De novo design of potent and resilient hACE2 decoys to neutralize SARS-CoV-2

A decoy to neutralize SARS-CoV-2 Many efforts to develop therapies against severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) are focused on the interaction between the spike protein, which decorates the surface of the virus, and its host receptor, human angiotensin-converting enzyme 2 (hACE2). Linsky et al. describe a de novo design strategy that allowed them to engineer decoy proteins that bind to the spike protein by replicating the hACE2 interface. The best decoy, CTC-445, bound with low nanomolar affinity, and selection of viral mutants that decrease binding is unlikely because this would also affect binding to hACE2. A bivalent version of CTC-445 bound even more tightly, neutralized SARS-CoV-2 infection of cells, and protected hamsters from a SARS-CoV-2 challenge. The stable decoy has the potential for respiratory therapeutic delivery. Science, this issue p. 1208

S ince its emergence as a global pandemic in December of 2019, severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) has caused millions of COVID-19 cases. The need for effective strategies to prevent and treat the disease remains urgent (1). There are multiple ongoing efforts to develop prophylactics and therapeutics using various approaches (2) such as vaccination (3), traditional protein engineering (1,4,5), de novo protein design (6), and small-molecule drug discovery (7). A challenge is that the high mutational rate of positive sense singlestrand RNA (+ssRNA) viruses (8)(9)(10) can often lead to viral escape (11), which could compromise the efficacy of many SARS-CoV-2 therapeutics under development. Several mutations have already occurred in the S protein of SARS-CoV-2 in the infected population (12,13). Deepsequencing studies of the receptor-binding domain (RBD) have shown that simple mutations can enable the virus to escape known netralizing antibodies or to increase its binding affinity for human angiotensin-converting enzyme 2 (hACE2) (14,15), the membrane protein that the virus exploits to gain entry into the cell. There is thus a pressing need to develop new therapeutics that can be more resistant to SARS-CoV-2 mutational escape.
Traditional approaches to combatting viruses (e.g., vaccination and monoclonal antibodies) rely on molecules interacting with the pathogens in a way that is fundamentally different from how the pathogen engages with its cellular targets (16,17). Viruses can be selected to evade neutralization, undergoing protein mutations that prevent recognition by the neutralizing molecules (e.g., antibodies) while preserving viral fitness. To address these challenges, we have developed a computational protein design strategy that enables the rapid and accurate design of hyperstable de novo protein "decoys" that replicate the protein receptor interface to which a virus binds to infect a cell. The decoys can achieve a similar or even higher affinity than the original protein receptor by stabilizing the binding interface. Therefore, at an optimal concentration, the decoys can outcompete viral interaction with the cell.
SARS-CoV-2 invades host cells in a two-step process (18)(19)(20). The S protein RBD attaches to the cell by binding to hACE2, a membraneassociated protein, triggering protease-mediated fusion with the cell membrane (21). The process is similar to the beta-coronaviruses HCoV-NL63 and SARS-CoV-1, which also target hACE2 for cellular entry (22). In principle, inhibiting the viral interaction with hACE2 should prevent infection. We applied our design strategy to engineer, validate, and optimize de novo hACE2 decoys to neutralize SARS-CoV-2 infection ( Fig.   1J and fig. S1). The design of the decoys started by identifying the structural motifs that form the hACE2 binding interface with the SARS-CoV-2 RBD. We based our effort on three publicly available structures of hACE2 in complex with the RDB of the S protein for SARS-CoV-1 (PDB: 6CS2) and SARS-CoV-2 (PDBs: 6VW1 and 6M17) (23)(24)(25). Four discontiguous binding elements were identified (Fig. 1A) and the three largest interacting motifs were selected to build the de novo decoys: two long alpha helices (H1 and H2) and a short beta hairpin (EE3) (Fig. 1A and fig. S2). To generate molecules that are biologically inert for humans, our computational design strategy avoided incorporating elements of hACE2 that are known (or predicted) to be biologically active, such as the catalytic site. Inspired by recent developments in the design of de novo structural elements (26)(27)(28)(29), we built new disembodied de novo secondary structure elements tailored to support the target structural elements in a way that is both compatible with globular folding and would stabilize the binding interface ( Fig. 1B and materials and methods). Then, in a strategy similar to the design of Neoleukin-2/15 (Neo-2/15) (26, 30), a combinatorial design approach based on Rosetta's "protein_mimic_designer" was used to generate multiple fully connected protein topologies containing all of the desired structural and binding elements (26). The design of the protein decoys was constrained to fully preserve (intact up to each amino acid's conformation) the target binding interface (Fig. 1, A and B,  and fig. S2) so that the de novo proteins would be resilient to viral mutational escape. Rosetta (31) was then used to generate amino acid sequences predicted to fold into the target structures, and the designs were evaluated with an automatic filtering pipeline based on nine computational parameters, including predictions of smooth folding funnels into a stable native state (Fig. 1, C and D) (32).
Approximately 35,000 computational ACE2 decoys were generated, and the top-ranking 196 designs (see the materials and methods) were selected for experimental testing for binding to SARS-CoV-2 RBD using yeast display (Fig. 1E). With no further optimization, the design CTC-445 showed strong (nanomolar) and specific binding for SARS-CoV-2 RBD (Fig. 1E,  fig. S3, and materials and methods). CTC-445 is a 160-amino acid protein comprising 18 of the natural amino acids; it does not contain cysteine or tryptophan residues. It exhibited 10-fold weaker binding affinity for SARS-CoV-2 than did hACE2 [disassociation constant (K D )3 57 nM, K D~3 1 nM, respectively; table S1] and, as a result, CTC-445 was a weak competitor of SARS-CoV-2 RBD binding to hACE2 [median inhibitory concentration (IC 50 @ hACE2[0.4nM] ) = 1.7 mM; Fig. 1I). We determined that low potency of CTC-445 was due to a certain degree 2 of 7 ...

Experimental characterization CTC-445
Directed evolution/rational selection 50 CTC-445 variants N C (gray) and its binding motifs (H1 19-52, orange; H2 55-84, green; EE3 346-360, blue) in complex with SARS-CoV-2 RBD (pink). Three starting structures were simultaneously used as targets (see main text); 6VW1 is shown. (B) De novo secondary structure elements (magenta) were computationally generated to stabilize H1, H2, and EE3. Seven combinations of secondary structure elements were considered. Circles are a-helices, triangles are b-sheets, filled circles are helices oriented forward, and empty circles are helices oriented backward. We used Rosetta to generate fully connected backbones (using the "protein_ mimic_designer" algorithm) and amino acid sequences predicted to fold into the target structure. In all cases, the binding interface of ACE2 with the SARS-CoV-2 RBD was preserved intact (see the materials and methods). Timewise, green indicates phases that we believe were performed optimally, red indicates those that can potentially be avoided in future efforts, and yellow indicates phases that can potentially be expedited by using more advanced and/or automated methods for gene synthesis, cloning, and high-throughput screening.  Fig. 2A), a bivalent version of CTC-445.2, had an~10-fold improvement in binding affinity for both SARS-CoV-2 RBD (K D~3 .5 nM; table S1) and SARS-CoV-1 RBD (K D~5 87 nM; Fig. 2C and table S1), and a similar increase in its ability to compete with hACE2 binding to SARS-CoV-2 RBD (IC 50 @ hACE2[0.4nM]~7 00 pM; Fig. 1I). A trivalent version of CTC-445.2 resulted in even higher (picomolar) binding affinity and a matching hACE2 competition potency (K D 270 pM, IC 50 @ hACE2[0.4nM]~1 0 pM; fig. S10 and table S1). In a cross-reactivity binding assay containing >21,000 human proteins, we confirmed that CTC-445.2d bound to the SARS-CoV-2 RBD with high selectivity (fig. S11 and materials and methods).
Single-particle cryo-EM structures of CTC-445.2 in complex with the SARS-CoV-2 S trimer showed that the de novo decoy is capable of simultaneous binding to all three RBDs of the SARS-CoV-2 trimeric S protein, both in the "up" and "partially down" RBD conformations (Fig. 3, A to D, and fig. S12). To accurately model the CTC-445.2-RBD interactions, we used focused classification and local refinement on the subset of particles that showed CTC-445.2 bound to a partially down RBD, which yielded a 4.1-Å map with improved CTC-RBD features relative to CTC-RBD regions on the up RBDs (Fig. 3, A to D, and figs. S12 and S13). The computationally derived model of CTC-445.2 closely matched the cryo-EMdetermined structure [Ca root mean square deviation (RMSD) = 1.1 Å], with minor differences observed in the N-terminal EE3 and H2 helix (Fig. 3, E to H). As designed, the binding interface of the SARS-CoV-2 RBD with CTC-445.2 closely mirrored the target hACE2 interface. We used site saturation mutagenesis (SSM; see the materials and methods) (33,34) to explore the effect of single-amino acid substitutions in CTC-445.2 on its binding to the SARS-CoV-2 RBD (Fig. 3, I and J). The experiment showed that mutations in the core of the design are disallowed, and mutations in surface or exposed residues are generally tolerated (Fig. 3, I and J). The SSM experiment also revealed that there is room to further improve the affinity of the protein by introducing mutations in the binding interface (Fig. 3I), although doing so would break the hACE2 structural mirroring of the de novo decoy.
We also performed an SSM experiment for the SARS-CoV-2 RBD binding interface to compare the effect of single-amino acid substitution on binding to hACE2 or CTC-445.2. As predicted, the effects of~1700 SARS-CoV-2 RBD mutations showed a strong correlation between binding to hACE2 and CTC-445.2 (R 2 = 0.84, Pearson's r = 0.92; Fig. 4 and fig. S14), highlighting the decoy's intrinsic resiliency to mutational escape. At low target concentrations (100 pM), CTC-445.2 had a large binding advantage over ACE2 for many of the RBD mutations ( fig. S14), likely a result of both its higher stability and smaller size. Although CTC-445.2 was resilient to viral mutations in the RBD-binding interface, we observed some decoy-binding-weakening mutations that had a lesser effect on hACE2 binding. Therefore, viral mutational escape might still be possible if multiple (decoy-bindingweakening) RBD mutations are combined.
