FDA-approved Pralatrexate identied by virtual drug screening inhibits SARS-CoV-2 replication in vitro

The severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) pandemic poses serious threats to the global public health and leads to an unprecedented worldwide crisis. Unfortunately, no effective drugs or vaccines are available till now. Since the RNA-dependent RNA polymerase (RdRp) of SARS-CoV-2 is a promising therapeutic target, a deep learning and molecular simulation based hybrid drug screening procedure was proposed and applied to identify potential drug candidates targeting RdRp from 1906 approved drugs. Among the four selected FDA-approved drug candidates, Pralatrexate and Azithromycin were conrmed to effectively inhibit SARS-CoV-2 replication in vitro with EC 50 values of 0.008µM and 9.453 µM, respectively. For the rst time, our study discovered that Pralatrexate is able to potently inhibit SARS-CoV-2 replication with a stronger inhibitory activity than Remdesivir within the same experimental conditions. The paper demonstrates the feasibility of accurate virtual drug screening for inhibitors of SARS-CoV-2 and provides potential therapeutic agents against COVID-19.


Introduction
The Coronavirus Disease 2019 (COVID-19) caused by severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) has developed into a global pandemic with millions of people infected and tens of thousands of lives being lost [1]. SARS-CoV-2 can be transmitted from person to person with stronger infection ability than SARS-CoV [2,3]. The presence of asymptomatic transmission and a long incubation period results in an extra challenge in preventing SARS-CoV-2 transmission and spread [4]. As of 11th July 2020, more than 12,556,000 human infections with more than 558,000 deaths have been reported.
To date, there are no clinically approved drugs or vaccines available for the treatment or prevention of COVID-19. Therefore, identifying antiviral agents that can combat SARS-CoV-2 is of high signi cance.
De novo drug development process is time-consuming and costly, which cannot meet the urgent need to combat COVID-19. Given current emergencies, repurposing existing FDA-approved drugs for COVID-19 provides a shortcut [5]. Recently, some drugs under clinical trials such as Remdesivir were shown to inhibit the replication of SARS-CoV-2 in vitro [6][7][8]. The structural basis of the RNA-dependent RNA polymerase (RdRp) inhibited by Remdesivir is well illustrated in a recent work [9]. Some severe SARS-CoV-2 patients treated with compassionate-use Remdesivir targeting RdRp, have shown a signi cant clinical improvement [10]. As a core component of the RNA synthesis machinery, RdRp is believed to be one of the most promising therapeutic targets for COVID-19 [11,12]. Small compounds binding with the catalytic site of RdRp have potential to interfere with SARS-CoV-2 RNA synthesis [13].
Several computational drug screening methods relying on molecular docking, deep learning or Molecular Dynamics (MD) simulations have been applied in drug repositioning studies for COVID-19 [12]. However, most studies normally rely on a single technique or lack experimental validation. Each computational technique has different pros and cons, and a proper combination and modi cations of such methods may provide a better solution. Previously, we have developed two deep learning-based models to estimate the protein-ligand interaction, the DFCNN [14] and DeepBindBC (http://hpcc.siat.ac.cn/DeepBindBC/). Instead of using protein-ligand docking conformation, DFCNN uses molecular vectors of protein pocket and ligand to estimates the protein-ligand pair as binding or non-binding with a probability value between 0 and 1. On the other hand, DeepBindBC estimates the binding possibility from the interface atom contact information of the modelled 3D protein-ligand complex. The input of DeepBindBC contains spatial information of the protein-ligand interface, thus it is strongly complementary to DFCNN.
In the present work, we propose a deep learning and molecular simulation based hybrid screening method which consists of DFCNN [14], DeepBindBC, Autodock Vina [15], pocket molecular dynamics simulation, and metadynamics to explore the potential 1906 drugs from TargetMol-Approved_Drug_Library, a collected drug library by TargetMol. We systematically selected four FDA-approved drugs for experimental validation, Pralatrexate was found to e ciently inhibit the replication of SARS-CoV-2 in vitro.

