The beginning of third decade in the 21st century was marred by the spread of the COVID-19 pandemic. Patients with pneumonia-like symptoms were initially reported in Wuhan, China as early as 8 December 2019 with a rapid, global spread, leading to over 21 million cases and 761,779 deaths reported by the World Health Organization (WHO) in August 2020 (Muralidar et al., 2020), with the WHO reporting global mortality rates of about 6.5 million by October 2022.

COVID-19 is caused by infection of The Severe Acute Respiratory Syndrome-Coronavirus-2 (SARS-CoV-2), which belongs to the family of coronaviruses and is highly homologous to SARS-CoV (Zhou et al., 2020). Infection normally begins with transmission through aerosol droplets followed by a latent period and respiratory tract infection (Salzberger et al., 2021). The majority of infections (∼90%) do not require hospitalization and approximately 40% are asymptomatic (Salzberger et al., 2021). Fatality rates increase with the presence of risk factors, with Case Fatality Rates (CFR) reaching double digit percentages (10%–20%) in patients over 70 years of age (Salzberger et al., 2021). High CFR rates and widespread infection rates have led to control measures such as lockdowns, which proved effective (Salzberger et al., 2021), albeit with a social and economic price tag. The understanding that in addition to the acute disease, tissue damage, autoimmunity and chronic inflammation may result in Long COVID (Yong, 2021), has increased the urgency for drug discovery.

The SARS-CoV-2 virion is made up of four structural proteins: nucleocapsid N), membrane M), envelope E) and spike S). Infection begins with the viral spike protein binding the human angiotensin-converting enzyme 2 (ACE2) receptor (Zhou et al., 2020; Jackson et al., 2022), the same receptor that mediates SARS-CoV infection (Li et al., 2003). The spike protein is a homotrimer in which each monomer is composed of a viral membrane binding S2 domain, and the S1 domain, containing the Receptor Binding Domain (RBD) which binds ACE2 and following furin-mediated cleavage, initiates cell fusion towards infection (Huang et al., 2020). Cryo-electron microscopy has demonstrated that spike can exist with all 3 RBD domains exposed, binding ACE2 receptors (PDB: 7A98), spike with 2 unbound RBDs erect (PDB: 7A93), spike with one RBD exposed, binding one ACE2 receptor (PDB: 7A94) and spike with 2 RBDs binding 2 ACE2 receptors (PDB: 7A97) (Benton et al., 2020). The diversity of the spike/ACE2 complex models raises the questions of biological relevance and implications to inhibitor discovery. In addition to the ACE2 binding conformations, there is also a furin-cleaved closed conformation, with no exposed RBDs (PDB: 6ZGI) (Wrobel et al., 2020). Additionally, it has also been shown that an infection pathway independent of ACE2 exists, in which spike binds activated α5β1 integrins, a process suggested to be more significant following initial ACE2 infection and resulting cytokine mediated integrin activation (Liu et al., 2022). Of the current therapeutic approaches only monoclonal antibodies and Convalescent Plasma Therapy target spike (Raman et al., 2021). These treatments are costly and need to be administered professionally, limiting their application in a future major outbreak, making additional drug discovery essential. The combination of a diverse viral spike protein, capable of infection via two independent pathways makes the discovery of viral fusion inhibitors a challenging task.

Peptides composed of 20 natural amino-acid monomers often suffer poor absorption, low stability and fast clearance that often limits their progress as therapeutic agents, to partially overcome these limitations, “non-natural” amino-acids and backbone modifications have been introduced to form peptidomimetics (Li Petri et al., 2022). The addition of monomers greatly increases the chemical space that needs to be covered by a discovery platform. The polymer chemical space can be described using the equation below: C h e m i c a l   S p a c e = M N Where M is the number of monomers available and N is the number of monomers composing the polymer. Non-natural amino-acid monomers are constantly being discovered, increasing M from 20 to over 500. So, for example, a polymer of five amino-acids has gone from a chemical space of approximately 10⁶ natural peptides to over 10¹³ peptidomimetics! Larger polymers, such as 15-mer macrocycles reach chemical spaces that exceed 10⁴⁰. Naturally, high-throughput screening systems that deploy robotic physical screening of 10⁵-10⁶ molecules cover too small a fraction of the chemical space to be efficient in peptidomimetic discovery. Even when limited to the natural monomers, the large peptide chemical space was not well explored before phage display libraries (Smith, 1985). Display technologies, have utilized tRNA-monomer attachment to increase chemical space (Kourouklis et al., 2005; Taki et al., 2006; Ohuchi et al., 2007), with some constraint/cyclization capacity (Cary et al., 2017). Nevertheless, physical techniques have a limited capacity to fully explore the chemical space, requiring more flexible screening methods.

Pepticom Ltd. was founded with the vision of allowing in silico peptide, peptidomimetic and macrocycle screening. As the enormous polymer chemical space is too large for explicit screening using standard “docking” software, even in the age of cloud computing, Pepticom took an Artificial Intelligence (AI) approach where reinforcement learning guides cloud based docking algorithms. In this methodology, docking is performed by agents or “bots” that upload to the cloud with a set of docking algorithms, heuristic, grid and sequence preferences. Bots may explore an innovative solution or alternately exploit an existing one. Once bots complete the docking task, they update the system with their findings and get a reward on the quality and originality of their solution. To further enhance this process, the concept of risk/return efficiency was translated from the financial world, making the exploitation of existing docked structures for additional sequences more efficient (Lerner et al., 2017). This screening approach has yielded numerous novel polymers; including GABA-A receptor allosteric modulators (Michaeli et al., 2020), bi-specific MD2/CD14 macrocycle agonists of Toll-Like Receptor 4 (TLR4) (Michaeli et al., 2018) and a rescue peptide for a destabilizing Glycogen Branching Enzyme (GBE1) mutation (Froese et al., 2015). The latter case showed another advantage for in silico screening; as protein rescue assays had to be conducted in vivo, on primary patient cells, barring the option for massive physical screening (Froese et al., 2015).

We used spike and spike/ACE2 protein structures to virtually screen macrocycle peptidomimetics with the potential to inhibit SARS-CoV-2 viral entry.

Virtual screening on different conformations was performed using Pepticom software. Each screening run focused on a specific area of a solved spike structure. The structures and docking regions were selected according to three strategies:

1) Allosteric inhibition of the spike trimer that keeps spike in the closed conformation, preventing RBD exposure.

For each virtual screening, a docking area was selected, A.I. screening was performed and the results were clustered into frequently exploited, high scoring conformational clusters. A set of high scoring, diverse, members of each cluster was selected for synthesis and physical screening. The number of molecules per set was determined according to score distribution and diversity.

In this study we performed virtual screening of small (5–15 amino-acid), N to C cyclic, peptidomimetics for three different inhibition strategies for the spike/ACE2 complex: allosteric inhibition of the spike trimer that keep spike in the closed conformation; direct ACE2 inhibition; and direct spike RBD inhibition. Output peptidomimetics for each strategy were clustered and representatives of each cluster were selected for synthesis and screening. Biochemical screening was performed using a spike RBD/ACE2 AlphaLISA assay and functional screening was performed using the SARS-CoV-2 RVP.

Of the three strategies, allosteric inhibition had the most attractive stoichiometry. Whereas, in the RBD competitive binding strategy, complete inhibition requires all spike RBDs to be bound, allosteric inhibitors can theoretically lock all the spike RBDs with one molecule. The ACE2 strategy is challenged by the broad tissue expression of ACE2. The critical morbidity factor in SARS-CoV-2 infection arises from the lungs, ACE2 is however, broadly expressed, in tissues including the small intestine, colon, gallbladder and others (Hikmet et al., 2020). Systematic administration of molecules that bind ACE2, will also bind ACE2 at sites remote to viral infection, requiring a high dosage. The allosteric strategy can theoretically deploy one peptidomimetic to constrict all 3 RBDs to the closed state, making it attractive. Another theoretical advantage of this approach, is the non-competitive mode of action, allowing a relatively small molecule to inhibit a high affinity protein-protein complex.

Of the 12 peptides representing 4 clusters in this effort, we reached one inhibitor that showed specific SARS-CoV-2 RVP inhibition. We believe that the partial inhibition was caused by partial RBD escape from inhibition that is most likely caused by lack of shape complementarity between the three-pointed star shape of the pocket and the relatively flat small macrocycle conformation (Figure 1A). Other options, such as branching new chains out of macrocycle side-chains are currently being explored to address this issue and divide the modeled binding energy somewhat more equally between the monomers. Another possible cause for potential RBD escape is the lack of symmetry between the calculated binding free-energy between the RBD petals. So, for example, in peptide 2 of cluster 4, the aspartate at position 405 of the closed structure (PDB: 6VXX) is met by a macrocycle salt bridge on chain C, but not chains A or B, while sterically preventing the access of another identical peptide to these positions. An alternative solution to this issue could be screening for molecules that bind the corners in which 2 petaLs meet. Such potential solutions could provide a better free energy distribution, albeit, requiring 3 peptides to inhibit spike/ACE2 complex formation.

The human receptor ACE2, spike binding region is rather convex in relation to a typical 5–15 AA macrocycle (Figure 2A), when compared to the corresponding spike region (Figure 3A). As small macrocycles are rather rigid, they are less likely to bind a convex region. The ACE2 convex geometry is further complexed by more concave areas that are near the spike binding interface, these cause the virtual screening agents to “runoff” away from areas where they can have complete efficacy. Our efforts yielded one cluster of weak inhibitors, with similar IC₅₀ values in AlphaLISA and cellular RVP assay, suggesting that our cell assay lacked integrin mediated penetration. Interestingly, the small, linear, fibronectin derived, non-competitive α5β1 integrin binding peptide, ATN-161 can inhibit complex formation and decrease infection at the cellular level (Beddingfield et al., 2021). In an in silico screening effort followed by lab validation, Merck’s antifungal macrocycle, Caspofungin, was also predicted to bind the ACE2 spike binding site at a low-level micromolar activity, but showed no lab screening activity, whereas the non-peptide bases macrocycle Zotarolimus, showed a low micromolar Kd and inhibitory activity (Day et al., 2021). Small molecules of the same screening effort such as the amino acid Levodopa showed an ACE2 Kd in the double digit nanomolar range, with over 36% competition with spike-RBD at 1 μM (Day et al., 2021). These published results concur with our findings that macrocycles are less than optimal for ACE2 targeted SARS-CoV-2 viral entry inhibitors.

Targeting the spike RBD in its ACE2 binding region yielded more clusters at the computational stage than in the corresponding ACE2 region, with RBD clusters 4 and 5 showing several molecules with better docking scores than the top scoring molecules o the ACE2 cluster (Tables 2, 3), suggesting it is a more plausible macrocycle target. Of the five selected clusters, two clusters, cluster 4 and 5, showed molecules with single digit micromolar inhibition constants in both AlphaLISA and SARS-CoV-2 RVP. These molecules initiated our ongoing effort to reach lead stage molecules.

Computationally derived, 56 amino acid RBD binding miniproteins, based on ACE2 helices, yielded inhibition constants reaching the picomolar range with proven live virus efficacy (Cao et al., 2020). Nevertheless, in our opinion, miniproteins do not offer distinct pharmaceutical differentiation from antibodies, some of which are in clinical use (Raman et al., 2021), or their derived nanobodies which show similar potency (Xiang et al., 2020). Hence, achieving smaller, non-natural molecules, could produce more novel drug candidates. A discovery effort for small peptide inhibitor, based on the ACE2 interface helix yielded a 13 amino acid helical peptide with ∼40% inhibition at 100 µM (Rajpoot et al., 2021). Like macrocycles, helical peptides “suffer” from geometric constraints that limit their flexibility in interacting with the target protein, albeit with lower loss of entropy, as was shown using molecular dynamics (Rajpoot et al., 2021). Small molecules screened against the spike/ACE2 complex also showed a proof of concept that drug candidates without a tertiary structure are possible (Bojadzic et al., 2021; Day et al., 2021). The physical screening of molecules in the organic dye chemical space yielded several low micromolar inhibitors, shown to bind the spike RBD domain and provide similar simulated pseudo viral entry blockage as cluster 4 and 5 RBD targeting macrocycles (Bojadzic et al., 2021). Interestingly, Lifitegrast was shown to bind spike RBD at ∼ 2 nM, and inhibit complex formation, with a millimolar viral infection IC₅₀ (Day et al., 2021). The Evans Blue dye was also shown to bind both ACE2 (Kd ∼2 nM) and spike RBD (Kd ∼2 μM), with a viral inhibition IC50 of ∼28 μM. With the exception of Evans Blue these small molecules all share a relatively poor abundance of hydrogen bond donors and acceptors and that can possibly lead to multiple non-specific interactions when assayed in biological assays. We feel that the data presented here, along with published efforts suggest that spike RBD can be useful drug target, in which a compromise between size, target specific binding and efficacy can produce innovative drug candidates.