The high and specific binding affinity of the optimized de novo protein decoys translated into effective and specific in vitro neutralization of SARS-CoV-2 viral infection (Fig. 5). In vitro, the presence of the de novo decoys had no impact on mammalian cell viability (Fig. 5A  and fig. S15) or the enzymatic activity of hACE2 ( fig. S16)  The deep mutational scanning revealed that there is still room to further improve the binding affinity of CTC-445.2, including mutations in the binding interface that in principle could afford higher potency and selectivity at the cost of compromising the decoy's mutational escape resiliency (see Fig. 4).
neutralize viral infection in in vitro systems of cell infection. Briefly, in a vesicular stomatitis virus (VSV) pseudovirus system expressing the SARS-CoV-2 S protein, the decoys specifically protected human embryonic kidney (HEK) 293T cells overexpressing hACE2 from infection ( fig. S15). The decoys also were able to fully neutralize infection by SARS-CoV-2 (SARS-CoV-2 nanoLuc; see the materials and methods) in the lung epithelial cell line Calu-3 expressing both ACE2 and the transmembrane protease serine 2 (TMPRSS2) (35,36) [median effective concentration < 5 nM at a multiplicity of infection (MOI) of 1.0; Fig. 5A].
In an in vitro time-of-addition assay using the Vero E6 cell line, CTC-445.2 and CTC-445.2d were most effective at neutralizing SARS-CoV-2 infection when continuously present in the cell media throughout the full course of infec-tion (as opposed to only before or after infection; Fig. 5A and figs. S16 to S18), confirming that their mechanism of viral inhibition is extracellular neutralization of the virus.
To determine the potential of our molecules to be used as respiratory-delivered therapeutics, we intranasally administered a single dose of CTC-445.2d to Balb/c mice (100 mg dose of CTC-445.2d in a 30-mL droplet) and observed the presence of the fully functional decoy for >24 hours in the lungs and respiratory tract of mice (Fig. 5B and fig. S19). A 14-day course of daily CTC-445.2d intranasal administration in mice (100 mg of CTC-445.2d in a 30-mL droplet) was well tolerated, causing no adverse effects (Fig. 5B). In a Syrian hamster model for SARS-CoV-2 infection, a single prophylactic intranasal dose of CTC-445.2d (560 mg of CTC-445.2d in a 100-mL droplet) administered 12 hours before the viral challenge afforded 100% survival from a lethal SARS-CoV-2 challenge (5 × 10 5 plaqueforming units of SARS-CoV-2; Fig. 5C). Specifically, by day 7, all control animals that received the viral challenge but not CTC-445.2d (n = 7) exhibited severe distress and required euthanasia. By contrast, hamsters that received a single dose of CTC-445.2d 12 hours before challenge all survived (n = 8), with modest weight loss and few or no clinical signs of distress ( Fig. 5C and table S5).
Other recent protein-engineering efforts have generated neutralizing proteins characterized by extremely high binding affinities for SARS-CoV-2, with K D s ranging from low nanomolar to femtomolar [e.g., mAb 2B04 (47); LCB1 (6); and the nanobody Nb6 (48)]. Nevertheless, the de novo decoy's resilience to viral escape is a distinctive feature of our design strategy ( Fig. 4 and figs. S14 and S20). A possible shortcoming is that a decoy's requirement to replicate a natural binding interface can intrinsically limit the maximum binding affinity attainable. However, we have demon-strated that the binding affinity (and potency) of the de novo decoys can be increased both by further sequence optimization (e.g., CTC-445.3d; fig. S21) or through avidity, allowing our trivalent decoy CTC-445.2t to reach the picomolar affinity range (Fig. 3I and fig. S10). It is possible that avid versions of CTC-445.2 coupled with more refined linkers (rigid and with proper spacing for binding simultaneously to multiple RBD subunits) might lead to larger increases in binding potency. We demonstrate rapid design of a therapeutic lead; further speed improvements to our pipeline are theoretically attainable, for example by using high-throughput experiments to rapidly select and optimize the designs (Fig. 1G).  Weight data shown are for the remaining mice (n = 18,15,12,9,6, and 3 at days 1, 2, 4, 8, 11, and 14, respectively). No significant weight loss or lung abnormalities were observed. Error bars indicate the standard deviation. (C) In vivo Syrian hamster SARS-CoV-2 challenge. Left: Body weight measurements through day 10 for unchallenged hamsters (n = 5, red) compared with SARS-CoV-2challenged hamsters treated either with a single dose of CTC-445.2d (day 0 at -12 hours; n = 8, orange) or PBS (day -1, day 0 at -12 hours, day 1, and day 2; n = 7, gray). Right: Survival plot. Hamsters were euthanized when they displayed clinical signs of distress according to protocol clinical scoring criteria (see the materials and methods). At the end of the experiment, all hamsters treated with the de novo decoy CTC-445.2d survived, exhibiting moderate weight loss, whereas hamsters treated with vehicle did not survive past day 7 because of severe weight loss and other complications from the viral infection (see table S5).