In this study, five types of comprehensive human protein–protein interactions (PPIs) including a yeast two-hybrid system, protein three-dimensional interactome, literature-derived kinase–substrate interactions, literature-derived signaling networks, and affinity-purification mass spectrometry (AP-MS) were integrated from 15 widely used databases. The elaborate descriptions of constructing the human interactome can be found in our recent studies (Cheng et al., 2019; Fang et al., 2020a).

To build a comprehensive drug–target network for existing drugs, high-quality physical DTIs were integrated from six authoritative databases, including DrugBank database (v4.3; Wishart et al., 2018), the PharmGKB database (Barbarino et al., 2018), Therapeutic Target Database (TTD, v4.3.02; Yan et al., 2018), ChEMBL (v20; Gaulton et al., 2017), BindingDB (Gilson et al., 2016), and IUPHAR/BPS Guide to PHARMACOLOGY (Harding et al., 2018). Reported binding affinity data comprising dissociation constant (K_d), inhibition constant/potency (K_i), median effective concentration (EC₅₀), or median inhibitory concentration (IC₅₀) equal to or less than 10 μM were applied to define a physical DTI. Only those physical DTIs conforming to the following criteria were retained in our study: (i) protein targets owning a unique UniProt accession number; (ii) protein targets labeled as “reviewed” in the UniProt database; and (iii) protein targets marked as Homo sapiens species. Consequently, we assembled 15,367 DTIs interacting 1,608 approved drugs with 2,251 unique human targets/proteins.

To construct a compound–protein network for natural products, we further integrated DTIs from several kinds of data sources. First, we collected direct DTI databases derived from ChEMBL (v21; Gaulton et al., 2017) and BindingDB (accessed in September 2017; Gilson et al., 2016), as well as indirect DTI databases obtained from STITCH5 (Szklarczyk et al., 2016), the Herbal Ingredients Target Database (Ye et al., 2011), and the Traditional Chinese Medicine Integrated Database (Huang et al., 2018). Second, we also integrated the experimentally validated DTIs from an in-house database that was collected from over 2,000 natural product-specific pharmacological references (Huang et al., 2019). After filtering all the DTIs according to the criteria mentioned in our previous research (Fang et al., 2017c), 38,220 DTIs connecting 3,882 natural products and 5,643 protein targets were finally preserved.

Given T, the set of host protein targets related with disease, and C, the set of drug targets for a compound, d_TC (the nearest distance) means the average nearest route distance of all the proteins in the host protein set (T) to drug targets for a compound (C) within the human protein interactome, as below (Guney et al., 2016):

where d(t,c) means the nearest route distance between host protein t and drug target c.

To evaluate the network length between T and C, we randomly extracted two sets of proteins from the human protein interactome that are same with the original protein sets in terms of size and degree distribution. After repeating this process for 1,000 times, the Z-score (z_d) was calculated as follows:

where d and d¯ are d(t,c) and the mean, and σ_d represents the standard deviation of the reference distribution.

p values were computed during the permutation test. For a given host protein set, compound–disease (COVID-19) pairs were regarded as significantly proximal if they satisfied the condition (Z < −2 and p < 0.05). We further defined that a drug or natural product was a promising candidate against COVID-19 if it was predicted as positive by at least two host protein sets of three.

The virus–host protein interaction networks of coronavirus (CoV) in our study were obtained from three latest public literature (Bojkova et al., 2020; Gordon et al., 2020; Zhou et al., 2020). Firstly, Zhou et al. integrated more than 100 high-quality host proteins which were validated with experiments for network-based drug reposition against SARS-CoV-2 (Zhou et al., 2020). Additionally, Gordon et al. (2020) systematically mapped the physical interactions between SARS-CoV-2 proteins and human protein targets utilizing AP-MS, and they identified 332 high-confidence SARS-CoV-2-human PPIs. Bojkova et al. (2020) revealed a proteomics profile of the cellular response to SARS-CoV-2 virus-infected human cells and identified 489 host proteins associated with SARS-CoV-2. Altogether, we assembled three host protein sets, resulting in 906 host proteins for further analysis (Supplementary Table 1).

Vero cells were purchased from the American Type Culture Collection (ATCC, CCL-81) and cultured in Dulbecco’s modified Eagle’s medium (DMEM, Gibco, Grand Island, NE, United States) supplemented with 10% fetal bovine serum (FBS, Gibco) in a humidified incubator of 5% CO₂ atmosphere at 37°C. Subsequently, cells were digested with trypsin (at the concentration of 0.25%) and 1 × 10⁴ cells in 100 μl of complete medium were uniformly seeded in each well of 96-well plates. After incubating for 24 h, Vero cells were infected with virus or treated with drug. The infectious virus (C-Tan-nCoV Wuhan strain 01) utilized in this study was a SARS-CoV-2 virus strain, which was isolated from clinical cases. All drugs were provided by the Chinese Academy of Medical Sciences and Peking Union Medical College.

In vitro antiviral efficiency of promising candidates against SARS-COV-2 virus on the Vero cell line was evaluated. Specifically, cells (at a density of 1 × 10⁴ cells/well) were plated into 96-well plates. After 24 h growth, Vero cells were pretreated with the drug candidates for 1 h. Subsequently, Vero cells were infected with the SARS-COV-2 virus (100 PFU/well) at a MOI of 0.01 for 2 h at 37°C. Following incubation, cells were maintained in medium with various concentrations of drugs for 48 h. Total viral RNA was quantified and analyzed utilizing the RT-PCR after extracting from the supernatant of the Vero cells.

To decipher whether these gene sets share positive correlations at a biological level, the correlation analysis was performed on the three sets of host genes utilizing Fisher’s exact test approach. As presented in Figures 2A,B, the host genes from Zhou et al. are significantly correlated with those from Gorden et al. since they share six overlapped genes (p = 0.013, Fisher’s exact test). Moreover, it looks apparently that there are high correlations between gene sets from Gorden et al. and Bojkova et al. as they have the largest number of common genes, with the p-value of 0.003. The total number of their overlapped genes has reached up to 17, including SRP54, GDF15, IMPDH2, DNMT1, TARS2, DCTPP1, RHOA, GGCX, DDX21, RAB5C, DDX10, PRIM2, CRTC3, PSMD8, ALG11, FBN1, and TOR1A. Also, there are 12 overlapped host genes between gene sets from Zhou et al. and Bojkova et al. (p = 2.843E-05, Fisher’s test). Taken together, these host gene sets do own significant positive correlations, indicating the rationality of the selected gene datasets.

To further elucidate the specific mechanism of actions underlying these host genes, we performed biological process and molecular function annotations using 119 high-quality experimental validated host genes from Zhou et al. As exhibited in Figure 2C, these host genes are involved in five groups of biological processes, such as nuclear import and interaction with the host. Interestingly, a recent study has demonstrated that SARS-CoV-2 efficiently blocks STAT1 and STAT2 nuclear import in order to impair transcriptional induction of IFN-stimulated genes (ISGs; Miorin et al., 2020). Additionally, SARS-CoV-2 is well-known to interact with host cells and employ the host cell receptor (ACE2) for cell entry (Hoffmann et al., 2020b; Walls et al., 2020). Likewise, according to the enrichment analysis of molecular function (Figure 2D), these host genes may participate in the regulation of virus receptor activity and single-stranded RNA binding.

We firstly developed two comprehensive drug–target (D–T) networks through integrating natural products and approved drugs that specifically targeted coronavirus host proteins. As displayed in Figure 3, the D–T network focusing on approved drugs comprises 490 DTIs connecting 331 drugs to 137 human host proteins, with the average degree of 1.48 on each drug. The host protein with the largest degree is SIGMAR1 (D = 83), followed by ABCC1 (D = 21), and SMAD3 (D = 13). Emerging evidence indicates that these targets play vital roles in COVID-19. For example, SMAD3 is the main transcription factor for TGF-β1 signals and acting on TGF-β-mediated Smad2/3 signaling may prevent or treat pulmonary fibrosis in COVID-19-infected cases (Chen et al., 2020). Moreover, as one of the top 10 host proteins in the DTI network, HDAC2 is a potential regulator of ACE2 in the human lung that serves as an important part for SARS-CoV-2 virus to bind with and enter host cells (Pinto et al., 2020). Meanwhile, 10 approved drugs targeting host proteins have degree (K) higher than 5: nintedanib (K = 11), sunitinib (K = 11), ruxolitinib (K = 8), bosutinib (K = 8), crizotinib (K = 7), flavitan (K = 6), dasatinib (K = 6), mitoxantrone (K = 6), hexachlorophene (K = 5), and glutathione (K = 5). Notably, recent evidence has demonstrated high potential (e.g., ruxolitinib and mitoxantrone) with COVID-19 therapies (Goker Bagca and Biray Avci, 2020; Lokhande et al., 2020). For instance, ruxolitinib has been reported to exert potential effectiveness on COVID-19 via inhibiting the JAK/STAT pathway (Goker Bagca and Biray Avci, 2020).

Similarly, Figure 4 comprises 2,002 interactions connecting 997 natural products with 259 human host proteins after removing ions and organic solvents. The average target degree (K) of a compound is 2.01, while the average compound degree (D) of a host protein is 7.73. Among the 997 natural products, the top 10 with the greatest target degree are berberine (K = 28), citric acid (K = 23), resveratrol (K = 23), epigallocatechin gallate (EGCG, K = 22), curcumin (K = 21), quercetin (K = 20), luteolin (K = 18), staurosporine (K = 13), andrographolide (K = 12), and wogonin (K = 11). Literature reveals that these natural products show the promising efficacy toward COVID-19. Taking resveratrol as an example, it inhibits the replication of SARS-CoV-2 virus in cultured Vero cells with an EC₅₀ value of 4.48 μM (Yang M. et al., 2020). Thus, resveratrol could be considered as a potential therapeutic during the treatment against SARS-CoV-2. Besides, network analysis suggests that curcumin interacts with 21 host proteins and recent study demonstrates that curcumin may be assessed to inhibit cellular entry and replication of SARS-CoV-2 virus as well as to benefit for the damage of pneumocytes that are associated with COVID-19 (Soni et al., 2020), indicating its potential treatment option for patients with coronavirus disease.

Interestingly, we found that the average degree (K = 2.01) of natural products in the compound-CoV host protein network is slightly higher than that of approved drugs (K = 1.48) in the drug–target network (P = 1.97 × 10^–5, one-sided Wilcoxon test), suggesting the “multi-target” effect of natural products. The detailed information of these two DTI networks is found in Supplementary Table 2. As the COVID-19 natural products and approved drugs could act on multiple host proteins according to the network analysis, thus, it is likely that most of the drug candidates may own the therapeutic potential toward COVID-19. Taken together, drug–target, or compound–protein network analyses contribute to the network-based identification of drug candidates and promote to further discover the potential therapeutic agents for COVID-19 through investigating both approved drugs and natural products.

The notion of the proposed network-based methodology posits that systematic identification and characterization of the virus–host interactome among the comprehensive human interactome network could facilitate effective therapeutic development for COVID-19. Here, we applied the advanced network proximity approach to measure the relevance between compounds and COVID-19 through assembling DTI networks, HCoV host proteins, and human PPI network. In total, 229 approved drugs were computationally identified to be associated with COVID-19 according to the criteria mentioned above. After searching for the previous literature, we found that 13 out of 229 approved drugs (Supplementary Table 3) were confirmed to possess anti-SARS-CoV-2 effects while 52 out of 229 approved drugs were reported to exert antiviral properties. Since cytokine storm is considered as one of the major pathological characteristics in patients infected with SARS-CoV-2 virus (Patra et al., 2020; Stebbing et al., 2020; Tian et al., 2020), we further sought evidence of anti-inflammation for these drug candidates and found that 113 out of 229 approved drugs were verified against inflammation. Taken together, there are 139 out of 229 (60.7%) approved drugs (Supplementary Table 3) that have been confirmed to possess anti-SARS-CoV-2, antiviral, or anti-inflammatory effects.

Among the 139 approved drugs, we selected 53 drug candidates, including 3 with experimental evidence of anti-SARS-CoV-2 effects and 50 that were satisfied with the following (Figure 5) factors: (i) confirmed antiviral effects; (ii) owning anti-inflammatory properties; and (iii) strength of the network-based prediction. As presented in Figure 5, 53 approved drugs can be grouped into 9 categories according to the first level of the Anatomical Therapeutic Chemical Classification (ATC) code. There are seven drugs for anti-infectives for systemic use, occupying the largest proportion (13.2%) among all the approved drug classes except “Others.” Interestingly, we found that several well-known anti-SARS-CoV-2 drugs (e.g., ribavirin and atazanavir) which belong to anti-infectives for systemic use were predicted to have significant associations on COVID-19 via our approach. For instance, atazanavir (Z-score = −2.815, by Bojkova et al. host protein set) was identified as a promising candidate, which has been reported to inhibit the replication of SARS-CoV-2 virus, alone or together with ritonavir in both Vero cells and a human pulmonary epithelial cell line (Fintelman-Rodrigues et al., 2020). We next systematically retrieved available evidence of antiviral effects in vivo for the 50 drug candidates to narrow up the scope of potential approved drugs (Figure 5). Among them, 19 network-predicted drugs were reported to exert antiviral effects in vivo. Additionally, potential candidates will also be kept if it was simultaneously predicted to be positive by all the three host gene sets. Altogether, 20 drug candidates were preserved for further validation (Figure 5).

Meanwhile, 432 natural products were predicted to be associated with COVID-19.76 out of them (17.6%) were reported to show anti-antiviral effects, while 156 of them (36.1%) were confirmed to own anti-inflammatory properties (Supplementary Table 4), indicating a high hit rate of our network proximity approach. Likewise, 74 potential natural products were selected (Figure 6) according to the same subject matter expertise above. Scaffold analysis reveals that these natural products can be divided into five clusters. Among them, cluster 5 has the largest number of natural products (n = 29), followed by cluster 3 (n = 22), and cluster 4 (n = 18). Indeed, a recent study has demonstrated that fluoxetine (cluster 4, Z-score = −3.153, by Bojkova et al. host protein set), a well-known antidepressant, could efficiently inhibit the entry and propagation of the SARS-CoV-2 virus in both Calu-3 and Vero E6 cells (Schloer et al., 2020). We next turned to focus on 23 potential natural products (Figure 6) with both of public evidence, including antiviral in vivo and anti-inflammatory effects. Literature evidence shows that 4 out of 23 natural products are confirmed to exert anti-SARS-CoV-2 effects, including colchicine, fluoxetine, mycophenolic acid, and fucoidan. For example, mycophenolic acid, a highly effective immunosuppressant which owns broad antiviral activity against different viruses, has been reported to exhibit a high antivirus activity (EC₅₀ = 0.87 μM) on SARS-CoV-2-infected VeroE6/TMPRSS2 monolayers (Kato et al., 2020). Similarly, fucoidan, commonly used in medicines and nutritional supplements for their healthy functions, also shows the inhibitory activity against SARS-CoV-2 virus with concentration of 15.6 μg/ml (Song et al., 2020). Detailed information of the predicted natural products is provided in Supplementary Table 4.

Taken together, our network-based approach offers novel and promising approved drugs and natural products with multiple supportive evidence for anti-SARS-CoV-2 effects, which deserves to be verified with various experimental assays or even randomized controlled clinical trials in the future.

To test the anti-SARS-CoV-2 activity of potential drug candidates against COVID-19, we firstly excluded drugs and natural products with confirmed efficacy on COVID-19 and only selected those with specific anti-inflammatory and antiviral properties in vivo for further assay validations. After applying subject matter expertise criteria mentioned above, 19 natural products and 17 approved drugs (Figures 5, 6) were consequently obtained for testing anti-SARS-CoV-2 activity using RT-qPCR. During the process, remdesivir was used as the positive control drug (Figure 7A, EC₅₀ = 0.54 μg/ml). In total, four promising candidates including hesperidin, isorhapontigenin, salmeterol, and gallocatechin-7-gallate were demonstrated to show inhibitory activity on SARS-CoV-2 virus-infected Vero cells. As presented in Figure 7, hesperidin (a flavanone glycoside isolated from various citrus fruits, such as orange; Küçükler et al., 2021) restrained the growth of the SARS-CoV-2 virus with an EC₅₀ value of 1.25 μg/ml (Figure 7B) while isorhapontigenin (originated from plants and fruits, such as Chinese herb Gnetum cleistostachyum; Chu et al., 2020) inhibited the replication of virus with an EC₅₀ value of 0.3 μg/ml (Figure 7C). As a citrus bioflavonoid, hesperidin possesses a variety of biological activities such as antioxidant, anti-inflammatory, and antivirus properties (Ding et al., 2018; Afzal et al., 2020). Recently, several computational methods have predicted hesperidin binds with the key proteins of the SARS-CoV-2 virus, including the “spike” protein and the main protease (Bellavite and Donzelli, 2020). In addition, salmeterol and gallocatechin-7-gallate both exhibited antiviral effects at EC₅₀ values of 0.88 and 0.98 μg/ml, respectively, (Figures 7D,E). As a novel host cdc2-like kinase 1 inhibitor, gallocatechin-7-gallate which was isolated from P. clypearia (Li et al., 2016) exhibited potent in vivo and in vitro activities against influenza virus (Li et al., 2018). For instance, gallocatechin-7-gallate could significantly improve the survival rate of the mice that were infected with H1N1 virus and reduce the lung virus titer, indicating a promising drug candidate for antiviral therapy (Li et al., 2018). Taken together, the high activities of four promising drug candidates, especially isorhapontigenin in the antiviral experimental assays, imply that these compounds may act as promising drug candidates for SARS-CoV-2, which deserve to be further exploited in the future.

In this section, we selected salmeterol (an FDA-approved drug for asthma) and isorhapontigenin (a well-known natural product) to further decipher their mechanism of actions on COVID-19. Since COVID-19 is an infectious respiratory system disease with common clinical manifestations of fever and dry cough (Tay et al., 2020), respiratory system drugs such as salmeterol are naturally developed as the potential treatment options for infected patients. As depicted in Figure 8A, salmeterol targets on seven non-n-COV host proteins (ADRB1, ADRB2, ADRB3, AOX1, DRD3, KCNH2, and THPO), meaning that it may not act on COVID-19 via directly targeting COV host proteins. After assembling its PPIs, 11 PPI partners (e.g., CEACAM1 and STAT3) were added to this network and salmeterol was computationally predicted to show noticeable correlations on COVID-19 (Z = −2.86, p < 10^–5). Interestingly, recent study has highlighted salmeterol as a potential drug against SARS-CoV-2 infection via in silico machine learning approach (Nunnari et al., 2020). However, the detailed molecular mechanism remains unclear. Figure 8A suggests that salmeterol could interact with STAT3 through integrating PPIs, providing the potential anti-SARS-CoV-2 mechanism of salmeterol. Indeed, STAT3 has been reported to affect ACE2 expression, which may be considered as a potential target for treating COVID-19 patients (Matsuyama et al., 2020; Shamir et al., 2020). Thus, it is likely that salmeterol could fight against COVID-19 via acting on STAT3.

Similarly, isorhapontigenin, a natural bioactive compound isolated from various plants and fruits (such as Chinese herb Gnetum cleistostachyum and grape), has shown several physiological activities including antiviral and anti-inflammatory effect in vitro and in vivo (Chu et al., 2020). The exact molecular mechanism against COVID-19 by isorhapontigenin is still indistinct. Network analysis (Figure 8B) suggests that isorhapontigenin interacts with one host protein and 20 PPI partners (e.g., CD209 and PARP1). Network proximity predicted isorhapontigenin with high association with COVID-19 (Z = −4.69, p = 0.005). CD209 has been reported to bind with S-RBD and also mediate SARS-CoV-2 S-pseudotyped virus entry, which can be regarded as alternative receptors for the SARS-CoV-2 virus (Amraie et al., 2020). Additionally, increasing evidence shows that the activation of PARP1 is the terminal point among a sequence of events culminating in patient mortality which should serve as the focus of COVID-19 immunotherapy (Badawy, 2020). Figure 8A shows that isorhapontigenin targets CD209 and PARP1 through protein–protein interaction, implying an anti-SARS-CoV2 mechanism of isorhapontigenin on COVID-19. Notably, we found that isorhapontigenin could also target STAT3, suggesting an alternative mechanism of isorhapontigenin on COVID-19. Taken together, aforementioned examples indicate that systems pharmacology-based network analyses offer new perspective for elucidating the anti-SARS-CoV-2 mechanisms of drug candidates. However, experimental validation is warranted to confirm its underlying mechanisms in future.