Repositioning of 8565 Existing Drugs for COVID-19

Abstract

The coronavirus disease 2019 (COVID-19) pandemic caused by severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) has infected over 5 million people and led to over 0.3 million deaths. Currently, there is no specific anti-SARS-CoV-2 medication. New drug discovery typically takes more than ten years. Drug repositioning becomes one of the most feasible approaches for combating COVID-19. This work curates the largest available experimental dataset for SARS-CoV-2 or SARS-CoV main protease inhibitors. Based on this dataset, we develop validated machine learning models with relatively low root mean square error to screen 1553 FDA-approved drugs as well as other 7012 investigational or off-market drugs in DrugBank. We found that many existing drugs might be potentially potent to SARS-CoV-2. The druggability of many potent SARS-CoV-2 main protease inhibitors is analyzed. This work offers a foundation for further experimental studies of COVID-19 drug repositioning.

Introduction

Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) outbreak in Wuhan, China, in late December 2019 and has rapidly spread around the world. By June 11, 2020, over 7.1M individuals were infected, and more than 408K fatalities had been reported. Currently, there is no specific antiviral drug for this epidemic. It is worth noting that recently, an experimental drug Remdesivir, has been recognized as a promising anti-SARS-CoV-2 drug. However, the high experimental value of IC₅₀ (11.41 μM)¹ indicates that it must be used in a large dose in treating COVID-19, which is subject to side effects.

Considering the severity of this widespread dissemination and health threats, panic patients misled by media flocked to the pharmacies for Chinese medicine herbs, which were reported to “inhibit” SARS-CoV-2, despite no clinical evidence supporting the claim. Although there is also no evidence for the claimed Chloroquine’s curing effect, some desperate people take it as “prophylactic” for COVID-19. Many researchers are engaged in developing anti-SARS-CoV-2 drugs.^2,3 However, new drug discovery is a long, costly, and rigorous scientific process. A more effective approach is to search for anti-SARS-CoV-2 therapies from existing drug databases.

Drug repositioning (also known as drug repurposing), which concerns the investigation of existing drugs for new therapeutic target indications, has emerged as a successful strategy for drug discovery due to the reduced costs and expedited approval procedures.^4–6 Several successful examples unveil its great values in practice: Nelfinavir, initially developed to treat the human immunodeficiency virus (HIV), is now being used for cancer treatments. Amantadine was firstly designed to treat influenza caused by type A influenza viral infection and is being used for Parkinson’s disease later on.⁷ In recent years, the rapid growth of drug-related datasets, as well as open data initiatives, has led to new developments for computational drug repositioning, particularly structural-based drug repositioning (SBDR). Machine learning, network analysis, and text mining and semantic inference are three major computational approaches commonly applied in drug repositioning.⁸ The rapid accumulation of genetic and structural databases,⁹ the development of low-dimensional mathematical representations of complex biomolecular structures,¹⁰ and the availability of advanced deep learning algorithms have made machine learning-based drug reposition a promising approach.⁸ Due to the urgent need for anti-SARS-CoV-2 drugs, a computational drug repositioning is one of the most feasible strategies for discovering SARS-CoV-2 drugs.

In SBDR, one needs to select one or a few effective targets. Study shows that SARS-CoV-2 genome is very close to that of the severe acute respiratory syndrome (SARS)-CoV.¹¹ The sequence identities of SARS-CoV-2 3CL protease, RNA polymerase, and the spike protein with corresponding SARS-CoV proteins are 96.08%, 96%, and 76%, respectively¹² (See S1.1 of the Supporting Material). We, therefore, hypothesize that a potent SARS 3CL protease inhibitor is also a potent SARS-CoV-2 3CL protease inhibitor. Unfortunately, there is no effective SARS therapy at present. Nevertheless, the X-ray crystal structures of both SARS and SARS-CoV-2 3CL proteases have been reported.^13,14 Additionally, the binding affinities of SARS-CoV or SARS-CoV-2 3CL protease inhibitors from single-protein experiments are available in various databases or the original literature. Moreover, the DrugBank contains about 1600 drugs approved by the U.S. Food and Drug Administration (FDA) as well as more than 7000 investigational or off-market ones.¹⁵ The aforementioned information provides a sound basis to develop an SBDR machine learning model for SARS-CoV-2 3CL protease inhibition. It is worth clarifying that SBDR machine learning models are driven by the data and do not explicitly form the energy terms related to some biophysical characteristics such as electrostatics and hydrogen bonding. Instead, those biophysical interactions are implicitly encoded in the fingerprints and their impacts on the binding affinity are regulated by machine learning scoring functions.