Results
The overall work ow of virtual drug screening against RdRp is illustrated in Fig. 1a. The interaction patterns (taken from the last frame of 100 ns MD simulation) between RdRp and Azithromycin, Pralatrexate are shown in Fig. 1b(i)(ii), respectively. The interaction between Remdesivir in its monophosphate form and RdRp (PDB ID: 7BV2) is also given for comparison in Fig 1b(iii). Azithromycin forms 2 hydrogen bonds with GLN573 and ILE494 through keto and hydroxyl groups respectively, and many hydrophobic related interactions with the RdRp binding site (e.g. LYS577, LEU576, ALA685) through alkyl groups, whereas Pralatrexate shows enhanced and more stable interactions with RdRp binding site, including 6 hydrogen bonds with GLN573, ARG569, ASN496, ASN497, LYS500 and GLY590. Pralatrexate also forms Alkyl or Pi-Alkyl interaction with LYS577 and LEU576, and salt bridges with ARG569 and LYS500. Azithromycin and Pralatrexate share 10 common neighbor residues (62.5%) of RdRp, as shown in Fig 1b(i)  A recent study shows the Remdesivir in its monophosphate form interacts with the RdRp involving with partial double-stranded RNA template, and covalently incorporated into the primer strand at the +1 position [9]. Shown in Fig. 1b 1) selected by the molecular vector-based and structure-based screening process. Sofosbuvir is both a nucleoside analogue and an antivirus drug. These 14 drugs are subject to force eld-based screening in the next stage.
Among 2 antibiotic drugs, Azithromycin, a drug used to treat a variety of bacterial infections, showed top Autodock Vina score of -8.6 kcal/mol, good DFCNN score and DeepBindBC score (0.9093 and 0.8589), respectively. Gautret, P. et al claimed that combined with Hydroxychloroquine, Azithromycin can have good e ciency in treating COVID-19 with signi cant viral load reduction [16]. However it should be noticed that currently there is no evidence of the effectiveness of Azithromycin in the treatment of COVID- 19 and have many debates about effective of Azithromycin on COVID-19 [17].
The top two predicted molecules by DeepBindBC are nucleotide analogues. Sofosbuvir is a nucleotide analogue inhibitor of hepatitis C virus (HCV) NS5B polymerase to treat infectious liver disease, [18] whereas Clofarabine is a purine nucleoside antimetabolite used for treating refractory acute lymphoblastic leukaemia [19]. More nucleotides analogues in the candidate list were selected by our method, such as Adenosine, Vidarabine, and Gemcitabine, indicating some RdRp-nucleotides interaction patterns have been implicitly recognized by the proposed hybrid drug screening method. The binding free energy vs coordination number (CV: collective variable) from metadynamics simulations is shown in Supplementary Fig. 4a. The lowest energy conformations of protein-drug complexes for Amoxicillin, Azithromycin, Pralatrexate and Sofosbuvir showed more contacts in the interface region, as indicated by the high coordination numbers ( Supplementary Fig. 4a (iii)), while most other compounds favor smaller coordination number (close to zero) indicating no or weak interactions (Supplementary Fig.   4a (i)).
Detailed interaction patterns between RdRp and the four most optimal compounds (Azithromycin, Pralatrexate, Amoxicillin and Sofosbuvir) are shown in Fig. 1b(i)(ii) and Supplementary Fig. 4b (i)(ii), whose structures are taken from the last frame of the 100 ns MD simulation. Azithromycin and Pralatrexate interact with 16 amino acid residues of RdRp to form a stable complex. The RdRp-Azithromycin complex is mainly dominated by van der waals interactions, whereas Pralatrexate involves more polar and charge interactions. According to the calculated free energy difference (ΔG) values from the metadyanmics simulations between the unbound state and the binding state for Amoxicillin, Azithromycin, Pralatrexate and Sofosbuvir (Supplementary Table 3), Azithromycin and Pralatrexate (-305.76 kJ/mol, -128.58 kJ/mol) show more favorable binding energy than Amoxicillin and Sofosbuvir (-67.33 kJ/mol and -89.88kJ/mol).
It is noted that all the nucleoside analogues highly recommended by deep learning-based screening methods were excluded from the force eld based screening process. The possible explanation is that our protein-drug systems do not contain the RNA primers during the MD simulation, and covalent bond formation, such as the Remdesivir in its monophosphate form, could not be estimated by traditional MD simulation.

