Computational design of nanomolar-binding antibodies specific to multiple SARS-CoV-2 variants by engineering a specificity switch of antibody 80R using RosettaAntibodyDesign (RAbD) results in potential generalizable therapeutic antibodies for novel SARS-CoV-2 virus

Abstract

The human infectious disease COVID-19 caused by the SARS-CoV-2 virus has become a major threat to global public health. Developing a vaccine is the preferred prophylactic response to epidemics and pandemics. However, for individuals who have contracted the disease, the rapid design of antibodies that can target the SARS-CoV-2 virus fulfils a critical need. Further, discovering antibodies that bind multiple variants of SARS-CoV-2 can aid in the development of rapid antigen tests (RATs) which are critical for the identification and isolation of individuals currently carrying COVID-19. Here we provide a proof-of-concept study for the computational design of high-affinity antibodies that bind to multiple variants of the SARS-CoV-2 spike protein using RosettaAntibodyDesign (RAbD). Well characterized antibodies that bind with high affinity to the SARS-CoV-1 (but not SARS-CoV-2) spike protein were used as templates and re-designed to bind the SARS-CoV-2 spike protein with high affinity, resulting in a specificity switch. A panel of designed antibodies were experimentally validated. One design bound to a broad range of variants of concern including the Omicron, Delta, Wuhan, and South African spike protein variants.

Highlights

• •· Identification of well characterized antibodies that bind to SARS-CoV-1 but not to SARS-CoV-2.
• •The use of RosettaAntibodyDesign to design variants of the anti-SARS-CoV-1 antibodies bind to SARS-CoV-2 with high affinity.
• •Experimental validation of the antibodies designed in silico.

1. Introduction

COVID-19 remains a rapidly spreading global pandemic that is lethal for many individuals [[1], [2], [3]]. Reports indicate that the SARS-CoV-2 virus which causes COVID-19 is associated with pathologies such as coagulopathies of the lung and disseminated intravascular coagulation [[4], [5], [6], [7], [8], [9]]. There is a clear, global need for therapeutics to combat SARS-CoV-2. As with all viral diseases, a vaccine represents the most effective public health measure against the COVID-19 pandemic. However, given the scope and severity of the SARS-CoV-2 pandemic it is beneficial to have multiple alternate approaches. It is also important to develop strategies that can be rapidly deployed to design clinical interventions to combat future epidemics and pandemics. Breakthrough infections following vaccination, the emergence of new variants and vaccine hesitancy all contributed to COVID 19 cases and hospitalizations even after vaccination programs were initiated [[10], [11], [12]]. Unlike vaccines, which are preventive, antivirals and monoclonal antibodies provide a clinical strategy for individuals who have contracted the disease. Unfortunately, monoclonal antibodies that target one strain of SARS-CoV-2 often do not bind to new variants that emerge [[13], [14], [15], [16], [17]]. Similarly, antivirals are not always successful in neutralizing new emergent strains. For instance, studies suggest that emerging variants of SARS-CoV-2 show mutations in the protease targeted by nirmatrelvir (PAXLOVID), an anti-viral used currently, resulting in drug resistance [[18], [19], [20]]. Thus, a proven method for rapidly engineering (or re-engineering) and evaluating antibodies that bind to viral targets addresses an important unmet need.

One conventional approach has been to identify, isolate, and sequence antibodies from patients who have recovered from SARS-CoV-2 [21]. An alternative approach uses humanized mice to produce antibodies that neutralize viral entry [22]. A third approach uses in vitro display methods, such as phage or yeast display, to isolate antibodies that bind to SARS-CoV-2 [23]. The methods require large-scale screening of libraries, expensive animal models, and long development timelines. Importantly these approaches are dependent on a biological system that is essentially a “black box” and thus do not allow for engineering of the antibody sequence to target new variants as they appear. Moreover, antibodies developed in animals (even when humanized) can elicit immune responses in patients that significantly affect their safety and efficacy [24].

Here we present an alternate approach using a computational method for the design of antibodies that bind to SARS-CoV-2 targets with high affinity followed by experimental validation. Our approach is based on using well characterized antibodies to related viruses as a template and reengineering these to target a novel virus. Our strategy is based on studies [[25], [26], [27], [28], [29]] demonstrating that receptor-binding domains of the spike protein of the virus surface glycoprotein for both SARS-CoV-2 and SARS-CoV-1 are similar in sequence and structure with few amino acid differences. Despite the similarity of the receptor binding domains (RBDs); antibodies, that bind SARS-CoV-1, do not bind to SARS-CoV-2 [30]. For example, monoclonal antibodies m396 [31], 80R [32], and CR3022 [25], which bind with high affinity to SARS-CoV-1, have been shown to either not bind to SARS-CoV-2 or to have a much lower binding affinity [22,33]. For this proof-of-concept study we selected an antibody (80R) that binds with nano-molar affinity to the spike protein of SARS-CoV-1 but exhibits no measurable binding to the spike protein of the SARS-CoV-2 virus.

