Identification of potential binders of the main protease 3CL^pro of the COVID-19 via structure-based ligand design and molecular modeling

Graphical abstract

Abstract

We have applied a computational strategy, using a combination of virtual screening, docking and molecular dynamics techniques, aimed at identifying possible lead compounds for the non-covalent inhibition of the main protease 3CL^pro of the SARS-CoV2 Coronavirus. Based on the X-ray structure (PDB code: 6LU7), ligands were generated using a multimodal structure-based design and then docked to the monomer in the active state. Docking calculations show that ligand-binding is strikingly similar in SARS-CoV and SARS-CoV2 main proteases. The most potent docked ligands are found to share a common binding pattern with aromatic moieties connected by rotatable bonds in a pseudo-linear arrangement.

At the beginning of this year, the world was dismayed by the outbreak of a severe viral acute respiratory syndrome (SARS), known as COVID-19, that rapidly spreads from its origin in the Hubei Chinese district to virtually whole China and, as of today, to more than 200 countries and territories around the world [1]. The new coronavirus, named SARS-CoV2 and believed to have a zoonotic origin, has infected thus far about 2450000 people worldwide with thousands in critical conditions, causing the death of 166000 people. The SARS-CoV2’s genome [2], [3] has a large identity [4] with that of the SARS-CoV whose epidemic started in early in 2003 and ended in the summer of the same year.

Most of the Coronaviridae genome encodes two polyproteins, pp1a and, through ribosomal frameshifting during translation [5], pp1ab. These polyproteins are cleaved and transformed in mature non-structural proteins (NSPs) by the two proteases 3CL^pro (3C-like protease) and PL^pro (Papain Like Protease) encoded by the open reading frame 1 [6]. NSPs, in turn, play a fundamental role in the transcription/replication during the infection [5]. Targeting these proteases may hence constitute a valid approach for antiviral drug design. The catalytically active 3CL^pro is a dimer. Cleavage by 3CL^pro occurs at the glutamine residue in the P1 position of the substrate via the protease CYS-HIS dyad in which the cysteine thiol functions as the nucleophile in the proteolytic process [7]. While dimerization is believed to provide a substrate-binding cleft between the two monomers [8], in the dimer the solvent-exposed CYS-HYS dyads are symmetrically located at the opposite edges the cleft, probably acting independently [9]. As no host-cell proteases are currently known with this specificity, early drug discovery was directed towards the so-called covalent Michael inhibitors [10], via electrophilic attack to the cysteinyl residue. On the other hand, the consensus in drug discovery leads to excluding electrophiles from drug candidates for reasons relating to safety and adverse effects such as allergies, tissue destruction, or carcinogenesis [11].

In spite of the initial effort in developing compounds with anti-coronavirus activity following the SARS outbreak [12], no anti-viral drug was ever approved or even reached the clinical stage due to a sharp decline in funding of coronavirus research after 2005–2006, based on the erroneous conviction that chance of a repetition of a new zoonotic transmission was extremely unlikely [6], The most potent non-covalent inhibitor for 3CL^pro, ML188, was reported nearly ten years ago [13] with moderate activity in the low micromolar range [14].

According to the latest report of the structure of 3CL^pro from SARS-CoV2 [15] (PDB code 6LU7) and the available structure of 3CL^pro from SARS-CoV [12], (PDB code 1UK4), the two proteases differ by only 12 amino acids, with $\alpha$ carbon atoms all lying at least 1 nm away from the 3CL^pro active site (see Fig. 1 a). The substrate-binding pockets of two coronavirus main proteases are compared in Fig. 1b, exhibiting a strikingly high level of alignment (RMSD = 0.99 Å) of the key residues involved in substrate binding, including the CYS145$\cdots$HIS41 dyad, and THR45, MET49, PHE140, ASN142, ASP187, ARG188, GLN189, MET165, HIS172, GLU166. The latter are believed to provide the opening gate for the substrate in the active state [12].

Fig. 1.

(a) SARS-CoV2 (orange) and SARS-CoV (green) main proteases. Violet spheres correspond to the alpha carbons of the 12 differing residues. Grey spheres indicate the CYS-HIS dyad (b) binding pocket with the main residues in bond representation (green and red for SARS-CoV2 and SARS-CoV, respectively). The shaded region mark the binding site for the substrate. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 1(a,b) strongly suggest that effective non-covalent inhibitors for SARS-CoV and SARS-CoV2 main proteases should share the same structural and chemical features. In order to investigate this matter, we have performed a molecular modeling study on both the 6LU7 and 1UK4 structures. 6LU7 is the monomer of the main protease in the active state with the N3 peptidomimetic inhibitor [15] while 1UK4 is the dimer with the protomer chain A in the active state [12]. The main protease monomer contains three domains. Domains I and II (residues 8-101 and 102-184) are made of antiparallel $\beta$-barrel structures in a chymotrypsin-like fold responsible for catalysis [16].

