Validation of our network-based knowledge mining results

To demonstrate that our computational pipeline for drug repositioning can yield reasonably accurate prediction results, we also validated our network-based knowledge mining algorithm (Fig. ​(Fig.1b)1b) using cross-validation and a retrospective study on the past data of the two coronaviruses that are closely related to SARS-CoV-2 and had been relatively well studied in the literature, i.e., SARS-CoV and MERS-CoV. We first performed ten-fold cross-validation on all virus-related drug–target interactions (DTIs) from our collected data to evaluate the prediction performance of our network-based knowledge mining algorithm CoV-DTI. To mimic a realistic scenario in which the goal is to predict the drug candidates for novel virus targets, we randomly split the virus targets into ten folds and masked the labels of one fold for testing. All the negative samples with unknown DTI labels were used during training when computing the Bayesian personalized ranking loss (Methods). The receiver operation characteristics (ROC) curves and the curves of recall scores with respect to a list of top k drug candidates predicted by our model CoV-DTI and DTINet,⁸ our previously developed network-based DTI prediction algorithm which served as a state-of-the-art baseline method, were obtained by averaging the predicted scores from five repeats of ten-fold cross-validation (Supplementary Fig. S1). Our cross-validation results showed that CoV-DTI achieved an area under the ROC (AUROC) score of 0.8273, which was 8% higher than that of DTINet. In addition, over 50% of the drugs known to act on a virus target can be accurately predicted by CoV-DTI from the top 200 drug candidates, with a better performance compared to DTINet. All these cross-validation results demonstrated that our computational method CoV-DTI can make reasonably accurate predictions on the reconstructed virus-related knowledge graph.

In our retrospective study using the past data of SARS-CoV and MERS-CoV, with the aid of our developed text-mining tool BERE, we found that many of the drugs that had been reported previously in the literature to have antiviral activities against the corresponding coronavirus, were also among the top list of our predicted results (Table ​(Table1).1). For example, chloroquine, an FDA-approved drug for treating malaria,⁹ which was previously reported to exhibit micromolar anti-SARS-CoV activity in vitro,¹⁰ was also repurposed to target the same virus by CoV-DTI. Gemcitabine, which was originally approved for treating certain types of cancers,¹¹ was also predicted by CoV-DTI to target SARS-CoV and can be further validated by previous in vitro studies.¹² Cyclosporine, a calcineurin inhibitor approved as an immunomodulatory drug,¹³ which was previously observed to block the replication of SARS-CoV,¹⁴ can also be successfully predicted by our approach. Among the top list of predicted drugs against MERS-CoV, miltefosine, which was approved for treating leishmaniasis,¹⁵ had been previously identified to have anti-MERS-CoV activities.¹⁶ Chlorpromazine and imatinib, which were used for treating schizophrenia¹⁷ and leukemia,¹⁸ respectively, were also selected by CoV-DTI as anti-MERS-CoV drugs and can be validated by previous in vitro evidences.¹² Thus, the above retrospective study illustrated that our computational framework is able to predict effective drug candidates against a specific coronavirus.

PARP1 inhibitors as potential drug candidates for COVID-19

Through careful examination of the screening results derived mainly by CoV-DTI and BERE and the connectivity map analysis results derived using the transcriptomic profiles of SARS-CoV-2/SARS-CoV-infected patients, we found that poly-ADP-ribose polymerase 1 (PARP1) inhibitors (Supplementary Fig. S2), such as PJ-34 and olaparib, were highlighted as potential therapeutic agents that may have the antiviral activities against SARS-CoV-2 (Table ​(Table2,2, Supplementary Table S1 and Supplementary Table S2). We then focused on this drug class and selected those drug candidates with both potential efficacy and acceptable pharmacokinetic and toxicological profiles for further investigation. We first noticed that PJ-34 was the most preferred drug derived from our computational framework, as both our network-based knowledge mining algorithm CoV-DTI and connectivity map analysis pointed to this drug (Table ​(Table22 and Supplementary Table S1). After checking the clinical status of PJ-34, we found that it only reached the preclinical trial stage (DrugBank ID: DB08348,¹⁹) which may raise the concern about its safety issue and thus limit its timely application for the current clinical usage against COVID-19. Therefore, we compared the molecular characteristics between PJ-34 and other PARP1 inhibitors that had been already approved by the Food and Drug Administration (FDA) or at least entered the clinical trials (Methods), and found that rucaparib and CVL218 (mefuparib) share the highest similarities with PJ-34. Considering the poor accessibility of rucaparib in the lung tissue (https://www.ema.europa.eu/en/medicines/human/EPAR/rubraca), it may be hard to apply this small molecule in treating the pneumonia caused by SARS-CoV-2. Thus, we mainly selected CVL218 to conduct further experiments, and this drug was later validated to exhibit a much better tissue distribution with high concentration in lung (see Supplementary Materials for more details). We also included olaparib, the first FDA-approved PARP1 inhibitor, in our further analyses, as this drug also achieved a high rank in our connectivity map analysis using the transcriptomic data of COVID-19 patients (Table ​(Table22).