Our specificity switch workflow involves: (i) identification of antibody-antigen co-crystal structures which share structural and sequence homology between target antigen (SARS-CoV-2) and the antigen present in the co-crystal structure (SARS-CoV-1). (ii) The use of RosettaAntibodyDesign (RAbD) [[34], [35], [36]] to design variants of the co-crystal structure antibody that bind with high affinity and specificity to SARS-CoV-2. (iii) Experimental characterization of designed antibody candidates consisting of expression, purification, and affinity measurement assays that can evaluate antibodies for binding to the SARS-CoV-2 virus with high affinity. Here we demonstrate the success of this approach in engineering antibodies that bind with high affinity and specificity to the wild type spike protein of SARS-CoV-2 as well as several variants that have emerged.

2. Results

2.1. Computational antibody designs using RabD

The computational workflow for designing antibodies with improved binding to the spike protein of SARS-CoV-2 is illustrated in Fig. 1A and described in the Materials & Methods section. The RAbD protocol computed 4000 designed antibody structures and sequences using the antibody 80R structure aligned to SARS-CoV-2 rather than SARS-CoV-1 as an input structure. The RAbD protocol implemented several different score functions (Fig. 1A) and design options to generate a variety of designs. Antibody 80R was selected as a template because it satisfied the following criteria: (i) a crystal structure is available for the 80R/SARS-CoV-1 antibody-antigen complex, (ii) 80R has been shown to bind to the related SARS-CoV-1 virus, and (iii) unlike some other anti-SARS-CoV-1 antibodies that bind to SARS-CoV-2 with lower affinity, 80R shows no measurable binding to SARS-CoV-2 [29,37]. The structure of the 80R/SARS-CoV-1 complex and the amino acid residues associated with binding to the spike protein of SARS-CoV-1 are depicted in Fig. 1B.

Fig. 1.

Computation design strategy A) the computational workflow for designing antibodies with improved binding to the spike protein of SARS-CoV-2 is illustrated. B) The complex structure of 80R antibody (grey) and spike protein of SARS-CoV-1 (green) is shown along with the amino acid residues associated with interface binding (magenta on 80R). (For interpretation of the references to color in this figure legend, the reader is referred to the Web version of this article.)

The 80R antibody was reengineered using various design strategies described in the Materials & Methods section. The primary filter for narrowing down which designs to manually inspect for structural analysis was the predicted binding free energy (dG_separated) score calculated in Rosetta. This score predicts whether the mutations are likely to increase affinity according to the Rosetta energy function. In addition to using predicted binding free energy to rank order the different designs of 80R, we considered the total energy (total_score) for the best-scoring design models and identified top-scoring individual mutations that occurred frequently amongst design runs while having a favorable interface energy for the individual residue. Fig. S1 shows all the computational metrics provided by the RosettaInterfaceAnalyzer for the top 80% total score designs and their distribution for each metric. The predicted binding free energy vs total score was used as the final ranking metric used for design selection. Fig. S2 provides a sequence logo for the top 80% Complementarity-determining regions (CDR) H3 mutations showing the diversity of the mutations in the CDR H3 and identifies those that occur frequently (e.g., R110 N and S111A), which were considered during the manual inspection. These mutations were incorporated into a subset of the designed sequences regardless of strategy, while wildtype positions were retained in the other subset of designs in order to hedge against the chance that these mutations would disrupt binding at the interface. Also important to note that position R110 also did have serine appear in the top 80% of total_score ranked sequences, but this mutation was not included because of its lower per-residue energy. Fig. S3 shows the predicted interface energies (y-axis) for each residue (x-axis) across various design runs (columns 1–3), colored by amino acid, and parsed out by CDR type (columns 1–6, CDR1-3 on the light and heavy chains). This plot summarizes the energy landscape of the predicted binding free energy for each residue at each CDR position coming out of all designs. This plot was used to identify residues that frequently occur regardless of design run, and have favorable predicted interface energy (less than 0). Fig. S4 shows the predicted binding energy versus total energy for single point mutation models of each of the identified point mutations from Fig. S3, as well as for combinations of single mutations modeled into the wildtype antibody structure for both the heavy and light chains of antibody 80R.

