Neutralizing antibodies (NAbs) against SARS-CoV-2 protect against infection in animal models1,3,4,10,11 and are being evaluated for prophylaxis and as therapeutic agents in humans7,8. These antibodies target the SARS-CoV-2 spike (S) trimer3,5,10,12,13,14,15,16,17, a viral glycoprotein that mediates binding to the angiotensin-converting enzyme 2 (ACE2) receptor18,19. The S trimer comprises three copies of an S1 subunit that contains the RBD and three copies of S2, which includes the fusion peptide and transmembrane regions20,21. The RBDs of SARS-CoV-2 and other coronaviruses exhibit flexibility, such that they bind to ACE2 only when they are in an ‘up’ conformation, compared with the ‘down’ RBD conformation of the closed, prefusion S trimer20,21,22,23,24,25.

Many human NAbs isolated from COVID-19-convalescent donors target the RBD, binding to distinct, sometimes non-overlapping, epitopes3,4,5,10,12,13,14,17. A subset of these antibodies blocks viral entry by binding to the ACE2-binding site on the RBD6,11,13,15,26,27. A family of recurrent ACE2-blocking human NAbs is composed of heavy chains encoded by the VH3-53 or VH3-66 gene segment3,12,13,16,17,27,28,29, most of which are known or predicted15,26,28,30,31 to exhibit a common RBD binding mode that results from the use of germline-encoded residues within the complementarity-determining regions 1 and 2 (CDRH1 and CDRH2) and a CDRH3 that is shorter than the average length (15 amino acids; IMGT32 complementarity-determining region (CDR) definition) in human antibodies33. Other SARS-CoV-2 RBD-binding antibodies are encoded by VH3-305, and these have also been isolated from donors infected with SARS-CoV34, and antibodies with a variety of the other Vh gene segments3,5,10,12,13,14,15,16,17.

To classify commonalities and differences among RBD-binding human NAbs isolated from COVID-19-convalescent individuals5, we solved complexes of NAbs with stabilized (2P and 6P versions)35,36 soluble S trimer. Subsequently, we used high-resolution details of the binding orientations of NAbs encoded by the VH1-2, VH1-46, VH3-30, VH3-53, VH4-34 and VH5-51 gene segments to determine rules for binding by four distinct anti-RBD antibody classes (Supplementary Table 2). The NAbs chosen for structures are highly potent, achieving 90% neutralization in pseudotype virus assays at concentrations ranging from 22 to 140 ng ml−1 (ref. 5), and thus our structural analyses and classifications directly relate to understanding mechanisms of neutralization and potency differences between human NAbs.

Class 1 VH3-53 NAbs block ACE2 and bind to up RBDs

We solved Fab and Fab–RBD crystal structures of C102 (Supplementary Table 1), which we compared to our previous26 cryo-electron microscopy (cryo-EM) structure of S trimer complexed with the related human NAb C105 (Extended Data Figs. 1, 2). Both C102 and C105 are VH3-53 NAbs with short (11 and 12 residues, respectively) CDRH3 loops (Extended Data Fig. 1g) that were isolated from the same donor5. They share structural similarities with each other and with other VH3-53-encoded short CDRH3 human NAb structures solved as complexes with RBDs12,30,37,38 (Extended Data Fig. 2a). Notably, the C102–RBD structure resembled the analogous portion of the C105–S structure26 (Extended Data Fig. 2a). These results establish that Fab–RBD structures can reproduce interactions with RBDs in the context of an S trimer; however, Fab–RBD structures do not reveal the state(s) of the antibody-bound RBD in the complex (up versus down) or the potential inter-protomer contacts by Fabs.

Because the C105 Fab bound to either two or three up RBDs on S with no observed interactions with down RBDs or with adjacent RBDs26 (Extended Data Fig. 1f), we used the higher-resolution C102 Fab–RBD structure to deduce a more accurate epitope and paratope than was possible using the C105–S cryo-EM structure with flexible up RBDs (Extended Data Fig. 1a–e). Buried surface area calculations showed that the C102 CDRH3 region had a relatively minor role in the paratope: of 1,045 Å2 of buried surface area on the antibody (786 Å2 on the heavy chain; 259 Å2 on the light chain), CDRH3 accounted for only 254 Å2 (Extended Data Fig. 2b). This contrasts with most antibodies in which CDRH3 contributes equally or more to the interface with antigen than the sum of CDRH1 and CDRH2 contributions39. The epitopes on RBD for all available VH3-53-encoded short CDRH3 human NAbs span the ACE2-binding site15,26,28,30,31 and show common RBD-binding interactions, represented by the C102 epitope (Extended Data Fig. 1b–e), which buried 1,017 Å2 on RBD (Extended Data Fig. 2b). The ACE2-blocking epitope for these NAbs is sterically occluded in the RBD down conformation (Fig. 1b, Extended Data Fig. 1f); therefore, class 1 NAbs can only bind to up RBDs, as observed in the C105–S structure26, and as previously discussed, IgGs in this class could crosslink adjacent RBDs within a single trimer to achieve tighter binding through avidity effects26.

