Coronavirus disease 2019 (COVID-19) pandemic caused by the severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) [1
] has become a serious threat to public health management. At least 55 million cases have been reported in more than 210 countries and territories for about eight months since December 2019, and the COVID-19 outbreak is currently in progress (Worldometer.info, 17 November 2020). While it is urgent to develop vaccines and treatments to prevent the spread of viral infections, unfortunately, no effective drug or therapy has been found for COVID-19 yet despite extensive efforts by researchers around the world except remdesivir (GS-5734). Remdesivir, developed as a treatment for Ebola virus disease, has been approved for emergency use in patients with severe COVID-19 by U.S. Food and Drug Administration (FDA) [2
]. More recently, only remdesivir, now commercially named as Veklury, has been approved as the first treatment for COVID-19 by FDA (FDA press release, www.fda.gov
, 22 October 2020). Multinational randomized clinical trials show that remdesivir does not make a significant difference in mortality, but is known to shorten the recovery period for patients [3
]. However, more effective medications are still necessary for patients suffering from symptoms of SARS-CoV-2 infection.
Our group previously predicted several antiviral (atazanavir, remdesivir, lopinavir/ritonvir) and non-antiviral drugs (rapamycin, tiotropium) that might inhibit SARS-CoV-2 using a new deep drug–target interaction (DTI) prediction model called Molecule Transformer (MT)-DTI [4
]. In the previous analysis, these drugs showed strong binding affinity values for viral proteins of SARS-CoV-2. Indeed, several antiviral drugs, such as atazanavir and remdesivir, predicted by MT-DTI, showed potent in vitro and in vivo inhibitory effects on SARS-CoV-2 infection [5
]. Recently, we further examined that top-ranked non-antiviral drug, tiotropium bromide, has an inhibitory effect on some of the SARS-CoV-2-induced genes through transcriptome-based analysis [7
]. While analyzing tiotropium bromide, it was suggested that there is a possibility of more therapeutic targets and options for COVID-19 that may be predicted by deep learning technology to fight against the on-going pandemic. To carry out the task, angiotensin-converting enzyme 2 (ACE2) and transmembrane protease serine subtype 2 (TMPRSS2) that the most studied host targets at the moment, were selected.
SARS-CoV-2 uses the ACE2 receptor and TMPRSS2 for infection into host cells [8
]. ACE2, which serves as a binding receptor of the spike protein of SARS-CoV-2, involves in blood pressure regulation and cardiovascular regulation through enzymatic hydrolysis of angiotensin (Ang) II to Ang (1–7). Ang II cleavage by ACE2 counteracts the function of ACE that promotes Ang II formation [10
]. Interestingly, since ACE2 is already known as a binding receptor for SARS-CoV [12
], it supports that SARS-CoV-2, which maintains a very close evolutionary relationship with SARS-CoV, also utilizes ACE2 for host cell entry. Another proposed target protein, TMPRSS2, is a cell membrane-bound serine protease [13
] which function is yet unclear since its identification, but the involvement of TMPRSS2 during SARS-CoV-2 entry is suggested as the host protease for the viral spike-ACE2 complex proteolytic activation [8
Due to the role of ACE2 and TMPRSS2, which are essential for virus entry, they have been proposed as host targets to block SARS-CoV-2 entry [14
]. Therefore, in this study, we predicted drug candidates that can control these two potential targets by using a deep-learning approach.
To identify potent FDA-approved drugs capable of inhibiting the entry of SARS-CoV-2, we performed an in silico screening method using the MT-DTI deep learning-based model [4
]. The MT-DTI model can predict the binding affinities of Kd
values from on chemical sequences (SMILES) and amino acid sequences (FASTA) of a target protein. The sequence-based drug-target affinity prediction approach predicted drugs with a strong binding affinity (Kd
< 100 nM) against ACE2 and TMPRSS2 from 1400 FDA-approved drugs (ACE2, Table 1
and Table S1
; TMPRSS2, Table 1
and Table S2
). The NCATS OpenData portal, which was developed to provide insight into drug repurposing for the treatment of COVID-19 (https://opendata.ncats.nih.gov/covid19
], provides drug results that affect protein–protein interaction between ACE2 and RBD (Receptor-binding domain) of the SARS-CoV-2 S protein through AlphaLISA and TruHit (counterscreen) ACE2-Spike protein–protein interaction proximity assay. We first compared the results obtained from the MT-DTI model with the drugs provided by the OpenData portal. Of the 460 drugs that may bind ACE2 with strong affinity, 17 drugs showed the interference of the interaction of ACE2 with SARS-CoV-2 spike protein (Table 2
). Although AC50
values of the proximity assay are not affinity values, submicromolar to one-digit micromolar AC50
results supports the ACE2 affinity prediction of 17 selected drugs. Interestingly, enalaprilat, an ACE inhibitor, was identified as a chemical compound that has a strong affinity for ACE2 and is likely to inhibit the entry of SARS-CoV-2. According to previous results, the angiotensin-converting enzyme (ACE) and its homologous ACE2 both belong to the ACE family of dipeptidyl carboxypeptidase but provide the opposite effect for the renin-angiotensin system. ACE produces Ang II, which induces enhanced inflammation, elevated blood pressure, and increased coagulation. However, ACE2 is responsible for the breakdown and inactivation of Ang II [10
]. Based on our predictions, enalaprilat suggests that it is a promising drug candidate that can reduce inflammation and blood pressure by inhibiting ACE activity and simultaneously inhibit the entry of SARS-Cov-2 through interaction with ACE2. Interestingly, venetoclax (Kd
6.12 nM), posaconazole (Kd
17.11 nM), daclatasvir (Kd
6.65 nM), and ombitasvir (Kd
5.91 nM) also predicted to have strong affinity for TMPRSS2, as shown in Table S2
. In particular, daclatasvir and ombitasvir are hepatitis C virus (HCV) inhibitors, suggesting that these compounds may exhibit antiviral effects on HCV and SARS-CoV-2.