Following the design process, we manually inspected the mutations introduced in the RAbD-based designs to determine if any favorable interactions were introduced at the antibody-antigen interface. Fig. 2A depicts the model of the 80R antibody in complex with SARS-CoV-2 spike protein with specific CDRs. The RAbD complex input model for designing the variants is shown in Fig. 2B. In the manual evaluation of the antibody variants, consideration was given to factors that could enhance binding affinity such as additional hydrogen bonds, filling hydrophobic pockets and increased contact between the two proteins. Interactions between variant 80R_5 of the antibody 80R and the spike protein of SARS-CoV-2 that enhance parameters associated with increased binding affinity are depicted in Fig. 2C and D. Here we visualize all of the hydrophobic residues in the interface of the 80R_5 design and the SARS-CoV-2 RBD spike protein and additional favorable pi-stacking (Y110 of CDR L3 to Y449 of SARS-CoV-2 RBD) interactions in the interface. Thus, while the initial part of the design process involved rank ordering mutants based on the interface and total scores, a subset of 7 designs (80R 1–2, 80R 13–17) were selected based on a combination of manual inspection of the interactions between the antibody and its target in homology models. The remaining designs were selected by choosing top 8 designs coming from each of the three separate runs of RAbD, with or without the two mutations (R110 N and S111A) in CDR H3. The final list of the antibody sequences selected (Table S1), and additional design information is provided in the Supplementary information.

Fig. 2.

Input model for design using the 80R antibody and wild-type RBD and examples of mutations chosen. A) Model of 80R antibody (heavy chain in green and light chain in cyan) in complex with SARS-CoV-2 spike protein (shown in grey) with specific CDRs. B) RAbD complex input with binding interface depicted in magenta C) 80R_5 design model bound to the wild-type RBD of the SARS-CoV-2 spike protein. Mutations in the 80R_5 design are shown as yellow sticks and contact residues on the RBD are shown in magenta. D) Show examples of mutations on both the heavy chain and light chain on the 80R antibody with hydrophobic residues shown in purple and pi-stacking type of interactions shown with Tyrosine residue (Y110 on antibody). (For interpretation of the references to color in this figure legend, the reader is referred to the Web version of this article.)

3.2. Expression and purification of antibody designs

A total of 30 antibody variants based on 80R, were expressed in mammalian HD 293F cells and purified with 85–95% purity.

2.3. Binding of computationally designed antibodies to SARS-CoV-2 wild-type spike protein and variants

Antibody-antigen binding kinetics for the original template and engineered antibody variants were measured using Biolayer Interferometry (BLI). A total of 30 antibody designs were evaluated for binding to the full-length trimer spike protein as well as to the spike variants and a typical binding curve, for 80R_5 is shown in Fig. 3A). The K_D values for those 80R variants which exhibited measurable binding-affinities for the spike protein of SARS-CoV-2 are depicted in Table 1, a sequence alignment is provided in Fig. S5, and further representative binding curves shown in Fig. S6. As a positive control we have included the B38 antibody which was isolated from a convalescent COVID-19 patient and demonstrated to bind to the spike protein and neutralize the SARS-CoV-2 virus [38]. Consistent with previous reports, the 80R antibody showed no measurable binding to the wild-type RBD. However, 10 of the 80R variants re-engineered by us exhibited binding to the wild-type full-length trimeric spike protein with a K_D values ranging from 0.468 nM to 114 nM. The positive control full length B38 antibody bound to the full length spike trimer protein of SARS-CoV-2 with a K_D of <0.0001 and it was reported to neutralize the SARS-CoV-2 virus [38]. Moreover, the most interesting of the 80R variants designed by us (80R_5), bound to the wild type (K_D, 0.468 nM) and all the full-length spike protein variants tested. The variants tested included Wuhan S6P (K_D, <0.001 nM), Delta (K_D, 0.441 nM), South African S6P (K_D, 8.32 nM), and Omicron (K_D, 6.12 nM) (Table 1). We have also determined that 80R_5 binds to the Omicron variant BA.5 (K_D, 1.12 nM). In contrast to the antibody 80R_5, designed by us, which shows broad binding to diverse SARS-CoV-2 variants, monoclonal antibodies that are currently used clinically do not cross-react with the new SARS-CoV-2 variants that have emerged [17]. To exclude the possibility of non-specific binding, we used the MERS Spike trimer protein as a negative control. The antibody 80R_5 exhibited no measurable binding to the MERS Spike trimer protein (Fig. S7).

Fig. 3.

Binding curves to SARS-CoV-2 full length spike protein variants for the 80R antibody designs. A) BLI binding curve between SARS-CoV-2 wild-type spike full-length trimer (10 nM) and 80R_5 antibody (10 nM). B-D) ELISA assay showing binding of B38, 80R and 80R_5 to spike protein of wild type SARS-CoV-2 and the beta, delta, South African and Omicron variants. The plates were coated with SARS-CoV-2 spike protein of wild type (Blue), beta (red), delta (green), Omicron (beige) and South African (orange)variants and incubated with increasing concentrations (15.6 ng/mL to 1000 ng/mL) of B38 (B), 80R (D) and 80R_5 (D) antibodies. (For interpretation of the references to color in this figure legend, the reader is referred to the Web version of this article.)

Table 1.

