key: cord-0978063-3gobqr8e
authors: Aronskyy, Illya; Masoudi-Sobhanzadeh, Yosef; Cappuccio, Antonio; Zaslavsky, Elena
title: Advances in the computational landscape for repurposed drugs against COVID-19
date: 2021-07-30
journal: Drug Discov Today
DOI: 10.1016/j.drudis.2021.07.026
sha: a31f4f83f8870ef295050fd732c891bb60347c35
doc_id: 978063
cord_uid: 3gobqr8e

The COVID-19 pandemic has caused millions of deaths and massive societal distress worldwide. Therapeutic solutions are urgently needed but de novo drug development remains a lengthy process. One promising alternative is computational drug repurposing, which enables the prioritization of existing compounds through fast in silico analyses. Recent efforts based on molecular docking, machine learning, and network analysis have produced actionable predictions. Some predicted drugs, targeting viral proteins and pathological host pathways are undergoing clinical trials. Here, we review this work, highlight drugs with high predicted efficacy and classify their mechanisms of action. We discuss the strengths and limitations of the published methodologies and outline possible future directions. Finally, we curate a list of COVID-19 data portals and other repositories that could be used to accelerate future research.

The emergence of the novel coronavirus, known as SARS-CoV-2, became a defining event for 2019. The virus was first discovered in Wuhan, China. Transmitted via respiratory secretions, it quickly spread to neighboring countries and the rest of the world, causing COVID-19 (Coronavirus Disease 2019). To date, the COVID-19 pandemic has resulted in over four million deaths. The relatively high mortality, alongside the massive societal distress, caused by the virus stimulated intensive research on its biology, associated disease, and the development of vaccines and therapeutics.

SARS-CoV-2 is an enveloped, non-segmented, positive-sense RNA virus. It is part of the Coronaviridae family, which are known to cause respiratory infections in human hosts. It is a beta coronavirus most closely related to SARS-CoV-1, with which it shares 79% genetic similarity, and more distally related to MERS-CoV, with which it shares approximately 50% of its genomic sequence. 1, 2 Several proteins encoded by the SARS-CoV-2 viral genome have been implicated in its pathogenesis. The spike protein, which is responsible for viral entry, binds to the angiotensin converting enzyme 2 (ACE2) receptor, which is broadly expressed in human tissues. 3, 4 Other viral proteins have key roles in the viral transcriptase-replicase complex (RdRp, Mpro/3CL, PLPro, helicase). [5] [6] [7] The defining features of COVID-19 are a high rate of asymptomatic spread and high incidence of life-threatening complications. 1, [8] [9] [10] [11] Owing to these complications and the lack of immunity in the population, the scientific and medical communities launched a large-scale research response that led to the development of multiple vaccines. 12 Despite the proven efficacy of these vaccines, there is an urgent need for COVID-19 therapeutics due to vaccine distribution delays, failed immunity, vaccine hesitancy, and the possible emergence of resistant strains. 13, 14 De novo development of therapeutics is a lengthy process, and is not a practical way to address the current pandemic. 15 Drug repurposing is an emerging strategy in which existing therapeutics, that have already tested safe in humans, are repositioned to treat other diseases. 16, 17 For example, clinical trials showed that remdesivir, an antiviral known for its effectiveness against the Ebola virus, was superior to placebo in shortening the time to recovery in COVID-19 patients. 17, 18 Nevertheless, a systematic experimental screening of the potentially very large pool of drugs to be considered for repurposing is not feasible, due to both lengthy time-scales and prohibitive cost. To overcome this limitation, one promising alternative is computational drug repurposing.

Computational drug repurposing, broadly speaking, uses algorithms and massive database information to prioritize drugs for repositioning. [19] [20] [21] [22] [23] [24] [25] One major advantage of computational approaches is the potential for fast large-scale screenings. Another major advantage is the potential to identify previously unknown benefits of existing drugs in disease contexts that are radically different from the original intended target conditions. By contrast, experimental drug repurposing is typically restricted to closely related disease contexts.

Multiple reviews in the area of computational repurposing of drugs have focused on drugs and mechanisms of action with potential for use against COVID-19, regardless of the specific computational methods used for prediction. [26] [27] [28] [29] Other works have reviewed progress and perspectives relating to artificial intelligence-based methods. 30,31 Additional, more comprehensive, reviews have extended the methodological classes further to include network-, structure-, and signature-based drug repurposing methods. [32] [33] [34] Our review, like that of Galindez et al., 33 is positioned to include both a careful curation of the many published methods and the identification of the most promising drug predictions. However, Galindez et al. 33 focused on generalized workflows in each method class, whereas we provide a more in-depth review of selected works, aiming to illustrate the details of drug repurposing strategies that have resulted in experimentally validated drug predictions. In addition, a novel aspect of our review is the introduction of early progress in the category of text mining for COVID-19 drug repurposing.

The objectives of this review are: (1) the identification of published works that produce drug predictions that have been validated either in vitro or in vivo; (2) classification of the corresponding computational techniques, with a detailed presentation of selected works; (3) summarization of drug candidates and the mechanisms of pathogenesis that would be modulated; and (4) curation of the main data sources and COVID-19-specific portals. Finally, this review aims to guide future work in the field.