CVL218 exhibits in vitro inhibitory activity against SARS-CoV-2 replication

We first conducted a pilot experimental test in vitro (Methods) on the anti-SARS-CoV-2 activities of CVL218, olaparib and several other related drugs (Fig. ​(Fig.2a).2a). We found that both PARP1 inhibitors olaparib and CVL218 exhibited inhibitory effects against SARS-CoV-2 replication. Nevertheless, CVL218 showed a much higher inhibition rate than olaparib. More specifically, olaparib inhibited SARS-CoV-2 replication by 15.31% at a concentration of 3.2μM, while CVL218 reached 34.64% reduction at a concentration of 3μM.

Open in a separate window

Fig. 2

The in vitro anti-SARS-CoV-2 activities of the tested drugs in Vero E6 cells. a The preliminary in vitro antiviral activities of oseltamivir at 3μM, zanamivir at 3μM, baricitinib at 3.2μM, olaparib at 3.2μM, arbidol at 3μM and 30μM, and CVL218 at 3μM and 30μM, respectively, were detected in Vero E6 cells infected with SARS-CoV-2 at an MOI of 0.05. The viral yield in the cell supernatant was then quantified by qRT-PCR. Results are shown as mean±SD over four replicates. b The concentration-dependent inhibition curve of CVL218 against SARS-CoV-2 replication and its cytotoxicity results. Viral infection and drug treatment at different concentrations were performed as mentioned above. Cytotoxicity of CVL218 to Vero E6 cells was measured by the CCK8 assays. c Visualization of virus nucleoprotein (NP) expression of the infected cells upon treatment of CVL218 at 48h post the SARS-CoV-2 infection using fluorescence microscopy. d Time-of-addition results on the inhibition of CVL218 and remdesivir against SARS-CoV-2 in vitro. The viral inhibitory activities of CVL218 and remdesivir were measured at “full-time”, “entry”, and “post-entry” stages, respectively. Results are shown as mean±SD over four replicates. e Virus NP expression in the infected cells upon the treatment of CVL218 and remdesivir was analyzed by western blot. f In vitro inhibitory activities against SARS-CoV-2 replication of favipiravir (30μM), CVL218 (3.5μM) and a combination of both drugs (30μM favipiravir+3.5μM CVL218). The concentrations were selected according to the EC₂₅ values of individual drugs against SARS-CoV-2 in vitro. Viral infection and drug treatment were performed as mentioned above. Results are shown as mean±SD over three replicates, and the significances were measured by p-values from t-tests. * and **** stand for p-value<0.05 and p-value<0.0001, respectively

Notably, the antiviral efficacy of CVL218 even surpassed arbidol, which has been used in clinical studies for the treatment of COVID-19.²⁰ In particular, arbidol inhibited SARS-CoV-2 replication by 21.76% at 3μM, much lower than that of CVL218 at the same concentration (Fig. ​(Fig.2a).2a). In contrast, oseltamivir, zanamivir (drugs used for preventing influenza virus infection), and baricitinib (JAK1/2 inhibitor, which was recommended in ref. ²¹ to treat COVID-19) showed no inhibitory activities against SARS-CoV-2 at the concentration of 3μM or 3.2μM.

Based on the above preliminary results in the pilot experimental test, we then chose CVL218 for subsequent experimental studies. Our further in vitro assays (Methods) showed that CVL218 exhibited effective inhibitory activity against SARS-CoV-2 replication in a dose-dependent manner, with an EC₅₀ of 5.194μM (Fig. ​(Fig.2b).2b). We also assessed the cytotoxicity of CVL218 by the CCK8 assay (Methods), and found that CVL218 had a CC₅₀ of 90.64μM in Vero E6 cells. Furthermore, immunofluorescence microscopy (Methods) revealed that, at 14h post SARS-CoV-2 infection, virus nucleoprotein (NP) expression in the CVL218-treated cells demonstrated a dose-response relationship with the treated drug concentrations, and was significantly lower upon CVL218 treatment compared to that in the DSMO treated cells (Fig. ​(Fig.2c).2c). Lower expression level of N gene compared to DMSO was also observed through RT-PCR experiments (Supplementary Fig. S3). In addition, no obvious cytopathic effect was observed in the infected cells treated with CVL218 at 25μM (Fig. ​(Fig.2c2c).

