Computational Screening Using a Combination of Ligand-Based Machine Learning and Molecular Docking Methods for the Repurposing of Antivirals Targeting the SARS-CoV-2 Main Protease

Abstract

Background

COVID-19 is an infectious disease caused by SARS-CoV-2, a close relative of SARS-CoV. Several studies have searched for COVID-19 therapies. The topics of these works ranged from vaccine discovery to natural products targeting the SARS-CoV-2 main protease (M^pro), a potential therapeutic target due to its essential role in replication and conserved sequences. However, published research on this target is limited, presenting an opportunity for drug discovery and development.

Method

This study aims to repurpose 10692 drugs in DrugBank by using ligand-based virtual screening (LBVS) machine learning (ML) with Konstanz Information Miner (KNIME) to seek potential therapeutics based on M^pro inhibitors. The top candidate compounds, the native ligand (GC-376) of the M^pro inhibitor, and the positive control boceprevir were then subjected to absorption, distribution, metabolism, excretion, and toxicity (ADMET) characterization, drug-likeness prediction, and molecular docking (MD). Protein–protein interaction (PPI) network analysis was added to provide accurate information about the M^pro regulatory network.

Results

This study identified 3,166 compound candidates inhibiting M^pro. The random forest (RF) molecular access system ML model provided the highest confidence score of 0.95 (bromo-7-nitroindazole) and identified the top 22 candidate compounds. Subjecting the 22 candidate compounds, the native ligand GC-376, and boceprevir to further ADMET property characterization and drug-likeness predictions revealed that one compound had two violations of Lipinski’s rule. Additional MD results showed that only five compounds had more negative binding energies than the native ligand (− 12.25 kcal/mol). Among these compounds, CCX-140 exhibited the lowest score of − 13.64 kcal/mol. Through literature analysis, six compound classes with potential activity for M^pro were discovered. They included benzopyrazole, azole, pyrazolopyrimidine, carboxylic acids and derivatives, benzene and substituted derivatives, and diazine. Four pathologies were also discovered on the basis of the M^pro PPI network.

Conclusion

Results demonstrated the efficiency of LBVS combined with MD. This combined strategy provided positive evidence showing that the top screened drugs, including CCX-140, which had the lowest MD score, can be reasonably advanced to the in vitro phase. This combined method may accelerate the discovery of therapies for novel or orphan diseases from existing drugs.

Graphical abstract

Supplementary Information

The online version contains supplementary material available at 10.1007/s40199-023-00484-w.

Introduction

COVID-19 is a disease caused by infection with the SARS-CoV-2 virus. As of 29th October, 2022, it reached a prevalence of 629,740,541 cases and a mortality of 6 587 818 people worldwide [1]. SARS-CoV-2 is a close relative of SARS-CoV and MERS-CoV, which caused epidemics in 2002 and 2013, respectively [2]. A unique property that caused COVID-19 to become a global pandemic is the ACE2 receptor protein affinity of SARS-CoV-2, which is 10–20 times higher than that of SARS-CoV [3]. Therefore, the replication and spread rates of SARS-CoV-2 have become higher than those of its two predecessors. Various parties have developed vaccines to treat COVID-19 [4–7]. In addition, several studies applying a drug repurposing approach have suggested using existing drugs as alternative therapies for COVID-19 [8–11].

All the studies targeting several authentic SARS-CoV-2 proteins showed that the SARS-CoV-2 main protease (M^pro), 3CL^pro, and NSP5 are the most potent of these proteins due to their essential physiological functions in viral replication and their relatively conserved sequences (99.02% identity with SARS-CoV) [12]. Several inhibitors of M^pro, such as GC-373 and 376, have been developed [13–17]. Numerous studies have conducted drug discovery using molecular docking (MD), a popular structure-based virtual screening method, to identify hydroxychloroquine, remdesivir, lopinavir, and several other drugs as alternative therapies against COVID-19 [18–21]. However, some of these drugs lack adequate efficacy and could have unknown side effects due to the different dosage regimens used [22]. Therefore studies exploring the potential and safety of alternative M^pro-targeting treatments need to be conducted. Such studies can use all approved and experimental drugs subjected to trials in combination with another virtual screening method, such as ligand-based screening.

In this study, we retrieved a database of drug compounds from DrugBank. Hence, 11,300 compounds with various functions and uses ready to be applied as alternative antiviral therapies targeting M^pro were obtained. The initial stage of ligand-based virtual screening (LBVS) was conducted starting from ligand preparation by performing standardization in accordance with Gadaleta et al. and classifying ligands as active and inactive on the basis of cutoffs [23]. Active ligands were later used as input for scaffold mining and hierarchical clustering as a preparation for constructing the maximum common substructure (MCS) in SMiles ARbitrary Target Specification (SMARTS) format [24]. We used the SMARTS format as a screening reference by equating the similarity of the MCS with phytochemical substructures. Meanwhile, we applied active and inactive ligands as input to establish a machine learning (ML) model based on fingerprint similarity. Molecular access system (MACCS) and ECFP4, often known as Morgan, were used as the fingerprint algorithm [25–27]. Supervised classification was applied as the ML algorithm. We determined the activity of the screened chemicals by ranking their similarity scores with the ligands that interact with M^pro by using the best model. The top 22 candidate compounds generated by the best ML model as well as the native ligand GC-376 of the M^pro inhibitor and the positive control boceprevir were further tested on the basis of absorption, distribution, metabolism, excretion, and toxicity (ADMET) characterization, drug-likeness prediction, and MD. The results of these tests were compared with those of ML for potential activity. The findings of our study may serve as the foundation for performing preclinical and clinical trials as well as for identifying a rapid and affordable alternative or combination therapy with other medications while waiting for a robust and specific COVID-19 antiviral treatment.