For a global view of efforts on computational drug repurposing, we conducted a manual curation of the literature and analyzed the COVID-19 Drug and Gene Set Library. 35 This resource includes both computationally predicted and experimentally validated drug candidates along with the associated publications. The COVID-19 Drug and Gene Set Library is one of the most comprehensive crowdsourcing initiatives to integrate drug and gene sets related to COVID-19 research.

We defined selection criteria for inclusion in this review to identify the most methodologically sound manuscripts that have undergone rigorous peer review, sufficient methodological diversity and the most promising drug candidate predictions ( Figure 1A) . First, we applied two selection criteria to the COVID-19 Drug and Gene Set Library: the presence of a computational prediction for at least one experimentally validated compound; and publication of the associated paper in a peer-reviewed journal (excluding preprints, technical reports or other resources). Second, we manually curated relevant published works that are not yet in the COVID-19 Drug and Gene Set Library to increase coverage of the different methodological areas and to incorporate other major publications. In all, we include 44 peer-reviewed manuscripts ( Figure  1A) . Consistent with previous reviews in the field of computational drug repurposing, [19] [20] [21] [22] [23] [24] [25] we group the selected works into the following categories: molecular docking, molecular docking with follow-up molecular dynamics simulations, structure-guided machine learning, networkbased methods, hybrid methods, and text mining.

Our review of the literature required the manual curation of manuscripts and drug predictions that, given the urgency of the pandemic, are being published at an exceptionally rapid rate. Thus, we restricted our curation to a time frame between the onset of the pandemic and January 2021. Analysis of publication trends during this time showed a very early production of works on molecular docking, with and without molecular dynamics simulations, followed by the publication of works on network-based approaches, structure-guided machine learning, and hybrid methods ( Figure 1B) . Currently, the majority of the published works are in the docking category (39%), docking with follow-up molecular dynamics simulations (27%), network-based approaches (20%), structure-guided machine learning (7%), and hybrid methods (5%) (Figure 1C ). To date, the text-mining category has received relatively little attention, with only one published manuscript. 36 Text-mining approaches may be underrepresented in the context of COVID-19 because they look for previously unrecognized connections between drug targets and diseases, but the genes involved in COVID-19 pathogenesis are still largely unknown.

In the context of drug repurposing for the development of therapeutics against viral infections, docking techniques assess the fitness of an interaction between small molecules and viral proteins with a known 3D structure. 22 The fitness (or binding affinity) is commonly computed as a potential energy, which is a function of force fields acting on the interacting molecular particles. Lower potential energy values (higher binding affinity) correspond to more stable conformations for complexes involving the small molecule and the viral protein, and are more likely to be associated with neutralization of the viral protein function by the ligand. Thus, small molecules that have high binding affinity for viral proteins are prioritized as potentially repurposable drugs. The well-established methods and software packages for docking approaches include GLIDE, 37 AutoDock Vina, 38 and SwissDock. 39 The packages differ from one another in terms of the employed algorithms, scoring methods, docking type (flexible or rigid), and docking elements (for example, protein-protein, protein-ligand, protein-peptide). Parks and Smith 40 provided a recent perspective on the potential advantages of molecular docking for quick repurposing of drugs for use against SARS-CoV-2.

The earliest studies faced the challenge of the lack of a solved structure for SARS-CoV-2 viral proteins. One way to circumvent this problem is to rely on the conservation of sequence and 3D structure between SARS-CoV-2 and other coronaviruses, particularly SARS-CoV-1. For example, some early works predicted drug candidates targeting SARS-CoV-2 Mpro/3CL by leveraging the sequence and structure conservation of key protein residues in SARS-CoV-1 and SARS-CoV-2 inhibitor-binding pockets. 41 A more principled way to overcome the challenge of unsolved viral protein structure is homology modelling. 42 Homology modeling refers to the construction of an atomic-resolution model of the protein of interest (e.g. SARS-CoV-2 Mpro/3CL) based on its amino acid sequence and an experimental three-dimensional structure of a homologous protein (e.g. SARS-CoV-1 Mpro/3CL). 43 For example, Wu et al. 44 systematically applied homology modeling to all of the proteins encoded by SARS-CoV-2 genes, predicting 19 structures. The authors then used these predicted structures to model their molecular docking against compound libraries such as the ZINC drug database 45 and a database of natural products. The study identified 78 commonly used antiviral drugs, including those currently on the market and undergoing clinical trials, as potential targets for SARS-CoV-2 Mpro/3CL (almitrine, carvedilol, telmisartan, conivaptan), RdRp (idarubicin), and PLpro (ribavirin). Additional studies have followed a similar strategy, leveraging homology modeling of the SARS-CoV-2 proteins RdRp, 46, 47 Mpro/3CL, 48 Plpro, 49 helicase, 50 and the spike protein, 51 as well as the human protein TMPRSS2 52, 53 in their docking studies. After the crystal structures of SARS-CoV-2 viral proteins were resolved, a second wave of papers applied docking methods directly to these solved structures to predict additional drug targets. [54] [55] [56] [57] [58] [59] [60] Interestingly, the predictions made by docking methods that used solved structures largely differed from those made using methods that leveraged homology modeling, with ribavirin being the most notable prediction made by methods in both categories.

