ZINC28467879 (1305–1375), ZINC611516532 (1310–1375), ZINC739681614 (1295–1360), and ZINC1602963057 (1300–1350) complexes showed a smooth increase in protein–solvent H-bonds without precise periodic jumps throughout the simulation.

The strong binding of inhibitors at the binding site of protein is crucial for the drug to perform its inhibitory action. H-bond analysis of selected small molecules was performed to investigate the most important residues involved in protein–ligand interaction ( Fig. 12). The suramin-bound RdRp (7D4F) demonstrated stable H-bonds with Lys547 Ser551, Lys553, Arg555 Lys623, Cys815, Ser816, Ser863, and Asp867. In contrast to 7D4F, ZINC611516532 formed stable H-bond contacts with Asp619, Lys799, Glu812, and Ser815 of RdRp. During the simulation time, the ZINC739681614 established H-bonds with Asp453, Lys546, Ser550, Arg554, Arg556, Lys622, and Ser815, while ZINC98208626 mediated H-bonds with Asp619, Ser760, Asp762, Ala798, Lys799, Trp801, and Glu812 for consideration duration of simulation. Two molecules, ZINC28467879 and ZINC85540154, did not mediate stable H-bond with the residues during the 100 ns simulation time. However, these small molecules encountered several protein residues: the former with Ser550, Ala763, and Trp801, while the latter with Tyr547, Ala548, Arg554, and Arg556. ZINC1166211307 formed hydrogen bonding with Lys552, Asp619, Lys622, Asp762, and Glu812 during simulation. Similarly, ZINC1398350200 displayed hydrogen bonding for significant period of simulation time with several residues including Lys546, Asp624, Asp761, Asp762, Trp801, and Glu812. Moreover, strong H-bond interactions were observed between ZINC1602963057 and Lys546, Ser550, Lys552, Arg554, Glu812, Ser815, and Arg837of RdRp throughout the simulation timescale.

Fig. 12.

Hydrogen bond occupancy of small molecule inhibitors (SARS-CoV-2 RdRp inhibited states) with the active site of RdRp during the course of 100 ns MD simulation.

Discussion

The nsp12 subunit of RdRp is a crucial component of SARS-CoV-2 replicative machinery. In vitro and in vivo studies have validated the significance of RdRp as a suitable target protein for inhibiting SARS-CoV-2 replication [67], [72], [73], [74]. In this study, we attempted to identify RdRp non-nucleoside inhibitors against the novel coronavirus by screening large small-molecule databases (ZINC20 and DrugBank). Recently,Yin et al. reported suramin as a potential inhibitor of template/primer-strand RNA attachment to the RdRp of SARS-CoV-2 [67]. Our study initiated with the interaction analysis and binding free energy estimations of both suramin molecules attached at the protein's template/primer-strand RNA sites. We selected the suramin attached with primer-strand RNA binding site as a drug target site in our study due to its better interaction profile, higher binding energy value (–9.94 kcal/mol) over the suramin attached in template-strand RNA (–3.15 kcal/mol) of the protein, and the chance of dual inhibition mechanism (hindering the RNA attachment and nucleotide entry). During the virtual screening of large-scale databases, pharmacophore, an ensemble of steric and electronic characteristics, ensures optimum supramolecular contacts. Compared to molecular docking, it is a more promising and efficient strategy to find compounds against a specific target in order to modulate macromolecular activity [75]. Cele and colleagues [76], Kumalo and Soliman [77], used a more reliable pharmacophore approach based on highly contributing residues to inhibitor binding estimated from molecular dynamics ensembles to screen anti-HIV and anti-Alzheimer agents from chemical databases. To further increase the pharmacophore model accuracy, we preferred interacting amino acid residues of RdRp with suramin that had high free energy contribution to binding and significant H-bond frequency. The resultant library of potential hits (over 3500) was more concise and direct. Molecular docking is a fundamental and crucial technique for drug discovery that enables the prediction of protein-ligand molecular interactions in the bound state [78]. Previous studies have implemented a molecular docking approach to target the SARS-CoV-2 RdRp with FDA-approved drugs, alkaloids, and ZINC database compounds [72], [79], [80], [81]. Shortlisted non- nucleoside inhibitors in the present study possessed good docking scores and strongly interacted through H-bond with one or more critical residues (Lys553, Arg555, Arg557, Cys815, Ser816) of the RNA primer strand binding site, indicating their high potential to interrupt the RNA attachment, thereby blocking the function of SARS-CoV-2 RdRp. Due to the unfavorable parameters of pharmacokinetics and toxicity, many potential therapeutic agents have failed to reach clinical trials [82]. Thus, an early prediction of these properties of chemical entities can be advantageous for the drug discovery process to save cost and time [83]. Here, the selected lead compounds revealed acceptable absorption, distribution, metabolism, excretion profile, and no toxicity, indicating their safety and efficacy to use as anti-SARS-CoV-2 agents. MD simulation is a powerful technique that can capture atomic position and motion at every time point, which is otherwise very difficult at the experimental level. Structural alterations, ligand attachment, and protein folding are essential biomolecular processes that can be captured using these simulations [84]. This study's stability analysis revealed lower average RMSD scores of Inhibitor bound RdRp systems (>3 Å) than the free protein. Compared to the suramin bound complex and the apo-RdRp, only the ZINC1166211307 and ZINC28467879 complexes remained unstable till the end of the simulation with maximum RMSD value of 2.61 Å and 2.57 Å, respectively. There was no visible change in the simulation run, validating the importance of our simulation results. RMSF fluctuation analysis showed less flexibility of RdRp-complex systems than the apo-protein. During the 100 ns simulation period, ligand-bound systems displayed higher RMSF fluctuations, mostly in nonbinding residues (corresponding to loop regions), whereas other residues showed lesser fluctuations. This signifies that those drug candidates which bind at the suramin (or primer-strand RNA) binding region of RdRp, were stable and did not compromise the protein's backbone stability. The lower motion of selected compounds than the free state is due to the differential dynamics following inhibitor binding. Consistent with the RMSF, Rg analysis depicted stable behaviour and compact structure of inhibited protein systems throughout the simulation timescale. Besides, the regular fluctuations in electrostatic energy changes and the altered pattern of the SASA in inhibited complexes than the free protein indicated the dominant structural changes in RdRp caused by the binding of small molecule inhibitors. Only a few modes among the hundreds of protein modes comprise over 50% variation in the system. PCA efficiently extracts the slow (most important) modes of the protein motion that defines its biological function [85], [86]. PCA analysis in the current study revealed that ligand-bound RdRp complexes remained in the unstable conformation state (red) (Fig. 8) than the apo-state, which reflects significant conformational instability of the inhibitor’s attached complexes, which is consistent with the study of Md. Sorwer Alam Parvez et al. [87]. Dynamic cross-correlation (DCC) analysis, which is extensively employed to compute the inter-atomic correlation coefficient of motion [88], showed substitution effects on the internal protein dynamics of the protein due to the ligand binding. In addition, DCC analysis indicated that selected RdRp inhibitors induce dynamic variability and structural changes in the protein, showing their affinity for SARS-CoV-2 RdRp. Compared to apo-form, the inhibitor-bound RdRp showed more transitional conformations in the protein, as shown by their free energy landscapes (Fig. 11). As distributed by high and low energy barriers, multiple metastable states were noticed during structural alterations in the inhibitor-bound systems. Furthermore, as per the H-bond analysis, the chosen inhibitors demonstrated stable H-bonds with SARS-CoV-2 RdRp during the MD simulations, indicating their significance in stabilizing the complexes [89]. MM/GBSA is a popular approach to computing binding free energy of protein-ligand complex because it is based on molecular dynamics simulations. This approach is accurate and successfully applied to reproduce and justify the experimental results and improve the docking and virtual screening results [90]. By calculating MM/GBSA binding free energies, negative ΔG values of RdRp-inhibitor complexes were obtained (Table 2), exhibiting a strong binding affinity of small molecule inhibitors for RdRp's RNA primer strand binding site [91]. Binding free energy results are consistent with our docking results. Non-nucleoside drug candidates of SARS-CoV-2 RdRp predicted herein need expermental evaluation to treat and prevent COVID-19.

