Nanobody screening and machine learning guided identification of cross-variant anti-SARS-CoV-2 neutralizing heavy-chain only antibodies

Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV-2) continues to persist, demonstrating the risks posed by emerging infectious diseases to national security, public health, and the economy. Development of new vaccines and antibodies for emerging viral threats requires substantial resources and time, and traditional development platforms for vaccines and antibodies are often too slow to combat continuously evolving immunological escape variants, reducing their efficacy over time. Previously, we designed a next-generation synthetic humanized nanobody (Nb) phage display library and demonstrated that this library could be used to rapidly identify highly specific and potent neutralizing heavy chain-only antibodies (HCAbs) with prophylactic and therapeutic efficacy in vivo against the original SARS-CoV-2. In this study, we used a combination of high throughput screening and machine learning (ML) models to identify HCAbs with potent efficacy against SARS-CoV-2 viral variants of interest (VOIs) and concern (VOCs). To start, we screened our highly diverse Nb phage display library against several pre-Omicron VOI and VOC receptor binding domains (RBDs) to identify panels of cross-reactive HCAbs. Using HCAb affinity for SARS-CoV-2 VOI and VOCs (pre-Omicron variants) and model features from other published data, we were able to develop a ML model that successfully identified HCAbs with efficacy against Omicron variants, independent of our experimental biopanning workflow. This biopanning informed ML approach reduced the experimental screening burden by 78% to 90% for the Omicron BA.5 and Omicron BA.1 variants, respectively. The combined approach can be applied to other emerging viruses with pandemic potential to rapidly identify effective therapeutic antibodies against emerging variants.

We leveraged a robust experimental pipeline for heavy chain-only (HCAb) library screening to identify 59 potent HCAbs that can cross-neutralize different SARS-CoV-2 variants. Several of these HCAbs with efficacy against different variants also bind to different SARS-CoV-2 epitopes, suggesting they could be used in antibody cocktails or be engineered as bispecific antibodies that are cross-variant and resistant to viral escape. Using existing published data and data generated from our library of HCAbs against diverse pre-Omicron SARS-CoV-2 variants, we developed an ML model to rapidly predict which HCAbs would be effective against Omicron BA.1 and Omicron BA.5. Using this ML model three HCAbs with efficacy against Omicron BA.1 and BA.5 were identified without requiring additional biopanning. This integrated computational and experimental pipeline can be leveraged for future outbreaks to rapidly identify cross-variant countermeasures that are effective against potential emerging variants.

Potent HCAbs tightly bind SARS-CoV-2 RBD VOIs and VOCs

To identify Nb sequences with affinity for SARS-CoV-2 VOIs and VOCs, our previously described synthetic humanized Nb phage display library was used to perform four rounds of selection against the RBD of the Alpha, Beta, Delta, Epsilon, Lambda, and Kappa SARS-CoV-2 variants [18]. To inform development of sequence-based ML models and highlight the potential for cross-variant efficacy, SARS-CoV-2 VOIs and VOCs were organized into three groups based on their common amino acid substitutions and mutation sites in the spike RBD (Fig 1). Within the Group 1 SARS-CoV-2 variants (Alpha, Beta, Gamma, & Mu), RBD mutations that potentially increase transmissibility and decrease the effectiveness of vaccines and therapeutics are prevalent. These include the E484K and N501Y mutations found in Beta (B.1.351) and Gamma (B.1.1.28) VOCs [20]. The E484K substitution has been shown to increase the number of serum antibodies needed to prevent infection of cells [21], while the N501Y substitution results in increased affinity of the viral spike protein for the angiotensin-converting enzyme 2 (ACE2) cellular receptor [22]. Within the Group 2 SARS-CoV-2 variants, which include the Delta, Kappa, Epsilon, and Lambda variants, the L452R/Q substitution on the neck of SARS-CoV-2 RBD led to higher infection rates of both vaccinated and unvaccinated people than previous VOIs and VOCs (Fig 1) [23–26]. The Delta variant, which also has the T478K mutation, potentially increased viral infectivity and raised concerns about vaccine efficacy. Here, we present a comparative evaluation of broadly neutralizing HCAbs with efficacy against the original variant, VOIs, and VOCs, including more recent Omicron variants BA.1 and BA.5 (Group 3).

Four rounds of selection were used to identify clones that bind to the RBD region of SARS-CoV-2 variants. Single colonies were selected and used to generate clonal phage, which were expanded in deep-well 96 well plates. Phage-displayed Nbs were tested for RBD-binding by ELISA against the available Group 1 and Group 2 SARS-CoV-2 variant RBDs and were further tested for neutralization against SARS-CoV-2 variant pseudotyped vesicular stomatitis virus encoding an enhanced green fluorescent protein (eGFP) (VSV-SARS-CoV-2-GFP). Ninety-eight clones bound one or more SARS-CoV-2 variant RBDs and neutralized the corresponding VSV-SARS-CoV-2-GFP expressing the spike protein of the VOCs. The CDR sequences of these 98 clones were evaluated for common motifs which might be effective against sets of mutations in the SARS-CoV-2 RBD. Most of the diversity observed within the clonotype groups occurs in the complementarity determining region 3 (CDR3) which can be seen from the shift in positional diversity within each respective CDR (Fig 2).

