We previously generated two SARS-CoV-2 spike proteins that each contain six RBD changes that were detected during persistent infection of an immunocompromised individual infected with a SARS-CoV-2 strain containing the D614G_S mutation (14-16). This individual received treatment with REGN-COV2 (17, 18), but several of the RBD substitutions had occurred even prior to administration of this therapeutic antibody cocktail (14-16). Lentivirus pseudotypes bearing these spike proteins, denoted day 146* and day 152* (Fig. 1A and table S2), were refractory to neutralization by V_H3-53-heavy chain gene-derived neutralizing antibodies, a potent class of neutralizing antibodies that have been repeatedly isolated from convalescent donors (19-25). These pseudotypes were also resistant to neutralization by components of REGN-COV2 (17, 18) and by polyclonal immunoglobulins (IgG) purified from the serum of COVID-19 convalescent donors (14). Substitutions in the day 146* and day 152* spike proteins, noted in samples sequenced from this individual in the spring and summer of 2020, foreshadowed those in currently circulating VOCs at three positions: N501_RBD, E484_RBD, and T478_RBD (Fig. 1A). The day 146* and day 152* spike proteins also contain substitutions that are not in current dominant strains but could have serious effects if acquired. For example, the S494P_RBD substitution is a therapeutic antibody (LY-CoV555) escape mutation (26) that as of September 27^th, 2021, was present in over 12,000 human-derived SARS-CoV-2 sequences on public research databases (GISAID) (27). Additionally, the Q493K_RBD mutation, which is found in over one hundred human-derived SARS-CoV-2 sequences as of September 27^th, 2021, on GISAID, confers resistance to multiple therapeutic antibodies [REGN10933, CB6 (LY-CoV016), and LY-CoV555] and V_H3-53-gene-derived antibodies (14, 16, 17, 28).

To determine the impact of their combined mutations on human ACE2 binding, we generated recombinant RBDs for the day 146* and day 152* spike protein mutants. The affinity of the day 152* mutant monomeric RBD for monomeric ACE2 ectodomain was substantially lower (K_D of 2.4 μM) than that of wild-type (Wuhan-Hu-1) RBD (54 nM, consistent with other reports) (3, 29), suggesting that its mutations also compromise ACE2 binding (fig. S1 and table S3). For comparison, the affinity we measured of the SARS-CoV RBD for human ACE2 was 0.26 μM, about nine-fold higher than the affinity for the day 152* RBD (fig. S1 and table S3). The affinity of the day 152* RBD for ACE2 is comparable to that of the RBDs of some bat coronaviruses that are closely related to SARS-CoV-2 and bind human ACE2 (e.g., RaTG13 virus RBD affinity of 3.9 μM) (30). The day 146* RBD, however, had a similar affinity (K_D of 46 nM) for ACE2 as that of the Wuhan-Hu-1 SARS-CoV-2 RBD (fig. S1 and table S3).

We determined the X-ray crystal structure of the day 146* RBD bound to the human ACE2 ectodomain (Fig. 1B, fig. S2, and table S4). This structure is similar to previously determined structures of ACE2/SARS-CoV-2 RBD complexes (2, 3), except we observed contacts between two N-linked glycans on ACE2 (attached to N53_ACE2 and N90_ACE2) and the RBD (fig. S3). Removing the N90_ACE2 glycan, which interacts with the RBD in both copies of the crystal asymmetric unit (fig. S3), increased Wuhan-Hu-1 SARS-CoV-2 and day 146* RBD affinity for ACE2, although the effect was modest (fig. S1 and table S3). This finding is consistent with prior work implicating the N90_ACE2 glycan, which is removed in a human polymorphism (T92I_ACE2), as a barrier to SARS-CoV-2 RBD binding to ACE2 (31, 32).

