Computationally restoring the potency of a clinical antibody against SARS-CoV-2 Omicron subvariants

Abstract

The COVID-19 pandemic underscored the promise of monoclonal antibody-based prophylactic and therapeutic drugs^1-3, but also revealed how quickly viral escape can curtail effective options^4,5. With the emergence of the SARS-CoV-2 Omicron variant in late 2021, many clinically used antibody drug products lost potency, including Evusheld^™ and its constituent, cilgavimab^4,6. Cilgavimab, like its progenitor COV2-2130, is a class 3 antibody that is compatible with other antibodies in combination⁴ and is challenging to replace with existing approaches. Rapidly modifying such high-value antibodies with a known clinical profile to restore efficacy against emerging variants is a compelling mitigation strategy. We sought to redesign COV2-2130 to rescue in vivo efficacy against Omicron BA.1 and BA.1.1 strains while maintaining efficacy against the contemporaneously dominant Delta variant. Here we show that our computationally redesigned antibody, 2130-1-0114-112, achieves this objective, simultaneously increases neutralization potency against Delta and many variants of concern that subsequently emerged, and provides protection in vivo against the strains tested, WA1/2020, BA.1.1, and BA.5. Deep mutational scanning of tens of thousands pseudovirus variants reveals 2130-1-0114-112 improves broad potency without incurring additional escape liabilities. Our results suggest that computational approaches can optimize an antibody to target multiple escape variants, while simultaneously enriching potency. Because our approach is computationally driven, not requiring experimental iterations or pre-existing binding data, it could enable rapid response strategies to address escape variants or pre-emptively mitigate escape vulnerabilities.

Figure ED1:. Optimized, single-concentration screening data in Gyrolab immunoassays allow selection of candidates from Set 1 (n=230) for down-stream characterization.

Parental mAb COV2-2130 (orange circles) and positive control mAb S309 (magenta squares) serve as references for computationally designed mAbs in single-concentration immunoassays. Target antigens are (A) wild type WA1/2020, (B) Delta, (C) Omicron BA.1 and (D) Omicron BA.1.1. Each screened antibody and antigen combination was evaluated in two replicate assays; each of the controls was replicated in two replicate assays for each of three groups of antibodies, resulting in six replicate assays for each point on the control curves. All error bars indicate standard deviation.

Figure ED2:. ELISA screening allows set for down selection of candidates from Set 2 (n=204), for down-stream characterization.

Parental mAb COV2-2130 (orange circles; 3 technical replicates) and positive control mAb S309 (magenta squares; 3 technical replicates) serve as references for computationally designed mAbs (black curves). Purple triangles are 2130-1-114-112; blue diamonds are 2130-1-0104-024. Each designed antibody had a single measurement (n=1) at each concentration. All curves are 4-parameter logistic fits, produced in GraphPad Prism. Target antigens are RBD from (A) wild type WA1/2020, (B) Omicron BA.1 (Acro), which is biotinylated, (C) Omicron BA.1, and (D) Omicron BA.1.1. For the biotinylated antigen in (B)(Acro Biosystems, cat. SPD-C82E4), an additional coat and wash cycle was required to prepare the ELISA plate with streptavidin.

Figure ED3:. Normalized pseudoviral neutralization of recently emerged VOCs by 2130-1-0114-112.

(B) COV2-2130 and 2130-1-0114-112 potently neutralize WA1/2020 D614G and (C) only 2130-1-0114-112 potently neutralizes BA.1, consistent with other pseudoviral neutralization assays. (D) 2130-1-0114-112 potently neutralizes BA.2.75, outperforming COV2-2130 by 90-fold. (E-F) 2130-1-0114-112 loses substantial potency in the context of BA.4.6 and artificially-produced BA.2.75+R346T but retains measurable neutralization, demonstrating mitigation of this critical weakness of COV2-2130. All sub-figures (B-F) show the mean of two technical replicates and curves show a fitted four-parameter logistic curve with fixed minimum and maximum values. (G) Neutralization IC50 values in ng/ml. “>” indicates IC50 greater than 10,000 ng/mL. All analysis conducted in GraphPad Prism.

Figure ED4.. Normalized pseudoviral neutralization of recently emerged VOCs

Pseudoviral neutralization assays testing COV2-2130 and 2130-1-0114-112 against D614G and BA.2.75 R346T F486S, BQ.1.1, XBB, and BN.1 variants. COV2-2130 has no detectable activity against all variants tested. In contrast, 2130-1-0114-112 maintains neutralization of BA.2.75 R346T F486S, but loses detectable neutralization activity against BQ.1.1 and XBB and exhibits a near-complete loss of neutralization activity against BN.1. All data points represent the mean of two technical replicates, while error bars denote the standard deviation. Curves represent a four-parameter logistic curve fit to the data using GraphPad Prism with fixed minimum and maximum values (0 and 100, respectively).