Next, we illustrate how homology modeling, molecular docking and experimental techniques can be successfully combined by looking in depth at the study by Jin et al. 61 In an effort to develop drugs against human coronaviruses, an earlier study by Yang et al. 62 had found an MPro/3CL substrate recognition pocket that is highly conserved in the genus Coronavirus. 62 On the basis of this finding, Yang et al. 62 designed a wide-spectrum inhibitor of Mpro/3CL, named N3, that showed antiviral activity against SARS-CoV-1 and MERS. Jin et al. 61 started by constructing a homology model for SARS-CoV-2 Mpro/3CL, and used molecular docking to ascertain that N3 can target the modeled Mpro/3CL. Next, Jin et al. 61 experimentally assessed the efficacy of N3 for SARS-CoV-2 Mpro/3CL through analysis of binding kinetics, showing potent inhibition. In addition, the study determined the crystal structure of SARS-CoV-2 Mpro/3CL in complex with N3. The determined structure served as a model for in silico screening of candidate inhibitors to target SARS-CoV-2 Mpro/3CL. Using the Glide docking software package, the analysis identified cinanserin, a well-characterized serotonin antagonist previously shown to inhibit SARS-CoV-1 Mpro/3CL, 63 as a potential inhibitor of SARS-CoV-2 Mpro/3CL. To discover other potential Mpro/3CL inhibitors, Jin et al. 61 {AuQ: Edit OK?} used a fluorescence resonance energy transfer assay to screen a library of about 10,000 compounds consisting of approved drugs, clinical-trial drug candidates and natural products. The primary candidates included the approved drug disulfiram, as well as the clinical-trial drug candidate ebselen, which had shown efficacy against other viruses. [64] [65] [66] Further in vitro experiments measuring viral replication in cell-based assays confirmed strong antiviral activity for N3 and ebselen.

Although docking methods can assess the fitness of an interaction between a drug and its intended target, they typically predict a large number of potential drug candidates. An effective way to restrict this pool of candidates is provided by Molecular Dynamics (MD) simulations. Docking methods are based on a static description of the interactions between drugs and targets, whereas MD gives a view of their dynamic interactions by analyzing the physical movements of atoms and molecules over time. Because molecular systems typically consist of a vast number of particles, the computational cost of MD simulations can be considerable. Therefore, it is computationally prohibitive to employ MD simulations to screen systematically against large drug databases, and MD simulations are typically applied to the most promising targets preselected through docking. MD and docking methods complement each other, leading to more reliable predictions and to less follow-up experimental work. Multiple software packages for MD simulations are available, including GROMACS, 67 and AMBER. 68 MD simulations, in combination with molecular docking, have been used to identify drug candidates against SARS-CoV-2 Mpro/3CL, 53,69-74 RdRp, 75 nsp16 and nsp10, 76 against the human protein ACE2, 77 and against the complex between the SARS-CoV-2 spike protein and ACE2. 51, 78 Structure-guided machine-learning approaches The recent rapid growth of drug-related data sets, as well as of open data initiatives, has led to new developments in machine-learning methods, especially in deep-learning techniques. Docking methods compute binding affinity using the detailed molecular structures of drugs and their targets, whereas machine-learning models use simplified, vector representations of these structures. In the simplest case, for example, the so-called 'fingerprint' representation of a molecule is a vector of the structural fragments in the molecule. Machine-learning models require a training set that pairs up input features, that is molecular descriptors of drugs and their targets (such as their molecular fingerprints), and outputs such as binding affinities {AuQ: Edit OK?}. Like docking methods, trained machine-learning models produce a prediction that is an estimate of the binding affinity between a drug candidate and its target proteins, either viral or host. When compared to docking, machine-learning models have additional advantages in that: they can be applied in the absence of a solved 3D protein structure, which was the case for some early papers using this approach; and the calculation of the binding affinity is more computationally efficient.

Focusing on the SARS-COV-2 Mpro/3CL, Gao et al. 79 proposed a machine-learning-based approach in which they compiled a carefully selected training set of 314 SARS-CoV-1 or SARS-CoV-2 Mpro/3CL protease inhibitors and their binding affinities collected from single protein experiments. The input variables for each inhibitor were constructed by encoding its highdimensional biomolecular interactions with Mpro/3CL in a low-dimensional representation. A gradient boosting decision tree model showed good performance (Pearson correlation between predicted and true binding affinities of 0.79). Gao et al. 79 then used the model to predict the binding affinities between the SARS-CoV-2 Mpro/3CL protease and 1553 US Food and Drug Administration (FDA)-approved drugs as well as 7012 investigational or off-market drugs. Among the FDA-approved drugs with predicted high-affinity scores were chloroxine (an antibacterial drug that belongs to the same family as chloroquine, but with a different predicted mechanism of action), demexiptiline (a tricyclic antidepressant) and fluorouracil (an antineoplastic). Among investigational compounds, the top result was debio-1347 (a fibroblast growth factor receptor (FGFR) inhibitor under investigation for the treatment of solid tumors). Interestingly, the predicted binding affinity of debio-1347 for SARS-CoV-2 Mpro/3CL was higher than that of any of the FDA-approved drugs, indicating the potential of debio-1347 for COVID-19 repurposing.

Batra et al. 80 trained a structure-based machine-learning model to repurpose drug candidates against the spike protein and the spike-human ACE2 interface complex. The model used as input includes a set of structural and physicochemical features that capture the interactions between the drug molecule and the target proteins at various spatial scales. As output, the model predicted the corresponding Autodock Vina score. Predictions were based on a random forest model trained on docking-generated binding affinities. Top-scoring ligands, including FDA-approved antivirals (such as valaciclovir) and antibacterials (such as sulfamethoxazole), were further validated in silico by all-atom docking studies.

Studies by Beck et al. 81 and by Ton et al. 82 employed deep-learning-based models, called Molecule Transformer-Drug Target Interaction (MT-DTI) and Deep Docking, respectively, to predict drug-target interactions. MT-DTI was trained with a database of approximately a billion compounds combining data from the BindingDB database 83 and the Drug Target Common database, 84 and was used to predict the binding affinity between SARS-CoV-2 proteins and 3411 FDA-approved drugs. Interestingly, the predictions included drug candidates that are able to target multiple viral proteins simultaneously. For example, atazanavir was predicted to bind SARS-CoV-2 Mpro/3CL, RdRp, and helicase. Similarly, ganciclovir was predicted to bind to three subunits of the replication complex. Deep Docking, a recently developed deep-learning platform shown to provide accurate and fast approximations of docking outcomes, enabled a virtual screening of the entire ZINC library of over 10 9 compounds against Mpro/3CL. 45 Analysis of the top 1000 candidate inhibitors showed good chemical diversity and superior docking scores relative to those of known Mpro/3CL inhibitors. This method demonstrates the plausibility of performing exhaustive virtual screening, although the large number of predictions make it difficult to evaluate the success rate and the overall clinical applicability.

During the viral life cycle, the virus interacts with multiple components of the host cell. Previously described approaches have aimed to identify inhibitory molecules that directly target viral proteins, but an alternative therapeutic option is to target the host or the virus-host interactome, typically using techniques from network pharmacology. The host interactome consists of human proteins as nodes and their interactions as edges, drawing upon various repositories of proteinprotein interactions (PPI, see Table 1 ). The virus-host interactome extends this network by including viral proteins and their interactions with host proteins, typically leveraging highthroughput experiments. Additional data, such as transcriptomics, are used to derive subnetworks that are relevant to the disease under study, in this case COVID-19. Finally, databases of drug targets are utilized to predict druggable nodes in these subnetworks, identifying potential therapeutics that can perturb key processes of COVID-19 pathology.

Early papers aimed to identify COVID-19 disease-relevant subnetworks that were needed to overcome the lack of data on interactions between SARS-CoV-2 and human proteins. Alakwaa 85 and Cava et al. 86 focused solely on host networks centered around ACE2, which is broadly expressed and is critical for SARS-CoV-2 cellular entry. They identified ACE2-associated networks derived from the analysis of gene expression in lung tissue transcriptomes in the public domain (e.g. GTEx 87 ), and mapped databases of drug perturbations onto these gene sets. The work of Zhou et al. 88 built upon a similar strategy and constructed a broader pan Coronavirus-host interactome network, which included a set of 119 host proteins associated with previously characterized human, mouse, and avian Coronaviruses. The drug repurposing methodology rested on the idea that for a drug to be effective, its targets should be either in the Coronavirushost interactome or in close proximity to the interactome within the global human PPI network. Guided by this idea, alongside additional information such as transcriptomics data, literaturereported antiviral evidence and known side effects, the authors prioritized four classes of drugs. These included estrogen receptor modulators, angiotensin receptor blockers, immunosuppressant or antineoplastic agents, and anti-inflammatory agents. Finally, Zhou et al. 88 predicted three potentially synergistic drug combinations for COVID-19: sirolimus (immunosuppressor) + dactinomycin (RNA synthesis inhibitor), toremifene (estrogen receptor modulator) + emodin (antiviral agent), and mercaptopurine (antineoplastic agent) + melatonin (anti-inflammatory agent).

The success of network-based repurposing strategies relies on solid information on interactions between viral and human proteins. In contrast to the computationally derived interactome in the Zhou et al. 88 study, Gordon et al. 89 experimentally identified 332 highconfidence interactions between SARS-CoV-2 and human proteins. For each viral protein, a set of interacting host proteins was determined using affinity purification-mass spectrometry, and analyzed via immunological pathway analysis. This analysis revealed that the interacting host proteins participate in major cellular processes including DNA replication, epigenetic and transcriptional regulation, multiple innate immune pathways, and host translational machinery. Next, the authors that predicted 69 agents, including FDA-approved drugs, would be able to disrupt host proteins and cellular processes that are potentially hijacked during SARS-CoV-2 infection. Further experiments revealed the antiviral activity of two broad classes of active compounds: those impinging on the translational machinery of the host needed for viral biogenesis, and those modulating the sigma1 and sigma2 receptors (chaperone proteins at the endoplasmic reticulum). Translation inhibitors (such as zotatifin and aplidin/plitidepsin) and sigma receptor ligands that have antiviral activity (such as haloperidol and clemastine/cloperastine) were identified as potential candidates for repurposing of drugs against COVID-19.

Bouhaddou et al. 90 performed the first proteomics and phosphoproteomics analysis to explore how host pathways rewire upon SARS-CoV-2 infection. The study revealed a prominent role of phosphorylation events in modulating the cell cycle and cytoskeletal organization of infected cells. The strategy for drug repurposing hinged on detecting changes in phosphorylation states between SARS-CoV-2-infected and control cells. These identified changes were then mapped to changes in the activity of 97 kinases on the basis of database knowledge linking kinases with their substrates. 91 Of note, several inferred kinases were members of the virus-host PPI map generated by Gordon et al. 89 Finally, Bouhaddou et al. 90 looked for drugs that were able to inhibit the identified kinases, producing 87 drug candidates. In vitro assays involving 68 out of these 87 kinase inhibitors, including silmitasertib (an inhibitor of CK2), gilteritinib (an inhibitor of AXL), and ARRY-797 (an inhibitor of p38) demonstrated strong antiviral activity.

More recent studies took advantage of the increasing availability of transcriptomics data, which make it possible to identify SARS-CoV-2-regulated genes and pathways, and their overlap with drug-modulated genes. 92, 93 By applying network proximity measures to multi-omics data from diverse tissues, Zhou et al. 94 revealed similarities between COVID-19 and other pathologies, including inflammatory bowel disease and asthma. The shared pathological processes could suggest candidates for repurposing in COVID-19.

The important study of Bojkova et al., 95 which makes and validates multiple predictions, does not fit the standard computational drug repurposing classification. We place this study in the 'Hybrid' category because it integrates diverse computational techniques as well as experimental work.

To identify host pathways that are modulated by SARS-CoV-2 infection, Bojkova et al. 95 established a novel cell culture model using a human epithelial carcinoma cell line. The SARS-CoV-2 infection profile, comprising both viral and host proteins, was quantified through translatome and proteome analysis at multiple time points within 2-24 hours after infection. First, the authors identified host proteins whose translation kinetics showed high correlation with those of viral proteins. Network and enrichment analyses of these proteins showed a strong increase in the host translation machinery, suggesting that translation inhibitors may reduce SARS-CoV-2 replication. This hypothesis was validated using two translation inhibitors, cycloheximide and emetine. To gain further insight into host proteome changes after infection, Bojkova et al. 95 performed hierarchical clustering of the host proteins' expression profiles over time. This revealed an upregulated cluster of proteins that was enriched in spliceosome components and carbon metabolism, suggesting that targeting these two pathways may prevent SARS-CoV-2 replication. Both hypotheses were validated using, respectively, pladienolide B (a spliceosome inhibitor) and 2-deoxy-D-glucose (an inhibitor of hexokinase implicated in carbon metabolism). Additional correlation analysis of the kinetics of host and viral proteins showed that SARS-CoV-2 replication can be also inhibited using ribavirin and NMS-873 (a compound targeting the protein homeostasis machinery). Altogether, this work revealed that SARS-CoV-2 infection modulates critical host functions such as splicing, diverse metabolism sub-pathways, and proteostasis. Drugs that interfere with these pathways can be successfully repurposed to prevent SARS-CoV-2 replication in a human cell line and are good candidates for further testing in clinical settings for COVID-19.

The compilation of knowledge on COVID-19 has grown exponentially since the beginning of the pandemic. Several initiatives, such as the COVID-19 Open Research Dataset (CORD-19) and LitCovid, have made the published work more easily accessible. 96, 97 This opens up great opportunities for analysis based on text mining and natural language processing. A recent publication by Wang et al. 97 provides a review of the different text-mining systems and their applications to COVID-19 research, including information retrieval, automated annotation, and drug repurposing. 98 The COVID-19 Drug Repository is the first significant work on computational drug repurposing based on text mining. 36 The goal of this repository, which integrates data from PubMed, clinical trials and drug databases, is to centralize information on drugs against COVID-19 in a structured framework. The available information includes drug descriptions, side effects, pharmacology-oriented data, and drug-gene interactions. The resource makes it possible to identify putative repurposing candidates by analyzing omics data in conjunction with mined drug-target interactions.

More generally, text mining can be used to derive 'knowledge graphs'. These are models of entities and relationships in a particular domain (for example, COVID-19 research) that can support the inference of novel relationships. In our context, a novel relationship could represent the potential efficacy of an existing drug against COVID-19. 98 Finally, another promising research direction is the mining of electronic health records, prescription records and medical insurance claims. 99 As these data types accumulate in the context of COVID-19, frameworks that identify new uses for on-market prescription drugs by systematically analyzing healthcare databases can be applied. 100 Although still in their early stages, the mining of text and healthcare records will become an increasingly important tool in addressing the information overload, and in generating new hypotheses for drug repurposing.

Computational drug repurposing for SARS-CoV-2 has homed in on two main mechanisms of action ( Table 2) . The most frequent mechanism is the direct inhibition of SARS-CoV-2 viral proteins (Figure 2A , seen in 31 of 44 reviewed works). The criterion for repurposing is the high binding affinity of the drugs to viral proteins, which is estimated using molecular docking or machine learning. This strategy has previously led to successful repurposing of antimicrobial drugs, including inhibitors of the Zika Virus. 17 In the context of SARS-CoV-2, potentially repurposable drugs include multiple inhibitors of RdRp, Mpro/3CL protease, and other viral proteins. We found extensive overlap between the drug compounds predicted by the reviewed works. Many commonly used antivirals that have garnered international attention and have been used clinically in the treatment of COVID-19, 101,102 such as ritonavir and ribavirin, were predicted by up to 12 different computational approaches (Figure 2B) .