As previously described, the input library had a diversity of 3.2 x10¹⁰, with three possible CDR3 lengths. Sequencing of the input library showed that the 9-amino acid CDR3 was the most prevalent (40%), with 12-amino acid (34%), and 15-amino acid CDR3s (25%) comprising the rest of the library [18]. After screening against the VOC and VOI RBDs, the distribution of our top clones showed an increase in the prevalence of 9-amino acid CDR3s from 40% in the designed library to 69.5% in Alpha and Beta binders, and 51.4% in Delta, Lambda, Epsilon, and Kappa binders, whereas only around 8.5–10% of 15-amino acid length CDR3 were observed in enriched sequences (S1A Fig). Despite the diverse array of sequences, binding activity, and screened antigens, further analysis revealed that the amino acids that form these CDRs display remarkable conservation among output sequences targeting a functionally conserved epitope on various SARS-CoV-2 variant RBDs (Figs 2, S1B and S1C). Importantly, hit clones have highly conserved CDR1s and CDR2s regardless of whether they bind the Group 1 or Group 2 RBDs (Figs 2A, 3B, 3D and 3E). As expected, the diversity responsible for differential Nb binding to different RBD groups is found in their CDR3s. For example, hit clones from screening against the Alpha and Beta RBDs shared an Asp residue at position 4 in CDR3 of the heavy chain (Fig 2C). Overall frequency of each amino acid across CDR3 loops of the viral variants was evaluated to characterize enrichment of certain amino acids across panning rounds. The frequency of Asp at position 4 was 70.1% in sequences with affinity for Alpha and Beta RBDs (S1B Fig). Similarly, the sequences with affinity for Delta, Lambda, Epsilon, and Kappa RBDs have a common motif with two amino acids with negatively charged side chains at positions 5 (52.4%) and 7 in CDR3 (44.3%) (S1C Fig).

From our 98 ELISA positives selected for sequencing, 59 sequences were identified as unique (S1 data). Humanized HCAbs were expressed as fusion proteins of Nbs with the crystallizable fragment (Fc) of human IgG1 for further characterization [27]. To determine how different sets of amino acid changes impact binding, all 59 HCAbs were evaluated for binding to Group 1 and Group 2 variant RBDs by titration ELISA (S1 Table, Fig 3). The top 37 candidates were selected based on their ability to bind one or multiple variant RBDs, and these were tested in neutralization assays against VSV-SARS-CoV-2-GFP expressing variant spikes (S1 Table). Of these, the top 15 HCAbs with high binding affinity and/or neutralization potency against multiple variant RBDs, were analyzed in kinetic and competition assays to characterize binding affinity and determine HCAb antigen epitopes. Almost all the HCAbs have high binding affinity for WT SARS-CoV-2 RBD (Figs 3 and S2A). The most promising HCAbs identified from screening against Alpha and Beta RBDs (notably AP1C3, AP2B6, BP1B1, BP2A6, and BP2H9) had sub nM binding affinity for similar Group 1 variant RBDs (Beta, Gamma, and Mu) but decreased binding affinity for Group 2 variant RBDs, particularly the Delta and Lambda lineages (Figs 3 and S2B–S2F). Similarly, HCAbs identified from screens against RBDs from Delta, Lambda, Epsilon, and Kappa variants (particularly KP1C9, KP2C6, and DP4F2) preferentially bind to Lambda and Delta RBD with little to no binding to Group 1 variant RBDs. In addition to group-specific binders, we did observe some HCAbs with cross-variant binding, including AP1C3 (with an affinity for RBDs from Mu, Lambda, and Delta variants), and EP2G4 and KP1G11, both with sub-nM K_D binding affinity against all RBDs in this study.

Cross-variant HCAbs compete with ACE2 and neutralize SARS-CoV-2 variants

The top 37 HCAbs that demonstrated high binding affinity or cross-variant binding by ELISA to the VOI and VOC RBDs were tested for neutralization of VSV-SARS-CoV-2-GFP [28] expressing spike protein from the WT, Beta, Gamma, Delta, Lambda, and Mu variants (S1 Table). Figs 4 and S3 summarize the neutralizing efficacy of selected HCAbs candidates against VSV-SARS-CoV-2-GFP [28] expressing variant spikes. Consistent with ELISA results, HCAbs identified from screening against Alpha/Beta variants (AP1B6, BP1B1, and BP2A6), potently neutralized VSV-SARS-CoV-2-GFP expressing WT and Group 1 spikes (Beta/Gamma), while outputs from screening against Lambda/Delta variants potently neutralized VSV-SARS-CoV-2-GFP expressing WT and Group 2 Spikes, including Delta. The most potent neutralizing HCAbs of the 37 tested were AP1B6, with a 50% effective inhibitory concentration (EC₅₀) of 0.12 nM and 0.26 nM against Beta and Gamma pseudotyped variant viruses respectively, and AP1C3, with an EC₅₀ of 0.02 nM and 1.07 nM against Lambda and Delta pseudotyped variant viruses respectively (Figs 4, S3E and S3F).

