key: cord-0941297-c3pwn0zg
authors: Gorgulla, Christoph; Padmanabha Das, Krishna M.; Leigh, Kendra E.; Cespugli, Marco; Fischer, Patrick D.; Wang, Zi-Fu; Tesseyre, Guilhem; Pandita, Shreya; Shnapir, Alec; Calderaio, Anthony; Gechev, Minko; Rose, Alexander; Lewis, Noam; Hutcheson, Colin; Yaffe, Erez; Luxenburg, Roni; Herce, Henry D.; Durmaz, Vedat; Halazonetis, Thanos D.; Fackeldey, Konstantin; Patten, Justin J.; Chuprina, Alexander; Dziuba, Igor; Plekhova, Alla; Moroz, Yurii; Radchenko, Dmytro; Tarkhanova, Olga; Yavnyuk, Irina; Gruber, Christian; Yust, Ryan; Payne, Dave; Näär, Anders M.; Namchuk, Mark N.; Davey, Robert A.; Wagner, Gerhard; Kinney, Jamie; Arthanari, Haribabu
title: A multi-pronged approach targeting SARS-CoV-2 proteins using ultra-large virtual screening
date: 2021-01-05
journal: iScience
DOI: 10.1016/j.isci.2020.102021
sha: 2892dc848ef67cb3b36c73690457c8acd36c4440
doc_id: 941297
cord_uid: c3pwn0zg

The unparalleled global effort to combat the continuing severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) pandemic over the last year has resulted in promising prophylactic measures. However, a need still exists for cheap, effective therapeutics and targeting multiple points in the viral life cycle could help tackle the current as well as future coronaviruses. Here we leverage our recently developed, ultra-large scale in silico screening platform, VirtualFlow, to search for inhibitors that target SARS-CoV-2. In this unprecedented structure-based virtual campaign, we screened roughly 1 billion molecules against each of 40 different target sites on 17 different potential viral and host targets. In addition to targeting the active sites of viral enzymes, we also targeted critical auxiliary sites such as functionally important protein-protein interactions.

proteins, and ii) drugs that target host proteins, including the viral receptor, cellular proteins essential to the viral life cycle, and components of the immune system.

Currently available viral therapeutics include protease inhibitors, reverse transcriptase and polymerase inhibitors (both nucleoside analogues and non-nucleoside inhibitors), integrase inhibitors, and viral entry blockers ( Figure S46 ). Although there are no drugs approved by the FDA specifically for SARS-CoV, MERS-CoV or SARS-CoV-2, there have been a number of efforts to target proteins of the viral machinery, including the proteases, helicase and RNA-dependent RNA polymerase (RdRP). A partial list of these published small molecules is shown in Table S4 . A well publicized example, Remdesivir, is an RdRP inhibitor originally developed for hepatitis C that was also previously tested against both SARS-CoV and MERS-CoV. It should be noted that subsequent studies on its use in the treatment of COVID-19 have yielded conflicting results, with some suggesting that Remdesivir may reduce the time to recovery in SARS-CoV-2 patients (37, 38) , while others showed no appreciable effect (39) .

For every virus, there are targets such as the entry protein or the viral polymerase, which have direct effects on the ability of the virus to replicate, even under ideal conditions for the virus. There are also other targets that, if inhibited, would reduce the severity of the viral infection, often referred to as the viral virulence. Some of these targets may only have an effect when inhibitors are evaluated in the context of a whole organism. For example, several viruses paralyze the host immune system by hijacking or sequestering key signaling proteins, and targeting the viral proteins that mediate this effect would reduce the virulence of the virus. In this study we targeted an array of SARS-CoV-2 proteins in addition to two human proteins, and we posited that inhibitors, which target proteins that affect the viability and/or the virulence, could synergistically combine for more efficacious combination therapies.

In silico screening methods enabled us to identify lead drug candidates for SARS-CoV-2 in an expedited manner, fueled by the availability of high-resolution structures of many of the SARS-CoV-2 proteins. Previous work has shown that the number of compounds screened in a structure-based in silico screen affects the quality of the resulting hits and the potency of the hits derived increases with the number of compounds screened (40) . A flurry of recent papers have demonstrated that by using in silico screening to expand the chemical space (screening~100 million compounds), one can identify picomolar-affinity inhibitors . To become capable of screening billions of molecules in a relatively short period (weeks), we previously developed VirtualFlow, an open source computational drug discovery platform (43) . VirtualFlow, with its linear scaling feature, leverages the power of computing clusters to perform ultra-large scale screens. Here we used the computational power of the Google Cloud Platform (GCP) to identify molecules that could inhibit SARS-CoV-2. We screened over one billion molecules against each of fifteen SARS-CoV-2 proteins and two human proteins, ACE2 and TMPRSS2. For some of the proteins we target multiple functional sites on the protein surface; an approach which provides two distinct advantages: i) the possibility of increasing efficacy using a cocktail approach, and ii) the identification of alternative ways to target the same protein, a necessity when faced with the possibility of resistance mutations. Whenever available, we used recently determined high-resolution structures of SARS-CoV-2 proteins in our screens. A complete list of the target proteins along with the models used can be found in 1.

When performing in silico screening, the probability of finding a true hit (a hit that is validated experimentally) increases with the number of molecules screened and the accuracy of the chosen structure relative to what exists in solution inside the cell. Small changes in structure or dynamics can be accommodated by flexible docking, but large conformational changes decrease the ability of in silico screening methods to identify potent binders. Large conformational changes can be tackled by ensemble docking methods (molecular dynamics simulations), but the trade-off for these approaches is additional computational time. Targeting protein-protein interactions (PPI) is also challenging for in silico screening. Typically with PPIs the off-rate between the two interacting proteins is slow and the interaction interface is large compared to a small molecule inhibitor. Thus targeting PPIs was approached in this study with the rationale that a small molecule engaging the monomer at the hot spot of the PPI interface could prevent complex formation, particularly if driven by excess of the small molecule. In the absence of a solved dimeric/multimeric structure, we made the assumption that the structure of the monomer is similar to that of the protein in complex, a fact which may not always hold true. In light of such caveats, we describe any potential limitations that we anticipate could affect the experimental outcome for each of the targets that were screened, as well as generally for the study as a whole.

The spike protein forms the highly glycosylated trimeric receptor-binding protein that decorates the virion surface and facilitates entry into the host cell through interaction with its receptor ACE2 (44) .

Binding of the receptor binding domain (RBD) of S to ACE2 induces large conformational changes in the S2 domain. These changes are followed by processing by host proteases such as TMPRSS2 and cathepsins and result in the formation of a stable six-helix bundle by heptad repeats 1 and 2 (HR1 and HR2) culminating in the fusion of the viral and cellular membranes (8) . In order to develop compounds capable of preventing SARS-CoV-2 from entering cells, we focused on the following three targets: a) the interaction interface of the RBD and ACE2, b) the host protease TMPRSS2, and c) the HR2 hydrophobic binding groove of HR1.

Recent studies have revealed molecular details of ACE2 binding with spike-RBD and multiple highresolution structures of the RBD in complex with ACE2 have been released (PDB IDs: 6m17, 6m0j, 6vw1) (16, 45, 46) . As the primary mediator of host cell attachment, the spike trimer is a clear potential target for therapeutic intervention, and in order to find an effective inhibitor of viral entry into host cells, we targeted both ACE2 and the RBD of spike in parallel.

Targeting the spike protein interaction interface on ACE2

The primary physiological function of the receptor, ACE2, is the maturation of angiotensin, a signalling peptide with cardiovascular regulatory effects (47) , making any targeting of the catalytic site of ACE2 unfavourable. Fortunately, the structure of the ACE2-RBD complex shows that the peptidase (PD) domain of ACE2 has two distinct lobes that are essential for substrate engagement, and while the RBD-ACE2 binding interface includes one of these lobes, it does not obstruct the PD active site (16, 45) (Figures 2, S32 , S33, S34, S35). This suggests that targeting the RBD interaction interface should not inhibit ACE2 functionality and therefore not affect angiotensin homeostasis. In fact, previous biochemical studies with SARS-CoV have shown that the ACE2 inhibitor MLN-4760, which favors the closed state, does not interfere with S protein binding and vice versa, also arguing for the independence of ACE2 function and spike binding (48) .

We have targeted four different sites on the ACE2 surface. Two of them are located on the spike interaction interface, and would directly disrupt the ACE2-spike interaction (Figures 2, S32, S33). For the first of these two sites (centered around Glu37), two independent virtual screens were carried out (Screen IDs: 1 and 2): one using an ACE2 structure where no spike protein was bound (PDB ID: 6m17; Screen ID: 1), and one using an ACE2 structure to which the RBD domain of the spike protein was bound (PDB ID: 6m18; Screen ID: 2). The difference between these two conformations at the site of binding is relatively small (RMSD: 0.355 Å in the region of the first two N-terminal -helices), yet still significant for structure-based virtual screening. The second site that we targeted on the spike binding interface (Screen ID: 3) is centered around residue Gly354, and was chosen because there is a shallow hydrophobic pocket. The other two sites targeted were two dynamic pockets adjacent to the spike interface on ACE2 potentially capable of accommodating tighter binders (Screen IDs: 4 and 5; Figures S32, S33 ). In addition, compounds that bind to the two dynamic pockets could be chemically linked to each other in such a way that the linker, which then would pass over the spike interaction interface, could potentially block the binding of spike protein. In these two screens, we used the conformations of ACE2 derived from molecular dynamics (MD) simulations conducted by D.E. Shaw Research (49) . In these MD simulations ACE2 samples an open state and we have targeted two dynamic/cryptic pockets that were revealed in the MD simulations. These conformations are not confirmed by experimental methods and thus are of exploratory nature and hits from these screens can be considered high-risk high-reward.

In vitro measurements suggest that SARS-CoV-2 RBD binds to ACE2 with nanomolar affinity (17, 50) . Each peptidase domain of ACE2 can accommodate one RBD and the interaction is mediated through polar interactions. Furthermore, it has previously been shown that the "up" conformation of the RBD is required for receptor binding and that two spike protein trimers can bind to a single ACE2 dimer (16) (Figure S2 ). The RBD itself consists of a core containing an anti-parallel β-sheet (β 1 to β 4 and β 7), with three short -helices and an extended loop. This extended loop engages the α 1 helix of the ACE2-PD, and it is the center of this interface on the RBD that we have targeted by virtual screening (Screen ID: 7) ( Figure S2 ).

The spike interaction interface on ACE2 is relatively flat and polar, which makes it challenging to identify molecules that bind to it with high affinity. This difficulty is reflected in the docking scores resulting from the virtual screen of this binding interface: the average docking score of the top 100 virtual hit compounds from this screen was approximately -8 kcal/mol, which is amongst the weakest docking score averages of all the target sites screened ( Figure S45 ).

The SARS-CoV-2 RBD has significantly higher affinity for ACE2 as compared to that of the SARS-CoV RBD (46, 50) ; however, the full-length spike protein from SARS-CoV-2 shows either similar or weaker affinity for ACE2 (17, 51) when compared to that of the full-length SARS-CoV spike. This contrast in binding affinities between the RBD alone and the full-length spike protein could be a consequence of RBD dynamics where the protein switches between RBD "up" and "down" states, as is also seen in the case of SARS-CoV (52, 53) . There is limited information available as to whether the SARS-CoV-2 spike favors a 'lying-down'binding-inactive state or a 'standing up' state capable of receptor binding (17, 50) . Furthermore, while structural studies have looked at conformational flexibility in the spike protein (54, 55) , there is little information on the inherent dynamics within the RBD domain itself, which is where the targeted molecules are expected to bind. This means that it is possible that local protein dynamics could adversely affect binding of some of the hits in ways that the screen is unable to predict.

It should be noted that while targeting the flat surface of the spike-binding interface with ACE2 (Screen IDs 1-3) resulted in compounds with an average binding energy of -8.3 kcal/mol, targeting the dynamic pockets resulted in hits with the potential for higher affinity binding. The binding energy of hits in the dynamic pockets was as high as -10.3 kcal/mol. However, experimental validation of these hits would be required to confirm this apparent difference in affinity.

The essential priming of S during entry can be executed by the host serine protease TMPRSS2 in the case of SARS-CoV-2, making it a potential therapeutic target. In addition, recent research has shown that the TMPRSS2 inhibtor, camostat mesylate, can block viral entry in cell-based assays (44) . Effective inhibition of TMPRSS2 could have multiple applications, since processing of ACE2 by TMPRSS2 has also been shown to be crucial for activation of S in SARS-CoV (56) . Although TMPRSS2 has been shown to activate protease-activated receptor 2 (PAR-2) and hepatocyte growth factor (HGF), giving it possible roles in tumor metastasis and epithelial-mesenchymal transition (57), the exact physiological function of this protein is still unknown. TMPRSS2-deficient mice showed no significant phenotypic changes (58) , and thus TMPRSS2 is considered to be dispensable for normal cellular function, making a TMPRSS2-specific inhibitor less likely to have side effects.

TMPRSS2 itself contains a cytoplasmic domain, a transmembrane domain, an extracellular lowdensity-lipoprotein-receptor (LDLR) domain, a scavenger-receptor cysteine-rich (SRCR) domain and a C-terminal serine protease domain. The TMPRSS2 precursor consists of 492 amino acids and autocatalytic cleavage at Arg255 generates the active protease (59) . As a chymotrypsinogen-like protease, TMPRSS2 contains a catalytic triad consisting of residues His296, Asp345 and Ser441. As there are currently no experimental structures available for TMPRSS2, we used a homology model from SWISS-MODEL that is based on the human serine protease hepsin (PDB ID: 5ce1), which shares 33.82% sequence identity with TMPRSS2 (60) . Approximately 1 billion compounds were screened against the active site of this homology model ( Figure S1 ). Since a homology model was used to screen for inhibitors, success would depend on the how closely the model represents the actual structure. The active site of TMPRSS2 is well conserved when compared to the template, the serine protease hepsin , and also contains a catalytic triad consisting of Ser, His and Asp residues ( Figure S42 ). The residues around the binding pocket are also well conserved, making it more likely that the predicted structure is sufficiently similar to the true structure in the region that we targeted around the active site binding pocket (Screen ID: 6). Our top 10 hits had an average binding energy of -11.3 kcal/mol, which should ideally correspond to nanomolar binding affinities. It must be noted that experiments with the TMPRSS2 inhibitor camostst mesylate in cultured cells showed that complete inhibition of viral entry is achieved only in conjunction with an inhibitor of cathepsins B and L (CatB/L), E-64d (44) ; however, the impact of this residual priming of S by CatB/L on viral pathogenesis in vivo is yet to be determined.