The N501Y_RBD substitution is found in multiple variants of concern (Fig. 1A); once it surfaced in the immunocompromised individual, it was retained at later time points (14-16). As also shown in a cryo-electron microscopy (cryo-EM) structure of the SARS-CoV-2 spike protein containing the N501Y_RBD substitution bound to ACE2 (33), the side chain of Y501_RBD interacts with Y41_ACE2 and K353_ACE2 with no significant structural change (Fig. 1, C and D). E484_RBD is a critical target of antibodies against SARS-CoV-2 and is mutated in several variants (12, 34, 35). In structures of Wuhan-Hu-1 SARS-CoV-2 RBD bound to ACE2, E484_RBD is near but does not directly contact the receptor (Fig. 1E). In the day 146* RBD/ACE2 complex structure, the K493_RBD sidechain reaches over the RBD surface to recruit the E484_RBD side chain to form a new salt bridge with K31_ACE2 (Fig. 1F). The nearby Y489H_RBD mutation, which removes a polar contact with ACE2, better accommodates repositioning of E484_RBD because the histidine is smaller than the tyrosine sidechain and would avoid potential steric clashes with E484_RBD in this binding mode (Fig. 1, E and F). A second rotamer for residue H34_ACE2 forms new RBD contacts to fill a gap created by the reorganization of local interactions (Fig. 1, E and F). This structural plasticity may explain how the RBD tolerates a surprisingly large number of mutations during intra-host evolution yet retains the ability to bind ACE2 tightly. It is also consistent with the large sequence divergence in the RBD residues that contact ACE2 among SARS-related coronaviruses that share this cellular receptor.

RBD-targeting antibodies can be categorized into classes based on whether they bind an overlapping footprint with ACE2 and recognize only an open or both an open and closed RBD on the spike protein trimer (36). CB6 (equivalent to LY-CoV016 or etesevimab) is a class 1, V_H3-66-derived antibody that blocks ACE2 binding and can only bind the RBD when it is open, and LY-CoV555 (bamlanivimab) is a class 2 antibody that blocks ACE2 binding but can bind both open or closed RBDs (21, 37). LY-CoV016 and LY-CoV555 are used as a cocktail and bind epitopes that partially overlap on the RBM such that both cannot bind simultaneously (21, 37). REGN10933 is a class 1 antibody, and REGN10987 is a class 3 antibody that sterically blocks ACE2 binding but binds the RBM outside the main ACE2 binding site; both are used as a cocktail (REGN-COV2) (17, 18).

Structural plasticity at the RBD-ACE2 interface suggests that the RBD could tolerate many more mutations than found in current variants of concern. We next generated pseudotypes for spike protein variants that contain composite mutations. The Delta variant, which contains the L452R_RBD and T478K_RBD substitutions, has become a dominant strain across the globe (38). We generated pseudotype for the Delta AY.2 variant, which contains the K417N_RBD mutation that is usually found in the Beta variant, and a Delta variant containing the N501Y_RBD, E484K_RBD, and F490S_RBD mutations usually found in Beta, P.1 (Gamma), and C.37 (Lambda) variants (referred to here as “Delta +3”) (Fig. 1A, table S1, and table S2). The set of RBD mutations for the latter strain occurred in deposited sequences from samples collected in Turkey (table S1). We also generated pseudotypes in which we combined spike protein substitutions detected in the immunocompromised host with mutations found in the Beta variant, which we chose because this VOC is highly resistant to antibody neutralization (10, 12, 39). Starting with a day 146* spike protein sequence, which contains an NTD deletion, we incorporated either one (E484K_RBD) or two additional substitutions (E484K_RBD and K417N_RBD); these are referred to as receptor binding mutant-1 (RBM-1) and RBM-2, respectively (Fig. 1A and table S2). Additionally, starting with the Beta variant spike protein sequence, we generated a variant pseudotype that contains two additional mutations associated with immune evasion (L452R_RBD and N439K_RBD) (40, 41). This pseudotype is referred to as RBM-3 (Fig. 1A and table S2). An ACE2-Fc fusion protein neutralized RBM-1, RBM-2, and RBM-3 pseudotypes, suggesting that all entered cells by binding ACE2 (Fig. 2B and fig. S4A).

We tested the activity of therapeutic antibodies against Delta AY.2, Delta +3, RBM-1, RBM-2, RBM-3, and additional variant pseudotypes with known resistance profiles to serve as comparators in the same assay (Fig. 2, A and B, and fig. S4A). LY-CoV555 was the most impacted by escape mutations, followed by CB6 (from which LY-CoV016 is derived) (Fig. 2, A and B, and fig. S4A). The Q493K_RBD mutation conferred absolute resistance to LY-CoV555, generated eightyfold resistance to CB6, and also compromised REGN10933 activity, consistent with previous reports (Fig. 2, A and B, and fig. S4A) (14, 16, 17, 26). In addition to the expected loss of activity of LY-CoV555 and CB6 against Beta and Gamma variants (9, 11, 12, 42), LY-CoV555 and CB6 lost all activity against day 146*, day 152*, RBM-1, RBM-2, and RBM-3 pseudotypes (Fig. 2, A and B, and fig. S4A). While the Delta variant is known to resist neutralization by LY-CoV555 but retain sensitivity to neutralization by CB6/LY-CoV016 (38), Delta AY.2 pseudotype was resistant to both agents (Fig. 2, A and B, and fig. S4A). This is expected because CB6/LY-CoV016 is derived from a V_H3-66 antibody (21), and the additional mutation the Delta AY.2 variant contains with respect to Delta (K417N_RBD) confers resistance to CB6/LY-CoV16 and other members of the V_H3-53/V_H3-66 class of neutralizing antibodies (9, 14, 16, 26, 43). Delta +3 pseudotype, which despite containing six RBD mutations does not contain the K417N_RBD substitution, only escaped neutralization by LY-CoV555 (Fig. 2, A and B, fig. S4A, and table S2). Although the distribution of LY-CoV016 and LY-CoV555 was paused in the United States in the summer of 2021 as the prevalence of Gamma and Beta VOCs increased, the distribution of this antibody cocktail has since been resumed with the rise of Delta as the predominant strain. However, our findings emphasize the importance of close monitoring of Delta AY.2 and of other Delta variants for acquisition of the K417N_RBD mutation.

Although REGN10933 lost substantial activity against the Beta variant, which is consistent with other reports (9, 12, 42), it still had an IC₅₀ value of less than 1 μg ml⁻¹ in our assays (Fig. 2, A and B, and fig. S4A). However, resistance markedly worsened with the day 146*, day 152*, RBM-1, RBM-2, and RBM-3 pseudotypes, with 800– to 1900–fold loss in neutralizing activity (IC₅₀ values ranging from 20 to 47 μg ml⁻¹). REGN10987 potently neutralized many of the variant pseudotypes we examined. While we observed expected resistance to REGN10987 neutralization by variants containing the N439K_RBD or the adjacent N440D_RBD substitutions (14, 16), we also observed some loss of activity against Epsilon and B.1.617.1 (Kappa), which was less expected because none of their substitutions fall within the REGN10987 RBD footprint (Fig. 1A and Movie 1). Other reports, nonetheless, have also noted varying degrees of modest in vitro resistance of Epsilon and Kappa variants to REGN10987 neutralization (39, 42). Notably, the day 146* and RBM-3 pseudotypes were the only ones to gain resistance to both antibodies in REGN-COV2 because they contain substitutions in the REGN10933 (e.g., Q493K_RBD, or E484K_RBD and K417N_RBD) and the REGN10987 binding sites (N439K_RBD or N440D_RBD) (Fig. 2, A and B, fig. S4A, and Movie 1) (14). We observed on GISAID instances of “day-146*-like” viruses that would be expected to resist neutralization by LY-CoV555, LY-CoV016, REGN10933, and REGN10987 because they contain the Q493K_RBD and N439K_RBD substitutions. One strain contains the N501Y_RBD, Q493K_RBD, and N439K_RBD mutations (sequenced once in South Africa), and the other contains the N501Y_RBD, Q493K_RBD, L452R_RBD, N439K_RBD, and N440F_RBD mutations (sequenced once in the United Kingdom) (table S1).

The broadly neutralizing antibody S309 (44), a class 3 antibody that binds the RBD but does not interfere with ACE2 binding, and from which the therapeutic antibody sotrovimab is derived, was active against all variants we tested (fig. S4A). However, we could not calculate reliable neutralization IC₅₀ values because of variable non-neutralizable pseudotype fractions (fig. S4A). The presence of a non-neutralizable fraction is unexplained but has been noted in other reports when human cells overexpressing ACE2, as opposed to African green monkey (Vero) target cells, were used to examine S309 neutralizing activity (45, 46).

Messenger RNA (mRNA)-based vaccines encoding the SARS-CoV-2 spike protein elicit robust neutralizing antibody responses (47-49). We directly compared immune evasion of the day 146*, day 152*, and RBM-2 pseudotypes to the B.1.1.7 (Alpha), Beta, and Gamma pseudotypes in sera obtained from individuals who had received a 2-dose series of an mRNA vaccine (BNT162b2 or mRNA-1273) (Fig. 2, C and D, and fig. S5). In addition to RBD substitutions, day 146*, day 152*, RBM-1, and RBM-2 spike proteins all contain NTD deletions spanning residues 141-144, which are positioned near NTD mutations in Alpha, Beta, and Delta in a key antigenic supersite (table S2) (6, 7). As similar NTD deletions found in Alpha, Beta, and Delta prevent binding of some neutralizing antibodies (6, 7, 46), they would be expected to escape neutralization by some NTD-targeting antibodies in addition to escaping neutralization by RBD-targeting antibodies. After first immunization but prior to the second dose, we observed a loss in neutralizing activity for all variants, although the severity varied. Variants that contain any substitution at E484_RBD, combined with an NTD deletion (Beta, day 152*, and RBM-2), were more effective at evading antibody responses than variants that had an E484_RBD substitution without an NTD deletion (Gamma) or an NTD deletion but no E484_RBD substitution (day 146*) (Fig. 2, C and D, fig. S5, and table S2). These findings are consistent with the role of E484_RBD as a major driver in neutralization escape of polyclonal antibody responses to SARS-CoV-2 (35) and observations that Beta more robustly escapes antibody neutralization than Gamma (9, 13). They further suggest that variants that have an NTD supersite deletion and an E484_RBD substitution are the most concerning when it comes to resistance to polyclonal antibodies.

One quarter of sampled individuals had no detectable activity against the Beta and RBM-2 pseudotypes after a single immunization (Fig. 2, C and D). However, sampling at seven and twenty-eight days after the second immunization revealed detectable neutralizing activity against all variants in all vaccine recipients, including against the RBM-2 pseudotype, which contains seven RBD mutations (Fig. 2, C and D, and fig. S5). Thus, repeated administration of an mRNA vaccine encoding constructs of the SARS-CoV-2 spike protein used in current formulations may provide sufficient neutralizing antibody breadth and potency to yield baseline serum neutralizing activity against variants that are more extensively mutated than the current dominant strains.

The RBD is also the major target of neutralizing antibodies against SARS-CoV, which caused a small outbreak of viral pneumonia from 2003 to 2004, though with a much higher case fatality rate (50, 51). Polyclonal antibody responses against SARS-CoV-2 poorly cross-neutralize SARS-CoV (52, 53). To identify barriers that restrict neutralization breadth, we performed single memory B-cell sorting with the SARS-CoV spike protein to mine the memory B cell repertoire of a COVID-19 convalescent individual (“C1”). Polyclonal IgG from C1 plasma neutralized SARS-CoV-2 pseudotype but had weak activity against SARS-CoV pseudotype (fig. S6A). From C1 peripheral blood mononuclear cells (PBMCs), using a pre-fusion stabilized SARS-CoV spike protein (S2P) (54), we cloned 17 cross-reactive antibodies. Of these, eleven antibodies bound both SARS-CoV and SARS-CoV-2 spike protein in an ELISA (fig. S6C and table S5). Only the two RBD-binding antibodies, C1C-A3 (“A3”) and C1C-C6 (“C6”), neutralized SARS-CoV-2 pseudotypes in our assays (Figs. 2B and 3A, and fig. S6F). Despite binding to SARS-CoV spike protein and RBD by ELISA, A3 and C6 did not neutralize SARS-CoV pseudotype (fig. S6, F and G). We also included C1A-A6 (“A6”) in these assays, a SARS-CoV-2 neutralizing antibody we previously isolated from the C1 donor using pre-fusion stabilized SARS-CoV-2 S2P in single B-cell sorting experiments (14). Unlike A3 and C6, A6 neutralized SARS-CoV pseudotypes (Figs. 2B and 3A, and fig. S6F). We determined Fab RBD binding affinities using biolayer interferometry (BLI) (fig. S7 and table S3) and confirmed A3 and A6 activity against infectious SARS-CoV-2 in a plaque reduction neutralization assay (fig. S4B).