In responding to the pressing need for anti-SARS-CoV-2 medications, we have carefully collected 314 bonding affinities for SARS-CoV or SARS-CoV-2 3CL protease inhibitors, which is the largest set available so far for this system. Machine learning models are built for these data points.

Unlike most earlier COVID-19 drug repositioning works that did not provide a target-specific cross-validation test, we have carefully optimized our machine learning model with a 10-fold cross-validation test on SARS-CoV-2 3CL protease inhibitors. We achieve a Pearson correlation coefficient of 0.78 and a root mean square error (RMSE) of 0.79 kcal/mol on the test sets of 10-fold cross validation tasks, which is much better than that of similar machine learning models for standard training sets in the PDBbind database (around 1.9 kcal/mol).¹⁶ We systematically evaluate the binding affinities (BAs) of 1553 FDA-approved drugs as well as 7012 investigational or off-market ones in the DrugBank by our 2D-fingerprint based machine learning model. Besides, a three-dimensional (3D) pose predictor named MathPose¹⁷ is also applied to predict the 3D binding poses. With these models, we report the top 20 potential anti-SARS-CoV-2 3CL inhibitors from the FDA-approved drugs and another top 20 from investigational or off-market ones. We also discuss the druggability of some potent inhibitors in our training set. The information provides timely guidance for the further development of anti-SARS-CoV-2 drugs.

Results

Binding affinity prediction and ranking of 1553 FDA-approved drugs

With the SARS-CoV-2 3CL protease as the target, we predict the binding affinities of 1553 FDA-approved drugs using our machine learning predictor. Based on these predicted affinities, the top 20 potential SARS-CoV-2 inhibitors from the FDA-approved drugs are shown in Table 1. We also supply the corresponding IC₅₀ (μM) derived from the binding affinity X (kcal/mol) via the following conversion: IC₅₀ = 10^X/1.3633 × 10⁻⁶. A complete list of the predicted values for 1553 FDA-approved drugs is given in the Supporting Material (S2.2).

Table 1:

A summary of the top 20 potential anti-SARS-CoV-2 drugs from 1553 FDA-approved drugs with their predicted binding affinities (unit: kcal/mol), IC₅₀ (μM), and corresponding brand names.

DrugID	Name	Brand name	Predicted binding affinity	IC50
DB01123	Proflavine	Bayer Pessaries, Molca, Septicide	−8.37	0.72
DB01243	Chloroxine	Capitrol	−8.24	0.89
DB08998	Demexiptiline	Deparon, Tinoran	−8.14	1.06
DB00544	Fluorouracil	Adrucil	−8.11	1.11
DB03209	Oteracil	Teysuno	−8.09	1.16
DB13222	Tilbroquinol	Intetrix	−8.08	1.18
DB01136	Carvedilol	Coreg	−8.06	1.22
DB01033	Mercaptopurine	Purinethol	−8.04	1.26
DB08903	Bedaquiline	Sirturo	−8.02	1.29
DB00257	Clotrimazole	Canesten	−8.00	1.35
DB00878	Chlorhexidine	Betasept, Biopatch	−8.00	1.35
DB00666	Nafarelin	Synarel	−8.00	1.35
DB01213	Fomepizole	Antizol	−7.98	1.39
DB01656	Roflumilast	Daxas, Daliresp	−7.97	1.41
DB00676	Benzyl benzoate	Ascabin, Ascabiol, Ascarbin, Tenutex	−7.96	1.45
DB06663	Pasireotide	Signifor	−7.95	1.47
DB08983	Etofibrate	Lipo Merz Retard, Liposec	−7.94	1.48
DB06791	Lanreotide	Somatuline	−7.94	1.48
DB00027	Gramicidin D	Neosporin Ophthalmic	−7.94	1.48
DB00730	Thiabendazole	Mintezol, Tresaderm, and Arbotect	−7.93	1.51