While the S1 subunit of the spike protein is crucial for host cell attachment, the S2 subunit plays an essential role in membrane fusion and subsequent entry into the host cell. Heptad repeats HR1 (residues 912-984) and HR2 (residues 1163-1213) in S2 interact to form the classic six-helix bundle (6-HB) of a class I entry protein (61, 62) . The HR domains of the spike protein and their mode of interaction are known to be highly conserved across CoVs, making them an attractive target for the development of pan-coronavirus fusion inhibitors (63) .

The central trimeric coiled-coil of HR1 has three hydrophobic grooves ( Figure S3 ) that are proposed to be potential drug binding sites for inhibition of viral fusion. The recent crystal structure of HR1-L6-HR2 (PDB ID: 6lxt) displays a 6-HB structure with the HR1 domains forming a parallel trimeric coiled-coil around the anti-parallel HR2 domain. The hydrophobic groove formed by the HR1 helices is also the binding interface for the HR2 domain (64) . The hydrophobic interaction between HR1 and HR2 is in the area of the fusion core, and previous studies have shown that both HR1 and HR2 components of this fusion core interface could be effectively targeted using complementary peptides in SARS-CoV-2 and MERS-CoV (61, 65) .

We performed a blind docking using QuickVina W against the entire surface of the trimeric alpha helix (Screen ID: 8; Figure S3 ). The resulting hits could not only potentially have high inhibitory activity against multiples CoVs, but could also have both prophylactic and therapeutic applications as was seen previously in the case of a peptide inhibitor (66) . Ligands binding to each of the three different hydrophobic grooves could be identified and linked together in order to enhance the overall affinity and thereby increase the potency.

Once coronavirus enters the cell, translation of the replicase gene gives rise to two large polyproteins, PP1a and PP1ab, composed of nsp1-11 and nsp1-16, respectively. M pro and PL pro cleave these polyproteins into individual nsps, which facilitate genomic replication and transcription of subgenomic RNAs. In order to develop small molecules capable of inhibiting replication we targeted the following five proteins and protein complexes, selected based on the availability of high-resolution structures and available information about their molecular mechanisms: a) the coronavirus primary protease (M pro , nsp5), b) the papain-like protease (PL pro , nsp3), c) the RNA-replication complex, composed of the RdRp (nsp12) and its co-factors nsp7 and nsp8, d) the helicase (nsp13), and e) the ssRNA-binding protein nsp9.

M pro Mpro (3CL pro 3CLpro) M pro (3CL pro ), the main protease of coronavirus, cleaves PP1a and PP1ab into many of their constituent nsps (11 cleavage sites in PP1ab). The inhibition of M pro would not only inhibit the protease itself, but also hinder downstream processes by preventing the production of key viral proteins through inhibition of their proteolytic processing. Proteolytic cleavage by M pro via the catalytic dyad (Cys-His) occurs at the P1 glutamine residue in the cleavage motif (67) . Dimerization is thought to create a substrate binding cleft (68) , and simulations suggest that only one monomer is active at a time (69) . M pro is highly conserved not only in SARS-CoV and SARS-CoV-2, but also among other viruses of the Nidovirales order (enveloped positive-sense single-stranded RNA viruses that produce nested 3'-co-terminal genomic RNAs), making it a potential target for the development of antiviral drugs with broad efficacy (70) .

Using the same strategies that allowed crystallization of SARS-CoV M pro (71) , several groups determined the three-dimensional structure of SARS-CoV-2 M pro in its apo form or bound to various inhibitors (72) (73) (74) (75) (76) (77) (78) . These structures revealed that the N-terminus of SARS-CoV-2 M pro contains two domains (domain I, residues 10-99; domain II, residues 100-184) that adopt a chymotrypsin-like fold, whereas the M pro C-terminal Domain III (residues 201-303) is composed of five α-helices and is responsible for the intermolecular interactions critical for the dimerization. The protease active site of the picornavirus 3C-protease-like domain is located at the cleft formed between domains I and II and contains a Cys145-His41 catalytic dyad. A total of five protein segments comprised of residues 25-27 and 44-50 from domain I and residues 140-143, 165-168 and 188-190 from domain II form the walls of the active site ( Figure S41a ).

It is important to note that the M pro active site seems to have significant conformational flexibility (79) . Comparison of three apo structures and five structures in complex with inhibitors reveals changes in the conformation and position of the Gln189-containing loop and the short Ser46containing α-helix ( Figure S41b ). Specifically, the cleft between domains I and II is narrower in the apo structures than in the structures of M pro bound to inhibitors, consistent with an induced-fit or conformational selection upon ligand binding.

Given the plasticity of the active site, we used two different conformations of M pro for in silico screening (Screen IDs: 16 and 17; Figures S4, S5 ). The first conformation corresponded to the structure of M pro bound to the peptide-like inhibitor N3 (73) . For this search, the protein model used was identical to the protein structure reported in the PDB file, after removal of all water molecules and bound ligands (Screen ID: 16; Figure S41c ). The second conformation corresponded to the structure of M pro bound to inhibitor 11b, an inhibitor with peptide-like features (74) . However, the model used for screening differed slightly from the conformation reported in the PDB file. In the crystal structure, the side chain of Met49 projects into the active site and thus would potentially interfere with compound docking. To avoid this, a different Met49 side chain rotamer was used that did not project as much into the active site. This change resulted in steric clashes with the side chain of Ser46, for which selection of a different side chain rotamer resolved these incompatibilities. Finally, a rotamer for Cys145 was selected that results in its side chain pointing towards the core of the protein. This rotamer was selected to allow the possible docking of compounds that could form covalent bonds with the Cys145 thiol group ( Figure S41d ).

The two conformations screened, hereafter referred to as M pro -N3 (Screen ID: 16) and 6m0k* (Screen ID: 17), respectively( Figure S41 ), identified compounds that docked with high binding energies. Interestingly, the 6m0k* conformation resulted in predicted binding free energies that were approximately 0.5 kcal/mol lower than the binding free energies (docking scores) obtained from screening the M pro -N3 conformation. Moreover, amongst the top 1000 hits identified by screening each of the two conformations (with the application of the standard filtering criteria of logP less than 6, molecular weight less than 600 daltons, and no reactive groups), only 12 hits overlapped. These results illustrate how screening closely related protein conformations can lead to the identification of a diverse set of docked compounds and this may be a generalizable strategy for in silico screening of proteins that exhibit conformational flexibility, such as SARS-CoV-2 M pro (79) .

Since the dimerization of M pro is known to be essential for its catalytic activity (80) (81) (82) , we also targeted the dimerization interface, which relies primarily on the α1 helix of the C-terminal domain (Screen ID: 18) . Previous studies have shown that the inter-conversion of M pro from monomer to dimer is mediated by an order-to-disorder transition of α5, and this aids domain swapping without exposure of the protein's hydrophobic core (83) . Since the C-terminal region of M pro is fully conserved between SARS-CoV and SARS-CoV-2, we used an NMR ensemble of the SARS-CoV M pro C-terminal domain (PDB ID: 2liz) to build a hybrid model of a disordered C-terminus. We subsequently used this model and screened against the attachment site of the α5 helix (Screen ID: 19) . Potential binders at this position would be expected to trap the protein in its inactive monomeric form by preventing dimerization.

PL pro PLpro SARS-CoV-2 nsp3 is a multi-domain protein containing an N-terminal ubiquitin-like fold, a glutamicacid-rich acidic domain, a phosphatase domain, a SARS unique domain (SUD), a catalytically active PL pro domain, a marker domain (G2M), and two transmembrane domains (84) . In addition to the protease activity, PL pro also possesses deubiquitination and de-ISGylation activity that aids in the disruption of the interferon regulatory factor 3 (IRF3) pathway and host innate immune responses (85) . The central α-helical "thumb" domain of PL pro harbors the catalytic Cys112, while the Cterminal domain is mostly β-sheet and includes the "palm" domain, which harbors the other two residues of the catalytic triad, His275 and Asp287, and the "fingers", which coordinate a Zn 2+ ion (PDB ID: 6w9c). PL pro also has two blocking loops (BL1 and BL2), targeting of which could prevent the substrate from entering the active site.

Comparison of the crystal structures of PL pro in its unbound and peptide inhibitor-bound states (PDB IDs: 6w9c and 6wx4, respectively) revealed conformational changes in both the "finger" domain as well as the BL2 loop. We therefore targeted the active site in the unliganded and the bound states, which correspond to the open and closed conformations of the active site (Screen IDs: 12 and 15, respectively; Figures S6 and S9 ). The side chain rotamer of Leu162 of structure 6wx4 was rotated to open one end of the tunnel that forms the active site. Without this modification the end of the tunnel would not be accessible with the docking programs we used. However, the co-crystal structure 6wx4 has a ligand extending through that tunnel, thus justifying the possibility of a small molecule passing through the tunnel. In addition to this, we also targeted the accessory cleft leading to the active site which would prevent substrate access (Screen ID: 13; Figure S7 ). In addition to the three sites that are described here, the ubiquitin contact site in Pl pro , which is critical for the deubiquitination and deISGylation functions, was also targeted and is described in detail in a later section.

Replication complex: nsp7, 8 and 12

Nsp7, nsp8, and nsp12 form a minimal 160-kDa complex capable of nucleotide polymerization with additional nsps playing important roles in RNA modification (86, 87) . nsp12 encodes for the RNAdependent RNA polymerase (RdRp) while nsp7 and nsp8 are cofactors that may form a hexadecameric supercomplex (88) . When compared to the SARS-CoV replicase, key functional residues are found to be fully conserved in SARS-CoV-2; however, substitutions have been found on the surface of the complex. The polymerase domain of nsp12 has an N-terminal nidovirus-unique Nterminal extension (NiRAN) in addition to the "finger", "palm", and "thumb" domains common to polymerases (89) (Figures S18, S19, S20, S21, S22). The seven catalytic motifs found in other known viral polymerases, and which are involved in template and nucleotide binding and catalysis, are also well conserved in the SARS-CoV-2 polymerase (89, 90) .

The SARS-CoV-2 RNA synthesis machinery is mediated largely by protein-protein interactions, and this provides us with an opportunity to target these interfaces using small molecules. The C-terminal head domain of nsp8 wraps around the nsp7 helical bundle, forming a hexadecameric complex. nsp8 has been suggested to possess RdRp activity of its own, and produce the primers utilized by the primer-dependent nsp12 RdRp. This nsp12 activity requires the formation of a large oligomeric complex that brings active site residues into proximity with each other (91, 92) . A previously determined crystal structure (PDB ID: 2ahm) shows nsp7 and nsp8 capable of interacting in at least two ways: i) residues on the C-terminal end of the nsp8 shaft form a hydrophobic core with residues in helix 1 of nsp7; and ii) helix 1 and the short helix of nsp7 interact with one of the helices of the nsp8 head domain ( Figure S10 ). The heterodimer formed by nsp7-nsp8 interaction binds to nsp12 on the polymerase "thumb" domain facing the NTP entry channel, thus placing the nsp12 polymerase "index finger" loop between nsp7-nsp8 and the polymerase "thumb" domain ( Figure S22 ). In addition to this interaction, a second nsp8 monomer binds to the nsp12 interface domain in the vicinity of Leu117. We targeted the nsp12 binding interface on nsp8 (Screen ID: 22; Figure S12 ), the nsp7 binding interface on nsp8 (Screen ID: 21; Figure S11 ), the entire alpha-helical surface of nsp7 which includes the nsp8 and nsp12 binding interfaces (Screen ID: 20; Figure S10 ), the nsp8 binding interface on nsp12 (Screen ID: 31; Figure S18 ), two sites in the RNA binding region on nsp12 (Screen IDs: 28 and 29; Figures S19, S21), nucleotide binding site (Screen ID: 30; Figure S20 ) and the nsp7 binding interface on nsp12 (Screen ID: 32; Figure S22 ).

The nsp7 binding interface on nsp12 is relatively shallow, and to a lesser degree, the RNA binding sites are as well. One of the principal challenges of targeting an RNA binding interface is the strong avidity effect, frequently in the nanomolar range, that restricts the ability of ligands to displace the RNA. Moreover, biological processes involving RNA-protein interactions are also known to often be dynamics-dependent. The average binding free energies of the top 100 hits from the screens of the RNA binding interfaces of nsp12 (Screen IDs: 28 and 29) were -9.9 and -9.3 kcal/mol, where as the the cocking score for the Nucleotide binding site (Screen ID: 30) was -9.7 kcal/mol. The screens targeting the PPIs (Screen IDs: 30 and 31) resulted in hits with potential for much higher affinity binding with an average binding energy of -11.1 and -10.4 kcal/mol ( Figure S45 ).

nsp13: helicase nsp13, which is one of the few proteins that is fully conserved in SARS-CoV and SARS-CoV-2, has known activity as an RNA helicase, NTPase, dNTPase, RTPase, and DNA helicase (93) . It is also known to interact with nsp12 (94) . The helicase is 603 amino acids long, contains an N-terminal zinc binding domain, a stalk domain, RecA-like domains 1A, 1B and 2A, and forms a triangular shaped structure ( Figure S23 ) (95) . There are multiple potential sites that could be targeted on this protein.

We therefore chose to target the active (ATP binding) site (Screen ID: 33; Figure S23 ), and two sites on the RNA interaction interface (Screen IDs: 34 and 35; Figures S24 and S24 ).

We made use of a homology model derived from SWISS-MODEL repository (96) for these screens (97) ; this model is expected to be reliable because the helicase of SARS-CoV-2 shares 99.8% sequence identity with the chosen template (PDB ID: 6jyt, structure of SARS-CoV nsp13). In fact, the only difference between the helicases of SARS-CoV and SARS-CoV-2 is a valine to isoleucine missense mutation in the C-terminus that is distant from our targeted sites. While this work was under review, additional structures of the helicase as a monomer and in a heteroligomeric complex with nsp12 and RNA were published (98, 99) . The backbone RMSD of the newly published apo structure (PDB ID: 6zsl) when compared to the model we used is 1.46 Å.