Pralatrexate and Azithromycin inhibit the replication of SARS-CoV-2 in vitro.
To further con rm the e ciency of the hits from the virtual screening, we tested the antiviral activity of the Azithromycin, Pralatrexate, Amoxicillin and Sofosbuvir in vitro. Vero cells were infected with SARS-CoV-2 (BetaCoV/Shenzhen/SZTH-003/2020, GISAID No. EPI_ISL_406594) at a MOI of 0.02 ( the cytopathic effect was mild at 48 hours post-infection with this MOI) in the presence of varying concentrations of the tested drugs, and the inhibition rates were evaluated by quanti cation of viral copy numbers in the cell supernatant via quantitative reverse transcription polymerase chain reaction (qRT-PCR) and con rmed with immuno uorescence assay (Fig. 2). The results showed that Pralatrexate and Azithromycin could e ciently inhibit the replication of SARS-CoV-2, with half-maximal effective concentration (EC 50 ) values of 0.008 and 9.453 μM (Fig. 2a), whereas Remdesivir achieved an inhibitory activity with EC 50 value of 8.777 μM within the same experimental system ( Supplementary Fig. 5). IFA showed similar results with qRT-PCR assay (Fig. 2b). CCK-8 assay of the two drugs showed that the halfcytotoxic concentration (CC 50 ) values of Pralatrexate and Azithromycin on Vero cells were 0.167 μM and > 100 μM, respectively, and the calculated the selectivity indexes (SI) of Pralatrexate and Azithromycin were 20.878 and >10.579, respectively. Whether the two drugs worked at the stage of viral entry or post entry was analyzed using time-of-addition assay as previously reported [8]. The results showed that Pralatrexate functioned at a stage post virus entry, while Azithromycin functioned at both entry and postentry stages of the SARS-COV-2 infection in Vero cells (Fig. 2c). Furthermore, surface plasmon resonance (SPR) experiments were performed to test the in vitro binding of Pralatrexate and Azithromycin with immobilized RdRp protein of SARS-CoV-2. Both drugs showed obvious binding response in Supplementary Fig. 6.

