To construct a cγAA-containing macrocyclic peptide library, γ1 and γ2 (Fig. 1a) were assigned to GAG and GUG codons using tRNAPro1E2CUC and tRNAPro1E2CAC, respectively (Fig. 1b), employing the FIT system. Two cyclic β2,3-amino acids (cβAAs), (1R,2S)-2-aminocyclopentane carboxylic acid (β1) and (1S,2S)-2-aminocyclopentane carboxylic acid (β2), were introduced at the GUU and UGU codons using tRNAPro1E2GAC and tRNAGluE2GCA, respectively, because we considered that the incorporation not only of cγAA, but also of cβAA, would give more diverse folding possibilities25. For the macrocyclization of the library via thioether bonds, we introduced N-chloroacetyl-d-tyrosine (ClAcy) and d-cysteine (c) at the initiator AUG codon using tRNAfMetCAU and elongator AUG codon using tRNAPro1E2CAU, respectively. The thioether bond spontaneously forms between the N-terminal chloroacetyl group on ClAcy and the thiol group of the c residue (Fig. 1c). These amino acids were precharged onto the respective tRNAs using flexizymes. The peptide library comprised a repeat of 11−14 random residues encoded by GWG and NNU codons (W = A or U; N = A, U, G or C) flanked by the cyclizing ClAcy and c residues. Since multiple cγAA incorporations could not be efficiently achieved, as reported in our previous study29, the library was designed to have only a single γ1 or γ2 appear at the GWG codon. The two cβAAs and the thirteen proteinogenic l-α-amino acids (A, D, F, G, H, I, L, N, P, R, S, T and Y) were assigned at NNU codons. The C-terminal SGGSGG sequence following the c residue was a linker peptide connected to the 3′ end of mRNA via a puromycin linker. To demonstrate the fidelity of translation, model macrocyclic peptides containing any of cγAA and cβAA (γ1, γ2, β1 and β2 assigned at GAG, GUG, GUU and UGU codons, respectively) were synthesized, and their identities were confirmed by matrix-assisted laser desorption/ionization coupled with time of flight mass spectrometry (MALDI-TOF MS; Supplementary Fig. 2).

The macrocyclic peptide library was then applied to the RaPID selection against recombinant SARS-CoV-2 Mpro. Translation of the random mRNA library into the peptide library, conjugation of the peptide with the parent mRNA via a puromycin linker and reverse transcription of the mRNA into the cDNA yielded peptide/mRNA/cDNA conjugates (Extended Data Fig. 1a). The library was first subjected to naked magnetic bead treatment to remove bead binders, and then applied to Mpro-immobilized magnetic beads to recover Mpro binders. The cDNA moiety of the recovered peptide/mRNA/cDNA conjugate was amplified by the polymerase chain reaction (PCR) and transcribed into the mRNA library for the next selection round (details in Methods). By repeating this affinity selection, the recovery rate of the Mpro binders was greatly increased at the third round of selection and, even more obviously, in the fourth round, while the recovery of the bead binders did not increase (Extended Data Fig. 1b). Deep sequencing of the cDNA library at the fourth round revealed that the library was enriched with several families of peptides bearing cγAA (Supplementary Table 1 shows the top 100 sequences). Among these, seven macrocyclic peptides containing cγAA were selected for further analysis of their binding affinities, inhibitory activity and proteolytic stability (Table 1, GM1−GM7). GM1, GM2, GM4 and GM6 contain a γ1 residue in their sequences, while GM3, GM5 and GM7 contain γ2. Note that the cβAA-containing peptides were not selected because they were not included in the major peptide families with high read numbers. GM1−GM7 were chemically synthesized on a large scale using the standard solid-phase method without the C-terminal SGGSGG linker, and their purities and identities were confirmed by ultra-performance liquid chromatography (UPLC) and MALDI-TOF MS, respectively (Supplementary Figs. 3 and 4).