Materials and Methods

Retrieval and analysis of proteins, targets, and drug compound data

A total of 16 473 ligands interacting with M^pro and their Simplified Molecular-Input Line-Entry System (SMILES) formats were obtained from PDB (https://www.pdb.org), ChEMBL (https://www.ebi.ac.uk/chembl/), and PubChem (https://pubchem.ncbi.nlm.nih.gov) databases. A total of 11 300 drug compounds were also obtained from the DrugBank database (https://www.drugbank.com) with their SMILES formats. Both data were standardized in accordance with the protocol of Gadaleta et al. by removing stereoisomers from SMILES; removing ions and salts; and filtering compounds with only H, C, N, O, F, Br, I, Cl, P, and S atoms. For ligand selection before standardization, the potency of the active/inactive compound must be classified through binary labeling: 0 for inactive and 1 for active, wherein an active compound has IC50/Ki activity < 1 µM [28] and %inhibition > 80% [29]. Specifically, in ligand preparation for scaffold mining, the valid protein–ligand interactions of PDB were filtered in accordance with resolution < 2.05 Å [30].

Scaffold mining, hierarchical clustering, and MCS alignment

The prepared and standardized active ligands were forwarded to scaffold-making with an unsupervised algorithm by using the RDkit Find Murcko Scaffold node to generate the Murcko framework [31]. After the scaffold collection was formed, it was further clustered via the average linkage method [32] with a distance threshold of 0.5 set at the cluster assigner node. Finally, the scaffold with the cluster number was forwarded to MCS creation and calculation by using the RDkit MCS node to produce an optimized scaffold in SMARTS format.

Candidate compound screening with the created MCS

The optimized scaffold was then used as a reference standard for screening the substructure of all candidate compounds that were filtered and standardized with the RDkit Substructure Filter and RDkit Molecular Highlighter nodes. This process was repeated every round with one cycle representing a scaffold that matched itself with the existing candidate compound structure such that a collection of screened compounds, which may contain more than one different scaffold, may be produced [33].

Ligand-based machine learning model analysis and testing

The pre-prepared ligands were classified as active and inactive on the basis of predetermined cutoff parameters. All active and inactive ligands were then forwarded to fingerprinting and ML similarity models. The ligand data in the form of SMILES and binary labels 0 and 1 were entered into fingerprint fragmentation by using the RDKit Fingerprint node with the fingerprint bitstring format. The two fingerprinting methods used were MACCS and Morgan [27]. SMILES was divided into several bitstrings and used to form an array, which was passed to three supervised ML algorithms that are often used for classification: random forest (RF), artificial neural network (ANN), and support vector machine (SVM).

The three algorithms were configured with a k-cross value of 10. The sampling method used by each algorithm was random stratified sampling. In addition, the Synthetic Minority Over-sampling Technique (SMOTE) node was used to help create an artificial ligand dataset to balance the number of training ligands. Furthermore, all ML configurations of the three algorithms were set to default by the application. After the ML model creation was completed, the models of the three algorithms for MACCS and Morgan were evaluated through the statistical analysis of the confusion matrix and receiver operating characteristic (ROC) curve [34]. The model with the best statistical results was used as a predictor to determine the similarity score of the previously screened candidate compound.

ADMET and drug-likeness prediction

In drug discovery, researchers must examine the ADMET properties of lead compounds. These properties play a crucial role in predicting how ligands will act in their active form as drugs when they reach the disease target. In this study, the ADMET properties of ligands and boceprevir were determined by using SwissADME (http://www.swissadme.ch/) and pkCSM (https://biosig.lab.uq.edu.au/pkcsm/prediction). Both websites are freely accessible and user-friendly. The features of both websites complement each other, covering molecular properties, pharmacokinetic properties, lipophilicity, solubility, and drug-likeness [35–37].

MD of the candidate drug compounds

The binding properties of the top 22 candidate compounds for 3CL^pro or M^pro were predicted via MD as previously described [38]. MOE 2010 (licensed from the Faculty of Pharmacy UGM) was utilized for docking simulation, RMSD docking score calculation, and interaction visualization. Computational prediction simulation was performed with Windows 11 operating system with 8 GB of RAM and Intel Core™ i5-10th Gen as a processor. The PDB ID of 7D1M (resolution 1.35 Å) was searched at http://www.rcsb.org. The top 23 candidate compounds, including boceprevir, were previously prepared. Microsoft Excel was used to construct a database of SMILES codes, which was saved in.csv format. Boceprevir is an HCV NS3-4A protease drug that was repurposed as an M^pro (3CL^pro) inhibitor [39, 40]. Structures were then subjected to conformational search and minimized in MOE by using the "Energy Minimize" menu. For docking simulation setting, London dG was used for Rescoring 1 and Rescoring 2. Triangle Matcher and Forcefield were applied for placement and refinement settings, respectively. Both placement and refinement of the docking result from the 30 retain setting. The results were used to determine which conformation had the lowest binding interaction between the ligand and its receptor.