Discussion
To perform the drug screening process e ciently and accurately is still a challenge for computer-aided drug design. Though a recent deep learning-based approaches has demonstrated its potential to be e cient/accurate by learning from a su cient amount of training data, problems such as over tting, and the discrepancy between training data and real-world data remain [20]. The proposed deep learning and molecular simulation based drug screening method was able to select 4 FDA-approved drug candidates targeting RdRp from 1906 drugs, and 2 out of 4 (Pralatrexate and Azithromycin) can effectively inhibit SARS-CoV-2 replication in vitro with EC 50 values of 0.008µM and 9.453 µM. The molecular vector-based deep learning method and the structure-based deep learning method are complementary to each other in the sense that high e ciency and accuracy are both achieved.
For the rst time, Pralatrexate is found to potently inhibit SARS-CoV-2 replication in vitro with a stronger inhibitory activity (EC 50 value 0.008µM) than Remdesivir (P<0.0001) within the same experimental system. Compared with the GHDDI drug list (The Global Health Drug Discovery Institute: https://ghddiailab.github.io/Targeting2019-nCoV/preclinical/.) that inhibit SARS-CoV-2 in vitro, Pralatrexate showed the smallest EC 50. Among the 154 current reported drugs by GHDDI, NSC319726 have top inhibitory activity over SARS-CoV-2 (EC 50 value <0.02µM) [21]. Pralatrexate is a folate analogue metabolic inhibitor, which was approved by FDA in 2009 for the treatment of patients with relapsed or refractory peripheral T cell lymphoma (PTCL). Pralatrexate inhibits the folate metabolism pathway through inhibition of dihydrofolate reductase (DHFR) [22]. The peak concentration in plasma (Cmax) can achieve 10.5 μM from a standard dosing regimen [23]. Its Cmax is around 800-fold higher than the EC 90 of antiviral activity, suggesting a great potential for clinical implications.
Pralatrexate was selected by the virtual screening pipeline based on its potential acts of inhibiting the RNA dependent RNA protease (RdRp) enzyme, whereas, it's extremely low EC50 for the virus replication compared to Remdesivir (RdRp inhibitor) may have multiply mechanism of action. Pralatrexate is known to be an antifolate that e ciently prevents synthesis of DNA and presumably also RNA [24], which may explain inhibition of SARS-CoV-2 replication. Pralatrexate was approved by FDA in spite of its toxicity, therefore, we should be aware that FDA approval does not guarantee the possibility of immediate use of the drug against COVID-19.
Though both Pralatrexate and Azithromycin inhibit SARS-CoV-2 replication in vitro, the time-of-addition experiment showed that they functioned at different stages of SARS-CoV-2 infection. Similar to Remdesivir, Pralatrexate mainly inhibited the replication of SARS-CoV-2 at the stages of post-entry. On the other hand, Azithromycin inhibited the replication of SARS-CoV-2 at both entry and post-entry stages like chloroquine [8]. This indicates the Azithromycin may also have multiple mechanism of action.
Out of the 4 selected drug candidates targeting RdRp, Amoxicillin and Sofasbuvir have failed to inhibit SARS-CoV-2 replication in vitro. Molecular dynamic simulations show they have deviated from its initial binding position ( Supplementary Fig. 4 a(i) a(ii)) with their ligand RMSD > 1.5 nm for most of the simulation time and large uctuation was observed ( Supplementary Fig. 2). The calculated free energies difference between binding state and unbound state (coordination number around 0) also indicates better binding for Pralatrexate and Azithromycin than Amoxicillin and Sofasbuvir, shown in Supplementary Fig.   4 a(iii) a(iv) and Supplementary Table 3.
To examine why Sofosbuvir can e ciently inhibit RdRp of hepatitis C virus (HCV) [18] while not RdRp of SARS-CoV-2, we have carried a sequence and structural comparison between RdRp of HCV and RdRp of SARS-CoV-2 virus (Supplementary Fig. 7). In addition to the low sequence identity (23.75%) between RdRp of HCV and RdRp of SARS-CoV-2 virus, binding pockets of both complexes showed a quite different composition. For instance, there are 5 vs 3 ASPs, 2 vs 1 LYSs, 1 vs 3 GLUs, 0 vs 6 ARGs in RdRp pocket of SARS-CoV-2 and RdRp pocket of HCV, respectively. The RdRp pocket of SARS-CoV-2 is more negatively charged, while the RdRp pocket of HCV is more positively charged.
Full system protein-ligand MD simulations for RdRp-Pralatrexate, RdRp-Azithromycin were performed to validate the robustness of the pocket MD method. Compared to pocket MD simulation, similar hydrogen bond numbers as well as similar low RMSD uctuations in full MD simulation were observed according to Supplementary Fig. 3 and Supplementary Fig. 8a(i)(ii). Some key neighbor residues in pocket MD simulation for Azithromycin and Pralatrexate were also kept during the full system MD simulation, according to Fig. 1b(i)(ii) and Supplementary Fig. 8b(i)(ii). For instance, LEU576, ILE589, ALA580 and ALA685 have formed alkyl related hydrophobic interaction with Azithromycin in the last frames of both simulations, and ARG569, ASN496 and LYS500 of RdRp have formed salt bridge or hydrogen bonds with Pralatrexate in the last frames of both simulations.
The e ciency and effectiveness of the DFCNN method have been examined by screening about 10 million drugs targeting 8 representative protein targets taken from the DUD.E diverse data set. DFCNN was able to screen the 10 million drugs within 5 hours using a workstation with 80 Intel CPU cores (2.00 GHz ) and 60 GB RAM. The effectiveness is evaluated by the prediction-random ratio (Ratio 0.9 ), shown in Supplementary Table 4. For 6 out of 8 protein targets, Ratio 0.9 is greater than 1.4, indicating DFCNN is able to enrich the active compounds in ten million compound pools. Among the 8 test cases, the DFCNN achieved best performance on HIVPR (Human immunode ciency virus type 1 protease) with Ratio 0.9 of 860 ( about 860 times better than random guess in selecting active compounds in terms of TPR). DFCNN performed worse for GPCR proteins (such as CXCR4) and protein with small inner pocket (such as AKT1). The possible reason is that GPCRs have limited number of reliable structure of protein-ligand complexes in our training dataset and membrane proteins may have very different binding mechanism compared to other type proteins. The poor performance for proteins with small inner pocket is likely due to the special physical-chemical and spatial features. As an enzyme, RdRp has large ligand binding cavity and should be suitable for virtual screening by DFCNN.
To study how the molecular vector-based deep learning screening method selects the 139 candidate drugs from 1906 drugs, 1906 drugs were clustered into 20 groups (Supplementary Fig. 9). Group 20 has the highest ratio of drugs being selected (31/89 drugs, Supplementary Fig. 9a). The drugs in the Groups 19, Groups 20, Group 17 and Group 15 with high selection ratio tend to contain many electrical donors and electrical acceptors, likely due to the RdRp pocket containing many charged groups ( Supplementary  Fig. 10), including 5 ASPs, 2 LYSs, and 1 GLU. The percentage of charge and polar residues in the RdRp pocket reaches 54.35% (Supplementary Fig. 10), which explains why DFCNN prefers to select polar and charged drugs for the RdRp. The structure-based screening (Autodock Vina plus DeepBindBC) selected 14 drugs from 139 drugs, 6 drugs belong to Group 15 (Supplementary Table 5), including nucleotide analogues as well as Pralatrexate. Groups 17, 18, 19 all have 2 drugs selected after structure-based screening. Belonging to different clusters, Pralatrexate has many hydrogen donors and acceptors while Azithromycin contains a macrocycle, which tends to form hydrogen bond (or salt bridge) and macrocyclic hydrophobic interactions, respectively.

Conclusion
Identifying effective drugs that can treat COVID-19 is important and urgent, especially the approved drugs that can be immediately tested in clinical trials. In this work, we have developed a hybrid protocol of combining deep learning methods with molecular simulations to search for potential drug candidates against RdRp that can inhibit the replication of SARS-CoV-2. Four potential drugs were systematically selected for experimental validation, and Pralatrexate and Azithromycin showed an inhibiting effect with EC 50 values of 0.008µM and 9.453 µM, respectively. Experimental results from qRT-PCR, CCK-8 assay, indirect immuno uorescence assay (IFA), Time-of-addition and Surface plasmon resonance (SPR) assay show the proposed screening protocol successfully identi ed a new therapeutic agent Pralatrexate against COVID-19 by targeting RdRp. The hybrid strategy of combining deep learning, molecular docking, MD simulation in a virtual screening pipeline can effectively help with drug repurposing application and facilitate virtual drug screening against other targets in SARS-CoV-2.

Materials And Methods
In this paper, a deep learning and molecular simulation based hybrid strategy is proposed for virtual drug screening against RdRp over the TargetMol-Approved-Drug-Library, an approved drug library with 1906 compounds collected by TargetMol, resulting in four candidates (Pralatrexate, Azithromycin, Sofosbuvir, Amoxicillin) for drug repurposing. qRT-PCR assay, indirect immuno uorescence assay (IFA) and CCK-8 assay were carried out to validate the e cacy for Pralatrexate, Azithromycin which inhibit SARS-CoV-2 replication in vitro. Surface plasmon resonance (SPR) assay was used to evaluate the RdRp-drug binding a nity.

Structural modeling of RdRp and drug compound dataset
The RdRp sequence and its modelled structure were obtained from https://zhanglab.ccmb.med.umich.edu/C-I-TASSER/2019-nCov/. The RdRp-ligand model was constructed by I-TASSER [25]. The ligand was taken from the template protein (PDB ID: 3BR9) [26] by COFACTOR algorithm [27] within the I-TASSER using structure comparison and protein-protein networks. We extract the amino acids within 1 nm of the ligand as the binding pocket. RMSD between the modeled structure and the recent experimental RdRp structure (PDB ID 6M71) is calculated (~0.516 Å), shown in Supplementary Fig. 11a) [13]. RNA polymerase superfamily region is also very similar between these two structures (RMSD=0.456 Å, Supplementary Fig. 11a). TargetMol