Biolayer interferometry binding affinities (K_D) for B38 and selected 80R designs to the SARS-CoV-2 spike trimer of wild-type, Wuhan S6P, Delta, South African S6P, and Omicron. N: no binding. This table contains only designs that bind to at least one full length spike variant.

	SARS-CoV-2 spike trimer
Antibody	Wild-type (KD nM)	Wuhan S6P (KD nM)	Delta (KD nM)	South African S6P (KD nM)	Omicron (KD nM)
B38	<0.001	3.68	0.16	N	N
80_wt	N	N	N	N	N
80R_5	0.468	<0.001	0.443	8.32	6.12
80R_6	4.16	N	N	N	N
80R_7	2.85	N	N	N	N
80R_8	114	N	N	N	N
80R_10	1.82	N	N	N	N
80R_18	2.68	N	N	N	7.99
80R_19	0.845	N	N	0.024	N
80R_21	3.1	N	N	N	N
80R_22	1.39	N	N	N	N
80R_23	0.858	N	0.625	0.001	6.35

Comparison of the computationally designed 80R_5 antibody with the B38 antibody identified from a convalescent COVID-19 patient.

We evaluated the binding of the positive control antibody, B38 and the engineered 80R_5 to the wild type spike protein and variants using an ELISA assay (Fig. 3B–D). Both B38 and 80R_5 showed binding to the wild type and Delta variants of SARS-CoV-2. However, B38 showed no measurable binding to the Wuhan S6P, South African S6P and Omicron variants. In contrast, the computationally designed 80R_5 antibody binds to all variants of the spike protein (albeit with varying affinities).

3.4. Computational analysis of antibody binding sites on the spike protein of the SARS-CoV-2 variants

To further understand the results of the binding assays, we generated models (using RosettaFastRelax with constraints and no constraints) of our designed antibody (80R_5) to the full-length spike proteins of the Delta, South African and Omicron variants (Fig. 4, energy parameters shown in Table S1).

Fig. 4.

(A) Predicted RosettaFastRelax models of 80R_5 designs bound to the omicron SARS-CoV-2 spike protein, Delta, and South African S6P, with antibody and RBD interface residues on the antibody shown in yellow. B-D) Omicron model superimposed onto wild-type RBD model with mutations on the RBD shown in purple and interacting residues on the antibody shown in orange. E-G) South African model superimposed onto wild-type RBD model with mutations on the RBD shown in purple and interacting residues on the antibody shown in orange. H and I) Delta variant model superimposed onto wild-type RBD model with mutations on the RBD shown in purple and interacting residues on the antibody shown in orange. (For interpretation of the references to color in this figure legend, the reader is referred to the Web version of this article.)

For the Omicron variant antibody bound models, we observed in the no constraints models that compared to wild-type RBD protein, the N440K mutation in the Omicron variant breaks a bidente hydrogen bond to the Y67 on the Heavy chain of the 80R antibody (Fig. 4B). The mutation Q493R on the Omicron variant also creates a potential spatial clash with neighboring residues T39, D58, A68, and I69 residues on the 80R_5 antibody (Fig. 4C). In addition, Q493R on the Omicron variant exchanges a sidechain hydrogen bond with T39 on the L chain of the 80R_5 antibody for a bidente backbone hydrogen bond with D58, while also forming an intra-chain backbone bond with F490 (Fig. 4C). Lastly, G496S and N501Y on the Omicron variant form an intra-chain sidechain hydrogen bond, replacing the backbone bond that is in the wild-type RBD (Fig. 4D). Per-residue interaction energies demonstrate differences in residue energy (Figs. S5–S7, for 80R_5 antibody bound to Omicron, South African S6P, and delta RBD, Table S1). Overall, the mutations on the omicron RBD spike protein changes residue interactions with the 80R_5 antibody.

For the South African variant, in the unconstrained model, we observed that K417 N forms new intramolecular bonds with E406 and Q409 (Fig. 4E). This doesn't appear to be close enough to affect antibody affinity significantly. E484K maintains the intermolecular bond with S83 of the light chain but breaks the intramolecular bond with the backbone of F490 (Fig. 4F). Lastly, N501Y is in the middle of the interface. Intramolecular bonds to the backbones of G496 and Y505 are exchanged for bonds with Q498 and the backbone of G496 (Fig. 4G). The phenyl ring is sterically larger, leading to a number of small-distance shifts in the surrounding residues.

For the Delta variant, we observed that L452R forms an intramolecular bond with Y351 which does not appear to be close enough to affect antibody affinity significantly (Fig. 4H). In addition, T478K does not add any additional interactions besides changing the charge of that position and having the positively charged lysine pointing outwards (Fig. 4I). Furthermore, total per-residue interaction energies between the wild-type RBD and 80R_5 complex models in both the constraint and free of constraint RosettaFastRelax models are provided in Figs. S8–S10.