Experimentally validated ML model enables rapid identification of the best heavy-chain only antibodies against new variants

The vast combinatorial mutational space of both antigens and antibodies presents an obvious challenge for antibody engineering and development, particularly when developing cross-variant antibodies or HCAbs. Here we tested whether an ML model could be trained and used to rapidly identify antibodies with efficacy against a wide range of mutants. Random forest models trained on Log₁₀(K_D) for HCAb binding to SARS-CoV-2 variant RBDs (Beta, Gamma Beta, Gamma, Delta, Lambda, & Mu variants) were generated and used to predict HCAb binding efficacy on novel variants. Features, namely, the Balaban- and Wiener-type index from Z, mass, van der Waals, electronegativity and polarizability weighted distance matrices, were derived from both the HCAbs CDRs, and from the predicted log escape fraction based on public datasets of conventional antibodies [29].

The random forest models were then used to efficiently explore the functional impact of single critical mutations in the SARS-CoV-2 Beta, Gamma, Delta, Lambda, and Mu variant RBDs overcoming the experimental limitations of testing the combinatorically large numbers of potential mutations in Omicron BA.1 and Omicron BA.5. Fig 5A shows the five-fold cross-validation results for a random forest model trained only on pre-Omicron variants. This model had a root-mean-square error (RMSE) of 0.90 log₁₀(μg/mL), Pearson correlation of 0.69, and Q² of 0.47 (Fig 5A). Using these pre-Omicron variants as a training set and Omicron BA.1 as a test set yielded an RMSE of 0.57 log₁₀(μg/mL), correlation of 0.9 (90% accuracy), and Q² of 0.77 (Fig 5B). Interestingly, model performance for Omicron BA.1 binding prediction was better (Fig 5B) than what would be expected based on the cross-validation (Fig 5A); possibly due to the size of the test set. For the next iteration of the model, we added data for Omicron BA.1 to the training data (Fig 5C). The addition of BA.1 data to the training set resulted in an RMSE of 0.8 log₁₀(μg/mL), correlation of 0.76, and Q² of 0.58 (Fig 5C). Application of this random forest model to the BA.5 data yielded an RMSE of 0.86 log₁₀ (μg/mL), correlation of 0.78 (78% accuracy), and Q² of 0.58 (Fig 5D). Not surprisingly, addition of the Omicron BA.1 data resulted in significantly better results (Fig 5C vs 5A), for two reasons: (1) inclusion of a larger training set and (2) addition of already well-predicted Omicron BA.1 data points to the test set. The predictions for the BA.5 variant shown in Fig 5D are well in line with what would be expected from the model result shown in Fig 5C.

The HCAbs identified as effective against Omicron using our ML model were experimentally validated and the best candidates against Omicron BA.1 and BA.5 were identified as BP1B1, BP2A6, and BP3D5. HCAbs were characterized through ELISA (Fig 6A and 6B) and neutralization assays with VSV-SARS-CoV-2-GFP expressing Group 3 (Omicron BA.1 and BA.5) spike variants (Fig 6C and 6D). All three HCAbs bound both Omicron variants, with BP3D5 displaying the highest binding affinity with a K_D of 1.12 and 1.83 nM against Omicron BA.1 and BA.5 RBD, respectively. BP3D5 was also the most potent in neutralization assays, with an EC₅₀ value of 4.50 nM and 0.08 nM against VSV-SARS-CoV-2-GFP expressing Omicron BA.1 and BA.5 (Fig 6E). BP3D5 was not one of the original top 37 HCAbs selected for neutralization assays prior to running the ML models. Once the ML-predicted BP3D5 was experimentally confirmed to be a strong binder and neutralizer of Omicron BA.1 and BA.5 we evaluated BP3D5 for neutralization of pseudotypes in group 1 and group 2. BP3D5 had neutralization activity against Group 1 pseudotyped viruses (Beta EC₅₀ = 0.5 nM, Gamma EC₅₀ = 1.4 nM, and Mu EC₅₀ = 15.4 nM), but not against Group 2 pseudotyped viruses (Delta, Lambda) (S1 Table and S1, S3 Figs)

To further determine the mechanism of action, we performed competition assays between top candidate HCAbs and ACE2-rbFc (rabbit Fc domain) fusion protein [18] with competitive antibodies showing a decrease in ACE2 binding to RBD. Like their ability to block infection with VSV-SARS-CoV-2-GFP expressing Omicron BA.1, BP2A6 and BP3D5 were strongly competitive with BA.1 RBD binding to ACE2 (Fig 7A). BP2A6 and BP3D5 also blocked ACE2 binding to BA.5 but not to the same extent as with BA.1. BP1B1 did not seem to compete with ACE2 for binding to either BA.1 or BA.5 (Fig 7A and 7B), and may be decreasing infection by a non-competitive mechanism.