Protein–protein interaction analysis of target proteins

BIOGRID (https://thebiogrid.org) [41] was searched for M^pro interactions by using the keyword "3clpro" by selecting "network" as the search term. The overall interaction of M^pro with other proteins was presented. In addition, protein interactions were filtered by using the minimal evidence "2" criterion to identify those with a high level of support in the literature. The curated protein was downloaded in.txt format for further customization in Cytoscape.

Txt files were exported to Cytoscape with Cytohubba plugin as previously described [42, 43] by using the "Drag network files here" menu. After exportation, the gene was sorted to the top 10 with "Degree" parameter in Hubba Nodes configuration. The steps included selecting the "cytoHubba" tab then clicking "Calculate" in the "Nodes’ scores" section. Subsequently, in the "Hubba nodes" section, the top 11 nodes were selected (with NSP5 included) and sorted based on the highest degree with the shortest path. The "Submit" button was then pressed.

Results

Protein and ligand preparation

The preparation of ligands that interact or have bioactivity with M^pro produced 269 active ligands from 16 473 available ligands, which can be used as materials for scaffold construction (Table 1). In addition, 3,969 active and inactive ligands from bioactivity data were used as materials to create ML and fingerprint similarity models. The division of labels based on cutoffs is essential considering that the reference for LBVS lies in ligand processing. We determined the cutoffs as resolution < 2.05 Å, IC50/Ki bioactivity < 1 µM, and %inhibition > 80% to ensure that the screened candidate compounds and ligands share some similar characteristics. Although the approach used is fingerprint similarity, the substructures with pharmacophore properties have potential bioactivity and position.

Table 1.

Results of collecting data on the ligands and compounds of M^pro. Data were collected by submitting an online request to each database then labeled and standardized

Protein Name	Compound Retrieval Result			Number of Active Compounds	Number of Standardized Active Compounds (SAC)	Number of SAC used for Scaffold Mining
	PDB	ChEMBL	PubChem
Mpro	317	8819	7337	5,259	487	269

Another preparation is standardization by using the protocol of Gadaleta et al. to generalize drug compounds into 2D and normal forms (without ions; salts; and atoms other than H, C, N, O, F, Br, I, Cl, P, and S) [23]. Standardization is done to facilitate scaffold construction in substructure breakdown such that a Murcko framework is produced, which is less disruptive and reduces bias for ML when constructing a model.

MCS scaffold creation and clustering

Through scaffolding, clustering, and MCS optimization, we successfully screened 98 scaffolds for candidate compounds (Table 2). These scaffolds were divided into five groups: (1) alyll and derivatives, (2) N-containing cyclohexane and biphenyl and its derivatives, (3) benzo-azoles, (4) naphthalene and benzopyrene, and (5) flavonoids. The total number of compounds screened also decreased to 5,310 from 10,692 (49.6%), matching the 98 scaffolds above. The scaffold with the highest number of candidate compounds was obtained with alyll and derivatives (1,761), followed by N-containing cyclohexane and biphenyl and its derivatives (636) and benzo-azoles (178). Therefore, candidate compounds with familiar substructures, such as the above top three compounds, have great potential to become M^pro inhibitors.

Table 2.

List of 10 optimized scaffolds that produced the greatest number candidate compounds and their clusters. The scaffold sequence started from coumarin, then continued with allyl carbon 9, N-containing cyclohexane, benzodioxol, flavonoids, naphthalene, allyl carbon 7, and biphenyl

Murcko framework creation, cluster grouping, and scaffold optimization by using the MCS approach were sequentially included in the LBVS method by using 2D descriptors. The descriptors were used such that screening was conducted on the basis of the properties of existing ligands, which were then utilized to identify compounds with similarities to the ligand properties used as variables. This method is different from structure-based screening, which refers to the 3D structure and the position or coordinates of ligands. In this work, we employed two approaches to use descriptors: the formation of a general substructure by forming a scaffold as discussed in the results above, and the breakdown of the ligand structure into fingerprints representing active and inactive ligand substructures. The results given by the first approach have numerous disadvantages. The three main drawbacks are disorganized screening results, the bias of screening results due to reliance on scaffold similarity, and the absence of the quantification of which compounds are potent in inhibiting target proteins. Therefore, fingerprint similarity methods and ML support the prediction of all candidate compounds.

Evaluation results of the fingerprint similarity approach and machine learning

The second ligand-based screening approach that we used in this study combined fingerprint similarity with ML. We obtained statistical results by entering 3,969 active and inactive ligands, pre-processing, and conducting training and model testing (Table 3). On the basis of the ML evaluation of the three algorithms and two fingerprint methods, we concluded that the RF MACCS model was the best model because all of its aspects had the highest values compared with those of other models. In addition, the value of each evaluation met the minimum criteria with overall accuracy > 90% [44] and ROC AUC within the quite acceptable range (0.7–0.8) [45]. However, it is different from Cohen’s kappa and balanced accuracy, where the kappa values ​​are in the low approval range (0.00–0.40) [46] and balanced accuracy < 1% because of an imbalance of active and inactive ligands from the start. Despite evaluating the RF model, SVM model for MACCS and Morgan showed an inability to classify inactive and active ligands as reflected by their overall accuracy and Cohen’s kappa results, which did not produce results, and their ROC AUC curves were less than 0.5 (Fig. 1). These two values ​​indicated that the SVM model failed [44–46]. After we determined that RF MACCS is the model to be used, we exported it to the last stage to predict the similarity of candidate compounds.

