R1-32-like public antibodies, characterized by shared IGHV1-69/IGLV1-40 usage, are elicited in more than 50% of individuals with COVID-19 and have been implicated in driving recurrent mutations at L452_SARS2 and F490_SARS2 within their convergent epitope in the SARS-CoV-2 spike receptor-binding domain. These mutations effectively mediate escape from non-affinity-matured R1-32-like antibodies with germline-like sequences. Here, we characterize four affinity-matured human R1-32-like antibodies, C092, C807, BD56-104, and BD56-597, that tolerate L452_SARS2 and F490_SARS2 mutations. We show that this tolerance arises from residues introduced by somatic hypermutation at convergent positions across multiple CDR loops and surrounding regions, thereby creating additional contacts that reinforce epitope binding. An unusual N354_SARS2 glycosylation site, which emerged in BA.2.86 and became fixed in its descendants, is linked to escape from affinity-matured R1-32-like antibodies, implying ongoing selection by this public antibody class. Using an AI model trained on extensive neutralization data, we further identified ZL525, an ultrapotent human R1-32-like antibody with pan-SARS-CoV-2 variant activity, including against the highly evasive KP.3 variant carrying the N354_SARS2 glycosylation, and broad sarbecovirus cross-reactivity extending to SARS-CoV-1. Together, these findings show how affinity maturation enables public antibodies to adapt to viral antigenic drift, reveal their role in shaping SARS-CoV-2 antigenic evolution, and demonstrate the potential of AI-empowered strategies for discovering broadly neutralizing antibodies.

Identification of R1-32-like public antibodies that tolerate spike mutations

We have previously defined IGHV1-69/IGLV1-40 encoded R1-32-like public antibodies as those that either exhibit a minimum of 80% amino acid similarity to R1-32 or FC08 in the HCDR3 region or contain a GYSGYG/D motif in HCDR3 [16,27]. In our previous study, we identified and characterized 10 R1-32-like antibodies isolated from ancestral (or wildtype, WT) SARS-CoV-2 convalescents or vaccinees [27,31,32,34,35]. Notably, among these, C092 and C807 showed no obvious reduction in binding to SARS-CoV-2 RBD variants with L452 or F490 mutations, including Delta (L452R_SARS2), Kappa (L452R_SARS2), and Lambda (L452Q_SARS2 + F490S_SARS2) RBDs [27]. We further identified BD56–104 and BD56-597 as mutation-resistant R1-32-like public antibodies that effectively neutralize Omicron subvariants BA.2.12.1 (L452Q_SARS2), BA.2.13 (L452M_SARS2), and BA.4/5 (L452R_SARS2), through the analysis of published sequence and neutralization data [28] (S1A Fig). Notably, both BD56-104 and BD56-597 were isolated from convalescents with Omicron BA.1 breakthrough infections [28]. In contrast to R1-32, which has minimal somatic hypermutation (SHM), the 4 mutation-resistant antibodies, C092, C807, BD56-104, and BD56-597, carry 6–15 SHMs in their heavy chains and 3–4 SHMs in their light chains, indicative of substantial affinity maturation (Fig 1A and 1B).

To further characterize these R1-32-like public antibodies, we performed antigen binding assays for C092, C807, BD56-104, and BD56-597 against a more comprehensive and updated panel of SARS-CoV-2 RBDs, which we classified into 4 categories: (1) RBDs with single amino acid substitutions at key residues (K417_SARS2, E484_SARS2, T478_SARS2, L452_SARS2, F490_SARS2); (2) RBDs from early VOC/VOI variants (Alpha to Lambda); 3) RBDs from early Omicron variants (BA.1 to BA.4/5); and (4) RBDs from later-emerging Omicron subvariants, including BQ, XBB, and EG variants. Within each category, mutants containing L452_SARS2 and F490_SARS2 mutations were further grouped together (Figs 1C and S1B–S1E).

The results showed that C092, C807, BD56-104, and BD56-597 retained high-affinity binding (K_D < 1.2 nM) to SARS-CoV-2 RBDs carrying L452_SARS2 and F490_SARS2 single mutations. In contrast, the germline-like R1-32 exhibited substantially reduced binding to these mutant RBDs (K_D = 11.8–54.2 nM) (Figs 1C and S1B–S1E). Substitutions at K417_SARS2, E484_SARS2, and T478_SARS2, commonly associated with the escape of epitope class 1 and class 2 antibodies [9,20,21], had minimal impact on the binding of C092, C807, BD56-104, and BD56-597 (Figs 1C and S1B–S1E), consistent with these residues being located at the periphery or outside of the epitope recognized by the R1-32 class of public antibodies.

C092, C807, BD56-104, and BD56-597 maintained very high affinities (K_D < 0.3 nM) towards RBDs of variants carrying no mutations at 452_SARS2 or 490_SARS2, including Alpha (B.1.1.7), Beta (B.1.351), Gamma (P.1), and multiple Omicron sublineages (BA.1, BA.1.1, BA.2, BA.3, BA.2.75, and BA.2.75.2). Consistent with their mutation tolerance, C092, C807, BD56-104, and BD56-597 also maintained high affinities (K_D < 2.1 nM) towards RBDs harboring L452_SARS2 substitutions, including Kappa (B.1.617.1), Delta (B.1.617.2), Omicron BA.2.11, BA.2.12.1, BA.2.13, BA.4/5, BF.7, and BQ.1.1. For comparison, germline-like R1-32 exhibited substantially reduced binding to these mutants (K_D > 55 nM).

Notably, C092, C807, BD56-104, and BD56-597 maintained high affinities (K_D < 4.5 nM) towards RBDs of the later-emerged XBB lineages (XBB, XBB.1.5), and EG.5 lineages (EG.5, EG.5.1), despite the presence of F490S_SARS2 mutations in these variants. All four antibodies exhibited the weakest binding to the Lambda (C.37) RBD carrying both the L452_SARS2 and F490_SARS2 mutations, although their K_D values are still below 6.8 nM (Figs 1C and S1B–S1E). For comparison, the germline-like R1-32 binds Lambda RBD with a K_D of ~ 1,000 nM, representing a reduction of ~1,000-fold.

These RBD binding results confirm that BD56-104 and BD56-597, like C092 and C807, are tolerant to diverse variants carrying L452_SARS2 and F490_SARS2. Among them, we found that C092 and BD56-597 consistently exhibited more robust binding across RBD variants compared to C807 and BD56-104. C807 showed moderately but noticeably reduced binding to RBDs harboring L452_SARS2 mutations, such as those in Kappa, Delta, Lambda, Omicron BA.4/5, BF.7, and BQ.1.1, characterized by accelerated dissociation kinetics (S1C Fig). BD56-104 also showed moderate but noticeable sensitivity to F490_SARS2 substitutions, displaying accelerated dissociation from RBDs of Lambda, XBB lineages (XBB, XBB.1.5), and EG.5 lineages (EG.5, EG.5.1) that carry this mutation (S1D Fig).

Mutation-tolerant antibodies retain potent neutralization against diverse SARS-CoV-2 variants

Neutralization activities of C092, C807, BD56-104, and BD56-597 were evaluated using VSV-based pseudovirus neutralization assays [36] against eight variants of SARS-CoV-2: the ancestral SARS-CoV-2 strain (WT), Delta (B.1.617.2), Lambda (C.37), Omicron BA.1 (B.1.1.529), BA.2, BA.4/5, BA. 2.75.2, BQ.1.1, and EG.5. The results showed that all four antibodies neutralized all tested variants, with IC₅₀ values ranging from 0.022 to 7.557 µg/mL (Fig 1D and 1E). However, compared with the WT, obvious reductions in neutralization activity were observed for Delta (L452R_SARS2), Lambda (L452Q_SARS2 + F490S_SARS2), BA.4/5 (L452R_SARS2), BQ.1.1 (L452R_SARS2), and EG.5 (F490S_SARS2) variants containing L452_SARS2, F490_SARS2, or both mutations. By contrast, germline-like R1-32 lost neutralization (IC₅₀ > 10 µg/mL) against Delta, Lambda, BA.4/5, BQ.1.1, and EG.5 pseudoviruses (Fig 1E).

Consistent with pseudovirus neutralization assays, C092, C807, BD56-104, and BD56-597 demonstrated potent neutralization (most IC₅₀ < 10 µg/mL) against authentic SARS-CoV-2 variants, including Delta (L452R_SARS2), BA.5 (L452R_SARS2), BQ.1(L452R_SARS2), XBB.1(F490S_SARS2), and EG.5 (F490S_SARS2), while R1-32 showed no detectable neutralization (IC₅₀ > 100 µg/mL) against these variants (Fig 1F and 1G). However, all five antibodies lost neutralizing activities against the recent circulating variant KP.3. Complementing the neutralization data, C092, C807, BD56-104, and BD56-597 exhibited strong, non-dissociating binding to the S-trimers of variants neutralized in the neutralization assays (S2 Fig). Binding affinity assays revealed moderately weakened binding with Lambda (L452Q_SARS2 + F490S_SARS2), BA.4/5 (L452R_SARS2), and BQ.1.1 (L452R_SARS2) S-trimers relative to WT, consistent with the observed reduction in neutralization potency against these variants. Altogether, our neutralization and binding data demonstrate that C092, C807, BD56-104, and BD56-597 exhibit markedly enhanced breadth and mutation tolerance compared to the germline-like R1-32, maintaining neutralization activity against SARS-CoV-2 variants carrying L452_SARS2, F490_SARS2, or even both escape mutations, with minimal to modest reductions in potency.

Binding of R1-32-like antibodies disassembles S-trimer without blocking ACE2 binding

To elucidate how C092, C807, BD56-104, and BD56-597 tolerate mutations to cross-neutralize SARS-CoV-2 variants, we prepared cryo-electron microscopy (cryo-EM) samples using stoichiometric mixtures of Fab, ACE2, and BA.4/5 S-trimer (Fab:ACE2:S-protomer = 1:1:1). Cryo-EM structures of C092, C807, BD56-104, and BD56-597 Fabs bound to BA.4/5 S-trimer in complex with ACE2 were determined (Figs 2A and S5–S9).