Previously reported hydrogen-deuterium exchange experiments suggest that the 1A and 2A domains of the helicase undergo a series of conformational changes during nucleotide binding and ATP hydrolysis (95) . This potentially inherent flexibility could impact the plasticity of the screened binding pockets and therefore the reliability of the model screened. Our two virtual screens for the RNA binding interface (Screen IDs: 34 and 35) resulted in top 100 hits with an average binding free energy of -10.2 and -10.7 kcal/mol, respectively, and the screen for the active site inhibitors (Screen ID: 33) resulted in an average binding energy of -10.2 kcal/mol. nsp9: ssRNA binding protein nsp9 is known to co-localize with nsp7, nsp8 and nsp10 within the replication complex and is presumed to play a role in RNA replication (100) . A study on SARS-CoV nsp9 has shown that nsp9 is an ssRNA-binding protein indispensable for viral RNA replication in mouse models (101) . There are several high resolution crystal structures available for nsp9 (102, 103) , including those of the protein from SARS-CoV-2 (PDB IDs: 6w4b and 6w9q). Since dimerization has been found to be essential for viral replication (104), we chose to target the dimerization interface as a method of inhibiting viral replication.

nsp9 contains seven antiparallel β-strands and an α-helix ( Figure S13 ). The dimer interface is formed mainly by the parallel association of the C-terminal α-helices ( Figure S14 ), and this surface, in particular the GXXXG motif, was targeted using VirtualFlow (Screen ID: 23). We also targeted a shallow cavity around Phe75 that partially overlaps with the first targeted site centered on the α-helix (Screen ID: 24). This second site interacts with the NTD of the other nsp9 monomer, and is more concave in shape than the first site, making it more pocket-like. The average docking score of the top 100 compounds of the second targeted site was -9.2 kcal/mol, which is a slight improvement over that of the first site with an average of -9.1 kcal/mol ( Figure S45 ).

The C-terminal α-helix, which is responsible for dimerization, has a high hydrophobic amino acid content and dimer formation is mainly facilitated by hydrophobic interactions. Targeting a region with these characteristics could result in hits that are also hydrophobic and that interact in a non-specific manner, something which would need to be addressed during follow-up experimental validation. The interface also lacks deep pockets, making it additionally challenging to find specific, tight binding compounds. Finally, targeting the dimerization interface is made even more difficult because it is not clear if the protein adopts different conformations between its monomeric form, which is to what an inhibitor might bind, and its dimeric form, which is the form that was used for the screen.

Disrupting the ability of coronaviruses to evade or subvert host defenses Initial studies on SARS-CoV-2 in ex-vivo systems and on clinical samples suggest that infection produces lower levels of interferon and differential cytokine/chemokine profiles, even when compared to the closely related SARS-CoV (105, 106) . This suggests that SARS-CoV-2 is able to escape, at least initial, host immune surveillance, and when combined with robust early replication (106, 107) , allows the virus to establish a strong foothold in the host, forcing the immune system play catch-up. In light of this, targeting SARS-CoV-2 immune evasion factors, particularly in combination with other inhibitors, could have clinical benefits.

Coronaviruses have an RNA genome and during its life cycle produce RNA species normally absent in the host cell's cytoplasm, such as dsRNA and RNA with a 5'-triphosphate. Sensing of RNA intermediates such as these is one way in which the host cell can detect viral infection and initiate an antiviral immune program (108) . An effective way in which coronaviruses avoid recognition by these intrinsic immune sensors is by using methyl transferases to add a cap-structure to viral RNA in order to mimic the cell's own mRNA (109) . Coronaviruses also encode an endoribonuclease, which prevents the accumulation of viral RNA that could activate dsRNA sensors in the cell (110) . The deubiquitination and deISGylation activities of PL pro further interfere with the host antiviral response by antagonizing the induction of type I interferon (IFN) pathways (111) . With the goal of preventing SARS-CoV-2 from being able to evade or subvert host defense mechanisms, we targeted the following four proteins: a) the guanine-N7 methyltransferase/ExoN (nsp10-14) complex, b) the 2'-Oribose methyltransferase (nsp16), c) the uridylate-specific endoribonuclease (NendoU/nsp15), and d) the deubiquitination and deISGylation activities of PL pro .

Guanine-N7 methyltransferase/ExoN (nsp10-14 complex) nsp14 is critical for viral replication and transcription, as it plays dual roles in proofreading and mRNA capping (112) (113) (114) . Disruption of nsp14 exonuclease activity (115) has been shown to result in increased sensitivity of the virus to Ribavirin (RBV) and 5-fluorouracil (5-FU) (88), demonstrating the importance of this 3'-mismatched dsRNA excision activity (116) . The ExoN activity relies on heterodimer formation of nsp14 with nsp10, and disruption of this interaction has been demonstrated to result in lowered replication fidelity (117) . Unlike many other RNA viruses, coronaviruses have a low mutation rate by virtue of this ExoN activity, which supported the expansion of coronavirus genomes (118) and is directly linked to its virulence (119) . Previous work with SARS-CoV has suggested that the limited efficacy of RBV in vivo may be due to excision by this ExoN activity (120) , and work in mouse hepatitis virus (MHV) has shown that a mutant strain lacking ExoN activity was more sensitive to Remdesivir (121) . These data suggest that use of a nucleoside analogue might be more effective if combined with an ExoN inhibitor (118) . nsp14 also functions as a SAM-dependent guanine-N7 methyltrasferase (N7-MTase) (114) . Mutation studies in a replicon system have shown that this N7-MTase activity is critical for viral replication and transcription (114) . This activity is used to add a cap 0 structure at the 5' end of viral mRNA, thus mimicking a defining structural feature of host mRNAs and assisting translation (122) as well as evasion of host defenses (123) . For nsp14 we targeted three sites: i) the active site of the ExoN domain (Screen ID: 37; Figure S28 ), ii) the active site on the methyltransferase domain (Screen ID: 38; Figure S27 ), and iii) the PPI interface with nsp10 (Screen ID: 36; Figure S26 ).

In the bimodular nsp14, the N-terminal domain of nsp14 up to Lys288 interacts solely with nsp10 and forms the ExoN domain, whereas the C-terminal domain possesses the N7-MTase activity. The ExoN domain is comprised of a twisted parallel β-sheet made up of five β-strands and are bordered by αhelices on both sides ( Figures S27, S26 , S28) (124) . The N7-MTase domain of nsp14 is made up of five β-strands, and the ligand binding pocket is situated between the β1 and β2 sheets (124) . At the time of screening, there were not yet any experimental structures available for the SARS-CoV-2 nsp10-14 complex, so we used a high confidence homology model available in the SWISS MODEL repository (97) constructed from the SARS-CoV ExoN/N7-MTase structure (PDB ID: 5c8s) (125) for our docking experiments. While nsp10 is 97% conserved between the two viruses, nsp14 has a slightly lower 95.1% identity.

The nsp10 interaction site on nsp14 has a fairly hydrophobic shallow pocket, resulting in relatively high docking scores for the top scoring compounds. However, due to the sparsity of polar atoms at this site, it is challenging to find compounds that form sufficient electrostatic interactions to provide the compounds with specificity. By comparison, the ExoN active site is more polar in nature, forming a wide shallow pocket. The polarity should lead to a larger proportion of hits that bind with high specificity, but the absence of a deep pocket could affect the binding affinity. The third target site, the active site of the N7-MTase domain of nsp14, also harbors a relatively large hydrophobic cavity, but still has polar elements. The suitability of each of these nsp14 pockets to accommodate ligands is reflected in their mean docking scores (for the top 100 hits for each target site): -10.4 kcal/mol for the PPI site, -13.4 kcal/mol for the N7-MTase active site, and -9.8 kcal/mol for the ExoN active site.

RNA cap modifications are known to play a role in the host cell's identification of self-RNA. For example, foreign RNA which lacks 2'-O methylation is inhibited by IFIT1 (126) . Pathogenic viruses such as coronaviruses that replicate in the cytoplasm have developed tactics to escape recognition by the host innate immune system. One such coronaviral mechanism is 2'-O-methyl capping of viral RNA by the nsp10-nsp16 complex. nsp16 is an m 7 GpppA-specific, SAM-dependent, 2'-O-MTase that must form a complex with nsp10 for its function (127, 128) . We therefore targeted the nsp10 binding interface on nsp16 ( Figure S31 ).

nsp16 has a catalytic core comprised of a β-sheet flanked by eleven α-helices ( Figures S30, S31 ). nsp10 has a small antiparallel β-sheet sandwiched between several α-helices and a large loop. The interaction surface of nsp10 with nsp16 is a mixture of hydrophobic, polar, and positively charged residues, and helps in the stabilization of the SAM binding site (PDB ID: 6w75). However, since the interface is relatively flat and polar, it is challenging to find small molecules that bind with sufficient strength.

Earlier structural studies showed that stabilization of the SAM-binding pocket by the nsp10-16 interaction also expands the RNA-binding pocket of nsp16 (129) , and short peptides derived from the interaction interface have been shown to inhibit 2'-O-methyltransferase activity of the nsp10-16 complex (130) . Therefore, in addition to targeting the interaction interface of nsp10 and nsp16 (Screen ID: 40), we also targeted the putative SAM binding site (Screen ID: 41, Figure S31 ). SAM binding has been shown to be essential for both nsp10-16 complex formation as well its catalytic function in another highly pathogenic coronavirus, MERS-CoV (131) . The screen targeting the nsp10 PPI and the SAM binding site resulted in the top 100 hits having an average docking score of -9.6 kcal/mol, and -11.4 kcal/mol, respectively.

Uridylate-specific endoribonuclease nsp15 is a uridylate-specific endoribonuclease (NendoU) carrying a C-terminal catalytic EndoU domain that has been described as having various roles in immune evasion in different coronviruses.

A recent study in mouse hepatitis virus (MHV), a murine coronavirus, showed that nsp15 was critical for evasion from host dsRNA sensors in macrophages (132) , and in porcine deltacoronavirus, nsp15 was found to inhibit IFN-β production (133) . The SARS-CoV-2 nsp15 monomer is composed of three distinct domains: a) an N-terminal domain containing an antiparallel β-sheet and two α-helices, b) a central domain composed of 10 β-strands and three short helices, and c) a C-terminal NendoU domain composed of two antiparallel β-sheets and five α-helices (PDB ID: 6vww) (134) . The SARS-CoV-2 nsp15 monomer is capable of forming dimeric, trimeric, and double-ring hexameric assemblies. It has been shown that the hexameric form is essential for enzymatic activity and that the hexamer is stabilized by interactions of the N-terminal oligomerization domain (135) . The NendoU active site itself is located in a groove between the two β-sheets, and conserved, crucial residues have been previously identified: His235, His250, Lys290, Thr341, Tyr343, and Ser294. With the aim of disrupting the endoribonuclease activity of nsp15, the active site was targeted in an in silico screen (Screen ID: 39; Figure S29 ), and this resulted in the top 100 hits with an average docking score of -10.9 kcal/mol ( Figure S45 ), which in theory corresponds to nanomolar inhibitors, however experimental validation is necessary to confirm this.

Deubiquitination and DeISGylation activities of PL pro PLpro PL pro (nsp3) cleaves at a consensus cleavage motif, LXGG, which is also the consensus sequence recognized by cellular deubiquitinating enzymes (DUBs). Coronaviruses attenuate the host anti-viral response by deubiquinating key components of the interferon-mediated immune response. While PL pro uses its catalytic site for its deubiquitination function, ubiquitin itself makes contacts with the "palm" and "fingers" regions of PL pro (136) . This protein-protein interface is well conserved in SARS-CoV-2 and could be targeted in order to inhibit the deubiquitination function of nsp3. SARS-CoV has also been shown to possess deISGylation activity in which ISG-15, a protein modifier consisting of two tandem ubiquitin-like domains involved in modulating the innate immune response, is cleaved (137) . We therefore targeted the proximal 'S1' Ubiquitin contact site (S1Ub) of the "palm" of nsp3 ( Figure  S8 ), inhibition of which could effectively block both the deubiquitination as well as deISGylation activities of PL pro (Screen ID: 14) . The binding interface is moderately hydrophobic and concave, making it an attractive drug target. This particular screen resulted in the top 100 ranking compounds exhibiting an average docking score of -9.2 kcal/mol ( Figure S45 ).

The macrodomain, sometimes called the X domain, is a highly conserved region of ~180 amino acids that binds to ADP ribose (138) and which has been shown to be dispensable in RNA replication, at least in the context of an RNA replicon (139) . However, it has also been shown to have possible roles in evading the host innate immune response (140, 141) . Viral macrodomains are known to have two major enzymatic activities: i) ADP-ribose 1''-phosphatase activity where the phosphate is removed from ADP-ribose 1''-phosophate, and ii) ADP-ribose hydrolase activity where ADP ribose is removed from mono-or poly-ADP-ribosylated proteins (142) .

Previous biochemical studies showed that the nsp3 macrodomain is responsible for processing of ADP-ribose 1''-phosphate, which is a byproduct of pre-tRNA splicing (143, 144) . This phosphatase activity has been found to be specific for ADP-ribose 1''-phosphate, however the turnover was found to be very low (145, 146) with a kcat of 1.7 to 20 min -1 ), making it unlikely that this activity has a strong direct effect on viral virulence.

ADP-ribosylation is a critical post-translational modification catalyzed mainly by poly (ADP-ribose) polymerases (PARPs) in which ADP-ribose is added to a protein (147) . Several PARPs are known to be interferon-stimulated genes (ISGs), and some PARPS have been shown to have antiviral activities. Studies in SARS-CoV and MERS-CoV have shown that these viruses were less virulent in the absence of the macrodomain and that this directly associated with changes in pro-inflammatory cytokine expression (148, 149) . Additional studies in HCoV-299E and SARS-CoV suggested that the nsp3 macrodomain may mediate the resistance to antiviral interferon responses (140) . Mutation of the macrodomain in SARS-CoV also induced a strong interferon and inflammatory cytokine response in the lungs of infected mice (148) . However, the exact mechanism linking the hydrolase/phosphatase activity of nsp3 and the observed cytokine production is unknown. The fact that a) the macrodomain participates in reversing antiviral ADP ribosylation (142, 150) , and b) mutation of a highly conserved asparagine residue abolishes phosphatase and hydrolase activites and mitigates viral virulence (148, 151) , suggests that targeting the ADP-ribose binding pocket could be an effective therapeutic strategy, particularly in combination with other targets. We observed two different conformations for the active site in the same X-ray structure, which indicated protein dynamics that could affect the ability of the docking to accurately recapitulate the solution-state structure. We therefore targeted both conformations of the active site of the macrodomain: one in which the active site is more closed (Screen ID: 10; Figure 4) , and a second conformation in which the active site is more open (Screen ID: 11; Figure S36 ). It is also important to note that the active site itself is very hydrophobic, which increases the risk that the resulting hits could bind non-specifically to membranes and hydrophobic pockets in other proteins, and this should be accounted for in any follow-up screening. The top 100 ranking compounds from screens with Screen IDs 10 and 11 have an average docking score of -10.8 kcal/mol and -12.7 kcal/mol, respectively ( Figure S45 ).