Table 3.

Summary of machine learning parameters with Consideration to Use the Model for Prediction for 3CL^pro Targeted Therapy. RF MACCS fingerprinting was used as a model predictor because it had the highest score among all ML algorithms and fingerprinting methods

ML Evaluation Parameter	MACCS			Morgan
	RF	ANN	SVM	RF	ANN	SVM
Overall Accuracy (%)	95.94	93.32	N/A	89.29	95.99	N/A
Balanced Accuracy (%)	0.880	0.924	N/A	0.940	0.98	N/A
Cohen’s Kappa (κ)	0.223	0.161	0.000	0.107	0.129	0.000
ROC AUC	0.762	0.703	0.418	0.748	0.584	0.433

Fig. 1.

ROC curves of each MACCS fingerprint machine learning algorithm. The RF model outperformed the other two models, whereas the SVM model was unsuitable for predicting the scores of candidate compounds

Application of machine learning models to compound screening

The RF MACCS model was recruited as a predictor then used to generate the predictive scores of candidate drug compounds. Drug compounds standardized in SDF format were broken down by the RDKit Fingerprint node, expanded by the expand bit vector node, and passed to the predictor node to generate a predictive score. The prediction results were produced from 3,166 compounds. The model obtained 16 drug compounds with confidence > 90%. Of these compounds, 3-bromo-7-nitroindazole received the highest score of 0.95 and 3-(2,6-difluorophenyl)-2-(methylthio)quinazolin-4(3H)-one received the lowest score of 0.20 (Fig. 2 and Appendix). Compounds containing the azole structure (indazole, thiadiazole, phenylpyrazole, and phenyltetrazole), pyrazolopyrimidine, benzenesulfonamide, dipeptide, and carboxylic acid have potential pharmacological effects targeting M^pro. These substructures have the potential to be advanced to the in vitro phase but require further comparison through a structure-based method like MD.

Fig. 2.

List of screened candidate compounds with confidence score >90%. The list was ranked from highest to lowest by the RF MACCS model. Out of 3,166 compounds, 16 had confidence score >90%

ADMET properties and drug-likeness prediction

Among 23 candidate compounds, only ethyl (2E)-4-({(2S)-2-[(N-{[(2-methyl-2-propanyl)oxy]carbonyl}-L-valyl)amino]-2-phenylacetyl}amino)-5-[(3S)-2-oxo-3-pyrrolidinyl]-2-pentenoate has two violations of Lipinski’s rule and was thus excluded from selection (Additional file). Among the top six compounds, boceprevir and GC-376 showed low gastrointestinal absorption rates of only 55.08% and 19.74%, respectively (Table 4). All compounds were considered to have low skin permeability and were not considered as blood–brain barrier (BBB) permeants. They also had relatively low steady-state volumes of distribution. Only boceprevir and GC-376 were predicted to be Pgp substrates. GC-376 was categorized as the most soluble compound and was not predicted to be a CYP450 substrate or inhibitor. SC-236 was predicted to be a renal organic cation transporter 2 (OCT2) substrate, indicating that it has potential adverse interactions with coadministered OCT2 inhibitors [37].

Table 4.

ADMET properties and drug-likeness predictions of the top 6 candidate compounds

Properties, Parameters	CCX-140	SC-558	Celecoxib	Boceprevir	SC-236	GC-376
Physicochemical properties
MW (g/mol)	495.86	446.2	381.4	519.7	401.8	485.6
Heavy atoms	33	26	26	37	26	37
Arom. Heavy atoms	21	17	17	0	17	0
Rotatable bonds	6	4	4	14	4	14
H-bond acceptors	9	7	7	5	7	8
H-bond donors	12	1	1	4	1	5
Molar Refractivity	113.7	92.7	89.96	144.4	90.01	122.7
TPSA	126.1	86.36	86.36	150.7	86.36	179.5
Absorption
GI absorption (%absorbed)	94.54	91.16	92.73	55.08	91.23	19.74
Log Kp skin permeability (cm/s)	-6.09	-6.37	-6.21	-7.23	-6.15	-8.76
Pgp substrate	No	No	No	Yes	No	Yes
Distribution
VDss (human) (log L/kg)	0.026	-0.07	-0.287	-1.189	-0.087	-1.711
BBB permeant	No	No	No	No	No	No
Lipophilicity
Consensus Log Po/w	3.81	3.69	3.4	2.06	3.51	0.66
Water Solubility
Consensus Solubility Class	Poorly soluble	Moderately soluble	Moderately soluble	Moderately soluble	Moderately soluble	Soluble
Metabolism
CYP450 substrate	-	CYP3A4	CYP3A4	CYP3A4	CYP3A4	-
CYP450 Inhibitor	CYP1A2, CYP2C19, CYP2C9, CYP3A4	CYP1A2, CYP2C19, CYP2C9	CYP1A2, CYP2C9	CYP3A4	CYP1A2, CYP2C19, CYP2C9	-
Excretion
Total Clearance (log ml/min/kg)	0.062	-0.077	0.461	0.158	-0.055	1.35
Renal OCT2 substrate	No	No	No	No	Yes	No
Toxicity
AMES toxicity	No	No	No	No	No	No
Max, tolerated dose (human) (log mg/kg/day)	0.561	0.52	0.261	-1.248	0.517	1.382
hERG inhibitor	hERG II	-	-	-	-	-
Oral Rat Acute Toxicity (LD50) (mol/kg)	2.053	2.741	2.389	3.692	2.728	1.959
Oral Rat Chronic Toxicity (LOAEL) (log mg/kg_bw/day)	2.526	0.412	1.027	2.516	0.422	4.325
Hepatotoxicity	Yes	Yes	Yes	Yes	Yes	Yes
Skin Sensitisation	No	No	No	No	No	No
T.Pyriformis toxicity (log ug/L)	0.285	0.332	0.518	0.285	0.333	0.285
Minnow toxicity (log mM)	1.677	-1.555	0.191	4.891	-1.409	3.613
Drug-likeness
Lipinski, violation	Yes, 0 violation	Yes, 0 violation	Yes, 0 violation	Yes, 1 violation	Yes, 0 violation	Yes, 1 violation
Bioavailability Score	0.55	0.55	0.55	0.55	0.55	0.11
Med. Chemistry
Synthetic accessibility	3.23	2.68	2.74	5.41	2.63	5.0