To systematically assess the inhibitory activities of CVL218 against SARS-CoV-2, we also performed a time-of-addition assay (Methods) to determine at which stage CVL218 inhibits viral infection. Remdesivir, which has entered the clinical trials for the treatment of COVID-19 (https://clinicaltrials.gov/ct2/show/NCT04257656), was also tested in this assay for comparison. In particular, compared to the DMSO control group, both CVL218 and remdesivir showed potent antiviral activities during the full-time procedure of SARS-CoV-2 infection in Vero E6 cells (Fig. ​(Fig.2d).2d). In addition, this time-of-addition assay indicated that CVL218 can partially work against the viral entry and significantly inhibit the replication of virus post-entry, while the remdesivir can only function at the post-entry stage (Fig. 2d, e). All together, results of the above in vitro assays indicated that CVL218 can be considered a potential therapeutic agent for treating COVID-19.

We further examined the drug combination effect of CVL218 with another anti-SARS-CoV-2 drug candidate favipiravir, which was originally designed to target the RNA-dependent RNA polymerase (RdRp) of influenza A virus,²² and had been recently reported to exhibit in vitro anti-SARS-CoV-2 activity with an EC₅₀ of 61.88μM²³ and a higher 7-day recovery rate than arbidol in clinical trials for treating COVID-19.²⁴ In particular, the inhibitory activities against SARS-CoV-2 replication of favipiravir at 30μM, CVL218 at 3.5μM and a combination of both were measured in Vero E6 cells (Methods). We found that, while favipiravir exhibited no inhibitory activity and CVL218 showed an inhibition rate of 36% at the experimental concentrations, the combination of both CVL218 and favipiravir achieved a significantly higher inhibition rate (64%, Fig. ​Fig.2f).2f). This result indicated that CVL218 can also be potentially combined with other drug candidates to enhance the therapeutic efficacy against SARS-CoV-2.

CVL218 interacts with the nucleocapsid protein of SARS-CoV-2

To further investigate the potential target of SARS-CoV-2 that CVL218 acts on, we next performed an in vitro surface plasmon resonance (SPR) assay to measure its binding affinity with the nucleocapsid (N) protein of SARS-CoV-2 (SARS-CoV-2-N) (Fig. ​(Fig.3,3, Methods). In this SPR assay, we also tested another two PARP1 inhibitors, including PJ-34 and olaparib, and two related anti-SARS-CoV-2 drugs, including remdesivir and arbidol, for their possible binding to SARS-CoV-2-N. As expected, no binding was observed for remdesivir and olaparib (Fig. 3d, e). This result was consistent with the known facts that remdesivir mainly performs its antiviral activity by targeting the RNA-dependent RNA polymerase (RdRp) of coronavirus²⁵ and arbidol was originally designed to inhibit the virus-host cell fusion against influenza virus.^26,27 Among all the three tested PARP1 inhibitors, including CVL218, PJ-34 and olaparib, CVL218 exhibited the highest binding strength towards SARS-CoV-2-N with a dissociation constant (K_D) of 3.1μM (Fig. ​(Fig.3a).3a). On the other hand, PJ-34 displayed a much lower binding affinity to SARS-CoV-2-N (K_D=68.1μM), which was about 22-fold lower than that of CVL218 (Fig. ​(Fig.3b).3b). Surprisingly, we observed no binding signal with SARS-CoV-2-N for the PARP1 inhibitor olaparib (Fig. ​(Fig.3c).3c). All these binding results derived from our SPR assay demonstrated that CVL218 interacts with SARS-CoV-2-N with a higher binding affinity than the other two tested PARP1 inhibitors, implying a potential antiviral mechanism mediated by the interplay between CVL218 and the nucleocapsid protein of SARS-CoV-2.