Epitope binning and structure-based cross-docking analysis of HCAb–antigen interactions

Epitope binning by Bio-layer interferometry (BLI) was used to determine whether our top HCAbs had overlapping epitopes on SARS-CoV-2 RBD. Loading of KP1C9 onto Delta RBD did not seem to interfere with binding of AP1C3. (Fig 8A). However, when AP1C3 was loaded before KP1C9 it blocked binding of KP1C9 to Delta RBD, suggesting that while there is some overlap between the two HCAbs, AP1C3 can bind outside of the KP1C9 binding region (S4 Fig). Similarly, when BP1B1 was first bound to the Omicron BA.1 RBD it did not interfere with binding of BP2A6, But loading Omicron BA.1 with BP2A6 prevented binding of BP1B1 (Figs 8B and S4). In addition, binding of BP2A6 to Omicron BA.5 RBD prevented binding of BP3D5 to BA.5 suggesting that BP1B1 and BP3D5 have overlapping epitopes with BP2A6 but BP2A6 can bind outside of those overlapping regions. (Fig 8C).

To assess the potential to resist future mutation, we docked the HCAbs and hACE2 with Delta and Omicron BA.1 RBD variants to determine how the mutations in the Delta and Omicron variants impacted HCAb binding. In the Delta variant, there are two critical mutations in the RBD, L452R and T478K [30,31]. The L452R mutation has been reported to critically enhance the binding affinity to the host entry receptor ACE2, as well as transmissibility, fitness, and infectivity, and so improves viral replication [32,33]. In docking AP1C3 and ACE2 to the Delta RBD, we found that AP1C3 shares a strong electrostatic interaction that could compete with ACE2 in binding with L452R through electrostatic interaction at D103 (Fig 9). The hydrophobic residue F490 (RBD) is clamped between two aromatic residues, Y101 and Y104 of AP1C3 which form a cation-π stacking interaction and contribute strongly to binding to the RBD and competition with ACE2.

The HCAb BP2A6, selected by the ML model, showed multi-epitope binding to Omicron BA.1. The primary interaction between this antibody and the RBD surface is through a central residue, E484K, and its network of adjacent residues, including F490, F489, N487, F486, and V483. Notably, BP2A6 binds in an orientation that markedly differs from other neutralizing HCAbs in this study (Fig 10B). Its CDR3 adopts an extended beta-hairpin conformation that inserts into the receptor binding site (RBS) using both polar and hydrophobic interactions as well as bridging water molecules (Fig 1B). The flexibility of CDR3 is involved in a hydrophobic and aromatic patch that interacts with the RBD. Integrating these docking studies with HCAbs, RBD variants, and hACE2 interactions offers a comprehensive approach to understanding the antiviral mechanisms of these HCAbs and enables prediction of how potential mutations may impact efficacy.

To verify the docking result, we generated a set of RBDs with mutations in the residues identified by docking simulation. As expected, mutating the WT SARS-CoV-2 RBD at 452 from L to R increased binding to AP1C3, similar to AP1C3 binding to the Delta SARS-CoV-2 RBD which correlates with the docking results in Fig 9 (S5A Fig). While mutating F490 to H in the WT SARS-CoV-2 RBD reduced binding to AP1C3, The WT SARS-CoV-2 RBD containing F490M and F490V (other hydrophobic side chains) maintained higher binding to AP1C3 in comparison to WT SARS-CoV-2 RBD with no mutations, which suggested that a hydrophobic residue at 490 on RBD is critical for the higher binding of the Delta SARS-CoV-2 RBD to AP1C3 (S5A Fig).

For docking with Omicron BA.1 SARS-CoV-2 RBD, we created Omicron BA.1_R493Q and Omicron BA.1_R493E and compared their binding affinities with Omicron BA.1 SARS-CoV-2 RBD and WT SARS-CoV-2 RBD. BP2A6 binding to Omicron BA.1 RBD with the R493Q or the R493E mutation was lower than BP2A6 binding to WT SARS-CoV-2 RBD and Omicron BA.1 SARS-CoV-2 RBD, confirming the docking prediction that hydrogen bonds between BP2A6 and these residues are important for ultra high binding affinity (S5B Fig).