We briefly describe the top 10 predicted potential anti-SARS-CoV-2 drugs from the FDA-approved set. The most potent one is Proflavine, an acriflavine derivative. It is a disinfectant bacteriostatic against many gram-positive bacteria. Proflavine is toxic and carcinogenic in mammals and so it is used only as a surface disinfectant or for treating superficial wounds. Under the circumstance of the SARS-CoV-2, this drug might be used to clean skin or SARSCoV-2 contaminated materials, offering an extra layer of protection. The second drug is Chloroxine, also an antibacterial drug, which is used in infectious diarrhea, disorders of the intestinal microflora, giardiasis, inflammatory bowel disease. It is notable that this drug belongs to the same family with Chloroquine, which was once considered for anti-SARS-CoV-2. However, according to our prediction, Chloroquine is not effective for SARS-CoV-2 3CL protease inhibition (BA: −6.92 kcal/mol). The third one, Demexiptiline, a tricyclic antidepressant, acts primarily as a norepinephrine reuptake inhibitor. The next one, Fluorouracil, is a medication used to treat cancer. By injection into a vein, it is used for colon cancer, esophageal cancer, stomach cancer, pancreatic cancer, breast cancer, and cervical cancer. The fifth drug, Oteracil, is an adjunct to antineoplastic therapy, used to reduce the toxic side effects associated with chemotherapy. The next one, Tilbroquinol, is a medication used in the treatment of intestinal amoebiasis. The seventh drug, Carvedilol, is a medication used to treat high blood pressure, congestive heart failure, and left ventricular dysfunction. The number eight drug, Mercaptopurine, is a medication used for cancer and autoimmune diseases. Specifically, it treats acute lymphocytic leukemia, chronic myeloid leukemia, Crohn’s disease, and ulcerative colitis. The next one is Bedaquiline, which is a medication used to treat active tuberculosis, specifically multi-drug-resistant tuberculosis along with other tuberculosis. The number ten drug, Clotrimazole, is an antifungal medication, which is used to treat vaginal yeast infections, oral thrush, diaper rash, pityriasis versicolor, and types of ringworm including athlete’s foot and jock itch.

Binding affinity prediction and ranking of 7012 investigational or off-market drugs

Using our validated machine elarning model, we present the binding affinity prediction and ranking of 7012 investigational or off-market drugs. We list the top 20 from the investigational or off-market drugs in Table 2. A complete list of the predicted values can be found in the Supporting Material (S2.3).

Table 2:

A summary of top 20 potential anti-SARS-CoV-2 drugs from 7012 investigational or off-market drugs with predicted binding affinities (BAs) (unit: kcal/mol), IC₅₀ (μM), and corresponding trade names.

DrugID	Name	Predicted BA	IC50
DB12903	DEBIO-1347	−9.02	0.24
DB07959	3-(1H-BENZIMIDAZOL-2-YL)-1H-INDAZOLE	−9.01	0.24
DB07301	9H-CARBAZOLE	−8.96	0.27
DB07620	2-[(2,4-DICHLORO-5-METHYLPHENYL)SULFONYL]	−8.89	0.30
	−1,3-DINITRO-5-(TRIFLUOROMETHYL)BENZENE
DB08036	6,7,12,13-tetrahydro-5H-indolo[2,3-a]pyrrolo[3,4-c]carbazol-5-one	−8.89	0.30
DB08440	N-1,10-phenanthrolin-5-ylacetamide	−8.83	0.33
DB01767	Hemi-Babim	−8.80	0.35
DB06828	5-[2-(1H-pyrrol-1-yl)ethoxy]-1H-indole	−8.73	0.39
DB14914	Flortaucipir F-18	−8.69	0.42
DB15033	Flortaucipir	−8.69	0.42
DB13534	Gedocarnil	−8.67	0.44
DB02365	1,10-Phenanthroline	−8.64	0.45
DB09473	Indium In-111 oxyquinoline	−8.64	0.45
DB08512	6-amino-2-[(1-naphthylmethyl)amino]-3,7	−8.60	0.48
	-dihydro-8H-imidazo[4,5-g]quinazolin-8-one
DB01876	Bis(5-Amidino-2-Benzimidazolyl)Methanone	−8.60	0.49
DB07919	7-METHOXY-1-METHYL-9H-BETA-CARBOLINE	−8.59	0.49
DB02089	CP-526423	−8.59	0.50
DB07837	[4-(5-naphthalen-2-yl-1H-pyrrolo[2,3-b]	−8.53	0.55
	pyridin-3-yl)phenyl]acetic acid
DB08073	(2S)-1-(1H-INDOL-3-YL)-3-{[5-(3-METHYL-1H-INDAZOL-5-YL)	−8.53	0.55
	PYRIDIN-3-YL]OXY}PROPAN-2-AMINE
DB08267	6-amino-4-(2-phenylethyl)-1,7-dihydro-8H-imidazo[4,5-g] quinazolin-8-one	−8.52	0.56