In the toxicity prediction test, all compounds were negative for AMES toxicity and skin sensitization but tested positive for hepatotoxicity. CCX-140 alone was predicted as a potential human ether-a-go-go gene (hERG) II inhibitor, leading to fatal ventricular arrythmia. Lastly, all compounds had no more than one violation of Lipinski’s rule, meaning all of them exhibited good drug-likeness. However, GC-376 showed the worst bioavailability score of 0.11. In terms of medicinal chemistry, SC-236 is the most synthetically accessible among the compounds with a score of 2.63 [35].

MD

After we obtained the top 22 potential compounds on the basis of the RF algorithm, ADMET properties, and drug-likeness predictions, we conducted MD to examine each compound’s potency against the M^Pro receptor protein complexed with GC-376 (PDB code: 7D1M resolution 1.35 Å). Out of the 22 compounds (including GC-376, the native ligand of 7D1M, and boceprevir, the positive control) selected on the basis of ADMET properties and drug-likeness predictions, only five compounds had more negative binding energies than the native ligand (− 12.25 kcal/mol). CCX-140 exhibited the most negative score docking score of − 13.64 kcal/mol, followed by followed by (2) 1-phenylsulfonamide-3-trifluoromethyl-5-parabromophenylpyrazole (SC-558, − 13.24), (3) celecoxib (− 12.52), (4) boceprevir (− 12.47), and (5) SC-236 (− 12.30) (Fig. 3 and Table 5). A low docking score indicates the strong binding affinity of a ligand to a receptor [47]. The inhibitory mechanism of boceprevir and GC-376 against M^pro has been proven, reinforcing previous research conducted by Fu et al. According to previous reports, boceprevir undergoes hydrogen bonding interactions with the residues of His41, Gly143, His164, and Glu166 in the M^pro binding site. It also undergoes hydrophobic interactions with the residues of Gln189, Gln192, and Thr190 in the S4 subsite [39]. In this study, we found that boceprevir showed interactions with Glu166 at a distance of 2.02 Å and Gln189 at distances of 1.73 and 2.10 Å. In addition, GC-376 created a covalent bond with Cys145 and experienced hydrogen bonding interactions with the carboxyl group of Glu166. It also underwent hydrophobic interactions with the residues of Arg40, His41, Met49, Tyr54, and Asp187. However, we found slight differences in the interactions that involved five residues: His41 at 3.96 Å, Gly143 at 2.09 Å, His163 at 1.97 Å, Glu166 at 2.11 Å, and Gln189 at 1.89 Å.

Fig. 3.

Visualization of the interaction of the crystal structure of M^Pro (PDB ID: 7D1M resolution 1.35 Å) with several ligands, such as (A) CCX-140, (B) SC-558, (C) celecoxib, (D) boceprevir, (E) SC-236, and (F) its native ligand (GC376)

Table 5.

Molecular docking analysis results of top 22 candidate compounds and boceprevir

Protein–protein interaction analysis

During transcription and translation, M^pro is coexpressed with other proteins, even with host proteins and the process does not end as they undergo replication to form a new virus; these two proteins interact with other proteins. Therefore, the protein–protein interactions (PPIs) of M^pro or NSP5 must be analyzed. PPI analysis with the BIOGRID database produced M^pro network interaction results with 316 nodes (15 from SARS-CoV-2 and 301 from Homo sapiens), 405 interactions, and 1,137 edges (1,124 physical edges and 13 physical/genetic edges) (Fig. 4 A). The top 10 interactors, other interacting proteins, and the relationship between the interactors and the repurposed drug candidate were identified (Fig. 4B). The following is a brief description of the function of each protein that interacted with M^pro.

Fig. 4.