Molecular vector-based drug screening
A deep learning-based method, DFCNN (Dense fully Connected Neural Network), has been developed for predicting protein-drug binding probability [14] and used in this paper for the initial drug screening (Fig.  1a). DFCNN utilizes the concatenated molecular vector of protein pocket and ligand as input representation, and the molecular vector are generated by Mol2vec [28] which is inspired by the word2vec model in natural language processing. DFCNN model was trained on a dataset extracted from PDBbind database [29]. Negative data samples in the dataset were generated by cross-combination of proteins and ligands from PDBbind database and positive data samples were taken from protein-ligand pairs in experimental structure. The details of the method were described in our previous paper [14], and DFCNN achieved an AUC value around 0.9 for the independent testing set [14]. The model is about ~100,000 times faster than Autodock Vina in predicting protein-ligand binding probability (range 0~1), because it does not rely on the protein-drug complex conformation.
To examine the e ciency and effectiveness of the DFCNN method, we screen a large scale chemical compound dataset (about 10 million compounds) targeting 8 representative protein targets taken from the DUD.E diverse data set. For each target, the corresponding dataset contains some active compounds (between 40 and 536) in the DUD.E dataset and 10,402, 895 drug-like compounds from ZINC database. The effectiveness is measured by the prediction-random ratio (Ratio 0.9 ), de ned as TPR 0.9 /Random 0.9 , where TPR 0.9 indicates the ratio (N 0.9 /Active_num) between the number of active compounds with a DFCNN score larger than 0.9 (N 0.9 ) and the active number of compounds (Active_num) . The total number of the compounds (Total_num) with score above 0.9 is de ned as NN. The random selection rate (Random 0.9 ) is de ned as NN/Total_num. Using cutoff score of 0.9, the prediction-random ratio measures the ratio of predicted TPR and random selection TPR.

Structure-based drug screening
DeepBindBC, an in-house deep learning-based software, is used for structure based drug screening. Unlike the DFCNN, the input of DeepBindBC includes both the physical-chemical information and spatial information between the protein-ligand interfaces (Fig. 1a), hence DeepBindBC is able to achieve higher accuracy, but requires protein-drug complex structure information as input generated by Autodock Vina.
Autodock Vina is used to dock the target with the potential ligands [15]. The pocket is determined by the location of ligand in the template protein (PDB ID: 3BR9) [26]. We set the cavity volume space with 3.5 nm, 3.5 nm and 3.5 nm in x, y, z dimensions from the pocket mass center. AutoDock Tools were used to convert PBD le format to PDBQT le format [30]. The exhaustiveness was set to 8; the num_modes was set to 20, and energy_range was set to 3. The scoring function and optimization algorithm of Autodock Vina have been well discussed in a previous article [15]. In this study, we selected the most likely targets for further validation by setting a binding energy threshold value of -7 kcal/mol.
The DeepBindBC is a ResNet model trained over the PDBbind database. In DeepBindBC, the protein-ligand interaction interface information will be converted into gure-like metric, similar to DeepBindRG [20]. By incorporating the cross-docking (docking proteins and ligands from different experimental complexes) conformation as negative training data, DeepBindBC is highly possible to distinguish non-binders. Since DeepBindBC relies on docking conformation and DFCNN only uses molecular vector information, these two methods are complementary to each other and DeepBindBC takes much more time than DFCNN.
We also proposed a pocket molecular dynamics simulation (pocket MD, Supplementary Fig. 11b) to facilitate the simulation process by only keeping the binding pocket region for simulation. Binding free energy calculation can be estimated by metadynamics simulations to explore whether protein-ligand will bind in solution. Metadynamics relies on addition of a bias potential to sample the free energy landscape along a speci c collective variable of interest [31] , [32]. Noted that the binding free energy calculations from Metadynamics may only be suitable for detect the general trend of binding in virtual screening.
The pocket MD is same as the classical MD simulation, except that we only using the pocket region to reduce system size for simulation ( Supplementary Fig. 11b), which is inspired by a previous dynamic undocking (DUck) method [33]. An in-house script was used to extract the pocket region of the protein (1nm within the binding ligand), the N terminal and C terminal ends were capped with the ACE and NHE, respectively. Terminals will be applied a position restrain to maintain the relative conformation of the pocket. MD simulation was carried out by Gromacs with AMBER-99SB force eld [34,35]. The topology of ligand and the partial charges of ligand was generated by ACPYPE [36], which relies on Antechamber [37]. Firstly, we created a dodecahedron box and put the target-ligand complex at the center. A minimum distance from the protein to box edge was set to 1 nm. We lled the dodecahedron box with TIP3P water molecules [38], the counter ions was added to neutralize the total charge using the Gromacs program tool [39]. The long-range electrostatic interactions under the periodic boundary conditions was calculated with Particle Mesh Ewald approach [40]. A cutoff of 14 Å was used for van der Waals non-bonded interactions.
Covalent bonds involving hydrogen atoms were constrained by applying the LINCS algorithm [41].
We performed the energy minimization steps with a step-size of 0.001ns, 100 ps simulation with isothermal-isovolumetric ensemble (NVT), and 10ns simulation with isothermal-isobaric ensemble (NPT) for water equilibrium. After that, a 100ns NPT production run (step size 2 fs) was carried out. The Parrinello-Rahmanbarostat and the modi ed Berendsen thermostat were used for simulation with a xed temperature of 308 K and a pressure of 1 atm. RMSD and hydrogen bond number of the trajectory were calculated using Gromacs tools.
The simulation was continued using the metadynamics approach for exploring the free energy landscape. The interface coordination number of atoms of protein ligand complex was used as collective variable (CV). The protein-ligand interface coordination numbers correlate with the numbers of atom contact, and larger coordination number usually indicates protein-ligand is in binding state.
The coordination number C is de ned as follows by Plumed: see formulas 1 and 2 in the supplementary les.