ORF7a is a non-essential accessory protein with a transmembrane helix at the C-terminus that has been shown to localize to the ER, Golgi, and cell surface (152, 153) . The assembly of ORF7a into viral particles suggests that the protein is important in the viral replication cycle, and that it might have a function early on in infection (154) . In SARS-CoV, ORF7a is known to interact with a 'host restriction factor', bone marrow stromal antigen 2 (BST-2) (153), which inhibits viral replication by preventing virus budding from the plasma membrane (155) . ORF7a has been shown to localize to the ER when co-expressed with BST-2. Moreover, ORF7a interferes with the glycosyslation of BST-2, which is essential for its tethering/restriction function (153) . There are also indications that ORF7a may play a role in virus-induced apoptosis (156) ; however, further studies are needed to understand the mechanism of this role.

Structurally, ORF7a has an N-terminal signal peptide, an 80 amino-acid long luminal domain, a transmembrane domain, and a short C-terminal cytoplasmic tail (152) . In the recent experimental structure of the SARS-COV-2 ORF7a (PDB ID: 6w37, (157)), the luminal domain displays structural characteristics similar to that of previously published ORF7a structures, and is composed of a sevenstranded β-sandwich ( Figure S25 ) with topological similarities to the Ig super family. Interestingly, recent genomic analysis of a sample picked up during sentinel surveillance in the state of Arizona in the USA revealed a 27 amino-acid in-frame deletion in ORF7a, in a region corresponding to the Nterminal signal peptide (158) . The functional consequences of this deletion, however, are not known.

Since the molecular mechanism of the ORF7a interaction is largely unknown, we pursued a blind docking strategy for this protein, covering the entire protein structure (Screen ID: 9; Figure S25 ). Interestingly, almost all top scoring compounds bound at a shallow cavity around Phe46, while a few bound to the opposite side near Leu31.

During experimental validation, it would need to be verified whether these hits can inhibit the interaction of ORF7a with small glutamine-rich tetratricopeptide repeat-containing protein (SGT) (159) or lymphocyte associated antigen 1 (LFA-1) (160), both of which are thought to play a role in virus-host interactions and host cell-cycle modulation. The presence of ORF7a on the viral surface has been suggested to mean that LFA-1 could be a potential viral attachment factor in leucocytes (154, 160) ; however there is currently insufficient data available about the possible residues and structural dynamics that would be involved in this interaction. Moreover, the rather flat surface of ORF7a, devoid of pockets or clefts, make it challenging to find small molecules with selective and potent binding.

Disrupting the viral assembly and packaging

As the viral lifecycle progresses, the structural proteins (S, E, M) are also expressed. These structural proteins are initially inserted into the endoplasmic reticulum (ER), but transition to the ER-Golgi intermediate compartment (ERGIC) where they aid the formation of mature virions (161) . Nucleocapsid protein (N) forms large oligomeric complexes with the replicated viral genome. The resulting ribonucleoprotein (RNP) complexes associate with M, facilitating packaging of the genome into a complete virion assembly. With limited structural information available about M and E, our efforts to find compounds that can disrupt the viral assembly and packaging were centered on the nucleocapsid protein.

The nucleocapsid protein (N) packages the viral RNA into a helically ordered RNP (162) and plays a critical role in virion assembly by interacting with M. N is also known to play key roles in the regulation of viral RNA synthesis and modification of the host cell metabolism (163) . The structure of N consists of an N-terminal RNA-binding domain (NTD, Figures S37, S38) , an intrinsically disordered SR-rich linker, and a C-terminal dimerization domain (CTD, Figures S40, S39) . The published crystal structure of the RNA-binding domain reveals a β-sheet core sandwiched between two loops. The β-sheet core has five antiparallel β-strands plus a short helix, and a longer β-hairpin between strands β2 and β5 (PDB ID: 6m3m) (164) . The central β-strands of the β-sheet core contain a purine/pyramidine monophosphate binding site, and this ribonucleotide binding mechanism is essential for virulence, making it an attractive drug docking site (165) . The NTD is also expected to interact with the CTD during RNA packaging. We have therefore targeted two minimally overlapping regions of the NTD, which in combination cover the entire surface of the NTD (Screen IDs: 42 and 43) ( Figures S37, S38 ). The first NTD site includes the ribonucleotide-binding site reported in (165) .

The CTD of N aids dimerization of the protein as well as RNA binding. The monomer consists of eight α-helices and a β-hairpin. The dimerization interface itself is composed of three helices (α5, α6 and α7) and one β-hairpin from each monomer, and involves a combination of hydrogen bonds and hydrophobic interactions, which result in a very stable dimer (PDB ID: 6wji). That the C-terminal domain has also been found to bind RNA (166) points to the critical role of this domain in the overall helical packaging of the viral genome (167) . We targeted the dimerization interface of the CTD by removing one monomer from the crystallized dimer structure (Screen ID: 44) ( Figure S39 ), and also screened the dimerization interface via a blind docking procedure against the entire dimer surface (Screen ID: 45) ( Figure S40 ). The goal of these screens was to identify a small molecule that will bind to the monomer at the dimer interface and disrupt the formation of the nucleocapsid. However, the structure of the monomer alone has not been solved and given the extent of the interaction at the dimer interface, we anticipate considerable dynamics in the monomer. Therefore the structure extracted from the dimer may not adequately describe the monomer structure, yielding imperfect hits. The other challenge of this target is that each virion has ~65 spike trimers and with a ratio of S trimer to N of 1:4 (168, 169) , this amounts to ~260 monomers or ~130 dimer interfaces. The abundance of contacts that would need to be disrupted means that a higher concentration of the inhibitor would be required for efficacy.Despite these reservations, it is worth noting that an HIV nucleocapsid inhbitor lenacapvir(GS-6207/GS-CA1) which targets the dimerization interface of the capsid protein has shown significant efficacy and has made it to clinical trials recently (170) (171) (172) . 

We screened against 17 proteins, 15 proteins of SARS-CoV-2 and two human proteins, targeting a total of 40 sites across these proteins in 45 screens with approximately 1.1 billion molecules per screen. Due to the unprecedented global situation, we wanted to make the screening results available to researchers prior to experimental validation in the hope that others may also benefit from our large multi-target dataset. The top 100 filtered hits as ranked by the docking score for each target site can be seen in Figure S45 . The top 1000 hits per virtual screen, along with additional information such as the docking scores, molecular properties, and docking poses, can be found online at https://vf4covid19.hms.harvard.edu/. In addition, the top million hits for each target site are available for download from the same website in DataWarrior format. The top 200 in silico hits for each target site are presented online (see the Resource Availability section for more details) . We hope the researchers can take advantage of our results.

As discussed earlier, not all the functional surfaces that were targeted had similar potential to accommodate a small molecule. The targeted site that yielded the weakest virtual hit compounds is the first (site 1) of the two targeted sites located on the spike interaction interface of ACE2. The reason for the relatively weak predicted binding affinities of the top scoring compounds is the flat and polar nature of this target site. These characteristics make it very challenging for small molecules to bind sufficiently well to the surface so as to disrupt the interaction with spike protein. The target yielding the virtual hits with the highest predicted binding affinities is the active site of the ExoN domain of nsp14. The high affinities are likely a product of the remarkably deep, buried pocket of this active site and the relatively high number of hydrophobic sites within that pocket. If we consider the top 1000 filtered hits, the weakest screen (ACE2, Screen ID: 2) had a mean docking score of -7.6 kcal/mol (2.3 µM) while nsp14 (Screen ID: 37) had the best, with a mean docking score of -12.9 kcal/ mol (0.3 nM). Although the trend was the same, the mean docking scores were better for the top 100 hits: -8.2 kcal/mol (1.1 µM) for ACE2 (Screen ID: 2) and -13.4 kcal/mol (0.14 nM) for nsp14 (Screen ID: 37). Finally, as expected, the mean docking scores were even better for the top 10 hits: -8.5 kcal/ mol (0.5 µM) for ACE2 (Screen ID: 2) and -14 kcal/mol (0.03 nM) for nsp14 (Screen ID: 37). The mean docking scores of the top 1000, 100, and 10 hits for each screen are shown in Figure S45 . Out of the 45 virtual screens, 41 had a mean docking score for the top 1000 compounds higher than -8.2 kcal/ mol. These scores indicate that sub-micromolar binders should be identifiable from the hits for each target site.

The hits derived from the rigid-docking procedures described here could be further improved for potency by a second-stage screen, in which the target proteins are allowed to be flexible. This could be achieved either by ensemble docking or allowing flexible side chains, for example by using GWOVina (173, 174) . Another way of improving the docking results in second stage screens consists of using alternative docking programs, such as AutoDock Vina, Smina or Gnina (175) (176) (177) . For special types of ligands such as compounds containing carbohydrates or halogens, specialized docking programs such as VinaXB or Vina-Carb could be utilized to further increase the accuracy of the results (178, 179) .

Small-molecule drug candidates that have been previously pre-approved by the FDA or other regulatory bodies for other indications, and those molecules that have been vetted for use in humans, including preclinical candidates, neutraceuticals, metabolites, and select natural products are particularly desirable hits in any inhibitor screen (180) . This is because these compounds can be fasttracked to clinical trials on the basis of previous data. Here we mined for such compounds in the top 1% of our hits from the "in stock" compounds of the ZINC library. We found 161 drugs (with 491 occurrences across all the screens) that are "world approved" in our top 1%. We further filtered these drug candidates to only those with a docking score better than -8 kcal/mol and show these 80 approved drugs along with their screening target and docking score in Table S5 . Of these 80 drugs, 16 of them are being considered in clinical trials for COVID-19. The complete list of "world approved" hits along with their binding targets and docking scores is available at https://vf4covid19.hms.harvard.edu/world-approved-drugs.

Several hits from the "world approved" drugs are from the steroid family. Corticosteroids such as betamethasone, ciclesonid, flumetasone and meprednisone acetate are already being investigated as COVID-19 treatments (181) (182) (183) . For example, betamethasone was identified as a hit for nsp14, the bifunctional enzyme that acts as a methyltransferase and an exoribonuclease. Betamethasone bound in the simulation to the active site of the exoribonuclease of nsp 14 ( Figure S47 ). While the usefulness of corticosteroids is largely attributed to their anti-inflammatory effects in the host, these screening results suggest that additional mechanisms of action that involve viral proteins may be possible. Other steroid hormones in clinical trial such as estradiol or dutasteride also came up as hits in our screens.

Of the non-steroid approved drug hits, tyrosine kinase inhibitors occupy a significant fraction. Drugs such as imatinib (Glivec ® ) are being tested in clinical trials due to their reported in vitro efficacy against SARS-CoV and MERS-CoV (184, 185) ; however, their mode of action is not yet fully understood. In our screens, imatinib binds to two targets: i) nsp14, but this time at the active site of the methyltransferase, and ii) the RNA-binding interface of the helicase, nsp13.

Of the 3897 "investigational" drugs in the ZINC15 database, 137 were hits (500 occurrences) and of the 101378 "in-man" compounds, 401 were hits (1061 occurrences). A full list of these molecules, along with their respective target sites and docking scores is available at (https://vf4covid19.hms.harvard.edu/investigational-drugs). Multiple occurrences of a hit molecule partly stem from the different docking scenarios that we used for the same target site. In addition, some of the compounds may be represented more than once in the list because of differences in tautomeric, isomeric, or protonation state, any of which earns the molecule a unique ID in the ZINC database. Despite the abundance of approved hits, some of which are already involved in or related to drugs in clinical trials, experimental confirmation of the specificity of hits from our screen against SARS-CoV-2 proteins is still needed to determine an antiviral mechanism of action.

Comparison to previously identified inhibitors SARS-CoV-2 is not the first coronavirus to cause serious disease in the human population, and these other coronaviruses (SARS-CoV and MERS-CoV) have also been the subject of research into small molecule inhibitors. There have been several efforts in the past, both experimental and computational, to target specific viral proteins from these viruses. The proteins that were most often targeted in these efforts were the proteases, M Pro (73, (186) (187) (188) and PL Pro (186, (189) (190) (191) , the polymerase (192, 193) , the spike protein (66, 194, 195) , the endoribonuclease (196) , and the helicase (197) (198) (199) S4. Although direct binding affinities were not available for most of the molecules, the inhibitory potential, IC50, or the effective potential, EC50, were reported.

In order to determine if previously published SARS-CoV and MERS-CoV hits for which an exact binding site had been reported were also top hits against SARS-CoV-2, we calculated docking scores using the same computational methods used in the screens described here, allowing direct comparison with our hits. One of the best inhibitors for SARS-CoV M Pro is an asymmetric aromatic disulfide molecule with an IC50 of 0.52 ± 0.06 µM (187) . When we docked this molecule against SARS-CoV-2 M pro , we got a docking score of -7.7 kcal/mol S4, whereas the top 100 hits from two of our recent screens described here, which targeted the active site of M pro (Screen IDs: 16 and 17), resulted in mean docking scores of -10.8 and -11.4 kcal/mol. However, this apparent improvement in binding affinity must still be validated experimentally. Similarly, in the case of PL pro , the best previously described active site inhibitor (189) had a binding free energy of -7.2 kcal/mol when re-screened, while the top 100 compounds from the two corresponding VirtualFlow screens (Screen IDs: 12 and 15) have a mean binding scores of -8.9 and -10.0 kcal/mol, respectively. S45. Lee, et al. described a novel helicase inhibitor that can block both the dSDNA unwinding and ATP hydrolysis activities of the SARS-CoV helicase (197) . Docking of this molecule to the SARS-CoV-2 helicase gave a score of -7.9 kcal/mol. We also targeted the active site of the SARS-CoV-2 helicase (Screen ID: 33) here and that screen resulted in a top 100 hits with a mean docking sore of -10.2 kcal/mol. In addition, there have been multiple attempts to develop small molecule inhibitors that target the active site of the RdRp, considered a promising target, from both SARS-CoV and MERS-CoV (192, 193) . Docking of the best molecules from these previously published studies against the SARS-CoV-2 RdRp resulted in docking scores of -6.1 and -6.8 kcal/mol, while our screens (Screen IDs: 28 and 29) resulted in compounds with mean docking scores of -11.1 and -9.7 kcal/mol, respectively, for the top 100 hits. It must be noted that, as stated before, the comparisons provided here are based on docking scores and therefore any apparent improvements in binding still need to be experimentally validated.