Open in a separate window

Fig. 3

CVL218 binds to the nucleocapsid protein of SARS-CoV-2 (SARS-CoV-2-N). a–e Kinetic analyses of the binding between SARS-CoV-2-N and the tested drugs, including CVL218 (a), PJ-34 (b), olaparib (c), remdesivir (d), and arbidol (e), measured by a SPR-based Biacore instrument. The derived dissociation constants between SARS-CoV-2-N and the tested drugs (CVL218 and PJ-34) are also shown

The network-based knowledge mining algorithm

The initial list of drug candidates targeting SARS-CoV-2 was first screened using a network-based knowledge mining algorithm, called CoV-DTI, which was modified from our previous work.^8,91 The goal of CoV-DTI is to capture the hidden virus-related feature information and accurately predict the potential drug candidates from the constructed knowledge graph, which is realized through learning a network topology-preserving embedding for each node.

More specifically, CoV-DTI uses a graph convolution algorithm⁹² to gather and update feature information for each node in the constructed heterogeneous knowledge graph network from neighborhoods so that the network topology information can be fully exploited. Suppose that we perform T iterations of graph convolution. At iteration 1≤t≤T, the message $m_v^t$ passed to node v can be expressed as:

where a_v,u,r stands for the weight for edge e=(u, v, r), $A_{u,v,r}=\frac{a_{v,u,r}}{{\sum }_u{a}_{v,u,r}}$, ${\mathbf{W}}_r^t\in {\mathbb{R}}^{d\times d}$, and ${\mathbf{b}}_r^t\in {\mathbb{R}}^d$ stand for the learnable parameters, $ReLU\left(x\right)=\max \left(0,x\right)$, and $N_r\left(v\right)=\left\{u,u\in V,u\ne v,\left(u,v,r\right)\in E\right\}$ denotes the set of adjacent nodes connected to vV through edges of type rR.

Then the feature $h_v^t$ of node v is updated by

where ${\mathbf{W}}^t\in {\mathbb{R}}^{d\times d}$ and ${\mathbf{b}}^t\in {\mathbb{R}}^d$ stand for the learnable parameters, and concat (·,·) stands for the concatenation operation.

Finally, the confidence score s_u,v of the relation r between node u and node v is derived from the learned node embeddings and the corresponding projection matrices, that is,

where $G_r,H_r\in {\mathbb{R}}^{d\times k}$ stand for the edge-type-specific projection matrices.

CoV-DTI minimizes the Bayesian personalized ranking (BPR) loss⁹³ for drug–target interaction reconstruction, by regarding those edges not in the edge set E as missing values rather than negative samples, that is,

where, s_{u, v} and s_{w, x} stand for the confidence scores of the relation r between u and v and between w and x, respectively, and σ(·) stands for the sigmoid activation function. Intuitively, in the above loss function, the confidence scores of the node pairs (u, v) in the edge set (i.e., (u, v, r) E) are encouraged to be higher than those of unseen pairs (w, x) (i.e., (w, x, r) E).

We predicted the confidence scores under the relation of virus target–drug interactions for each virus target–drug pair using Eq. (3). Then the confidence scores were averaged across all the proteins of a certain virus (e.g., SARS-CoV, MERS-CoV, or SARS-CoV-2), and the corresponding p-values were obtained by z-test. For each virus, we selected those predictions with a p-value<0.05 as drug candidates.

Automated relation extraction from large-scale literature texts

We used a deep learning based relation extraction method named BERE⁴ to extract the coronavirus-related drugs from large-scale literature texts. More specifically, the sentences mentioning the two entities of interest, i.e., name (or alias) of a coronavirus or coronavirus target, or name (or alias) of a drug, are first collected using a dictionary-based name entity recognition method (string matching). For each pair of entities (e₁, e₂), there are usually more than one sentence describing the underlying relations. Therefore, we use a bag of sentences $S_{e_1,e_2}$, denoting the set of all the sentences mentioning both e₁ and e₂, to predict the relation between these two entities.

We first encode each sentence $s\in S_{e_1,e_2}$ in a semantic and syntactic manner using a hybrid deep neural network ($h:s\to {\mathbb{R}}^d$), including a self-attention module,⁹⁴ a bi-directional gated recurrent unit (GRU) module⁹⁵ and a Gumbel tree-GRU module.^4,96 Each sentence representation h(s) is then scored by a sentence-level attention module to indicate its contribution to the relation prediction, that is,

where $\beta \left(s\right)\in \mathbb{R}$ stands for the weight score, and $W_s\in {\mathbb{R}}^{d\times 1}$ stands for the learnable weight parameters. Finally, the relation is predicted by a binary classifier, based on the weighted sum of sentence representations, that is,

where $r_{e_1,e_2}$ stands for the probability of the relation of interest between entities e₁ and e₂ mentioned by the bag of sentences $S_{e_1,e_2}$.

The training corpus we used was curated automatically from nearly 20 million PubMed (http://www.pubmed.gov) abstracts by a distant supervision technique.⁹⁷ In detail, the names (or aliases) of drugs or targets in sentences were first annotated by a dictionary-based named entity recognition method (string matching), in which the name dictionary was derived from DrugBank (version 5.1.0),¹⁹ with ambiguous names (e.g., common words) removed. Next, the label for each bag of sentences co-mentioning a drug–target pair of interest was annotated automatically by the known drug–target interactions in DrugBank. The unlabeled corpus that we used in this work for text mining the coronavirus-related drugs was obtained from approximately 2.2 million PMC full-text articles, with entities of interest annotated using the aforementioned named entity recognition approach. A coronavirus-related drug was extracted as a hit candidate if the model found a bag of sentences describing a relation between this drug and a target in the coronavirus of interest.

Connectivity map analysis

We used the transcriptome analysis approach to further filter the potential drug candidates for treating the COVID-19 patients infected by SARS-CoV-2. We collected the gene expression profiles of the samples from SARS-CoV-2 infected patients to screen the drug candidates against COVID-19. We also used those from SARS-CoV infected patients to filter out the potential therapeutic agents targeting at COVID-19. Such a strategy is reasonable as SARS-CoV and SARS-CoV-2 are two closely related and highly similar coronavirus. First, the genome of SARS-CoV-2 is phylogenetically close to that of SARS-CoV, with about 79% of sequence identity,⁹⁸ and the M (membrane), N (nucleocapsid) and E (envelope) proteins of these two coronaviruses have over 90% sequence similarities.⁹⁹ In addition, the pathogenic mechanisms of SARS-CoV-2 and SARS-CoV are highly similar.¹⁰⁰

In particular, we collected the gene expression profiles of the peripheral blood mononuclear cells (PBMCs) and the bronchoalveolar lavage fluid (BALF) samples from three SARS-CoV-2 infected patients (NGDC: PRJCA002326)⁶ and PBMC samples from ten SARS-CoV infected patients (GEO: GSE1739).⁷ For samples from SARS-CoV infected patients, the raw gene expression values were first converted into logarithm scale, and then the differential expression values (z-scores) were computed by comparing to those of healthy persons using the same protocol as described in ref. ⁵ that is,

where Z_infected stands for the z-scores of the SARS-CoV infected patients, X_infected and X_healthy stand for the gene expression values in logarithm scale of the infected and healthy persons, respectively, median(·) stands for the median operation, MAD(·) stands for the median absolute deviation operation, and C=1.4826 is a constant for normalization. The p-values for all the genes with measured expression values during the analysis were also computed based on the z-scores. For samples from those SARS-CoV-2 infected patients, due to the limited sample sizes (i.e., three samples for the healthy, three PBMC and two BALF samples for the infected patients), using z-score for differentially expressed gene analysis can be inaccurate and misleading. This is because the MAD mentioned above can be of high variance when the number of healthy samples is low (i.e., three), resulting in inaccurate estimation of the z-score. Here we mainly used the fold-change to analyze the differential expressed genes, which was obtained from the original paper.⁶ The up- and downregulated genes were determined by p-value <0.01 and sorted according to the z-score/log2-fold-change values. We used the connectivity map (CMap),⁵ which contains the cellular gene expression profiles under the perturbation of 2428 well annotated reference compounds, to measure the associations of gene expression patterns between SARS-CoV-2/SARS-CoV infected patients and the reference compound-perturbed cells. The connectivity map scores were computed based on the up- and downregulated gene sets of SARS-CoV-2/SARS-CoV infected patients using the web tool (https://clue.io/query). Under the hypothesis that the gene expression pattern resulting from the perturbation by a therapeutic compound should be negatively correlated with that resulting from the coronavirus infection, we selected those compounds that have significant negative connectivity map scores, that is, the list of drug candidates predicted to treat the coronavirus infected patients was obtained by selecting the compounds with the connectivity map scores <−90, which was suggested by the original paper.⁵

Comparing molecular characteristics between PARP1 inhibitors

We compared the molecular characteristics between PJ-34 and other available PARP1 inhibitors that are FDA-approved or in clinical trials, including niraparib, rucaparib, veliparib, nicotinamide, olaparib, iniparib, theophylline, talazoparib, and CVL218. In particular, near 200 molecular characteristics were calculated using RDKit,⁸⁸ by enumerating all the property descriptor functions in rdkit.Chem.Descriptors.descList. These functions calculate the scalar features such as molecular weight, logP and number of heteroatoms. Those functions with the returned invalid values were discarded. As the output molecular features of different functions were not always in comparable scales, each dimension of the molecular characteristics of every PARP1 inhibitor was normalized into range [0,1]. Finally, the Euclidian distance between the normalized feature vectors was used to measure the similarities between two PARP1 inhibitors.

Cells and virus

The African green monkey kidney Vero E6 cell line was purchased from the Cell Resources Center of Shanghai Institute of Life Science, Chinese Academy of Sciences (Shanghai, China) and cultured in DMEM medium (Gibco Invitrogen, no. 12430-054) containing 10% fetal bovine serum (FBS; Gibco Invitrogen) at 37°C with 5% CO₂ atmosphere. BetaCoV/JS03/human/2020 (EPI_ISL_411953), a SARS-CoV-2 virus strain, was isolated from nasopharyngeal swab of a 40-year old female confirmed as COVID-19 case by reverse transcriptase polymerase chain reaction (RT-PCR) in December 2019. The virus was propagated in Vero E6 cells, and the viral titer was determined by the 50% tissue culture infective dose (TCID50) based on microscopic observation of cytopathic effects. All the in vitro SARS-CoV-2 infection experiments were performed in a biosafety level-3 (BLS-3) laboratory in Jiangsu Provincial Center for Diseases Control and Prevention, Jiangsu, China.

Antiviral drugs

Potential antiviral drugs, including zanamivir, oseltamivir, remdesivir, baricitinib, olaparib, and arbidol, were all provided by MCE (Medchem Express, China). The PARP1 inhibitor mefuparib (CVL218) with a purity of more than 99.0% was provided by Convalife, Shanghai, China.

Cytotoxicity test and virus infection assay

The cytotoxicity of the tested drugs on Vero E6 cells was determined by the CCK8 assays (Beyotime, China). At 48h post addition of the tested drugs, 20μL CCK8 was added to each well and incubated at 37°C for 1h. Then optical density was measured at 450nm. The 50% cytotoxic concentration (CC₅₀) values were calculated using GraphPad Prism 5 (GraphPad Software, USA). Vero E6 cells were seeded into 96-well plates with a density of 5×104 cells/well for incubation in DMEM medium supplemented with 10% FBS for 16h in an incubator with 5% CO₂ at 37°C, for cells to reach 80% confluent. Then, cell culture medium of each well was removed, and PBS was used to wash the cells once, before evaluating the antiviral efficacy of the tested drugs. Four duplicated wells were made for each dose of drugs, and the cells were pretreated with different doses of antiviral drugs diluted by the cell maintenance solution (50μL per well) for 1h. For the virus control and cell control wells, cell medium containing DMSO or only medium of 50μL per well was added. Next, pretreated or untreated cells in each well were infected with the virus with a multiplicity of infection (MOI) of 0.05 for 2h. After that, the virus-drug mixture was removed and cells were further cultured with fresh drug-containing medium at 37°C with 5% CO₂ atmosphere for 48h. Then culture supernatant per well was harvested and inactivated at 56°C for 30min to further extract and quantify viral RNA. The preliminary in vitro antiviral activities of the tested drugs were first screened at individual concentrations with oseltamivir at 3μM, zanamivir at 3μM, baricitinib at 3.2μM, olaparib at 3.2μM, arbidol at at 3μM and 30μM, and CVL218 at 3μM and 30μM, respectively. Then, the concentration-dependent inhibition activity of CVL218 against SARS-CoV-2 replication was performed with 5-fold serial dilutions of the maximum concentration at 50μM.

Viral RNA extraction and quantitative real-time PCR (qRT-PCR)

Viral RNA was extracted from culture supernatant using the HP RNA Isolation Kit (Roche) according to the manufacturer’s instructions. RNA was eluted in 30μL RNase-free water. Reverse transcription was performed with a SARS-CoV-2 nucleic acid detection kit (BioGerm, China) according to the manufacturer’s instructions. The PCR reaction system was configured as follows: 6μL of qRT-PCR reaction solution, 2μL of qRT-PCR enzyme mixture, 2μL of primer probe and 2.5μL of template, and the reaction was performed as follows: 50°C for 10min, 95°C for 5min, followed by 40 cycles of 95°C for 10s, 55°C for 40s. The values of 2^−ΔCT were calculated according to the CT value measured from the PCR instrument, to represent the relative virus copies of the drug group to the control group. The virus replication inhibition rate (%) was calculated as (1–2^−ΔCT) × 100%. The dose-response curves were plotted according to viral RNA copies and the drug concentrations using GraphPad Prism 5 (GraphPad Software, USA).

