Introduction

Infections by the severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) resulted in approximately 1.8 million deaths in 2020 (WHO 2021) and led to the coronavirus 2019 (COVID-19) pandemic (Kucharski et al. 2020; Wu et al. 2020; Zhu et al. 2019). SARS-CoV-2 is a zoonotic betacoronavirus highly similar to SARS-CoV and MERS-CoV, which caused outbreaks in 2002 and 2012, respectively (Bermingham et al. 2012; Kuiken et al. 2003; Zaki et al. 2012). SARS-CoV-2 encodes its proteome in a single, positive-sense, linear RNA molecule of ~ 30 kb length, the majority of which (~ 21.5 kb) is translated into two polypeptides, pp1a and pp1ab, via ribosomal frame-shifting (Thiel et al. 2003; Bredenbeek et al. 1990). Key viral enzymes and factors, including the reverse-transcriptase machinery, inhibitors of host translation and molecules signalling for host cell survival, are released from pp1a and pp1ab via post-translational cleavage by two viral cysteine proteases (Hilgenfeld 2014). These proteases, a papain-like enzyme cleaving pp1ab at three sites, and a 3C-like protease cleaving the polypeptide at 11 sites, are primary targets for the development of antiviral drugs.

The 3C-like protease of SARS-CoV-2, also known as the viral main protease (Mpro), has been the target of intense study owing to its centrality in viral replication. Mpro studies have benefited from previous structural analyses of the SARC-CoV 3C-like protease and the earlier development of putative inhibitors (Ghosh et al. 2007; Verschueren et al. 2008; Yang et al. 2003,2005). The active sites of these proteases are highly conserved, and peptidomimetic inhibitors active against Mpro are also potent against the SARS-CoV 3C-like protease (Zhang et al. 2020; Rut et al. 2020). However, to date no Mpro-targeting inhibitors have been validated in clinical trials. In order to accelerate Mpro inhibitor development, an international, crowd-funded, open-science project was formed under the banner of COVID Moonshot (Achdout et al. 2020), combining high-throughput crystallographic screening (Douangamath et al. 2020), computational chemistry, enzymatic activity assays and mass spectrometry (El-Baba et al. 2020) among the many methodologies contributed by collaborating groups.

As part of COVID Moonshot, we utilised saturation transfer difference nuclear magnetic resonance (STD-NMR) spectroscopy (Mayer and Meyer 1999; Becker et al. 2018; Walpole et al. 2019) to investigate the Mpro binding of ligands initially identified by crystallographic screening, as well as molecules designed specifically as non-covalent inhibitors of this protease. Our goal was to provide orthogonal information on ligand binding to that which could be gained by enzymatic activity assays conducted in parallel by other groups. STD-NMR is a proven method for characterising the binding of small molecules to biological macromolecules, able to provide both quantitative affinity information and structural data on the proximity of ligand chemical groups to the protein. Here, we provide detailed documentation on the NMR protocols used to record these data and highlight the advantages, limitations and assumptions underpinning our approach. Our aim is to assist the comparison of Mpro STD-NMR data with other quantitative measurements, and facilitate the consideration of these data when designing future Mpro inhibitors.

Materials and methods

Protein production and purification

We created a SARS-CoV-2 Mpro genetic construct in pFLOAT vector (Rogala et al. 2015), encoding for the viral protease and an N-terminal His6-tag separated by a modified human rhinovirus (HRV) 3C protease recognition site, designed to reconstitute a native Mpro N-terminus upon HRV 3C cleavage. The Mpro construct was transformed into Escherichia coli strain Rosetta(DE3) (Novagen) and transformed clones were pre-cultured at 37 °C for 5 h in lysogeny broth supplemented with appropriate antibiotics. Starter cultures were used to inoculate Terrific Broth Autoinduction Media (Formedium) supplemented with 10% v/v glycerol and appropriate antibiotics. Cell cultures were grown at 37 °C for 5 h and then cooled to 18 °C for 12 h. For 15 N isotopically enriched protein production transformed E. coli clones were grown overnight at 37 °C in 200 mL M9 minimal media starter cultures supplemented with antibiotics and 15 N NH4Cl. These cultures were then used to inoculate 4–8 L of similarly supplemented M9 minimal media cultures. Cells were grown at 37 °C until OD600 of ~ 0.6, at which point protein expression was induced by addition of 0.25 mM isopropyl β-d-1-thiogalactopyranoside and was allowed to proceed for 12 h at 18 °C. Bacterial cells were harvested by centrifugation at 5000×g for 15 min.

Cell pellets were resuspended in 50 mM trisaminomethane (Tris)-Cl pH 8, 300 mM NaCl, 10 mM imidazole buffer, incubated with 0.05 mg/ml benzonase nuclease (Sigma Aldrich) and lysed by sonication on ice. Lysates were clarified by centrifugation at 50,000×g at 4 °C for 1 h. Lysate supernatants were loaded onto a HiTrap Talon metal affinity column (GE Healthcare) pre-equilibrated with lysis buffer. Column wash was performed with 50 mM Tris–Cl pH 8, 300 mM NaCl and 25 mM imidazole, followed by protein elution using the same buffer and an imidazole gradient from 25 to 500 mM concentration. The His6-tag was cleaved using home-made His6-tagged HRV 3C protease. The HRV 3C protease and the cleaved tag were removed by reverse metal affinity using a HiTrap Talon column. Flow-through fractions were concentrated and applied to a Superdex75 26/600 size exclusion column (GE Healthcare) equilibrated in NMR buffer (150 mM NaCl, 20 mM Na2HPO4 pH 7.4).

Nuclear magnetic resonance (NMR) spectroscopy

All NMR experiments were performed using a 950 MHz solution-state instrument comprising an Oxford Instruments superconducting magnet, Bruker Avance III console and TCI probehead. A Bruker SampleJet sample changer was used for sample manipulation. Experiments were performed using TopSpin (Bruker). For direct STD-NMR measurements, samples comprised 10 μM Mpro and variable concentrations (20 μM–4 mM) of ligand compounds formulated in NMR buffer supplemented with 10% v/v D2O and deuterated dimethyl sulfoxide (d6-DMSO, 99.96% D, Sigma Aldrich) to 5% v/v final d6-DMSO concentration. In competition experiments, samples comprised 2 μM Mpro, 0.8 mM of ligand  x0434 and variable concentrations (0–20 μM) of competing compound in NMR buffer supplemented with D2O and d6-DMSO as above. Sample volume was 140 μL and samples were loaded in 3 mm outer diameter SampleJet NMR tubes (Bruker) placed in 96-tube racks. NMR tubes were sealed with POM balls. For heteronuclear 2D spectra samples were formulated in 310 μL final volume in NMR buffer supplemented with D2O and d6-DMSO as above, 25 μM 15N-enriched Mpro unless otherwise indicated, and either 50 μM ligand or no ligand present. 15N Mpro samples were placed in 5 mm outer diameter advanced NMR microtubes (Shigemi) matched to D2O.

STD-NMR experiments were performed at 10 °C using a pulse sequence described previously (Mayer and Meyer 1999) and an excitation sculpting water-suppression scheme (Hwang and Shaka 1995). Protein signals were suppressed in STD-NMR by the application of a 30 ms spin-lock pulse. We collected time-domain data of 16,384 complex points and 41.6 μsec dwell time (12.02 kHz sweepwidth). Data were collected in an interleaved pattern, with on- and off-resonance irradiation data separated into 16 blocks of 16 transients each (256 total transients per irradiation frequency). Transient recycle delay was 4 s and on- or off-resonance irradiation was performed using 0.1 mW of power for 3.5 s at 0.5 ppm or 26 ppm, respectively, for a total experiment time of approximately 50 min. Data were processed using TopSpin (Bruker). Reconstructed time-domain data from the difference of on- and off-resonance irradiation (STD spectra) or only the off-resonance irradiation (reference spectra) were processed by applying a 2 Hz exponential line broadening function and twofold zero-filling prior to Fourier transformation. Phasing parameters were derived for each sample from the reference spectra and copied to the STD spectra. 1H peak intensities were integrated in TopSpin using a local-baseline adjustment function. Data fitting to extract Kd values were performed in OriginPro (OriginLab). The folded state of Mpro in the presence of each ligand was verified by collecting 1H NMR spectra similar to Fig. 1a from all samples ahead of STD-NMR experiments.

Fig. 1
figure 1

1D and STD-NMR spectra of SARS-CoV-2 Mpro. a Methyl regions from 1H NMR spectra of recombinant SARS-CoV-2 Mpro. The spectrum on the left was recorded from a 10 μM protein concentration sample in a 5 mm NMR tube at 25 °C using an excitation sculpting water-suppression method (Hwang and Shaka 1995). 512 acquisitions with recycle delay of 1.25 s were averaged, for a total experiment time of just over 10 min. The spectrum on the right was recorded from a 10 μM Mpro sample in a 3 mm NMR tube at 10 °C, using the same pulse sequence and acquisition parameters. For both spectra, data were processed with a quadratic sine function prior to Fourier transformation. Protein resonances are weaker in the 10 °C spectrum due to lower temperature and the reduced amount of sample used for acquisition in the smaller NMR tube. The position where on-resonance irradiation was applied for STD spectra is indicated. b Vertically offset 1H STD-NMR spectra from ligand  x0434 binding to Mpro. The reference spectrum is in black with the  x0434, H2O and DMSO 1H resonances indicated. The STD spectrum of  x0434 in the presence of Mpro is shown in red while that in the absence of Mpro is in green. STD spectra are scaled up 64 × compared to the reference spectrum. Bottom panels correspond to magnified views of the indicated spectral regions, with  x0434 resonances assigned to chemical groups of that ligand as shown

Full size image

Heteronuclear 2D 1H-15N spectra of Mpro were recorded at 25 °C using a SOFAST-HMQC pulse sequence from the biomolecular NMR group of the Institut de Biologie Structural in Grenoble (Favier and Brutscher 2019). We collected time-domain data of 756 complex points and 32.9 μsec dwell time (15.18 kHz sweepwidth) in the 1H dimension, and 64 complex points and 346.4 ms incrementation (2.89 kHz sweepwidth) in the indirect 15N dimension. Recycle delay was set to 0.2 s and 750 transients were collected per indirect increment, for a total experiment time of 7.4 h. Data were processed using NMRPipe (Delaglio et al. 1995) by applying a quadratic sine-function and tenfold zero filling along each spectra dimension prior to Fourier transformation. Spectra were visualised and overlaid in Sparky (Goddard et al. 2007).

Ligand handling

Compounds for the initial STD-NMR assessment of crystallographic fragment binding to Mpro were provided by the XChem group at Diamond Light Source in the form of a 384-well plated library (DSI-poised, Enamine), with compounds dissolved in d6-DMSO at 500 mM nominal concentration. 1 μL of dissolved compounds was aspirated from this library and immediately mixed with 9 μL of d6-DMSO for a final fragment concentration of 50 mM, from which NMR samples were formulated. For titrations of the same crystallographic fragments compounds were procured directly from Enamine in the form of lyophilized powder, which was dissolved in d6-DMSO to derive compound stocks at 10 mM and 100 mM concentrations for NMR sample formulation.

STD-NMR assays of bespoke Mpro ligands used compounds commercially synthesised for COVID Moonshot. These ligands were provided to us by the XChem group in 96-well plates, containing 0.7 μL of 20 mM d6-DMSO-disolved compound per well. Plates were created using an Echo liquid handling robot (Labcyte) and immediately sealed and frozen at −20 °C. For use, ligand plates were thoroughly defrosted at room temperature and spun at 3500×g for 5 min. In single-concentration STD-NMR experiments, 140 μL of a pre-formulated mixture of Mpro and NMR buffer with D2O and d6-DMSO were added to each well to create the final NMR sample. For STD-NMR competition experiments, 0.5 μL of ligands were aspirated from the plates and immediately mixed with 19.5 μL of d6-DMSO for final ligand concentration of 0.5 mM from which NMR samples were formulated. For 2D heteronuclear NMR spectra ligands were provided by the XChem group in 96-well plates containing pre-aliquoted compounds as above, where the pre-formulated mixture of protein and buffer was added.

Molecular dynamics (MD) simulations

The monomeric complexes of Mpro bound to chemical fragments were obtained from the RCSB Protein Data Bank entries 5R81 (ligand  x0195), 5REB ( x0387), 5RGI ( x0397), 5RGK ( x0426), 5R83 ( x0434) and 5REH ( x0540) for MD simulations with GROMACS version 2018 (Abraham et al. 2015) and the AMBER99SB-ILDN force field (Lindorff-Larsen et al. 2010). All complexes were inserted in a pre-equilibrated box containing water implemented using the TIP3P water model (Lindorff-Larsen et al. 2010). Force field parameters for the six ligands were generated using the general Amber force field and HF/6 – 31G*– derived RESP atomic charges (Bayly et al. 1993). The reference system consisted of the protein, the ligand, ~ 31,400 water molecules, 95 Na and 95 Cl ions in a 100 × 100 × 100 Å simulation box, resulting in a total number of ~ 98,000 atoms. Each system was energy-minimized and subsequently subjected to a 20 ns MD equilibration, with an isothermal-isobaric ensemble using isotropic pressure control (Bussi et al. 2009), and positional restraints on protein and ligand coordinates. The resulting equilibrated systems were replicated 4 times and independent 200 ns MD trajectories were produced with a time step of 2 fs, in constant temperature of 300 K, using separate v-rescale thermostats (Bussi et al. 2009) for the protein, ligand and solvent molecules. Lennard–Jones interactions were computed using a cut-off of 10 Å and electrostatic interactions were treated using particle mesh Ewald (Darden et al. 1993) with the same real-space cut-off. Analysis on the resulting trajectories was performed using MDAnalysis (Michaud-Agrawal et al. 2011; Gowers et al. 2016). Structures were visualised using PyMOL (DeLano 2002).

Notes

The enzymatic inhibition potential of Mpro ligands, measured by RapidFire mass spectrometry (Achdout et al. 2020), was retrieved from the Collaborative Drug Discovery database (CDD database 2021).

Results

STD-NMR assays of Mpro ligand binding

Mpro forms dimers in crystals via an extensive interaction interface involving two domains (Zhang et al. 2020). Mpro dimers likely have a sub-μM solution dissociation constant (Kd) by analogy to previously studied 3C-like coronavirus proteases (Grum-Tokars et al. 2008). At the 10 μM protein concentration of our NMR assays Mpro is, thus, expected to be dimeric with an estimated molecular weight of nearly 70 kDa. Despite the relatively large size of Mpro for solution NMR, 1H spectra of the protease readily showed the presence of multiple up-field shifted (< 0.5 ppm) peaks corresponding to protein methyl groups (Fig. 1a). In addition to demonstrating that Mpro is folded under the conditions tested, these spectra allowed us to identify the chemical shifts of Mpro methyl groups that may be suitable for on-resonance irradiation in STD-NMR experiments. Trials with on-resonance irradiation applied to different methyl group peaks showed that irradiating at 0.5 ppm (Fig. 1a) produced the strongest STD signal from ligands in the presence of Mpro, while simultaneously avoiding ligand excitation that would yield false-positive signals in the absence of Mpro (Fig. 1b). Further, we noted that small molecules abundant in the samples but not binding specifically to Mpro, such as DMSO, produced pseudo-dispersive residual signal lineshapes in STD spectra, while true Mpro ligands produced peaks in STD with absorptive 1H lineshapes. We surmised that STD-NMR is suitable for screening ligand binding to Mpro, requiring relatively small amounts (10–50 μgr) of protein and time (under 1 h) per sample studied.

The strength of STD signal is quantified by calculating the ratio of integrated signal intensity of peaks in the STD spectrum over that of the reference spectrum (STDratio). The STDratio factor is inversely proportional to ligand Kd, as \({STD}_{ratio}\propto \frac{1}{{K}_{d}+[L]}\) where [L] is ligand concentration. Measuring STDratio values over a range of ligand concentrations allows fitting of the proportionality constant and calculation of ligand Kd. However, time and sample-amount considerations, including the limited availability of bespoke compounds synthesized for the COVID Moonshot project, made recording full STD-NMR titrations impractical for screening hundreds of ligands. Thus, we evaluated whether measuring the STDratio value at a single ligand concentration may be an informative alternative to Kd, provided restraints could be placed, for example, on the proportionality constant.

Theoretical and practical considerations suggested that three parameters influence our evaluation of single-concentration STDratio values towards an affinity context. Firstly, the STDratio factor is affected by the efficiency of NOE magnetisation transfer between protein and ligand, which in turn depends on the proximity of ligand and protein groups, and the chemical nature of these groups (Mayer and Meyer 1999; Becker et al. 2018; Walpole et al. 2019). To minimize the influence of these factors across diverse ligands, we sought to quantify the STDratio of only aromatic ligand groups, and only consider those showing the strongest STD signal; thus, that are in closest proximity to the protein. Second, STD-NMR assays require ligand exchange between protein-bound and -free states in the timeframe of the experiment; strongly bound compounds that dissociate very slowly from the protein would yield reduced STDratio values compared to weaker ligands that dissociate more readily. Structures of Mpro with many different ligands show that the protein conformation does not change upon complex formation and that the active site is fully solvent-exposed (Douangamath et al. 2020), which suggests that ligand association can proceed with high rate (107–108 M−1 s−1). Under this assumption, the ligand dissociation rate is the primary determinant of interaction strength. Given the duration of the STD-NMR experiment in our assays, and the ratios of ligand:protein used, we estimated that significant protein—ligand exchange will take place even for interactions as strong as low-mM Kd. Finally, uncertainties or errors in nominal ligand concentration skew the correlation of STDratio to compound affinities; as shown in Fig. S1, STDratio values increase strongly when very small amounts of ligands are assessed. Thus, overly large STDratio values may be measured if ligand concentrations are significantly lower than anticipated.

Quantitating Mpro binding of ligands identified by crystallographic screening

Mindful of the limitations inherent to measuring single-concentration STDratio values, and prior to using STD-NMR to evaluate bespoke Mpro ligands, we used this method to assess binding to the protease of small chemical fragments identified in crystallographic screening experiments (Douangamath et al. 2020). In crystallographic screening campaigns of other target proteins such fragments were seen to have very weak affinities (> 1 mM Kd, e.g. Davies et al. 2016), thereby satisfying the exchange criterion set out above. 39 non-covalent Mpro interactors are part of the DSI-poised fragment library to which we were given access, comprising 17 active site binders, two compounds targeting the Mpro dimerisation interface and 20 molecules binding elsewhere on the protein surface (Douangamath et al. 2020). We initially recorded STD-NMR spectra from these compounds in the absence of Mpro to confirm that we obtained no or minimal STD signal when protease is omitted, and to verify ligand identity from reference 1H spectra. Five ligands gave no solution NMR signal or produced reference 1H spectra inconsistent with the compound chemical structure; these ligands were not evaluated further. Samples of 10 μM Mpro and 0.8 mM nominal ligand concentration were then formulated from the remaining 34 compounds (Table S1), and STD-NMR spectra were recorded, from which only aromatic ligand STD signals were considered for further analysis.

We observed large variations in STD signal intensity and STDratio values in the presence of Mpro across compounds (Fig. 2a, b; Table S1), with many ligands producing little or no STD signal, suggesting substantial differences in compound affinity for the protease. However, we also noted that ligand reference spectra differed substantially in intensity (Fig. 2c), despite compounds being at the same nominal concentration. Integrating ligand peaks in these reference spectra revealed differences in per-1H intensity of up to ~ 15-fold (Table S1). Such differences in ligand signal may arise from parameters of the NMR experiment, such as sample centering and calibration, from errors in sample formulation, or alternatively from concentration inconsistencies in the compound library and ligand aggregation in solution. To evaluate these possibilities we integrated the residual 1H signal of d6-DMSO in our reference spectra, which acts as internal control being sensitive to the same NMR parameters and formulation errors as the ligands. We found that DMSO signal varied by less than 35% across any pair of samples (11% average deviation). Thus, we concluded that NMR parameters and sample formulation errors may have contributed differences in ligand signal of up to ~ 1/3, but did not account for the ~ 15-fold signal differences observed. This suggests that effective ligand concentrations in solution vary substantially.

Fig. 2
figure 2

Assessment of fragment binding to Mpro. a STDratio values for chemical fragments identified by crystallographic screening as binding to Mpro (Douangamath et al. 2020). Ligands binding to the Mpro active site are coloured orange, at the Mpro dimer interface in red, and elsewhere on the protein surface in blue. b Overlay of STD-NMR spectra from fragments  x0305,  x0387 and  x0434, which bind the Mpro active site, showing the ligand aromatic region in the presence of Mpro. Spectra are colour coded per ligand as indicated. As seen, the three fragments yield significantly different STD signal intensities captured in the STDratio values shown in (a). c Overlay of reference spectra from fragments  x0305,  x0376 and  x0540, showing the ligand aromatic region. Peak intensities vary substantially, suggesting significant differences in ligand concentration

Full size image

Given that differences in effective compound concentration can skew the relative STDratio values of ligands (Fig. S1), and that such concentration differences were also observed among newly designed Mpro inhibitors (see below), we questioned whether recording STDratio values under these conditions can provide useful information. To address this question we attempted to quantify the affinity of crystallographic fragments to Mpro, selecting ligands that showed clear differences in STDratio values in the assays above and focusing on compounds binding at the Mpro active site; hence, that are of potential interest to inhibitor development. We performed Mpro binding titrations monitored by STD-NMR of compounds  x0195,  x0354,  x0426 and  x0434 in 50 μM–4 mM concentrations (Fig. S2), and noted that only compounds  x0434 and  x0195, which show the highest STDratio (Fig. 2a), bound strongly enough for an affinity constant to be estimated (Kd of 1.6 ± 0.2 mM and 1.7 ± 0.2 mM, respectively). In contrast, the titrations of x0354 and  x0426, which yielded lower STDratio values, could not be fit to extract a Kd indicating weaker binding to Mpro.

To further this analysis, we assessed the binding of fragments  x0195, x0387,  x0397,  x0426,  x0434 and  x0540 to the Mpro active site using atomistic molecular dynamics (MD) simulations of 200 nsec duration. As shown in Fig. S3a, b, and Movies S1 and S2, fragments with high STDradio values ( x0434 and  x0195) always located in the Mpro active site despite exchanging between different binding conformations (Fig. S4), with average ligand root-mean-square-deviation (RMSD) of 3.2 Å and 5.1 Å respectively after the first 100 nsec of simulation. Medium STDratio value fragments ( x0426 and  x0540, Fig. S3c, d, and Movies S3 and S4) show average RMSDs of approximately 9 Å in the same simulation timeframe, frequently exchanging to alternative binding poses and with  x0540 occasionally exiting the Mpro active site. In contrast, fragments showing very little STD NMR signal ( x0397 and  x0387, Fig. S3e, f, and Movies S5 and S6) regularly exit the Mpro active site and show average RMSDs in excess of 15 Å with very limited stability. Combining the quantitative Kd and MD information above, we surmised that, despite limitations inherent in this type of analysis and uncertainties in ligand amounts in solution, STDratio values recorded at single compound concentration can act as proxy measurements of Mpro affinity for ligands.

Assessment of Mpro binding by COVID Moonshot ligands

We proceeded to characterise by STD-NMR the Mpro binding of bespoke ligands created as part of the COVID Moonshot project and designed to act as non-covalent inhibitors of the protease (Achdout et al. 2020). Similar to the assays of crystallographic fragments above, we focused our analysis of STD signals to aromatic moieties of ligands binding to the Mpro active side and extracted STDratio values only from the strongest STD peaks. Once again, we noted substantial differences in effective compound concentrations, judging from reference 1H spectral intensities (Fig. 3a). These differences, which may reflect ligand aggregation, could not be attributed to errors in NMR parameters or sample preparation as the standard deviation of residual 1H intensity in the d6-DMSO peak did not exceed 5% in any of the ligand batches tested. Crucially, out of 650 different molecules tested, samples of 35 compounds (7.6%) yielded no detectable NMR signal and 86 (13.2%) very little signal (Fig. 3a). In these cases, NMR assays were repeated using a separate batch of compound; however, 96.2% of repeat experiments yielded the same outcome of no or very little NMR signal from the ligands.

Fig. 3
figure 3

STD-NMR of COVID Moonshot ligands binding to Mpro. a Overlay of reference spectra from the indicated COVID Moonshot ligands, showing the ligand aromatic region in each case. Spectra are colour coded per ligand as indicated. As seen, peak intensities vary substantially, suggesting significant differences in ligand concentration. Peaks of ligand EDJ-MED-c8e7a002-1 (green) are indicated by arrows; ligand EDJ-MED-e4b030d8-12 (red) produced no peaks in the NMR spectrum. b Plot of STDratio values from COVID Moonshot ligands assessed by STD-NMR against their IC50 value estimated by RapidFire mass spectrometry enzymatic assays (Achdout et al. 2020). Ligands in blue show weak correlation between the two methods (red line, corresponding to an exponential function along the IC50 dimension). Ligands in grey represent outliers of the STD-NMR or enzymatic method as discussed

Full size image

We measured STDratio values from samples where ligands produced sufficiently strong reference 1H NMR spectra to be readily visible, and deposited these values and associated raw NMR data to the Collaborative Drug Discovery database (CDD database 2021). Some of these ligands were assessed independently for enzymatic inhibition of Mpro using a mass spectrometry method as part of the COVID Moonshot collaboration (Achdout et al. 2020). Where both parameters are available, we compared the STDratio values and 50% inhibition concentrations (IC50) of these ligands. As shown in Fig. 3b, STDratio and IC50 values show weak correlation (R2 = 30%) for most ligands tested; however, a subset of ligands displayed conspicuously low or even no STD signals considering their effect on Mpro activity, and presented themselves as outliers in the correlation graph. As these outlier ligands had IC50 values below 10 μM, suggesting that their affinities to the protease may be in the μM Kd region, we considered whether our approach gives rise to false-negative STD results, for example through slow ligand dissociation from Mpro.

To address this question, we derived an assay whereby the bespoke, high-affinity Mpro inhibitor would outcompete a lower-affinity ligand known to provide strong STD signal from the protease active site. In these experiments the lower-affinity ligand would act as ‘spy’ molecule whose STD signal reduces as function of inhibitor concentration. We used fragment  x0434, which yields substantial STD signal with Mpro (Figs. 1b and 2a), as ‘spy’, and tested protease inhibitors EDJ-MED-a364e151-1, LON-WEI-ff7b210a-5, CHO-MSK-6e55470f-14 and LOR-NOR-30067bb9-11 as  x0434 competitors. Of these inhibitors, EDJ-MED-a364e151-1 gave rise to substantial STD signal in earlier assays, whereas the remaining produced little or no STD signal; yet, all four inhibitors were reported to have low-µM or sub-µM IC50 values based on Mpro enzymatic assays. In these competition experiments, both EDJ-MED-a364e151-1 and LON-WEI-ff7b210a-5 yielded Kd parameters comparable to the reported IC50 values (Fig. 4a, b), showing that at least in the case of LON-WEI-ff7b210a-5 the absence of STD signal in the single-concentration NMR assays above represented a false-negative result. In contrast, CHO-MSK-6e55470f-14 and LOR-NOR-30067bb9-11 were unable to compete x0434 from the protease active site (Fig. 4c, d), suggesting that in these two cases the reported IC50 values do not reflect inhibitor binding to the protease, and that the weak STD signal of the initial assays was a better proxy of affinity. We surmised that although some low STDratio values of Mpro inhibitors may not accurately reflect compound affinity to the protease, such values cannot be discounted as a whole as they may correspond to non-binding ligands.

Fig. 4
figure 4

Quantification of interaction affinities for COVID Moonshot ligands binding to Mpro. Shown here are plots of STDratio values derived from two NMR peaks of fragment  x0434 as function of a competing COVID Moonshot ligand concentration. For each ligand, the STDratio values derived from single-point assays and the IC50 values from RapidFire enzymatic assays are indicated. The titrations of COVID Moonshot ligands EDJ-MED-a364e151-1 (a) and LON-WEI-ff7b210a-5 (b) could be fit to an ideal competition model to derive binding affinities for these compounds; in contrast, ligands CHO-MSK-6e55470f-14 (c) and LOR-NOR-30067bb9-11 (d) failed to displace fragment  x0434 from the Mpro active site in a significant manner

Full size image

To further this analysis, we attempted to evaluate binding of the same protease inhibitors to Mpro using protein-observed heteronuclear 2D spectra. In such spectra protein resonances are expected to shift as a result of compound binding due to alterations in the chemical environment of the binding site caused by the ligand. We produced 15N isotopically enriched Mpro and recorded spectra using a number of different NMR experiments optimised for rapid data acquisition and signal accumulation, including SOFAST-HMQC, BEST-HSQC and BEST-TROSY (Favier and Brutscher 2019). We obtained the best Mpro spectra using the SOFAST-HMQC pulse sequence; however, we noted that spectral quality, including signal-to-noise obtained per unit time, was reduced as function of protein concentration, which is indicative of Mpro aggregation under the conditions used (Fig. S5a, b). As a result, we proceeded to record spectra at 25 μM Mpro concentration, which necessitated almost 7.5 h of acquisition time per spectrum to accumulate adequate signal-to-noise.

Despite these efforts, the resulting SOFAST-HMQC spectra of Mpro were of relatively low resolution and signal quality (Fig. S5b), and displayed in the order of tens of discrete peaks (for reference, we would expect one peak per amino acid residue except for prolines; hence, more than 290 discrete peaks for Mpro). Addition of ligands in concentrations that should result in Mpro saturation based on reported ligand IC50 values yielded no changes to the vast majority of peaks in the NMR spectra, with only a small number of resonances showing slight perturbations (Fig. S5c, d). As assignments of Mpro NMR resonances have not yet been reported, we were unable to confirm whether the slight perturbations observed upon ligand addition corresponded to binding events at the protease active site, or alternatively non-specific interactions elsewhere on the protein. Given the difficulty in obtaining these 2D spectra, and the ambiguous nature of their interpretation, we concluded that such protein-observed experiments are not a suitable NMR method for characterising ligand binding to Mpro.

Discussion

Fragment-based screening is a tried and tested method for reducing the number of compounds that need to be assessed for binding against a specific target in order to sample chemical space (Erlanson et al. 2016). Combined with X-ray crystallography, which provides information on the target site and binding pose of ligands, initial fragments can quickly be iterated into potent and specifically-interacting compounds. The COVID Moonshot collaboration (Achdout et al. 2020) took advantage of crystallographic fragment-based screening (Douangamath et al. 2020) to initiate the design of novel inhibitors targeting the essential main protease of the SARS-CoV-2 coronavirus; however crystallographic structures do not report on ligand affinity and inhibitory potency in enzymatic assays does not always correlate with ligand binding. Thus, supplementing these methods with solution NMR tools highly sensitive to ligand binding can provide a powerful combination of orthogonal information and assurance against false starts. Our primary aim in the study presented here was to provide exactly this type of information for COVID Moonshot. However, we recognise that many of the problems encountered in this work are likely to reoccur in the context of other ligand screening efforts; hence, we hope that this report will prove informative to a more general audience.

A key issue encountered early in our effort to characterise Mpro-interacting ligands was compound quality. An initial assessment of Mpro ligands drawn from a commercial library showed large variation of up to 15-fold in effective ligand concentration (Fig. 2c). Ligand quality control is a recognised problem in screening campaigns stemming from a multitude of factors, including the exact amount of compounds provided by vendors, storage conditions, and ligand degradation and aggregation in aqueous buffers, among others (Lepre 2011). A recent study aiming to develop a robust fragment library for high-throughput screening reported that up to approximately 30% of fragments failed to pass one or more quality-control metrics, including ligand concentration (Sreeramulu et al. 2020). Our own experience with bespoke Mpro ligands was that ~ 20% of compounds had low or very low effective concentration in solution (Fig. 3a). Although in the context of the COVID Moonshot project this proportion of poorly-behaved ligands could be tolerated in the interest of rapid progression, it is important to underline the effect of library quality in screening as this can give rise to both false-negative and false-positive results.

Despite these deficiencies, we showed that STD-NMR is a suitable method for characterising ligand binding to Mpro, allowing us to assess ligand interactions using relatively small amounts of protein and in under 1 h of experiment time per ligand (Fig. 1b). However, screening compounds in a high-throughput manner is not compatible with the time- and ligand-amount requirements of full STD-NMR titrations. Thus, we resorted to using an unconventional metric, the single-concentration STDratio value, as proxy for ligand affinity. Although this metric has limitations due to its dependency on magnetisation transfer between protein and ligand, and on relatively rapid exchange between the ligand-free and -bound states, we demonstrated that it can nevertheless be informative. Specifically, the relative STDratio values of chemical fragments bound to the Mpro active site provided insight on fragment affinity (Fig. 2a), as crosschecked by quantitative titrations (Fig. S2) and MD simulations (Fig. S3). Furthermore, STDratio values of COVID Moonshot compounds held a weak correlation to enzymatic IC50 parameters (Fig. 3b), although false-negative and -positive results from both methods contribute to multiple outliers. Thus, in our view the biggest limitation of using the single-concentration STDratio value as metric relates to its supra-linear sensitivity to effective ligand concentration (Fig. S1), which can vary substantially across ligands in a large project (Fig. 3a).

How then should the STD data recorded as part of COVID Moonshot be used? Firstly, we showed that at least for some bespoke Mpro ligands the STDratio value obtained is a better proxy for compound affinity compared to IC50 parameters from enzymatic assays (Fig. 4). This, inherently, is the value of employing orthogonal methods thereby minimizing the number of potential false results. Thus, when one is considering existing Mpro ligands to base the design of future inhibitors, a high STDratio value as well as low IC50 parameters are both desirable. Second, due to the aforementioned limitations of single-concentration STDratio value as proxy of affinity, and the influence of uncertainties in ligand concentrations, we believe that comparisons of compounds and derivatives differing by less than ~ 50% in STDratio is not meaningful. Rather, we propose that the STDratio values of Mpro ligands measured and available at the CDD database should be treated as a qualitative metrics of compound affinity.

In conclusion, we presented here protocols for the assessment of SARS-CoV-2 Mpro ligands using STD-NMR spectroscopy, and evaluated the relative qualitative affinities of chemical fragments and compounds designed as part of COVID Moonshot. Although development of novel antivirals to combat COVID-19 is still at an early stage, we hope that this information will prove valuable to groups working towards such treatments.