(A). Protein-protein interaction network of NSP5 or M^pro analyzed by using BIOGRID. (B). Top 10 proteins interacting with NSP5 or M^pro analyzed by the degree Hubba node of Cytoscape. Dark colors indicate strong interactions

The first protein is SDHA, a succinate dehydrogenase complex flavoprotein subunit A protein that participates in cellular respiration at the Krebs cycle stage and oxidative phosphorylation and even acts as a tumor-suppressor gene [48]. The second protein is DEAD-box helicase 39B (DDX39B) with a vital role in ATP hydrolysis during pre-mRNA splicing and a gene belonging to the gene cluster encoding TNF-α [49]. The third protein is NSP9, a part of the RP1a protein of SARS-CoV-2 that acts in viral replication [50]. The fourth protein is SUPT16H, a component of the FAcilitates Chromatin Transcription complex, a chromatin-specific factor required for transcription elongation in DNA replication and repair [51]. The fifth protein is LRPPRC, a leucine-rich PPR-motif-containing protein with an important role in the stability, regulation, splicing, and translation of RNA in humans and is frequently involved in tumors and viral infections [52]. The sixth protein is MYO1B, which has a vital role in cell migration and proliferation in the body [53]. The seventh protein is TFAM, a protein that encodes transcription factor A, which is a DNA-binding protein that activates transcription at promoters in mitochondrial DNA [54]. The eighth protein is EIF2S1, the alpha subunit of the initiation factor eIF2 protein complex that initiates the binding of the initiator tRNA to the 40 s ribosomal subunit [55]. The ninth protein is the eukaryotic initiation factor 4A-I protein, which has a vital role in ribosome recruitment and translation initiation [56]. The final protein is the nonerythrocytic beta-spectrin 2 (SPTBN2) protein, which participates in neurotransmitter traffic in the central nervous system, especially the cerebellum [57].

The analysis of the 10 proteins that interacted with NSP5 showed that SARS-CoV-2 infection, apart from hijacking the host’s cells and inhibiting the immune system, has numerous other potentially harmful effects on the human body, including the initiation of the following pathologies: (1) NF-κB-mediated inflammation as evidenced by DDX39B and LRPPRC; (2) the potential to elicit cognitive ability and memory syndromes as evidenced by SPTBN2; and (3) cancer progression and angiogenesis as evidenced by SDHA, SUPT16H, LRPPRC, and MYO1B. Several clinical reports also mentioned the potential harmful pathologies of M.^pro, although their evidence is unclear [58–64]. 

Discussion

In this study, we repurposed candidate drugs targeting M^pro from drug compound databases for COVID-19. A total of 98 MCS-ed scaffolds were obtained through the processing of 269 active ligands specifically targeting M^pro. Of the 10,692 compounds retrieved from DrugBank, 3,166 targeting M^pro were successfully screened on the basis of the MCS reference. These drugs were divided into six groups in accordance with their status, with the highest number of drugs originating from the experimental (1,525 compounds), investigational (965 compounds), and approved (489 compounds) groups. Prediction results obtained by using the RF MACCS models showed that the drug with the best confidence score was 3-bromo-7-nitroindazole (0.95), which is an indazole (pyrazole). Through literature analysis, six compound classes with high affinity to M^pro and reasonable safety were discovered, namely, benzopyrazole, azole, pyrazolopyrimidine, carboxylic acids and derivatives, benzene and substituted derivatives, and diazone. The drug screening results identified drugs that were different from the findings of MD studies published on the Internet. Therefore, in this study, we also compared the MD results and IC₅₀ from ML to provide accurate data and selected four groups of drugs as follows:

1. CCX-140.

CCX-140 is a CCR2 chemokine inhibitor, which acts as a potent antidiabetic therapeutic by reducing hyperglycemia and insulinemia and increasing insulin sensitivity and circulated adiponectin levels (T. J. [65]. Similar to Celecoxib, which can suppress the IL-6 cytokine, CCX-140 has the potential to be repurposed as a COVID-19 drug through the same pathway. This is reinforced by its ML RF prediction score of 0.91 and MD, showing that CCX-140 had a binding affinity of − 12.58 kcal/mol, which exceeded that of the native ligand (− 12.39 kcal/mol), and four hydrogen bonds involving the amino acids His41 (3.11), Met165 (4.16), Glu166 (3.96), and Gln189 (2.05) (Fig. 3A). CCX-140 also acts as a drug for diabetic kidney disease by decreasing albuminuria [66, 67].

CCX-140 was predicted to inhibit four different CYP450 enzymes, namely, CYP1A2, CYP2C19, CYP2C9, and CYP3A4. CYP450 enzymes play important roles in drug detoxification, cellular metabolism, and homeostasis because they can metabolize endogenous and exogenous substances [68, 69]. The expression and activity of CYP450 enzymes are also influenced by immune responses. In this case, the metabolism of CCX-140 is similar to that of remdesivir, an FDA-approved drug for COVID-19 treatment [70]. Remdesivir has been shown to inhibit CYP1A2, CYP2C9, CYP2C19, CYP2D6, and CYP3A4 [71]. Given that CYP3A4 is responsible for metabolizing a substantial portion (approximately 70%) of administered drugs, its inhibition due to concomitant inflammatory conditions and the concurrent use of other drugs metabolized by CYPs could potentially reduce the elimination of CCX-140 and lead to unpredictable dose-related toxicity [72].

Additionally, the pkCSM AMDET prediction test showed that only CCX-140 has potency as a hERG II inhibitor. The development of hERG inhibitors poses a challenge in drug discovery because hERG inhibition can lead to cardiotoxicity and the subsequent rejection of potential drug candidates [73]. hERG inhibition is also associated with QT prolongation and lethal ventricular tachycardia, which is known as torsades de pointes [74]. Therefore, further studies are necessary to investigate the metabolic activity and hERG inhibition potency of CCX-140 before considering its repurposing as a COVID-19 treatment.

• b.b.SC-558, celecoxib, and SC-236.

