Computational anti-COVID-19 drug design: progress and challenges

Abstract

Vaccines have made gratifying progress in preventing the 2019 coronavirus disease (COVID-19) pandemic. However, the emergence of variants, especially the latest delta variant, has brought considerable challenges to human health. Hence, the development of robust therapeutic approaches, such as anti-COVID-19 drug design, could aid in managing the pandemic more efficiently. Some drug design strategies have been successfully applied during the COVID-19 pandemic to create and validate related lead drugs. The computational drug design methods used for COVID-19 can be roughly divided into (i) structure-based approaches and (ii) artificial intelligence (AI)-based approaches. Structure-based approaches investigate different molecular fragments and functional groups through lead drugs and apply relevant tools to produce antiviral drugs. AI-based approaches usually use end-to-end learning to explore a larger biochemical space to design antiviral drugs. This review provides an overview of the two design strategies of anti-COVID-19 drugs, the advantages and disadvantages of these strategies and discussions of future developments.

1 Introduction

The emerging coronavirus disease 2019 (COVID-19) is a disease caused by severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2), which has structural similarities with SARS-CoV-1 and poses a massive crisis to global public health. SARS-CoV-2 can be spread through air droplets, etc. [1]. Due to the high infection rate of COVID-19, a dramatic increase occurred in terms of case and death numbers. As of 10 June 2021, there were 174 061 995 confirmed cases and 3 758 560 deaths, according to the World Health Organization.

Vaccines are considered one of the most effective methods to help human society return to normal. Several vaccines have been authorized for emergency use. Although vaccines have produced very positive effects in many countries, they still face several significant challenges. Studies have reported that a small number of patients presented with venous thrombosis and thrombocytopenia after receiving the first dose of the SARS-CoV-2 vaccine against COVID-19 [2–5]. Moreover, the vaccine protection effect is not 100%. For example, the efficacy of the two whole-virus inactivated vaccines designed by the China National Biotech Group Company Limited on symptomatic COVID-19 cases was 72.8% and 78.1% in the third phase, separately [6, 7]. The last and most important concern is whether the vaccines are still effective against the emerging SARS-CoV-2 variants. At present, delta has quickly become the dominant SARS-CoV-2 variant. Research suggests that vaccines offer slightly reduced protection against delta, and vaccinated people with breakthrough infections can spread the delta variant [8, 9]. Whether vaccines can effectively slash the spread of delta remains unknown.

Drug development is another important way to defend against viruses. Drug repositioning has been a critical direction in the field of drug research. In recent months, researchers have searched for drugs to treat COVID-19 by finding new therapeutic targets and discovering often unknown relationships among apparently distant diseases. However, most drugs, such as remdesivir, dexamethasone and hydroxychloroquine, fail to display efficacy in treating COVID-19 [10].

The design of exclusive therapeutic drugs is a trend from the perspective of structural biology and molecular physics [11, 12]. Drug design may have a more therapeutic effect than drug repositioning and drug combinations in the disease of complex, rare or chronic, etc. [13, 14]. Currently, many specific drugs have been designed by virtual screening techniques based on specific pharmacological insights or an end-to-end framework to control the generation of molecules [15–20].

With the in-depth study of diseases and exploration of the structure of target proteins [21–25], many new technologies have been applied to drug design in extensive practice of medicinal chemistry by improving potency, altering physical properties and eliminating or modifying toxicophores [26–30]. All drug designs are the interaction between a drug and its target (usually proteins). Therefore, improving methods of predicting the magnitude of protein–ligand interactions can improve the efficiency of drug development [31–34]. Drug discovery steps require structural optimization of lead compounds to establish the highest possible level of selectivity, potency, and appropriate physicochemical and pharmacokinetic characteristics [26, 35, 36]. Critically, surveying binding hotspots in protein surfaces can help guide the exploration of potential ligand-binding regions [37–41].

The world is plagued by the emergence of the SARS-CoV-2 virus. Unfortunately, many methods face data scarcity when designing drugs for new targets [42]. With the development of big data and computer technology, the application of machine learning and deep learning as drug design algorithms has grown in recent decades [43–46]. Deep learning has become more active in the preparation and process of drug design, such as predicting molecular properties and activity [47, 48], identifying drug–target interactions [49–51] and planning chemical syntheses [44, 52–54]. Therefore, the potential of deep learning and molecular modeling methods helps develop drug design pipelines, especially where there are limited or unavailable target-specific ligand datasets [55, 56]. The designed drug must have an excellent inhibitory effect on the disease. Nevertheless, the prediction of the pharmacokinetics and toxicity characteristics of the scheduled drug can avoid the failure of clinical trials [57, 58]. Now, a set of tools incorporating in silico and deep learning are used to advance sequence-based or structure-based drug design problems in the computer-aided drug design area [19, 59–64].