A3 neutralized almost all SARS-CoV-2 variant pseudotypes with a neutralization IC₅₀ value of less than 1 μg ml⁻¹, including Beta, Gamma, Delta AY.2, Delta +3, RBM-1, RBM-2, and RBM-3 pseudotypes; the Epsilon variant was the only exception, with an IC₅₀ value of 1.9 μg ml⁻¹ (Fig. 2B and 3A, and fig. S4A). C6 and A6 also broadly neutralized variants but with higher baseline IC₅₀ values, even against D614G_S pseudotypes (ranging from 2.0 to 11.4 μg ml⁻¹) (Fig. 2B and 3A, and fig. S4A).

To determine where on the RBD A3, C6, and A6 bind, we carried out competition studies with C1A-B12 (14), a class 1 antibody; REGN10987 (17, 18) and S309 (44), two class 3 antibodies; and CR3022 (55), a class 4 antibody (Fig. 3B, fig. S8, and Movie 1). A3 competed with CR3022 and REGN10987 for RBD binding, C6 competed with CR3022, and C6 and A6 competed with each other (Fig. 3B and fig. S8). A6 did not compete with any of the other antibodies tested. Among A3, C6, and A6, only A3 competed with binding of an ACE2-Fc fusion protein, suggesting that A3 blocks cellular attachment.

We determined the 3.1 Å cryo-EM structure of the A3 Fab bound to the SARS-CoV-2 spike protein ectodomain (Fig. 4A, figs. S9 and S10, and table S6). A3 binds the RBD core with the spike protein trapped in the three open RBD conformation (Fig. 4A). In agreement with competition assays (Fig. 3B), A3 interacts with RBD residues that overlap significantly with those of CR3022 (Fig. 3, C, D and F, and Movie 1). A3 is therefore a class 4 antibody, a class that includes CR3022, S2A4, S304, S2X35, H014, COVA1-16, S2X259, and DH1047 (4, 56-59) (Movie 1 and fig. S11). Although the A3 and S309 footprints on the RBD do not overlap and S309 (a class 3 antibody) can bind the closed spike protein trimer (44), both antibodies contact the N-linked glycan attached to N343_RBD but approach it from different faces (Fig. 3C and Movie 1).

The A3 Fab avoids the RBD-ACE2 interface, which contains the majority of key antibody escape mutations, but like other class 4 antibodies, nonetheless binds the RBD in a manner that would sterically interfere with ACE2 binding (Fig. 4, B-E, and fig. S11). Based on its epitope, in addition to retaining activity against all variants we tested, A3 would also have activity against emergent and pre-emergent SARS-CoV-2 variants; these include a virus sequenced from travelers from Tanzania that contains the E484K_RBD, T478R_RBD, and R346K_RBD mutations, and B.1.621 (Mu), a variant detected early in 2021 in Colombia that has since spread internationally and contains the E484K_RBD, N501Y_RBD, and R346K_RBD mutations (Fig. 1G , Fig. 4D, and table S1). The R346K_RBD mutation falls within the RBD core and is in the S309 binding site but is not within A3’s footprint (Fig. 3, D and E, and Movie 1). However, S309 would likely retain activity against SARS-CoV-2 variants that contain the R346K_RBD mutation, as the residue that is at the position analogous to SARS-CoV R346_RBD is a lysine in SARS-CoV, and S309 neutralizes both SARS-CoV and SARS-CoV-2 (44, 60).

Despite A3’s breadth against SARS-CoV-2 variant pseudotypes (Figs. 2B and 3A), A3 does not neutralize SARS-CoV pseudotype (fig. S6, F and G). The A3 epitope is highly conserved between SARS-CoV-2 and SARS-CoV; however, N370_RBD is a site of N-linked glycosylation in SARS-CoV (N357_RBD in SARS-CoV numbering) and in animal coronaviruses but not in SARS-CoV-2 (Fig. 5, A, B, C and F) (61). An N-linked glycan attached to SARS-CoV-2 N370_RBD would introduce steric clashes with the A3 antibody heavy and light chains (Fig. 5D). Furthermore, calculations of antibody accessible surface areas using molecular dynamic trajectories of a fully glycosylated SARS-CoV-2 spike protein with a modeled N370_RBD glycan suggest that its addition would restrict A3 epitope access and could also impact binding of other class 4 antibodies (fig. S12) (61, 62).

Partial occupancy of the glycan attached to SARS-CoV N357_RBD in recombinant protein preparations may explain how we observed spike protein and RBD binding but lack of SARS-CoV pseudotype neutralization (fig. S6, F and G). In surface plasmon resonance binding assays, A3 IgG bound tightly to the SARS-CoV RBD only when the RBD was enzymatically deglycosylated (fig. S13 and table S7). Consistent with the role of the SARS-CoV N357_RBD N-linked glycan as a barrier to A3 neutralization, introducing a substitution that would prevent its addition (T359A_RBD) sensitized SARS-CoV pseudotypes to A3 neutralization (IC₅₀ value of 5 μg ml⁻¹) (Fig. 2B and fig. S4A).

The A372S/T_RBD mutations, which would introduce an N-linked glycosylation motif and allow for modification of N370_RBD in the SARS-CoV-2 spike protein, are found in human-derived SARS-CoV-2 sequences (GISAID) (27), including on sequences for VOCs Alpha and Delta, without apparent geographic restriction (48 sequence counts as of October 10^th, 2021, and detected in at least fourteen countries) (Fig. 5F and table S8). Although the mutations are currently rare, their presence in sequence databases suggest that SARS-CoV-2 strains containing these mutations can replicate in humans. To confirm that an N-linked glycan could be added to N370_RBD, we conducted glycan analysis on recombinant SARS-CoV-2 RBD containing the A372S_RBD substitution and observed 90% occupancy of an N-linked glycan at position N370_RBD (Fig. S13B)

Because acquisition of a putative N-linked glycan at N370_RBD was the most frequent on the Alpha variant at the time of our initial analysis, we generated an Alpha pseudotype that contains the A372T_RBD substitution (Alpha A372T). We tested the effect of this substitution on three class 4 antibodies: A3, the antibody we isolated here, S2A4, an antibody that does not cross-react with the SARS-CoV RBD (4) and COVA1-16, an antibody that has weak cross-neutralizing activity against SARS-CoV (57). The mutation resulted in eightfold resistance to A3 neutralization (IC₅₀ value of 1.1 μg ml⁻¹, as compared to 0.14 μg ml⁻¹ with Alpha pseudotype) and complete resistance to S2A4 neutralization (Fig. 2B, Fig. 3A, Fig 5, G and H, and fig. S4A). S2A4 and COVA1-16 neutralized variants with potency that was overall comparable to A3 in most cases (Fig. 2B, Fig. 3A, and fig. S4A). COVA1-16, probably because it has some activity against SARS-CoV (above the limit of detection in our assays but 29 μg ml⁻¹ as reported by Liu et al.) (57), retained activity against Alpha A372T pseudotype (Fig. 2B, Fig. 3A, Fig. 5H, fig. S4A). The Fab binding pose of certain class 4 antibodies, therefore, may allow them to avoid steric hindrance from an N-linked glycan attached to N370_RBD (S2X259 is an example of one such antibody) (Movie 1) (56).

Coronaviruses that circulate in animals and have spike protein RBDs that can bind human ACE2 are a continued threat. RaTG13 virus, which is closely related to SARS-CoV-2 phylogenetically, is an example of one such virus (63). The RaTG13 virus spike protein contains a threonine at RBD position 372, which would allow for N370_RBD glycosylation (Fig. 5F). Despite the presence of the N-linked glycan, A3 potently neutralized RaTG13 virus pseudotype (neutralization IC₅₀ value of 21 ng ml⁻¹), suggesting that A3 neutralization breadth extends to pre-emergent coronaviruses that are closely related to SARS-CoV-2 (Fig. 5, G and H). Structural superposition reveals that the N370_RBD glycan on the RaTG13 RBD is positioned in a manner that may not block A3 epitope access but, rather, could interfere with binding of other antibodies that bind nearby epitopes on the RBD core (Fig. 5E).