Protein production of SARS-CoV-2 variant RBDs

SARS-CoV-2 variant RBD plasmids and proteins (WT, Alpha, Beta, Gamma, Delta, Epsilon, Lambda, Kappa, Mu) were produced in the following way. These plasmids were constructed from ligation between the pSF-CMV plasmid and gBlocks of the SARS-CoV-2 variant RBD sequence (IDT), with edits made from the original SARS-CoV-2 RBD in the Snapgene software. Double-stranded DNA fragments encoding the SARS-CoV-2 variant RBD, an N-terminal Kozak sequence and signal peptide, and a C-terminal TEV cleavage site and 10xHisTag were commercially synthesized (IDT). The RBD DNA fragment and pSF-CMV plasmid, restriction digested with NcoI and BamHI were assembled using HiFi MasterMix (NEBuilder). The resulting plasmid was transformed into 10β Escherichia coli cells (NEB), amplified in overnight cultures, and purified using Miniprep (Epoch Life Sciences) and Midiprep kits (Qiagen).

SARS-CoV-2 variant RBDs were expressed in Expi293F cells (Thermo, A14527) in 50–100 ml transfections at 1 μg DNA/mL. Transfections were fed 18–22 hours later according to the manufacturer’s specifications and harvested six days post-transfection by centrifugation at 4000 xg for 20 min and filtering through a 0.22 μm filter (Nalgene). The filtered supernatant was run on a HisTrap Excel 5 ml Column (Cytiva) equilibrated in 20 mM phosphate, 300 mM NaCl buffer at pH 7.4 and subsequently through a Superdex 200 Increase 10/300 GL SEC column (GE) equilibrated in PBS. Purified proteins were concentrated using 10 kDa MWCO concentrators (Amicon) and quantified using the DC assay (Biorad). The biotinylated versions of the SARS-CoV-2 WT, Delta, Omicron BA.1 and Omicron BA.5 RBDs were purchased from Sino Biological.

Phage preparation and purification

The phage display library was constructed as previously described by Stefan et al. (2021) [18]. The phage library for biopanning was produced from TG1 E. coli stock diluted to an initial OD₆₀₀ of 0.1 in 2xYT-GA (31 g/L 2xYT, 100 mM glucose, 100 μg/mL Carbenicillin), and grown to an OD₆₀₀ of 0.4–0.6 after which, CM13 phage was added to a final concentration of 2.0 × 10⁹ CFU/mL. The culture was incubated at RT for 15 min and then shaken at 37°C at 250 rpm for 30 minutes. E. coli were then pelleted at 5500 xg for 10 minutes, resuspended in 2xYT-AK (100 μg/mL Carbenicillin, 50 μg/mL Kanamycin), and incubated at 37°C overnight. To purify the phage, E. coli were pelleted at 5500 xg for 10 minutes at 4°C, and 1/4^th volume of 20% PEG, 2.5 M NaCl solution was added to supernatant containing phage. This solution was incubated at 4°C for 4 hours or overnight. The phage were then pelleted at 5500 xg for 1 hour at 4°C and resuspended in 10 mL PBS. The solution was centrifuged at 5500 xg for 5 minutes at 4°C 1–2 times further to pellet any remaining E. coli, after which a second 20% PEG precipitation step was performed on ice for 30 minutes. Finally, the phage was pelleted at 17900 xg for 10 min at 4°C, resuspended in PBS, sterile filtered through a 0.22 μm filter (Nalgene) and titered through serial dilution after incubation with TG1 E. coli.

Biopanning

Four rounds (R1-4) of biopanning were performed against each SARS-CoV-2 variant RBD (Alpha, Beta, Delta, Epsilon, Lambda, Kappa). All rounds of panning were conducted in a 96-well plate format (Immulon 4 HBX Flat Bottom 1x12 strips). For the initial round 3 μg/well of the RBD was coated in each well using blocking buffer (100 mM NaHCO₃, 150 mM NaCl; pH 8.3) overnight at 4°C; for subsequent rounds (R2-4), wells were coated with 2 μg/well of antigen as described above. In round 1, we coated 24 wells with 50 μl/well; in round 2–4, we coated 12 wells with 50 μl/well. Plates were extensively washed with PBST (0.05% Tween-20) on an EL 406 Plate Washer (Biotek). After washing, the block was added to each well for 2 hrs, which, depending on the round, was either Pierce Protein-Free Block (for R1; Thermo), 1% BSA in PBST (R2 & R4), or 2% milk in PBST (R3). Phage dilutions were made up in the corresponding block with either 2.5 × 10¹² CFU/mL of the original Nb-phage library (R1) or 6.67 × 10¹⁰ CFU/mL of the previous round’s phage library for subsequent steps (R2-4); 150 μL of the phage dilution was added to the plate and shaken at 400 rpm at RT for 2 hours. The plate was washed with PBS and Tween percentages that were adjusted during different biopanning rounds as indicated in S2 Table.