Compared FDA-approved drugs, investigational or off-market drugs might be more promising SARS-CoV-2 inhibitors. Among them, DEBIO-1347 has a binding affinity of −9.02 kcal/mol (IC₅₀: 0.24 μM). Another top-ranking drug is 3-(1H-BENZIMIDAZOL-2-YL)-1HINDAZOLE, which has a binding affinity of −9.01 kcal/mol (IC₅₀: 0.24 μM). However, we note that drug discovery is a complex and challenging issue. Having a favorable binding affinity is a necessary but not sufficient condition. Many FDA approved drugs selected for their other characteristics including toxicity, partition coefficient (log P), solubility (log S), synthesizability, pharmacodynamics, pharmacokinetics, etc. These are the issues that prevent many investigational/experimental drugs from becoming the FDA approved ones. Many off-market drugs might have toxicity and/or side-effect issues, in addition to the availability of better alternatives.

Pose prediction

The prediction of binding poses is another important task in drug discovery. The goal pose prediction is to find out the binding-conformation of small-molecule ligands to the appropriate target binding site. The availability of binding poses enables researchers to understand the molecular mechanism of protein-drug interactions and elucidate fundamental biochemical processes. For example, protein-ligand pose and binding affinity predictions are major tasks in D3R Grand Challenges.¹⁷ Molecular docking is one of the most frequently used methods for pose predictions. In this work, utilizing the MathPose developed in recent work,¹⁷ we predict and analyze the binding poses of our predicted top 3 FDA approved drugs and predicted top 3 investigational or off-market ones. More detail of the MathPose is given in the section of Materials and methods. The predicted poses are described in the next section.

Discussion

Analysis of predicted top FDA-approved drugs

Top 3 FDA-approved drugs

The first-ranking candidate from the FDA-approved drugs is Proflavine (see Figure 1(a)), with a predicted binding affinity to the SARS-CoV-2 main protease of −8.37 kcal/mol. The predicted binding pose using our MathPose¹⁷ is illustrated in Figure 1(b). It reveals that there are two hydrogen bonds formed between the drug and the SARS-CoV-2 main protease. The first one is between one amino of Proflavine and the O atom in the main chain of the residue Glu166 of the protease. The second one is between the other amino of the drug and the five-member ring in the side chain of the residue His41 of the protease. As a result, the binding affinity is promising.

Figure 1:

Proflavine, Chloroxine, Demexiptiline and their complexes with SARS-CoV-2 main protease.

The predicted second-best drug is Chloroxine (see Figure 1(c)). Its predicted binding affinity is −8.24 kcal/mol. Between the drug and the protease, there are two hydrogen bonds (see Figure 1(d)): One is formed by the H atom of the hydroxy of the drug with the main-chain O atom of the residue Leu141. The other one is between the hydroxy O atom of the drug and the amino in the main chain of Cys145.

The third one, Demexiptiline (see Figure 1(e)), has a predicted binding affinity of −8.14 kcal/mol. The hydrogen bond between this drug and the protease are formed by the H atom of the amino on the tail of the drug with the side-chain O atom of Ser144. Hydrophobic interactions also play a critical role in the binding.

The binding affinities of the protease-based drugs

It is interesting to analyze the binding affinities of the existing drugs developed as protease inhibitors. Table 3 shows their predicted binding affinities. The predicted values by a recent study¹⁸ are given in the parenthesis, it appears that these values are overestimated. Notably, the current protease inhibitors do not have a substantial effect on the SARS-CoV-2 main protease. A possible reason is that SARS-CoV-2 3CL protease is genetically and structurally different from most other known proteases.