Limitations and future perspective

The present research employed various computational techniques, such as virtual screening, molecular docking, and molecular dynamics simulation, to identify non-nucleoside inhibitors of RdRp of SARS-CoV-2. Although these techniques are widely used for computer-aided drug discovery, there are some limitations associated with such methods, including model accuracy, size of the search space, ignoring flexibility of protein or ligand molecule, low enrichment, size of the modelled systems, and accuracy of the scoring function and force field [92], [93], [94], [95]. Therefore, the next step should be in vitro and in vivo validation of the identified compounds to establish their safety and efficacy profile. We leave the experimental evaluation of identified non-nucleoside inhibitors in the present study for the follow-up wet-lab studies. Future research may also focus on targeting the RNA-primer strand binding site as well as the RNA-template binding site with different categories of inhibitor molecules.

Conclusion

RNA-dependent RNA polymerase enzyme of SARS-CoV-2 is considered a major therapeutic target. We implemented a distinct structure-based virtual screening method to search non-nucleoside inhibitors of SARS-CoV-2. Based on docking scores, protein-ligand interaction, and physiological and pharmacokinetics parameters, eight compounds (including five novel and three existing) were identified as promising inhibitors of RdRp. Thus, molecular dynamics simulation was employed to explore the structural and dynamic behavior of RdRp upon binding with these eight compounds. MM/GBSA calculations revealed high binding affinity of ZINC285540154, ZINC28467879, ZINC611516532, ZINC1398350200, ZINC739681614, and ZINC1602963057) in range of −12.87 kcal/mol to −37.74 kcal/mol, whereas ZINC1166211307 and ZINC98208626 also showed good binding affinities for RdRp. Based on our computational results, we propose that these compounds can hinder the replication of SARS-CoV-2 by specifically preventing RNA-primer strand attachment with the RdRp enzyme. Hence, these compounds can act as potential non-nucleoside inhibitors of COVID-19 infection and warrant in vitro and in vivo testing to validate our findings.

Funding

Work was supported by grant from The Oman Research Council (TRC) through the funded project (BFP/RGP/EBR/21/005). The authors would like to thank the Deanship of Scientific Research at Umm Al-Qura University for supporting this work by Grant Code (22UQU4331128DSR55).

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A. Supplementary material

Supplementary material.

.