For the C092 Fab:ACE2:S-trimer sample, each S-trimer was found to adopt either 2 or 3 RBD “up” conformations, allowing simultaneous binding of either 2 or 3 Fab and ACE2 molecules. All other structures showed S-trimers with all three RBDs in the “up” position, with each RBD concurrently engaging both ACE2 and Fab (S5–S8 Figs). Consistent with previous conclusions [27], these S-trimer structures confirm that R1-32-like public antibodies do not interfere with ACE2 binding, in agreement with results from the antibody-ACE2 competition assays (S3 Fig). In addition to ligand bound S-trimer structures, we also observed isolated S1 domains bound to both Fab and ACE2, suggesting that a portion of S-trimers disassembles upon simultaneous antibody and ACE2 binding. (S5–S8 Figs). Based on these disassembled particles, structures of C092, C807, BD56-104, and BD56-597 bound to ACE2 bound S1 domain were determined to high-resolutions (2.58–2.83 Å) (Figs 2B and S5–S9). We previously concluded that R1-32 neutralizes SARS-CoV-2 by inducing disassembly of S-trimers, thereby rendering them incapable of mediating membrane fusion [27]. Based on the imaging data, we infer that C092, C807, BD56-104, and BD56-597 similarly retain the ability to induce disassembly of S-trimers from variant viruses.

R1-32-like antibodies impair spike fusogenic conformational change

To further confirm the neutralization mechanism, we conducted ligand-induced conformational change assays. The results showed that, like the WT, BA.4/5 S-trimers can be triggered by ACE2-Fc or YB9-258 [37], an epitope class 1 antibody that can mimic ACE2 function, to undergo the fusogenic conformational change, generating the proteinase K-resistant S2 fusion core (S4A and S4C Fig). ACE2-Fc retains the ability to trigger the EG.5.1 S-trimer, whereas YB9-258 loses this ability, likely due to epitope mutations in EG.5.1 that disrupt YB9-258 binding (S4E Fig). In line with our previous findings, S-trimers (WT, BA.4/5, and EG.5.1) pre-incubated with R1-32 and the four antibodies (C092, C807, BD56-104, and BD56-597) were unable to undergo structural transition into a post-fusion conformation (S4A, S4C, and S4E Fig).

Further, assay results also showed that WT, BA.4/5, and EG.5 S-trimers pre-incubated with the four R1-32-like antibodies (C092, C807, BD56-104, or BD56-597) were unable to undergo the fusogenic conformational change upon subsequent addition of ACE2 (S4B, S4D, and S4F Fig). In contrast, although R1-32 efficiently impairs the fusogenic conformational change of the WT S-trimer, it only partially inhibits this transition in the BA.4/5 and EG.5 S-trimers (S4B, S4D, and S4F Fig), likely due to the presence of L452_SARS2 or F490_SARS2 mutations in these variants.

Together with our cryo-EM data showing disassembled S-trimers, the results from ligand-induced conformational change assays confirm that the four identified R1-32-like antibodies retain the ability to inhibit ACE2-induced fusogenic conformational changes of S-trimers from variant viruses, likely by destabilizing or disassembling S-trimers, as previously proposed for R1-32 against the WT virus [27]. The results are further consistent with neutralization assays demonstrating superior resilience of C092, C807, BD56-104, and BD56-597 against L452_SARS2- and F490_SARS2-bearing variants compared to R1-32. Furthermore, although R1-32 retains residual binding and partially inhibits ligand-induced spike conformational changes, its loss of detectable neutralizing activity against these variants suggests that the reduced binding and incomplete inhibition are insufficient to induce the extent of spike inactivation or disruption required to effectively block viral entry.

Broadly neutralizing public antibodies maintain binding to the mutated convergent epitope

The newly determined C092, C807, BD56-104, and BD56-597 complex structures were aligned to our previously reported R1-32 complex structure (PDB: 7YDI), using the RBD as the reference (Fig 2C). Based on this alignment, the root mean square deviations (RMSDs) of the Fab variable regions were calculated as follows: C092 (1.5 Å), C807 (1.8 Å), BD56-104 (1.7 Å), and BD56-597 (1.5 Å). The low RMSD values indicate that these antibodies adopt binding modes and recognize epitopes nearly identical to those of R1-32 (Fig 2C, left panel).

Antibody interface analysis reveals that all four R1-32-like antibodies engage the RBD through both their heavy and light chains, with a predominant contribution from the heavy chains: C092 binds an epitope area of 1,350 Ų (HC: 903 Ų, LC: 447 Ų); C807 binds an epitope area of 1,193 Ų (HC: 820 Ų, LC: 373 Ų); BD56-104 binds an epitope area of 1,148 Ų (HC: 765 Ų, LC: 383 Ų); and finally BD56-597 binds an epitope area of 1,290 Ų (HC: 937 Ų, LC: 353 Ų). For comparison, the germline-like R1-32 binds an epitope area of 1,214 Ų (HC: 813 Ų, LC: 401 Ų).

All four R1-32-like antibodies utilize their characteristic IGHV1-69 HCDR2, featuring hydrophobic residues at positions 52_H, 54_H, and 55_H (subscripts denote heavy (H) or light (L) chain), to engage the hydrophobic RBD residues F490_SARS2 and L492_SARS2 (Fig 3A–3D, left panels). Notably, within the epitopes of R1-32-like antibodies in the BA.4/5 RBD, the epitope substitution L452R_SARS2 is present, with the aliphatic portion of R452_SARS2 side chain contacted by the hydrophobic HCDR2 residue 55_H. HCDR3 residues 102_H, 103_H, 104_H, and 107_H form polar interactions with RBD residues E465_SARS2, R466_SARS2, and I468_SARS2 (Fig 3A–3D, right panels). Additionally, HCDR1 (Fig 3A–3D, left panels) and LCDR1 (Fig 3A–3D, middle panels) contribute to RBD binding in all four antibodies. Notably, in C092, unlike the other three antibodies, both the light chain framework region 3 (LFR3) and LCDR3 participate in RBD binding (Fig 3A, middle panel). Also notably, in BD56-104 and BD56-597, the heavy chain backbone region 3 (HFR3) participates in RBD binding (Fig 3C and 3D, left panels).

We previously found that somatic hypermutation (SHM)-introduced residues play a key role in enhancing antigen binding and conferring mutation tolerance [22,37]. Compared with the germline antibody R1-32 (S9E Fig), all four antibodies harbor SHM-introduced residues in HCDR1 and LCDR1 (Fig 3A–3D). In C092, SHM-introduced residues are also present in LFR3 and LCDR3 (Fig 3A). In BD56-104 and BD56-597, SHM-introduced residues are also found in heavy chain framework region 3 (HFR3) (Fig 3C and 3D). Many of these residues form additional contacts with the RBD, likely contributing to enhanced antigen binding (see below).

Convergent somatic hypermutations confer enhanced binding and mutation tolerance

C092, C807, BD56-104, and BD56-597 likely resist epitope mutations by having undergone affinity maturation through somatic hypermutation. To investigate the affinity maturation pathway underlying their mutation tolerance, we curated 98 R1-32-like antibodies from sequences reported by Cao and colleagues [28] and analyzed the correlation between SHM patterns and neutralization activity (IC₅₀ values). We examined whether specific mutations are enriched in the R1-32-like antibodies isolated from individuals with different exposure histories (S10 Fig), as well as antibodies escaped by or tolerant to BQ.1.1 (L452R_SARS2) and XBB (F490S_SARS2) (Fig 4A). We found several SHMs are highly enriched in R1-32-like antibodies that tolerant to BQ.1.1 and XBB, such as A33P/S_H, S35T_H, and N59D_H on the heavy chain and S26N_L, G52A/V_L, and S54N/T_L on the light chain (Fig 4A). In addition, these SHMs were enriched in R1-32-like antibodies isolated from patients with breakthrough infections (S10 Fig). Several of these SHM-enriched sites are also directly involved in epitope binding in C092, C807, BD56-104, and BD56-597 (Fig 3).

In the C092:RBD complex, SHM-introduced HCDR1 residues N31_H and H32_H form a hydrogen bond and a salt bridge with RBD residues T470_SARS2 and E471_SARS2, respectively (Fig 3A, left panel). SHM introduced HCDR1 residue F33_H contacts RBD residues I468_SARS2 and Y351_SARS2, as well as HCDR2 residue I52_H, thereby reinforcing the hydrophobic interactions involving I52_H, I54_H, and L55_H of HCDR2 and RBD residues F490_SARS2 and L492_SARS2 (Fig 3A, left and right panels). Additionally, SHM-introduced residues in the light chain further enhance antigen binding: N26_L of LCDR1 and D70_L of LFR3 form hydrogen bonds and salt bridges with R357_SARS2, while D99_L of LCDR3 forms salt bridges with R346_SARS2 (Fig 3A, middle panel). To confirm the function of the SHM-introduced residues, we reverted them to their germline sequences and evaluated antigen binding by the revertants. Reverting C092 HCDR1 to germline reduced binding to L452R, Kappa (L452R_SARS2), Delta (L452R_SARS2), Lambda (L452Q_SARS2 + F490S_SARS2), and BA.4/5 (L452R_SARS2) RBDs by ~ 10-fold (Figs 4B and S11A). Likewise, reverting the light chain SHM-introduced residues (N26S_L, D70G_L, and D99G_L) to germline reduced binding to the same set of RBDs by ~10–100-fold, demonstrating that light chain residues make important contributions to the breadth of C092 (Figs 4B and S11A).