These three compounds have similar structures, and the functional group of 3-phenyl-1H-pyrazole is the only feature that can be used to distinguish among them. Specifically, SC-558 has a bromo group, celecoxib has a methyl group, and SC-236 has a chloro group. The three compounds are also known as cyclooxygenase-2 (COX-2) inhibitors. Their ML RF scores were 0.93, 0.92, and 0.92. MD results showed that among the three drugs, celexocib has the highest binding affinity (− 12.69 kcal/mol), followed by SC-558 (− 12.58 kcal/mol), and SC-236 (− 12.54 kcal/mol), exceeding the binding affinity of the native ligand (− 12.39 kcal/mol). Among all COX-2 inhibitors, celecoxib has the most prominent evidence to become a potential repurposed drug for COVID-19. The two pathophysiologies of COVID-19 considerably influence hyperinflammation and cytokine hyperactivity [75]. Celecoxib could suppress the effect of the proinflammatory cytokines IL-6 and IL-1β [76].

Furthermore, celecoxib has viral replication inhibition activity. It could target the spike and RdRp proteins of SARS-CoV2 [75], synergistically inhibit 5N1 viral infection in combination with zanamivir [77], and execute antiviral activities against the FIP coronavirus [78]. Hence, celecoxib has dual therapeutic targets in the treatment of COVID-19: decreasing cytokine hyperactivity and inhibiting viral replication [75]. Moreover, as proven by Gimeno et al. [79], celecoxib has high potency as a M^pro inhibitor due to the in vitro testing of M^pro inhibition at 50 μM (11.90%).

The interesting aspect of the ADMET prediction test is its metabolism part, which revealed that all three compounds, along with boceprevir, are CYP3A4 substrates. Two antiviral drugs, lopinavir and ritonavir, have been proposed for emergency COVID-19 treatment [80]. Notably, lopinavir and ritonavir are also CYP3A4 substrates [81]. In the context of COVID-19 treatment, lopinavir levels increase possibly due to the inflammation caused by COVID-19, which can downregulate CYP3A4 expression. This result has been demonstrated by Croxtall and Perry [82] and Marzolini et al. [83], who observed that lopinavir levels are 3.5-fold higher in patients with COVID-19 than in those with HIV infection [82, 83]. Considering the findings for lopinavir and the fact that SC-558, celecoxib, and SC-236 are also CYP3A4 substrates, further studies must be conducted to specify the appropriate dose regimen for each of these compounds to avoid drug toxicity.

• c.c.GC-373, GC-376, and boceprevir.

In this study, GC-376 and boceprevir acted as the native ligand and positive control, respectively. GC-376 was proven to act as a M^pro inhibitor in vitro and in vivo. In vitro, GC-376 has excellent antiviral activity against SARS-CoV2 with the M^pro inhibition of 0.0030 ± 0.008 μM [84]. In vivo studies reported GC-376 as a promising lead candidate causing mild tissue lesions, decreasing the presence of viral antigens and suppressing viral loads in the K18 hACE2 transgenic mouse model [85]. Boceprevir is a FDA-approved anti-HCV drug [86]. In Vero cells, GC-376 and Boceprevir efficaciously inhibited SARS-Co-V-2 by targeting M^pro [39].

Although GC-376 and GC-373 are broad-spectrum inhibitors of 3C and 3C-like proteases of picornaviruses, noroviruses, and coronaviruses [87], these two ligands showed different MD simulation results. The binding affinity to the protein receptor shown by GC-373, which has a prediction score of 0.89 from ML RF, is not as strong (− 11.89 kcal/mol) as that of the native ligand. Boceprevir has the docking score of − 12.47 kcal/mol. GC-376 itself, as the native ligand of M^Pro with a prediction score of 0.85, has a docking score of − 12.25 kcal/mol.

Drug-likeness prediction showed that GC-376 tends to have low gastrointestinal absorption, skin and BBB permeability, and bioavaliability even though it has considerably better solubility than other compounds. The interesting aspect of the drug-kikeness prediction is the metabolism evaluation, which shows that only GC-376 lacks CYP450 substrate and inhibition potency. This result may indicate that GC-376 is not metabolized by CYP enzymes but may be metabolized by other enzymes, such as UDP-glucuronosyltransferases, sulfotransferases, or N-acetyltransferases. In this case, further research is necessary to delve deeply and identify the metabolic pathway of GC-376 because it is important for understanding the drug interactions of this compound. Meanwhile, GC-376 and boceprevir were predicted to be Pgp substrates, indicating that both compounds utilize the Pgp transporter to pump out zenobiotics from cells [88], leading to decreased drug concentrations at the target site.

Next, the molecular mechanism of the screened drugs in M^pro inhibition was proposed. The results of PPI analysis highlighted that the 10 proteins with interactions with M^pro have the potential to have molecular mechanisms for four pathologies: (1) downregulating antiviral activity (antiviral evasion), (2) increasing the risk of NFΚB-mediated inflammation, (3) impairing cognitive and memory impairment due to ataxia or another syndrome, and (4) increasing risk factors for cancer in long COVID-19 cases. There is some molecular mechanism evidence of M^pro in causing these four pathologies and how the illustration of these class compounds from screening and previous literature reviews can help reduce the four pathologies.

