Structural basis of ribosomal frameshifting during translation of the SARS-CoV-2 RNA genome

We captured a 0 frame, preframeshift ribosomal complex by introducing a stop codon in place of the second codon of the slippery site (U_UUA_AAC to U_UUA_UAA) (Fig. 1A) and adding mutant eukaryotic release factor 1 [eRF1 (AAQ)] that is unable to release the nascent polypeptide. Translating complexes were prepared in an in vitro translation reaction using an in-house–generated rabbit reticulocyte lysate (RRL) system that supported efficient frameshifting in the previously reported range of around 50% (19) according to dual luciferase experiments (see methods). The ribosomes were programmed with mRNA encoding an affinity tag and harboring a region of the SARS-CoV-2 genome that encodes proteins Nsp10 (C terminus), Nsp11, and most of Nsp12. Western blotting showed that when using the wild-type (WT) RNA template, frameshifting was efficient, whereas the stop codon mutation prevented frameshifting and led to ribosome pausing. This effect was further enhanced when eRF1 (AAQ) was present in excess over endogenous WT eRF1 (Fig. 1B).

The cryo-EM three-dimensional (3D) reconstruction of ribosome–nascent chain complexes affinity-purified from the reactions supplemented with eRF1 (AAQ) revealed two distinct ribosomal complexes captured in the process of translating the slippery sequence (figs. S1 and S2). One represented a termination complex that contained the ATP-binding cassette transporter 1 (ABCE1), which is known to be involved in termination and recycling together with mutant eRF1 interacting with the stop codon (fig. S3). The second reconstruction resolved translating 80S ribosomes containing bound P- and E-site tRNAs (fig. S2). This reconstruction at 2.2-Å resolution allowed us to build the most accurate structure of a mammalian 80S ribosome so far and directly visualize many protein and virtually all rRNA modifications identified for the human ribosome based on quantitative mass spectrometry and as interpreted in a recent human ribosome structure (20, 21), consistent with the complete conservation of all modified residues between rabbit and human ribosomal RNAs (rRNAs) (figs. S4 and S5; and tables S1 to S3). Importantly, this reconstruction also featured additional density at the entrance to the mRNA channel suggestive of a structured RNA, which, after focused classification, revealed a prominent density for a complete 3′ frameshifting stimulatory pseudoknot at the entry of the mRNA channel on the 40S subunit (Fig. 1, C and D). The resolution of this reconstruction ranged from 2.4 Å at the core of the ribosome to ~7 Å at the periphery, where the most flexible regions of the pseudoknot are located (figs. S2 and S6). Based on the high-resolution maps that allowed visualization of the codon-anticodon interactions and modifications in the tRNA (Fig. 1E and fig. S6, A and B), we could unequivocally determine that a Phe-tRNA(Phe) was bound at the P-site (22). The mRNA does not adopt any unusual structure in the A-site of the ribosome as was observed for the HIV-1 frameshifting sequence visualized on the bacterial ribosome (23). This implied that the ribosome is paused by the downstream pseudoknot located at the entrance to the mRNA channel such that the P-site tRNA interacts with the UUU codon just prior to the first codon, UUA, of the slippery site (Fig. 2A).

The observation that the pseudoknot acts as an obstacle to slow down translation as the ribosome approaches the slippery site is mechanistically reasonable. Because the pseudoknot is a stable structural element in the mRNA, it will resist unfolding and consequently generate a back-pull on the viral RNA, resulting in an increased chance of −1 frameshifting as the tRNAs are translocated. A pause in translocation at a codon that precedes the slippery site, characterized by a >10 times longer occupancy prior to the slippage event, was observed in an analogous case of heptanucleotide −1 frameshifting on the bacterial dnaX gene using single-molecule experiments (24). According to this model, it would be anticipated that a further round of translocation results in unwinding of Stem 1 of the downstream stimulatory pseudoknot structure. Consistently, in our structure of the eRF1 (AAQ)–bound ribosome that advanced one codon further along the mRNA, no clear secondary structure is visible at the entrance to the mRNA channel because the mRNA now becomes disordered at this position (figs. S1 and S3, A and B).

To investigate the slowdown of translation on the WT slippery sequence, we performed disome footprint profiling, a method that identifies translational pause sites through the analysis of transitory ribosome collisions (25–27) (see methods). Notably, recent studies using conventional ribosome profiling methodology reported a lack in monosome footprint coverage across the frameshifting region on the SARS-CoV-2 RNA (11, 28), possibly because ribosomes in this area became trapped in temporary collisions. Moreover, the highly structured pseudoknot at the entry to the mRNA channel would likely preclude efficient trimming by ribonuclease I (RNase I), the enzyme used for footprint generation, further reducing efficient monosome footprint capture. Using a modified nuclease treatment protocol (see methods) that recovered monosome footprints from the frameshift region (Fig. 3, A and C), our experiments revealed that ribosome collisions occur as a result of ribosomal pausing at the same position that is observed in the structure of the pseudoknot-engaged ribosome (Fig. 3, B and D). Apparently, although the base substitutions creating a stop codon in the 3′ adjacent slippery site did not change the features of pausing, they increased the dwell time of the ribosomes at the pause site sufficiently to allow visualization in the cryo-EM experiment.

The results of our disome profiling experiments prompted us to structurally investigate disomes by cryo-EM. We were able to visualize the pseudoknot-paused ribosome followed by a closely trailing ribosome. Upon focused refinement, we obtained a high-resolution (3.1 Å) structure of the trailing ribosome in a rotated state (fig. S1). In congruence with our estimated positioning of the ribosomes in disome profiling (Fig. 3D), the purine-pyrimidine pattern of codon-anticodon pairs in the structure of the colliding ribosome revealed that the pause occurs with CCC and AUG triplets in the P- and A-sites, respectively (Fig. 3C).

The intermediate local resolution (5 to 7 Å) of the cryo-EM map in the area of the pseudoknot allowed us to visualize the overall fold of the RNA and readjust its previously predicted secondary structure (14–17, 19) (Fig. 1, C, D, and F). The stimulatory pseudoknot forms an H-type pseudoknot with Stem 1 and Stem 2 coaxially stacked on top of each other to form a quasi-continuous helix, whereas Stem 3 stands out almost perpendicular to this plane (Figs. 1D and 2B). This corkscrew-like formation provides a bulky and well-structured obstacle wedged at the mRNA entry channel, which has the potential to resist unwinding by the helicase activity of the ribosome and generate tension on the upstream mRNA up to the decoding center. Stem 1 of the pseudoknot forms a 9–base pair helix that is GC rich at the bottom (Fig. 1F). The penultimate nucleotides of the “spacer region” before Stem 1 are located at the mRNA entry tunnel, where they interact with several basic residues in the C-terminal domain of uS3 on one side and are supported by uS5 from the other, with an additional weak contact contributed by the C-terminal end of eS30. uS3 and eS30 are primary components of the ribosome helicase, and uS5 has been proposed to be a component of the ribosomal helicase processivity clamp at the mRNA entry site (29, 30). The observed distance between the P-site UUU codon and Stem 1 of the pseudoknot underscores the critical dependence of the frameshifting efficiency on the length of the spacer region (31). Translocation to the next codon would place the frameshifting codon UUA into the P-site, with a simultaneous increase in the tension of the mRNA and unwinding of the GC-rich base of Stem 1 upon entering the mRNA entry channel, comparable to the situation when the ribosome proceeds to the engineered stop codon, as observed in our eRF1 (AAQ)–stalled structure (fig. S3).

The pseudoknot structure also reveals a hitherto unobserved and possibly unappreciated role for the distal site of the mRNA entrance channel in helicase activity. Although mRNA unwinding studies outside the mRNA entrance channel have so far implicated only a helix in the C-terminal domain of uS3 (32), we noticed that Loop 1 of the pseudoknot contacts the N-terminal domain of uS3 as well as the C-terminal tail of eS10 (Fig. 2B and fig. S6D), whereas the flipped-out base G13486 in this loop forms specific interactions (Fig. 2B). Furthermore, because the pseudoknot is located at the entry to the mRNA channel, helix h16 of the 18S rRNA is noticeably pushed outward owing to a direct contact with the minor groove of Stem 1 (Fig. 2B and fig. S7A). Because the pseudoknot wedges between the head and the body of the small ribosomal subunit, it would restrict their relative motions that need to take place during translocation. This is consistent with the studies on dynamics of coronavirus frameshifting, which revealed that the mechanism of −1 frameshifting involves restriction of small subunit head motion (33).

The structure also reveals another key aspect of the architecture of the pseudoknot as the ribosome encounters it. The start of the pseudoknot is shifted relative to the predicted secondary structure (14–17, 19) by two nucleotides. The two opposed nucleotides, which were assumed to base pair with Stem 1, are actually forming the start of Stem 3 by pairing with bases predicted to be in the single-stranded linker 2 (Fig. 1F and fig. S7, B and C). Our cryo-EM density reveals that Loop 3 accommodates a total of four nucleotides, three of which were originally attributed to Stem 2. Thus, we observe that Loop 3 is shifted and expanded relative to the initially predicted secondary structures (14–17, 19).

To functionally support our structural findings and confirm the nature and specificity of the pseudoknot interactions, we performed structure-guided mutagenesis experiments using dual luciferase reporter assays in human embryonic kidney (HEK) 293T cells (see methods) and monitored the frameshifting efficiency relative to the WT (Fig. 2C). Mutation of G13486 of Loop 1 to another purine reduced the frameshifting efficiency to 30% of the WT level, and mutation of this base to a pyrimidine further reduced frameshifting to 15%. As expected from our structural data, deletions of the nucleotides of the spacer regions also had a deteriorating effect on frameshifting. Loss of Loop 1 entirely abolished frameshifting. Deletion of a single nucleotide of Loop 3 in agreement with its proposed role in forming the base-pairing interactions diminished the frameshifting rate to 25% of the WT level. Loss of the entire Loop 3 reduced frameshifting to 10% of WT levels.

In SARS-CoV-2, the 0 frame stop codon is located five codons downstream of the frameshift site and is a constituent of Stem 1. The placement of the stop codon in such proximity to the frameshift site is a common feature in coronaviruses, and its presence in a critical region of the stimulatory pseudoknot prompted us to probe the effect of the distance of the 0 frame stop codon on frameshifting. To this end, knowledge of the 3D structure of the pseudoknot helped us to confidently manipulate the stop codon without hampering pseudoknot formation. We introduced mutations to incrementally extend the stop codon from the WT position and to completely remove the occurrence of a stop codon in the 0 frame (Fig. 2D and fig. S8). Whereas introducing a stop codon six nucleotides downstream of the WT position only marginally decreased the frameshifting rate (98% of WT), a stronger attenuation was observed when the distance of the stop codon was increased to 15 nucleotides from the WT stop (80% of WT). Finally, removal of the stop codon by two different point mutations led to a reduction of frameshifting efficiency to 50% of WT levels. To test whether reduced ribosomal loading rescues the effect of stop codon removal, we analyzed the frameshifting efficiency in the context of weaker initiation codons such as CUG and AUU (Fig. 2D). These constructs led to a 45% rescue of the reduction in frameshifting compared with stop codon mutants initiating at an AUG start.

Taken together, these observations suggest that the stop codon position plays an important role in maintaining optimum frameshift efficiency. We propose that the stop codon serves to prevent the closely trailing ribosome from encountering a viral RNA that was unfolded by the leading ribosome. In this case, upon encountering a stop codon, termination and subunit disassembly will occur, which will provide an opportunity for the pseudoknot to refold without the constraints of the mRNA channel (see Conclusions). According to this model, although the WT stop codon will make the frameshifting efficiency less sensitive to ribosome loading in the “no-frameshifting” scenario, the frameshifting events that occur after a −1 frameshift will nevertheless be more likely when the ribosomes are spaced further apart. Our measurements of the efficiency of frameshifting for the WT sequence in the context of different rates of translation initiation are in agreement with this hypothesis (fig. S9). This mechanism, consistent with our biochemical data, increases the efficiency of frameshifting to the levels required by SARS-CoV-2 and may be used by viruses in general when high-efficiency frameshifting is required.

Notably, in the reconstruction of the paused translating ribosome, the nascent chain that corresponds to the viral polyprotein was visible along the entire length of the ribosomal exit tunnel (Fig. 4A). The density corresponds to the C-terminal region of Nsp10, which is the activator of the viral proofreading exonuclease and N7-methyltransferase Nsp14 (34, 35), and then (depending on the frameshifting event) continues as either the viral RNA-dependent RNA polymerase Nsp12 (6) or as protein Nsp11, whose function is still unknown (Figs. 1A and 4B). The nascent chain makes several specific interactions with the ribosomal tunnel, one of which is at the constriction site where Arg⁴³⁸⁷ of Nsp10 interacts with A¹⁵⁵⁵ of the 28S rRNA [corresponding to A¹⁶⁰⁰ in humans, numbering according to PDB 6EK0 (36)] and is stabilized by the preceding Leu⁴³⁸⁶ (Fig. 4C). Notably, these two amino acids are highly conserved across multiple coronaviruses (Fig. 4G), although they are located in the unstructured C-terminal region of Nsp10 and therefore considered not to be important for the fold of the protein (37).

Further down the tunnel, the C-terminal end of Nsp10 adopts a partially folded zinc finger motif (Fig. 4, D and E), which, upon superposition, reveals similarity with the corresponding fully folded C-terminal domain previously observed in the crystal structure of SARS-CoV-1 Nsp10 (37). Trp⁴³⁷⁶, which is located between the two pairs of cysteines that form the zinc finger, stacks with A²²⁶¹ (A²⁴¹⁸), an interaction that might serve to promote the change of nascent chain direction and facilitate folding of the zinc finger at the end of the exit tunnel. Cotranslational events, such as insertion of a transmembrane domain at the exit of the ribosomal tunnel, were shown to promote −1 ribosomal frameshifting in alphaviruses (38).

To investigate whether the observed contacts between the nascent chain and the ribosomal tunnel are specific and whether these interactions and cotranslational folding of Nsp10 might play a role in modulating the frameshifting process, we used our dual luciferase reporter assay to measure the frameshifting efficiency of WT and mutant nascent chain sequence constructs. Because our measurements in HEK293T cells did not reveal an appreciable change of frameshift efficiency, we carried out the same experiments in vitro using RRL to monitor the effects in a single mRNA setup. Replacement of the entire nascent chain with an unrelated sequence leads to a 35% increase in frameshifting (Fig. 4F). Importantly, this effect was provoked by the change in peptide sequence and not simply by the loss of the 5′ attenuator loop, given that a reporter containing silent attenuator loop mutations resulted in only a slight increase in frameshifting (Fig. 4F). Mutation of the Leu⁴³⁸⁶ and Arg⁴³⁸⁷ to alanine led to a considerable (30%) increase in frameshifting (Fig. 4, F and G), implying that these nascent chain interactions with the ribosomal exit tunnel play an important role in regulating frameshifting levels, possibly mechanistically akin to the well-studied SecM stalling system in bacteria (39), where it was shown that cotranslational folding and the translocon-induced mechanical force can rescue the stall induced by interactions between the nascent chain and the ribosomal tunnel (40). These observations also suggest that any cellular nascent chain factors (41, 42) might influence the rate of frameshifting.

The sensitivity of the coronavirus to the finely controlled frameshifting levels (13) may present an opportunity to develop compounds that interfere with the frameshifting process and thus inhibit replication of the virus. Using computational modeling and reporter assays, compounds that have been predicted to bind the pseudoknot and inhibit SARS-CoV-2 frameshifting were described (19, 43) but never tested with respect to their ability to inhibit viral replication. Furthermore, the fluoroquinolone compound merafloxacin was recently reported to also inhibit −1 frameshifting efficiency of SARS-CoV-2 and other betacoronaviruses (44). To demonstrate that the inhibition of frameshifting is a plausible strategy for drug development, we compared two of the previously described compounds with respect to their ability to reduce viral levels in infected African green monkey VeroE6 cells (fig. S10 and methods). Our experiments demonstrate that merafloxacin is a better candidate compound because it showed a concentration-dependent inhibition of frameshifting, whereas, contrary to earlier reports (19, 43), the small-molecule ligand MTDB did not specifically inhibit frameshifting under our experimental conditions (fig. S10). The two compounds showed no cellular toxicity and resulted in a three to four orders of magnitude reduction of SARS-CoV-2 titer, with a half-maximal inhibitory concentration (IC₅₀) of 48 μΜ for MTDB and an order of magnitude higher efficacy for merafloxacin, with an IC₅₀ of 4.3 μΜ (fig. S10). Because MTDB did not appear to affect frameshifting in our reporter construct experiments in vitro and in vivo, it is possible that it inhibits SARS-CoV-2 replication by a different mechanism. Although the potency range for these compounds is not what would be expected from potential drug candidates, it nevertheless provides a starting point for high-throughput screening and establishes that frameshifting is a viable target for therapeutic intervention against SARS-CoV-2.

Our results provide a mechanistic description of frameshifting that occurs during translation of the SARS-CoV-2 genome and reveal the features that may be exploited by the virus to finely control the stoichiometry of viral proteins at different stages of infection (Fig. 5). Interfering with the frameshifting process at the level of nascent chain interactions with the ribosomal tunnel, at the level of RNA folding that leads to the formation of the frameshift stimulatory pseudoknot, or to perturb the interactions between the pseudoknot and the mRNA channel represent viable strategies in our search for new drugs against SARS-CoV-2, the virus that is currently causing the global COVID-19 pandemic. Our results will also be useful for understanding the mechanism of programmed ribosomal −1 frameshifting (4), including that used by many other medically important viruses.

We thank A. Jomaa for advice on cryo-EM data processing, N. Casey for support with technical resources, and A. Picenoni for help with the grid preparation. We are indebted to the ETH scientific center for optical and electron microscopy (ScopeM) for access to electron microscopes and, in particular, to M. Peterek and D. Boehringer. We thank J. Guo for helpful discussions and for providing the merafloxacin compound. Funding: This work was supported by the Swiss National Science Foundation (SNSF) to N.B. and V.T. (grants 31003A_182341 and 310030_173085) through the National Center of Excellence in RNA and Disease (project funding 182880) to N.B., V.T., and D.G.; by the Federal Ministry of Education and Research, Germany (BMBF; grant RAPID, #01KI1723A) to A.K. and V.T.; and by the Irish Research Council Advanced Laureate (IRCLA/2019/74) to J.F.A., who thanks G. F. Fitzgerald for funds from Carbery Group Ltd. Author contributions: P.R.B. and N.B. initiated the project and, together with J.F.A., designed the experiments. P.R.B. carried out biochemical experiments and sample preparation. A.S. and P.R.B. prepared grids. A.S. carried out data collection and calculated EM maps. A.S. and M.L. performed molecular model building and refinement. A.S., M.L., P.R.B., and N.B. interpreted the structure. P.R.B., A.S., M.L., and N.B. drafted the manuscript. A.S. prepared the figures. G.L. and K.M.O. performed dual luciferase reporter assays. A.M. prepared and purified the drug-like ligand MTDB. A.K. and V.T. formulated experiments to test viral infected cells. A.K. carried out testing of ligand on viral-infected cells. R.M., R.D., and D.G. designed, carried out, analyzed, and interpreted ribosome profiling experiments. All authors contributed to the final version of the manuscript. Competing interests: The authors declare no competing interests. Data and materials availability: The structure and cryo-EM map for the high-resolution reconstruction is available in the Protein Data Bank (PDB) as PDB ID 7O7Y and in the Electron Microscopy Data Bank (EMDB) as EMD-12756, respectively. The structure refined into the further classified particle set reconstruction and the corresponding maps are available as PDB-7O7Z and EMD-12757, respectively. The structure of the ribosome bound with eRF1 (AAQ) and ABCE1 reconstruction and the corresponding maps are available as PDB-7O80 and EMD-12758, respectively. The structure of the colliding ribosome reconstruction and the corresponding maps are available as PDB-7O81 and EMD-12759, respectively. The structure of the composite map of the disome is available in the EMDB as EMD-12760. Ribosome profiling data have been deposited at GSE167421. This work is licensed under a Creative Commons Attribution 4.0 International (CC BY 4.0) license, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. To view a copy of this license, visit https://creativecommons.org/licenses/by/4.0/. This license does not apply to figures/photos/artwork or other content included in the article that is credited to a third party; obtain authorization from the rights holder before using such material.