Table 3:

A summary of the predicted binding affinities (BAs) (unit: kcal/mol) and IC₅₀ (μM) of the existing protease inhibitors. Numbers in parenthesis are predictions from the literature.¹⁸

DrugID	Predicted Binding Affinity	IC50	DrugID	Predicted BA	IC50
Remikiren	−7.42	3.57	Moexipril	−6.55	15.63
Candoxatril	−7.22	5.05	Trandolapril	−6.54	17.70
Darunavir	−7.16	5.55	Lopinavir	−6.50	16.92
Isoflurophate	−7.09	6.28	Spirapril	−6.49	17.16
Atazanavir	−7.03 (−9.57)	6.96	Dabigatran etexilate	−6.46	17.96
Argatroban	−7.02	6.98	Apixaban	−6.44	18.84
Sitagliptin	−6.93	8.22	Tipranavir	−6.39	20.36
Fosamprenavir	−6.92	8.26	Lisinopril	−6.35	21.87
Quinapril	−6.91	8.45	Perindopril	−6.34	22.10
Amprenavir	−6.82	9.83	Cilazapril	−6.31	23.36
Benazepril	−6.81	10.05	Ritonavir	−6.26 (−8.47)	25.50
Rivaroxaban	−6.74	11.21	Ximelagatran	−6.24	26.14
Fosinopril	−6.74	11.28	Vildagliptin	−6.15	30.38
Telaprevir	−6.73	11.54	Cilastatin	−6.15	30.40
Captopril	−6.72	11.68	Indinavir	−6.11	32.91
Ramipril	−6.66	12.84	Saxagliptin	−6.07	35.27
Enalapril	−6.66	12.93	Nelfinavir	−6.05	36.23
Alogliptin	−6.62	13.90	Boceprevir	−6.00	39.16
Linagliptin	−6.58	14.73	Simeprevir	−5.77 (−8.29)	58.25
Saquinavir	−6.56	15.26	Ecabet	−5.71	64.15

Comparison to experimental data

In this section, we are interested in comparing our predicted binding affinities to the corresponding experimental ones of some existing drugs outside our training set. Table 4 lists our predictions along with the experimental values of these drugs. These experimental data are extracted from the recent literature.^1,19,20 The RMSE of experimental values and predicted ones is 0.87 kcal/mol, showing a good agreement. It is to point out that all these data were obtained from cell-culture experiments, leading to discrepancies when comparing these experimental values to our results only tailoring to the inhibition of the SARS-CoV-2 3CL protease. For example, the target of Remdesivir is the RNA-dependent RNA polymerase rather than the 3CL protease.

Table 4:

A summary of our predicted binding affinities (BAs) and the corresponding experimental ones of some existing drugs against SARS-CoV-2. All numbers are in kcal/mol unit.

DrugID	Experiment	Prediction	DrugID	Experiment	Prediction
Remdesivir	−6.741	−6.29	Perhexiline	−7.081	−6.67
Chloroquine	−7.001	−6.92	Loperamide	−6.861	−6.98
Lopinavir	−6.871	−6.51	Mefloquine	−7.311	−6.89
Niclosamide	−8.931	−7.66	Amodiaquine	−7.211	−6.93
Proscillaridin	−7.751	−6.50	Phenazopyridine	−6.211	−7.51
Penfluridol	−7.231	−6.54	Clomiphene	−7.191	−7.12
Toremifene	−7.421	−7.20	Digoxin	−9.161	−7.00
Hexachlorophene	−8.241	−7.37	Thioridazine	−7.051	−6.96
Salinomycin	−9.021	−7.00	Pyronaridine	−6.131	−6.68
Ciclesonide	−7.311	−7.04	Ceritinib	−7.561	−6.77
Osimertinib	−7.481	−6.62	Lusutrombopag	−7.391	−6.78
Gilteritinib	−7.051	−5.57	Berbamine	−6.961	−6.87
Ivacaftor	−7.071	−6.74	Mequitazine	−7.001	−6.41
Dronedarone	−7.371	−6.19	Eltrombopag	−6.931	−6.17
Fluphenazine	−7.0819	−6.29	Benztropine	−6.6319	−6.94
Chlorpromazine	−7.5019	−7.00	Terconazole	−6.7119	−7.18
Simeprevir	−6.6720	−5.77	Boceprevir	−7.3420	−6.00
Narlaprevir	−7.1420	−6.38

Analysis of top 3 investigational or off-market drugs

Among the investigational or off-market drugs, the top-ranking candidate is DEBIO-1347 (see Figure 2(a)). Its binding affinity with the SARS-CoV-2 main protease is predicted to be −9.02 kcal/mol. The MathPose predicted pose is illustrated in Figure 2(b). It indicates a hydrogen bond network formed between the drug and the protease leads to the moderately high binding affinity. This network consists of two hydrogen bonds: the first hydrogen bond is between one N atom in the Pyrazole of the drug and the main-chain amino of the residue Glu166 of the protease; the second one is between one N atom in the 1H-1,3-benzodiazole of the drug and the main-chain amino of the residue Gly143 of the protease.