3. Discussion

In the introduction we have provided a rationale for why recombinant antibodies to treat SARS-CoV-2 infections are a potentially advantageous class of drugs for the treatment of COVID-19 patients. We also delineate the drawbacks to the current approaches for identifying monoclonal antibody candidates that bind to SARS-CoV-2 proteins with high affinity.

The novel approach described here leverages the RAbD computational method to rationally design antibodies that bind with high affinity to SARS-CoV-2 spike protein. The most advantageous aspect of this workflow, which sets it apart from current methods is that it can be rapidly deployed to design antibodies that target new strains of the virus as soon as they appear.

Computational redesign of antibodies has obvious advantages with respect to speed and cost. Our approach only requires: (i) suitable template antibody-antigen structures with minimal homology to the design target as a starting point for computation design, and (ii) a structure of the target antigen. In this study we implemented our workflow to design antibodies that could target SARS-CoV-2. We used the 80R that was initially identified to bind and neutralize the SARS-CoV-1 spike protein during the 2002 outbreak [39]. The structures of 80R binding to the SARS-CoV-1 are available and while the antibody binds with high affinity to SARS-CoV-1, there is no measurable binding to the virus of current interest, SARS-CoV-2. Thus, 80R antibody provides a better demonstration of the proof-of-concept than other anti-SARS-CoV-1 antibodies that bind to SARS-CoV-2 but with a lower affinity.

The antibody used as a template, 80R, was subjected to computational redesign using RabD with different input settings that together provide changes diverse enough to increase antibody binding affinity to the spike protein of SARS-CoV-2. Following the workflow described in Fig. 1A, we generated 30 design variants of 80R that were likely to be good candidates for in vitro evaluation.

Of the 30 redesigned variants that were experimentally validated 10 variants showed measurable binding (K_D 0.4 nM to 114 nM) to the full-length spike protein of the SARS-CoV-2 virus. Three variants, 80R_5, 80R_19, and 80R_23, bound to the wild-type full-length spike protein with K_Ds <1 nM. Thus, the mutations introduced in the 80R antibody by our computational workflow resulted in successfully increasing the affinity to the spike protein in a third of the mutants. As the parental template antibody, 80R, does not bind to either the RBD or spike domain of SARS-CoV-2, the successful computational reengineering of the antibody, 80R provides a strong demonstration of the power of the RAbD driven redesign.

It is important to emphasize that the computational workflow has no direct estimation of affinity. Rather, rank-scoring of antibody mutants is based on altered biophysical characteristics that serve as surrogate markers for affinity. Nonetheless, the workflow allows exploring a very large search space and then quantitatively rank ordering for optimal binding affinities. Thus, we can move a substantively reduced number of antibody designs to be tested in an in vitro experimental system. We expect that when we transition to experimental determination of binding affinities, some variants will result in improved binding affinity while others will not. The unsuccessful variants could also be useful in subsequent iterations of the computational re-engineering of the antibody and in improving the methodology. The power of the computational workflow is evidenced by the fact that using as a template an antibody that exhibited no measurable binding to the spike protein of SARS-CoV-2, we successfully designed a variant with sub-nanomolar binding affinity (80R_5, K_D of 0.46 nM). In all we identified 10 antibody variants that bound to the spike protein of SARS-CoV-2 (Table 1).

As alluded to above, novel variants of the virus that arise spontaneously add to the challenge of treating patients with monoclonal antibody therapies. The new variants of the SARS-CoV-2 often cannot be treated with existing therapies. For instance, a recent study demonstrated that the two therapeutic antibody cocktails that are authorized for emergency use in the United States did not neutralize the Omicron variant [17]. This report clearly emphasizes the real-world need for the rapid discovery and development of antibodies that bind with high affinity to emerging variants of SARS-CoV-2. Such antibodies can then be used to develop/select for antibodies that can neutralize the SARS-CoV-2. The large number of putative anti-SARS-CoV-2 designed using our computational approach offer a ready source of antibodies that can be tested against emerging viral variants.

We tested the binding affinity of the computationally designed antibody variants to the Wuhan S6P, Delta (B.1.617.2), Beta (South African S6P, B.1.351), and Omicron (B.1.1.529 and BA.5) strains of SARS-CoV-2. We found that of the 30 antibody variants experimentally characterized by us the 80R_5 design was the most interesting. This design showed binding to all the variants tested including wild type, Wuhan S6P, Delta, Beta, and Omicron with K_Ds of 0.4 nM–8 nM. The 80R_5 antibody is thus a potential generalizable antibody that can bind to current variants of SARS-CoV-2.

The different antibody designs exhibit diversity in their capacity to bind to the SARS-CoV-2 variants. Thus, the antibody 80R_5 binds with high affinity to all variants tested, the 80R_18 antibody binds to the wild type and Omicron variants, and the 80R_19 antibody binds to the wild type and South African variants, while the antibody 80R_23 binds to the wild type and Delta variants. Several other antibody designs bind only to the wild type strain of SARS-CoV-2. Thus, these antibodies have potential research and translational applications in distinguishing between the different SARS-CoV-2 strains.

Currently monoclonal antibodies that target a particular virus are identified from among those generated naturally by patients. Once such well characterized monoclonal antibody derived from a convalescent patient is B38 [38]. Using an ELISA assays we compared the naturally occurring B38 with the computationally designed R80_5 (Fig. 3). Both B38 and 80R_5 showed binding to the wild type and Delta variants of SARS-CoV-2. Although B38 has higher affinity for the spike protein of SARS-CoV-2, the 80R_5 variant binds to all variants tested (Wuhan, Delta, South African and Omicron). B38 on the other hand binds only to the wild type and Delta variant. Additionally, the K_Ds for the binding of 80R_5 to the spike protein is well within the range for antibodies that neutralize the SARS-CoV-2 virus.

The modeled structures of 80R_5 with the Omicron, South African, and Delta RBD variants demonstrate that most of the mutations either replace existing protein-protein interactions with new protein-protein interactions such as backbone hydrogen bonds or changing charges. Most differences were observed in the Omicron and South African variants, since these have a larger number of mutations in the spike protein. However, these particular amino acid changes on the RBD interface were not sufficient to eliminate the binding to the 80R_5 antibody but did decrease the K_Dvalues to ∼6–8 nM. In addition, we observed that the wild type RBD and 80_5 antibody interface rosetta score was slightly more favorable compared to the Rosetta score for the omicron RBD and 80R_5 complex. This is interesting because the initial design approach was to design favorable interactions to the wild-type RBD.

Taken together our findings indicate that RAbD driven rational design of antibodies offers a rapid and effective method for generating novel antibodies which can bind to targets with high affinity. Importantly, this approach could allow the management of emerging strains of the virus. Such a workflow could find broad applications in the rapid development of therapeutics to emerging pathogens responsible for pandemics. We would like to emphasize that the finding that the recombinant antibody design 80R_5 binds to multiple SARS-CoV-2 variants is fortuitous. The advantage of our workflow is the ability to rapidly design antibodies that bind to the spike protein of new variants as they arise.

A drawback of our approach is that the computational workflow allows the design of antibodies with increased affinity for the target. This need not necessarily translate to increased neutralization. This limitation is further seen when we used the same design approach on two additional antibodies initially identified during the SARS-CoV-1 outbreak, CR3022 and m396. The CR3022 designs did not show an increased binding capacity to SARS-CoV-2, while m396 did not show any binding capabilities to SARS-CoV-2 spike protein (data not shown). Indicating that this computational workflow would need to be implemented with many antibodies through high-throughput protein expression or display technologies to increase the likelihood of success.

We have provided a proof-of-concept for the rational design of antibodies that can target biologically and clinically important proteins on emerging pathogens. The computational workflow if combined with high throughput antibody synthesis and screening offers a viable strategy for deployment against emerging viral pathogens. While antibodies produced by recombinant DNA technology have been some of the most successful therapeutics in the last decade, the results have not been as promising for treating existing and emerging viral infections. The workflow provided here could be foundational to developing such therapies. Additionally, well characterized antibodies that bind to spike proteins from different strains of the SARS-CoV-2 are also useful reagents and can potentially be leveraged to develop diagnostic tools.

4. Material and methods

4.1. Preparing input files to use for the RosettaAntibodyDesign (RAbD) program

We used a co-crystal structure of the 80R single-chain Fv bound to the SARS-CoV-1 RBD (PDB 2GHW) as a starting point for design. We separately energy minimized the AHo-renumbered [40] antibody Fv region and the SARS-CoV-2 RBD (PDB 6M0J) into the Rosetta energy function (beta_nov16) [41,42] using CDRDihedralConstraintMover [34] for the CDR regions and dihedral coordinate constraints for the antibody region (please refer to the “FastRelax files for 80R″ in Computational_Antibody_Design_files.zip). We then brought both structures into contact (using PyMOL [43]) by aligning the RBD domain of the refined 6M0J structure to the RBD domain of the 2GHW co-crystal structure to create an input complex structure. Visually the epitope regions of the SARS-CoV-2 and SARS-CoV-1 RBD domains align well, lending more confidence to this input structure as a design starting point.

To maintain diversity in the input complex structure, we used three different approaches to further minimize this input structure into the Rosetta energy function:

• 1.We used the Rosetta FastRelax protocol [44] and the beta_nov16 score function [45].
• 2.We used Rosetta's gradient-based minimization [46] (moving a structure to its nearest local energy minimum) and the beta_nov16 score function [45].
• 3.We used the Rosetta FastRelax protocol [44] and the current default score function REF15 [41,42].

