Broadly neutralizing SARS-CoV-2 antibodies through epitope-based selection from convalescent patients

Epitope-based selection of human memory B cells using mutant SARS-CoV-2 RBDs

We utilized peripheral blood mononuclear cells (PBMCs) from convalescent patients from the COSIN study (New South Wales COVID-19 patient cohort; patients diagnosed by RT-PCR in March 2020 and follow-up samples collected between May and November 2020)²², at 1- and 4-months post SARS-CoV-2 infection. In order to rapidly identify antibodies binding to different epitopes, we sorted memory B cells based on their capacity to bind to RBD variants carrying epitope class-specific mutations. We initially investigated whether to utilize full spike glycoprotein (trimeric; ancestral strain;²³ randomly biotinylated) or recombinant RBD (single biotin-tag) to select human memory B cells. Using fluorescently-labeled tetramerized spike and tetramerized RBD, we observed a population of CD3^–CD19⁺CD20⁺CD10^–IgD^–IgG⁺ B cells with a high mean of fluorescence intensity (MFI) for both the RBD and spike (albeit lower MFI for spike, Fig. 1a left panel and Supplementary Fig. 1 – Sort 1). Subsequent sorting (Fig. 1a right panels, Supplementary Fig. 1; Sort 2 and Sort 3) focused on the use of tetramerized RBD and a series of SARS-CoV-2 RBD mutants to target different epitope classes. We specifically designed RBD mutants to differentiate antibodies blocking the ACE2 interactions and antibodies binding outside the ACE2 binding site. For this purpose, we utilized the following mutant SARS-CoV-2 RBDs: Mut1 (T500A/N501A/Y505A), perturbing the ACE2 binding site (targeting class 1 and 2 antibody binders); Mut2 (L455A/F456A), blocking a different surface of the ACE2 binding site (further targeting further class 1 and 2 binders); and Mut3 (K378S), blocking the CR3022 antibody epitope (targeting class 4 binders) (Fig. 1b)²⁴. In addition to mutant RBDs, we also utilized SARS-CoV-1 RBD in order to identify broadly neutralizing antibodies (all RBDs were tetramerized with four distinct fluorescent dyes (Supplementary Fig. 1). Having observed RBD/spike cross-specificity (Sort 1), the next strategy (Sort 2) employed SARS-CoV-1, SARS-CoV-2 and Mut1 and Mut3 RBDs, classifying cells as falling into class 1/2, 4 or unknown epitope classes (Fig. 1a and Supplementary Fig. 1). We consolidated these findings using a third sorting strategy (Sort 3) employing WT SARS-CoV-2, Mut1, Mut2 and Mut3 (Fig. 1a and Supplementary Fig. 1 – Sort 3), strengthening delineation of binders that compete with the ACE2 interface.

a Memory B cells were selected from convalescent patient PBMCs by gating on CD3^–CD19⁺CD20⁺CD10^–IgD^–IgG⁺ cells (Supplementary Fig. 1; upper panel) and single cells sorted by binding to fluorescently labeled SARS-CoV-2 RBD or trimeric spike initially (left panel). Red dots represent the B cells for which monoclonal antibodies were amplified and characterized (Supplementary Fig. 1; Sort 1). In addition, using b mutant RBD protein (mutations colored in orange), single cells were sorted based on FACS epitope bins (a, right panels, Supplementary Fig. 1; Sort 3). c Monoclonal antibodies analyzed, showing affinity for SARS-CoV-2 RBD (ancestral strain), neutralization potential and epitope class (initial FACS epitope bin and validation). nn = non neutralizing. In vitro neutralization of live (d) ancestral and (e) Delta VOC SARS-CoV-2 virus. n = 4 technical replicates, data are presented as mean values. Source data are provided as a Source Data file.

Epitope binning and characterization of human monoclonal antibodies

Following the indexed sorting strategies, human antibody variable regions were amplified from isolated B lymphocytes, cloned into human IgG1 expression vectors, and the encoded monoclonal antibodies expressed and purified^25,26. For an initial screen, we assessed antibody binding by ELISA and found that 80% (16/20) of the antibodies selected bound to WT SARS-CoV-2 RBD (Supplementary Fig. 2a). We then measured the antibody binding affinities to SARS-CoV-2 RBD by BioLayer Interferometry (BLI), which ranged from 290 nM to 200 pM (Fig. 1c and Supplementary Fig. 2b).

To further validate class 1/2 antibodies, we performed BLI competition assays with recombinant human ACE2. Amongst the 16 antibodies screened, we identified seven that competed with ACE2 and thus could be considered as class 1/2 candidates (GAR04, 05, 06, 07, 09, 15 and 20) (Supplementary Fig. 3). All but one of these (GAR04) were capable of binding the E484K RBD mutant, a residue pivotal to the binding mode of many class 2 binders⁶, hence we tentatively assigned GAR04 as a class 2 binder, while the other six antibodies were assigned as class 1 binders (Fig. 1c). IGHV3-53, IGHV1-2 and IGHV3-30 germlines have been commonly observed among class 1 and class 2 SARS-CoV-2 binders²⁷, and we also observed these germlines among several of these selected clones (including GAR01, GAR04, and GAR15; Supplementary Table 3 and Supplementary Table 4).

To further validate class 4 antibodies, we first investigated cross-reactivity with other sarbecovirus RBDs, which are known to be relatively conserved within this epitope (SARS-CoV-1, pangolin CoV and bat RaTG12-CoV)^12,13. We performed BLI binding experiments with these three sarbecovirus RBDs and identified seven cross-reactive antibodies (GAR01, 03, 11, 13, 14, 16 and 20) (Supplementary Fig. 4). BLI competition assays performed on these cross-reactive antibodies with the class 4 monoclonal antibody CR3022^28,29 (known to bind both SARS-CoV-1 and SARS-CoV-2), confirmed that all, except GAR01 and GAR03, bind to the class 4 epitope (Supplementary Fig. 3). In addition to competing with CR3022, GAR20 also blocked binding of recombinant ACE2 to SARS-CoV-2 RBD, as observed for antibody 3467¹⁴, however this was not observed for the other class 4 antibodies.

The finding that GAR01 and GAR03 are broadly cross-specific and bind to epitopes conserved in SARS-CoV-2, SARS-CoV-1, pangolin CoV and bat RaTG12-CoV (but do not compete with ACE2 nor CR3022) indicate that these two antibodies may be rare binders falling into the class 5 epitope bin. To investigate this hypothesis, we performed BLI competition assays with the class 5 antibody S2H97¹², and confirmed that binding of both antibodies is indeed consistent with the class 5 classification (Supplementary Fig. 3).

Overall, our epitope-binning strategy assigned most of the antibodies to established class 1-5 epitope bins (Fig. 1c)³⁰. In addition, antibodies originally sorted into unknown epitope bins were further identified as falling either into epitope class 1/2 (unaffected by T500A/N501A/Y505A and L455A/F456A RBD mutations) or alternately as falling into a new epitope class (class 6, as below).

Next, we evaluated the capacity of these antibodies to neutralize live virus. We observed that 12/16 of the monoclonal antibodies neutralized SARS-CoV-2 ancestral (D614G) strain with IC₅₀s ranging from 20 µg/ml to 12 ng/ml (Fig. 1c, d). Several (6/16) antibodies also neutralized the Delta VOC (GAR05, GAR07, GAR09, GAR12, GAR15 and GAR20) (Fig. 1e). Three of these antibodies, GAR05, GAR12 and GAR20, were particularly broad in their specificity and neutralized all analyzed variants of concern, with GAR05 neutralizing with IC₅₀s ranging from 115 to 26 ng/ml, and GAR12 and GAR20 neutralizing with IC₅₀s ranging from 255 to 128 ng/ml and 12 to 4 µg/ml respectively (Supplementary Fig. 5).

GAR05 is a broadly neutralizing class 1 antibody

Based on its broad neutralization potential, antibody GAR05 was further characterized by cryo-EM and X-ray crystallography. For cryo-EM studies, GAR05 Fab was incubated with stabilized D614G ancestral trimeric spike (“VFLIP”)³¹ at a 3:1 molar ratio. The complex was subsequently flash-frozen and examined using cryo-EM (Supplementary Fig. 6 and Supplementary Table 1). The cryo-EM map of GAR05 bound to trimeric spike (Fig. 2a) shows three antibody Fabs bound to the RBD domains of the trimeric spike, where all three RBDs are in the up conformation, consistent with its classification as a class 1 antibody (similar to what has been observed for antibody C105³²). We also solved the crystal structure of GAR05 as a Fab fragment in complex with SARS-CoV-2 RBD to 3.2 Å resolution (Fig. 2b, Supplementary Table 2), essentially confirming the binding mode suggested by the cryo-EM map. GAR05 straddles the middle of the saddle-like RBD surface known to interact with the ACE2 receptor (Fig. 2b). The antibody forms a large RBD-interface (~1275 Ų, green surface in Fig. 2d), 815 Ų of which is contributed by the heavy chain. Heavy-chain complementarity determining regions (CDRs) form a cleft bordered on one side by two finger-like loops (CDRs H1 and H2), opposite a thumb-like loop (extended CDR H3) between which the saddle of the RBD (residues 470-492, Fig. 2c) is bound. CDRs L1 and L3 of the light chain also contact the RBD, forming an additional cleft accommodating the side chain of residue Y505, which projects from an adjacent surface of the RBD (Fig. 2c). The large surface area buried at the interface is at the upper end of what is commonly observed for antibody/antigen interactions^33,34, and is likely reflected by the high affinity of GAR05 towards the RBD (540 pM). However, the surface coverage of GAR05 is in close proximity to multiple positions mutated in various Omicron lineages (Fig. 2d). Rationalisation of why tight binding of GAR05 is nevertheless observed can be made by superposition of the GAR05 complex with cryo-EM and crystal structures of an ensemble of Omicron variants (B1.1.529, BA.2, BA.4/5) from a range of biophysical circumstances (RBD down, RBD up, RBD up and complexed by ACE2, see Supplementary Fig. 7). Of the constellation of positions mutated in Omicron VOCs, eight are likely to interface with GAR05. Four of these (S477N, T478K, E484A and F486V) adorn a loop demonstrating considerable flexibility and which forms one end of the saddle bound by ACE2 (Supplementary Fig. 7b). Some positions are mutated to smaller side chains and thus are unlikely to present a steric clash (E484A, F486V, K417N and Y505H; Supplementary Fig. 7, panels b-f). Some positions are mutated to larger side chains albeit presenting in a variety of side chain conformations, some of which might accommodate GAR05 binding (Q493R and N501Y; panel f). Ten hydrogen bonds exist between GAR05 and the RBD (panels g-i). Two of these involve the light chain epitope centered on Y505, both of which will likely be lost by Omicron mutants Y505H and N501Y (panel g). The remaining 8 are heavy-chain centric, 7 involved in contacts with conserved RBD targets (both main-chain and side-chain), and only one will be lost by the E484A mutation (panel i). Hence, the ability of GAR05 to maintain tight binding to Omicron VOCs appears due to combinatorial effects of; coverage of a large epitope surface, the redundancy of heavy and light chains effectively targeting different surfaces (heavy chain targeting the RBD saddle feature, and the light chain targeting the adjacent Y505 feature), the apparent flexibility noted in part of the RBD heavy-chain epitope, the lack of any obvious side-chain mutation likely to present as an unavoidable steric block, and the bulk of hydrogen bonds targeting conserved features.

a Cryo-EM structure of GAR05 Fab bound to trimeric spike (3.27 Å resolution) showing full antibody Fab occupancy of all the RBDs in the “up” conformation. Two perspectives are shown b Structure of GAR05 bound to SARS-CoV-2 RBD based on the X-ray crystal structure, outlining the antibody bound to the ACE2 “saddle” of the RBD (ACE2 interface shaded black). c Interaction of the CDR regions of the VH and VL domains with the RBD saddle and the Y505 side chain of the RBD. d Comparison of the ACE2 interface on the surface of RBD and the GAR05 epitope, showing high overlap within the ACE2 saddle region. The large, buried surface area of GAR05 (1275 Ų) indeed blankets many key residues identified in VOCs, colored in yellow, yet remarkably still binds all VOCs with high affinity.

The binding mode of GAR05 is similar to that of mAb P2C-1F11, which has been described as an ACE2 mimetic due to sharing extensive steric clash volume with ACE2³⁵. A primary difference, however, is the higher affinity and extremely long HCDR3 loop of GAR05, which wraps further than P2C-1F11 around the RBD saddle targeted by ACE2 (Supplementary Fig. 8, top panels). A further class I targeting antibody, S2K146³⁶, has also been described as an ACE2 mimetic as it targets multiple evolutionarily conserved residues utilised by ACE2, and displays broad resistance to VOCs. The overall binding mode of S2K146 resembles that of GAR05, with the notable exception that the orientation of the antibody heavy and light chains are reversed (approximately 180-degree rotation; Supplementary Fig. 8, lower panels).

Structural identification of a novel class 6 epitope

We next performed structural studies on two broadly cross-reactive antibodies, GAR03 and GAR12, binding conserved epitopes in proximity to that of antibody S2H97¹². Initial cryo-EM studies using a GAR03-spike mix that had been incubated for 30 min were unsuccessful due to significant aggregation and sample degradation. However, a shorter 1 min incubation of GAR03:spike at a 3:1 ratio improved sample integrity and this preparation was utilized to solve the structure of GAR03 bound to spike glycoprotein (Supplementary Fig. 6 and Supplementary Table 1). Examination of the cryo-EM map revealed that density could not be observed for the N-terminal domain (NTD) and RBD of one of the three protomers (essentially an S1 domain), with the GAR03 Fab instead occupying the space of the NTD while being bound to the RBD of an adjacent protomer (Fig. 3a). Given that the S2 domain of all three protomers are visible, we hypothesize that binding of GAR03 to an RBD causes the NTD and RBD of an adjacent protomer to be displaced and become flexible, with the averaging methods employed in single particle analysis cryo-EM making their location unidentifiable. Subsequently, we crystalized the GAR03 Fab in complex with the RBD and solved the structure by X-ray crystallography to 2.75 Å resolution (Fig. 3a, middle and right-hand panels, and Supplementary Table 2), confirming the binding mode suggested by cryoEM, whereby the bulk of the contact interface is mediated by the antibody light chain, and where the binding epitope is well separated from Omicron variant mutation positions (Fig. 3a, positions highlighted in yellow). Competition binding assays suggested that GAR03 binds to an epitope shared with the class 5 antibody S2H97 (Supplementary Fig. 3). Surprisingly, however, mapping of the GAR03 structure onto the S2H97-RBD structure (PDB 7m7w)¹² revealed that, although GAR03 and S2H97 block each other from binding to the RBD, the epitope overlap is minimal (Fig. 3c, d – overlap colored green). More specifically, the GAR03 epitope occupies RBD surface extending from the class 5 epitope exemplified by S2H97 to the class 3 epitope site, exemplified by the monoclonal antibody Sotrovimab (S309, PDB 7r6x¹², Fig. 3c – S2H97 in yellow, S309 in pink, and GAR03 in shades of blue). This epitope is highly conserved among sarbecoviruses, as highlighted by the cross-reactivity of GAR03 with SARS-CoV-2, SARS-CoV-1, Pangolin CoV and Bat RaTG12-CoV RBD. BLI competition experiments revealed that two other effectively neutralizing antibodies, GAR12 and GAR17, competed both with GAR03 and S309 (class 3 epitope) (Supplementary Fig. 9). However, the lack of competition with S2H97 suggests that GAR12 and GAR17 presumably target regions more oriented towards the class 3 epitope site (Fig. 1c). To validate this hypothesis, we solved the crystal structure of the GAR12 Fab in complex with RBD to 2.25 Å resolution (Fig. 3b and Supplementary Table 2). Indeed, GAR12 binds to an epitope on the surface of the RBD that overlaps with S309 and abuts the ACE2 contact surface (Fig. 3b–d, orange cartoon and surfaces). The GAR12 epitope broadly avoids residues mutated in VOCs, thus demonstrating the broad and effective neutralization of this antibody (Fig. 3b). We designate this epitope, shared by GAR03 and GAR12 and spanning a triangular surface between class 5, class 3 and class 1, as class 6 (Fig. 3c).

a Cryo-EM structure of spike trimer (three shades of grey surface) complexed with GAR03 (blue cartoon and transparent surface) revealing a single GAR03 antibody Fab bound to trimeric spike (resolution 3.39 Å). Binding displaces the S1 domain (N-terminal domain and RBD) of one of the protomers (disordered and not visible in the cryo-EM map), with the GAR03 Fab binding to the RBD of the adjacent protomer. Middle and right panels show two perspectives of the RBD-GAR03 crystal structure. The binding interface is dominated by the GAR03 light chain (light blue cartoon; CDRs L1, L2, and L3) with more minor contribution from the heavy chain (dark blue; CDR H3). The ACE2 interface of the RBD (grey surface and cartoon) is shaded black, whilst Omicron VOC mutation positions are indicated by yellow sticks and surface, and N343 linked carbohydrate is shown as sticks. b Crystal structure of GAR12 (heavy and light chains colored dark and light orange) in complex with RBD (grey cartoon and surface, with ACE2 interface colored black, and Omicron VOCs colored yellow). Two perspectives are shown. c Global juxtaposition of GAR03 (blue surfaces) and GAR12 (orange surfaces) with class 5 representative S2H97 (yellow) and class 3 representative S309 (pink) shown against the RBD surface (grey, and black indicating the RBD-ACE2 interface). d RBD-interface comparisons. Left and middle panels including GAR03 (blue) and S2H97 (yellow) and their overlap (green). The right panel additionally includes the GAR12 (orange) and S309 interfaces (pink), and their overlap (red). e Spike trimer (three shades of grey surface) presented with all three RBDs in the down conformation (PDB entry 6xm5) as viewed down the approximate 3-fold axis (left and middle panels); docked Fabs (middle panel) GAR03 (blue cartoons), GAR05 (green cartoons), GAR12 (orange cartoons), as well as the primary helix from ACE2 (yellow cartoon). Modelling indicates that only GAR12 is capable of binding to spike with RBD in the down conformation (right panel, viewed from the side).

Superposition of GAR12, GAR03 and GAR05 antibodies onto the trimeric spike protein with the RBDs in the down position indicates that GAR12 is unimpeded and capable of binding to spike in the fully down position (Fig. 3e – orange Fab cartoons). In contrast, neither GAR03 (blue Fab cartoons) nor GAR05 (green Fab cartoons) can access their respective RBD epitopes in the down position due to steric obstruction from either neighboring N-terminal (both) or RBD domains of adjacent protomers (GAR05) within the spike trimer (Fig. 3e).

Live virus challenge in the K18-hACE2 mouse model

To further validate monoclonal antibodies identified in this study, live virus challenge experiments were carried out using the K18-hACE2 mouse model of SARS-CoV-2 infection in prophylaxis and therapeutic treatments^37,38. For this purpose, we selected three monoclonal antibodies with broad neutralization of VOCs (Fig. 1c, e, Supplementary Fig. 5) and non-overlapping epitopes (Supplementary Fig. 9), to enable future application as combination therapy. More specifically, we selected GAR05 for class 1, GAR12 for class 6 and GAR20 for class 4.

We initially evaluated the antibodies in a prophylaxis model using live SARS-CoV-2 virus (ancestral). Mice were injected intraperitoneally with 30 mg/kg of monoclonal antibody three days prior to challenge with 1 × 10³ PFU live virus. Mice administered with monoclonal antibodies maintained consistent body weight throughout the challenge compared to mice administered with human IgG1 isotype control (Fig. 4a). Moreover, treatment provided considerable protection from clinical symptoms, with only control mice displaying severe scores (Fig. 4b). Viral titers in lung homogenates from mice treated with monoclonal antibodies were significantly reduced compared to isotype control and were generally below the detection limit (Fig. 4c). Investigation of the bronchoalveolar lavage fluids (BALF) revealed that all three treatment groups had statistically significant reductions in inflammatory innate immune cells (macrophages and neutrophils), while maintaining undetectable levels of lymphocytes (Supplementary Fig. 10).

Animals were injected interperitoneally with 30 mg/kg of GAR05, GAR12, GAR20 or human IgG1 isotype control 3 days prior to viral challenge (-3 dpi) for prophylactic studies (a–c, g–i) or +1 dpi post challenge for therapeutic studies (d–f). Mice were challenged (d0) with 1 × 10³ PFU of intranasal ancestral SARS-CoV-2 or Delta VOC (B.1.167.2) virus and monitored for 6 days, and euthanized (d6) for tissue collection. Measurements of weight (percentage loss from initial weight, a [n = 4 mice per group measured on 6 consecutive days. 2-way ANOVA with Tukey’s multiple comparisons test. *p < 0.05], d [n = 6 mice per group measured on 6 consecutive days. A single isotype control mouse was removed from the study on day 5 as it reached a humane endpoint. Mixed-effects analysis with Šídák’s multiple comparisons test. *p = 0.0149], g [n = 7 istoype control mice and n = 6 GAR05 treated mice measured on 6 consecutive days. 2-way ANOVA with Tukey’s multiple comparisons test. **p = 0.0049]), clinical score (weight loss, eye closure, appearance of fur and posture, and respiration, b/e/h), and viral titers (lung homogenates; plaque assay, c [n = 4 mice per group. Kruskal-Wallis test of non-parametric data with Dunn’s multiple comparisons test. *p < 0.05, **p < 0.01], f [n = 6 mice per group. Two-tailed unpaired t-test. ****p < 0.0001], i [n = 7 istoype control mice and n = 6 GAR05 treated mice. Two-tailed unpaired t-test. *p = 0.0179]) shown. Data are presented as mean values + /- SEM. Source data are provided as a Source Data file.

In a subsequent study, mice were challenged with ancestral SARS-CoV-2 in a therapeutic setting, with GAR05 administered post-viral challenge at 30 mg/kg. This experiment revealed that, as had been observed in a prophylactic setting, mice treated with GAR05 were fully protected from viral challenge with no measurable weight loss (Fig. 4d), reduction in clinical scores (Fig. 4e) or detectable lung viral titers (Fig. 4f). In a third in vivo setting, GAR05 was used as a prophylactic modality (as above), and mice challenged with the Delta (B.1.617.2) VOC. As had been observed for the initial live virus challenge (ancestral), mice treated with GAR05 were protected from the viral challenge with the Delta VOC: indeed, no measurable reduction of either weight (Fig. 4g), clinical scores (Fig. 4h) could be observed, with undetectable viral titers in the prophylaxis group (Fig. 4i).

Taken together, these findings demonstrate that the monoclonal antibodies developed here, targeting non-overlapping epitopes, broadly and effectively protect human ACE2 mice from SARS-CoV-2 live virus challenge, highlighting their potential for therapeutic applications.

Broad neutralization of Omicron VOCs

The Omicron VOC and its subvariants have caused worldwide outbreaks, highlighting the importance of broad antibody neutralization and mutational resistance^39,40. When Omicron BA.1 (B.1.1.529) first emerged in November 2021, we and others had shown that only 1/6 monoclonal antibodies in clinical practice fully maintained activity, namely the broadly neutralizing class 3 monoclonal Sotrovimab (S309)^10,41,42. To assess the potency of GAR05 and GAR12 against Omicron BA.1 and subvariants, we performed in vitro neutralization assays using live virus. We observed that GAR05 and GAR12 maintained their activity against the analyzed Omicron VOCs (Fig. 5a), with IC₅₀ values ranging from 16.17 ng/ml to 337.6 ng/ml, considerably exceeding those of Sotrovimab (S309)⁴¹, particularly for the BA.2 subvariant which is not neutralized by Sotrovimab (Fig. 5b).

a SARS-CoV-2 neutralization (live virus), ancestral (A.2.2), Omicron VOCs BA1, BA2, BA5 (n = 4 technical replicates). b IC₅₀ values (ng/ml) reported for S309, GAR05 and GAR12 on ancestral and Omicron VOCs. nn = non-neutralizing. Data are presented as mean values + /- SD for measurements performed over a range of concentrations. Source data are provided as a Source Data file.