There have also already been multiple attempts to employ virtual screening of varying library sizes to identify approved drugs as well as discover new molecules that can effectively inhibit SARS-CoV-2 viral infection (73, (200) (201) (202) (203) (204) (205) (206) (207) (208) , and more than one screen is focused solely on M pro (200, 203, 206, 207) .

It is important to note that in the case of M pro , the conformational plasticity of the active site means that screening against a single conformation, as is done in these studies, may be less effective at identifying functional hits. One work of note is that by Ton, et al. 1.3 billion compounds from the ZINC 15 library were screened against M pro using a Deep Docking algorithm (204) . While Deep Docking significantly accelerates the screening of ultra-large compound libraries, it relies heavily on the docking scores of a small random subset of compounds to eliminate a large chunk of the compounds early in the docking workflow, and this can be detrimental to finding an absolute docking score for individual compounds.

A recent mass-spectrometry-based study has identified several host proteins that interact with SARS-CoV-2 viral proteins, leading to the identification of multiple druggable host targets and a series of known modulating compounds for these targets that were then shown to have antiviral activity in cellculture-based assays (209) . The efficacy of these already approved host-directed compounds lends support to the development of these drugs as therapeutic interventions for SARS-CoV-2 infection. Coincidentally, some of the compounds from this mass spectrometry study were also identified as hits in our screens against diverse viral proteins. For example, 'Nafamostat', an approved anticoagulant identified in Gordon, et al. as having a possible impact on cell entry, was also identified in our screens as a hit with good scores against three viral proteins: ORF7a, PL pro , and nucleoprotein. This compound is already undergoing clinical trials for COVID-19 (Ref:NCT04352400, NCT04418128) . Similarly, 'progesterone', a common birth control drug that is also known to have anti-inflammatory properties was identified in the proteomics study as a ligand with antiviral activity that targets sigma-1 or sigma-2 receptors. In our screens, this compound was identified as an inhibitor for nsp16, nsp7 and spike. Progesterone is also in clinical trials for treatment of COVID-19 (NCT04365127). Silmitasertib (CX-4945), an inhibitor of casein kinase II (CK2), was identified both in the proteomics study as well as in another recent study that mapped the host phosphorylation landscape in the context of SARS-CoV-2 infection (209, 210) . Silmitasertib was originally developed to be used in combination with chemotherapy. Here, we identify that Silmitasertib could potentially bind to nsp16. If some of these host-protein-targeted compounds can also directly inhibit some viral proteins, as suggested by the docking scores derived from our screens, this could form the basis for a unique regimen of multi-target drugs to treat COVID-19 that acts on both viral as well as host proteins simultaneously. A list of compounds that are common to both the interactome study and our virtual screens, along with their potential targets are provided in Table S6 .

To verify if the docking scenarios and procedures that were used in this study worked in principle, we carried out two types of evaluation studies with compounds which were previously identified as inhibitors and/or binders. These verification studies should be interpreted on a qualitative level as they are not be able to recapitulate the unique advantages of the ultra-large scale of the screens presented here. First, we completed re-dockings of ligand-protein complexes for the macrodomain, M pro , PL pro , nsp15 and nsp16. Here, the similarity of the binding geometry of the ligand, derived from re-dock, is compared with that of the available high-resolution experimental structure for the target protein and ligand. Approximately 35% of the 29 compounds that we re-docked have an RMSD below 2.5Å when compared to the experimental structure. As a relative benchmark of our performance, in the original AutoDock Vina publication an average re-docking success rate of around 50% was reported when using rigid receptors (175) . However the docking accuracy in the aforementioned publication was set to a much higher value (i.e., exhaustiveness parameter set to 8), while we used QuickVina 2 in its fastest setting (exhaustiveness of 1) in our study due to the computational costs of screening a billion compounds per target. The key idea of our approach is to use the scale of the screen to increase the quality of the resulting top scoring compounds (40, 43, 211) . One of the successfully re-docked compounds for nsp16 can be found in Figure S44 .

In addition, we carried out enrichment studies for M pro and PL pro , as these were the only two targets with sufficient non-covalent binders reported. Enrichment studies assess the ability of the docking routines to re-identify and enrich previously identified binders when mixed with a pool of random compounds or decoys. Two different flavors of enrichment studies were performed: i) using experimentally identified molecules, and ii) using the virtual hits identified in this study. The DUD-E database was used for generating decoy molecules based on the active compounds (212, 213) . The complete details of these studies can be found in the Supplement on page Enrichment and re-docking verification studies. The experimentally active compounds that were used in the enrichment studies are listed in Table S1 and the virtual hits that were used as active compounds are listed in Table S2 .

The enrichment factors at 2%, EF2, that were obtained for the experimental binders are 20 and 25 for M pro and PL pro , respectively. The maximum possible EF2 is 50 for each of the two target proteins. The results obtained here are typical of similar docking procedures (212, 213) . When the enrichment studies were performed with ten virtual hits, the enrichment factors at 2% are equal to the theoretical maximum value of 50, which reflects the effect of ultra-large scale screening.

Our ultra-large library size and the multi-pronged screening approach we present here is not only unprecedented in terms of scale, significantly enhancing the chances of finding an effective antiviral, but also provides a pool of potential antiviral compounds that could be developed further as pancoronavirus drugs as well as combination therapies. Combination regimens of reverse transcriptase inhibitors and protease inhibitors have been found to be effective as anti-retroviral therapy for human immunodeficiency virus (HIV) and has lead to dramatic improvement in HIV infected patient health and survival worldwide (214) (215) (216) . Clinical studies have also shown that in cases of severe influenza, a combination of multiple antiviral drugs is much more effective than a monotherapy (217, 218) . Therefore, it could also be beneficial to develop a combination therapy to treat severely affected COVID-19 patients. Considering that vaccine deployment will take time and that the long-term efficacy of any SARS-CoV-2 vaccines is still unknown, combination therapeutic strategies for the control and treatment of coronavirus infections should continue to be developed in parallel.

The mutating nature of coronaviruses and the potential for the emergence of novel coronaviruses in the future make the development of broad spectrum antivirals desirable. For a protein to qualify as a candidate target for pan-coronavirus drug development, it should be i) critical for the viral replication or virulence, and ii) highly conserved across all known CoVs and considered likely to be conserved in emerging CoVs. While the structural proteins (S, N, M and E) are not as well conserved, the accessory proteins are conserved to an even lesser extent across different CoV species (ProViz, (219)). Highly conserved nonstructural proteins and conserved host proteins would therefore be ideal targets for the development of broad spectrum CoV drugs (220) . While studies (221) have suggested that the S and N genes of SARS-CoV-2 are undergoing episodic selection during human transmission, recent analyses of genome samples from across the world suggest that the greatest genetic diversity is occurring in S, ORF3a and ORF8 (222) . Structural genomic analysis across multiple virus classes, including SARS-CoV-2, have shown that intra-viral interactions are more conserved than viral hostinteractions (223) (224) (225) (226) . The protein-protein interaction interfaces involved in the heteromeric RNA polymerase complex (composed of nsp7,nsp8 and nsp12) are evolutionarily conserved (223) and our top hits against these crucial PPIs represent promising candidates for the development of pancoronavirus inhibitors. Similarly, structural genomic analysis of the methyl transferase complexes (nsp10-nsp14 and nsp10-nsp16) also reveals PPIs that are fully conserved with mutations limited to the surface (223) . Moreover, the substrate binding regions we used for PL pro , M pro , nsp14, nsp16 and the RNA binding site of the nucleocapsid protein have all been shown to be highly conserved (226) , and the correspondingly identified small molecules could also potentially be developed as pancoronavirus therapeutics. Since most human coronaviruses differ in their use of host cell receptors and priming proteases (18), identifying a cell-based assay or functional animal model for developing pancoronavirus therapeutics could be challenging. The use of normal human bronchial epithelia (NHBE) cells as a potential universal screening platform for multiple classes of coronaviruses shows promise (227) (228) (229) , and the general need for high-throughput ex vivo testing models will hopefully drive development in this area.

It is important to remember that the work described here is only the first step in the drug discovery process. Virtual screening seeks to use structure and modelling-guided approaches to identify higher quality starting hits. Hits from these screens still need to be experimentally validated for binding, inhibition, and activity in the context of viral infection. Some of these targets, which affect viral immune evasion rather than replication, may need more complex evaluation in cell culture or in animal models to determine efficacy. While our own experimental validation of hits from these screens has begun, we believe that making these results available to a broader audience for experimentation is important in light of the current pandemic.

The alarming global spread of SARS-CoV-2 and the dearth of effective prophylaxis or treatment methods has resulted in a pressing need for accelerated drug discovery. Here we make use of the recent chemical space revolution that resulted in ultra-large make-on-demand libraries, to prioritize lead molecules and minimize the transition time from discovery to clinic. We have used the availableto-date SARS-CoV-2 structome to carry out detailed structure-based virtual screening against 17 different target proteins, frequently at multiple target sites, by leveraging the versatility of our recently developed in silico screening platform, VirtualFlow. The set of small molecule hits that we present here can aid drug development happening at multiple institutions worldwide, and the set of prospective drug molecules derived from this ultra-large multi-target screening campaign will potentially contribute to the building of a future arsenal of anti-coronaviral drugs, ready to tackle future outbreaks.

We present here results from our ultra-large-scale computational effort and the hits have not been experimentally validated for their efficacy. Given the multipronged approach targeting 40 different sites on 17 proteins, it is a challenge for any one research group to validate all the targets in the timely manner that the urgency of the current pandemic warrants. We hope that the results provided here can be leveraged collectively by the research community to identify potent inhibitors of SARS-CoV-2. In addition, we have also discussed the potential and plausible challenges, if any, for each target site from a docking perspective that researchers should bear in mind while validating hits.

When performing in silico screening, we generally choose a 3D surface on the target protein, referred to as the "docking box", and evaluate the energetics of a molecule bound to that chosen surface. The docking box is judiciously chosen to target a known functional domain, either an active site, an allosteric site, or a protein-protein interaction interface. In practice, the docking box is typically extended to accommodate secondary interactions with the small molecule, especially when competing with a substrate. However, occasionally this results in identifying a tight binder that does not harbour any inhibitory function. It is important to take into account this possibility during follow-up experiments and test for binding using a direct binding assay such as ITC, SPR, MST or NMR. In particular, NMR and X-ray crystallography structures would provide detailed information on the binding site and how the small molecule engages the target. It could still be possible to functionalize these non-inhibitory binders with warheads to degrade the targeted protein by established methods, referred to as proteolysis-targeting chimeras (PROTACS) (230, 231) . However the effectiveness of these PROTACs would depend on the concentration, rate of turnover, and accessibility of the target viral proteins.

As noted before, when testing these compounds in viral assays, one should be cognizant of the nature of the target and the possibly differential effects on viral viability versus viral virulence. It may also be possible to correlate "pre-approved" hits from the virtual screens with patient data, for example to determine if patients' regular medication and supplement regimens reveal any hits that positively correlate with either a milder disease course or a faster time to recovery. This could help narrow down targets and provide justification for a more detailed and time consuming experimental analysis of a hit.

Further information and requests for resources should be directed to and will be fulfilled by the Lead Contact, Dr. Haribabu Arthanra, hari@hms.harvard.edu.

This study did not generate nor use any new or unique reagents.

The virtual screening results and special subsets (e.g. the hits which are natural products) are available on the project homepage at https://vf4covid19.hms.harvard.edu/. Table 1 : Overview of the Performed Virtual Screens. A total of 45 virtual screens were performed, involving 17 different target proteins and 40 unique target sites. In each screen, approximately 1 billion molecules from the Enamine REAL library, and approximately 10 million compounds from the in-stock library ZINC 15 database were screened. 6wx4* The side chain rotamer of Leu162 in PDB ID 6wx4 was rotated to open up the end of the tunnel that forms the active site (for more details please refer to the main text). 6m0k* The side chain rotamers of the residues Met49, Ser46, and Cys145 of the PDB ID 6m0k crystal structure were modified to give increased access to the binding pocket (for more details please refer to the main text and Figure S41 ).

VirtualFlow was used to screen around 1 billion on-demand (synthesizable) compounds from the Enamine REAL library, and around 10 million in-stock compounds from the ZINC 15 library (232) for each targeted site. Both libraries were previously prepared with 'VirtualFlow for Ligand Preparation' (VFLP) (43) . For each virtual screen a single receptor structure was used, and the protein was held rigid. QuickVina W was used to perform a blind docking procedure (233) for the HR1 domain of the spike protein, the RNA binding interface of the nucleoprotein, the RNA binding site of nsp12, as well as for nsp7 and ORF7a. For all the other docking routines, we used QuickVina 2 (234) . Both docking programs are based on AutoDock Vina (175) , and the exhaustiveness parameter was set to 1 during all virtual screens. This setting minimizes the search of the conformational space within the docking box and maximizes the computational efficiency. The receptor structures were prepared with AutoDockTools (235) from the PDB format to the PDBQT format.

To conduct VirtualFlow screens we needed to leverage a significant amount of high performance computing (HPC) time and infrastructure in order to complete multiple virtual screens in parallel. For this purpose we have taken advantage of Slurm (https://slurm.schedmd.com), a highly scalable and fault-tolerant workload manager capable of handling multiple jobs in parallel. In addition, the Elastifile Cloud File Storage (ECFS) was used, which is a POSIX-compliant and highly scalable NFS file service mounted to the Slurm cluster compute nodes. ECFS provides a full suite of cloud-native NAS features (including snapshots and multi-zone accessibility). The Google Cloud Platform (GCP, https://cloud.google.com), which combines all of these components, including Google Compute Engine preemptible virtual machines, was used to provide the scale of computation required for the ultra-large virtual screens reported here. Up to 160,000 CPUs (Intel Xeon or AMD EPYC) were used in parallel per virtual screen. Google Cloud Storage was used to durably store the generated data.

We operated within a Google Cloud project that has a single Virtual Private Cloud to provide networking functionality to all the virtual instances. To set up this infrastructure we leveraged automated deployment tools like Terraform (https://www.terraform.io/) and the Google Cloud Deployment Manager to replicate this architecture across 3 different zones belonging to two different geographic regions. This resulted in three different clusters ( Figure S43 ) that were used for virtual screens against different targets in parallel. This globally distributed architecture allowed us to run hundreds of thousands of cores in parallel, consuming a total of approximately 100 million CPU hours in a total of about 3 weeks for all the virtual screens reported here. The types of nodes used are listed in Table S3 .

The virtual hits resulting from these screens were filtered to remove compounds that have reactive functional groups, a molecular weight greater than 600 daltons, a cLogP greater than 6, or more than 10 hydrogen bond acceptors using the open-source software DataWarrior (236) . A smaller number of non-druglike molecules from the ZINC library were also removed after visual inspection.

In addition, for each target site, we showcase one of the top hits. This hit was chosen based on consideration of a number of parameters including, but not limited to, docking score, 'drug-likeness' (which includes criteria such as Lipinski's rule of five and the tendency to aggregate, as estimated by a fraction of sp3-hybridized carbon atoms between 0.25 and 1), and the presence of PAINS (pan-assay interference compounds) or other potentially toxic moieties. The hits depicted were also chosen to highlight particularly plausible small molecule-protein interactions that can be accommodated by the docking site. Due to these filtering criteria, it is therefore important to note that although the example hit molecule displayed in each figure is one of the top hits for the represented screen, it is often not the hit with the highest docking score.

In in silico screening efforts the docking score is the metric that informs on how well a molecule is predicted to bind to the target protein. The docking score of the AutoDock Vina scoring function is expressed in kcal/mol and represents the free energy of binding , which can be converted into the binding affinity (dissociation constant, ) using the equivalence provided by the Gibbs equation

, where R is the ideal gas constant and T is the temperature (237) . The docking programs including the ones we have used (QuickVina 2, QuickVina-W) are calibrated with respect to the free energy of binding based on experimental binding data, as both use the original AutoDock Vina scoring function (234, 233, 175) . This implies that a docking score of -10.0 represents a predicted binding free energy of -10.0 kcal/mol, which can be converted into a dissociation constant using the above equation. We used the docking score to rank the hits and associate it with a corresponding binding affinity, a metric commonly used by experimental researchers. However, it should be noted that there are inconsistencies between computationally calculated binding affinities and experimentally derived ones. This is partly due to the fact that the computational method, especially in the least exhaustive mode required to ensure high-throughput, cannot capture all the factors that contribute to binding and is therefore only an estimation of the affinity. The docking score and associated binding constant should only be taken as comparative metric in the absence of further experimental validation.

We carried out re-docking studies of known ligands to some of the SARS-CoV-2 proteins that were targeted in this work to verify whether the docking scenarios we used worked in principle. We found that only a few identified ligands with experimental ligand-complex structures were available. Furthermore, it is important to note that the published studies do not recapitulate the scale of the screens that we have carried out. This is an important point as the scale improves the quality of the top scoring compounds (211, 43, 40) .

In the re-dockings, the same docking program (QuickVina 2) and docking settings where used as in the primary virtual screens (e.g., exhaustiveness 1). A docked compound was classified as successfully reproducing the experimental binding geometry when the RMSD in comparison to the experimental ligand binding mode was below 2.5 Å. Based on this definition, 10 out of the 29 redocked compounds were successful, which corresponds to a success rate of roughly 35%.

PL pro . For PL pro a total of 10 ligands were re-docked, based on the following co-crystal structures with PDB IDs: 7cjm, 7d7t, 7jir, 7jit, 7jiv, 7jiw, 7jn2, 7jrn, 7koj, 7kol. All of the ligand were reported to bind to the accessory pocket, so the docking scenario used in Screen ID 13 was used for the redocking. The obtained RMSD values were as follows: 7cjm: 8.64 Å, 7d7t: 3.24 Å, 7jir: 9.10 Å, 7jit: 0.92 Å, 7jiv: 2.22 Å, 7jiw: 7.61 Å, 7jn2: 10.15 Å, 7jrn: 0.35 Å, 7koj: 6.13 Å, 7kol: 1.45 Å.

M pro . For M pro a total of 10 ligands were re-docked, based on the following co-crystal structures with PDB IDs: 5ret, 5reu, 5rev, 5rf7, 7c7p, 7c8t, 7cx9, 7d3i, 7jkv, and 7k40. The obtained RMSD values were as follows: 5ret: 4.64 Å, 5reu: 3.63 Å, 5rev: 6.81 Å, 5rf7: 1.87 Å, 7c7p: 5.06 Å, 7c8t: 6.75 Å, 7cx9: 4.33 Å, 7d3i: 4.42 Å, 7jkv: 2.41 Å, and 7k40: 2.37 Å.

For the macrodomain a total of 2 ligands were re-docked, based on the co-crystal structures with PDB IDs 6wcf and 7jme. The obtained RMSD values were as follows: 6wcf: 1.57 Å, and 7jme: 7.78 Å.

nsp15. For nsp15 a total of 5 ligands were re-docked, based on the following co-crystal structures: PDB IDs 6w01, 6wlc, 6x1b, and 6wxc. The obtained RMSD values were as follows: 6w01: 2.33 Å, 6wlc: 3.67 Å, 6x1b: 4.66 Å, and 6wxc: 2.95 Å.

nsp16. For nsp16 a total of 4 ligands were re-docked, based on the following co-crystal structures: PDB IDs 7jhe, 6wvn, and 6yz1. The obtained RMSD values are as follows: 7jhe: 2.25 Å, 6wvn: 3.26 Å, and 6yz1: 3.63 Å. The re-docking results related to PDB ID 7jhe are illustrated in Figure S44 .

We carried out enrichment studies for M pro and PL pro , in which 10 active compounds from the literature for each of the two proteins were used as positive hits. For each of these active compounds, approximately 50 decoys were obtained from them DUD-E database (212, 213) , resulting in a total of 500 decoys for each of the two target proteins. The active compounds used are listed in Table S1 . Docking of the decoys and of the active compounds was carried out with the same settings as was used in the original virtual screens (Screen IDs 13 and 16). For M pro an enrichment factor at 2% (EF2) of 20.0 was obtained, and for PLpro an EF2 of 25.0 was obtained. The maximum possible EF2 is approximately 50 for each of the two target proteins.

For comparison we also carried out enrichment studies for both Mpro (Screen ID 16) and PLpro (Screen ID 13) where the positive/active ligands were the top 10 virtual screening hits for each of the targets (see Table S2 ). Again, for each ligand we obtained 50 decoys from the DUD-E database, resulting in a total of 500 decoys per target site. In both cases the EF2 equaled 50.0, which is the theoretical maximum possible value. Although limited in the scope for interpretation, these results demonstrate the effect of the scale on the virtual screening results. The results would be expected to be similar for the other target sites due to the extraordinarily high docking scores of the top scoring compounds, which result from the ultra-large scale. Table 1 .a, M pro (violet) and an example compound (light pink) from the virtual screen bound to the active site. Here, a modified version of the structure wth PDB ID 6m0k was used (details described in the main text)(Screen ID: 17). b, Electrostatic surface of the target protein to which an example compound (light pink) is bound. c, An overview of the interactions between the inhibitor and the protease structure. d, Residues within 4 Å of the inhibitor. e, Distribution of the docking scores of the top 100 virtual screening hits. Figure S6 : PL pro and an example compound from the top 0.0001% of screened compounds bound at the enzymatic active site. Related to Table 1 .a, PL pro (violet) and an example compound (light pink) from the virtual screen bound to the active site (Screen ID: 12). b, Electrostatic surface of the target protein to which an example compound (light pink) is bound. c, An overview of the interactions between the inhibitor and the protease structure. d, Residues within 4 Å of the inhibitor. e, Distribution of the docking scores of the top 100 virtual screening hits. Figure S7 : PL pro and an example compound from the top 0.0001% of screened compounds bound at the accessory pocket near the active site. Related to Table 1 .a, PL pro (violet) and an example compound (light pink) from the virtual screen bound at the accessory pocket near the active site (screen ID: 13). b, Electrostatic surface of the target protein to which an example compound (light pink) is bound. c, An overview of the interactions between the inhibitor and the protease structure. d, Residues within 4 Å of the inhibitor. e, Distribution of the docking scores of the top 100 virtual screening hits. Figure S8 : PL pro and an example compound from the top 0.0001% of screened compounds bound at the DUB binding site. Related to Table 1 .a, PL pro (violet) and an example compound (light pink) from the virtual screen bound at the DUB binding site (Screen ID: 14) . b, Electrostatic surface of the target protein to which an example compound (light pink) is bound. c, An overview of the interactions between the inhibitor and the receptor structure. d, Receptor residues within 4 Å of the inhibitor. e, Distribution of the docking scores of the top 100 virtual screening hits. Figure S9 : PL pro and an example compound from the top 0.0001% of screened compounds bound at the enzymatic active site (tunnel region). Related to Table 1 .a, PL pro (violet) and an example compound (light pink) from the virtual screen bound to the active site (tunnel region) (Screen ID: 15). b, Electrostatic surface of the target protein to which an example compound (light pink) is bound. c, An overview of the interactions between the inhibitor and the protease structure. d, Residues within 4 Å of the inhibitor. e, Distribution of the docking scores of the top 100 virtual screening hits. Table 1 .a, nsp7 (violet), where an example compound (light pink) from the virtual screen is bound at the nsp8 (light gold) interface. In the virtual screen, a blind docking was carried out for each screened compound over the entire surrounding helical surface. This helical surface includes large parts of the nsp8 (light gold, cyan) and nsp12 (lavender) binding interfaces (Screen ID: 20). b, Electrostatic surface of nsp7 to which an example compound (light pink) is bound. Also shown are the protein-protein interaction partners nsp8 (light gold, cyan) and nsp12 (lavender). c, An overview of the interactions between the inhibitor and the nsp7 structure. Table 1 .a, nsp10 (violet) bound to nsp14 (light gold) and an example compound (light pink), at the nsp14/nsp16 binding interface of nsp10 (Screen ID: 25). b, Electrostatic surface of nsp10 to which an example compound (light pink) as well nsp14 (light gold) are bound. c, An overview of the interactions between the inhibitor and the nsp10 structure. d, Residues within 4 Å of the inhibitor. e, Distribution of the docking scores of the top 100 virtual screening hits. Table 1 .a, nsp12 (violet) bound to nsp7 (light gold) and an example compound (light pink) at the nsp7 binding interface (Screen ID: 32). b, Electrostatic surface of nsp12 to which an example compound (light pink) is bound to the nsp7 binding site. nsp7 is shown in light gold. c, An overview of the interactions between the inhibitor and the nsp12 structure. d, Residues within 4 Å of the inhibitor. e, Distribution of the docking scores of the top 100 virtual screening hits. Figure S23 : The helicase (nsp13) and an example compound from the top 0.0001% of screened compounds bound at the enzymatic active site. Related to Table 1 Table 1 .a, nsp16 (violet) bound to an example compound (light pink) at the active site (Screen ID: 41). b, Electrostatic surface of nsp16 to which an example compound (light pink) is bound. c, An overview of the interactions between the compound and the nsp16 structure. d, Residues within 4 Å of the inhibitor. e, Distribution of the docking scores of the top 100 virtual screening hits. Figure S31 : nsp16 and an example compound from the top 0.0001% of screened compounds bound at the nsp10 protein-protein interaction interface. Related to Table 1 .a, nsp16 (violet) bound to an example compound (light pink) at the nsp10 (light gold) protein-protein interface (Screen ID: 40). b, Electrostatic surface of nsp16 to which an example compound (light pink) and nsp10 (light gold) are bound. c, An overview of the interactions between the compound and the nsp16 structure. d, Residues within 4 Å of the inhibitor. e, Distribution of the docking scores of the top 100 virtual screening hits. Table 1 .a, The target protein ACE2 (gold) in a closed conformation bound to the RBD of the spike protein (magenta) and an example compound (light pink) from the virtual screen bound to target site 3 (around Phe40), which contains a dynamic pocket located adjacent to the RBD interface (Screen ID: 4). b, Electrostatic surface of the target protein (ACE2) bound to the RBD domain of the spike protein (magenta) and an example compound (light pink). c, An overview of the interactions between the ligand and the receptor structure. d, Residues within 4 Å of the ligand. e, Distribution of the docking scores of the top 100 virtual screening hits. Table  1 .a, The target protein ACE2 (gold) in a closed conformation bound to the RBD of the spike protein (magenta) and an example compound (light pink) from the virtual screen bound to target site 4 (around Leu39), which represents a dynamic pocket adjacent to the RBD binding interface (Screen ID: 5). b, Electrostatic surface of the target protein (ACE2) bound to the RBD domain of the spike protein (magenta) and an example compound (light pink). c, An overview of the interactions between the ligand and the receptor structure. d, Residues within 4 Å of the ligand. e, Distribution of the docking scores of the top 100 virtual screening hits. Table 1 .a, Secondary structural elements that form the active site of M pro . The catalytic dyad Cys145 (C145)-His41 (H41) at the base of the active site is surrounded by secondary structural elements contributed by domains I and II of the protease. The side chains of selected residues are shown: Thr25 (T25), Ser46 (S46), Asn142 (N142), Pro168 (P168) and Gln189 (Q189). The structure of M pro bound to the inhibitor N3 is shown (PDB ID: 6lu7). b, Overlay of M pro active site structures showing that binding of inhibitors induces conformational changes in two secondary structural elements, which form the walls of the active site: the loop harboring Gln189 and the short α-helix harboring Ser46 (see panel a) . The main chain is colored gray for the three apo structures of M pro (PDB IDs: 6y2e, 6yb7, 6m2q), purple for the structure of M pro bound to the inhibitor N3 (PDB ID: 6lu7), blue for the structure of M pro bound to the inhibitor 11b (PDB ID: 6m0k) and green for three other structures of M pro bound to inhibitors (PDB IDs: 6m2n, 6y2g, 7buy). c, Van der Waals surface of the active site of M pro bound to the inhibitor N3 (PDB ID: 6lu7), after removing all water and ligand molecules. The side chains of Met49 (M49), Ser46 (S46) and Cys145 (C145) are colored. d, Van der Waals surface of the active site of M pro bound to the inhibitor 11b (PDB ID: 6m0k), after removing all water and ligand molecules (left) and the same structure after selecting different rotamers for the side chains of Met49, Ser46 and Cys145 (right). The images in panels a-d show M pro from the same orientation and were generated using PyMOL. Figure S42 : Alignment of extracellular domain of TMPRSS2 against extracellular region of the transmembrane serine protease hepsin represented by WebLogo images (https://weblogo.berkeley.edu/logo.cgi). The structure of hepsin (PDB ID: 1z8g) was used as a template for making the homology model used in Screen ID 6. The residues which comprise the catalytic triad are marked, and are fully conserved. Figure S43 : Google Cloud Platform (VPC-1) Figure S44 : Redocking geometries for PDB structure 7jhe (nsp16). Related to Table 1 . SARS-CoV-2 nsp16. Redocking of the ligand S-Adenosyl-L-homocysteine (SAH) from the PDB structure 7jhe was performed using the same docking settings as in the primary virtual screening for this target. The docking program used was QuickVina 2, and the exhaustiveness was set to 1. The experimentally reported binding geometry is shown in violet, while the docked compound is shown in cyan. The RMSD between the two geometries is 2.25Å. Table 1 . Average docking scores of the top 100 compounds of each virtual screen that was carried out. Only compounds with cLogP less than 6, molecular weight less than 600 dalton, no reactive groups, and at most ten hydrogen bond acceptors were considered (a smaller number of non-druglike molecules from the ZINC library were also removed by visual inspection). The docking scores estimate the free energy of binding in kcal/mol. These are computational values that can diverge significantly from the experimental binding free energies. The dashed lines indicate the corresponding Kd values of 1 µM, 100 nM, and 10 nM under standard conditions.

Slurm controller node n1-standard-16

Slurm login node n1-standard-8

Slurm compute nodes n2d-highcpu-64 (cluster 1 and 2), n1-highcpu-32 (cluster 3)

Elastifile management node n1-standard-8 ECFS storage node custom (16 vCPUs, 96 GB memory) (193) 1600 Table S4 : Known inhibitors of SARS-CoV-2.The inhibitors that were previously reported for SARS-CoV-2 proteins were docked against respective targets and the docking scores were calculated using the same computational methods we used in the screens presented here.

Generic name COVID clinical trial Targets (Screen IDs) Table S5 : Approved drugs among the virtual screening hits. Related to Table 1 . A table of "world approved" drugs that bind to SARS-CoV-2 proteins as identified by VirtualFlow. These drugs are mined from the top 1 % of the hits from "in stock" compounds of the ZINC library. The drugs presented here are the top 1 % of hits that were further filtered for a docking score of better than -8 kcal/mol. Some of the drugs are predicted to bind to multiple protein targets, and for some of the drugs (denoted by *) only the targets of the top five virtual screens as judged by the docking score are shown. Some of the drugs shown here are currently in clinical trials for the treatment of COVID-19 (highlighted in light green). Table S6 : A list of drug hits that found to be in common between a previously published SARS-COV-2 interactome study (209) and our virtual screens. These drugs, upon experimental validation, could be potentially developed into multi-target drugs capable of modulating both host and viral targets simultaneously. Three of the drugs listed here are currently undergoing clinical trials for the treatment of COVID-19 (highlighted in light green). 

A pneumonia outbreak associated with a new coronavirus of probable bat origin

A new coronavirus associated with human respiratory disease in China

The proximal origin of SARS-CoV-2

A novel coronavirus from patients with pneumonia in China

An interactive web-based dashboard to track COVID-19 in real time

COVID-19 dashboard by the Center for Systems Science and Engineering (CSSE) at Johns Hopkins University (JHU), 2020b

The species severe acute respiratory syndrome-related coronavirus: classifying 2019-nCoV and naming it SARS-CoV-2

Clinical features of patients infected with 2019 novel coronavirus in wuhan, china. The Lancet

Virological assessment of hospitalized patients with COVID-2019

Smell and taste dysfunction in patients with COVID-19. The Lancet Infectious Diseases

Transmission of 2019-nCoV infection from an asymptomatic contact in germany

Follow-up of asymptomatic patients with SARS-CoV-2 infection

The molecular biology of coronaviruses

Supramolecular architecture of the coronavirus particle

Structural basis for the recognition of SARS-CoV-2 by full-length human ACE2

Structure, function, and antigenicity of the SARS-CoV-2 spike glycoprotein

Host cell proteases: Critical determinants of coronavirus tropism and pathogenesis

The nonstructural proteins directing coronavirus RNA synthesis and processing

Roles of Host Gene and Non-coding RNA Expression in Virus Infection

Epidemiology, genetic recombination, and pathogenesis of coronaviruses

A novel coronavirus associated with severe acute respiratory syndrome

Newly discovered coronavirus as the primary cause of severe acute respiratory syndrome

Coronavirus as a possible cause of severe acute respiratory syndrome

Identification of a novel coronavirus in patients with severe acute respiratory syndrome

Middle East respiratory syndrome coronavirus (MERS-CoV): Announcement of the Coronavirus Study Group

Genomic characterization of a newly discovered coronavirus associated with acute respiratory distress syndrome in humans

Isolation of a novel coronavirus from a man with pneumonia in Saudi Arabia

Origin and evolution of pathogenic coronaviruses

Genomic characterisation and epidemiology of 2019 novel coronavirus: implications for virus origins and receptor binding

Vaccines for the 21st century

Clinical assessment of a novel recombinant simian adenovirus ChAdOx1 as a vectored vaccine expressing conserved influenza a antigens

Modified mRNA-based vaccines elicit robust immune responses and protect guinea pigs from ebola virus disease

Safety and efficacy of the ChAdOx1 nCoV-19 vaccine (AZD1222) against

-2: an interim analysis of four randomised controlled trials in Brazil, South Africa, and the UK. The Lancet, 2020

An mRNA vaccine against SARS-CoV-2 â " preliminary report

Safety and efficacy of the BNT162b2 mRNA Covid-19 vaccine

Remdesivir in adults with severe COVID-19: a randomised, double-blind, placebocontrolled, multicentre trial. The Lancet

Remdesivir for the treatment of Covid-19 â " final report

Ultra-large library docking for discovering new chemotypes

Discovery of lysine-targeted eIF4E inhibitors through covalent docking

Virtual discovery of melatonin receptor ligands to modulate circadian rhythms

An open-source drug discovery platform enables ultra-large virtual screens

SARS-CoV-2 cell entry depends on ACE2 and TMPRSS2 and is blocked by a clinically proven protease inhibitor

Structure of the SARS-CoV-2 spike receptorbinding domain bound to the ACE2 receptor

Structural basis of receptor recognition by SARS-CoV-2

A novel angiotensin-converting enzyme-related carboxypeptidase (ACE2) converts angiotensin i to angiotensin 1-9

Receptor and viral determinants of SARS-coronavirus adaptation to human ACE2

Molecular Dynamics Simulations Related to SARS-CoV-2

Cryo-EM structure of the 2019-nCoV spike in the prefusion conformation

Cell entry mechanisms of SARS-CoV-2

Cryo-EM structures of MERS-CoV and SARS-CoV spike glycoproteins reveal the dynamic receptor binding domains

Cryo-electron microscopy structures of the SARS-CoV spike glycoprotein reveal a prerequisite conformational state for receptor binding

situ structural analysis of SARS-CoV-2 spike reveals flexibility mediated by three hinges

Structures and distributions of SARS-CoV-2 spike proteins on intact virions

ADAM17 cleave ACE2 differentially and only proteolysis by TMPRSS2 augments entry driven by the severe acute respiratory syndrome coronavirus spike protein

The androgen-regulated protease TMPRSS2 activates a proteolytic cascade involving components of the tumor microenvironment and promotes prostate cancer metastasis

Phenotypic analysis of mice lacking the Tmprss2-encoded protease

Catalytic cleavage of the androgen-regulated TMPRSS2 protease results in its secretion by prostate and prostate cancer epithelia

Fusion mechanism of 2019-nCoV and fusion inhibitors targeting HR1 domain in spike protein

Structures and mechanisms of viral membrane fusion proteins: Multiple variations on a common theme

Interaction between heptad repeat 1 and 2 regions in spike protein of SARS-associated coronavirus: implications for virus fusogenic mechanism and identification of fusion inhibitors

Inhibition of SARS-CoV-2 (previously 2019-nCoV) infection by a highly potent pan-coronavirus fusion inhibitor targeting its spike protein that harbors a high capacity to mediate membrane fusion

Structure-based discovery of Middle East respiratory syndrome coronavirus fusion inhibitor

A pancoronavirus fusion inhibitor targeting the HR1 domain of human coronavirus spike

Coronavirus main proteinase (3CLpro) structure: Basis for design of anti-SARS drugs

Mechanism for controlling the dimer-monomer switch and coupling dimerization to catalysis of the severe acute respiratory syndrome coronavirus 3C-like protease

Only one protomer is active in the dimer of SARS 3C-like proteinase

Molecular dynamic simulations analysis of ritronavir and lopinavir as SARS-CoV 3CLpro inhibitors

The crystal structures of severe acute respiratory syndrome virus main protease and its complex with an inhibitor

Crystal structure of SARS-CoV-2 main protease provides a basis for design of improved -ketoamide inhibitors

Structure of M pro from SARS-CoV-2 and discovery of its inhibitors

Structure-based design of antiviral drug candidates targeting the SARS-CoV-2 main protease. Science, page eabb4489

Structural basis for the inhibition of SARS-CoV-2 main protease by antineoplastic drug carmofur

SARS-CoV-2 3CL protease (3CL pro) in complex with a novel inhibitor

SARS-CoV-2 main protease with unliganded active site (2019-nCoV

SARS-CoV-2 3CL protease (3CL pro) apo structure (space group C21)

Structural and evolutionary analysis indicate that the SARS-CoV-2 mpro is a challenging target for small-molecule inhibitor design

Activation and maturation of SARS-CoV main protease

Characterization of SARS main protease and inhibitor assay using a fluorogenic substrate

Critical assessment of important regions in the subunit association and catalytic action of the severe acute respiratory syndrome coronavirus main protease

Foldon unfolding mediates the interconversion between Mpro-C monomer and 3D domain-swapped dimer

The SARS-coronavirus papain-like protease: Structure, function and inhibition by designed antiviral compounds

The papain-like protease of severe acute respiratory syndrome coronavirus has deubiquitinating activity

Discovery of an essential nucleotidylating activity associated with a newly delineated conserved domain in the RNA polymerase-containing protein of all nidoviruses

Insights into RNA synthesis, capping, and proofreading mechanisms of SARS-coronavirus

Coronaviruses lacking exoribonuclease activity are susceptible to lethal mutagenesis: Evidence for proofreading and potential therapeutics

Structure of the RNA-dependent RNA polymerase from COVID-19 virus. Science, page eabb7498

Structure of the SARS-CoV nsp12 polymerase bound to nsp7 and nsp8 co-factors

A second, noncanonical RNA-dependent RNA polymerase in SARS coronavirus

The SARS-coronavirus nsp7+nsp8 complex is a unique multimeric RNA polymerase capable of both de novo initiation and primer extension

Multiple enzymatic activities associated with severe acute respiratory syndrome coronavirus helicase

Mechanism of nucleic acid unwinding by SARS-CoV helicase

Delicate structural coordination of the severe acute respiratory syndrome coronavirus nsp13 upon ATP hydrolysis

The SWISS-MODEL repository-new features and functionality

mini replication and transcription complex

Crystal structure of the SARS-CoV-2 helicase at 1.94 angstrom resolution

Proteolytic processing, deubiquitinase and interferon antagonist activities of Middle East respiratory syndrome coronavirus papain-like protease

Molecular determinants of severe acute respiratory syndrome coronavirus pathogenesis and virulence in young and aged mouse models of human disease

The nsp9 replicase protein of SARS-coronavirus, structure and functional insights

Structural basis for dimerization and RNA binding of avian infectious bronchitis virus nsp9

Severe acute respiratory syndrome coronavirus nsp9 dimerization is essential for efficient viral growth

Imbalanced host response to SARS-CoV-2 drives development of COVID-19

Comparative replication and immune activation profiles of SARS-CoV-2 and SARS-CoV in human lungs: An Ex Vivo study with implications for the pathogenesis of COVID-19

Temporal profiles of viral load in posterior oropharyngeal saliva samples and serum antibody responses during infection by SARS-CoV-2: an observational cohort study. The Lancet Infectious Diseases

Nucleic acid recognition by the innate immune system

2'-O methylation of the viral mRNA cap evades host restriction by IFIT family members

An "old" protein with a new story: Coronavirus endoribonuclease is important for evading host antiviral defenses

Deubiquitinating and interferon antagonism activities of coronavirus papain-like proteases

High fidelity of murine hepatitis virus replication is decreased in nsp14 exoribonuclease mutants

Functional screen reveals SARS coronavirus nonstructural protein nsp14 as a novel cap N7 methyltransferase

Discovery of an RNA virus 3'-¿5' exoribonuclease that is critically involved in coronavirus RNA synthesis

RNA 3'-end mismatch excision by the severe acute respiratory syndrome coronavirus nonstructural protein nsp10/nsp14 exoribonuclease complex

Mutations in coronavirus nonstructural protein 10 decrease virus replication fidelity

The curious case of the nidovirus exoribonuclease: Its role in RNA synthesis and replication fidelity

A live, impaired-fidelity coronavirus vaccine protects in an aged, immunocompromised mouse model of lethal disease

Structural and molecular basis of mismatch correction and ribavirin excision from coronavirus RNA

Coronavirus susceptibility to the antiviral Remdesivir (GS-5734) is mediated by the viral polymerase and the proofreading exoribonuclease