To elute bound phage, 50 μL of 0.1 mg/mL trypsin (Sigma) in TBSC (10 mM Tris, 137 mM NaCl, 1 mM CaCl₂; pH 7.4) was added to each well and incubated at 37°C for 30 minutes. Trypsin supernatant was collected, and this step was repeated for another 20 minutes. Log-phase TG1 E. coli was added to the pooled trypsin supernatant (3 mL of culture) and the trypsin-treated wells (100 μL per well). Both phage and E. Coli solutions were incubated at 37°C without shaking for 30 minutes, then shaking for another 30 minutes. Finally, the two preparations were combined, plated on 2xYT-GA plates, and titered as above. Colonies were scraped, resuspended in 10 mL 2xYT-GA, and regrown and purified as done above for the original phage library. After the final round, individual colonies were picked and grown in three Masterblock 2mL 96-well plates (Greiner) with CM13 infection, as done on the larger scale above.

A round of negative selection was carried out after the fourth panning round for additional screening and to identify further Nb candidates specific to a particular SARS-CoV-2 variant. Ten wells of a 96-well plate strip (Immulon) were coated overnight at 4°C with 2 μg/well of each variant SARS-CoV-2 RBDs except for the original panning antigen and structurally similar variants. As above, plates were washed with PBST using the Biotek plate washer and blocked with Pierce Protein-Free Block (Thermo) for 2 hours. Then 2.5 × 10¹² CFU/mL (or if phage concentration was too low, undiluted) R4 phage library was added to row A and incubated shaking at 400 rpm for 1 hour. The phage supernatant was then transferred from row A to row B and incubated by shaking for 1 hour. This process was repeated for each variant RBD coated in a row of the plate until the phage supernatant was collected, used to infect TG1 E. coli, plated, and titered as above.

Phage and heavy chain-only antibodies ELISAs

Nb-phage candidates were chosen based on performance in a monoclonal phage Enzyme-Linked Immunosorbent Assay (ELISA) against various SARS-CoV-2 variant RBDs, both before and after the negative selection round. Maxisorp Nunc 384-well plates were coated with 0.05 μg/well of each respective RBD in coating buffer overnight (100 mM NaHCO₃, 150 mM NaCl; pH 8.3) at 4C. Plates were blocked with Pierce Protein-Free Block (Thermo) for 2 hours, after which 25μl of 1:10 dilution of phage preparation was added and incubated with shaking for 2 hours. Plates were washed as above, and mouse anti-M13 phage primary antibody (Invitrogen; 1:4000) was added for one hour of incubation. Plates were washed, and the HRP-Rabbit anti-mouse IgG secondary antibody (Abcam; 1:2000) was incubated for an additional hour. Finally, plates were washed and developed using TMB Ultra (Thermo). The reaction was stopped with equal volume 2 M H₂SO₄. Absorbance at 450 nm was monitored on a Tecan Spark plate reader.

Maxisorp Nunc 384 well plates were coated with 0.05 μg/well of SARS-CoV-2 variant RBD in coating buffer overnight at 4°C. Plates were washed as above between each step; wells were blocked with 50 μl/well Thermo Pierce Protein-Free commercial block for two hours. After washing, plates were incubated for 2 hours with primary Nb-huFc dilutions starting at 10 μg/ml and diluted 1:3 down the plate. Then, HRP-Goat anti-human IgG secondary antibody (1:15,000; Thermo) was added for 1 hour. Plates were developed with equal volumes of TMB and 2 M H₂SO₄. Absorbance at 450 nm was read on a Tecan Spark plate reader.

Viruses and cell culture

VSV-ΔG-GFP pseudotyped viruses (except Omicron BA.1, Omicron BA.5) with VSV-G were produced in BHK-21 cells through transfection of pVSV-G at 20 μg DNA per 10cm plate and infection with ppVSV-reporter at an MOI of 0.1. The virus was harvested from the supernatant after centrifuging at 1000 xg for 10 min and filtering at 0.45 μm PES (GE). Using a serial virus dilution, VSV-ΔG-GFP was titered over Vero cells transfected with pVSV-G and incubated for 16–20 hours at 37°C. The SARS-CoV-2 variant spike plasmids were produced in the pCAGGS plasmid using gBlocks (IDT) for the matching spike variant. VSV-SARS-CoV-2 variant pseudo viruses (except Omicron BA.1, Omicron BA.5) were produced in Expi293 cells in 25-50ml transfections at 1 μg DNA per ml of culture of the respective pCAGGS plasmid form of the variant spike protein. Transfections were fed per the manufacturer’s instructions 18–22 hours post-transfection. At 72 hours post-transfection, Expi293 cells were infected with VSV-ΔG-GFP at an MOI of 3 for one hour, shaking at 37°C; cells were then pelleted at 1000 xg for 5 min and resuspended in fresh prewarmed Expi293 media supplemented with 0.137 μg/ml anti-VSV antibody (ATCC, from CRL-2700). At 18–24 hours post-infection, the cells were pelleted at 1000 xg for 10 minutes, and the viral supernatant was harvested, snap-frozen, and stored at -80°C.

A replication-competent vesicular stomatitis virus (VSV) expressing eGFP and the SARS-CoV-2 spike gene (rVSV-SARS-CoV-2-GFP) was provided by Dr. Sean Whelan for the Omicron BA.1 [36] and Omicron BA.5 variants. Replication-competent VSV chimeras were expanded in MA-104 cells in Dulbecco’s Modified Eagle medium (DMEM) supplemented with 2% fetal bovine serum (FBS), 100 units/mL penicillin, and 100 μg/mL streptomycin (Invitrogen, 15070063). Viral supernatants were harvested at the time of maximum cytopathic effect (approximately 48 hours post-infection), centrifuged at 1000 xg for 10 minutes to remove cellular debris, filtered over a 0.45 μm filter, and flash frozen with 10mM HEPES buffer. Virus aliquots were stored at -80°C. Viruses were titered by foci-forming unit assay over Vero-E6-TMPRSS-T2A-ACE2 cells. BHK-21 cells (Golden hamster kidney, ATCC CCL-10) were cultured in DMEM containing sodium pyruvate and supplemented with 10% FBS, 100 units/mL penicillin, and 100 μg/mL streptomycin. Vero cells (African green monkey kidney, ATCC CCL-81) were maintained in minimum essential medium alpha (αMEM) supplemented with 10% FBS, 100 units/ml penicillin, and 100 μg/ml streptomycin. Vero E6-TMPRSS-T2A-ACE2 cells (BEI NR-54970) were maintained in DMEM supplemented similarly to the above with 10 μg/ml puromycin. MA-104 cells (African green monkey kidney epithelial, ATCC CRL-2378.1) were maintained in DMEM with 10% FBS, 100 units/mL penicillin, and 100μg/mL streptomycin. All mammalian cell lines were grown at 37°C in 5% CO₂.

Foci-forming unit assays

Serial dilutions of virus aliquots were prepared in triplicate over Vero E6-TMPRSS-T2A-ACE2 cells in a 384-well format, with 25μL of dilution added per well. Cells were incubated at 37°C for 18 hours before fixation in 4% PFA. Fixed cells were counterstained with DAPI nuclear stain. Whole wells were imaged on a Tecan Spark Cyto plate reader, and the number of infected cells was determined. Viral FFU titers were determined as (number of infected cells)/(dilution factor*volume of dilution per well (mL)).

VSV-SARS-CoV-2-GFP fluorescent reporter neutralization assays

Neutralization assays were carried out using confluent Vero cells (or Vero E6-TMPRSS2-T2A-ACE2 cells for Omicron subvariants) plated in black, coated, clear-bottom 384-well-plates (Corning) in supplemented αMEM (or supplemented DMEM for Omicron). Neutralizing HCAbs were serially diluted from a final concentration of 50 μg/ml; VSV-GFP virus (either rVSV-SARS-CoV-2-GFP or variant VSV-SARS-CoV-2-GFP) was added to a final MOI of 0.2. HCAbs and virus were incubated at 37°C shaking at 120 rpm for 30 min to 1 hour, and added to confluent cells. At 16–20 hours later, the plate was washed with PBS, fixed with 4% paraformaldehyde (Electron Microscopy Sciences), 1 μg/ml Hoechst 33342 nuclear stain (Invitrogen, H3570) solution for 30 min in the dark, rewashed, and imaged using the DAPI and FITC channels to visualize nuclei and GFP-positive cells on a Tecan Spark Cyto multi-mode plate reader and image cytometer. Images were analyzed using the Spark Control Image Analyzer Software. Fluorescence values were normalized using no-treatment infection values and dose-response curves, and EC₅₀ values were generated using GraphPad Prism 9.

Gyros HCAb-ACE2 competition assay

Competition assays between top candidate HCAbs and ACE2-rbFc (rabbit Fc domain) fusion protein [18] were carried out on the Gyrolab xPlore system in Bioaffy 1000 HC CDs. Samples were diluted in RexxipA buffer (Gyros Protein Technologies) and run using the General PK program. Biotinylated SARS-CoV-2 variant RBDs (Delta (Sino), Omicron BA.1 or BA.5 (Acro)) were used as the capture reagent at 10 μg/ml. A seven-point dilution series of HCAb (starting at 20 μg/ml; diluted 1:5 down) was pre-mixed with ACE2-rbFc (0.2 μg/ml in all samples & blank), and AlexaFluor 647 goat anti-rabbit IgG (H+L) (2 μg/ml; Invitrogen) was used for detection.

Epitope binning using bio-layer interferometry

Epitope binning assays were carried out on the Octet RED384 system. Samples were prepared in the same buffer and plate as above. Streptavidin biosensors (Sartorius) were used to load biotinylated SARS-CoV-2 variant RBDs (Sino Biological; WT, Delta, Omicron BA1) at 0.15 μg/ml (5 nM). Remaining unbound streptavidin was quenched with 10 μM biotin solution, and association readings were read for two analytes (25 nM) in sequence.

Model antibody/RBD structures

Structures for the antibodies were homology modeled using the SWISS-MODEL web server [37], which has greater than 60% sequence identity to all target sequences, as a template. The protonation states of titratable residues were predicted using PROPKA at pH 6.0 [38,39]. Modeled structures were further refined using Molecular dynamic (MD) simulations performed using NAMD ver. 2.1 [40] and the CHARMM36 forcefield [41]. The models were solvated in a box of TIP3P water molecules that extended 10 Å from the surface of the protein structure. Counterions (Na⁺, Cl^-) were added to achieve a NaCl concentration of 0.1 M. MD simulations were run with periodic boundary conditions at a constant temperature of 333.15 K [42,43]. Solvated proteins were energy minimized for 10 ps and then gradually heated to the desired temperature (333.15 K) over 300 ps. The system was then equilibrated at 333.15 K for 5 ns under an isothermal−isobaric ensemble (NPT), followed by a 50 ns production run in the canonical (NVT) ensemble. Snapshots were collected at 5 ps.

Machine learning model

To calculate dissociation constants (K_D) for each HCAb/variant combination, concentration-response data were fit to the titration curve equation with a background term using a maximum likelihood estimation nonlinear regression. A 4-degree-of-freedom t-distribution was used to model the error to improve robustness. A competing constant model was also fit, and if it had a lower AIC for a given HCAb/variant combination, the data was removed from the set. Out of 399 HCAb/variant combinations, 4 were removed. Log₁₀(K_D) were used as the endpoint for a random forest model [44]. HCAb heavy chain CDR’s were combined into one long sequence and used to generate features with protR [45]. 8 features were based on sequence-order-coupling numbers with a maximum lag of four, and 8 features were scales-based descriptors based on a selection of ten Dragon [46] topological descriptors (namely, the Balaban- and Wiener-type index from Z, mass, van der Waals, electronegativity and polarizability weighted distance matrices). The top two principal components and a maximum lag of 2 were used. Two features were based on a model for antibody log escape fraction built from a public dataset [47] in particular, we used its log escape fraction predictions for Cov2-A2050 and Cov2-A2082 for the given variant’s sequence.

S1 Table. Several HCAbs characterized by titration ELISA and neutralization assay against several SARS-CoV-2 variant RBDs.

The data are from experimental conditions performed in triplicateAn HCAb is considered non-binding (n.b.) for a variant RBDs when the respective kD is > 125 nM (10 ug/ml). An HCAb is considered non-neutralizing (n.n.) for a particular VSV pseudotyped viral variant when the respective EC50 is > 625 nM (50 μg/ml). Some HCAbs were assigned as non-tested (n.t.) because they did not show binding affinity with their biopanned variants.

https://doi.org/10.1371/journal.ppat.1012903.s001

(XLSX)

S1 Fig. Distribution of amino acid residue lengths of CDR3 loops across libraries (A). Amino acid frequencies by position in CDR loops of various libraries (B, C).

We classified 5 amino acid families by side-chain properties, Hydrophobic side chain: amino acids A, V, I, L, M, F, Y, W; Positively charged side chain: R, H and K; Negatively charged side chain: D, E; Special cases: C, G and P.

https://doi.org/10.1371/journal.ppat.1012903.s003

(TIF)

S2 Fig. Several HCAbs characterized by titration ELISA against SARS-CoV-2 variant RBDs.

A) WT, B) Mu, C) Beta, D) Gamma, E) Lambda, and F) Delta. The data are from experimental conditions performed in triplicate; the error is the standard deviation from the mean.

https://doi.org/10.1371/journal.ppat.1012903.s004

(TIF)

S3 Fig. Neutralization efficacy of highest affinity HCAbs against VSV-SARS-CoV-2 expressing variant spike proteins.

(A), Mu (B), Beta (C), Gamma (D), Lambda (E), and Delta (F). Effectiveness is measured in a half-maximal effective inhibitory concentration EC₅₀ (nM) and the data shown is normalized to the infection rate in the absence of the HCAbs. The data are from experimental conditions performed in triplicate, the error is the standard deviation from the mean.

https://doi.org/10.1371/journal.ppat.1012903.s005

(TIF)

S4 Fig. Competition assay was conducted by using Bio-layer interferometry (BLI) to study binding against two epitopes on SARS-CoV-2 Delta (A), Omicron BA.1 (B) and Omicron BA.5 (C) RBDs.

Epitope binning was performed by injecting a first HCAb for 600 s, followed by injection of a second HCAb for 600 s, with a final dissociation step for 600 s.

https://doi.org/10.1371/journal.ppat.1012903.s006

(TIF)

S5 Fig. Titration ELISA of RBDs from WT, Delta, and Omicron BA.1 SARS-CoV-2 variants with single point mutations with AP1C3 (A) and BP2A6 (B).

The data are from experimental conditions performed in triplicate; the error is the standard deviation from the mean.

https://doi.org/10.1371/journal.ppat.1012903.s007

(TIF)