We first evaluated the binding affinity of GM1−GM7 to Mpro by surface plasmon resonance (SPR). All the peptides exhibited low-to-moderate nanomolar kinetic equilibrium (KD) values (Table 1 and Supplementary Fig. 5). Notably, GM1 and GM4, which contain γ1, and GM5, which contains γ2, have a conserved four residue motif, yFHγX (γX = γ1 or γ2), at their N termini and showed potent affinities, with KD values of 2.3, 5.2 and 5.2 nM, respectively; GM2 has the same motif, but has a higher KD (71 nM). GM6 and GM7, which had relatively low read frequencies, exhibited weak binding compared to GM1−GM5, indicating selection for tight binding peptides. To evaluate contributions of the cγAA residues to potency, we synthesized peptides where the cγAA was substituted with alanine (Table 1, GM1γ14A, GM3γ210A, GM4γ14A and GM5γ24A). In addition to GM1, GM4 and GM5, which share the conserved yFHγX motif, GM3 was selected for alanine substitution because of its high read percentage among peptides without the yFHγX motif. SPR measurements of the alanine mutants revealed that the KD values of GM1γ14A, GM4γ14A and GM5γ24A were comparable or one order higher than those of the original peptides, indicating that the cγAA residue in the conserved yFHγX motif is important for binding. To our surprise, GM3γ210A exhibited an eightfold stronger binding affinity (KD = 7.5 nM), revealing that the contribution of cγAA to the binding affinity is sequence context dependent.

We next evaluated the inhibitory activity of GM1−GM7 against the hydrolytic activity of SARS-CoV-2 Mpro using a reported MS-based method39. GM1−GM7 exhibited inhibition of Mpro; in particular, the yFHγX motif containing GM1, GM4 and GM5, which manifest single-digit nanomolar KD values, showed particularly potent inhibition, with half-maximal inhibitory concentration (IC50) values of 40, 50 and 40 nM, respectively (Table 1, Fig. 2a and Extended Data Fig. 2). The peptides with weaker binding affinity, that is, GM2, GM3, GM6 and GM7, had IC50 values of 1,750–7,930 nM, implying correlation between IC50 and KD values. The inhibitory activities of the alanine mutants were then determined. Notably, despite GM1γ14A having only a fourfold weaker binding affinity compared with GM1, inhibition was ablated, revealing the importance of the cγAA. By contrast, GM3γ210A, GM4γ14A and GM5γ24A retained inhibition activity, showing the context-dependent effects of the cγAA (IC50 = 120, 10 and 50 nM, respectively; Table 1, Fig. 2a and Extended Data Fig. 2).

a, Dose response analysis of peptides against Mpro. The Mpro inhibitory activities of peptides were investigated by solid-phase extraction purification coupled to MS analysis using a RapidFire 365 high-throughput sampling robot (Agilent) connected to an iFunnel Agilent 6550 accurate mass quadrupole TOF mass spectrometer39. IC50 values were determined by the mean of three or five independent replicates each performed in technical duplicate. Data are presented as mean values ± standard deviation, s.d. (n = 5 for GM4, GM4γ14A and GM4H3Q; n = 3 for GM4γ14N, GM4H3A and GM4H3E). Extended Data Fig. 2 shows other peptides. b, Serum stability assay of macrocyclic peptides. GM4, GM4γ14A and GM4H3Q were co-incubated with an internal standard peptide in human serum (37 °C). At each time point, the relative intensity of each peptide to the standard peptide was estimated by LC/MS. The relative intensity at 0 h was defined as 100%. Half-lives (t1/2) were determined by analysing the mean of three technical replicates of each sample by nonlinear regression using GraphPad Prism 9. Data are presented as mean values ± s.d. (n = 3). Extended Data Fig. 3 shows results for other peptides.

We evaluated the half-life of potent cγAA-containing peptides and their mutants in human serum, because in vivo stability is a critical factor in the development of therapeutic peptides. Each peptide and an uncleavable internal standard peptide were co-incubated in human serum at 37 °C, and the relative amount of the remaining sample peptide was estimated by liquid chromatography/MS (LC/MS). The potent inhibitors GM1 and GM4, containing γ1, and GM5, containing γ2, exhibit high peptidase resistance with half-lives (t1/2) of 90 h, 126 h and 32 h, respectively (Table 1, Fig. 2b and Extended Data Fig. 3a). Importantly, by contrast, their alanine mutants show substantially shorter half-lives (t1/2 = 8.3 h, 11 h and 21 h for GM1γ14A, GM4γ14A and GM5γ24A, respectively; Table 1). The enhancement of serum resistance observed for GM4, for instance, was 12-fold. Notably, despite the substitution of γ1 to an Ala residue not diminishing inhibition relative to GM4, the enhancement in serum stability by the γ1 residue is substantial. Moreover, the less potent inhibitors GM2 and GM3 have more than nearly three orders of magnitude shorter half-lives (1.1 h and 2.5 h, respectively) compared to GM4, suggesting that the higher (binding and inhibitory) activity of GM4 is reflected in its higher serum stability, possibly due to its more conformationally constrained fold.

To identify peptidase cleavage sites in serum, the products of GM4 and GM4γ14A after 24 h incubation were analysed by LC/MS. In the case of GM4, six fragments (GM4-f1−GM4-f6, Extended Data Fig. 3b) containing the c-(thioether)-AcyFHγ1L motif were detected, but no fragmentation of the motif itself was apparent. With GM4γ14A, three shorter fragments containing c-(thioether)-AcyF were detected (GM4γ14A-f1−GM4γ14A-f3; Extended Data Fig. 3c), indicating cleavage between F and H of the motif. These results show that the presence of non-canonical residues (one cγAA (γ1), two d-amino acids and a thioether bond) can enable high peptidase resistance by preventing the proteolysis of flanking residues, leading to a remarkable improvement in serum stability.

To gain structural insights into how GM4, including its γ1 residue, interacts with Mpro, we obtained an X-ray crystal structure of their complex to 1.7 Å resolution (Fig. 3a and Supplementary Table 2). The structure reveals GM4 bound with high occupancy and in the same manner at the active site regions of both monomers in the Mpro dimer. Excellent electron density was observed for all residues in the GM4 macrocycle (Fig. 3c). The overall conformations of Mpro in the GM4 and nirmatrelvir38 (used for comparison) complex structures are very similar (backbone root mean square deviation, 0.34).

a, GM4 binds in the substrate binding cleft of both protomers A and B in the Mpro dimer. b, Structure of GM4; the H3GM4 carbonyl O is indicated with an arrow. c, Polder omit map of GM4 contoured at a level of ±1.5 standard deviation (σ). d, View of GM4 at the active site. The H41 and C145 catalytic dyad is in orange. γ14GM4 (magenta) occupies the S1′ pocket; the side chains of H3GM4, F2GM4 and Acy1GM4 occupy the S1, S2 and S4 pockets, respectively. e,f, Close-ups of the S1 pocket, showing the H3GM4 backbone carbonyl in the oxyanion hole (C145, G143 backbone amides). The backbone NH of H3GM4 is positioned to interact with the H164 backbone CO. The H3GM4 imidazole is positioned to hydrogen bond with the H163 side chain and to interact with the E166 side chain, which interacts with S1 of protomer B, as observed in apo-Mpro structures. Y9GM4 is 3.1 Å and 3.3 Å from the E166 and S1B side chains, respectively. g, Nirmatrelvir with the residues occupying the S1−S4 subsites labelled P1−P4 (the reactive nitrile is indicated with an arrow). h, Superimposition of views from crystal structures of Mpro with GM4 and nirmatrelvir (Protein Data Bank 7VH8 (ref. 60)). GM4 non-covalently interacts at the active site, while the nitrile of nirmatrelvir reacts with C145. Note the similar locations of the nirmatrelvir nitrile-derived N and the H3GM4 carbonyl O.

The four residues in the yFHγ1 motif of GM4 occupy the S4, S2, S1 and S1′ substrate binding subsites, respectively40 (Fig. 3d). H3GM4 is accommodated in the S1 subsite, forming hydrogen bonds with Mpro C145, H164 and H163 (GM4 residues are identified by a subscript; Fig. 3e,f). The F2GM4 residue binds in the S2 subsite, making hydrophobic interactions with M49, M165 and H41. The y1GM4 side chain forms hydrogen bonds with T190 and Q192 in the S4 subsite (Extended Data Fig. 4a).

The c13GM4–thioether link, which enables macrocyclization, and the adjacent G12GM4 enable a turn connected to P11GM4, R10GM4 and Y9GM4; none of these residues are positioned to make direct contacts with Mpro in the structure, though these cannot be ruled out in solution. The backbone of GM4 forms three intramolecular hydrogen bonds, that is between γ14GM4 and L7GM4, Acy1GM4 and R10GM4, and P11GM4 and c13GM4 (Extended Data Fig. 4b). The side chains of P11GM4 and Y9GM4 are arranged such that they ‘stack’ with each other, an arrangement positioning the phenolic side chain of Y9GM4 adjacent to the imidazole ring of H3GM4, appearing to lock it into the S1 pocket; the phenolic side chain of Y9GM4 also forms interactions with Ser1, from the other protomer of the Mpro dimer (Fig. 3f).

G8GM4, L7GM4, N6GM4 and L5GM4 form a loop leading to γ14, which is linked to H3GM4. γ14GM4 is positioned to make hydrophobic contact with T24 and T25; the side chains of L5GM4 and N6GM4 are positioned close to the surface of Mpro, with L7GM4 projecting towards H41, C44 and T45. The observation of the unnatural γ14GM4 residue, the cyclobutane ring of which was refined in a near planar conformation, in the P1′ position is striking and implies it contributes to the inhibition or lack of efficient GM4 hydrolysis.

At the active site region, the sulfur of the nucleophilic C145 is positioned close (3.4 Å) to the imidazole ε-nitrogen of H41, which acts as a general base/acid in Mpro catalysis, and to the carbonyl carbon (4.4 Å) of the amide linking H3GM4 and γ14GM4; this carbonyl carbon is positioned in a similar manner to the imine (anion) derived by reaction of the nirmatrelvir nitrile with C145 (Fig. 3g,h).

Like the binding modes of substrates (predicted) and inhibitors (as observed by crystallography), the carbonyl of the amide linking H3GM4 and γ14GM4 is positioned in the oxyanion hole, making hydrogen bonds with the C145 NH (2.9 Å) and with the G143 NH (2.9 Å). The ~100° angle between the C145 sulfur and the carbon and oxygen of the amide linking H3 and γ14 is close to that predicted for nucleophilic attack onto an amide carbonyl as observed in studies on serine proteases41. Despite this apparently catalytically productive arrangement, analysis of the electron density map implies a lack of substantial covalent reaction, consistent with solution studies showing a lack of efficient GM4 hydrolysis by Mpro. Lack of reaction appears to be a consequence of multiple interactions made between the peptide macrocycle and Mpro, likely involving P11GM4, Y9GM4 and H3GM4, resulting in tight binding in a catalytically non-productive manner.

SARS-CoV-2 Mpro cleaves after a conserved glutamine residue (P1) in substrates with the following sequence motif: (P2, L/F/V/M (hydrophobic))(P1, Q)↓(P1′, S/A/G/N (small)), where ↓ represents the cleavage site40,42. Interestingly, rather than a glutamine (Q), GM4 has a histidine (H3) at the analogous ‘P1’ position. To investigate the contribution of the H3/P1 residue for activity, we evaluated peptides bearing Q, A or E residues at the analogous position. Among these, GM4H3Q exhibited a sixfold stronger binding affinity (KD = 0.86 nM) and fivefold more potent inhibitory activity (IC50 = 10 nM; Table 1, Fig. 2 and Supplementary Fig. 5). By contrast, GM4H3A manifested no inhibition, and inhibition of GM4H3E was observed only with the highest tested concentration (25 µM). These results are consistent with the known substrate selectivity of Mpro at the P1 position40.

In the case of the P1′ residue, Mpro prefers substrates with small residues, such as G, A and S, although N is also tolerated40. Strikingly, GM4 has γ14 at the P1′ position, likely reflecting the compact size of the γ1 four-membered-ring main-chain residue. In support of this, GM4γ14A (γ1 → A) exhibited a comparable inhibitory activity to that of GM4 (IC50 of GM4γ14A = 10 nM), and GM4γ14N was 20-fold less active (IC50 = 1,000 nM). Despite the almost equivalent Mpro inhibition activities of GM4 and GM4γ14A, the serum stability of GM4 is 12-fold higher (t1/2 = 126 h) than that of GM4γ14A (Fig. 2b). Furthermore, the most potent mutant GM4H3Q bearing γ14 at P1 also exhibited a comparable serum stability (t1/2 = 82 h). These results show how the introduction of a nonstandard amino acid at the P1′ position can prolong the serum stability to non-targeted protease-catalysed hydrolysis while maintaining potency against the targeted protease (Mpro in our case).