In the C807:RBD complex, SHM-introduced HCDR1 residue I33_H contacts RBD residues I468_SARS2 and Y351_SARS2, as well as I52_H of HCDR2, via hydrophobic interactions (Fig 3B, left and right panels). Therefore, I33_H of C807, like F33_H in C092, enhances the hydrophobic interactions mediated by HCDR2. Reverting HCDR1 of C807 into the germline sequence reduced binding of RBDs with 452_SARS2 and 490_SARS2 mutations (Figs 4C and S11B).

In the BD56-104:RBD complex, SHM-introduced HFR3 residue D59_H forms salt bridges with the mutated RBD residue R452_SARS2, while SHM-introduced HFR3 residue E62_H forms salt bridges with R346_SARS2 (Fig 3C, left panel). Germline reversion of HFR3 residues (D59N_H and E62Q_H) reduced binding to RBDs carrying 452_SARS2 and 490_SARS2 mutations. In particular, the combined BD56-104 HFR3 revertant (D59N_H/E62Q_H) showed an ~100-fold reduction in binding to the BA.4/5 RBD harboring the L452R_SARS2 mutation (Figs 4D and S11C).

In the BD56-597:RBD complex, similar to C092 and C807, SHM-introduced HCDR1 residue P33_H provides extra hydrophobic interactions to enhance antigen binding (Fig 3D, left and right panels). Further, similar to BD56-104, SHM-introduced HFR3 residue D59_H forms salt bridges with the mutated RBD residue R452_SARS2 (Fig 3D, left panel). Notably, SHM-introduced LCDR1 residue D26_L interacts with R357_SARS2 via salt bridges (Fig 3D, middle panel). Germline reversion of HCDR1, HFR3 (D59N_H), and LCDR1 (D26S_L) reduced binding to the RBDs carrying 452_SARS2 and 490_SARS2 mutations. Notably, towards the Lambda RBD, carrying both L452Q_SARS2 + F490S_SARS2 mutations, a 10-fold reduction in binding was observed (Figs 4E and S11D).

HCDR3 combines with somatic mutations to confer mutation tolerance

In addition to the extra interactions mediated by SHM-introduced residues, HCDR3 also contributes to enhanced antigen binding in C092, C807, and BD56-597. Antigen-binding assays showed that replacing the HCDR3s of C092, C807, or BD56-597 with that of R1-32 (97_H-ARENGYSGYGAAANFDL-113_H, with A109_H underlined in bold) markedly reduced binding to the tested RBDs, with C807 being the most affected (Figs 4B–4E and S11). In contrast, HCDR3 substitution in BD56-104 had a comparatively minor impact on binding. Structural data revealed that the HCDR3s of C092, C807, and BD56-597 all contain a shared Y109_H, which forms hydrogen bonds and hydrophobic interactions with RBD residues Y351_SARS2 and I468_SARS2 (Fig 3A, 3B, and 3D, right panels). Substituting Y109_H with the alanine present in R1-32 also notably impaired antigen binding in these three antibodies. Moreover, combining HCDR1 germline reversion with the HCDR3 Y109A_H substitution in C092, C807, and BD56-597 produced an additive reduction in binding (Figs 4B–4E and S11), indicating that distinct antigen-contact regions jointly contribute to high-affinity binding in a combinatorial manner.

Artificial affinity maturation of R1-32 validates SHM-introduced residues in enhancing mutation tolerance

Our structural and antigen-binding data confirmed seven residues enriched with mutations and associated with enhanced antigen binding, A33F_H, N59D_H, Q62E_H, and A109Y_H in the heavy chain, and S26D_L, G70D_L, and G99D_L in the light chain. We introduced these residues into R1-32 to generate an antibody called termed “R1-32-AAM” (R1-32-artificial-affinity-matured). Notably, these SHM-introduced residues not only mediate additional RBD interactions in C092, C807, BD56-104, and BD56-597, but most of them are also enriched in R1-32-like antibodies that retain tolerance to BQ.1.1 and XBB (Fig 4A).

Binding assays showed that R1-32-AAM exhibits markedly enhanced antigen-binding activity compared with R1-32. Relative to R1-32, R1-32-AAM bound to RBDs bearing L452 mutation, including L452R_SARS2, Kappa (L452R_SARS2), Delta (L452R_SARS2), Lambda (L452Q_SARS2 + F490S_SARS2), BA.4/5 (L452R_SARS2), BF.7 (L452R_SARS2), and BQ.1.1 (L452R_SARS2) with faster association, slower dissociation, and resulting in ~100-fold enhanced affinity (Fig 4F). It also displayed enhanced binding to RBDs carrying F490S_SARS2 or F490W_SARS2, and XBB variant (F490S_SARS2), with ~10–100-fold higher affinity than R1-32. Consistent with these binding data, R1-32-AAM neutralized authentic Delta (L452R_SARS2), BA.5 (L452R_SARS2), XBB.1 (F490S_SARS2), and EG.5 (F490S_SARS2) viruses (Figs 1F, 1G, 4G, and 4H), confirming that SHM-introduced residues enriched among R1-32-like antibodies confer tolerance to L452_SARS2 and F490_SARS2 substitutions. Together with structural differences to R1-32 (S9E Fig), these findings indicate that the breadth of C092, C807, BD56-104, and BD56-597 is largely attributable to affinity-maturation-introduced residues, particularly within and around the CDRs, which mediate additional antigen interactions.

Affinity maturation broadens the activity of R1-32-like antibodies toward SARSr-CoV RBDs

We further explored potential cross-reactivities of C092, C807, BD56-104, and BD56-597 toward SARSr-CoV RBDs using antigen-binding assays (Figs 5A and S12), testing representative type-1 to type-3 RBDs as previously defined by the extent of deletions within the receptor-binding motif (RBM) [38]. All four antibodies bound the “type-1” RBDs of Pangolin-GD-2019, Bat-RaTG13, Pangolin-GX-2017, Laos-20-52, Laos-20-236, and Bat-RsSHC014 with affinities ranging from 0.016 to 48 nM. Notably, C092, C807, and BD56-597 cross-reacted to WIV1 RBD, whereas BD56-104 did not. Only C807 and BD56-597 bound the SARS-CoV-1 RBD, albeit with much compromised affinities (K_D < 341 nM) (Fig 5A). Remarkably, all four antibodies also bound the “type-2” BtKY72 RBD, despite its RBM sequence differing notably from the “type-1” SARS-CoV-2 RBD. Moreover, they retained binding to the “type-3” RBDs of GX2013, HeB2013, and RmYN02, which contain substantial loop deletions within the RBM region (Fig 5A). Among them, C092 and C807 exhibited tighter binding affinities than BD56-104 and BD56-597, demonstrating broader cross-reactivity towards SARSr-CoV RBDs. In contrast, R1-32 showed much weaker binding to the RBDs of Bat-RaTG13, Pangolin-GX-2017, and Bat-RsSHC014 (4.76−297 nM) than C092, C807, BD56-104, BD56-597, and failed to bind the RBDs of Bat-WIV1, SARS-CoV-1, BtKY72, GX2013, HeB2013, and RmYN02 (Fig 5A), indicating that affinity maturation confers broader SARSr-CoVs binding to R1-32-like antibodies.

We have established that C092, C807, BD56-104, and BD56-597 can well tolerate and maintain almost uncompromised neutralization towards variants carrying 1 mutation at RBD residues 452_SARS2 or 346_SARS2, including Delta (L452R_SARS2), BA.5 (L452R_SARS2), and BA.2.75.2 (R346T_SARS2). Neutralization was somewhat reduced towards variants with 2 mutations within the RBD epitope, including Lambda (L452Q_SARS2 + F490S_SARS2), BQ.1.1 (R346T_SARS2 + L452R_SARS2), XBB.1 (R346T_SARS2 + F490S_SARS2), and EG.5 (R346T_SARS2 + F490S_SARS2) (Fig 1C–1G). We noted that all four R1-32-like antibodies exhibited markedly impaired binding to the SARS-CoV-1 RBD. Within the R1-32-like antibody epitope, SARS-CoV-1 and SARS-CoV-2 differ at L452_SARS2, F490_SARS2, R346_SARS2, R357_SARS2, and T470_SARS2. To assess the mutation tolerance of the four R1-32-like antibodies (S13 Fig), we introduced R346T_SARS2, R357T_SARS2, and T470N_SARS2 mutations into the Lambda (L452Q_SARS2 + F490S_SARS2) RBD, towards which C092, C807, BD56-104, and BD56-597 already exhibited 9.5-178-fold reduction in binding affinity (Fig 1C). Of note, although R346T_SARS2 lies at the periphery of the R1-32-like antibody epitope, it has been identified as a key immune-escape mutation in several previous studies [29,39]. The R346_SARS2 side chain directly forms salt bridges with residues in C092 and BD56-104 (Fig 3A and 3C). Introducing every single mutation into the Lambda RBD further reduced binding affinity by 4.6- to 59-fold (S13B Fig). We previously found that antibody neutralization is more impacted by reduced antigen association than by accelerated antigen dissociation, as the latter can be compensated by avidity [22]. Notably, these mutated RBDs impacted both antibody association and dissociation, with association rates (k_on) reduced by 2.6-12-fold (S13C Fig), thereby 3 epitope mutations likely impact neutralization. For BD56-104, Lambda+R346T_SARS2 + T470N_SARS2, completely abolished binding. For C092, C807, and BD56-597, we found that only when all three mutations (R346T_SARS2, R357T_SARS2, and T470N_SARS2) were introduced simultaneously into the Lambda RBD did the antibodies completely lose binding (S13B and S13C Fig). These data suggest that C092, C807, BD56-104, and BD56-597 can still maintain antigen binding, withstanding 3–4 mutations in the epitope.

To further investigate cross-reactivity, the C092 Fab and D1H4 Fab [22] were incubated with the GX2013 S-trimer. Cryo-EM imaging of the sample revealed S-trimer disassembly. A 3.45 Šresolution structure of the C092 Fab:D1H4 Fab:GX2013 S1 complex was determined from disassembled S-trimer particles (Figs 5B, 5C, and S14). The complex was aligned to the C092 Fab:BA.4/5 RBD complex using the RBDs as references to compare the difference of C092 binding to GX2013 and BA.4/5 RBDs. The result shows that C092 binds to the same epitope area on both the GX2013 and BA.4/5 RBDs (Fig 5D). C092 binds an epitope area of 1,003 Ų on GX2013 RBD, with 675 Ų buried by HCDRs, and 328 Ų buried by LCDRs. This binding interface is smaller than the C092:BA.4/5 RBD interface (1,350 Ų), consistent with weaker binding of C092 to GX2013 RBD than BA.4/5.

In the C092:GX2013 RBD complex, the HCDR2 loop contains hydrophobic residues I52_H, I54_H, and L55_H and interacts with residues Y438_GX (L452_SARS2), Y463_GX (F490_SARS2), and L465_GX (L492_SARS2) on GX2013 RBD via hydrophobic interaction. These hydrophobic interactions are further strengthened by the SHM-introduced HCDR1 residue F33_H via extra hydrophobic contacts (Fig 5E, left panels). The HCDR3 contacts GX2013 RBD residues - E451_GX (E465_SARS2), R452_GX (R466_SARS2), L454_GX (I468_SARS2) and Y342_GX (Y351_SARS2), with Y102_H, S103_H, G104_H, S107_H, and Y109_H via extensive hydrogen bonds (Fig 5E, right panel). Further, SHM-introduced HCDR1 N31_H and LCDR1 N26_L form hydrogen bonds with S456_GX (T470_SARS2) and K348_GX (R357_SARS2), respectively, to strengthen antigen binding. (Fig 5E, left and middle panels). Therefore, we conclude that affinity maturation is a key driver for broadening the reactivity of R1-32-like antibodies toward SARSr-CoV RBDs, with SHM-introduced residues mediating additional antibody-antigen interactions.

AI-guided discovery of an ultrapotent broadly neutralizing antibody

Structure-guided artificial affinity maturation of R1-32 markedly enhanced its tolerance to viral escape mutations, whereas natural affinity maturation broadened the activity of this antibody class against sarbecoviruses with considerably diverged RBD epitopes, highlighting their potential for further expansion of breadth against emerging SARS-CoV-2 variants. Although C092, C807, BD56-104, and BD56-597 retained neutralizing activity against variants generally with 2 epitope mutations, including Lambda (L452Q_SARS2 + F490S_SARS2), BQ.1.1 (R346T_SARS2 + L452R_SARS2), XBB.1 (R346T_SARS2 + F490S_SARS2), and EG.5 (R346T_SARS2 + F490S_SARS2), the emerging KP.3 variant completely escaped their neutralization (Fig 1F and 1G).

To discover antibodies with broader reactivity, we employed an AI-driven approach to identify R1-32-like antibodies with ultrapotent activities. We developed a deep learning model to predict antibody-antigen binding affinities and used the predicted affinity rankings to virtually screen and prioritize antibody candidates (Fig 6A). We employed the evolutionary-scale ESM2 model to derive sequence representations for antigens. For antibody sequences, besides ESM2, we also adopted an antibody-specific pre-trained language model [33] as an alternative to capture sequence characteristics of antibodies. Embeddings derived from both antibody and antigen language models were processed through convolutional neural networks (CNNs) to calculate antibody-antigen binding affinities. This deep learning model was fine-tuned using 752 experimentally determined neutralization measurements of antibodies against wild-type SARS-CoV-2. The dataset was partitioned into training, validation, and test sets at a ratio of 80%, 10%, and 10% using a stringent similarity-based splitting strategy. Antibody HCDR3 sequence similarity between each test-set sample and all samples in the training and validation sets was maintained below 70%.

We evaluated our model against competing approaches, PIPR [40] and one-hot-based baseline, using multiple metrics, including Pearson’s correlation coefficient, Spearman’s correlation coefficient (Fig 6B), RMSE (Root Mean Squared Error), and MAE (Mean Absolute Error) (S16A Fig). The proposed model achieved Pearson’s and Spearman’s correlation coefficients of 0.6266 and 0.5969, with RMSE and MAE values of 0.7394 and 0.5634, respectively. In contrast, the sequence-based method PIPR obtained corresponding Pearson’s correlation, Spearman’s correlation, RMSE, and MAE values of 0.4655, 0.4456, 0.8333, and 0.6674. These results demonstrate the effectiveness of our proposed method. Furthermore, we established a baseline model by replacing the language model features with one-hot encoded vectors. Compared with language model-based methods, the baseline model showed a 40.81% decrease in Pearson’s correlation and a 34.33% decrease in Spearman’s correlation, further demonstrating the critical contribution of the learned language-model features.

The affinity prediction model was subsequently applied to a curated panel of 190 R1-32-like antibodies (S16B Fig). Notably, the AI model assigned the highest affinity score to ZL525, whereas the antibodies R1-32 and FC08 ranked among the lowest-scoring candidates. To experimentally evaluate the predictive performance of the model, we selected the top five AI-ranked antibodies, including ZL525, together with nine low-ranked antibodies for functional validation. Among the five top-ranked antibodies, four exhibited broad binding activity, whereas only one (T45) showed limited breadth. In contrast, only one of the nine low-ranked antibodies (T49) displayed broad binding, while the remaining antibodies exhibited limited activity against antigenically divergent variants (S16C and S16D Fig). These results demonstrate that the AI model effectively enriched for antibodies with broad binding activity and accurately prioritized candidates with improved breadth. Among all candidates, ZL525 displayed the highest predicted score together with the broadest binding profile and was therefore selected for further characterization.

Authentic virus neutralization assays showed that ZL525 effectively neutralized all tested SARS-CoV-2 variants, including Beta, Delta, BA.1, BA.5, BQ.1, XBB.1, EG.5, and KP.3, with IC₅₀ values ranging from 0.044 to 7.193 µg/mL (Fig 6C and 6D). Of note, ZL525 exhibited an excellent IC₅₀ value of 0.107 µg/mL against KP.3, in contrast to C092, C807, BD56-104, and BD56-597, which completely lost neutralizing activity against the emerging KP.3 variant (IC₅₀ > 100 µg/mL) (Fig 1F and 1G), underscoring ZL525 as an ultrapotent broadly neutralizing antibody. Consistent with the neutralization assays, ZL525 exhibited high affinities (below 0.71 nM) for the RBDs carrying L452_SARS2, F490_SARS2, or both mutations, including Delta, Kappa, Lambda, BA.4/5, BQ.1.1, XBB.1.5, and EG.5.1(Figs 6E and S17C). Furthermore, ZL525 maintained strong binding affinities (below 1.11 nM) for newly emergent variants, including HK.3, JD.1.1, BA.2.86, JN.1, KP.2, and KP.3 (Figs 6E and S17D), confirming its potential as a broadly therapeutic candidate. In addition, ZL525 demonstrated enhanced cross-reactivity towards sarbecoviruses, by comparison with C092, C807, BD56-104, and BD56-597, binding to all tested sarbecovirus RBDs, including SARS-CoV-1(Fig 5A).

Structural basis of ZL525 tolerance to glycan-mediated immune escape

To understand ZL525 mutation tolerance, the LP.8.1 S-trimer was incubated with ZL525 Fab and ZL58 Fab (of a IGHV3-53 antibody [41]). S-trimers in a 3-RBD “up” conformation with each RBD bound to a ZL525 Fab and a ZL58 Fab were imaged. By focused refinement, a 2.93Å ZL525 Fab: ZL58 Fab:LP.8.1 RBD complex structure was determined (Figs 6F and S15). As expected, ZL525 recognizes the convergent epitope bound by other R1-32-like antibodies. Of note, its HCDR2 contains two characteristic SHM-introduced residues A55_H and M57_H to contact the RBD hydrophobic patch W452_SARS2, F490_SARS2, and L492 _SARS2 on RBD (Fig 6F, left panel). Structural superposition of the BA.4/5 RBD:C092 and LP.8.1 RBD:ZL525 complexes reveals a conformational perturbation in the HCDR2 loop of ZL525. This conformational perturbation appears to allow HCDR2 to better contact the mutated L452W_SARS2 residue (S17A Fig). The above hydrophobic interactions are further enhanced by SHM-introduced HCDR1 residue P33_H. Similar to C092, SHM-introduced residues LCDR1 residue N26_L and LFR3 residue D70_L form hydrogen bonds and salt bridges with R357, respectively, to further enhance interaction (Fig 6F, middle panel). Compared with C092, C807, BD56-104, and BD56-597, ZL525 exhibits a distinctive structural feature: SHM-introduced residues within HCDR2 (A55_H and M57_H) directly contact epitope. Although BD56-104 and BD56-597 contain SHM-introduced residues in HCDR2 (I51V_H and A58P_H in BD56-104, A58T_H in BD56-597), these residues are not involved in epitope contact with their side chains orienting away from the epitope.

Notably, our structure reveals that the newly introduced N354_SARS2 glycosylation [42,43] maps directly onto this public antibody epitope (Figs 6F, middle panel and S17B), suggesting that glycan shielding may obstruct antibody recognition. To test this, we engineered T356K_SARS2 mutants of BA.2.86, JN.1, KP.2, and KP.3 RBDs to eliminate the N354_SARS2 glycosylation. All four antibodies (C092, C807, BD56-104, and BD56-597) exhibited restored high-affinity binding to these glycan-revertants, with markedly improved association rates (k_on) (S17D Fig), confirming that the N354_SARS2 glycosylation contributes to their escape. In contrast, the AI-identified antibody ZL525 exhibits greater tolerance to this glycan-mediated shielding, likely due to structural adaptations introduced by somatic hypermutation.

To define the molecular basis of ZL525 glycan tolerance, we systematically reverted its SHMs to their corresponding germline residues. Structural analysis identified A55_H and M57_H in HCDR2 as potentially important SHM-introduced residues involved in epitope recognition (Fig 6F). Germline reversion of these two residues (ZL525-A55L_H + M57I_H) reduced binding affinity for N354_SARS2-glycosylated RBDs by approximately 3.2- to 10-fold, demonstrating their substantial contribution to glycan tolerance (S18 Fig). Nevertheless, this revertant retained higher affinity for N354_SARS2-glycosylated RBDs than the previously identified point-mutation-tolerant antibodies C092, C807, BD56-104, and BD56-597, suggesting that additional SHMs or other sequence features contribute to the exceptional glycan tolerance of ZL525 (S17D Fig).

Consistent with this interpretation, additional reversion of the HCDR2 residues T58_H and HFR3 residue D59_H (ZL525-A55L_H + M57I_H + T58A_H + D59N_H) caused a further 1.1- to 2.6-fold reduction in binding affinity compared with reversion of the two HCDR2 residues alone. Simultaneous reversion of SHMs in HCDR1, HCDR2, and HFR3 (ZL525-A55L_H + M57I_H + T58A_H + D59N_H + H32Y_H + P33A_H) resulted in an additional 1.4- to 3.6-fold loss of affinity relative to the HCDR2 and HFR3 revertant. Structural inspection further identified four SHM-derived residues, A55_H and M57_H in HCDR2, D59_H in HFR3, and P33_H in HCDR1, that directly contribute to epitope recognition. Combined reversion of these four residues (ZL525-A55L_H + M57I_H + D59N_H + P33A_H) caused an affinity loss comparable to that observed upon complete reversion of the HCDR1, HCDR2, and HFR3 SHMs, indicating that these four residues account for most of the heavy-chain SHM-mediated effect (S18 Fig).

Additional reversion of LCDR SHMs in the background of the four-residue heavy-chain revertant further reduced binding affinity for the KP.2 and KP.3 RBDs by approximately 1.1- to 1.7-fold, indicating that light-chain SHMs also contribute to recognition of N354_SARS2-glycosylated RBDs (S18 Fig). Together, these results identify A55_H and M57_H as the principal determinants of ZL525 binding to N354_SARS2-glycosylated RBDs, while additional SHMs and sequence features distributed across both the heavy and light chains collectively enhance its exceptional glycan tolerance.

Collectively, these findings establish N354_SARS2 glycan shielding as a key antigenic change driving immune evasion in emerging KP.3 lineage variants, consistent with previous reports [44], and highlight ZL525 as a broadly neutralizing antibody capable of overcoming this glycan-mediated escape mechanism.

Ethics statement

The antibody isolation procedures in this study were approved by and conducted under the supervision of the Ethics Committee of the First Affiliated Hospital of Guangzhou Medical University (Ethics ID: ES-2020-65). All participants provided written informed consent before enrollment.

COVID-19 patient and vaccine recipient enrollment

Between December 2022 and December 2023, we enrolled a cohort of 68 individuals with RT-PCR-confirmed SARS-CoV-2 infection. All participants had received three doses of an inactivated COVID-19 vaccine (Sinopharm or CoronaVac) before experiencing sequential breakthrough infections with the BA.5 and XBB subvariants, which corresponded to the two major epidemic waves during this period. Detailed procedures for antibody isolation and sequencing are described in a separate study [41].

Cells and viruses

Expi293F cells (Thermo Fisher Scientific, Cat# A14527) were cultured in FreeStyle^TM293 Expression Medium (Thermo Fisher Scientific, Cat# A1435101) in a shaking incubator at 37 °C with 8% CO₂. Human embryonic kidney (HEK) 293T (ATCC, Cat# CRL-3216), HEK293T-ACE2, and Vero E6 cells (ATCC, Cat# CRL-1586) were cultured in 10% FBS-supplemented Dulbecco’s Modified Eagle’s Medium (DMEM) at 37 °C, 5% CO₂. Expi293F cells were used for recombinant protein expression. HEK293T cells were used to generate VSV-based pseudoviruses, while HEK293T-ACE2 and Vero E6 cells were used for neutralization assays. VSV-based pseudotyped viruses were generated as previously described [36]. Authentic SARS-CoV-2 viruses, including WT, Beta, Delta, and Omicron (BA.1, BA.4/5, BQ.1, XBB.1, EG.5, and KP.3), were isolated from clinical specimens of COVID-19 patients and stored at the Guangzhou Customs District Technology Center BSL-3 laboratory. The SARS-CoV-2 Delta strain was kindly provided by the Guangdong Provincial Center for Disease Control and Prevention, China. All experiments involving authentic SARS-CoV-2 were conducted under biosafety level 3 (BSL-3) conditions at the Guangzhou Customs District Technology Center BSL-3 Laboratory.

Expression and purification of human monoclonal antibodies

The heavy chain (IgH) and light chain (IgL) genes of antibodies were amplified and cloned into the expression vector pCMV3 using ClonExpress Ultra One Step Cloning Kit (Vazyme, C115). When the density of Expi293F cells reached 2.5 × 10⁶ cells/mL, the paired IgH and IgL plasmids were transiently co-transfected into the cells at a ratio of 1:1 using transfection reagent Polyethylenimine (PEI). The transfected cells were transferred to a shaking incubator at 33 °C with 8% CO₂. Five days post-transfection, the IgG was purified from the cell supernatant using Protein A affinity chromatography. Firstly, the supernatant was harvested by centrifugation and incubated with Protein A Resin (Genscript, China) for 2 h to enable antibody binding. After the resin was washed with PBS (Gibco), pH 7.4, the IgG bound to the resin was eluted using 0.1 M citric acid, pH 3.0, and neutralized immediately with an equal volume of 1 M Tris-HCl, pH 8.0, to maintain the pH within the neutral range. Subsequently, the fractions containing IgG were concentrated and buffer-exchanged into PBS (pH 7.4) using a 50 kDa MWCO Amicon Ultra filtration device (Merck Millipore). Purified antibodies were aliquoted, flash-frozen, and stored at −80 °C until needed.

Fab fragment production

To produce Fab gene fragment, the VH and the human IgG1 constant region CH1 were amplified from the IgH gene. The plasmids of paired Fab-IgH and IgL were transiently co-transfected into the Expi293F cells at a ratio of 1:1 using the transfection reagent Polyethylenimine (PEI) when the cell density reached 2.5 × 10⁶ cells/mL. Five days post-transfection at 33 °C, the culture supernatant was collected by centrifugation and incubated with LambdaFabSelect (Cytiva) resin for 2 h to enable Fab binding. After the resin was washed with PBS, the Fab bound to the resin was eluted with 0.1 M glycine (pH 2.0) into 1/2th volume 1 M Tris-HCl, pH 8.0, to maintain the pH within the neutral range. Subsequently, the target elutions were concentrated by a 30 kDa MWCO Amicon Ultra filtration device (Merck Millipore). Finally, the Fab was further purified and buffer-exchanged into PBS using a Superdex 200 increase 10/300 GL column (Cytiva). Purified Fab was aliquoted, flash-frozen, and stored at −80 °C until needed.

Expression and purification of SARS-CoV-2 S trimer and RBD

SARS-CoV-2 Omicron BA.4/5 S trimer ectodomain (residues 14−1,211) with HexaPro [57] mutations (F817P, A892P, A899P, A942P, K986P, and V987P) and deletion of amino acids “PRRA” (681−684) at the fusion site while retaining R685_SARS2 was constructed into the expression vector pCDNA3.1 with an N-terminal µ-phosphatase signal peptide and a C-terminal TEV-cleavage site, a T4 foldon trimerization motif, and a hexahistidine tag [58,59]. Various SARS-CoV-2 variants S trimer ectodomain (S-R) with “PRRA” (681−684) deletion at the fusion site were also constructed using the same method. The coding sequence of SARS-CoV-2 RBD (residues 332−527) with an N-terminal µ-phosphatase signal peptide and a C-terminal hexahistidine tag was cloned into the expression vector pCDNA3.1 [58]. All expression plasmids of S-trimer and RBD were transiently transfected into Expi293F cells using transfection reagent Polyethylenimine [60]. Five days post-transfection at 33 °C, the cultural supernatant was collected and supplemented with 25 mM phosphate, pH 8.0, 300 mM NaCl, 5 mM imidazole, and 0.5 mM PMSF recirculated onto a HiTrap TALON crude column (Cytiva). Subsequently, the column was washed with buffer A (25 mM phosphate, pH 8.0, 300 mM NaCl, 5 mM imidazole), and target protein was eluted with a 100 mL linear gradient to 100% buffer B (25 mM phosphate, pH 8.0, 300 mM NaCl, 500 mM imidazole). Fractions containing the target protein were concentrated with an MWCO Amicon Ultra filtration device (Merck Millipore) and buffer-exchanged into PBS. Additionally, all RBDs were further purified and buffer-exchanged into PBS using a Superdex 75 increase 10/300 GL column (Cytiva) to remove RBD dimers. All SARS-CoV-2 spike and RBD variants were purified as described above. Purified S-trimers and RBDs were aliquoted, flash-frozen, and stored at −80 °C until needed.

Binding kinetics and affinity assessment of antibodies by BLI

Binding kinetics and affinities of mAbs against SARS-CoV-2 spikes or RBDs and SARSr-CoVs RBDs were assessed by Biolayer Interferometry (BLI) on an Octet RED96 instrument (FortéBio, USA). The assays were performed at 25 °C, and all proteins were diluted to PBST buffer (PBS with 1 mg/mL BSA and 0.02% v/v Tween-20). Biosensors were pre-equilibrated in the PBST buffer for 10 min before the experiments. Initially, antibodies (11 µg/mL) were immobilized onto Protein A biosensors to a level of ~1.5 nm. After a 60 s baseline step in PBST buffer, biosensors loaded with mAb were exposed (300 s) to the spikes (from 200 to 3.125 nM in 2-fold serial dilutions) or various RBDs (200 nM) to measure association. Subsequently, the biosensors were dipped into PBST (600 s) to measure dissociation of spikes or RBDs from the biosensor surface. A blank reference was set for each reaction. Data was reference-subtracted and analyzed using the FortéBio data analysis software HT v12.0.2.59 (FortéBio) by fitting to single- or double-phase association and dissociation kinetics to determine k_on, k_off and K_D. Raw data and fit data were plotted in GraphPad Prism 8.0.

Antibody competition assay by BLI

The competitive assays between antibodies or between antibodies and ACE2 were measured by Biolayer Interferometry (BLI) using an Octet RED96 instrument (FortéBio, USA). All experimental steps were performed at 25 °C. All proteins were diluted to PBST buffer (PBS with 1 mg/mL BSA and 0.02% v/v Tween-20), and anti-His biosensors were pre-equilibrated in PBST buffer for 10 min before being immobilized with purified SARS-CoV-2 RBD-His protein (30 μg/mL). Biosensors loaded with RBD were saturated with the first ligand (antibodies or ACE2 with a concentration of 200nM), and then binding of the second ligand (antibodies or ACE2 with a concentration of 200nM) or buffer as control was measured for 300 s. Data was analyzed using the FortéBio data analysis software HT v12.0.2.59 (FortéBio) and raw data was plotted in GraphPad Prism 8.0.

Pseudovirus neutralization assay

The neutralization potency of antibodies against SARS-CoV-2 wild-type and multiple variants (including Delta, Lambda, and Omicron sublineages BA.1, BA.2, BA.4/5, BA.2.75.2, BQ.1.1, and EG.5) was evaluated using a pseudovirus neutralization test following established protocols [36]. The antibody samples were diluted to an initial concentration of 10 μg/mL and subjected to eight gradient dilutions (3× serial dilutions), while the pseudovirus was diluted to 1.3 × 10⁴ TCID₅₀/mL. After incubating the antibody samples and pseudovirus at 37 °C for 1 hour, HEK293T-ACE2 cells were added. After 24 hours of culture, luciferase detection reagent was added and placed in the multifunctional micropore detector (PE Ensight) to read the luminous value. Results were analyzed, and IC₅₀ values were calculated by GraphPad Prism 8.0.

Authentic virus neutralization assay

All experiments involving authentic SARS-CoV-2 virus were conducted under biosafety level 3 (BSL-3) containment at the Guangzhou Customs District Technology Center. Antibodies were initially diluted to 100 μg/mL and then subjected to 4-fold serial dilutions. Each diluted mAb sample was combined with an equal volume of virus suspension containing 200 focus-forming units (FFU) of specified SARS‑CoV‑2 strains (WT, Beta, Delta, and Omicron subvariants BA.1, BA.5, BQ.1, XBB.1, EG.5, and KP.3). Mixtures were added to Vero E6 cells in 96-well cell culture plates for 1 h at 37 °C. After removal of the mixtures, fresh overlay medium consisting of MEM supplemented with 1.2% carboxymethylcellulose was added to each well. Following 24 hours of incubation, the overlay was discarded, and the cells were fixed with 4% paraformaldehyde (Biosharp, China, Cat# BL539A) and permeabilized using 0.2% Triton X-100 (Sigma, USA, Cat# T8787). Cells were incubated with a human anti-SARS-CoV-2 nucleocapsid protein monoclonal antibody (obtained by laboratory screening) at 37 °C for 1 h. After three washes with 0.15% PBST, cells were incubated with an HRP-labeled goat anti-human secondary antibody (Jackson ImmunoResearch Laboratories, Cat# 609-035-213) at 37 °C for 1 h. Following additional washes, foci were developed using TrueBlue Peroxidase Substrate (KPL, Gaithersburg, MD, Cat# 50-78-02), and counted with an ELISPOT reader (Cellular Technology Cleveland, OH). The foci reduction neutralization test titer (FRNT₅₀) was calculated by the Spearman–Karber method.

Cryo-EM sample preparation and data collection

The BA.4/5 S-R/6P S-trimer was mixed with the corresponding Fab and ACE2 at a 3(S-protomer):3(Fab):3(ACE2) molar ratio for 1 min to form the C092, C807, BD56–104, and BD56–597 Fab complexes. Similarly, the GX2013 S-trimer was mixed with C092 and D1H4, and the LP.8.1 S-trimer was mixed with ZL525 and ZL58, each at a 3:3:3 molar ratio for 1 min to assemble the C092 Fab:D1H4 Fab:GX2013 S-trimer and ZL525 Fab:ZL58 Fab:LP.8.1 S-R/6P complexes, respectively. All mixtures contained 0.1% octyl glucoside, except for the C092 Fab:D1H4 Fab:GX2013 complex, which contained 0.05% octyl glucoside. Samples were loaded onto freshly glow-discharged (15 mA, 30 s) holey carbon grids (Quantifoil, Cu R1.2/R1.3) at 4 °C and 100% humidity. The grids were blotted for 2.5 s with a force of 4 and samples were frozen immediately in liquid ethane using a Vitrobot (Thermo Fisher Scientific).

For the C092 Fab:ACE2:Omicron BA.4/5 S-R/6P and C092 Fab:D1H4 Fab:GX2013 S-trimer complexes, cryo-EM data were collected using a 200 keV Talos Arctica electron microscope (ThermoFisher Scientific) equipped with a K3 direct electron detector (Gatan). Movies were recorded using the SerialEM version 3.8.7 software at a nominal magnification of 45,000× with a calibrated pixel size of 0.88 Å and a defocus range from −0.8 to −2.5 μm. Samples were exposed for 1.83 s with a dose rate of 25.4 e⁻/pixel/s fractionated over 27 frames, giving a total electron dose of 60 e⁻/Ų.

To achieve higher resolution, a separately prepared C092 Fab:ACE2:BA.4/5 S-R/6P complex was additionally imaged in a Titan Krios electron microscope (Thermo Fisher Scientific) operating at 300 keV equipped with a SelectrisX energy filter (slit width 10 eV) and a Falcon 4 direct electron detector. The C807, BD56-104, and BD56-597 complexes, as well as the ZL525 Fab:ZL58 Fab:LP.8.1 S-R/6P complex, were also imaged in the Titan Krios electron microscope under the same setup. Movie stacks were automatically recorded using EPU with the electron event representation (EER) mode at a nominal magnification of ×165,000 with a calibrated pixel size of 0.73 Å and nominal defocus values ranging between −0.6 to −2.0 μm. Each stack was recorded and exposed at a dose rate of 6.53 e⁻/pixel/s for 4.08 s resulting a total dose of ~ 50 e⁻/Ų. All movie stacks were imported into cryoSPARC live (v3.3.2/v4.2.0) [61] for pre-processing, which includes patched motion correction, contrast transfer function (CTF) estimation, and bad image rejection.

Cryo-EM data processing

Data processing was carried out using cryoSPARC v3.3.2/v4.2.0. After removing bad images, particles were picked by blob-picking on images and extracted for 2D Classification. Well-defined 2D classes were selected as templates for template-picking.

For the C092 Fab:ACE2:BA.4/5 S-R/6P dataset collected at 200 keV (S5 Fig), template-picked particles were extracted and subjected to two rounds of 2D classification to remove low-quality particles. Particles of either disassembled spike (S1) or Fab bound S-trimers were separated by 2D classification into two sub-datasets. For the S1 sub-dataset, Ab-initio Reconstruction was performed to generate initial 3D models, before the selected models were used as references in Heterogeneous Refinement. A well-defined structure was constructed by Non-uniform Refinement. To improve local resolution, a soft mask was applied to the ACE2:RBD:Fab interface for Local Refinement, yielding a final map at 3.84 Å resolution. For the S-trimer sub-dataset, three rounds of Ab-initial Reconstruction were conducted to remove bad particles. All resulting models from the final round were used as references in Heterogeneous Refinement. Particles generating highly similar structures were combined and subject to Non-uniform Refinement. This process yielded two distinct S-trimer reconstructions, featuring either 2 or 3 RBD “up” conformations, allowing simultaneous binding of either 2 or 3 Fab and ACE2 molecules.

To improve resolution, a separate C092 Fab:ACE2:BA.4/5 S-R/6P complex dataset was collected on a Titan Krios operating at 300 keV (S6 Fig). Template-picked particles were similarly split into S1 and S-trimer subsets via 2D classification. For the S1 sub-dataset, Ab-initio Reconstruction and Heterogeneous Refinement were performed to generate 3D reconstructions. To improve reconstruction, Topaz Training and Picking was performed before picked particles were subjected to 2D Classification, Ab-initio Reconstruction, and Heterogeneous Refinement again. The final particles were subjected to Non-uniform Refinement, followed by Local Refinement with a mask focused on the ACE2:RBD:Fab interface regions to yield a final map at 2.58 Å resolution. For the S-trimer sub-dataset collected at 300 keV, A good 3D reconstruction was not obtained due to the absence of side-view projections in 2D classes, even after Topaz picking.

For the C807/BD56-104/BD56-597 Fab:ACE2:BA.4/5 S-R/6P datasets (S6 and S7 Figs), Template-picked particles were similarly divided into two sub-datasets. The S1 sub-datasets were processed in a workflow identical to the C092 Fab:ACE2:BA.4/5 S-R/6P S1 sub-dataset collected at 300 keV. The S-trimer sub-datasets were processed using similar workflows, each yielding a structure showing a S-trimer simultaneously bound by three Fabs and three ACE2s. For the C807 Fab bound S-trimer sub-dataset, Topaz training was performed to further pick S-trimer particles. The Topaz picked particles were subjected to 2D Classification, Ab-initio Reconstruction, Heterogeneous Refinement, and Non-uniform Refinement to reconstruct a final map. For the BD56–104 Fab bound S-trimer sub-dataset, Topaz training and picking was performed after Heterogeneous Refinement. For the BD56-597 S-trimer sub-dataset, reconstruction was obtained without performing Topaz picking.

For the C092 Fab:D1H4 Fab:GX2013 dataset collected at 200 keV (S14 Fig), Fab bound S1 particles from disassembled S-trimers were template-picked before subjecting to 2D Classification, Ab-initio Reconstruction, and Homogeneous Refinement. Topaz training and picking were performed to further pick particles. The final reconstruction was obtained using Topaz picked particles using Non-uniform Refinement, with a resolution of 3.45 Å.

For the ZL525 Fab:ZL58 Fab:LP.8.1 S-R/6P dataset (S15 Fig), only Fab bound S-trimers were observed by auto-picking and subsequent 2D classification. S-trimer particles were subject to Ab-initio Reconstruction, Homogeneous Refinement, and Non-uniform Refinement, imposing C3 symmetry to generate a map with a resolution of 2.75 Å. To improve the resolution of the RBD:Fab interfaces, 3D classification with a focused mask (ZL525 Fab:ZL58 Fab:LP.8.1 RBD) was applied to remove low-quality particles, followed by Local Refinement. Ultimately, the focused ZL525 Fab:ZL58 Fab:LP.8.1 RBD map was refined to a resolution of 2.93 Å.

All resolutions were estimated at the 0.143 criterion from phase-randomization-corrected Fourier shell correlation (FSC) curves calculated between two independently refined half-maps, multiplied by soft-edged solvent masks, in either RELION v4.0 or cryoSPARC [61]. Additional data processing details are summarized in S5–S8, S14, and S15 Figs and S11 Table.

Cryo-EM model building and structure refinement

A previously determined structure of R1-32 Fab:ACE2:SARS-CoV-2 RBD complex (PDB: 7YDI) was fitted into the Cryo-EM map in UCSF Chimera v1.15 [62] and used as the starting model for model building of the four complexes (C092/C807/BD56-104/BD56-597 Fab:ACE2:Omicron BA.4/5 RBD. Manual model building was carried out in Coot [63] in order to adjust Ramachandran, rotamers, bond geometry restraints, etc. Structure refinement was performed automatically using real-space refinement in PHENIX [64]. For C092/C807/BD56-104/BD56-597 Fab:ACE2:Omicron BA.4/5 S-timer complexes, the structures (PDB:7YE5/7YEG) were used as initial model. For C092 Fab:D1H4 Fab:GX2013 RBD and ZL525 Fab:ZL58 Fab LP.8.1 RBD complexes, the structures (PDB:8ZY6 and 8WXL) were used as initial model respectively. Model validation statistics were summarized in S11 Table.

Ligand-induced conformational change assay

SARS-CoV-2 S-R spike at 1 mg/mL (7.09 µM) was incubated with ACE2-Fc or antibodies at a 1:1.1 molar ratio for 1 h at room temperature. The samples were subsequently treated with 50 µg/mL proteinase K for 30 min at 4 °C. Non-reducing SDS-PAGE loading buffer (5×) was added to each sample immediately before boiling at 98 °C for 5 min to stop the reaction. Samples were separated by SDS–PAGE and transferred onto a PVDF membrane. After blocking with 5% (w/v) skimmed milk in TBST buffer, the membrane was incubated with a rabbit anti-SARS-CoV-2 S2 polyclonal antibody (1:2,500 dilution, Sino Biological, 40590-T62) and subsequently with a goat anti-rabbit IgG conjugated to horseradish peroxidase (1:1,000 dilution, Beyotime, A0208). The S proteins on the membrane were detected by chemical luminescence using Pierce ECL western blotting substrate (Thermo Fisher, 32106). To further assess the effect of non-ACE2 competing antibodies on S-R spike conformation, the S-R spike protein was pre-incubated with antibodies (1 h), followed by incubation with ACE2-Fc (1 h). Samples were analyzed by western blotting as described above.

Pre-trained ESM2 model, antibody language models, and virtual screening

We employed the pre-trained ESM2 (150M) model to capture sequence information from antigen sequences. This model was pre-trained on UniRef50 [65], enabling it to learn evolutionary relationships across a vast number of protein families. Given its strong capacity in representing general protein sequences, we applied ESM2 to extract embedding vectors for antigens. For antibody sequences, in addition to ESM2, we also adopted antibody-specific pre-trained language models as an alternative approach to capture comprehensive sequence characteristics. As described in our previously reported A2binder model [33], we pre-trained antibody-specific language models on 1.4 billion antibody sequences sourced from the OAS database. After removing duplicate sequences and those containing non-canonical amino acids, the curated dataset consisted of 1.2 billion heavy chain and 210 million light chain sequences. Using the RoFormer [66] architecture implemented in the A2binder model, separate models were pre-trained for heavy and light chains. A Unique Amino Acid (UAA) tokenizer was adopted to represent each residue as a discrete token. Self-supervised learning was performed via a Masked Amino Acid (MAA) objective, in which randomly obscured residues were predicted from contextual cues, enabling the model to infer functional properties of antibody sequences. Then, the embeddings generated by both the antibody and antigen models were processed using convolutional neural networks to predict the antibody–antigen binding affinity [33].

The model was fine-tuned on a consolidated IC₅₀-labeled neutralization dataset assembled from multiple public sources [1,12,28,31,34,67–70]. A total of 752 data entries regarding the neutralizing activity of antibodies against wild-type SARS-CoV-2 were collected. The entire dataset was split into training, validation, and test sets with a ratio of 80%, 10%, and 10%, respectively, while strictly ensuring that the sequence similarity of antibody HCDR3 regions between any sample in the test set and those in the training and validation sets remained below 70%. Here, similarity was computed based on the HCDR3 regions of the antibody sequences using MMseqs2.

Finally, we applied the model to screen a library of 190 R1-32-like monoclonal antibodies in silico, of which 102 antibodies were derived from published data and 88 antibodies were isolated from convalescent individuals. Detailed information on isolating the latter set of antibodies is described in a separate study [41]. Candidates were ranked by predicted binding affinities, and top-scoring antibodies were prioritized for experimental validation.

S1 Fig. Binding of mutation-tolerant antibodies to the RBDs of SARS-CoV-2 variants.

A, Identification of antibodies BD56-104 and BD56-597 that may tolerate the L452_SARS2 mutation from the reported IGHV1-69/IGLV1-40 encoded antibody database. The data underlying the figure can be found in the S4 Data file. B–E, Binding curves of C092 (B), C807 (C), BD56-104 (D), and BD56-597 (E) to a panel of SARS-CoV-2 RBDs as analyzed by BLI. Antibodies were immobilized onto Protein A biosensors and submerged into RBD solutions at a concentration of 200 nM. Detailed binding kinetic parameters are summarized in S1 Table. The K_D values and normalized K_D fold changes are shown in Fig 1C. The data underlying the figure can be found in the S4 Data file.

https://doi.org/10.1371/journal.pbio.3003996.s001

(TIF)

S2 Fig. Binding of mutation-tolerant antibodies to the S-trimers of SARS-CoV-2 variants.

Binding curves of C092, C807, BD56-104, and BD56-597 to SARS-CoV-2 wild-type, Delta, Lambda, Omicron BA.1, BA.2, BA.4/5, BA.2.75.2, and BQ.1.1 spikes are shown. Binding assays were performed using BLI, with antibodies immobilized onto Protein A sensors and surmeged into a dilution series of spike proteins (200 to 3.125 nM). Binding kinetics parameters are summarized in S2 Table. The data underlying the figure can be found in the S5 Data file.

https://doi.org/10.1371/journal.pbio.3003996.s002

(TIF)

S3 Fig. Competition assays between R1-32-like antibodies and ACE2-Fc or R1-32.

A–D, Competition binding to SARS-CoV-2 RBD between C092 (A), C807 (B), BD56-104 (C) and BD56-597 (D) and ACE2-Fc or R1-32 was assessed by BLI. SARS-CoV-2 RBD was immobilized onto HIS1K biosensors, and immobilized RBD was saturated with ACE2-Fc (left panels) or R1-32 (middle panels) before incubation with C092, C807, BD56-104, or BD56-597 (purple line). Binding of the immobilized RBD to only C092, C807, BD56-104, or BD56-597 (olive line) was assayed as controls. Conversely, the biosensors immobilized with the RBD were saturated with C092, C807, BD56-104, or BD56-597 (right panels) before incubation with ACE2-Fc (purple line) or R1-32 (red line). Binding of the immobilized RBD to only ACE2-Fc (olive line) or R1-32 (black line) was assayed as controls. The data underlying the figure can be found in the S6 Data file.

https://doi.org/10.1371/journal.pbio.3003996.s003

(TIF)

S4 Fig. Mutation-tolerant antibodies inhibit the conformational change of SARS-CoV-2 S-trimers.

A–B, Inhibition effect of R1-32-like antibodies on the conformational change of the SARS-CoV-2 wild-type spike (S-R/WT). S-R/WT (7.09 µM) was incubated with ACE2-Fc or antibodies (S-protomer:ACE2-Fc/IgG = 1:1.1 molar ratio) for 1 h (A). Samples were analyzed by western blotting to detect the generation of the 55 kDa proteinase K-resistant core, which is a signature of the post-fusion S2 structure. Only ACE2-Fc and class 1 antibody YB9-258 induced post-fusion structures. Inhibition of conformational change by non-ACE2 competing antibodies was assayed by antibody preincubation (1 h) before further incubation (1 h) with ACE2-Fc (B). Only R1-32-like antibodies were able to abolish fusogenic spike conformational change. C-D, Inhibition effect of R1-32-like antibodies on the conformational change of the SARS-CoV-2 BA.4/5 S-trimer (S-R/BA.4/5). E-F, Inhibition effect of R1-32-like antibodies on the conformational change of the SARS-CoV-2 EG.5.1 S-trimer (S-R/EG.5.1). Original uncropped western blot images can be found in the S1 Raw Images file.

https://doi.org/10.1371/journal.pbio.3003996.s004

(TIF)

S8 Fig. Resolution assessment of cryo-EM structures.

A, D, G, and J, Local resolution maps of C092 Fab (A), C807 Fab (D), BD56-104 Fab (G), and BD56-597 Fab (J) in complex with ACE2 and Omicron BA.4/5 S-R/6P. B, E, H, and K, Global resolution assessment by Fourier shell correlation (FSC) at the 0.143 criterion. The data underlying the figure can be found in the S7 Data file. C, F, I, and L, Correlations of model versus map by FSC at the 0.5 criterion. The data underlying the figure can be found in the S7 Data file.

https://doi.org/10.1371/journal.pbio.3003996.s008

(TIF)

S9 Fig. Representative cryo-EM densities showing epitope binding by R1-32-like antibodies.

A–D, Representative densities (gray surfaces) of the C092 (A), C807 (B), BD56-104 (C), and BD56-597 (D) epitopes and CDR loops. E, R1-32 HCDR2, HCDR3, and LC epitopes are shown from rotated views (PDB: 7YDI) for comparison. RBD, R1-32-H, and R1-32-L residues are colored in cyan, magenta, and purple, respectively. The backbone carbonyl oxygens and amide nitrogens are indicated by red and blue dots, respectively. Hydrogen bonds are shown as black dash lines.

https://doi.org/10.1371/journal.pbio.3003996.s009

(TIF)

S10 Fig. Frequency of somatic hypermutations (SHMs) in antibodies isolated from different exposure histories.

A–C, Frequency of SHMs in the heavy chains of antibodies isolated from WT convalescents or vaccinees (A), BA.1/BA.2 convalescents (B), and BA.5 convalescents (C). D, Frequency of SHMs in the heavy chains of antibodies escaped by BQ.1.1 and XBB variants. E–G, Frequency of SHMs in the light chains of antibodies isolated from WT convalescents or vaccinees (E), BA.1/BA.2 convalescents (F), and BA.5 convalescents (G). H, Frequency of SHMs in the light chains of antibodies escaped by BQ.1.1 and XBB variants.

https://doi.org/10.1371/journal.pbio.3003996.s010

(TIF)

S11 Fig. Binding assays of constructed R1-32-like antibody mutants to the RBDs.

A–D, Binding curves of mutants derived from C092 (A), C807 (B), BD56-104 (C), and BD56-597 (D) to the RBDs of SARS-CoV-2 wild-type, L452R, F490S, Kappa, Delta, Lambda and BA.4/5. Binding assays were performed by BLI, with detailed binding kinetics parameters summarized in S3 Table. Binding affinity changes to different mutants are shown in Fig 5B–5E. The data underlying the figure can be found in the S8 Data file.

https://doi.org/10.1371/journal.pbio.3003996.s011

(TIF)

S12 Fig. Binding of mutation-tolerant antibodies to the RBDs of SARSr-CoVs.

A–F, Binding curves of C092 (A), C807 (B), BD56−104 (C), BD56−597 (D), ZL525 (E), and R1-32 (F) to the RBDs of SARS-related coronaviruses (SARSr-CoVs). Binding assays were performed by BLI, with RBDs at a fixed concentration of 200 nM. The K_D values are shown in Fig 5A. Detailed binding kinetic parameters are summarized in S5 Table. The data underlying the figure can be found in the S9 Data file.

https://doi.org/10.1371/journal.pbio.3003996.s012

(TIF)

S13 Fig. RBD mutations and their effects on the binding of mutation-tolerant antibodies.

A, Binding of R1-32-like antibodies to RBDs with single amino acid substitutions. The Loop-R indicates the replacement of the SARS-CoV-2 ACE2 binding loop (474−486) with the corresponding SARS-CoV-1 residues (461−472). Bottom panels: Relative affinity changes are normalized to the K_Ds for wild-type RBD (set as 100%) and shown as bar graphs. Fold changes in K_D relative to the wild-type RBD are indicated above the bars. B, Binding affinities of R1-32-like antibodies to single and combination mutations constructed on the Lambda RBD. Bottom panels: Relative affinity changes are shown as bar graphs. Fold changes in K_D are indicated above the bars. C, Relative changes in association rate constant (k_on) are normalized to the k_on for wild-type RBD (set as 100%) and are shown as bar graphs. Fold changes in k_on are indicated above the bars. K_D and k_on values derived from fitting BLI binding curves in (A) and (B) are summarized in S6 and S7 Tables. The data underlying the figure can be found in the S10 Data file.

https://doi.org/10.1371/journal.pbio.3003996.s013

(TIF)

S14 Fig. Cryo-EM structure determination of the C092 Fab:D1H4 Fab:GX2013 S1 complex.

A, Cryo-EM data processing pipeline for the C092 Fab:D1H4 Fab:GX2013/5P dataset and local resolution assessment of the C092 Fab:D1H4 Fab:GX2013 S1 complex strucuture. B, Global resolution assessment by Fourier shell correlation (FSC) at the 0.143 criterion. The data underlying the figure can be found in the S7 Data file. C, Correlations of model versus map by FSC at the 0.5 criterion. The data underlying the figure can be found in the S7 Data file. D, Representative cryo-EM densities showing the C092 CDR loops and the C092 epitope on the GX2013 RBD.

https://doi.org/10.1371/journal.pbio.3003996.s014

(TIF)

S15 Fig. Cryo-EM structure determination of the ZL525 Fab:ZL58 Fab:LP.

8.1 S-R/6P complex. A, Cryo-EM data processing pipelines for the ZL525 Fab:ZL58 Fab:LP.8.1 S-R/6P dataset and local resolution assessment of the ZL525 Fab:ZL58 Fab:LP.8.1 RBD complex structure. B, Global resolution assessment by Fourier shell correlation (FSC) at the 0.143 criterion. The data underlying the figure can be found in the S7 Data file. C, Correlations of model versus map by FSC at the 0.5 criterion. D, Representative cryo-EM densities showing the ZL525 CDR loops and epitope. The data underlying the figure can be found in the S7 Data file.

https://doi.org/10.1371/journal.pbio.3003996.s015

(TIF)

S16 Fig. AI-guided affinity prediction and validation of representative antibodies.

A, Comparison of different methods in terms of root mean square error (RMSE) and mean absolute error (MAE). The data underlying the figure can be found in the S3 Data file. B, Histogram and kernel density estimation (KDE) curve of predicted affinity values. The data underlying the figure can be found in the S3 Data file. C, Binding affinities of representative top-ranked and low-ranked antibodies selected from B. Top-ranked antibodies are labeled in black, whereas low-ranked antibodies are labeled in red. RBDs carrying L452, F490, or both substitutions are colored in red, blue, and purple, respectively. RBDs carrying L452 substitution together with N354 glycosylation are colored in green. Dissociation constant K_Ds (nM) are shown, with values <1, 1–100, >100 nM, and weak or no binding colored in blue, green, yellow, and orange, respectively. Binding curves and detailed kinetic parameters are provided in D and S8 Table. The data underlying the figure can be found in the S3 Data file.

https://doi.org/10.1371/journal.pbio.3003996.s016

(TIF)

S17 Fig. Structural and functional characterization of N354 glycosylation on SARS-CoV-2 RBD.

A, Structural superposition of the LP.8.1 RBD:ZL525 Fab and BA.4/5 RBD:C092 Fab complexes, using the LP.8.1 RBD as the reference. The LP.8.1 RBD:ZL525 Fab complex is shown in color, and the BA.4/5 RBD:C092 Fab complex is depicted in gray. B, Mapping of the N354 glycosylation site on the RBD. The structure of the BA.2.86 RBD is colored gray. C, Binding curves of ZL525 to a panel of SARS-CoV-2 RBDs. The K_D values and normalized K_D fold changes are shown in Fig 6E. The data underlying the figure can be found in the S11 Data file. D, Binding curves of C092, C807, BD56-104, BD56-597, and ZL525 against RBDs from newly emergent variants and engineered T356K mutants. Binding assays were performed by BLI. Detailed binding kinetic parameters are summarized in S9 Table. The data underlying the figure can be found in the S11 Data file.

https://doi.org/10.1371/journal.pbio.3003996.s017

(TIF)

S18 Fig. Affinity maturation confers ZL525 glycan tolerance.

A–B, Heavy chain (A) and Light chain (B) sequence alignments of ZL525 with other R1-32-like antibodies (R1-32, C092, C807, BD56-104, and BD56-597). The sequences of R1-32 are used as the references, with somatic hypermutation-derived amino acids marked in red. The HCDRs and LCDRs residues are highlighted in salmon. C, Relative binding affinities of antibody mutants derived from ZL525 to the RBDs of SARS-CoV-2 wild-type, BA.2.86, JN.1, KP.2, and KP.3. Affinity changes of mutants are normalized to K_Ds of the corresponding unmodified ZL525 (set as 100%). Binding curves are shown in D, and K_D values used for comparisons are shown in S10 Table. The data underlying the figure can be found in the S12 Data file.

https://doi.org/10.1371/journal.pbio.3003996.s018

(TIF)

S12 Table. L452/F490 mutations and N354 glycosylation in SARS-CoV-2 variants.

A dash (–) indicates no amino acid substitution; amino acid changes are shown as the original and substituted residues. A check mark (√) indicates the presence of a glycosylation mutation at the indicated position.

https://doi.org/10.1371/journal.pbio.3003996.s030

(XLSX)

S7 Data. The individual numerical values for the following figure panels: S8B, S8C, S8E, S8F, S8H, S8I, S8K, S8L, S14B, S14C, S15B, and S15C Figs.

https://doi.org/10.1371/journal.pbio.3003996.s037

(XLSX)