Cocrystal structure of the messenger RNA 5' cap-binding protein (eIF4e) bound to 7-methyl-GDP

Conventional and unconventional mechanisms for capping viral mRNA

Structural basis and functional analysis of the SARS coronavirus nsp14-nsp10 complex

A viral RNA structural element alters host recognition of nonself RNA

Crystal structure and functional analysis of the SARScoronavirus RNA cap 2'-O-methyltransferase nsp10/nsp16 complex

Molecular mechanisms of coronavirus RNA capping and methylation

Biochemical and structural insights into the mechanisms of SARS coronavirus RNA ribose 2'-O-methylation by nsp16/ nsp10 protein complex

Short peptides derived from the interaction domain of SARS coronavirus nonstructural protein nsp10 can suppress the 2'-O-methyltransferase activity of nsp10/nsp16 complex

Binding of the methyl donor S-adenosyl-L-methionine to Middle East respiratory syndrome coronavirus 2'-O-methyltransferase nsp16 promotes recruitment of the allosteric activator nsp10

Coronavirus nonstructural protein 15 mediates evasion of dsRNA sensors and limits apoptosis in macrophages

Liurong Fang

Porcine deltacoronavirus nsp15 antagonizes interferon β-production independently of its endoribonuclease activity

Crystal structure of nsp15

endoribonuclease NendoU from SARS-CoV-2. Protein Science

Mutational analysis of the SARS virus nsp15 endoribonuclease: Identification of residues affecting hexamer formation

Structural basis for the ubiquitin-linkage specificity and deISGylating activity of SARS-CoV papainlike protease

Selectivity in ISG15 and ubiquitin recognition by the SARS coronavirus papain-like protease

The macro domain is an ADP-ribose binding module

A G-quadruplexbinding macrodomain within the "SARS-unique domain" is essential for the activity of the SARScoronavirus replication-transcription complex

The ADP-ribose-1''-monophosphatase domains of severe acute respiratory syndrome coronavirus and human coronavirus 229E mediate resistance to antiviral interferon responses

Mouse hepatitis virus liver pathology is dependent on ADP-ribose-1"-phosphatase, a viral function conserved in the alpha-like supergroup

Viral macro domains reverse protein ADP-ribosylation

Unique and conserved features of genome and proteome of SARS-coronavirus, an early split-off from the coronavirus group 2 lineage

ADP-ribose-1"-monophosphatase: a conserved coronavirus enzyme that is dispensable for viral replication in tissue culture

Adpribose-1"-phosphatase activities of the human coronavirus 229e and sars coronavirus x domains

Structural basis of severe acute respiratory syndrome coronavirus ADPribose-1''-phosphate dephosphorylation by a conserved domain of nsP3

SnapShot: ADP-ribosylation signaling

The conserved coronavirus macrodomain promotes virulence and suppresses the innate immune response during severe acute respiratory syndrome coronavirus infection

The nsp3 macrodomain promotes virulence in mice with coronavirus-induced encephalitis

Viral macrodomains: Unique mediators of viral replication and pathogenesis

ADP-ribosylhydrolase activity of chikungunya virus macrodomain is critical for virus replication and virulence

Structure and intracellular targeting of the SARS-coronavirus orf7a accessory protein

Severe acute respiratory of glycosylation interference

Severe acute respiratory syndrome coronavirus 7a accessory protein is a viral structural protein

Tetherin: Holding on and letting go

Severe acute respiratory syndrome coronavirus gene 7 products contribute to virus-induced apoptosis

ORF7a encoded accessory protein

An 81 nucleotide deletion in SARS-CoV-2 ORF7a identified from sentinel surveillance in arizona

Severe acute respiratory syndrome coronavirus protein 7a interacts with hSGT

SARS-CoV accessory protein 7a directly interacts with human LFA-1

Molecular interactions in the assembly of coronaviruses

Specific interaction between coronavirus leader RNA and nucleocapsid protein

Nucleocapsid protein recruitment to replication-transcription complexes plays a crucial role in coronaviral life cycle

Crystal structure of SARS-CoV-2 nucleocapsid protein RNA binding domain reveals potential unique drug targeting sites

Structural basis for the identification of the N-terminal domain of coronavirus nucleocapsid protein as an antiviral target

Assembly of severe acute respiratory syndrome coronavirus RNA packaging signal into virus-like particles is nucleocapsid dependent

Structure of the SARS coronavirus nucleocapsid protein RNA-binding dimerization domain suggests a mechanism for helical packaging of viral RNA

Supramolecular architecture of severe acute respiratory syndrome coronavirus revealed by electron cryomicroscopy

Architecture of the SARS coronavirus prefusion spike

A highly potent long-acting small-molecule HIV-1 capsid inhibitor with efficacy in a humanized mouse model

Safety and antiviral activity over 10 days following a single dose of subcutaneous gs-6207, a first-in-class, long-acting hiv capsid inhibitor in people living with hiv

Safety and pk of subcutaneous gs-6207, a novel hiv-1 capsid inhibitor

GWOVina: A grey wolf optimization approach to rigid and flexible receptor docking

Accounting of receptor flexibility in ultra-large virtual screens with VirtualFlow using a grey wolf optimization method

AutoDock Vina: improving the speed and accuracy of docking with a new scoring function, efficient optimization, and multithreading

Lessons Learned in Empirical Scoring with smina from the CSAR 2011 Benchmarking Exercise

Proteinligand scoring with convolutional neural networks

AutoDock VinaXB: implementation of xbsf, new empirical halogen bond scoring function, into AutoDock Vina

Vina-Carb: Improving glycosidic angles during carbohydrate docking

Rapid repurposing of drugs for COVID-19

Dexamethasone in hospitalized patients with covid-19 -preliminary report

Therapeutic potential of ciclesonide inhalation for COVID-19 pneumonia: Report of three cases

A retrospective cohort study of methylprednisolone therapy in severe patients with COVID-19 pneumonia

Abelson kinase inhibitors are potent inhibitors of severe acute respiratory syndrome coronavirus and middle east respiratory syndrome coronavirus fusion

Repurposing of clinically developed drugs for treatment of middle east respiratory syndrome coronavirus infection

Woo Song Lee, and Young Bae Ryu. Tanshinones as selective and slow-binding inhibitors for SARS-CoV cysteine proteases

Discovery of unsymmetrical aromatic disulfides as novel inhibitors of SARS-CoV main protease: Chemical synthesis, biological evaluation, molecular docking and 3D-QSAR study

Design of wide-spectrum inhibitors targeting coronavirus main proteases

Severe acute respiratory syndrome coronavirus papain-like novel protease inhibitors: Design, synthesis, protein-ligand x-ray structure and biological evaluation

Thiopurine analogues inhibit papain-like protease of severe acute respiratory syndrome coronavirus

Thiopurine analogs and mycophenolic acid synergistically inhibit the papain-like protease of middle east respiratory syndrome coronavirus

Small-molecule antiviral β-d-n4-hydroxycytidine inhibits a proofreading-intact coronavirus with a high genetic barrier to resistance

Broad-spectrum antiviral GS-5734 inhibits both epidemic and zoonotic coronaviruses

Tryptophan-dependent membrane interaction and heteromerization with the internal fusion peptide by the membrane proximal external region of SARS-CoV spike protein

Identification of novel small-molecule inhibitors of severe acute respiratory syndrome-associated coronavirus by chemical genetics

Small molecule inhibitors of the SARS-CoV nsp15 endoribonuclease. Virus Adaptation and Treatment

A novel chemical compound for inhibition of SARS coronavirus helicase

Identification of a novel small molecule inhibitor against SARS coronavirus helicase

Comparative genomic analysis MERS CoV isolated from humans and camels with special reference to virus encoded helicase

High throughput virtual screening to discover inhibitors of the main protease of the coronavirus SARS-CoV-2. Preprints

Networkbased drug repurposing for novel coronavirus 2019-nCoV/SARS-CoV-2

High throughput virtual screening reveals SARS-CoV-2 multi-target binding natural compounds to lead instant therapy for COVID-19 treatment

Potential inhibitors for novel coronavirus protease identified by virtual screening of 606 million compounds

Rapid identification of potential inhibitors of SARS-CoV-2 main protease by deep docking of 1.3 billion compounds

Structure-based drug designing and immunoinformatics approach for SARS-CoV-2. Science Advances, page eabb8097

Computational screening of antagonists against the SARS-CoV-2 (COVID-19) coronavirus by molecular docking

Coronavirus disease 2019 drug discovery through molecular docking

Identification of novel compounds against three targets of SARS CoV-2 coronavirus by combined virtual screening and supervised machine learning. Bulletin of the World Health Organization

A SARS-CoV-2 protein interaction map reveals targets for drug repurposing

The global phosphorylation landscape of SARS-CoV-2 infection

Free Energy Methods Involving Quantum Physics, Path Integrals, and Virtual Screenings: Development, Implementation and Application in Drug Discovery. Dissertation

Benchmarking sets for molecular docking

Directory of Useful Decoys, Enhanced (DUD-E): Better ligands and decoys for better benchmarking

The history of antiretroviral therapy and of its implementation in resource-limited areas of the world

Lamivudine in combination with zidovudine, stavudine, or didanosine in patients with HIV-1 infection. a randomized, double-blind, placebo-controlled trial

Declining morbidity and mortality among patients with advanced human immunodeficiency virus infection

Antiviral combinations for severe influenza. The Lancet Infectious Diseases

Efficacy of clarithromycin-naproxen-oseltamivir combination in the treatment of patients hospitalized for influenza a(h3n2) infection

ProViz -a web-based visualization tool to investigate the functional and evolutionary features of protein sequences

Broad-spectrum coronavirus antiviral drug discovery

The 2019-new coronavirus epidemic: Evidence for virus evolution

Genomic epidemiology of novel coronavirus -global subsampling

Structural genomics of SARS-CoV-2 indicates evolutionary conserved functional regions of viral proteins

Extreme evolutionary conservation of functionally important regions in H1N1 influenza proteome

From mosquitos to humans: Genetic evolution of zika virus

Patterns of amino acid conservation in human and animal immunodeficiency viruses

Culturing the unculturable: Human coronavirus HKU1 infects, replicates, and produces progeny virions in human ciliated airway epithelial cell cultures

Tropism and replication of middle east respiratory syndrome coronavirus from dromedary camels in the human respiratory tract: an in-vitro and ex-vivo study

An orally bioavailable broad-spectrum antiviral inhibits SARS-CoV-2 in human airway epithelial cell cultures and multiple coronaviruses in mice

Induced protein degradation: an emerging drug discovery paradigm

Structural basis of PROTAC cooperative recognition for selective protein degradation

ZINC 15 -ligand discovery for everyone

Protein-ligand blind docking using QuickVina-W with inter-process spatio-temporal integration

Fast, accurate, and reliable molecular docking with QuickVina 2

AutoDock4 and AutoDockTools4: Automated docking with selective receptor flexibility

DataWarrior: An opensource program for chemistry aware data visualization and analysis

Thermodynamics and Kinetics of Drug Binding. Methods and Principles in Medicinal Chemistry

Structure-based design, synthesis, and biological evaluation of a series of novel and reversible inhibitors for the severe acute respiratory syndrome-coronavirus papain-like protease

Inhibitor recognition specificity of MERS-CoV papain-like protease may differ from that of SARS-CoV

Evaluation of polyphenols from Broussonetia papyrifera as coronavirus protease inhibitors

A noncovalent class of papain-like protease/deubiquitinase inhibitors blocks SARS virus replication

The recent outbreaks of human coronaviruses: A medicinal chemistry perspective

Stable benzotriazole esters as mechanism-based inactivators of the severe acute respiratory syndrome 3CL protease

Design, synthesis, and evaluation of inhibitors for severe acute respiratory syndrome 3C-like protease based on phthalhydrazide ketones or heteroaromatic esters

synthesis and antiviral efficacy of a series of potent chloropyridyl ester-derived SARS-CoV 3CLpro inhibitors

Synthesis and evaluation of isatin derivatives as effective SARS coronavirus 3CL protease inhibitors

Isatin compounds as noncovalent SARS coronavirus 3C-like protease inhibitors

Identification, synthesis and evaluation of SARS-CoV and MERS-CoV 3C-like protease inhibitors

Synthesis and evaluation of pyrazolone compounds as SARS-coronavirus 3C-like protease inhibitors

Ultra-large in silico screen of targets important for the replication of SARS-CoV-2 (85) • Screening of multiple functional sites on individual target proteins

• 17 target proteins, 45 screens, ~50 billion docking instances to target SARS-CoV-2 (84) • Conservation in some target sites means hits could exhibit pan-coronavirus function

This work was supported by a Google Cloud COVID- 19 Research Grant, and we thank Anna-Bettis Padgett for help with formal issues related to the grant and appropriate attributions of credit. We thank Ankita Chaudhary for help with curating the data. We thank Brigitte Klein for proofreading the manuscript. We would like to thank Arthur Jaffe and Meng Zhang for their support. We thank Jerry Parks, Jeremy Smith, and Jerome Baudry for helpful discussions. Declaration of Interests M.C. and V.D. work for Innophore. C.C.G. is shareholder and CEO of Innophore, an Enzyme and Drug Discovery company. G.W. and C.G. are cofounders of the company Virtual Discovery, Inc., which provides virtual screening services. G.W. serves as the director of this company. G.W. is cofounder of PIC Therapeutics, Cellmig Biolabs and Skinap Therapeutics. H.A. is an equity holder in PIC Therapeutics. I.I. and D.R. work for Enamine, a company that is involved in the synthesis and distribution of chemical building blocks, fragments, and screening compounds. Y.M. is a scientific advisor for Enamine. Y.M., O.T. and A.P. work for Chemspace, a company that is involved in the distribution of chemical building blocks, fragments, and screening compounds. I.D. works for UkrOrgSyntez Ltd. (UORSY), a company that is involved in the synthesis of chemical building blocks, fragments, and screening compounds. G.T.,S.P.,A.S.,M.G.,N.L.,C.H.,E.Y.,R.L.,R.Y.,D.P. and J.K. work for Google, a company which also provides cloud computing services. The research described here is scientifically and financially independent of the efforts in any of the abovementioned companies.

11-beta-Hydroxyandrosterone-3glucuronide No nsp16 (41) , TMPRSS2