Figure ED5:. Authentic virus neutralization in plaque assays using Vero E6-TMPRSS2-T2A-ACE2 (VAT) cells.

Plaque assay neutralization of (A) Delta, (B) Omicron BA.1 and (C) Omicron BA.1.1 viruses. Data are represented as the normalized infection of mAb-treated virus to virus treated with control human IgG (Invitrogen). For S309, each point shows the mean of four technical replicates; all other points are means of two technical replicates; error bars show standard deviation. Curves are two-parameter (IC50, hill-slope) logistic fits to normalized response. (D) IC50 values (ng/ml, “>” indicates IC50 greater than 10,000 ng/mL) show that 2130-1-0104-024, while having only two mutations from COV2-2130, remains potent against BA.1 and suffers a 20-times loss in potency against BA.1.1. 2130-1-0114-112 is strongly potent against all three tested variants. All analysis performed in GraphPad Prism.

Figure ED6:. Cryo-EM structure of neutralizing antibody 2130-1-0114-112 in complex with Omicron BA.2 RBD.

(A) RBD residues in 7 Å distance form 2130-1-0114-112. RBD in red, with BA.2 mutated residues in orange, and 2130-1-0114-112 in yellow/green. (B) Left, 3D representation of the interaction plot between RBD and 2130-1-0114-112 HC. 2130-1-0114-112 is shown as stick (yellow) and RBD as gray spheres with the contact residues in red. Contact residues labelled and numbered. Right, 3D representation of the interaction plot between RBD and 2130-1-0114-112 LC. 2130-1-0114-112 show as stick (green) and RBD as gray sphere with the contact residues in red. Contact residues are labelled and numbered. (C) Glu112 and Lys440 with the EM map and distance between the sidechains. (D) Fab COV-2130 paratope and epitope residues involved in hydrogen bonding (dashed lines) and hydrophobic interactions. Hydrophobic interactions residues are shown as curved lines with rays. Atoms shown as circles, with oxygen red, carbon black, and nitrogen blue. Image was created with Ligplot+. (E) Representative micrograph.

Figure ED7:. CryoEM workflow of SARS-CoV-2 BA.2 spike bound to Fab.

At the bottom, Gold-standard Fourier shell correlation curves and maps colored by local resolution calculated using Relion, before and after local refinement.

Figure ED8:. Sequence logos.

The set of 376 designed IgG (A) includes mutation at 16 positions in the heavy chain (blue; mutations in green) and 9 positions in the light chain (magenta; mutations in pink). Height of each letter is proportional to the frequency of the amino acid in the group. This set of 376 sequences is divided into two overlapping sets, Set 1 (B; n=230) and Set 2 (C; n=204). From these two sets, a set of eight sequences (D) was selected for further evaluation. Selected sequences show reduction in mutations throughout the CDRH3 residues (103–118) mutated in (A), especially in residues 103–108. After eliminating sequences 2130-1-1231-194 (IH55W) and 2130-1-1231-199 (SL33Y), we produced the remaining eight sequences at larger scale and evaluated their thermostability and binding performance (Fig 2).

Figure ED8:. Sequence logos.

The set of 376 designed IgG (A) includes mutation at 16 positions in the heavy chain (blue; mutations in green) and 9 positions in the light chain (magenta; mutations in pink). Height of each letter is proportional to the frequency of the amino acid in the group. This set of 376 sequences is divided into two overlapping sets, Set 1 (B; n=230) and Set 2 (C; n=204). From these two sets, a set of eight sequences (D) was selected for further evaluation. Selected sequences show reduction in mutations throughout the CDRH3 residues (103–118) mutated in (A), especially in residues 103–108. After eliminating sequences 2130-1-1231-194 (IH55W) and 2130-1-1231-199 (SL33Y), we produced the remaining eight sequences at larger scale and evaluated their thermostability and binding performance (Fig 2).

Fig. ED9:. Bar plot representation of superpositions of the reference Omicron RBD structure (PDB ID 7t9k chain A) with 40 RBDs from WT, Omicron or Delta variants and comparison of experimentally solved RBD structures.

(A) The RMSD (Å) and LGA structure similarity scores (0–100) were calculated against the reference structure are provided in the right columns. Deviations of < 1Å are shown in green, 1–2 Å in yellow, 2–4 Å in orange, and > 4Å in red. The regions in RBD within the RBD-Fab interface where the major structural deviations between Omicron, Delta and WT are observed (positions 446 and 484) are marked at the top. (B) Structural clustering by LGA of 40 experimentally-solved RBDs. A red rectangle marks an identified centroid (an RBD from Omicron PDB ID 7t9k chain A) (C) A structure of WT RBD-COV2-2130 (PDB ID 7L7E, blue) superimposed with a model of Omicron RBD-COV2-2130 (Derived from PDB ID 7T9K, red). A significant deviation between two models is observed in the RBD-Fab interface in the region surrounding a mutation position WT to Omicron G446S (arrow).

Figure ED10.. The AbBERT model can be used to generate whole CDR replacements in COV2-2130.

For each sub-figure, 1,000 sequences were sampled by masking and replacing residues in loop regions: (A) CDRH1, (B) CDRH2, (C) CDRH3, (D) CDRL1, (E) CDRL2, and (F) CDRL3. Wild type COV2-2130 residues are in blue; mutations are in black. Letter heights correspond to frequency in the 1,000 sampled sequences. Because the residues are numbered in concatenated order, all residue numbers in the light chain are offset from their conventional values by +130, the length of the heavy chain. (G) Masking and replacing individual CDRH3 residues, executed separately for each position shown here, preserves considerably more structure and information about the COV2-2130 CDRH3 than complete masking and reconstitution of the entire CDRH3 as in (C).

Figure ED10.. The AbBERT model can be used to generate whole CDR replacements in COV2-2130.

For each sub-figure, 1,000 sequences were sampled by masking and replacing residues in loop regions: (A) CDRH1, (B) CDRH2, (C) CDRH3, (D) CDRL1, (E) CDRL2, and (F) CDRL3. Wild type COV2-2130 residues are in blue; mutations are in black. Letter heights correspond to frequency in the 1,000 sampled sequences. Because the residues are numbered in concatenated order, all residue numbers in the light chain are offset from their conventional values by +130, the length of the heavy chain. (G) Masking and replacing individual CDRH3 residues, executed separately for each position shown here, preserves considerably more structure and information about the COV2-2130 CDRH3 than complete masking and reconstitution of the entire CDRH3 as in (C).

Figure 1.. Overview of the GUIDE computationally driven drug engineering platform.

Given a parental antibody and target antigens, co-structures are estimated experimentally and/or computationally (left). Within the main computational loop (center left), a sequence generator proposes multi-point mutant antibody candidates, and a Bayesian optimization agent selects which proposed sequences to evaluate via a set of affinity prediction tools. A subset of 376 computationally evaluated sequences based on Pareto optimality, mutational distance, and sequence diversity were experimentally evaluated for binding affinity by Gyros or ELISA (center right). The top sequences are then evaluated for neutralization of SARS-CoV-2 variants (right). See Methods for details.

Figure 2.. Computationally designed IgG antibodies improve Omicron binding and maintain parental thermostability and binding to historical strains.

(A) The parental COV2-2130 (orange circles) and computationally designed antibodies (2130-1-0114-112, purple triangles; 2130-1-0104-024 blue diamonds; remainder in black) were assayed for thermal shift (n=3, technical replicates). Bars indicate the mean, and error bars indicate standard deviation. (B-E) The parental COV2-2130 antibody and computationally designed antibodies (represented by the same symbols as in A) and cross-reactive positive control antibody S309 (magenta squares) were analyzed for relative binding against four SARS-CoV-2 Spike-RBD variants in Gyrolab immunoassay: WA1/2020 (B), Delta (C), Omicron BA.1 (D) and Omicron BA.1.1 (E). Lines represent 4-parameter logistic regression model fit using GraphPad Prism to each titration, executed without technical replicates.

Figure 3:. Designed antibodies improve pseudoviral neutralization over COV2-2130.

The parental COV2-2130 antibody (orange circles), cross-reactive positive control antibody S2K146 (magenta squares), negative control antibody DENV-2D22 (gray x), and down-selected computationally designed antibodies were assayed by neutralization with lentiviruses pseudotyped with spike variants of WA1/2020 D614G (A), Omicron BA.1 (B), BA.1.1 (C), BA.2 (D), and BA.4 (E). (F) Symbols for each antibody are indicated in the legend. (G) IC50 values and 95% confidence intervals. “>” indicates a value > 10,000; NC indicates positive hill slope or failure to converge. Symbols indicate the mean and standard deviation of two technical replicates; curves are 4-parameter logistic regression models fit using GraphPad Prism.

Figure 4:. 2130-1-0114-112 is potent in focus reduction neutralization tests with authentic virus in Vero-TMPRSS2 cells.

2130-1-0114-112 potently neutralizes (A) WA1/2020 D614G (B) Delta B.1.617.2, (C) Omicron BA.1, (D) Omicron BA.1.1, (E) Omicron BA.2, (F) Omicron BA.2.12.1, (G) Omicron BA.4, (H) Omicron BA.5, and (I) Omicron BA.5.5 authentic viruses in focus reduction neutralization assays in Vero-TMPRSS2 cells. Symbols indicate the mean and standard deviation of two technical replicates; curves are 4-parameter logistic regression models fit of normalized data using GraphPad Prism. (J) IC50 values and 95% confidence intervals corresponding to (A)-(I). “>” indicates IC50 values > 10,000; “NC” indicates fits that were unconverged, unstable, or with positive hill slope. Analyses were performed in GraphPad Prism.

Figure 5:. 2130-1-0114-112 provides in vivo prophylactic protection against SARS-CoV-2 variants.

Eight-week-old female K18-hACE2 mice were administered 100 μg (~5 mg/kg) of the indicated mAb treatment by intraperitoneal injection one day before intranasal inoculation with 10⁴ FFU of WA1/2020 D614G (Left), Omicron BA.1.1 (Center) or BA.5 (Right). Tissues were collected four days after inoculation. Viral RNA levels in the lungs (Top), nasal turbinates (Center), and nasal washes (Bottom) were determined by RT-qPCR (lines indicate median of log₁₀ values); n = 9 (WA1/2020 D614G and BA.1.1 isotype control groups) or 10 (all others) mice per group, two experiments; Kruskal-Wallis ANOVA with Dunn’s multiple comparisons post-test; ns, not significant; *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001). All analyses conducted in GraphPad Prism.

Fig 6.. COV-2130 and 2130-1-0114-112 escape mapping using deep mutational scanning

(A-B) Comparison between IC50 values measured using deep mutational scanning for COV-2130 and 2130-1-0114-112 antibodies in (A) BA.1 and (B) BA.2 backgrounds with key mutations highlighted. Arbitrary units in both plots are on the same scale. Interactive plots that display each mutation can be found at https://dms-vep.github.io/SARS-CoV-2_Omicron_BA.1_spike_DMS_COV2-2130/compare_IC50s.html for BA.1 background and at https://dms-vep.github.io/SARS-CoV-2_Omicron_BA.2_spike_DMS_COV2-2130/compare_IC50s.html for BA.2 background. (C-D) Heatmaps of mutation escape scores at key sites for each antibody in (C) BA.1 and (D) BA.2 backgrounds. Escape scores are calculated relative to the wild-type amino acid in the same virus background. X marks wild-type amino acid in the relevant background. Amino acids not present in the deep mutational scanning libraries lack squares and gray squares are mutations that strongly impair spike-mediated infection. Mutations identified in (A) and (B) are shown with a heavy line surrounding the corresponding box. Interactive heatmaps for full spike can be found at for BA.1 background and https://dms-vep.github.io/SARS-CoV-2_Omicron_BA.2_spike_DMS_COV2-2130/COV2-2130_vs_2130-1-0114-112_escape.html for BA.2 background.

Figure 7:. Cryo-EM structure of neutralizing antibodies 2130-1-0114-112 in complex with Omicron BA.2 RBD.

(A) Cryo-EM map and model of the RBD-Fab complex. The map is transparent and colored by chain with RBD red, 2130-1-0114-112 HC yellow, and 2130-1-0114-112 LC green. (B) Atomic model of the RBD-Fab complex. Colors are the same as in A. Hydrogen bonds are shown as dashed lines. BA.2 RBD mutations are in orange. 2130-1-0114-112 mutation in cyan and blue (HC and LC). (C) Detail showing the 2130-1-0114-112 modified residues and the interaction with BA.2 RBD. Left, HCDR3 Glu112. Middle, LCDR1 Ala32 and Ala33 hydrophobic network. Right, LCDR2 Glu59 salt bridge with Arg498. Orange and green dashed lines indicate H-bonds and hydrophobic interactions, respectively; yellow dashed lines are labeled with distances. (D) Left, HCDR3 shown as in (C) with surface color by electrostatic potential, showing the positive and negative charges of Lys444 and Glu112. Right, A32 and A33 in LCDR1 with the nearby RBD surface colored by hydrophobicity (orange to cyan indicates hydrophobic to hydrophilic). (E) 2D diagram of Fab 2130-1-0114-112 paratope and epitope residues involved in hydrogen bonding (dashed lines) and hydrophobic interactions. Residues involved in hydrophobic interactions are shown as curved lines with rays. Atoms shown as circles, with oxygen red, carbon black, and nitrogen blue. Interacting residues that belong to CDR loops are colored in different shade. Asterisks correspond to mutated residues. Image created with Ligplot+.