In Rosetta, the energy of proteins is calculated by scoring it with an optimized energy function (score function called REF15) which is a weighted sum of different energy terms. Some of the terms represent physical forces such as electrostatics and Lennard-Jones potential for van der Waals interactions, while others represent probabilities such as Ramachandran space and backbone and side-chain torsional preferences [41]. In Brief, the beta_nov16 score function is an updated version of REF15 [41] with additional score terms to consider coordinated water molecules specifically at interfaces. This score function is not the default in Rosetta at this point but we reasoned that consideration bridging water molecules may be very relevant to our design objective and therefore chose to include both the default REF15 and the newer beta_no16 score functions in our design strategy [45].

For all FastRelax refinement runs, we produced a total of 100 refined structures and selected the best input complex structure for RAbD based on the lowest total_score (total energy of the system after sampling and optimizing the conformations of the backbone and side chains of the complex) and dG_separated score (the change in Rosetta energy of 80R Fv and the RBD separated vs when they are complexed, i.e. the estimated binding energy). Rosetta XML and flags files used for the relax protocol are located in the “Computational_Antibody_Design_files.zip”.

4.2. RAbD protocol allows for various design strategies

After identifying the best energy minimized input complex structure (as described above), we used that input pdb file for starting the RAbD design protocol. We generated a total of 4000 designs based on antibody 80R using two different RAbD design approaches. Each design approach used different RAbD design options which are detailed in the flags files in the Supplementary Information. Briefly, we set up design runs using either the REF15 default score function or the newer beta_nov16 score function, and ran these options with either one of the relaxed complex structures, or the Cartesian minimized complex structure as input. We generally allowed sequence design for the light chain CDR loops and the light chain DE loop, as well as heavy chain CDR loops 1 and 2. We further undertook a designated CDR H3 design run, where we allowed sequence design for the H3 loop, where the mutations were sampled from a manually curated set of CDR loops [34] (the PyIgClassify database of CDR sequences is already added in Rosetta and part of the RAbD design protocol) with the same canonical conformation that was identified for antibody 80R, by providing a resfile to Rosetta. By using this database (or a custom resfile in the case of CDR H3 design), we can limit design to only consider amino acid identities observed in experimentally solved antibody structures with CDR loops of the same conformation as the design target. These two design approaches with RAbD resulted in 4000 designs (1000 designs for each design strategy and starting input combination).

To select the most promising designs, we first filtered the designs by predicted total energy (total_score). Specifically, we discarded the 20% highest scoring (least favorable) designs to triage designs that are likely to suffer from stability problems. We next ranked the remaining designs based on the sum of the Z-scores for predicted binging energy (dG_separated) and total energy (total_score). Variants 80R 3–12 were selected based on this metric (Fig. S1) in addition to manually identifying favorable interactions (hydrophobic pockets, hydrogen bonding networks, electrostatic interactions) by visual inspection in PyMOL [47].

Designs 80R 18–30 all share two mutations in CDR H3 (R110 N and S111A, Fig. S2) which were observed in the majority of designs stemming from the CDR H3 design approach. We further manually combined CDR H3 mutations that were predicted to be beneficial based on a per-residue analysis (Fig. S3 described below) with the RAbD design outputs to finalize the 80R 18–30 designs. All 80R design sequence are shown in Table S1.

4.3. Description of the per-residue analysis

In addition to the systematic ranking of designs by dG_separated score (predicted binding energy) and total_score (predicted total energy), we also designed a set of variants (80R 1 + 2, 80R 13–17) with mutations selected on the basis of per-residue interface energy, where the most negative per-residue energies represent mutations that favorably contribute to an improved interface binding energy (negative difference between the energy of the relaxed complex and the sum of the energies of the separated components after relaxation for each approach described above) the most according to Rosetta. To select mutations for both the light chain and heavy chains, we identified mutations that had very favorable (more negative) per-residue binding energy across various design runs with differing structure preparation and/or Rosetta energy function selection (beta_nov16 vs REF15). To select mutations of interest we performed two tests, we performed a single mutation analysis; where for each mutated residue we generated 100 relaxed structures of the single mutant and the wild-type antibodies parent clone and compared scores (dG_separated and total_score). Mutations that scored lower than the wild type were then combined into combination mutants containing 4 to 7 mutations, relaxed, and compared to the wild type. (Fig. S4).

4.4. Antibody production and purification

Full length IgG antibodies were codon optimized, ordered, expressed, and purified. Designed heavy chain (HC/VH) and light chain (LC/BL) was used with default CH/CL (human IgG1 and human Ig kappa). The recombinant plasmids (pcDNA3.4 as the expression vector) encoding designed antibodies were transiently co-transfected into suspension mammalian HD 293F cells. All proteins expressed in high purity, purified by affinity purification column, and buffer exchanged into PBS buffer pH 7.2. The purified protein was analyzed by SDS-PAGE analysis to determine the molecular weight and purity which ranged between 85 and 95%.

4.5. Biolayer interferometry (BLI)

Recombinant SARS-CoV-2 RBD-His Tag protein (Genscript Cat. No. Z03483) was loaded onto Anti-Penta-HIS (HIS1K) biosensors at 20 nM concentrations, designed antibodies were added at various concentrations (20 nM and 10 nM) for real time association and dissociation analysis using the ForteBio Octet RED96e system and software. Recombinant wild-type SARS-CoV-2 full length spike trimer protein (Gift from Dr. Chris Bahl), Wuhan S6P full length trimer spike variant (Gift from Dr. Jesper Pallesen), South African S6P full length trimer spike variant (Gift from Dr. Jesper Pallesen), Delta full length trimer spike variant (Sino Biological Cat No. 40589-V08H10), Omicron B.1.1.529 full length trimer spike variant (Sino Biological Cat No. 40589-V08H26), Omicrom BA.5 full length trimer spike variant (ACROBiosystems Cat No. SPN-C522e) and MERS full length trimer spike variant (ACROBiosystems Cat No. SPN-M52H4) were loaded onto Anti-Human IgG-Fc capture (AHC) biosensors and real time association and dissociation experiments were conducted using the ForteBio Octet RED96e system and software. All experiments used 96-well, black, flat bottom, polypropylene microplates.

4.6. ELISA

Ninety-six well-ELISA plates (Catalog # 3455, ThermoFisher Scientific, Waltham, MA) were coated with 100 μL of SARS-CoV-2 spike protein of wild type, beta, delta, Omicron and South African variants in 1X Carbonate-Bicarbonate buffer (Catalog #C3041, Sigma) at a concentration of 1 μg/mL. Plates were incubated at 4 °C overnight and then washed three times with TBST (Catalog #T9039, Sigma). Plates were blocked using 200 μL per well of 1% bovine serum albumin (Catalog # 37520, ThermoFisher Scientific, Waltham, MA) in TBS buffer, incubated for 1 h at RT, and then washed three times with TBST. Antibody B38, 80R and 80R_5 were diluted in blocking buffer, and 100 μL of the diluted antibody was added to each of the experimental wells in triplicate. The plates were incubated for 1 h at RT and then washed three times with TBST. Goat anti human IgG Fc antibody conjugated to horseradish peroxidase (HRP) (Catalog # NB7448, Novus, Centennial, CO) was diluted 1:100,000 in blocking buffer and added to the wells to detect the binding to SARS-CoV-2 proteins. The plates were incubated for 1 h at RT and then washed three times with TBST. Substrate solution (Catalog #N301, ThermoFisher Scientific, Waltham, MA) was added, and the optical density (OD450 nm) was determined after 15 min by stopping the reaction with 1 N sulfuric acid-based stop solution (Catalog # A300, ThermoFisher Scientific, Waltham, MA). A blank without antibody was included in triplicates on each plate, and these values were subtracted from the values obtained in the presence of the antibody samples.

In Brief

COVID-19 is the greatest public health crisis of our time. Monoclonal antibody therapies significantly reduce COVID-19 hospitalization rates and mortality. Here Hernandez et al. provide a proof-of-concept study to design high affinity antibodies that bind to multiple variants of the SARS-CoV-2 spike protein using a computational approach. The workflow provided here could help to rapidly develop therapies for treating existing or emerging viral infections.

Disclaimer

This article reflects the views of the authors and should not be construed to represent FDA's views or policies.

Author contribution statement

Nancy E. Hernandez, Wojciech Jankowski, Rahel Frick, Simon P. Kelow: Conceived and designed the experiments; Performed the experiments; Analyzed and interpreted the data; Contributed reagents, materials, analysis tools or data; Wrote the paper.

Joseph H. Lubin: Performed the experiments; Analyzed and interpreted the data; Wrote the paper.

Vijaya Simhadri: Performed the experiments; Analyzed and interpreted the data; Contributed reagents, materials, analysis tools or data.

Jared Adolf-Bryfogle: Analyzed and interpreted the data; Contributed reagents, materials, analysis tools or data; Wrote the paper.

Sagar D. Khare: Analyzed and interpreted the data; Contributed reagents, materials, analysis tools or data.

Roland L. Dunbrack Jr, Jeffrey J. Gray, Zuben E. Sauna: Conceived and designed the experiments; Analyzed and interpreted the data; Wrote the paper.

Funding statement

Roland Dunbrack Jr was supported by National Institute of General Medical Sciences {R35 GM122517}.

Data availability statement

Data will be made available on request.

Declaration of interest’s statement

The authors declare the following conflict of interests: N.E.H, W.J, R.F, S.P.K, R.L.D, J.J.G, and Z.E.S. are co-inventors on a patent application related to this manuscript that was filed by the FDA. The remaining authors declare no competing financial interests.

Appendix A. Supplementary data

The following is the Supplementary data to this article: