Memory B cell repertoire for recognition of evolving SARS-CoV-2 spike

Abstract

Memory B cell reserves can generate protective antibodies against repeated SARS-CoV-2 infections, but with an unknown reach from original infection to antigenically drifted variants. We charted memory B cell receptor-encoded monoclonal antibodies (mAbs) from 19 COVID-19 convalescent subjects against SARS-CoV-2 spike (S) and found 7 major mAb competition groups against epitopes recurrently targeted across individuals. Inclusion of published and newly determined structures of mAb-S complexes identified corresponding epitopic regions. Group assignment correlated with cross-CoV-reactivity breadth, neutralization potency, and convergent antibody signatures. mAbs that competed for binding the original S isolate bound differentially to S variants, suggesting the protective importance of otherwise-redundant recognition. The results furnish a global atlas of the S-specific memory B cell repertoire and illustrate properties conferring robustness against emerging SARS-CoV-2 variants.

INTRODUCTION

Coronavirus (CoV) disease 2019 (COVID-19) caused by the severe acute respiratory syndrome (SARS) CoV-2 virus has rapidly become a pandemic of historic effect. Although vaccines have been developed in record time, new variants continue to emerge and threaten to evade immune responses. We need to understand immune recognition of SARS-CoV-2, especially as stored in B cell memory, to illuminate the requirements for broad protective immunity in humans. We focus on B cells, because antibodies, a key part of the immune defense against most viruses, are sufficient to protect against SARS-CoV-2 infection in animal models (1, 2).

Antibodies are both soluble effector molecules and the antigen-receptor component of the B cell receptor (BCR). BCRs evolve enhanced pathogen binding through immunoglobulin (Ig) gene somatic hypermutation (SHM) and selection in lymphoid tissue germinal centers (GCs), leading to antibody affinity maturation (3) and generation of both antibody-secreting plasma cells (PCs) and memory B cells. Higher avidity interactions encourage terminal differentiation of B cells into PCs; memory B cells frequently have lower avidity but more cross-reactive specificities (4).

Both PC-derived secreted antibody and memory B cells supply immune memory to prevent repeat infection, but with non-redundant roles. Secreted antibodies can prophylactically thwart pathogen invasion with fixed recognition capability, while memory B cells harbor expanded pathogen recognition capacity and can differentiate quickly into PCs to contribute dynamically to the secreted antibody repertoire (4). Moreover, memory B cells retain plasticity to adapt to viral variants through GC re-entry and SHM-mediated evolution (5).

The viral spike (S) glycoprotein binds ACE2 on host cells and mediates viral fusion with the host (6). Its fusogenic activity depends on a furin-mediated cleavage, resulting in N-terminal S1 and C terminal S2 fragments (7) and on a subsequent cleavage of S2 mediated either by cathepsins or by a serine protease, TMPRSS2 (8). The S glycoprotein is the principal neutralizing antibody target and the focus of most vaccines. SARS-CoV-2 S antibodies decline with time (9, 10) and can lose reactivity to emerging variants (11). Antibodies cloned from memory B cells target the S glycoprotein in redundant as well as unique ways, indicating cooperative and competitive recognition (12–17). Many of these antibodies have been identified and characterized; their positions within the distribution of practical cooperative recognition of SARS-CoV-2 S within the human memory B cell repertoire have not. Moreover, the recognition reach of memory B cells induced by one SARS-CoV-2 strain toward evolving stains across the major epitopic regions has not yet been defined.

We present here an unbiased global assessment of the distribution of memory B-cell encoded antibodies among cooperative and competitive recognition clusters on the SARS-CoV-2 S glycoprotein and assess features that direct their collaborative robustness against emerging SARS-CoV-2 variants. In a comprehensive competition analysis of 152 monoclonal antibodies (mAbs) from 19 subjects for binding with trimeric S ectodomain, we have identified 7 recurrently targeted competition groups -- three for antibodies with epitopes on the receptor-binding domain (RBD), two for epitopes on the N-terminal domain (NTD), and two for S2 epitopes. We show that these groups represent the major practical antibody footprints, with rare antibodies outside them. We map the clusters onto the S glycoprotein by including previously characterized antibodies and new cryo-EM determined structures. Ig repertoire analysis indicates both divergent and convergent clones with the competition groups.

Antibodies mapped to RBD-2 and NTD-1 were the most potent neutralizers, while the S2–1 group has the greatest recognition breadth across CoVs. The emerging SARS-CoV-2 variants, particularly the South Africa strain, strongly affected the antibodies in one of the RBD and one of the NTD clusters. The mutations in those variants differently influenced affinity of antibodies within a competition group, indicating that the depth of otherwise redundant mAbs to a given S variant confers recognition breadth for dynamically mutating S.

RESULTS

Monoclonal antibody (mAb) isolation

To identify the general pattern of SARS-CoV-2 S recognition by memory B cells in convalescent subjects, we sorted single CD19⁺ CD27⁺ IgG⁺ B cells recognizing soluble prefusion-stabilized S trimer (Fig. 1A, Fig. S1) from 19 individuals with a history of COVID-19 (Data S1). Because less is known about S-reactive antibodies that bind outside the RBD region, we also sorted S-reactive B cells that did not bind RBD from 3 individuals. S-reactive B cells made up 0.2% (0.07%–0.4%) of the total B cell population (Fig. 1A left panel), with RBD-binding cells representing about a quarter of S-reactive IgG⁺ B cells (Fig. 1A right panel) consistent with prior work (18).

Fig. 1. SARS-CoV-2 surface glycoprotein (spike) specificities of memory B cells from convalescent subjects.

(A) Cells recovered from two sorting strategies, shown in dot plots as percentages of total CD19⁺ cells. Left: IgG⁺CD27⁺ cells from 18 donors (one dot per donor) and the subset of those that expressed spike-binding BCRs. Right: cells from 3 donors expressing spike-binding BCRs and sorted to recover principally those that did not bind recombinant receptor-binding domain (RBD). Sorting protocols as described in Methods and shown in Fig. S1. (B) Summary of all antibodies (expressed as recombinant IgG1) screened by ELISA (with recombinant spike ectodomain trimer) and cell-surface expression assays (both 293T and yeast cells). Total numbers in the center of each of pie chart; numbers and color codes for the indicated populations shown to next to each chart. To the right of the charts for the two alternative sorting strategies are bar graphs showing frequencies of SARS-CoV-2 RBD and NTD binding antibodies for those subjects from whom at least 10 paired-chain BCR sequences were recovered. (C) Binding to a panel of spike proteins and SARS-CoV-2 subdomains, listed on the left, as determined by both ELISA (with recombinant spike ectodomain) and by association with spike expressed on the surface of 293T cells or with RBD or NTD expressed on the surface of yeast cells, for cells sorted just for spike binding (left) and for those sorted for positive spike binding but no RBD binding (right). The rows with pink highlighting are from the ELISA screen; those with blue highlighting, from the cell-based screens. Each short section of a row represents an antibody. The rows labeled VH mutation and VL mutation are heat maps of counts (excluding CDR3) from alignment by IgBLAST, with the scale indicated. (D) Dot plots of heavy- and light-chain somatic mutation counts in antibodies that bound RBD, NTD, S2, and a “broad CoV group” that included MERS, HKU1, and OC43. The significantly higher numbers of mutations in the last group suggest recalled, affinity matured memory from previous exposures to seasonal coronaviruses. ****P < 0.0001; one-way ANOVA followed by Tukey’s multiple comparison. Horizontal lines show mean ± SEM.

mAb binding

We cloned cDNAs encoding Ig heavy (H) and light (L) chains from individual, sorted memory B cells into human IgG1 and kappa or lambda vectors and expressed them in HEK 293T cells. We detected IgG in 255 of the culture supernatants, which we used to screen for binding of SARS-CoV-2 S (Fig. 1 and fig. S2). Of the 255 IgGs, 216 bound SARS-CoV-2 S expressed by HEK 293T cells, as assayed by flow cytometry (157 from the S⁺ sorting, 59 from S⁺/RBD⁻ sorting) (fig. S2A) and 166 of the 216 bound recombinant SARS-CoV-2 S, as assayed by ELISA (116 from the S⁺ sorting, 50 from the S +RBD⁻ sorting).

We estimated, by ELISA and, where possible, yeast display of the subdomains (fig. S2B and C), the proportion of mAbs that bound to RBD, NTD, and S2. Recombinant and yeast-displayed S2 protein could have any of several conformations, and all or parts of the polypeptide chain might be disordered; antibodies that bound S2 on ELISA plates might therefore tend to recognize linear epitopes or even the S2 post-fusion conformation. Indeed, most of the S2-binding antibodies had relatively low ELISA-determined affinities for intact, recombinant, prefusion S, although a few bound more tightly to S expression on the surface of 293T cells (fig. S2D). Of the 157 S-reactive mAbs sorted with SARS-CoV-2 stabilized S trimer, a total of 37 (23%) were RBD-specific as assayed by ELISA, by yeast display, or both (Fig. 1B). We detected 16 (10%) mAbs that bound the NTD and 49 (31%) that bound recombinant S2 (Fig. 1C). Eleven of the 49 S2 binders bound cell-surface-expressed, but not ELISA-based SARS-CoV-2 S.

We also assessed mAbs by ELISA for cross-reactivity to other CoV S glycoproteins. Those of SARS (GenBank: MN985325.1), MERS (GenBank: JX869059.2), and common cold β-CoVs HKU1 (GenBank:Q0ZME7.1) and OC43 (GenBank: AAT84362.1) have sequences with 75.8%, 28.6%, 25.1%, and 25.5 % amino-acid identity, respectively, with SARS-CoV-2 S; the more distantly related common cold a-CoVs, NL63 (AAS58177.1) and 229E (GenBank: AAK32191.1), just 18.3% and 20.2 %. Of the 157 S ectodomain-sorted mAbs, 47 (29.9%) bound to SARS-CoV S and 8 to other β-CoV S glycoproteins. These 8 cross-reactive antibodies have higher mutation levels than do RBD, NTD, and the other S2-binding mAbs from our cohort (Fig. 1C and D). Among the 59 S-binding mAbs cloned from the S⁺RBD⁻ sorted memory B cells, ELISA detected 23 (39%) mAbs that bound NTD (11 of which also bound NTD on yeast), 14 (23.7%) that bound S2, of which 7 (11.9%) cross-reacted with SARS-CoV S. One mAb bound RBD (Fig. 1B and C).

Global competition analysis defines seven epitopic regions on S

We used a competition ELISA to determine pairwise overlaps of the 105 antibodies in our panel for which we could detect signal at 1 μg/mL. By adding a biotinylated version of each mAb (1 μg/mL) together with 100-fold excess of each of the other mAbs individually into ELISA plates pre-coated with pre-fusion-stabilized SARS-CoV-2 S (19), we could detect competition of mAbs with up to 100-fold differences in affinity. We also included 15 published mAbs with known structures as references (fig. S3A).

We identified seven major clusters of competing mAbs—three RBD clusters, two NTD clusters, and two S2 clusters (fig. S3A). The three RBD clusters overlapped to varying extents, as expected for sites on a relatively small domain. Asymmetric competition (one mAb blocks binding of another, but the second does not block the first) occurred when one had much higher affinity than the other -- e.g., S309 (20), which binds more tightly than do most of the RBD-1 mAbs we isolated. The clusters define relatively broad epitopic regions, as the footprints of two antibodies within a cluster might not overlap with each other but both might overlap with the footprint of a third (e.g., REGN10933, REGN10987, both of which competed with CC12.1, although they have completely distinct footprints at either end of the RBD receptor binding motif, (RBM) (21). Some crosstalk between clusters is also evident (e.g. C93D9, which bound the RBD, blocked both RBD-2 and NTD-1 mAbs). The published 4–8, 4A8 and COVA1–22 mAbs (12, 13, 22), which have been shown to bind the NTD, compete with each other and map to NTD-1. NTD-2 mAbs cluster distinctly from NTD-1, indicating minimal spatial overlap of these two NTD regions. One NTD-1 mAb (C81H11) competed strongly with antibodies from S2–1, and a second could be assigned on the basis of competition either to NTD-1 or to S2–1, suggesting structural adjacency of at least some sites in these two clusters (fig. S3A). Several segments of S2 are in contact with either RBD or NTD, some with differential exposure depending on whether the RBD is “up” or “down”.

Thirty-six mAbs in the ELISA competition analysis cross-reacted with SARS-CoV. These mapped mostly to the RBD-1 (11 mAbs) and S2–1 (17 mAbs) clusters. Four mAbs that mapped to S2–1 (C15C3, C7A4, C7A9 and G32Q1) also bound the common cold b-CoVs, and two of these (C7A9, C15C3) also bound MERS S. Thus, S2–1 mAbs appear to recognize a region of S2 conserved among SARS-CoV-2, SARS-CoV, MERS, HKU1 and OC43, as shown previously for S2 antibodies (23, 24). Isolation of a single mAb (C12B3) that bound S2 but did not map to any of the seven major clusters suggests that the immune system may target additional regions of S2, but that those responses are subdominant.

Memory B cells dominant across individuals in natural infection

We probed the relative distribution of epitopes recognized by SARS-CoV-2 specific memory B cells in the population represented by our cohort by ELISA-based and cell surface-based assays. We clustered all the S⁺ mAbs (Fig. 2) and S⁺RBD⁻ mAbs from a separate sorting step (fig. S3B and C). In the former set, comprising 73 mAbs that bound strongly enough for the ELISA competition assay, the order of epitopic region frequencies was RBD-1 (27.4%), S2–1 (19.2%), NTD-1 (17.8%), RBD-2 (15.1%), NTD-2 (8.2%) (Fig. 2A). There were 36 more mAbs that had insufficient affinity for ELISA competition but that bound cell surface SARS-CoV-2 S (Fig. 2B and fig. S4A). To map ELISA-insufficient mAbs to the 7 clusters, we mixed the biotinylated antibodies with blocking antibodies selected from the ELISA-mapped competition assays, incubated the mixture with cells expressing SARS-CoV-2 S, and recorded the mean value florescent intensity (MFI) to calculate the blocking strength. We used twenty mAbs (2 RBD-3, 4 RBD-2, 4 RBD-1, 2 NTD-2, 4 NTD-1, 2 S2–1 and 2 S2–2) from the ELISA-mapped competition clusters as blocking antibodies, a non-COVID-19 related blocking antibody as negative control, and self-blocking as positive control (Fig. 2B and S3C). We used an additional 118 mAbs (distributed across the 7 clusters) to map the 13 mAbs that failed to compete with the initial 20 (fig. S4B).

Fig. 2. Competition epitope mapping.

(A) Cross competition matrix for 73 antibodies from the spike⁺ sort in Fig. 1 with affinity sufficient for detection by ELISA. Blocking antibodies (columns) added at 100 μg/ml; detection antibodies (rows), at 1μg/ml. Intensity of color shows strength of blocking, from 0 signal (complete blocking) to 70% full signal (top gradient at right of panel: orange). Hierarchical clustering of antibodies by cross competition into 7 groups (plus a singleton labeled S2–3), enclosed in square boxes, with designations shown and in colors from dark blue (NTD-1) to dark red (S2–3). Green arrows on the left designate antibodies newly reported here. The lower parts of the panel show: competition of blocking antibody with soluble, human ACE2 (second gradient at right: dark red); log(IC₅₀) in pseudovirus neutralization assay (third gradient at right: violet); area under the curve for ELISA binding (bottom gradient at right: brown); binding (ELISA) to recombinant domains and heterologous spike proteins. (B) Competition in cell-based assay for 36 antibodies with binding in ELISA format too weak for reliable blocking measurement (rows). Blocking antibodies (columns) selected from each of the 7 clusters in the ELISA assay (fig. S2). Strength of blocking shown as intensity of orange color, as in (A). (C) Distribution of antibodies from three individual subjects (expressed as percent of sequence pairs recovered from that subject) into the 7 principal clusters, plus a non-assigned (unknown) category (unk) and S2–3. Data are shown for only those subjects from whom we recovered at least 10 heavy- and light-chain sequence pairs. Heat map scale shown at right of panel. Top row shows total distribution, from panel (A) and (B).

Cell-based competition showed that mAbs with affinities too low to test by ELISA mapped primarily to S2 and to the NTD (Figs 2B and S3C). Results from including ELISA-mapped antibodies in the cell-based competition assay showed that the two assays were consistent (fig. S5A and B) and suggested that cell-surface binding simply extended the dynamic range of the ELISA competition assay to include less tightly binding antibodies, justifying use of the combined competition results in subsequent analyses. This combined approach showed that frequencies of cluster-targeting mAbs from the two individuals that contributed the most clones (C12 and G32) were largely similar to all others (Fig. 2C).

Structural features of competition groups

We included in the competition assays, antibodies for which published structures show their interaction with S. We also determined by cryo-EM structures of Fab fragments of four mAbs from the RBD-1 and NTD-1 clusters bound with S ectodomain, to fill gaps in the representation of antibodies from those clusters in published work. Two of those structures are at relatively high resolution (those of Fab C12C9, in NTD-1, and Fab G32R7, in RBD-1), a third (C81C10, at the periphery of NTD-1) at intermediate resolution, and a fourth (12C11, in NTD-1) at much lower resolution.

RBD-1.

The complex with Fab G32R7 (Fig. 3) has three RBDs in the “up” configuration, each bound with a Fab. The epitope is part of the RBD surface that faces outwards in the “down” configuration of the domain, but interference of the bound Fab with the NTD of the anticlockwise neighboring subunit (as viewed from “above” the spike in the orientation generally shown) would prevent binding to a down-oriented RBD. The connection of the RBD allows a range of orientations for the up configuration, and association with the G32R7 Fab does not fix the orientation of the RBD, blurring density in a 3-D reconstruction of the intact spike. Local refinement of an RBD-Fab subparticle then yielded a map that allowed us to build a good model of the interface (Fig. 3 and figs. S6, S7). RBD contacts are all with the heavy chain variable domain (V_H), principally CDRH2, framework residues in the C”, D and E strands, and CDRH3. The unusually long CDRH3 (24 residues) also interacts with three glycans -- one on RBD Asn343 and the others on NTD Asn122 and Asn165. Although V_H approaches the NTD closely enough to interact with the glycans, we could identify just one likely additional contact with an NTD side-chain (Phe157). The one published RBD-1 antibody structure is that of S309, a neutralizing antibody isolated from a convalescent SARS-CoV donor that also neutralizes SARS-CoV-2 (20). Its contacts with the RBD do not overlap those of G32R7, but the light chain of the latter would collide with it.

Fig. 3. Ab contact regions.

Surface regions of the SARS-CoV-2 spike protein trimer contacted by antibodies in four of the seven principal clusters, according to the color scheme shown (taken from the color scheme in Fig. 2), with a representative Fab for all except RBD-3. The C81C10 Fab defines an epitope just outside the margin of NTD-1, but it does not compete with any antibodies in RBD-2. The RBD-2 Fv shown is that of C121 (PDB ID: 7K8X: Barnes et al, 2020), which fits most closely, of the many published RBD-2 antibodies, into our low-resolution map for C12A2. Left: views normal to and along threefold axis of the closed, all-RBD-down conformation; right: similar views of the one-RBD-up conformation. C121 (RBD-2) can bind both RBD down and RBD up; G32R7 (RBD-1) binds only the “up” conformation of the RBD. The epitopes of the several published RBD-3 antibodies are partly occluded in both closed and open conformations of the RBD; none are shown here as cartoons. A cartoon of the polypeptide chain of a single subunit (dark red) is shown within the surface contour for a spike trimer (gray).

RBD-2 and RBD-3.

Most potently neutralizing antibodies cluster in RBD-2; many published structures show modes of antibody binding within this group (25, 26). Their epitopes include various parts of the ACE2 binding site (i.e. the RBM) at one apex of the domain. For the antibodies characterized here, low resolution structure of Fab C12A2 showed that its epitope was essentially identical to that of published antibody 2–4 (12). The same IGVH encodes the heavy chains of both antibodies, and the light-chain genes are closely related (overall amino-acid sequence identity). They contact the slight concavity in the center of the RBM, the site for most of the neutralizing antibodies represented by structures in the PDB. The probably immunosubdominant RBD3 class includes several antibodies for which published structures are available; we included CR3022, an antibody originally isolated from a SARS-CoV convalescent subject that cross-reacts with SARS-CoV-2 (27). Its epitope on the RBD is nearly opposite that of G32R7 (Fig. 3), in an epitopic region partly occluded in the down configuration of the RBD previously referred to as a “cryptic supersite” (26).

NTD-1.

NTD-1-cluster antibodies vary in neutralizing strength from strong (e.g., C12C9) to weak (C12C11). The latter, judging from the low-resolution map (fig. S8) appears to have a footprint that coincides with that of the published 4A8 antibody (PDB ID: 7C2L; (22)). Like the G32R7 complex, the C12C9 complex also required local subparticle refinement to yield a map interpretable at the level of side-chain contacts at the Fab-NTD interface. Its footprint overlaps that of 4A8, but it is displaced slightly toward the threefold axis of the S trimer. Both antibodies have principal contacts in two NTD surface loops, residues 140–160 and 245–260. The C81C10 mAb, which we have grouped in NTD-1 but which competes with only two of the most weakly binding members of that cluster, appears to represent a distinct and potentially subdominant subset. An 8 Å resolution structure (fig. S8) bound with spike trimer shows that its epitope is at the “bottom’ of the NTD, well displaced from the epitopes of C12C9, C12C11 and 4A8.

NTD-2.

We have so far no structures of NTD-2 antibodies bound with spike, but from non-competition with NTD-1, insensitivity to NTD loop deletions (see below), and exposure of NTD surfaces on the trimer, we suggest that the NTD-2 epitopes may be on one of the lateral faces of the NTD.

Representation of neutralizing antibodies in RBD and NTD clusters

Using two different pseudovirus assays, we determined neutralization by mAbs from each of the 7 clusters and found neutralizing antibodies in 5 of the 7 clusters (RBD-1, −2, −3, NTD-1, and NTD-2) (fig. S9). The most potent were in RBD-2, as expected from their co-clustering with known strong neutralizers such as REGN10987 and REGN10933, which are used as a mAb drug cocktail (28) and from many previous reports (12, 13, 29). Among Abs from that cluster, 52–58% neutralized with IC₅₀<0.1 μg/ml and 28.5–35% with IC₈₀<0.1 μg/ml (fig. S9D). The strongest of the RBD-1 antibodies had IC₅₀ in the range of 1 μg/ml; those in RBD-3 were in general much weaker. NTD-1 antibodies appeared more sensitive to the neutralization assay used, but a few, such as C12C9, approached or exceeded the strongest in RBD-1 (fig. S9A). NTD-2 antibodies were in general less potent than NTD-1 antibodies, and none of the S2 antibodies neutralized infection, with the possible exception of very weak neutralization by G32C5 (IC₅₀ of 22 ug/ml) (fig. S9C).

Molecular features of mAb recognition groups

Variable region exons of IgH and IgL genes are each assembled by V(D)J recombination from a diversity of gene segments. Preference of VH gene segment usage frequencies differed among the 7 mAb clusters (Fig. 4A). Enrichment for VH3–53, previously reported to be associated with SARS-CoV-2 S (21), was exclusively within the RBD-2 group. VH3 family antibodies are particularly abundant in all the clusters. VH3–30 and VH3-30-3, which have average frequency in the general human repertoire of 5.4% and 1.3% respectively, account between them for over 30% of the antibodies in RBD-1, and for 16 of the 19 antibodies in S2–2. The VH1 and VH4 families are co-dominant with VH3 in NTD-1 and NTD-2, respectively (Fig. 4A). VH1–69 encoded antibodies are enriched in S2–1, which contains most of the cross-reactive antibodies to other coronaviruses. VH1–69 encoded antibodies are frequently observed in antiviral responses to influenza virus, HCV, and HIV-1 (30), and previous work reported that SARS-CoV-2 spike-specific mAbs isolated from SARS-CoV infected patients also showed an enriched VH1–69 gene segment usage (31). VH1–69, which is well represented in heavy chains of “natural antibodies”, also associates strongly with polyreactivity. VH and VL somatic mutation levels were generally, but not significantly, greater in S2–1 (fig. S10A)

Fig. 4. Antibody sequence analyses.

(A) Heavy-chain variable-domain genes of the 167 mAbs characterized by binding SARS-CoV-2 spike in either ELISA or cell-surface expression format. The inner ring of each pie chart shows the VH family and the outer ring, the gene. PBMC repertoire is from 350 million reads of deep sequencing (37). S binders include 167 clones in Table S2. *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001; Bonferroni correction. Red asterisks: comparing to S binders; black asterisks: comparing to a non-selected B cell repertoire from PBMCs. (B) Maps of pairwise distances of CDRH3 (lower left triangle) and CDRL3 (upper right triangle) for the NTD-2 and S2–1 cluster antibodies from (A). Antibodies in both clusters arranged by VH usage. Clones converging on identical VH/VL alleles and closest distance of CDRL3 from the same cluster are shown. Pairwise distances analyzed by Mega X. Intensity of color shows the distance, from 0 (identical) to 1 (no identity). Sequence alignment for the antibodies from the indicated clusters with identical VH and VJ and similar CDR3s. Differences in CDR3s from the reference sequence (bold) are in red; dashes indicate missing amino acids; dots represent identical amino acids. (C) Summary of convergent sequences of anti-SARS-CoV-2 S and RBD antibodies from independent datasets. Ig sequences derived from binding to DIII of Zika virus E protein, and HA of influenza virus H1N1 were used as control datasets. Convergent sequences had identical VH and VL and >50% identity in CDRH3 and CDRL3. (D) Representative convergent clones from different individuals and independent datasets from Fig. 4C.

IgH and IgL variable regions harbor three complementary determining regions (CDRs), which serve as principal contact sites for antigen. CDRs 1 and 2 for H and L chain are encoded within the VH and VL gene segments, respectively. Highly diverse, non-templated sequences produced by VDJ_H junctions encode CDRH3 regions, which have dominant roles in most Ab-antigen interactions. CDRL3, which can contribute antigen contact surfaces, is also diverse due to VJ_L junctional heterogeneity, but has less non-templated sequence additions. Intracluster mAb CDR3 sequence comparisons showed little sequence similarity (fig. S10B). NTD-2 contained a subcluster of identical CDRL3 sequences that were associated with the same VH and VL segments from two different individuals (C81, C12) (Fig. 4B). S2–1 had a small subcluster of CDRH3 and CDRL3 sequence similarities from 5 different study participants (C83, C102, C163, C12, C53) (Fig. 4B). These data indicate substantial intracluster CDR3 diversity with rare instances of CDR3 sequence similarity between different individuals.

We also asked whether we could find sequences very similar (i.e. convergent) to any in our dataset, from other COVID-19 data sets for which paired IgH and IgL sequence data are available. Based on prior convergent sequence analysis (32), our criteria for convergence were (i) same VH and VL, and (ii) no less than 50% CDRH3 and CDRL3 identity, These criteria identified rare sequences very similar to representatives from RBD-1, RBD-2, and NTD-1 in independent datasets for SARS-CoV-2 (1, 12, 33), but not for antibodies against Zika (34) or influenza (35) viruses (Fig. 4C, D). We also found convergent pairs within our own dataset representing both S2–1 and S2–2 (Fig. 4D).

Finally, we note that antibody C93D9 represents a striking example of structural convergence. All but two of the 20 antibodies from the literature shown in fig. S11C have the same VH and a non-random selection of VL but divergent CDRH3 sequences and lengths. Nonetheless, all 20, as well as C93D9, bind the RBM in almost identical poses—consistent with germline encoded CDRs as the principal binding contacts (21).

Reactivity of the anti-S memory B cell repertoire for emerging SARS-CoV-2 variants

Emergence of SARS CoV-2 variants that enhance transmissibility, such as the variant B.1.1.7 (i.e. the UK variant) (36), and in some cases reduce the neutralization titers of convalescent sera, such as the variant B.1.351 (i.e. South Africa (SA) variant) (37), indicates more rapid evolution of the virus than expected from the error-correcting properties of coronavirus RNA-dependent RNA polymerases. In the case of the SA variant in particular, the clusters of three substitutions and one deletion in the NTD and three substitutions in the RBD concentrate at contacts of the most potent of the many well-characterized neutralizing antibodies. Moreover, recurrent deletions in loops of the NTD appear to accelerate SARS CoV-2 antigenic evolution (38).

We examined the effects of naturally occurring mutant spike protein on binding of mAbs in each competition group. The UK variant had lower affinity for various mAbs in the RBD-1 and NTD-1 cluster. None of the RBD-1 mAbs lost binding completely, and testing a variant with just the deletion at position 144 in the NTD showed that this single mutation caused loss of binding by nearly two-thirds of the mAbs in the NTD-1 cluster. Mutations in the variant B.1.351 had more pronounced effects, particularly on mAbs in the RBD-2 and NTD-1 clusters, as expected from the positions of the sequence changes. In addition to the N501Y substitution also present in the variant B.1.1.7, an E484K mutation lies at the center of the epitope for many of the most potent RBD-2 neutralizing antibodies. About one-third of the RBD-2 mAbs retained modest to high affinity, but the variant spike failed to bind any of the NTD-1 cluster, with the marginal exception of 4A8.

Differential effects on antibodies with overlapping but still distinct epitopes illustrate the potential importance of a redundant, polyclonal response. Although C12C9, C12C11, and 4A8 all contact the 140–160 loop (Fig. 3 and S7) and all are sensitive to the multi-position, D141–144 and D243–244, recurrent deletions, only the latter two are sensitive to the recurrent, single-position deletions at 144 or 146 (Fig. 5 and fig. S12). Moreover, although they are in the same convergent structural class whose members bind the RBM in nearly identical poses (fig. S11), CC12.1 fails to recognize B.1.351 (~0%) while C93D9 retains some marginal affinity (~27%) (Fig. 5). Thus, apparently redundant memory B cell clones can have non-redundant functional roles.

Fig. 5. Recognition of naturally occurring deletions and mutations in the spike.

Heat map showing binding of 119 mAbs to Nextstrain cluster 20A.EU1 (A222V), Danish mink variant (Δ69–70 and Y453F), UK B.1.1.7 (Δ69–70, Δ144, N501Y, A570D, P681H, T716I, S982A, D1118H) and SA B.1.351 (L18F, D80A, D215G, Δ242–244, K417N, E484K, N501Y, A701V) (top) and NTD deletion variants (bottom). The Wuhan-Hu-1 S sequence and all variants include the D614G mutation. Binding for each mAb was first normalized (“normalized IgG MFI”) by dividing the MFI for that mAb by the MFI for C81E2 (S2–2 cluster). The normalized MFI of for binding the Wuhan-Hu-1 spike was used as a reference (normalized Wuhan IgG MFI). The relative binding intensities of the tested mAbs for each variant, calculated as the ratio of the normalized variant IgG MFI and the normalized Wuhan IgG MFI, are shown in shades of blue.

DISCUSSION

Our results illustrate the landscape of memory B cell coverage of the SARS-CoV-2 S glycoprotein in convalescent donors. Unlike the terminally differentiated plasma cells that determine the profile of serum antibodies, memory B cells will clonally expand upon re-exposure to antigen, some differentiating into fresh antibody secreting cells and others re-entering germinal centers and undergoing further SHM-mediated diversification and affinity maturation. These outcomes offer a layer of flexibility for adaptation to drifted or related viral strains, if available secreted antibodies fail to prevent initial infection. Loss of protection against overt or severe disease is not an inevitable consequence of a waning serum antibody titer. This atlas of B cell memory therefore maps systematically a crucial component of the long-term immune response to SARS-CoV-2 infection.

The donors for this study experienced COVID-19 symptom onset between March 3 and April 1, 2020, and blood draws analyzed here were between April 2 and May 13, 2020, early in the pandemic. Immune responses in these SARS-CoV-2 naive donors were to early and relatively homogeneous variants circulating well before emergence of the UK and SA strains first reported in Dec. 2020 and probably before the spread, in New England, of the D614G variant that in any case did not substantially alter antigenicity (39, 40). This set of BCR sequences and corresponding mAbs thus represents responses to a relatively homogeneous infectious virus and provides a valuable tool for examining the degree to which these antibodies retain recognition of emerging variants and for studying the extent to which loss of neutralizing titer correlates with loss of longer-term protection.

The competition clusters we have identified are roughly analogous to genetic complementation groups. Competition can result from overlapping binding footprints or non-overlapping but neighboring footprints that lead to mutual exclusion of IgGs bound at the two adjacent epitopes. Competition can also result from stabilization by one antibody of a conformation (e.g., the up-down conformational isomerism of the RBD) that excludes or lowers affinity of the another. Any of these mechanisms may contribute to the clusters we have mapped, but the outcome in all cases is an apparent redundancy of binding capacity in a broadly polyclonal response that may nonetheless impart recognition breadth toward an evolving pathogen within a single individual.

Complementary recognition of non-overlapping viral targets by non-competing antibodies in the repertoire can reduce the likelihood of viral escape (41). Our data suggest an additional mechanism for preventing viral escape: competing antibodies may help retain recognition of a rapidly evolving antigen by their differential sensitivity to specific mutations. The potential dynamic reach of otherwise redundant mAb recognition, illustrated by selective retention of affinity for the UK variant by some antibodies within a cluster but not by others, may give selective advantage to immune mechanisms that yield multiple competing antibodies to critical epitopes, as those that retain adequate affinity can then re-activate, expand, and potentially undergo further affinity maturation. The emergence of strains that may have gained selective advantage by escape from neutralization emphasizes the importance of determining whether the level of retained affinity for the S protein by some antibodies in the immunodominant clusters influences protection from clinical disease.