Quantitative reverse transcription polymerase chain reaction
This assay was carried out as described previously [52]. Viral RNAs were extracted from the samples using the QIAamp RNA Viral Kit (Qiagen, Heiden, Germany), and quantitative reverse transcription polymerase chain reaction (qRT-PCR) was performed using a commercial kit (Genrui-bio) targeting the S and N genes. The specimens were considered positive if the Ct value was ≤ 38.0, and negative if the results were undetermined. Specimens with a Ct higher than 38 were repeated. The specimen was considered positive if the repeat results were the same as the initial result and between 38 and 40. If the repeat Ct was undetectable, the specimen was considered negative.
Indirect immuno uorescence assay (IFA) IFA was carried out as previously reported [53,54]. Vero cells were xed in 4% formaldehyde at 48 hours post infection. Then cells were permeabilized in 0.5% Triton X-100, blocked in 5% BSA in PBS, and then probed with the plasma of this patient or healthy control at a dilution of 1 (v/v) Igepal CA-630 (Anatrace) was added and incubated at 4 °C for 10 min. Cells were lysed by sonication and the lysate was clari ed by ultracentrifugation. Cleared lysates were passed through a 0.22-μm lter lm before further puri cation. The protein was puri ed by tandem a nity chromatography and SEC.
Surface plasmon resonance (SPR) assay The a nities between nsp12 and drugs were measured at room temperature (r.t.) using a Biacore 8K system with CM5 chips (GE Healthcare). The nsp12 protein was immobilized on the chip with a concentration of 100 μg/mL (diluted by 0.1mM NaAc, PH 4.0).
Drug samples were prepared according to procedure 29264621AA of GE Healthcare Life Sciences. 1×PBS solution plus 5% DMSO and 0.005% p20 was used for running and diluting drugs. A blank channel of the chip was used as the negative control. Serial diluted drugs were then owed through the chip surface.

Statistical analysis
Data are presented as the mean ± SD (Standard Deviation). All analyses were performed using GraphPad Prism version 7.0 for Windows (GraphPad Software, San Diego California, USA). Data were subjected to statistical analysis by two-way ANOVA or two-tailed Student's t-test. The P values less than 0.05 were considered statistically signi cant.

DATA AVAILABILITY
The structure for SARS-CoV-2 RdRp has been deposited in the Protein Data Bank with accession number 6M71, 6WTT. The template structure we used to construct RdRP-ligand complexes are deposited in the Protein Data Bank with accession number 3BR9).
Declarations Figure 1 Drug repurposing against RdRp for COVID-19 using a hybrid deep learning and molecular simulation strategy. a, 1906 approved drugs were subject to the proposed screening process which consists of molecular vector-based screening, structure-based screening and force eld-based screening. DFCNN and DeepBindBC are both deep learning-based methods. 4 candidate drugs were selected by the proposed method, including Pralatrexate, Azithromycin, Sofosbuvir, Amoxicillin. b, Key interactions between the studied drugs and RdRp from the last frame of MD simulation, for (i) Azithromycin and (ii) Pralatrexate. RdRp binding pocket is shown in green with surface representation and the corresponding drugs are shown in magenta. The 2D Schematic diagram of drug-RdRp interaction is given bottom, and neighbor residues (within 4 Å of the drug) are shown. b(iii), the experimental structure of Remdesivir in its monophosphate form with RdRp (PDB ID 7BV2), the 2D Schematic diagram of the interaction was also shown. hour before viral attachment, and at 2 hours post-infection, the virus-drug mixture was replaced with fresh