The 6LU7 structure was first fed to the PlayMolecule web application [17] using a novel virtual screening technique for the multimodal structure-based ligand design [18], called Ligand Generative Adversarial Network (LIGANN). Ligands in LIGANN are generated so as to match the shape and chemical attributes of the binding pocket and decoded into a sequence of SMILES enabling directly the structure-based de novo drug design. SMILES codes for ligands were obtained using the default LIGANN values for shapes and channels with the cubic box center set at the midpoint vector connecting the SH and NE atoms of the CYS-HIS dyad in the 6LU7 structure. The PlayMolecule interface delivered 93 optimally fit non-congeneric compounds, spanning a significant portion of the chemical space, whose SMILES and structures are reported in the Supporting Information (SI). Each of these compounds was docked to the 6LU7 and to the 1UK4 structures, using Autodock4 [19] with full ligand flexibility. For both structures, the docking was repeated by setting the dyad with the residue in their neutral (CYS-HIS) and charged state (CYS⁻/HIS⁺). Details on Docking parameters are given in the SI.

Results for the binding free energies of the 93 3CL^pro ligands are reported in Fig. 2 . Binding free energies are comprised in the range 4–9 kcal/mol and are found to be strongly correlated for the two protonation states of the CYS-HIS dyad. Correlation is still high when ligand binding free energies for the main proteases are compared, confirming that good binders for SARS-CoV are, in general, also good binders for SARS-CoV2 3CL^pro. For each of these compounds, using the XLOGP3 methodology [20], we computed the octanol/water partition coefficient (LogP) to assess the distribution in hydrophobic and cytosolic environments. LogP values range from −0.5 to 5 with a number of rotatable bonds from 2 to 12. Most of the LIGANN compounds bear from 2 to 5 H-bond acceptor or donors (Table S1 of the SI). In Fig. 3 , we show the probability distributions for $\mathrm{\Delta }G$ correlated in turn to the LogP, number of H-bond donor/acceptors and number of rotatable bonds. We note, on the left and central panel, sharp maxima for $\mathrm{LogP}=3:4,\mathrm{\Delta }G=-7:-8$ and for $\mathrm{H}-\mathrm{acc}/\mathrm{don}=3,\mathrm{\Delta }G=-6:-7$, respectively, suggestive of a ligand-protein association driven mostly by hydrophobic interactions. We must stress that the computed $\mathrm{\Delta }G$ pertains to the associations of the ligand with one protein whatever the state of association of the protein. At free ligand concentration equal to $K_d\equiv e^{-\mathrm{\Delta }G/\mathit{RT}}$, i.e. when half of the protein molecules are inhibited, the probability to have both monomers inhibited is equal to 1/4, whatever the dissociation constant of the dimer [21], hence the need for identifying nanomolar or subnanomolar inhibitors of 3CL^pro.

Fig. 2.

Correlation diagrams of autodock-computed binding free energies for 93 ligands of the SARS-CoV and SARS-CoV2 3CL^pro structures. $R,\mathrm{mue},\tau$ indicate the Pearson correlation coefficient, the mean unsigned error, and the Kendall rank coefficient, respectively. Upper panel: correlation diagram between ligand free energies obtained with the charged CYS⁻¹-HIS⁺ and with neutral CYS-HIS dyad. Lower panel: correlation diagram between ligand free energies of SARS-CoV2 and SARS-CoV. Larger symbols refer to ML188.

Fig. 3.

2D probability histograms $\mathrm{\Delta }G$ with LogP (left), H-bond acceptors or donors (center) and rotatable bonds (right) for the 93 compounds of Table S1 of the SI. The common color-coded z-scale on the right corresponds to the 2D probability. Results for SARS-CoV2 and SARS-CoV in upper and lower panels, respectively.

Table 1 shows the chemical structures of the five compounds exhibiting the highest binding affinity to the main protease of SARS-CoV2 when the CYS-HIS dyad is in the neutral state. None of these compounds is commercially available, although some of them (27, 31, 40) show a high degree of similarity with known structures according to the Tanimoto metrics.[22] The structures of Table 1, as well as many of those reported in Figs. S1-S5 of the SI, seem to share a common pattern with aromatic moieties connected by rotatable bonds in a pseudo-linear arrangement. Table 1 shows the binding free energy data of these five best ligands for both CoV proteases and both protonation states of the dyad.

Table 1.