Therefore, the rapid application of drug design on an increasingly broader scale with the advancement of biometrics and bioinformatics [65–67]. Efficient tools are now available for systematically designing compounds with biological activity as preliminary drug candidates [68–70]. Drug design strategies have been applied to various epidemic diseases [38, 71–74]. For example, to avoid the early infection of HIV, a generative adversarial autoencoder[75], which combines a neural network with virtual screening of a chemical database, was developed to design potential HIV-1 entry inhibitors. However, as COVID-19 is raging on a global scale, researchers have focused their attention on drug design against SARS-COV-2.

SARS-CoV-2 is a single-stranded RNA virus that includes two types of proteins: small envelope (E) glycoprotein, structural [spike (S) glycoprotein, membrane (M) glycoprotein, and nucleocapsid (N) protein] and nonstructural (NSP) protein (i.e. nsp1-16), which has a genome size of approximately 30 000 bp [76]. All SARS-CoV-2 proteins play an essential role in pathogenesis and virus replication. For example, S is a promising drug target because it attaches to human cells and participates in entering the cells [77]. NSP is contained in ORF1a and ORF1ab, and they produce two polyproteins, Pp1a and Pp1ab [78, 79]. The latter protein is produced by ribosomal transfer, enabling continuous translation of ORF1a and ORF1ab [80]. Specifically, ORF1ab is the largest protein of SARS-CoV-2, and the ORF1ab gene of human -coronavirus (HBC) species has a signature of a strong positive selection site in the genome analysis of SARS-CoV-2. The positively selected sites of ORF1ab could justify some clinical features of SARS-CoV-2 compared with other HBCs [81]. Moreover, mutational spectra should be considered when designing drugs [82]. The Pp1a protein contains two viral proteases, 3C-like main protease (Mpro, corresponding to nsp5) and papain-like protease (PLpro, a domain of nsp3)[83]. The main protease (M) of SARS-CoV-2 is a crucial enzyme of coronaviruses and has a pivotal role in mediating viral replication and transcription [20, 84–86]. PL is also critical to SARS-CoV-2 replication and represents a promising goal for drug design and development [87–89]. Among nonstructural proteins, the large, multidomain Nsp3 is encoded by SARS-CoV-2. One of its units is the ADP-ribose phosphatase domain (ADRP; also known as the macro domain, MacroD), which interferes with the host immune response [90].

It is well known that molecular inhibitors such as drugs can achieve inhibitory effects by targeting cancer or virus expression pathways. For example, seasonal and pandemic influenza have a substantial impact on global public health [91]. A study found that endonuclease activity exists in the independently folded N-terminal domain of PA (PAN) [92, 93], where PA is a subunit of RNA-dependent RNA polymerase (RdRp) that can catalyze viral transcription. Subsequently, the molecular mechanism of the inhibitor was structurally confirmed, and the interaction sites in the crystal structure of the virus strain and the inhibitor were obtained [94]. This information also provides a basis for drug discovery and design. The molecular mechanisms involved in immune regulation during the new coronavirus infection have played a major role in resisting the new coronavirus. Mitogen-activated protein kinase (MAPK) affects cell defense and apoptosis [95]. Azithromycin has been shown to control activation of the MAPK cascade, one of the molecular mechanisms involved in virus infection, thereby reducing virus replication [96].

In general, SARS-CoV-2-related proteins have been suggested as targets for drug design [97]. In particular, the proteins considered in the entire design of anti-coronavirus drugs are also related to cancer treatment and other diseases [98–100]. Indeed, potentially suitable drugs against this virus essentially affect signal transduction and the synthesis of macromolecules, which strongly interfere with the host immune response, particularly the proteins associated with COVID-19 [13].

2 Drug design strategies for COVID-19

Computational drug design approaches applied to COVID-19 can be broadly categorized as (i) structure-based approaches and (ii) AI-based approaches [101–105]. Some methods consist of both structure-based and AI-based approaches [106–109]. The drug design process is shown in Figure 1.

Figure 1.

Structure-based approaches and AI-based approaches for designing drugs. The decisive part of structure-based and AI-based approaches is the identification of the interaction between molecules and target proteins. Structure-based approaches of designing drugs rely on the three-dimensional structure of the target protein and its active site to identify the interaction between protein and molecule, while AI-based approaches rely on the knowledge of protein and molecule to understand the interaction between them through machine learning or deep learning algorithms.

2.1 Structure-based approaches

Insertion of halogen atoms on hit or lead compounds has been used to exploit their steric effects because the formation of halogen groups in ligand–target complexes favorably contributes to the stability of the protein–ligand complex [28, 110, 111]. The use of carbamates in medicinal chemistry has increased, and many derivatives are specifically designed to form drug–target interactions through their carbamate moiety [29, 112, 113]. Fragment-based drug discovery is an effective strategy for generating small-molecule protein inhibitors and drug candidates, which has led to three FDA-approved drugs and clinical trials for nearly 50 molecules [114]. To enable the screening of virtual libraries in search of active compounds, combinatorial chemistry and structure-based design need to be combined, which can exploit ligands and targets’ fundamental structural and physicochemical properties [115–117]. We believe that these approaches have been helpful for designing antiviral compounds.

2.1.1 Structure-based drug design for target CoV main proteases

The genus Coronavirus contains approximately 25 coronaviruses (CoVs), which are essential pathogens causing highly prevalent diseases [38]. By comparing four crystal structures and homologous models representing all three genetic clusters of the genus Coronavirus, it was found that the CoV main proteases (M or 3CL) are critical enzymes in viral gene expression and replication and share a highly conserved substrate recognition pocket [118, 119]. In addition, the active sites of M, S1, S1, S2 and S4 are highly conserved among all coronaviruses [38]. It is possible to design and synthesize inhibitors that target SARS-CoV-2 M by analyzing the substrate-binding pocket of SARS-CoV M (PDB ID 2H2Z). When designing a new inhibitor, Dai et al. chose an aldehyde as a new warhead to form a covalent bond with cysteine. In addition, they introduced cyclohexyl or 3-fluorophenyl in P2 to enhance the activity and introduced indole groups in P3 to form new hydrogen bonds with S4, which improves the drug-like properties [103]. UCI-1 is a cyclic peptide inhibitor that was designed based on the crystal structure of an inactive SARS-CoV M (C145A) variant. The purpose of designing UCI-1 is to mimic the conformation of a C-terminal autolytic cleavage site of the SARS-CoV M, a naturally occurring M substrate. In UCI-1, the carboxy-terminus of the P2’ residue is linked to the amino-terminus of the P2 residue with a [4-(2-aminoethyl)phenyl]-acetic acid (AEPA) group, creating a cyclophane. The (2-aminoethyl)phenyl group of AEPA is designed to act as a surrogate for a phenylalanine side chain at position P3’ and fill the S3’ pocket. Furthermore, research shows that UCI-1 tends to be nontoxic toward human embryonic kidney cells at concentrations that inhibit M, but compared with other M inhibitors, it shows lower activity [86].

Owing to the structural elucidation of the target, the application of structure-based drug design (SBDD) software is flourishing [120–122]; the basic steps of SBDD are described in detail in Figure 2. Using different complementary virtual screening and docking approaches can identify nonapproved active compounds as new potential inhibitors of 3CLpro from the ZINC15 library. Then, these compounds could be further optimized by using SBDD [123]. On the other hand, one study started with the X-ray crystal structure of SARS-CoV M [86] and then used UCSF Chimera software [124] to modify the substrate to create a cyclic peptide inhibitor within the M active site. Finally, AutoDock Vina [125] was used to evaluate this model by docking the inhibitor to SARS-CoV-2 M.

Figure 2.

The basic steps of SBDD. First, the preparation of the target macromolecule structure is fundamental to SBDD approaches, which can be obtained by structure identification techniques such as X-ray and NMR analysis, searching the PDB database or using other calculation methods. Next, the binding site of the target macromolecule is identified to determine the protein–ligand interaction, and ligands or drug-like compounds with the limitation of Lipinski’s rule of five’ are selected from a chemical database to construct a ligand library. When all preparations are ready, molecule docking programs are applied, and scoring functions are mostly used in the post-docking analysis. Finally, the top-ranked molecules are chemically synthesized, and biological evaluation is necessary.

2.1.2 SBDD for target PL

PL aids coronaviruses in their evasion of host innate immune responses because it has the additional function of stripping ubiquitin and interferon-stimulated gene 15 (ISG15) from host-cell proteins [126]. Inhibition of SARS-CoV-2 PL with GRL-0617 has three main tasks: impair the virus-induced cytopathogenic effect, maintain the antiviral interferon pathway and reduce viral replication in infected cells [127]. A recent investigation attempted to design potential SARS-CoV PL inhibitors containing naphthalene and 3,4-dihydro-2H-pyran moieties connected via-NHCO-linker [89].

Another study first used HyCoSuL, a novel chemical approach to perform comprehensive activity profiling of SARS-CoV-2 PL, and revealed the molecular rules governing PL substrate specificity [128]. Then, compared with other proteases, potent inhibitors (VIR250 and VIR251) were designed and biochemically characterized were shown to be more effective against SARS-CoV-2 PL and related SARS-CoV-1 PL, presenting high selectivity. In addition, it was surprisingly discovered that the P4 amino acids of VIR250 and VIR251 occupy both sides of the wide S4 pocket of SARS-CoV-2 PL, which will make contributions to future drug discovery [87]. It is worth noting that there is not enough information about SARS-CoV-2-PL, and HyCoSul is based on a hypothesis that SARS-CoV-2-PL is highly similar to SARS-CoV-PL. The crystal structure of the SARS-CoV-2 ADP-ribose phosphatase domain (ADRP) in multiple states was determined by many studies: in the apo form and in complexes with 2-(N-morpholino) ethane sulfonic acid (MES), ADP-ribose (ADPr) and AMP. Researchers have proposed a robust system to identify potential small-molecule inhibitors with apo crystals diffracting to atomic resolution based on structure-based experiments [90].

2.1.3 SBDD for target spike glycoprotein

A coronavirus was identified as the causative agent of SARS as early as 2003 [129]. Current research shows that the latest SARS-CoV-2 in 2019 has a very similar structure and function to SARS [130–132]. The COVID-19 genomes from isolates of China, India, Italy, Nepal and the USA have a sequence similarity of approximately 60% with the human SARS-CoV German isolate, but it has sequence similarity of 79–80% with bat SARS-CoV [133]. Spike glycoprotein (S) is crucial in the attachment of SARS-CoV-2 to host receptors and cell entry, leading to COVID-19 infection [134, 135]. An in silico pharmacophore modeling and virtual screening approach has been used to explore drugs against receptor-binding domain (RBD) of SARS-CoV-2. First, the 3D structure of RBD is modelled, the conservative area is used as a template and LigandScout is used to design the pharmacophore. Then, the Cambridge, DrugBank, ZINC and TIMBLE databases are used to screen lead compounds. Finally, AutoDock Vina Is used to dock the shortlisted lead compounds molecularly and visualize the interacting residues [77]. All the resulting lead compounds are preliminary findings, and clinical and investigational trials are still required. New research points out that the cell membrane receptor angiotensin-converting enzyme 2 (ACE2) plays a key role in the entry of SARS-CoV-2 into cells [136]. SARS-CoV-2 S interacts with ACE2 through the RBD [137]. To this end, a structure-based ACE2 variant dataset was combined with the SARS-CoV-2 RBD, resulting in a total of 242 structural models. These models can be used as a starting point for drug design and be used further to understand the recognition of SARS-CoV-2 S protein by ACE2 [138]. Research has identified a pair of key salt bridges formed by the side chains of K537 and E619. Drugs designed to prevent the formation of these salt bridges can effectively treat COVID-19, but for the simplicity of the protein complex, it is concluded to be applicable to the trimeric S protein [139].

For the different target proteins of SARS-COV-2, SBDD studies against the COVID-19 pandemic are summarized in Table 1. In addition, small peptides can be transported across the cell membrane by amino acid and peptide transporters [140]. A variety of small peptide compounds have been developed to inhibit ACE [141, 142]. Another promising method is to select and combine the substructures of highly active compounds to form new peptide analogs [143–146].

Table 1.

Structure-based approaches for drug design

Reference	Target protein	Description	Advantages	Possible designing drugs or pharmacophore
77	Spike glycoprotein	The conserved region was used as a template to design a pharmacophore using LigandScout.	Interacting residues were visualized.	6-(1H-imidazol-1-yl)-N-[1-(3,4,5-trifluorophenyl)ethyl]pyrimidin-4-amine
138	Spike glycoprotein	242 structural models of variants of human ACE2 bound to the receptor binding domain (RBD) of the S protein was built and refined their interfaces with HADDOCK.	The effects of these variations on the 3D structure of the protein and its complex with RBD have been systematic studied.	https://kastritislab.github.io/human-ace2-variants/
139	Spike glycoprotein	Mutations of K537Q and E619D reduced their side chain lengths and eliminated this pair of salt bridges.	The screened molecule is capable of blocking the formation of the key pair of salt bridges.	The side chains of K537 and E619.
103	M	The aldehyde groups of 11a and 11b are covalently bound to cysteine 145 of M.	Both compounds showed good pharmacokinetic properties in vivo, among which 11a has lower toxicity.	11a and 11b.
86	M	The cyclic peptide inhibitor was designed to mimic the conformation of a substrate at a C-terminal autolytic cleavage site of M.	The inhibitor is active against M in vitro and is nontoxic toward human cells in culture.	A first-in-class cyclic peptide inhibitor.
123	M	3CL proteases have a central role in polyprotein processing during replication.	The period up to approval as a therapeutic against SARS-CoV-2 could hopefully be shortened.	Hydrogen-bond acceptors or donors of pocket amino acids.
89	PL	Potential SARS-CoV PLpro inhibitors containing naphthalene and 3,4-dihydro-2H-pyran moieties connected via -NHCO- linker have been designed.	The designed ligands against the receptor SARS CoV-2 Papain-like protease (PL) have strong binding affinity and inhibition potential.	naphthalene based SARS-CoV PL inhibitors.
87	PL	Viral papain-like cysteine protease (PL, NSP3) represents a promising target for the development of antiviral drugs.	SARS-CoV-2-PL harbours deISGylating activities similar to SARS-CoV-1-PL, but its ability to hydrolyse K48-linked Ub chains is diminished	VIR250 and VIR251.
90	PL	Macrodomains may help to reduce the viral load and facilitate recovery.	A robust system has been developed to identify potential small-molecule inhibitors for structure-based experiments.	The anomeric C atom.

2.2 AI-based approaches

Designing novel drugs for new diseases is a complicated process necessary to find molecules that bind to specific biomolecular targets and have good physical and chemical properties over a broad chemical space [147]. First, knowledge of the virus and target is necessary. A number of methods have used machine learning or deep learning algorithms to identify the structure of the protein, but most of them could not obtain an accurate structure until AlphaFold [148] appeared. The input of AlphaFold is the given protein sequence, which can be more easily obtained than the structure through experiments. Then, it learns protein-specific potential by training a deep neural network and makes an accurate prediction of the structure by minimizing the potential by gradient descent. Another high-quality prediction of protein structure is C-I-TASSER [149]; it is an improvement based on I-TASSER [150], with three modules added on the original basis, which are a new multivariate sequence comparison protocol, an improved meta method NeBcon and an optimized contact potential.

Then, with the ever-increasing number of potential lead compounds derived through virtual screening studies, it is helpful to screen extensive virtual libraries of compounds with improved biological properties such as explicitness and selectivity toward the respective target, lower toxicity or reduced cost to discover drugs with essential biological features. It is laborious for drug designers to extract synthetic information from substantial compound libraries to devise drugs with important biological features; machine learning has become an effective tool to solve this problem [151–155]. For example, random forest analyses coupled with a unique approach to bioactivity and chemical data curation have led to a series of target-specific and cross-validated predictive feature fingerprints [156]. Moreover, in the study of reference [157], the desire to map small molecules to target interactions across multiple stages of SARS-CoV-2 infection drives the initial target selection.

Deep learning can be applied to extract complex and deeper features from simple representations [158]. For example, given a protein pocket and using deep generative modeling for compound design, potential binding compounds can be generated [159]. We know that effective noncovalent inhibitors of the major proteases of SARS-CoV and SARS-CoV-2 should have the same structural and chemical characteristics [84, 160, 161]. The ligand generative adversarial network (LIGANN) is a novel virtual screening technique that is applied for multimodal structure-based ligand design [61]. The ligand is generated in LIGANN to match the shape and chemical properties of the binding pocket and then is decoded into its SMILES sequence, thereby directly realizing AI-based drug design [73].

CogMol (controlled generation of molecules) is an end-to-end framework that was proposed to design new drug-like small molecules targeting novel viral proteins with high affinity and off-target selectivity [162]. CogMol combines the variational autoencoder in the molecular SMILES format and the multi-attribute controlled sampling scheme and applies this approach to three SARS-CoV-2 target proteins: the main protease, the RBD of the spike protein and the nonstructural protein. Compared with earlier deep-learning molecular methods, the new approach explores new molecules by adding a meaningful molecular fragment one by one rather than a single atom at a time [163, 164]. However, due to the deviation of the training data or the inaccuracy of the predictor used to control the generation, there may be a certain degree of unreliability and inability to meet the needs of generating molecules with specific properties. ADQN-FBDD [165] is another AI-based drug design approach against SARS-CoV-2. First, an initial molecular database targeting SARS-CoV-2 3CL is built by collecting SARS-CoV 3CL inhibitors (284 molecules). Then, this set of molecules is split into fragments, and the molecular weight of each fragment does not exceed 200 Daltons. Finally, the features of the fragments are input into the deep Q-learning network to generate potential lead compounds. Most of the computation-aided drug design methods need to be experimentally verified, and ADQN-FBDD is no exception. In conclusion, the major steps of AI-based drug design are described in Figure 3.

Figure 3.

The basic steps of AI-based drug design. The first step is target discovery, which means identifying the target protein that interacts with coronavirus. Next, a virtual screening technique is applied to search ligands from chemical databases, and knowledge of target protein–ligand interactions can be obtained from databases or calculated by other calculation methods. Then, AI methods, including machine learning and deep learning, such as generators and predictors, can be used to discover novel drugs. However, before machine learning or deep learning algorithms are used to generate lead molecules, the molecule fed into the learning system should be transformed into vectors, such as molecular descriptors, fingerprints, SMILES strings or grids.

3 Discussion and conclusion

Although vaccines against SARS-CoV-2 have been developed to help prevent COVID-19, there are still some potential problems. Some vaccines developed internationally continue to reveal potential risks and cannot effectively prevent the virus after vaccination. Drug relocation is also a popular way to treat COVID-19, but previous drugs often do not have a significant inhibitory effect on SARS-CoV-2. Therefore, the development of drugs against SARS-CoV-2 is more reasonable and attractive and can provide an effective first line of defense for emerging coronavirus-related drugs in the future. In recent years, technological developments have enabled the determination of more protein structures and expanded structural biology, thereby accelerating the progress of drug design.

It is estimated that the traditional pharmaceutical industry spends US$2.6 billion to design a new drug, with the failure rate reaching 90% between clinical trials and approval [169]. The emergence of structure-based and AI-based drug design has greatly reduced the cost and time and is more creative. Moreover, the key to structure-based and AI-based drug design is target–protein determination and target–drug interactions. However, SBDD relies more on the three-dimensional structure to optimize the drug compounds, which has a larger chemical space, whereas AI-based structures use machine learning and deep learning algorithms in combination with virtual screening to find the lead molecule. Finally, predicts the binding models between COVID-19 targets and designed molecules by [170].

We also found that structure-based approaches are more abundant than purely AI-based methods because SBDD methods have been developed for decades. As a result, the entire design–experiment–verification process is flawless, and the corresponding tools are also more comprehensive. We summarize the tools of drug design in Table 2. On the other hand, the contribution of AI-based methods to drug design may not achieve the expected results. The possible reason is that these methods are limited to our knowledge. Currently, we do not have enough knowledge to understand a virus that was only discovered less than 2 years ago. Therefore, basic research is essential to the analysis of research applied to biomedicine because basic research is an indispensable foundation for knowledge growth, but it is often undervalued. Nevertheless, powerful and effective calculation methods and pipelines for designing compounds can provide beneficial drug candidates for the treatment of SARS-CoV-2 infection, and these candidates or their variants are likely to produce effective anti-COVID-19 lead drugs. Therefore, we believe that these advances will help make more meaningful contributions to the fight against COVID-19 in the future.

Table 2.

The tools of drug design

Reference	Tools name & Website	Description
101	GOLEM N/A	GOLEM from the field of inductive logic programming was applied to the drug design problem of modelling structure–activity relationships.
121	Drug Guru N/A	Drug Guru (drug generation using rules) is a new web-based computer software program for medicinal chemists that applies a set of transformations, that is, rules, to an input structure.
59	ReLeaSE https://github.com/isayev/ReLeaSE	ReLeaSE is a novel computational strategy for de novo design of molecules with desired properties.
60	GuacaMol https://www.benevolent.com/guacamol	GuacalMol is a de novo design tool that seeks to generate molecules with required property profiles by virtual design-make-test cycles.
166	DeepScaffold https://github.com/deep-scaffold/deep_scaffold	DeepScaffold is proposed based on a wide spectrum of scaffold definitions, including Bemis-Murcko scaffolds, cyclic skeletons, and scaffolds with specifications on side-chain properties.
167	Cov_FB3D http://202.114.32.71:10280/wandox/Cov_FB3D/home.html	Cov_FB3D involves the in silico assembly of potential novel covalent inhibitors by identifying the active fragments in the covalently binding site of the target protein.
73	LIGANN N/A	LIGANN is developed by using a combination of virtual screening, docking and molecular dynamics techniques for noncovalent inhibition of the main protease 3CL of SARS-CoV-2.
114	FBDD https://github.com/PatWalters/fragment_expansion	FBDD has reviewed the different types of linkers published and obtained the lead compound by designing these linkers.
124	UCSF ChimeraX https://www.rbvi.ucsf.edu/chimerax	UCSF ChimeraX is the next-generation interactive visualization program from the Resource for Biocomputing, Visualization, and Informatics (RBVI), following UCSF ChimeraX.
171	PerSpect ML N/A	PerSpect ML considers 11 persistent spectral variables and feeds them into machine learning models to predict the protein–ligand affinity for drug design.
55	OpenChem https://github.com/Mariewelt/OpenChem	OpenChem offers easy and fast model development, modular software design, and several data preprocessing modules.
63	Target2DeNovoDrug https://github.com/bengeof/Target2DeNovoDrug	Target2DeNovoDrug offers researchers a fresh set of tools to approach SBDD problems by integrating data science and artificial intelligence (AI).
64	MolAICal https://molaical.github.io	MolAICal software is introduced to supply a way for generating 3D drugs in the 3D pocket of protein targets by combining the merits of a deep learning model and classical algorithm.
168	Target2DeNovoDrugPropMax https://github.com/bengeof/QPoweredTarget2DeNovoDrugPropMax	Target2DeNovoDrugPropMax includes a variety of AI methods along with incorporated in silico techniques to provide a holistic tool for automated drug discovery.

N/A: N/A means that the tool does not provide a link in their article, due to copyright or other reasons.

Key Points

• Drug design is rapidly being applied on an increasingly broader scale with the advancement of biometrics and bioinformatics. In particular, the design of anti-COVID-19 drugs is meaningful work for COIVD-19.

• Structure-based approaches use the basic structure, physicochemical properties of ligands and targets to screen virtual libraries to find active anti-COVID-19 drugs. In summary, develop corresponding design strategies of anti-COVID-19 drugs according to the characteristics of different target proteins.

• AI-based approaches can extract complex and deeper features from simple drug molecular representations and generate potential binding compounds. It is essential to extract features from compounds or proteins using interpretable methods.

5 Author contributions statement

J.W., Y.Z., Y.L. and W.N. analyze the data and wrote the manuscript. L.D. reviewed and revised the manuscript.

Jinxian Wang received the Bachelor’s degree from Hunan Agricultural University in 2019, and at present is studying for a master degree at Central South University supervised by Prof. Lei Deng. His study focuses on machine learning and bioinformatics.

Ying Zhang is a pharmacist at Department of Pharmacy, Heilongjiang Province land reclamation headquarters general hospital, Harbin, China. His study focuses on drug development.

Wenjuan Nie is a graduate student at School of Computer, Central South University, Changsha, China. Her research interests is using machine learning algorithms to predict protein–NA binding residues.

Yi Luo received BE degree in Computer Science from School of Computer Science and Technology, South West University, ChongQing,China. Currently,he is a MS candidate in Data Science from School of Science.the University of Auckland,Auckland, New Zealand. His research interests include data mining, bioinformatics and data analysis.

Lei Deng is a professor at School of Computer Science and Engineering, Central South University, Changsha, China. His research interests include data mining, bioinformatics and systems biology.

Funding

National Natural Science Foundation of China (grant No. 61972422).