SARS-CoV-2, which is endocytosed into host cells, releases RNA. RNA is then translated by ORF1a into RP1a, which consists of several NSPs, including M^pro (NSP 5). It is subsequently replicated and translated into viral proteins. In this process, M^pro in the mitochondria of host cells inhibits SDHA, causing oxidative stress and potentially increasing tumor risk [89]. M^pro then inhibits LRPPRC, increases the risk of NF-κB-mediated inflammation, and inhibits IFN-1 signaling by blocking TRAF3 and TRAF6 [90]. In addition, M^pro inhibits DDX39B, which regulates PSMB9, a NFΚB-mediated proinflammatory protein [91, 92], and inhibits TFAM, which indirectly inhibits ISG15 in the upregulation of IFN-1 production due to moderate mitochondrial stress [93]. M^pro inhibits SUPT16H, which indirectly inhibits ISG15, which M^pro also directly inhibits, thereby decreasing IFN-1 production [94, 95]. The interaction of M^pro with cleaved NEMO also has an essential role in brain endothelium damage [96], and SPTNB2 is part of the vulnerable 40 loci of ataxia genes, including ATXNs [63], suggesting how cognitive impairment occurs in patients with COVID-19.

The illustration of the mechanism in Fig. 5 was proposed by observing the complexity of the above molecular interactions. It briefly explains how M^pro regulates the protein that causes the four pathologies above and how the candidate drug compounds address these problems. M^pro inhibition is beneficial because it directly manages to stop various risk factors. The drugs are predicted to work on M^pro inhibition and related proteins interacting with M^pro. For example, CDK2 inhibitors inhibit the oncogene protein, and COX inhibitors inhibit not only COX but also IL-6 and the NF-κB proinflammatory pathway. However, the precise mechanisms of signaling processes in pathways, such as MAVS/RIG-1, and the role of downstream proteins in the relevant pathways are unknown and should be studied further.

Fig. 5.

Proposed molecular mechanism of RP1a and other proteins under treatment with the candidate compounds (created with PathVisio 3.3.0)

Scaffolding with MCS alignment generates worse results than scaffold hopping or pharmacophore hip-hop alignment (pharmacophore-based) [33] because in contrast to the pharmacophore-based approach, it does not generate structure–activity relationships and is reliant on scaffold similarity. However, the pharmacophore-based approach needs to be conducted in a commercial setting. This study indirectly emphasizes the use of an open-source pipeline and application to generate a drug repurposing approach.

Conclusion

We identified 98 MCS-ed scaffolds through the processing of 269 active ligands specifically targeting M^pro. Additionally, 3,166 therapeutic compounds targeting M^pro were successfully screened out of the 11,300 compounds obtained from DrugBank on the basis of the MCS reference, yielding a screening success rate of 28.01%. The MD data also revealed that five of the top 22 compounds identified on the basis of confidence scores exhibited negative binding energies compared with the native ligand. On the basis of their lowest MD score, these compounds were ranked as CCX-140, SC-558, ethyl (2E)-4-({(2S)-2-[(N-{[(2-methyl-2-propanyl)oxy]carbonyl}-L-valyl)amino]-2-phenylacetyl}amino)-5-[(3S)-2-oxo-3-pyrrolidinyl]-2-pentenoate, celecoxib, boceprevir, and SC-236. In general, all six compounds showed good ADMET properties and had no more than one violation of Lipinski’s rule, indicating they all exhibited good drug-likeness and can be reliably advanced to the in vitro phase. The top 10 interactors, other interacting proteins, and the relationship between the interactors and the repurposed drug candidate were identified. The proposed mechanism briefly explains how M^pro regulates the protein that causes the pathologies due to SARS-CoV-2 infection and how the candidate drug compounds address these problems. In this work, two virtual screening methods were successfully combined to accelerate drug repurposing and provided good results with greater accuracy than two methods used separately.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

ADMET

Absorption, distribution, metabolism, excretion, and toxicity

ANN

Artificial neural network

BBB

Blood–brain barrier

DDX39B

DEAD-box helicase 39B

COX-2

Cyclooxygenase-2

hERG

Human ether-a-go-go gene

LBVS

Ligand-based virtual screening

MACCS

Molecular access system

MCS

Maximum common substructure

MD

Molecular docking

ML

Machine learning

M^Pro

SARS-CoV-2 main protease

OCT2

Organic cation transporter 2

PPI

Protein–protein interaction

RF

Random forest

ROC

Receiver operating characteristic

SC-558

1-Phenylsulfonamide-3-trifluoromethyl-5-parabromophenylpyrazole

SMARTS

SMiles ARbitrary Target Specification

SMILES

Simplified Molecular-Input Line-Entry System

SMOTE

Synthetic Minority Over-sampling Technique

SPTBN2

Nonerythrocytic beta-spectrin 2

SVM

Support vector machine

Author Contributions

GPWCY contributed to the design, acquisition, the writing and revision of the article, drafted the article, and the finalized the version to be published. NH contributed to the data acquisition, writing and revision of the article. AH contributed to supervision, review and evaluation of the data and the final approval of the version to be published. This manuscript is a part of the bachelor thesis of GPWCY.

Funding

The authors did not receive any particular grant from the public, commercial, or non-profit funding agency.

Data Availability

All data generated or analyzed during this study are included in this published article [and its supplementary information files].

Declarations

Ethics approval and consent to participate

This work does not involve any studies with human participants or animals.

Competing interest

The authors declare that they have no competing interests.