For quantification of N gene, total RNA was also extracted from the infected cells using the Rneasy mini kit (Qiagen) according to the manufacturer’s instructions. The viral N gene was quantified by qRT-PCR by using a SARS-CoV-2 nucleic acid detection kit (bioPerfectus technologies, China) according to the manufacturer’s instructions.

Time-of-addition assay

To facilitate the observation of the antiviral effects of drugs against SARS-CoV-2 at different timing, relative high doses of the tested drugs (CVL218 at 20μM and remdesivir at 10μM) were used for the time-of-addition assay. Vero E6 cells with a density of 5×10⁴ cells per well were treated with the tested drugs, or DMSO as controls at different stages of virus infection. The cells were infected with virus at an MOI of 0.05. The “Full-time” treatment was to evaluate the maximum antiviral effects, with the tested drugs in the cell culture medium during the whole experiment process, which was consistent with the descriptions in the virus infection assay. For the “Entry” treatment, the tested drug was added to the cells for 1h before virus infection, and then cells were maintained in the drug-virus mixture for 2h during the virus infection process. After that, the culture medium containing both virus and the tested drug was replaced with fresh culture medium till the end of the experiment. For the “Post-entry” experiment, virus was first added to the cells to allow infection for 2h before the virus-containing supernatant was replaced with drug-containing medium until the end of the experiment. At 14h post infection, the viral inhibition in the cell supernatants of the tested drug was quantified by qRT-PCR, and calculated using the DMSO group as reference.

Indirect immunofluorescence assay

Vero E6 cells were treated with CVL218 at 5μM, 15μM, and 25μM, respectively, following the same procedure of “full-time” treatment. After 48h of culture, infected cells were fixed with 80% acetone in PBS and permeabilized with 0.5% Triton X-100, and then blocked with 5% BSA in PBS buffer containing 0.05% Tween 20 at room temperature for 30min. The cells were further incubated with a rabbit polyclonal antibody against a SARS-CoV nucleocapsid protein (Cambridgebio, USA) as primary antibody at a dilution of 1:200 for 2h, followed by incubation with the secondary Alexa 488-labeled goat anti-rabbit antibody (Beyotime, China) at a dilution of 1:500. Nuclei were stained with DAPI (Beyotime, China). Immunofluorescence was observed using fluorescence microscopy.

Western blot assay

NP expression in infected cells was analyzed by western blot. Protein samples were separated by SDS-PAGE and then transferred onto polyvinylidene difluoride membranes (Millipore, USA), before being blocked with 6% Rapid Block Buff II (Sangon Biotech, China) at room temperature for 10min. The blot was probed with the antibody against the viral nucleocapsid protein (Cambridgebio, USA) and the horseradish peroxidase-conjugated Goat Anti-Rabbit IgG (Abcam, USA) as the primary and the secondary antibodies, respectively. Protein bands were detected by chemiluminescence using an ECL kit (Sangon Biotech, China).

Inhibitor combination assay

To assess the potential synergistic effect, CVL218 at 3.5μM and favipiravir at 30μM were mixed, while CVL218 alone and favipiravir alone were included as controls. The concentrations were selected according to the EC₂₅ values of individual drugs against SARS-CoV-2 in vitro. The mixtures were tested for their inhibitory activities on the SARS-CoV-2 with a multiplicity of infection (MOI) of 0.05 as described above. Each sample was tested in triplicate.

SPR binding assay

Surface plasmon resonance experiments were performed with a BIAcore S200 (GE Healthcare) as previously described.¹⁰¹ The running buffer contained 1× PBS, 2% DMSO. The purified SARS-CoV-2-N protein was desalted using Zeba^TM spin desalting column (Thermo Scientific), diluted in 10mM sodium acetate (pH 4.0) to 20μg/mL, and immobilized on a CM5 sensor chip by amine coupling. The tested drugs in 2-fold serial dilutions were made in the running buffer. All measurements were performed at a flow rate of 30μL/min. Data processing and analyses were performed using BIAevaluation 1.1 software.

LPS-induced cytokine production in PBMCs

Peripheral blood mononuclear cells (Allcells) were cultured at 37°C at concentration 5% CO₂ atmospheric on a 96-well plate in RPMI1640 cell growth medium (Gibico, Cat. 11875-093). For stimulation, PBMC cells were incubated with 1μg/mL LPS (Sigma, Cat. L2880-25MG). To test whether CVL218 and olaparib can inhibit the production of IL-6, IL-10, IFN-γ, TNF-α, 1μM and 3μM concentrations of CVL218 or olaparib were added to cell culture medium for 6 and 24h, respectively. The concentration of each cytokine was determined by ELISA using a commercial kit (BioLegend, USA).

Pharmacokinetics and toxicity study

Sprague-Dawley rats were purchased from Shanghai Laboratory Animal Center, China. The animals were grouped and housed in wire cages with no more than six per cage, under good laboratory conditions (temperature 25±2°C; relative humidity 50±20%) and with dark and light cycle (12h/12h). Only healthy animals were used for experimental purpose. The pharmacokinetics and biodistribution study in Sprague-Dawley rats was approved by Center for Drug Safety Evaluation and Research, Shanghai Institute of Materia Medica, Chinese Academy of Sciences. A total of 144 Sprague-Dawley rats with each sex were used for toxicity study. Animals were randomly separated into four groups (18/sex/group). CVL218 was administered at doses of 20, 40, 60, and 160mg/kg. For all the groups, 20 rats (10/sex/group) were randomly selected and euthanized at day 28, and their sections of various tissues and organs were obtained and frozen. Ten (5/sex/group) animals were euthanized after a 28-day drug free period, and their sections of tissues and organs were obtained and frozen. Six (3/sex/group) were euthanized after the blood samples were obtained. For pharmacokinetic and toxicity evaluation, clinical symptoms, mortality and the animals’ body weight were examined. Serum (0.5mL) was collected to analyze toxicokinetics at different timepoints post drug administration. The plasma concentration-time data were analyzed using a noncompartmental method (Phoenix, version 1.3, USA) to derive the pharmacokinetic parameters.

Biodistribution study

Thirty Sprague-Dawley rats were randomly divided into three time point groups (3/sex/group). At 3, 6, and 8h after CVL218 administration, animals were sacrificed, and the brain, heart, lung, liver, spleen, stomach, and kidney tissues were collected. Tissue samples were washed in ice-cold saline, blotted with paper towel to remove excess fluid, and weighed. Tissue samples were fluid, weighted and stored at −20±2°C until the determination of drug concentration by LC-MS-MS.

Toxicity study in cynomolgus monkeys

Healthy male and female cynomolgus monkeys aged 3–4 years were purchased from Guangdong Landau Biotechnology, China. The animals were maintained in accordance with the Guide for the Care and Use of Laboratory Animals.

Cynomolgus monkey (5/sex/group) were selected using a computerized randomization procedure, and administered CVL218 by nasogastric feeding at dose levels of 0 (control), 5, 20, 80mg/kg. Individual dose volumes were adjusted weekly based on body weight of monkeys. The monkeys were observed twice daily for viability/mortality and for any change in behavior, reaction to treatment or ill-health. Electrocardiograms, intraocular pressure, rectal temperature and body weight were recorded. For all the groups, 2/3 of the animals were randomly selected and euthanized at day 28. The remaining animals were euthanized after a 28-day drug free period. Blood samples were taken before and at 0.5, 1, 2, 4, 8, and 24h post-dose on days 1 and 28 of the treatment period. Pharmacokinetic evaluation was performed using a noncompartmental method (Phoenix, version 1.3, USA) and pharmacokinetic parameters were calculated for individual monkeys.

This work was supported in part by the National Natural Science Foundation of China (61872216, 81630103, 31900862, 31725014), Jiangsu Provincial Emergency Project on Prevention and Control of COVID-19 Epidemic (BE2020601), the Nation Science and Technology Major Projects for Major New Drugs Innovation and Development (2018ZX09711003-004-002, 2019ZX09301010), Pudong New Area Science and Technology Development Foundation (PKX2019-S08), and the Turing AI Institute of Nanjing, and the Zhongguancun Haihua Institute for Frontier Information Technology. We thank Dr. Yu Zhou for providing gene expression profile data of COVID-19 patients. We thank Dr. Feixiong Cheng for email communications on the connectivity map analysis.