The second mechanism is the inhibition of host pathways using network-based and hybrid approaches (Figure 2A , seen in 9 of 44 reviewed works). These methods identified host targets involved in virus life cycle, particularly entry and RNA synthesis. Progress in identifying inhibitors of host pathways that contribute to pathogenesis is critical, particularly for a virus that has already mutated enough to render some previously effective therapeutics significantly less potent. Of note, current methods are less effective in identifying drugs that target other key host pathways that contribute to pathology, such as aberrant immune responses and thrombosis, whose therapeutic manipulation showed promising results. [103] [104] [105] [106] This may be due to a poor molecular understanding of the systems-level effects of SARS-CoV-2 infection. Finally, a few studies have identified both viral and host proteins as potential targets (Figure 2A , seen in 4 of 44 reviewed works). These predictions were produced by docking and structure-based machine learning, which are not intrinsically limited to the identification of viral proteins as targets.

The strengths, limitations, and successful applications of the methodologies included in this review have been discussed previously. [19] [20] [21] [22] [23] [24] [25] In the context of COVID-19, an additional challenge is the lack of solid prior biological knowledge, with the knowledge required for each of the different method classes being accumulated at different rates. For example, the SARS-CoV-2 protein structure, which is required for docking, was quickly solved for several important viral proteins. On the other hand, the mapping of the virus-host interactome, which is needed for network approaches, is a research objective that will require longer-term effort. Therefore, efforts to assess the success rates of the different methods for COVID-19 drug repurposing are still premature. Regardless of the specific method being used, the preconditions for success go beyond target activity to encompass other key pharmacological aspects, such as safety, toxicity and drug-drug interactions. Furthermore, as noted by Edwards, the success rate of hypothesisdriven repurposing research far exceeds that of large-scale, hypothesis-free screenings. 107 This concept is also evident in works of Gordon et al., 89 Jin et al. 61 and Bojkova et al. 95 reviewed here, which first produced data specifically for hypothesis generation, before validating the predictions in vitro. Going forward, the increasing accumulation of data, as well as improved knowledge of COVID-19 pathology, should improve our ability to strike the right balance between hypothesisdriven screenings and the predominantly data-driven repurposing strategies reviewed here.

With this objective in mind, we have curated a list of COVID-19-specific data portals. We also include a list of other relevant resources typically employed by computational repurposing methods, both for COVID-19 and more broadly ( Table 1) . These repositories contain data describing gene and protein sequences and structures; drugs and drug targets; properties of chemical compounds; and biological networks and pathways. Furthermore, several -omics datasets, profiling SAR-CoV-2 infection in various tissues, have been published. [108] [109] [110] [111] [112] [113] Although not exhaustive, our curated lists include the most commonly utilized repositories and provide a basis for future work.

Effective treatment of complex diseases, such as viral infections, may require the targeting of multiple host and viral processes. It is widely recognized that the severity and mortality COVID-19 are highly dependent upon biological factors, such as older age, and upon pre-existing conditions, such as hypertension, diabetes, cardiovascular disease, or chronic obstructive pulmonary disease (COPD). [114] [115] [116] [117] The optimal treatment is likely to depend on a specific patient's characteristics, so future repurposing methods may account for variation in risk factors, leading to personalized treatments.

Combination therapies increase the spectrum of disease modulation, improving drug efficacy/toxicity ratios and combating drug resistance. 118 A synergistic effect can be created by using several drugs that target different disease pathways. 119 For example, combinational drug therapy has become the gold standard in the treatment of cancer 120 and in the treatment of several virus infections, notably HIV. 121 The principle of 'complementary exposure', 122 which optimizes drug effects over a large portion of the disease network, has predicted several synergistic drug combinations for the treatment of COVID-19. One example is sirolimus and dactinomycin, which inhibit mTOR signaling and RNA synthesis pathways, respectively. An important challenge and a promising objective for future computational drug repurposing is the ability to design effective drug combinations that target multiple COVID-19 pathological pathways.

Computational drug repurposing has led to remarkable success stories in several disease areas, including cancer, autoimmune diseases, and viral infections. Despite the many challenges, our review lays out the early promising steps towards successful computational repurposing for COVID-19. To identify the most effective computational drug repurposing works, we used the COVID-19 Drug and Gene Set Library, 35 complemented by manual curation of the literature. The selection criteria aimed to identify methods that resulted in at least one predicted compound that was validated with further experiments. B. We classify the selected works into six groups: docking, docking/MD (molecular dynamics), network, structure-guided machine-learning (ML), hybrid and text mining. The publication trends for these groups in 2020 are shown. C. The final numbers of publications in each category included in our review. 