Another strategy to block the entry of SARS-CoV-2 by targeting ACE2 is to increase the human recombinant soluble ACE2 (hrsACE2), thereby inhibiting the membrane association of host cells with SARS-CoV-2 [27
]. Indeed, hrsACE2 has already undergone phase 1 and phase 2 clinical trials to treat acute respiratory distress syndrome and is now considered a candidate for treatment for COVID-19. In the cell, the soluble form of ACE2 is endogenously generated through ACE2 ectodomain shedding by the disintegrin metalloproteinase 17 (ADAM-17) [28
]. On the other hand, calmodulin inhibits ectodomain shedding by interacting with ACE2 [29
]. Therefore, inhibition of calmodulin-ACE2 interaction through the calmodulin antagonists may be an effective strategy to reduce the infectivity of the virus. We predicted that the calmodulin antagonist, the oxide of the amitriptyline, binds to ACE2 with a Kd
of 40.38 nM (Table S1
). However, it requires experimental verification that the drug’s interaction with ACE2 further promotes ACE2 shedding.
To compare MT-DTI affinity prediction results to widely used docking study tool AutoDock Vina, Gibbs free energy (∆G) values of each drug-target pair were predicted. For the ACE2 cross-prediction result, ∆G −8.0 kcal/mol is given as threshold, and 27 drugs show correlative prediction results out of 460 DTI predicted results (Table 3
, left). The TMPRSS2 cross-prediction result was screened with ∆G −7.0 kcal/mol threshold as the numbers of predicted drugs are less than 100, and 11 out of 75 drugs showed correlation to the MT-DTI prediction (Table 3
, right). However, TMPRSS2 docking results through a homology modeled TMPRSS2 may not be precise as ACE2 due to the absence of an actual 3D structure, therefore, it should be considered again when crystallography results are available. All ∆G values predicted through AutoDock Vina are annotated in Tables S1 and S2
We have reported in the previous MT-DTI studies that tiotropium bromide has a strong binding affinity to the SARS-CoV-2 viral proteins [4
]. A recent RNA-seq analysis of normal bronchial epithelium (NHBE) cells infected with SARS-CoV-2 has shown that the treatment of tiotropium can have an advantageous prescription effect [7
]. Interestingly, tiotropium bromide was also predicted to strongly bind with ACE2 (Kd
0.92 nM) in the present study, suggesting that the drug is a promising candidate for COVID-19.
We further investigated 75 drugs predicted to have an affinity of less than 100 nM for TMPRSS2 that is a serine protease for S protein priming of SARS-CoV-2. Clinically proven protease inhibitors, including bromhexine, aprotinin, camostat, and nafamostat, have been suggested as potential treatment options for COVID-19 [30
]. However, the MT-DTI results for TMPRSS2 showed affinity of more than 100 nM for these protease inhibitors. Intriguingly, we found that five of the top 30 drugs that strongly bind to TMPRSS2 are antiviral drugs with a Kd
value of less than 20 nM (Table S2
). Three out of five (ombitasvir, daclatasvir, and paritaprevir) are anti-hepatitis C virus (HCV) drugs, and recently, there are reports that antiviral agents used to treat HCV could be another option for SARS-CoV-2 treatment [32
]. Although HCV and coronavirus are not closely related viruses, it can be suggested that those drugs may act on life cycles of both viruses in a mode of action perspective as they are single-stranded RNA viruses. These three anti-HCV drugs predicted to have Kd
values of less than 10 nM for TMPRSS2, implicating a strong inhibitory potency (Table 1
To screen for more effective drugs that can inhibit TMPRSS2, we used the connectivity map (CMap) database [33
] that allows gene-expression signature matching between the drugs proposed as TMPRSS2 inhibitors and the top 30 drugs of the MT-DTI result. CMap is a database that provides transcriptome profiling information caused through genetic and chemical perturbations [33
]. We compared gene expression changes by the previously proposed TMPRSS2 inhibitors (Bromhexine [34
] and Probucol [35
]) and the top 30 drugs of MT-DTI through CMap Linked User Environment (CLUE, https://clue.io
) (Table 4
). Scores above +90 indicate a high positive correlation between the gene-expression signature of a given perturbagen and the gene-expression signature by the query. Among the top 30 drugs, the gene expression by eprosartan and lisuride was strongly correlated with the gene expression signatures of bromhexine and probucol, respectively. Interestingly, eprosartan is a type I angiotensin II receptor (AT1R) blocker, preventing Ang II from binding to the AT1R receptor in the renin-angiotensin system [36
]. As the binding of Ang II and AT1R causes fibrosis and damage of lung tissues, AT1R blocker, eprosartan, is considered a suitable therapeutic target for COVID-19 [35
]. In addition to eprosartan, neuropsychiatric agents, lisuride, have recently been predicted as a candidate drug for treating COVID-19 through coexpression-based drug enrichment analysis of COVID-19 induced genes [37
An AI-integrated drug repurposing approach, such as MT-DTI, is a promising solution for the development of potential drugs that can overcome life-threatening diseases caused by COVID-19. For example, atazanavir, predicted by MT-DTI in the previous study [4
], is registered in the REVOLUTIOn trial (NCT04468087) for the treatment of COVID-19. As a case study for ACE2 affinity prediction, a recent report on a clinical trial RECOVERY (NCT04381936) suggested dexamethasone, an anti-inflammatory corticosteroid, as a possible treatment option for COVID-19 [38
]. Interestingly, dexamethasone was predicted to have an affinity of Kd
9.50 nM to ACE2 (Table S1
). In addition, approved drugs sharing a similar polycyclic core structure of steroids such as obeticholic acid (Kd
0.98 nM), mestranol (Kd
1.02 nM), norethynodrel (Kd
1.51 nM), clobetasol (Kd
1.80 nM), norethisterone (Kd
1.92 nM), fluoxymesterone (Kd
2.29 nM), cholic acid (Kd
2.86 nM), and more were listed on top in the ACE2 affinity prediction. However, original indications and side effects must be considered since steroids are involved in hormonal regulations.
Furthermore, the MT-DTI prediction resulted that ACE inhibitors, including enalaprilat, zofenopril, lisinopril, benazepril, trandolapril, cilazapril, perindopril, ramipril, fosinopril, moexipril, spirapril, have a strong binding affinity of less than Kd
70 nM for ACE2 (Table S1
). Although there is concern that inhibition of ACE2 activity may promote an excessive inflammatory response resulting from an increase in Ang II [40
], ACE inhibitors can be considered as a treatment option for COVID-19 in two respects. First of all, inhibition of ACE function that elicits an immune response via Ang II may alleviate the side effects of loss of ACE2 function by ACE inhibitors. Another is that ACE inhibitors do not inhibit the activity of ACE2, but rather interfere with the interaction between ACE2 and the viral S protein. Indeed, ACE2 is not only known to be insensitive to ACE inhibitors [41
], but the NCATS OpenData portal provides experimental results that the enalapril maleate disrupts the protein–protein interaction between ACE2 and S protein with an AC50 of 7.5 µM. Therefore, our results suggest that the ACE inhibitor is a promising therapeutic option.
We further identified three anti-HCV drugs, including ombitasvir, daclatasvir, and paritaprevir, in the top 30 predictive candidates for TMPRSS2. Another group recently reported that anti-HCV reagents were potential drug repurposing candidates for COVID-19 [42
], supporting the results. Above all, we were able to present more effective drug candidates by merging AI-based prediction results and CMap-based drug repurposing approach in this study. Eprosartan and lisuride, drugs predicted to have overall strong binding affinity values for TMPRSS2, demonstrated that they could effectively control TMPRSS2 by showing gene expression signatures similar to the previously proposed TMPRSS2 inhibitors (bromhexine, a mucolytic drug for respiratory diseases; probucol, an anti-hyperlipidemic drug). Bromhexine and probucol are known to block TMPRSS2, but both drugs are currently used as treatments for different indications and have different mode of actions other than TMPRSS2 inhibition. Consistent with this, the connectivity score between the two previously proposed TMPRSS2 inhibitors is not high. Therefore, the drugs predicted by MT-DTI do not seem to show a strong correlation for both drugs simultaneously. Interestingly, the dasatinib, which was predicted to have the strongest affinity for TMPRSS2, had little correlation with bromhexine and probucol implicating that chemical and transcriptional influence of dasatinib may be different compared to both formerly proposed TMPRSS2 inhibitors. It suggests that CMap analysis is a useful approach for filtering out possible false-positive candidates that have opposite effect or little biological effect from DTI prediction results. However, at least, any positive connectivity score, better with higher positive connectivity score, should be considered and may guide researchers to pick drugs with priority for drug repurposing.
Finally, to go deeper to target-centered drug repurposing, drugs appearing both in ACE2 and TMPRSS2 such as aclidinium, buprenorphine, emtricitabine, lurasidone, and tiotropium may be investigated as dual targeting drugs in polypharmacological perspective [43
]. Therefore, we hope that our results can contribute to the therapeutic management of COVID-19 through evaluation efficacy and safety in vitro and in vivo studies.
Taken together, the drugs resulted through the MT-DTI affinity prediction for ACE2 and TMPRSS2 are suggested promising drug candidates for an effective drug repurposing strategy to treat COVID-19 in a target-wise perspective.