Computed binding free energies (kcal/mol), $\mathrm{\Delta }G$, of the best five binders for SARS-CoV2 3CL^pro. $\mathrm{\Delta }G$ values are reported for the two protonation states of the dyad and for SARS-CoV and SARS-CoV2 main protease. Below the 2D structures and ${\text{k}}_d$ values.

	CoV19		SARS
Comp.	H-CYS	H-HIS	H-CYS	H-HIS	LogP
27	−8.92	−8.79	−8.92	−9.46	4.90
30	−8.84	−8.19	−7.47	−7.86	3.74
39	−8.25	−7.08	−6.82	−6.72	6.06
77	-8.17	−7.25	−7.43	−7.21	2.03
19	−8.03	−8.26	−7.01	−7.12	5.58
ML188	−7.96	−7.63	−6.46	−6.22a	4.97

^aExperimental value for ML188 is[14]$\mathrm{\Delta }G=-7.98$ kcal/mol.

Inspection of Table 1 confirms that SARS-CoV2 best binders 27, 29, 39, 77, 19 are also good binders for SARS-CoV 3CL^pro. Remarkably, compound 27 is consistently the most potent ligand for the two proteases, irrespective of the dyad protonation state. In the Table 1 we also report the Autodock4-computed binding free energy for ML188 (−6.2 and −6.5 kcal/mol for the H-HIS and H-CYS tautomers), not too distant indeed from the experimentally determined value of −8 kcal/mol, lending support for the LIGANN-Autodock4 protocol used in identifying the lead compounds of Table 1.

In order to assess the stability of the 3CL^pro-27 association, we have performed extensive molecular dynamics simulations[23], [24] of the bound state with explicit solvent. The overall structural information was obtained by combining data from three independent simulations (for a total of about 120 ns), all started from the best docking pose of 27 on the 6LU7 monomeric structure. Further methodological aspects [25] are provided in the SI. In Fig. 4 , we show the probability distribution, $P(R)$, of the distance R between the center of mass (CoM) of the ligand and that of the domains I + II. The distribution has a Gaussian shape with a half-width of about 1 Å, exhibiting only a minor positive skewness and defining a tight binding site volume $V_{\mathrm{site}}$ of few ų at most.[26] The MD-determined $P(R)$ shows that the ligand never leaves the binding pocket at any stage during the whole simulation. In the inset of Fig. 4a, we show the potential of mean force (PMF) along the ligand-protein CoM distance R, computed as $v(R)=-\mathit{RT}\log (P(R)/\max [P(r)])$. As $1/K_d$ = ${\int }_{V_{\mathrm{site}}}{e}^{-\beta v(R)}d\mathbf{R}$,[27], the steepness of the curve is suggestive of a profound minimum and hence of a large association constant, confirming the indication obtained from the Docking calculations. Fig. 4b shows polar and hydrophobic residues found in contact with the ligand 27. All essential residues for binding are included, with the addition of MET165, PHE140 and LEU141 hydrophobic residues, consistently lingering near the pyrazolic or the chlorinated phenyl rings of 27, in agreement with the hydrophobic character of the interaction.

Fig. 4.

(a) Probability distribution of the distance between the CoM of compound 27 and domain I + II of $3{\text{CL}}^{\mathrm{PRO}}$, as obtained from MD simulations (inset: the corresponding PMF). (b) Binding pocket of 6LU7 with ligand 27. The time record of the minimal distances between ligand and the depicted nearby residues are reported in Figs. S6, S7 of the SI.

Fig. 3, Fig. 4 show possible avenues for improvement. For example, forcing the L-shaped binding structure in bulk also, by redesigning the rotatable connectors in the ligand, may reduce the penalty due conformational entropy loss upon binding,[26] hence boosting the ligand affinity for 3CL^pro. Building upon this knowledge, we plan to optimize the lead using MD simulations coupled to relative binding free energy calculation on congeneric variants [28], eventually providing in silico determined anti-viral compounds to be synthesized an experimentally tested in vitro and in vivo.

The infection rate for COVID-19 in China is currently declining for days. As the road for delivering an effective anti-viral drug is still a long one indeed, the SARS-CoV2 harsh lesson, nonetheless, should not be forgotten once the emergency will end, hoping that our contribution can pave the way for the design of effective non-covalent antiviral drugs for the present and future Coronavirus emergencies.

CRediT authorship contribution statement

Marina Macchiagodena: Data curation, Investigation, Writing-review & editing. Marco Pagliai: Investigation, Writing-review & editing. Piero Procacci: Conceptualization, Investigation, Writing- review & editing.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Supplementary material

The following are the Supplementary data to this article: