Computational redesign of nirmatrelvir yields ranked MERS-CoV protease inhibitor candidates

Researchers have built a large virtual library of nirmatrelvir analogues and used it to develop a fast scoring method for prioritizing candidate inhibitors of the MERS-CoV main protease, the viral enzyme that the drug nirmatrelvir was originally designed to block in SARS-CoV-2. The study, published as an open-access research article in Discover Molecules on 14 October 2026, is entirely computational: no new inhibitor was synthesized or tested in the laboratory, and the authors state explicitly that the predicted binding affinities should not be interpreted as experimentally confirmed inhibitory activities.

The team, led by Md Saidur Rahman and Shahidul M. Islam, started from a structural feature of coronaviruses. The main protease, also called 3CLpro, is essential for processing viral polyproteins and is therefore a prime target for small molecules. SARS-CoV-2 and MERS-CoV main proteases share roughly 50 percent sequence identity and closely similar active-site architectures, yet nirmatrelvir binds MERS-CoV Mpro poorly. Six amino acid residues differ between the two binding sites — M49L, N142/145C, H164/167Q, P168/171A, R188/191K and T190/193V — and the nirmatrelvir binding site contains 21 residues within 4 angstroms in SARS-CoV-2 Mpro but only 18 in MERS-CoV Mpro. Known resistance mutations, including S144A and E166A in SARS-CoV-2 and S142G, S142R and S147Y in MERS-CoV, further reduce nirmatrelvir binding, motivating the search for analogues that could remain effective against resistant or related proteases.

To generate candidates, the researchers used the open-source cheminformatics toolkit RDKit to systematically replace side chains at six positions on the nirmatrelvir scaffold, which correspond to the S1 through S4 substrate-binding pockets of the protease. Single-position substitutions produced 209 unique analogues, and simultaneous substitutions at two positions added 15,744 more, for a combined virtual library of 15,953 drug-like structures. The workflow enforced chemical validity, preserved the core scaffold, avoided multiple substitutions at the same position, and excluded fragments that would push molecular weight beyond 700 daltons. All 15,953 molecules were then docked into the MERS-CoV Mpro active site with AutoDock Vina, using the crystal structure PDB 7VTC, which contains nirmatrelvir bound to MERS-CoV Mpro at 2.54 angstrom resolution. Redocking the co-crystallized ligand reproduced the crystallographic binding mode, which the authors cite as evidence that the docking protocol was reliable.

A key analytical step compared the designed analogues with a control set of 13,090 randomly selected ChEMBL compounds curated to match the molecular weight distribution of the analogue library. Using both 2D Morgan fingerprint similarity (Tanimoto coefficient) and 3D shape similarity (Shape Tversky Index), the team found a clear positive relationship between structural resemblance to nirmatrelvir and docking performance among the designed compounds, while the random compounds showed a much weaker and more variable trend. Most random molecules clustered below 20 percent 2D similarity and 40 percent 3D similarity to nirmatrelvir. For the analogues, the average docking score improved progressively as similarity increased, with a noticeable enhancement around 50 to 60 percent similarity and average scores reaching approximately minus 7.5 kilocalories per mole for compounds above 90 percent similarity. The 3D shape similarity tracked docking performance almost monotonically, suggesting that preserving the overall molecular geometry of nirmatrelvir matters more for binding than topological resemblance alone.

From this similarity–affinity relationship the authors derived two related metrics. The first, a Binding Confidence Score (BCS), combines the fraction of molecules in a given similarity interval, the similarity itself, and the docking score, with empirical weighting parameters and a correction factor for sparse high-similarity bins. The second, a Binding Score (BS), is docking-independent: for a query molecule it uses only the molecule’s 2D similarity to nirmatrelvir and the average docking score of all analogues within the corresponding similarity interval, so a candidate can be ranked without running a docking calculation. The authors emphasize that these empirical parameters were derived from the nirmatrelvir analogue library itself, and that BCS and BS were used as pragmatic filters for compound prioritization rather than independently trained or validated models for predicting biological activity.

Applying the framework, six top-ranked candidates emerged, labeled M1 through M6. M1, M2, M3 and M4 came from two-position modifications, while M5 and M6 came from single-position substitutions. Static docking scores against the crystal structure were minus 9.0 kilocalories per mole for M1 and M2, minus 8.8 for M3 and M4, minus 8.5 for M5 and minus 8.2 for M6. The docking-independent BS values for M1, M3, M5 and M6 were 9.1, 9.2, 10.6 and 8.9, respectively, substantially higher than the scores of 4.7 for both M2 and M4. Molecular dynamics exposed why the confidence metric was useful: after simulation, M2’s docking score fell to minus 6.8 kilocalories per mole as all initial hydrogen bonds were lost, and M4’s score dropped to minus 7.6 as the ligand shifted away from the canonical active-site pocket. Two other compounds, M7 and M8, detached entirely from the protease during simulation despite respectable initial docking scores, underscoring the limits of static docking.

The four leading analogues, M1, M3, M5 and M6, were then stress-tested with four independent 250-nanosecond molecular dynamics simulations each, totaling 1 microsecond per protein–ligand system, alongside nirmatrelvir. Root mean square deviations of the protease and the complexes remained stable around 1.5 to 2.0 angstroms, ligand RMSD profiles stabilized after roughly 125 nanoseconds, and the active-site residues showed minimal fluctuations below 2 angstroms. Hydrogen-bond analysis showed that a glutamine-to-ligand contact at Glu169 anchored all complexes except M1, which instead relied on a Glu169 interaction with 47 percent occupancy; M3 showed the most persistent hydrogen bonding, with 96 percent occupancy for its key Glu169 contact. MM/GBSA binding free energies averaged over four trajectories came to minus 31.7, minus 42.9, minus 35.8 and minus 33.8 kilocalories per mole for M1, M3, M5 and M6, respectively, compared with minus 28.2 for nirmatrelvir, with van der Waals contributions reaching minus 54.87 kilocalories per mole for the top analogues. KDEEP, a 3D convolutional neural network trained on the PDBbind dataset, predicted absolute binding affinities of minus 9.5 to minus 10.6 kilocalories per mole for the top compounds, and the Boltz-2 deep-learning framework predicted pIC50 values of 10.0 to 11.8, corresponding to predicted IC50 values of 0.089, 0.016, 0.020 and 0.150 micromolar for M1, M3, M5 and M6 against a predicted 0.82 micromolar for nirmatrelvir.

As an external check, the authors applied the BS approach to published nirmatrelvir analogues with experimental IC50 values against SARS-CoV-2 Mpro. Replacing the P1 nitrile group with an alkyne, an aldehyde or a carboxylic acid yielded reported IC50 values of 0.14, 0.010 and 0.008 micromolar, respectively, versus at most 0.025 micromolar for unmodified nirmatrelvir, and the calculated BS values of 8.1, 8.9 and 9.0 tracked that ordering. The authors caution that this comparison provides supporting evidence but not full independent validation, because the experimental dataset is small and drawn from SARS-CoV-2 rather than MERS-CoV protease inhibitors. They also note that Boltz-2 outputs are predictions, not measurements, so agreement with it cannot count as experimental validation of the BS approach.

The study carries several stated limitations. Because nirmatrelvir was removed before docking but the receptor kept its ligand-bound conformation, the template may bias results toward structurally similar analogues. Entropic contributions were omitted from the MM/GBSA calculations for computational cost, a common approximation the authors argue is acceptable for ranking structurally related analogues. To make the method reproducible, the team released a web interface that computes BS values for any molecule from its similarity data and the average docking score of its similarity interval, along with the modification scripts in a public GitHub repository. The authors conclude that the prioritized analogues M1, M3, M5 and M6 are computationally prioritized candidates that now require biochemical inhibition assays, structural characterization and antiviral experiments before any claim of practical efficacy can be made.

Subject of Research: Computational side-chain optimization of nirmatrelvir to design and rank candidate inhibitors of coronavirus main proteases

Article Title: Computationally guided side chain optimization of nirmatrelvir for lead discovery against coronavirus main proteases

Article References: Rahman, M. S., & Islam, S. M. (2026). Computationally guided side chain optimization of nirmatrelvir for lead discovery against coronavirus main proteases. Discover Molecules, 3(1), Article 35. https://doi.org/10.1007/s44345-026-00079-3

Image Credits: AI Generated

DOI: 10.1007/s44345-026-00079-3

Keywords: nirmatrelvir, MERS-CoV, SARS-CoV-2, main protease, molecular docking, molecular dynamics, MM/GBSA, Boltz-2, KDEEP, drug discovery, cheminformatics, binding score