Figure 2:

DEBIO-1347, 3-(1H-BENZIMIDAZOL-2-YL)-1H-INDAZOLE, 9H CARBAZOLE and their complexes with SARS-CoV-2 main protease.

The second-best investigational drug is 3-(1H-BENZIMIDAZOL-2-YL)-1H-INDAZOLE (Figure 2(c)) with a predicted binding affinity of −9.01 kcal/mol. Figure 2(d) reveals that the drug forms two hydrogen bonds with the protease. One is between one N atom in the 1H-1,3-benzodiazole of the drug and the main-chain O atom of the residue Glu166 of the protease. The other is between one N atom in the 1H-indazole of the drug and the main-chain O atom of the residue His164 of the protease.

The third one, 9H-CARBAZOLE (see Figure 2(e)), also has a promising predicted affinity of −8.96 kcal/mol. One can see from Figure 2(f), a strong hydrogen bond is formed between the N atom of the drug and the main-chain O atom of the residue His164 of the protease. The hydrophobic interactions play an essential role in the binding as well.

Analysis of top SARS-CoV/SARS-CoV-2 3CL-protease inhibitors

Note that, in our training set collected from the existing experimental data, 21 samples have binding affinity values lower than −9 kcal/mol. Table 5 provides a list top 20 SARSCoV/SARS-CoV-2 3CL-protease inhibitors with their experimental binding affinities and estimated druggable properties. Moreover, 4 of these 21 samples have 3D experimental structures available. Although these inhibitors are not on the market yet, they serve as good starting points for the design of anti-SARS-CoV-2 drugs. A full list of our training compounds is given in the Supporting Material (S2.1).

Table 5:

A summary of top 20 SARS-CoV-2 main protease inhibitors in the training set with experimental binding affinities (unit: kcal/mol), IC₅₀ (μM), as well as calculated synthesizability, log P and log S.

ID	Binding Affinity	IC50	synthesizability	log P	log S
CHEMBL497141	−11.08	0.01	2.4	2.18	−3.65
PDB ID 2zu4	−10.12	0.04	4.04	2.35	−3.53
CHEMBL222234	−9.95	0.05	2.26	2.66	−3.59
CHEMBL2442057	−9.94	0.05	2.26	5.39	−6.22
CHEMBL213054	−9.92	0.05	4.2	3.15	−3.81
CHEMBL212080	−9.87	0.06	4.25	3.15	−3.76
CHEMBL222840	−9.85	0.06	2.23	2.55	−3.37
CHEMBL398437	−9.85	0.06	2.29	4.12	−5.39
CHEMBL222769	−9.82	0.06	2.16	4.87	−5.73
PDB ID 3avz	−9.80	0.07	4.65	−1.35	−2.33
CHEMBL225515	−9.80	0.07	2.22	3.44	−4.28
CHEMBL1929019	−9.80	0.07	4.23	−0.77	−2.41
CHEMBL222893	−9.57	0.10	2.21	4.17	−5.01
PDB ID 2zu5	−9.56	0.10	4.27	3.79	−4.39
PDB ID 3atw	−9.55	0.10	4.63	−0.46	−2.47
CHEMBL334399	−9.50	0.11	2.20	3.06	−4.17
CHEMBL253905	−9.43	0.12	2.43	4.78	−5.45
CHEMBL403932	−9.42	0.12	1.94	4.11	−4.97
CHEMBL254103	−9.25	0.16	2.10	2.35	−3.34
CHEMBL426898	−9.23	0.17	2.17	3.70	−4.72

Binding interaction analysis of the top 3 experimental structures

Among the SARS-CoV/SARS-CoV-2 3CL-protease complexes with their 3D experimental structures available, the one with the PDB ID 2zu4²¹ is the most potent one with a binding affinity over −10 kcal/mol. This high binding affinity is due to a strong hydrogen bond network between the inhibitor and the protease, which consists of as many as 7 hydrogen bonds. These 7 hydrogen bonds are formed by the inhibitor with protease residues Gln189, Gly143, His163, His164, and Glu166 of the protease.

The second-best 3D-experimental structure is the one with the PDB ID 3avz²² and its binding affinity is −9.80 kcal/mol. A hydrogen bond network, including 7 bonds, plays an essential role in this strong binding. This network is between the inhibitor and protease residues Gln192, Thr190, His164, His163, Glu166, and Gly143.

The PDB ID of the third one is 2zu5²¹ with a binding affinity of −9.56 kcal/mol. A strong hydrogen bond network with 7 bonds can also be found in the structure. The protease residues in the network are Glu166, Phe148, His163, His164, Gly143, and Gln189.

Since His163, His164, and Glu166 emerge in the hydrogen bond networks of all these three structures, it suggests that these three residues are critical to inhibitor binding.

The analysis of log P, log S, and synthesizability of 3 top SARS-CoV/SARSCoV-2 3CL-protease inhibitors

The partition coefficient (log P), aqueous solubility (log S), and synthesizability are also critical medical chemical properties for deciding whether a compound can be a drug or not. Notably, synthesizability is always in terms of synthetic accessibility score (SAS), for which 1 means the easiest, 10 means the hardest. Here, we first calculate the log Ps, log Ss, and SASs of the 1553 FDA-approved drugs (See the Supporting Material S2.2), then investigate whether the three properties of the inhibitors in the top 3 experimental structures are in the preferred ranges of that of the FDA-approved drugs.

According to the log P distribution of the FDA-approved drugs in the Supporting Material (S2.2), the log P interval with a large population of the FDA-approved drugs is between −0.14 and 4.96. The log P values of the top 3 inhibitors are 2.35, −1.35, and −0.46, respectively.

The log S distribution reveals that the preferred range of log S is between −5.12 and 1.76. The log S values of the top 3 inhibitors are −3.53, −2.33, and −4.39.

In the SAS distribution, most of the FDA-approved drugs have SASs between 1.84 and 3.94. The SAS values of the top 3 inhibitors are 4.04, 4.65, and 4.27.

So in summary, for the inhibitor in the first ranking PDB structure 2zu4, its log P and log S are quite good for a drug. The SAS is a little higher, but it is still not so hard to synthesize: 344 of the 1553 FDA-approved drug have larger SASs than this inhibitor, even 56 of them have SASs over 6.

Similarly, for the 3avz and 2zu5 inhibitors, their log Ss are very promising. The log Ps and SASs are some out of the preferred ranges, but still, many FDA-approved drugs have worse log Ps and SASs than theirs. As a result, these top 3 inhibitors, especially the first one, could be good starting points for developing anti-SARS-CoV2 drugs. Obviously, their toxicity will be a major concern for any further development.

Material and methods

The training set of SARS-CoV-2 3CL protease inhibitors

We collect the training set from single-protein experimental data of SARS/SARS-CoV-2 3CL protease in public databases or the related literature.

ChEMBL is a manually curated database of bioactive molecules.²³ Currently, ChEMBL contains more than 2 million compounds only in the SMILES string format. In ChEMBL, we find 277 SARS-CoV or SARS-CoV-2 3CL protease inhibitors with reported Kd/IC50 from single-protein experiments.

Another database is PDBbind. The PDBbind database includes all the protein-ligand complexes with the crystal structures deposited in the Protein Data Bank (PDB) and their binding affinities in the form of Kd, Ki or IC50 reported in literature.²⁴ The newest PDBbind v2019 consists of 17,679 complexes as well as the binding affinities. We find another 30 inhibitors in the PDBbind v2019.

Additionally, binding affinities for four other SARS-CoV main protease inhibitors and three other SARS-CoV-2 main protease inhibitors are extracted from Ref.²⁵ and Ref.,¹⁴ respectively. Therefore, we collected 314 SARS-CoV/SARS-CoV-2 3CL protease inhibitors with available experimental binding affinities.

The binding affinity range in this set is from −3.68 kcal/mol to −11.08 kcal/mol. The distribution is depicted in Figure S3. The top 20 inhibitors in the training set are summarized in Table 5.

The FDA-approved drugs and investigational or off-market ones

DrugBank (www.drugbank.ca)¹⁵ is a richly annotated, freely accessible online database that integrates massive drug, drug target, drug action, and drug interaction information about FDA-approved drugs as well as investigational or off-market ones. Due to the high quality and sufficient information contained in, the DrugBank has become one of the most popular reference drug resources used all over the world. In current work, we extract 1553 FDA-approved drugs and 7012 investigational or off-market ones from DrugBank, and evaluate their binding affinities to the SARS-CoV-2 3CL protease.

The log P, log S and synthesizability calculations

In this work, the log P and synthesizability values are calculated by RDKit;²⁶ the synthesizability in RDkit is in terms of synthetic accessibility score, 1 means the easiest, 10 means the hardest. The log S values are obtained via Alog PS 2.1.²⁷

MathPose

The 3D binding poses in this work are predicted by the MathPose, a 3D pose predictor which converts SMILES strings into 3D poses with references of target molecules. It was the top performer in D3R Grand Challenge 4 in predicting the poses of 24 beta-secretase 1 (BACE) binders.¹⁷ For one SMILES string, around 1000 3D structures can be generated by a common docking software tool, i.e., GLIDE. Moreover, a selected set of known complexes is re-docked by the three docking software packages mentioned above to generate at 100 decoy complexes per input ligand as a machine learning training set. All of those structures are optimized by minimization component in GLIDE with OPLS3 force field.²⁸ The machine learning labels will be the calculated root mean squared deviations (RMSDs) between the decoy and native structures for this training set. Furthermore, MathDL models¹⁷ are set up and applied to select the top-ranked pose for the given ligand.

The machine learning-based binding affinity predictor

In the current work, we develop a machine learning model for predicting the binding affinities of SARS-CoV-2 inhibitors. Our current model is classified as a ligand-based approach, the most popular framework in computer aided drug design owning to its simplicity in the data preparation but still delivering satisfactory performance.^29–32 Since the size of the training set in our current case study is only 314, we apply the gradient boosting decision tree (GBDT) model because of its accuracy for handling small datasets. This GBDT predictor is constructed using the gradient boosting regressor module in scikit-learn (version 0.20.1).

The 2D fingerprints of compounds are used as the input features to our GBDT predictor. Previous study shows that, the consensus of ECFP4, Estate1 and Estate2 fingerprints performs the best on binding-affinity prediction tasks.³³ In this work, we also make use of this consensus. The 2D fingerprints are calculated from SMILES strings using RDKit software (version 2018.09.3).²⁶

The 10-fold cross-validation of the binding affinity predictor

We validate the performance of our machine learning predictor for the 314 inhibitors in the SARS-CoV-2 BA training set. We use 10-fold cross-validation, which is carried out using 50 random splittings. In the Supporting Material S1.4 we show that our machine learning predictor is trained with the average Pearson correlation coefficient (R_p) 0.997, the Kendall’s τ(τ) 0.972, and RMSE 0.095 kcal/mol. These metrics are based on the averaged values across 10 folds and these results indicate our model is well trained. Their averaged test performances across the 10-folds of the whole SARS-CoV-2 BA set are found to be R_p=0.777, τ=0.586, and RMSE=0.792 kcal/mol. These results endorse the reliability of our model in the binding affinity prediction of SARS-CoV-2 inhibitors.

Conclusion

The current pneumonia outbreak caused by severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2), has evolved into a global pandemic. Although currently there is no effective antiviral medicine against the SARS-CoV-2, the 3CL proteases of SARS-CoV-2 and SARS-CoV have a sequence identity of 96.1% and the binding-site RMSD of 0.42 Å, which provides a solid basis for us to hypothesize that all potential anti-SARS-CoV chemotherapies are also effective anti-SARS-CoV-2 molecules. In this work, we curate 314 SARS-CoV-2/SARS-CoV 3CL protease inhibitors with available experimental binding data from various sources to form a machine learning training set. Using this training set, we develop a gradient boosted decision trees (GBDT) model to predict the binding affinities of potential SARS-CoV-2 3CL protease inhibitors. The 10-fold cross-validation shows our model has a Pearson correlation coefficient of 0.78 and a relatively low root mean square error of 0.80 kcal/mol. A total of 8565 drugs from DrugBank is evaluated by their predicted binding affinities. We highlight 20 FDA-approved drugs as well as 20 investigational or off-market drugs as potentially potent medications against SARS-CoV-2. We also analyze the druggability of some potent inhibitors in our training set. This work serves as a foundation for the further experimental development of anti-SARS-CoV-2 drugs.

Figure 3:

The inhibitors and complexes from the top 3 PDBbind structures, 2zu4, 3avz, and 2zu5.