The trinity of COVID-19: immunity, inflammation and intervention

The genetic sequence, origin, and diagnosis of SARS-CoV-2

Comprehensive analysis of drugs to treat SARS-CoV-2 infection: mechanistic insights into current COVID-19 therapies (Review)

ACE2 & TMPRSS2 expressions in head & neck tissues: a systematic review

Genotype and phenotype of COVID-19: their roles in pathogenesis

Structure-based design of antiviral drug candidates targeting the SARS-CoV-2 main protease

COVID-19 pandemic: insights into structure, function, and hACE2 receptor recognition by SARS-CoV-2

Immunity, endothelial injury and complement-induced coagulopathy in COVID-19

COVID-19 pathophysiology: a review

Neurological manifestations of COVID-19 (SARS-CoV-2): a review

Cardiac and arrhythmic complications in patients with COVID-19

SARS-CoV-2 vaccines: status report

Evidence of escape of SARS-CoV-2 variant B.1.351 from natural and vaccine-induced sera

Clinical outcomes of a COVID-19 vaccine: implementation over efficacy

Development times and approval success rates for drugs to treat infectious diseases

Drug repurposing: progress, challenges and recommendations

Drug repurposing for viral infectious diseases: how far are we?

Remdesivir for the treatment of Covid-19 -final report

Machine learning models for drug-target interactions: current knowledge and future directions

Drug databases and their contributions to drug repurposing

The COVID-19 Drug and Gene Set Library

COVID19 Drug Repository: text-mining the literature in search of putative COVID19 therapeutics

Glide: a new approach for rapid, accurate docking and scoring. 1. Method and assessment of docking accuracy

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

SwissDock, a protein-small molecule docking web service based on EADock DSS

How to discover antiviral drugs quickly

Potential inhibitors against 2019-nCoV coronavirus M protease from clinically approved medicines

Homology modeling in drug discovery: current trends and applications

SWISS-MODEL: homology modelling of protein structures and complexes

Analysis of therapeutic targets for SARS-CoV-2 and discovery of potential drugs by computational methods

ZINC 15 -ligand discovery for everyone

Anti-HCV, nucleotide inhibitors, repurposing against COVID-19

Tenofovir against SARS-CoV-2 RNA dependent RNA polymerase (RdRp): a molecular docking study

Prediction of the SARS-CoV-2 (2019-nCoV) 3C-like protease (3CL pro) structure: virtual screening reveals velpatasvir, ledipasvir, and other drug repurposing candidates

silico drug repurposing for COVID-19: targeting SARS-CoV-2 proteins through docking and consensus ranking

State-of-the-art tools unveil potent drug targets amongst clinically approved drugs to inhibit helicase in SARS-CoV-2

Molecular docking study of potential phytochemicals and their effects on the complex of SARS-CoV2 spike protein and human ACE2

Virtual screening of natural products against type II transmembrane serine protease (TMPRSS2), the priming agent of Coronavirus 2 (SARS-CoV-2)

Drug repurposing for coronavirus (COVID-19): in silico screening of known drugs against coronavirus 3CL hydrolase and protease enzymes

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

Drug repurposing for candidate SARS-CoV-2 main protease inhibitors by a novel in silico method

In silico studies on therapeutic agents for COVID-19: drug repurposing approach

Prediction of novel inhibitors of the main protease (M-pro) of SARS-CoV-2 through consensus docking and drug reposition

Potential covalent drugs targeting the main protease of the SARS-CoV-2 coronavirus

Virtual screening and repurposing of FDA approved drugs against COVID-19 main protease

Potential COVID-19 papain-like protease PLpro inhibitors: repurposing FDA-approved drugs

Structure of Mpro from SARS-CoV-2 and discovery of its inhibitors

Design of wide-spectrum inhibitors targeting coronavirus main proteases

Cinanserin is an inhibitor of the 3C-like proteinase of severe acute respiratory syndrome coronavirus and strongly reduces virus replication in vitro

The selenium-containing drug ebselen potently disrupts LEDGF/p75-HIV-1 integrase interaction by targeting LEDGF/p75

Ebselen inhibits hepatitis C virus NS3 helicase binding to nucleic acid and prevents viral replication

Ebselen alleviates testicular pathology in mice with Zika virus infection and prevents its sexual transmission

The Amber biomolecular simulation programs

Identification of potential molecules against COVID-19 main protease through structure-guided virtual screening approach

Using integrated computational approaches to identify safe and rapid treatment for SARS-CoV-2

Peptide-like and smallmolecule inhibitors against Covid-19

Computational studies of drug repurposing and synergism of lopinavir, oseltamivir and ritonavir binding with SARS-CoV-2 protease against COVID-19

Molecular docking and dynamics simulation of FDA approved drugs with the main protease from 2019 novel coronavirus

Identification of chymotrypsin-like protease inhibitors of SARS-CoV-2 via integrated computational approach

SARS-CoV-2 RNA dependent RNA polymerase (RdRp) targeting: an in silico perspective

Potential drugs targeting early innate immune evasion of SARS-Coronavirus 2 via 2'-O-methylation of viral RNA

In silico study the inhibition of angiotensin converting enzyme 2 receptor of COVID-19 by Ammoides verticillata components harvested from Western Algeria

An integrated drug repurposing strategy for the rapid identification of potential SARS-CoV-2 viral inhibitors

Screening of therapeutic agents for COVID-19 using machine learning and ensemble docking studies

Predicting commercially available antiviral drugs that may act on the novel coronavirus (SARS-CoV-2) through a drug-target interaction deep learning model