Fig. 1: Cryo-EM structure of the C144–S complex illustrates a distinct VH3-53 NAb binding mode.
figure 1

a, 3.2 Å cryo-EM density for the C144–S trimer complex revealing C144 binding to a closed (three down RBDs) spike conformation. LC, light chain; HC, heavy chain. b, Overlay of C102 Fab (from C102–RBD crystal structure) (Extended Data Fig. 1) and C144 Fab (from C144–S structure) aligned on a RBD monomer. RBD residues corresponding to the ACE2 epitope (orange-red cartoon) are shown on the same RBD for reference. C144 adopts a distinct conformation relative to the C102-like VH3-53-encoded short CDRH3 NAb class, allowing binding to the down RBD conformation on trimeric spike, whereas C102-like NAbs can only bind to up RBDs. c, Quaternary epitope of C144 involving bridging between adjacent RBDs via the CDRH3 loop (illustrated as thicker ribbon). d, e, Close-up view of CDRH3-mediated contacts on adjacent protomer RBD (dark grey). C144 CDRH3 residues F100D and W100E are buried in a hydrophobic pocket comprising the RBD α1 helix, residue F374RBD and the N343RBD glycan. f, Surface representation of C144 epitope (light blue) across two adjacent RBDs. RBD epitope residues (defined as residues containing atom(s) within 4 Å of a Fab atom) are labelled in black.

Class 2 NAbs recognize ‘up’ and ‘down’ RBDs

In addition to the recurrent VH3-53-encoded short CDRH3 NAb structures, a small subset of potently neutralizing VH3-53-encoded antibodies use longer CDRH3 regions5, 12 (more than 15 residues, IMGT definition32) (Extended Data Fig. 1g). A recent structure of a RBD complexed with a VH3-53-encoded long CDRH3 human NAb (COVA2-39) revealed a different RBD binding mode38, thus confirming predictions that binding with a C102-like interaction requires a short CDRH326,30. To further determine molecular mechanisms for binding of VH3-53-encoded long CDRH3 human NAbs, we solved a 3.2 Å cryo-EM structure of C144 (encoded by the VH3-53 and VL2-14 gene segments; 25-residue CDRH3) bound to an S trimer36 (Extended Data Fig. 3). Despite the ability of ligand-free stabilized S trimers to adopt up RBD conformations36 and modelling suggesting the C144 binding site would be accessible on up RBDs (Fig. 1b), the C144–S structure revealed three C144 Fabs bound to a completely closed S with three down RBDs (Fig. 1a). The C144 binding mode differs from class 1 NAbs, the binding orientation of which is incompatible with down RBD conformations (Fig. 1b). In addition, the binding orientation observed for C144 differs from the binding described for COVA2-39, the RBD epitope of which is predicted to be accessible only on up RBDs38 owing to steric hinderances imposed on the light chain by the N343RBD-associated glycan on the adjacent RBD (Extended Data Fig. 1h). Despite differences in orientation, the RBD epitopes of C144, C102 and COVA2-39 overlap with the ACE2-binding site, which suggests a neutralization mechanism that involves direct competition with ACE2 (Fig. 1b).

Despite overlapping with the ACE2-binding site on up RBDs, an interesting feature of C144 binding is that its long CDRH3 bridges between adjacent down RBDs to lock the spike glycoprotein into a closed, prefusion conformation, providing an additional neutralization mechanism in which S cannot open to engage ACE2 (Fig. 1c, d). The formation of the C144 quaternary epitope is driven by sandwiching CDRH3 residues F100D and W100E (in which subscripts denote numbering of the CDRH3 loop) into a hydrophobic RBD cavity at the base of an N-linked glycan attached to N343RBD. The cavity comprises the RBD α1 helix (337–344), α2 helix (364–371), and hydrophobic residues (F374RBD and W436RBD) at the edge of the RBD five-stranded β-sheet (Fig. 1e, f). In contrast to the CDRH3s of class 1 VH3-53-encoded short CDRH3 NAbs, the C144 CDRH3 contributed to most (approximately 60%) of the paratope and buried 330 Å2 of surface area on the adjacent RBD (Extended Data Fig. 2b), rationalizing observed escape at L455RBD (Fig. 1f) in C144 selection experiments40. Despite adjacent CDRH3 hydrophobic residues (F100D and W100E) likely to be solvent-exposed before antigen binding, C144 IgG showed no evidence of non-specific binding in a polyreactivity assay (Extended Data Fig. 1i).

Given the unusual binding characteristics of C144, we investigated whether antibodies that showed similar S binding orientations in low-resolution negative-stain electron microscopy reconstructions5 use similar neutralization mechanisms. We characterized Fab–S cryo-EM structures (overall resolutions from 3.4 to 3.8 Å) of potent NAbs (C002, C104, C119 and C121) predicted to compete with ACE2 binding5, which varied in their use of V gene segments and CDRH3 lengths (Fig. 2, Extended Data Figs. 3, 4, Extended Data Table 1). Fab–S cryo-EM structures of these class 2 NAbs showed bound RBDs in both up or down conformations, consistent with observations of similar human NAbs from negative-stain electron microscopy5,12 and single-particle cryo-EM studies10,34,41. By contrast, the C144–S structure showed Fabs bound only to down RBDs (Fig. 1), which suggests that C144 binding requires recognition of the closed S trimer, or that C144 Fab(s) initially bound to up RBD(s) could trap the closed (three RBDs down) S conformation through CDRH3-mediated interactions between adjacent RBDs.

Fig. 2: Cryo-EM structures of class 2 C002 and C121 NAbs show binding to up and down RBDs.
figure 2

a, b, Cryo-EM densities for C002–S (a; 3.4 Å) and C121–S (b; 3.7 Å) complexes, revealing binding of C002 or C121 to both down and up RBDs. Inset, alignment of C002 and C121 Fabs on the same RBD. ACE2 is represented as a green surface for reference. c, d, Surface representations of C002 epitope (orange, c) and C121 epitope (purple, d) on the RBD surface (grey). RBD epitope residues (defined as residues containing atom(s) within 4 Å of a Fab atom) are labelled in black. e, C002 forms inter-protomer contacts via binding to an adjacent up RBD conformation on the surface of the trimer spike (also observed for class 2 C121–S, C119–S and C104–S structures) (Extended Data Fig. 5). Red box shows close-up of adjacent up RBD and C002 light-chain interface.

To understand commonalities of class 2 RBD epitopes better, we further analysed two additional potent human NAbs, C002 (encoded by VH3-30 and VK1-39 gene segments; 17-residue CDRH3, half-maximal inhibitory concentration (IC50) = 8.0 ng ml−1)5 and C121 (encoded by VH1-2 and VL2-23 gene segments; 23-residue CDRH3, IC50 = 6.7 ng ml−1)5, for which cryo-EM Fab–S structures were solved to 3.4 Å and 3.6 Å, respectively (Fig. 2a, b), using crystal structures of unbound C002 and C121 Fabs for fitting (Supplementary Table 1). The C002 and C121 RBD epitopes are focused on the receptor-binding ridge, overlapping with polar and hydrophobic residues along the flat face of the RBD responsible for ACE2 interactions (Fig. 2c–e). Similar to C144, NAbs C002 and C121 buried most of their RBD epitopes against heavy-chain CDR loops, with light-chain CDR loops engaging the receptor-binding ridge (Fig. 3). Notably, Fab–S structures of C002, C121, C119 and C104 revealed a quaternary epitope involving an adjacent RBD (Extended Data Figs. 3, 4, 5a–c), albeit distinct from the quaternary binding of C144 (Fig. 1c–e). The C002/C121/C119/C104 type of secondary interaction was only observed when a Fab was bound to a down RBD and adjacent to an up RBD. The extent of secondary interactions varied depending on the antibody pose (Extended Data Fig. 5a–c). Bridging interactions between adjacent up and down RBDs would not allow the two Fabs of a single IgG to bind simultaneously to an S trimer. However, this class of antibodies could support bivalent interactions between two adjacent down RBDs (Extended Data Fig. 5h, Extended Data Table 1).

Fig. 3: Details of common RBD interactions among class 2 human NAbs.
figure 3

al, Conserved interactions between the RBD and CDRs of class 2 NAbs as observed for C144 (HC, cyan; LC, sky blue) (ad), C002 (HC, dark orange; LC, light orange) (eh), and C121 (HC, purple; LC, pink) (il). Primary and secondary epitopes on adjacent down RBDs are shown for C144. Secondary epitopes for C002 and C121, which require adjacent up RBDs, are shown in Extended Data Fig. 5. RBDs are grey; potential hydrogen bonds and π–π stacking interactions (d, Y33LC and F486RBD; h, Y92LC and F486RBD; l, Y91LC and F486RBD) are indicated by dashed lines.

Characterization of the highest resolution interface (C002–S structure) showed C002 light-chain framework regions 1 and 2 (FWR1 and FWR2) interfaced with the RBD residues comprising the five-stranded β-sheet and α-helix that spans residues 440–444 (Fig. 2e), which is typically located near the three-fold axis of a closed S trimer. In addition to contacting neighbouring RBDs, inter-protomer engagement with the N165NTD-associated glycan in the N-terminal domain (NTD) was observed for the class 2 NAb BD2313. If fully processed, the N165NTD glycan could adopt a conformation that would allow interactions with the heavy-chain FWR3 and CDRH1 (Fig. 2e). However, in the structures reported here, we did not observe N165NTD glycan density beyond the initial N-acetylglucosamine.

Given differences in class 2 human NAb V gene segments, CDRH3 lengths and antibody poses, we investigated sequence features that drive conserved interactions. Sequence differences between SARS-CoV-2 and SARS-CoV RBD, including at positions 486RBD and 493RBD (F and Q, respectively, in SARS-CoV-2), in the ACE2 receptor-binding motif allowed more favourable ACE2 binding to the SARS-CoV-2 RBD42. Analysis of interactions by C144, C002 and C121 revealed common interactions with these residues and also for E484RBD by both antibody heavy-chain and light-chain residues (Fig. 3). In particular, class 2 NAb interactions with F486RBD mimicked ACE2 interactions, in that F486RBD buries into a hydrophobic pocket typically involving CDRL1 or CDRL3 tyrosine residues43 (Fig. 3d, h, l). Mimicking of the ACE2 F486 binding pocket by SARS-CoV-2 human NAbs was observed across different light-chain V gene segments (Extended Data Table 1), which suggests that there is no restriction in light-chain V gene segment usage for class 2 NAbs. Notably, a germline-encoded feature described for VH3-53-encoded short CDRH3 class 1 NAbs, the CDRH2 SXXS motif, is also found in other class 2 NAbs (for example, C121 and C119) despite different Vh gene segment usage. Similar to VH3-53 NAbs C144 and COVA2-39, the C121 CDRH2 SXXS motif forms a potential hydrogen-bond network with E484RBD (Fig. 3b, j).

Overall, these results suggest a convergent mode of recognition by germline-encoded residues across diverse Vh/Vl gene segments for SARS-CoV-2, which may contribute to low levels of somatic hypermutation observed for these human NAbs (Extended Data Fig. 4i–n, Extended Data Table 1).

Class 3 NAbs bind outside the ACE2-binding site

C135 is a potent NAb that showed distinct binding properties from class 1, 2 and 4 NAbs, the latter of which bind a highly conserved buried epitope that is only accessible in up RBD conformations (Extended Data Table 1). To evaluate the mechanism of C135-mediated neutralization of SARS-CoV-2, we solved the cryo-EM structure of a C135–S complex to 3.5 Å (Fig. 4a, Extended Data Fig. 6), using an unbound C135 crystal structure for fitting (Supplementary Table 1). The structure revealed three C135 Fabs bound to an S trimer with two down and one up RBD, although the C135-bound up RBD conformation was weakly resolved and therefore not modelled. C135 recognizes a similar glycopeptidic epitope to the cross-reactive SARS-CoV NAb S30934, focusing on a region of the RBD near the N343 glycan and non-overlapping with the ACE2-binding site (Fig. 4b, Extended Data Fig. 6c, d). Despite differences in binding orientations between C135 and S309, targeting of the RBD epitope was mainly Vh-mediated (the surface area buried by RBD on the C135 heavy chain represented approximately 480 Å2 of the 700 Å2 total buried surface area) and included interactions with the core fucose moiety of the N343RBD glycan. The smaller C135 footprint relative to S309 (approximately 700 Å2 versus 1,150Å2 buried surface area, respectively) (Extended Data Fig. 6c, d) focused on interactions with R346RBD and N440RBD, which are engaged by residues from heavy-chain and light-chain CDRs (Fig. 4c, d) and are not conserved between SARS-CoV-2 and SARS-CoV RBDs, rationalizing the lack of SARS-CoV cross-reactivity observed for C1355.

Fig. 4: Cryo-EM structure of S complexed with the class 3 (non-ACE2 blocking) human NAb C135.
figure 4

a, 3.5 Å cryo-EM density of C135–S complex. b, Composite model of C135–RBD (blue and grey, respectively) overlaid with the SARS-CoV-2 NAb S309 (sand; PDB code 6WPS) and soluble ACE2 (green; PDB code 6M0J). The model was generated by aligning on 188 RBD Cα atoms. c, d, C135 CDRH (dark blue) and CDRL (light blue) interactions with R346RBD (c) and N440RBD (d). Potential π–π stacking interactions (c) and hydrogen bonds (c, d) are illustrated by dashed black lines. e, f, Model of RBD interactions of NAbs C135 (class 3) and C144 (class 2), demonstrating that both Fabs can bind simultaneously to a single monomeric RBD (e), but would clash if bound to adjacent down RDBs on S trimer (f). Steric clashes indicated by a red and yellow star in f. g, h, Model of RBD interaction of NAbs C135 (class 3) and C119 (class 2) demonstrating that both Fabs cannot bind simultaneously to a single monomeric RBD (g), but do not clash if bound to adjacent down RDBs on S trimer (h). Steric clashes indicated by a red and yellow star in g.

The discovery of class 3 NAbs such as C135 and S309 that were raised during SARS-CoV-2 or SARS-CoV natural infections, respectively, and bind outside of the ACE2-binding site, provides the potential for additive neutralization effects when combined with NAbs that block ACE2, while also limiting viral escape1,40. A pair of antibodies in human clinical trials that includes REGN109878, a human NAb that binds distal to the ACE2-binding site, prevented SARS-CoV-2 viral escape in vitro, but did not show synergistic neutralization6. Comparison of C135 and REGN10987 interactions with S showed similarities in epitopes (interactions focused on R346RBD and N440RBD) (Extended Data Fig. 7c, f). However, REGN10987 binding would sterically hinder ACE2 interactions, whereas C135 binding does not (Fig. 4b, Extended Data Fig. 6b). Notably, a structure of S complexed with C110 (encoded by the VH5-51 and VK1-5 gene segments), isolated from the same donor as the C102 and C105 (class 1) and C119 and C121 (class 2) NAbs5, showed a binding pose that resembled that of REGN10987 (Extended Data Fig. 6b, e, f). The C110 epitope showed similarities with both class 3 and class 2 NAbs, binding distal to the ACE2-binding motif, but like REGN10987, could potentially sterically interfere with ACE2 (Extended Data Fig. 7). For each of these class 3 NAbs, the Fab binding pose suggests that inter-protomer crosslinking by a single IgG is not possible (Extended Data Table 1).

Class 3 human NAbs add to the anti-SARS-CoV-2 antibody repertoire and could probably be effectively used in therapeutic combinations with class 1 or 2 NAbs. However, when using structures to predict whether NAbs have overlapping epitopes, it is sometimes not sufficient to only examine Fab–RBD structures or even static images of the S trimer because of the dynamic nature of the spike. Thus, what might appear to be non-overlapping epitopes on an isolated RBD could overlap in some (Fig. 4e, f), but not all (Extended Data Fig. 7), scenarios on a spike trimer, complicating interpretation of competition experiments using monomeric RBDs and S trimers. The opposite can also be true; that is, two Fabs predicted to be accommodated on a trimer could clash on an RBD monomer (Fig. 4g, h). Finally, adjacent monomers in different orientations could accommodate different antibodies that target overlapping sites (Extended Data Fig. 7).

RBD substitutions affect NAb binding

Vesicular stomatitis virus (VSV) reporter viruses pseudotyped with the SARS-CoV-2 S protein can escape by mutation(s) from the C121, C135 or C144 NAbs40 that we used for structural studies. RBD mutations that were selected in response to antibody pressure correlated with the epitopes mapped from the structures of their Fabs complexed with the S trimer (Figs. 1, 2, 4).

To further assess the effects of these mutations and other RBD substitutions, we assayed NAbs for which we obtained structural information (eight from this study; C105–S complex previously described26) for binding to mutated RBD proteins. The RBD mutants included two that induced escape from the class 3 hNAb C135 (R346S and N440K)40 (Fig. 4c, d), one found in circulating isolates44 that conferred partial resistance to C135 (N439K)40 (Fig. 4d), a circulating variant (A475V) that conferred resistance to class 1 and 2 VH3-53 NAbs44, two that induced escape from C121 or C144 (E484K and Q493R)40 (Fig. 3), and a circulating variant that conferred partial resistance to C121 (V483A)40. Kinetic and equilibrium constants for the original and mutant RBDs were derived from surface plasmon resonance (SPR) binding assays in which RBDs were injected over immobilized IgGs. Loss of binding affinity was consistent with RBD mutations that conferred escape (Extended Data Fig. 8). Comparing effects of point mutations between NAb classes showed that point mutations leading to a loss of binding for NAbs within one class did not affect NAbs in a different class, indicating that antibody pressure that leads to escape from one NAb class would be unlikely to affect a different class. These results suggest a therapeutic strategy involving human NAbs of different classes for monoclonal NAb treatment of individuals infected with SARS-CoV-2.


Here we report structural, biophysical and bioinformatics analyses of SARS-CoV-2 NAbs (Extended Data Fig. 9), providing information for interpreting correlates of protection for clinical use. The structures reveal a wealth of unexpected interactions of NAbs with the spike protein, including five antibodies that reach between adjacent RBDs on the protomers of a single trimer. A notable example of bridging between spike protomers involved the human C144 NAb that uses a long CDRH3 with a hydrophobic tip to reach across to an adjacent RBD, resulting in all three RBDs on the spike trimer being locked into a closed conformation. This example, and four other NAbs that contact adjacent RBDs, demonstrates that crystal structures of Fab–monomeric RBD complexes, although informative for defining primary epitopes on one RBD, do not reveal how antibodies recognize the flexible up or down RBD conformations on the spike trimer that are targeted for neutralization on a virus. Indeed, our cryo-EM structures of Fab–spike trimer complexes showed all possible up and down combinations of recognized RBDs, with structures showing either three or two Fabs bound per trimer. By analysing approach angles of Fabs bound to RBDs on spike trimers, we predicted whether an IgG can bind to a single spike trimer to gain potency through avidity, which would also render the antibody more resistant to spike mutations. In addition, structural information allowed us to assess RBD mutants that arose in circulating viral isolates and/or were obtained by in vitro selection. Together, this study provides a blueprint for the design of antibody cocktails for therapeutic agents and potential spike-based immunogens for vaccines.


No statistical methods were used to predetermine sample size. The experiments were not randomized, and investigators were not blinded to allocation during experiments and outcome assessment.

Cell lines

Expi293F cells (GIBCO) for protein expression were maintained at 37 °C and 8% CO2 in Expi293 Expression medium (GIBCO), transfected using Expi293 Expression System Kit (GIBCO) and maintained under shaking at 130 rpm. Cell lines were not specifically authenticated, but lines tested negative for contamination with mycoplasma.

Protein expression

Expression and purification of SARS-CoV-2 ectodomains were conducted as previously described26. In brief, constructs encoded the SARS-CoV-2 S ectodomain (residues 16–1206 of the early SARS-CoV-2 GenBank MN985325.1 sequence isolate with 2P35 or 6P36 stabilizing mutations, a mutated furin cleavage site between S1 and S2, a C-terminal TEV site, foldon trimerization motif, octa-His tag, and AviTag) were used to express soluble SARS-CoV-2 S ectodomains. Constructs encoding the SARS-CoV-2 RBD from GenBank MN985325.1 (residues 331–524 with C-terminal octa-His tag and AviTag) and mutant RBDs were made as described26, SARS-CoV-2 2P S, 6P S, and RBD proteins were purified from the supernatants of transiently transfected Expi293F cells (Gibco) by nickel affinity and size-exclusion chromatography26. Peak fractions were identified by SDS–PAGE, and fractions corresponding to S trimers or monomeric RBDs were pooled and stored at 4 °C. Fabs and IgGs were expressed, purified, and stored as previously described45,46.

X-ray crystallography

Crystallization trials were carried out at room temperature using the sitting drop vapour diffusion method by mixing equal volumes of a Fab or Fab–RBD complex and reservoir using a TTP LabTech Mosquito robot and commercially available screens (Hampton Research). Crystals were obtained in 0.2 M ammonium sulfate, 20% (w/v) PEG 3350 (C102 Fab), 0.2 M sodium citrate tribasic, 20% (w/v) PEG 3350 (C102–RBD), 0.2 M lithium sulfate monohydrate, 20% (w/v) PEG 3350 (C002 Fab), 0.04 M potassium phosphate, 16% (w/v) PEG 8000, 20% (v/v) glycerol (C135 Fab), 0.2 M ammonium citrate pH 5.1, 20% PEG 3350 (C121 Fab), or 0.2 M sodium tartrate dibasic dihydrate pH 7.3, 20% (w/v) PEG 3350 (C110 Fab). A C135 Fab crystal was directly looped and cryopreserved in liquid nitrogen. Other crystals were quickly cryoprotected in a mixture of well solution with 20% glycerol and then cryopreserved in liquid nitrogen.

X-ray diffraction data were collected for Fabs and the Fab–RBD complex at the Stanford Synchrotron Radiation Lightsource (SSRL) beamline 12-1 on a Eiger X 16 M pixel detector (Dectris) at a wavelength of 1.0 Å. Data from single crystals of C121 Fab and C110 Fab were indexed and integrated in XDS47 and merged using AIMLESS v.0.7.4 in CCP448 v.7.0.6 (Supplementary Table 1). Data from single crystals of C102 Fab, C135 Fab and C002 Fab were indexed and integrated using XDS47 and merged in Phenix49 (v.1.18). Data from a single crystal of C102 Fab–RBD complex were indexed and integrated using XIA250 v.0.3.8 implementing DIALS51,52 v.2.2 and merged using AIMLESS in CCP448. For C110 Fab and C121 Fabs, structures were determined by molecular replacement in PHASER53 v.2.8.2 using the coordinates for B38 (PDB 7BZ5) or an inferred germline form of the HIV-1 NAb IOMA54 inferred germline (unpublished), respectively, after removing CDR loops as a search model. For C002 Fab, C102 Fab, C102 Fab–RBD and C135 Fab, structures were determined by molecular replacement in PHASER53 using B38 Fab coordinates (PDB 7BZ5) after trimming heavy and light chain variable domains using Sculptor55 v.2.0 (and for the C102 Fab–RBD data, also RBD coordinates from PDB code 7BZ5) as search models. Coordinates were refined using Phenix49 and cycles of manual building in Coot56 (Supplementary Table 1).

Cryo-EM sample preparation

Purified Fabs were mixed with the SARS-CoV-2 S 2P trimer35 or SARS-CoV-2 S 6P trimer36 (1.1:1 molar ratio Fab per protomer) to a final Fab–S complex concentration of 2–3 mg ml−1 and incubated on ice for 30 min. Immediately before deposition of 3 μl of complex onto a 300 mesh, 1.2/1.3 UltrAuFoil grid (Electron Microscopy Sciences) that had been freshly glow-discharged for 1 min at 20 mA using a PELCO easiGLOW (Ted Pella), a 0.5% (w/v) octyl-maltoside, fluorinated solution (Anatrace) was added to each sample to a final concentration of 0.02%. Samples were vitrified in 100% liquid ethane using a Mark IV Vitrobot (Thermo Fisher) after blotting at 22 °C and 100% humidity for 3 s with Whatman No. 1 filter paper.

Cryo-EM data collection and processing

Single-particle cryo-EM data were collected on a Titan Krios transmission electron microscope (Thermo Fisher) operating at 300 kV for all Fab–S complexes except for C144–S, which was collected on a Talos Arctica (Thermo Fisher) operating at 200 kV. Movies were collected using SerialEM v.3.7 automated data collection software57 with beam-image shift over a 3-by-3 pattern of 1.2 μm holes with 1 exposure per hole. Movies were recorded in super-resolution mode on a K3 camera (Gatan) for the C144–S dataset on the Arctica (0.435Å per pixel) or on a K3 behind BioQuantum energy filter (Gatan) with a 20 eV slit on the Krios (0.418 Å per pixel) for all other datasets. Data collections parameters are summarized in Supplementary Table 2. In general, the dat- processing workflow described below was performed for all datasets in cryoSPARC v.2.1558.

Cryo-EM movies were patch motion corrected for beam-induced motion including dose weighting within cryoSPARC58 after binning super-resolution movies. The non-dose-weighted images were used to estimate CTF parameters using CTFFIND459 v.4.1.14 or with cryoSPARC implementation of the Patch CTF job, and micrographs with power spectra that showed poor CTF fits or signs of crystalline ice were discarded. A subset of images were randomly selected and used for reference-free particle picking using Blob picker in cryoSPARC58. Particles were subjected to 2D classification and the best class averages that represented different views were used to generate 3 ab initio models. The particles from the best classes were used in another 2D classification job, and the best set of unique views was used as templates for particle picking on the full set of images. Initial particle stacks were extracted, downsampled twice, and used in heterogeneous refinement against the three ab initio volumes generated with the smaller dataset (ab initio volumes used were interpreted as a Fab–S complex, free Fab or dissociated S protomers, and junk/noise class). Particles assigned to the Fab–S volume were further cleaned via iterative rounds of 2D classification to select class averages that displayed unique views and secondary structural elements. Resulting particle stacks were homogenously refined before being split into nine individual exposure groups based upon collection holes. Per particle CTF and aberration corrections were performed and the resulting particles further 3D refined. Additional processing details are summarized in Supplementary Table 2.

Given the known heterogeneity of spike trimers20,21, homogenously refined particles were used for 3D classification in cryoSPARC58 (ab initio job: k = 4 classes, class similarity = 0.3). This typically resulted in one or two majority Fab–S complexes, with the other minority populated classes representing junk or unbound S trimer. Particles from the good class(es) were further subjected to 3D classification (ab initio job: k = 4, class similarity = 0.7) to attempt to separate various Fab–S complex states. If several states were identified (as observed for the C002–S and C121–S complexes), particles were heterogeneously refined, followed by re-extraction without binning (0.836 Å per pixel) before homogeneous refinement of individual states. For all other datasets, most particles represented one state that was homogenously refined after re-extraction without binning.

Particle stacks for individual states were non-uniform refined with C1 symmetry and a dynamic mask. To improve resolution at the Fab–RBD interfaces, volumes were segmented in Chimera60 and the regions corresponding to the NTD and RBD domains of the S1 subunit and the Fab Vh–Vl domains were extracted and used to generate a soft mask (5-pixel extension, 10-pixel soft cosine edge). Local refinements with the mask resulted in modest improvements of the Fab–RBD interface, which allowed for fitting and refinement of this region. The particles were then subjected to CTF refinement and aberration correction, followed by a focused, non-uniform refinement with polished particles imposing C1 symmetry (except for the C144–S complex, in which C3 symmetry was used). Final overall resolutions were according to the gold-standard FSC61. Details of overall resolution and locally refined resolutions according to the gold-standard FSC61 can be found in Supplementary Table 2.

Cryo-EM structure modelling and refinement

Coordinates for initial complexes were generated by docking individual chains from reference structures into cryo-EM density using UCSF Chimera62 v.1.13. The following coordinates were used: SARS-CoV-2 S trimers: PDB codes 6VXX, 6VYB and 6XKL, up RBD conformations: PDB codes 7BZ5 or 6W41, and unbound C102, C002, C110, C121, C135 Fab structures (this study) (Supplementary Table 1). Initial models were then refined into cryo-EM maps using one round of rigid body refinement followed by real space refinement. Sequence-updated models were built manually in Coot56 v.0.8.9 and then refined using iterative rounds of refinement in Coot56 and Phenix49. Glycans were modelled at potential N-linked glycosylation sites in Coot56 using ‘blurred’ maps processed with a variety of B-factors63. Validation of model coordinates was performed using MolProbity64 (Supplementary Table 2).

Structural analyses

CDR lengths were calculated based on IMGT definitions32. Structure figures were made with PyMOL (v.2.2 Schrodinger, LLC) or UCSF ChimeraX60 v.1.0. Local resolution maps were calculated using cryoSPARC v.2.1558. Buried surface areas were calculated using PDBePISA v.1.4865 and a 1.4 Å probe. Potential hydrogen bonds were assigned as interactions that were less than 4.0 Å and with an A-D-H angle above 90°. Potential van der Waals interactions between atoms were assigned as interactions that were less than 4.0 Å. Hydrogen bond and van der Waals interaction assignments are tentative due to resolution limitations. r.m.s.d. calculations following pairwise Cα alignments were done in PyMOL without rejecting outliers. Criteria for epitope assignments are described in figure legends.

To evaluate whether intra-spike crosslinking by an IgG binding to a single spike trimer was possible (Extended Data Table 1), we first measured the Cα distance between a pair of residues near the C termini of adjacent Fab Ch1 domains (residue 222 of the heavy chain on each Fab) (Extended Data Fig. 5h). We compared this distance to the analogous distances in crystal structures of intact IgGs (42 Å, PDB code 1HZH; 48 Å, PDB code 1IGY; 52 Å, PDB code 1IGT). To account for potential influences of crystal packing in these measurements, as well as flexibility in the Vh–Vl/Ch1–Cl elbow bend angle and uncertainties in Ch1–Cl domain placement in Fab–S cryo-EM structures, we set a cut-off of ≤65 Å for this measured distance as possibly allowing for a single IgG to include both Fabs. Entries in the ‘potential IgG intra-spike binding’ column in Extended Data Table 1 are marked ‘no’ if all of the adjacent Fabs in cryo-EM classes of that structure are separated by more than 65 Å for this measured distance. Entries in the ‘potential IgG intra-spike binding’ column in Extended Data Table 1 are marked as ‘yes’ if at least one pair of the adjacent Fabs in cryo-EM classes of that structure are separated by ≤65 Å for this measured distance.

SPR binding experiments

SPR experiments were performed using a Biacore T200 instrument (GE Healthcare). IgGs were immobilized on a CM5 chip by primary amine chemistry (Biacore manual) to a final response level of approximately 3,000 resonance units. Concentration series of the original SARS-CoV-2 RBD and RBD mutants (six fourfold dilutions starting from a top concentration of 1,000 nM) were injected at a flow rate of 30 μl min over immobilized IgGs for a contact time of 60 s, followed by an injection of 0.01 M HEPES, pH 7.4, 0.15 M NaCl, 3 mM EDTA, 0.005% (v/v) surfactant P20 buffer for a dissociation time of 300 s. Binding reactions were allowed to reach equilibrium, and Kd values were calculated from the ratio of association and dissociation rates (Kd = kd/ka) derived from a 1:1 binding model (C002, C102, C105, C110 and C119 (except for C119 mutant E484K)), C121, C135 and C144), or from a two-state binding model (Kd = kd1/ka1 × kd2/[kd2 + ka2]) (C104 and C119 mutant E484K). Kinetic constants were calculated using Biacore T200 Evaluation Software v.3.2 using a global fit to all curves in each dataset. Flow cells were regenerated with 10 mM glycine, pH 2.0, at a flow rate of 90 μl min−1.

Polyreactivity assays

IgGs were evaluated for off-target interactions by measuring binding to baculovirus extracts containing non-specific proteins and lipids as described59. The assays were automated on a Tecan Evo2 liquid handling robot fitted with a Tecan Infinite M1000 plate reader capable of reading luminescence. Maxisorb 384-well plates (Nunc) were adsorbed overnight with a 1% preparation of recombinant baculovirus particles generated in Sf9 insect cells66. The adsorbed plate was blocked with 0.5% bovine serum albumin (BSA) in PBS, then incubated with 20 μl of a 1.0 μg ml−1 solution of IgG in PBS for 3 h. Polyreactivity was quantified by detecting bound IgG using an HRP-conjugated anti-human IgG secondary antibody (SouthernBiotech) at a 1:5,000 dilution and SuperSignal ELISA Femto Maxiumum Sensitivity Substrate (Thermo Scientific). RLUs were measured at 475 nm in the integrated plate reader. Engineered human anti-HIV-1 IgGs previously demonstrated to exhibit high levels of polyreactivity (NIH45-46(G54W) and 45-46m2)60,61 were used as positive controls. NIH45-46, which exhibited intermediate polyreactivity62, was also evaluated for comparisons. Negative control IgGs with low polyreactivity included the human HIV-1 antibodies N663 and 3BNC11762 and BSA. RLU values were plotted in GraphPad Prism v8.4.3 and presented as the mean and standard deviation of triplicate measurements (n = 3 biological replicates) with results for individual experiments shown as circles in Extended Data Fig. 1i.

Reporting summary

Further information on research design is available in the Nature Research Reporting Summary linked to this paper.