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

BindingDB: a web-accessible database of experimentally determined protein-ligand binding affinities

Drug Target Commons 2.0: a community platform for systematic analysis of drug-target interaction profiles

Repurposing didanosine as a potential treatment for COVID-19 using single-cell RNA sequencing data

In silico discovery of candidate drugs against Covid-19

The human transcriptome across tissues and individuals

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

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

The global phosphorylation landscape of SARS-CoV-2 infection

The functional landscape of the human phosphoproteome

Transcriptome-based drug repositioning for coronavirus disease 2019 (COVID-19)

A new advanced in silico drug discovery method for novel coronavirus (SARS-CoV-2) with tensor decomposition-based unsupervised feature extraction

A network medicine approach to investigation and population-based validation of disease manifestations and drug repurposing for COVID-19

Proteomics of SARS-CoV-2-infected host cells reveals therapy targets

Keep up with the latest coronavirus research

CORD-19: The Covid-19 Open Research Dataset

Text mining approaches for dealing with the rapidly expanding literature on COVID-19

Repurposing current therapeutics for treating COVID-19: a vital role of prescription records data mining

Framework for identifying drug repurposing candidates from observational healthcare data

A trial of lopinavir-ritonavir in adults hospitalized with severe Covid-19

Ribavirin therapy for severe COVID-19: a retrospective cohort study

COVID-19: complement, coagulation, and collateral damage

Dexamethasone treatment for the acute respiratory distress syndrome: a multicentre, randomised controlled trial

Repurposing existing drugs for the treatment of COVID-19

Harnessing adenosine A2A receptors as a strategy for suppressing the lung inflammation and thrombotic complications of COVID-19: potential of pentoxifylline and dipyridamole

What are the odds of finding a COVID-19 frug from a lab repurposing screen?

Identification of SARS-CoV-2 inhibitors using lung and colonic organoids

SARS-CoV-2 infection of human iPSC-derived cardiac cells predicts novel cytopathic features in hearts of COVID-19 patients

Temporal and spatial heterogeneity of host response to SARS-CoV-2 pulmonary infection

Two distinct immunopathological profiles in autopsy lungs of COVID-19

Single-cell landscape of immunological responses in patients with COVID-19

A singlecell atlas of the peripheral immune response in patients with severe COVID-19

Risk factors for severity and mortality in adult COVID-19 inpatients in Wuhan

COVID-19 diagnosis and management: a comprehensive review

Predictive symptoms and comorbidities for severe COVID-19 and intensive care unit admission: a systematic review and meta-analysis

Transmission and risk factors of OF COVID-19

Biomolecular network-based synergistics drug combination discovery

Drug combination therapy increases successful drug repositioning

Combination therapy in combating cancer

Current status and prospects of HIV treatment

Virus-CKB: an integrated bioinformatics platform and analysis resource for COVID-19 research

LitCovid: an open database of COVID-19 literature

COVID-KOP: integrating emerging COVID-19 data with the ROBOKOP database

HIT-COVID, a global database tracking public health interventions to COVID-19

Crowdsourcing drug discovery for pandemics

CIDO, a community-based ontology for coronavirus disease knowledge and data integration, sharing, and analysis

Exploring the SARS-CoV-2 virus-host-drug interactome for drug repurposing

The protein data bank

Pharmgkb: the Pharmacogenetics Knowledge Base

DrugBank 5.0: a major update to the DrugBank database for 2018

The European Nucleotide Archive

GeneCards Version 3: the human gene integrator

Complement Map Database

Molecular signatures database (MSigDB) 3.0

UniProt: a hub for protein information

The SWISS-PROT protein sequence data bank

PRIDE: the proteomics identifications database

CORUM: the comprehensive resource of mammalian protein complexes

KEGG: Kyoto encyclopedia of genes and genomes

Artificial intelligence approach fighting COVID-19 with repurposing drugs

Searching for target-specific and multi-targeting organics for Covid-19 in the Drugbank database with a double scoring approach

Antimalarial drugs inhibit the replication of SARS-CoV-2: an in vitro evaluation

Efficacy of almitrine in the treatment of hypoxemia in SARS-Cov-2 acute respiratory distress syndrome

Analysis of the efficacy of HIV protease inhibitors against SARS-CoV-2's main protease

Atazanavir, alone or in combination with ritonavir, inhibits SARS-CoV-2 replication and proinflammatory cytokine production

The inhaled steroid ciclesonide blocks SARS-CoV-2 RNA replication by targeting the viral replicationtranscription complex in cultured cells

Hydroxychloroquine in nonhospitalized adults with early COVID-19 : a randomized trial

Virtual and in vitro antiviral screening revive therapeutic drugs for COVID-19

Remdesivir, lopinavir, emetine, and homoharringtonine inhibit SARS-CoV-2 replication in vitro

Identifying and repurposing antiviral drugs against severe acute respiratory syndrome coronavirus 2 with in silico and in vitro approaches

Discovery of the FDAapproved drugs bexarotene, cetilistat, diiodohydroxyquinoline, and abiraterone as potential COVID-19 treatments with a robust two-tier screening system

Remdesivir and chloroquine effectively inhibit the recently emerged novel coronavirus (2019-nCoV) in vitro

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

COVID-19 treatment: close to a cure?