key: cord-0300424-1exwassg
authors: Zhao, Fangzhu; Keating, Celina; Ozorowski, Gabriel; Shaabani, Namir; Francino-Urdaniz, Irene M.; Barman, Shawn; Limbo, Oliver; Burns, Alison; Zhou, Panpan; Ricciardi, Michael J.; Woehl, Jordan; Tran, Quoc; Turner, Hannah L.; Peng, Linghang; Huang, Deli; Nemazee, David; Andrabi, Raiees; Sok, Devin; Teijaro, John R.; Whitehead, Timothy A.; Ward, Andrew B.; Burton, Dennis R.; Jardine, Joseph G.
title: Engineering SARS-CoV-2 neutralizing antibodies for increased potency and reduced viral escape
date: 2022-01-07
journal: bioRxiv
DOI: 10.1101/2022.01.06.475303
sha: 14e5f39b696e5641490bb496bb2c5563324bcba7
doc_id: 300424
cord_uid: 1exwassg

The rapid spread of SARS-CoV-2 variants poses a constant threat of escape from monoclonal antibody and vaccine countermeasures. Mutations in the ACE2 receptor binding site on the surface S protein have been shown to disrupt antibody binding and prevent viral neutralization. Here, we use a directed evolution-based approach to engineer three neutralizing antibodies for enhanced binding to S protein. The engineered antibodies showed increased in vitro functional activity in terms of neutralization potency and/or breadth of neutralization against viral variants. Deep mutational scanning revealed that higher binding affinity reduced the total number of viral escape mutations. Studies in the Syrian hamster model showed two examples where the affinity matured antibody provided superior protection compared to the parental antibody. These data suggest that monoclonal antibodies for anti-viral indications could benefit from in vitro affinity maturation to reduce viral escape pathways and appropriate affinity maturation in vaccine immunization could help resist viral variation.

Over the last year and a half, severe acute respiratory syndrome coronavirus 2 (SARS- 2) has had devastating consequences for global health and economies. Following the discovery of 41 the disease, there was a rush to produce protective vaccines and therapeutics. Multiple highly 42 effective vaccines have been developed that elicit immune responses against the SARS-CoV-2 43 spike (S) trimer 1 . The protective mechanisms for the coronavirus disease 2019 vaccines are still being deduced, however, several analyses have found that the elicitation of 45 neutralizing antibodies (nAbs) correlates with protection 2,3 , a finding consistent with many other 46 successful antiviral vaccines 4 . NAbs have been identified that target several distinct epitopes on the 47 S trimer, but the majority of nAbs target the receptor binding domain (RBD) 5,6 . While vaccines are 48 undisputedly the most effective strategy for control of COVID-19, recombinantly produced nAbs 49 offer the potential to supplement prophylactic coverage in populations that respond poorly to 50 vaccines, e.g. immunocompromised individuals, can be administered as a post-exposure 51

prophylactic and can be used therapeutically to prevent hospitalization 7, 8 . 52

One of the unique challenges in using a neutralizing monoclonal antibody (mAb) for antiviral 53 indications is addressing existing viral diversity and the high mutational propensity in viruses that 54 can give rise to resistant viral variants. Since the discovery of SARS-CoV-2 in 2019, thousands of 55 viral variants containing synonymous and nonsynonymous mutations have been documented 9 . A 56 growing number of these new variants (termed "Variants of Concern" or VOCs) 10 contain mutations 57 that increase infectivity and/or allow viral escape from monoclonal nAbs elicited against the original 58 SARS-CoV-2 11-14 . Several strategies are commonly used to mitigate the likelihood of viral escape 59 from nAbs. Investigators often select antibodies that target functionally important and conserved 60 regions, reducing the number of mutations that can allow viral escape without incurring a fitness 61

cost [15] [16] [17] . It is also common to use cocktails of at least two nAbs targeting different epitopes, so 62 multiple mutations are necessary for viral escape [18] [19] [20] [21] . A third approach that is less well explored is 63 to in vitro affinity mature the nAb against the target antigen to increase the binding affinity, helping 64 to mitigate the impact of the viral mutations 22 . Here, we explore how increased binding affinity

Structural analysis of nAbs CC6.30 and CC6.33 71 Previously, we reported the structure of nAb CC12.1, which binds to the RBS-A or Class 1 72 epitope site and competes directly with angiotensin-converting enzyme 2 (ACE2) 5,24,25 but 73 specificities of two other nAbs of interest, CC6.30 and CC6.33, were limited to epitope-binning 74 data 23 . To better understand the molecular contacts of antibodies, we used cryoelectron microscopy 75 (cryoEM) to solve the structures of two of the antibodies in complex with stabilized SARS-CoV-2 S 76 trimers: 1) nAb CC6.30, which targets the RBS-B or Class 2 epitope site, binds RBD with an affinity 77 of 1.7 nM (IgG format) and directly competes with ACE2, and 2) nAb CC6.33, which targets the 78

Class 3 epitope site that is distinct from the ACE2 binding site and binds RBD with an affinity of 257 79 nM ( Fig.1 and Fig.2 ). Several clinical-stage nAbs also recognize these epitopes such as CB6/LY-80

CoV16 (RBS-A or Class 1) 25,26 , REGN10933 (RBS-B or Class 2) 14,27 and S309 (Class 3) 5,28 . In each 81 complex, we used a stabilized, uncleaved Spike trimer (HPM7) based on the Wuhan strain 82 containing 6 proline (hexaproline or HP) mutations 29 and an engineered interprotomer disulfide 83 (mut7 or M7) between residues 705 and 883 of the S2 subunit. The global resolution of the two 84 complexes was 3.6 Å (CC6.30/HPM7) and 3.3 Å (CC6.33/HPM7) and model building was assisted 85 by local refinement maps of just the RBD and fragment antigen binding (Fab) variable region 86 (Supplementary Fig.1 contributions from all heavy chain (HC) and light chain (LC) complementarity determining regions 91 (CDRs) except for LCDR2 ( Fig.1a,b) . The HCDR3 contains a disulfide bond (C99-C100d, Kabat 92 numbering) that makes it more rigid, perhaps in turn acting to stabilize this more flexible region of 93 the RBD around the "ACE2-binding ridge" (approximately Spike residue numbers 470-490) 94 (Supplementary Fig.1b, c) . Important interactions at the binding interface include hydrophobic 95 packing of the RBD F490 side chain with antibody HC side chains I31, I52, I53 and I54 96 ( Supplementary Fig.1d ), and predicted hydrogen bonding between the side chain of HCDR3 R97 97 and both the RBD I492 backbone carbonyl and Q493 side chain ( Supplementary Fig.1e) . A salt 98 bridge is formed between the side chains of RBD E484 and LCDR3 R96 (Fig.1c) . This interaction 99 is predicted to decrease the effectiveness of CC6.30 against certain VOCs, specifically Beta 30 and 100

Gamma 31 VOCs, which contain an E484K mutation that abolishes the salt bridge and causes a 101 possible charge-charge repulsion. Of similar concern, the L452 side chain of the RBD has 102 hydrophobic interactions with HCDR1 I31 and HCDR2 I53 which would be incompatible with a long, 103

positively-charged side chain that results from the L452R mutation found in the Delta variant 32 104 (Fig.1d) . 105

CC6.30 appears to only bind the RBD-up state of the Spike (Fig.1a) . While the ligand-free 106 cryoEM structures of HPM7 spike reveal a 3 RBD-down conformation, our data suggest that the 107 antibody is capable of shifting the equilibrium by capturing RBD in the up conformation . The most  108  stable cryoEM reconstruction contains 2 up RBDs, each bound by CC6.30, while the unoccupied  109 RBD remains in the down position (Fig.1a) . Superposition of the RBD:CC6.30 portion of the model 110 onto the all-down RBD ligand-free structure predicts that clashes would occur between HCDR3 and 111 the RBD and N343 glycan from an adjacent protomer ( Supplementary Fig.1f ). Finally, E484 is a 112 major part of the epitope as defined in the Class 2 or RBS-B nomenclatures 14 , consistent with our 113 other observations for CC6.30. 114 CC6.33 binds a non-overlapping epitope to that of CC6.30, with the heavy and light chain 115 interface centered on the N343 glycan ( Supplementary Fig.2a,b) . In contrast to CC6.30, this nAb 116 binds the RBD-down conformation and the most stable reconstruction has 2 down RBDs bound by 117 the antibody, while the third RBD is in the up position (Fig.2a) . Indeed, the binding of CC6.33 to a 118 down RBD requires slight opening of the apex to relieve a clash with the RBD ridge of the adjacent 119 protomer, and is likely the driving force for the unoccupied RBD shifting to the up position after the 120 binding of two Fabs ( Supplementary Fig.2e ). While modeling suggests that CC6.33 should be able 121

to bind an up RBD, we did not observe this in our dataset, possibly due to the HPM7 spike design 122 preferentially displaying 3 down-RBDs ( Supplementary Fig.2f) . Also, portions of HC framework 123 regions 1 (HFR1) and HFR3 contact the RBD ridge of the neighboring protomer (still in the down 124 position) in a manner that mimics the interaction between the unoccupied up-RBD and an adjacent 125 RBD-down ridge, further stabilizing the interaction between antibody and spike ( Fig.2c,d) . Binding 126 of CC6.33 to its epitope is largely governed by hydrophobic interactions involving the HC, including 127 HCDR2 residues I52, I53 and L54 packing against RBD residues L335, V362 and P527 128 ( Supplementary Fig.2g ). HCDR3 W98 reaches into an aromatic pocket lined with RBD residues 129 F338, F342, A363, Y365 and L368, while also donating a hydrogen bond to the backbone carbonyl 130 of D364 ( Supplementary Fig.2h ). Fewer hydrogen bonds are predicted between RBD and CC6.33 131 compared to CC6.30. Those contributed by CC6.33 often involve bonds to RBD main chain atoms 132 (e.g. HCDR3 Q97 with RBD backbone C336, V362 and D364), making such interactions less 133 susceptible to changes in side chains resulting from VOC mutations ( Supplementary Fig.2i ). The 134 antibody epitope itself is largely positioned away from the common RBD mutations that could affect 135 binding, a property shared with other Class 3 RBD antibodies (Fig.2d) . Lastly, the LC is mostly 136 involved via LCDR2 packing against and providing hydrogen bonds to the viral N343 glycan, and a 137 single peptide-peptide hydrogen bond between the side chains of LCDR1 Y32 and RBD E340 138 ( Fig.2b and Supplementary Fig.2j ). 139

140 Affinity maturation of CC12.1, CC6.30 and CC6.33 was achieved using our rapid affinity 141 maturation strategy, SAMPLER 33 . Briefly, HC and LC libraries were synthesized containing one 142 mutation per CDR loop from the starting sequence, for up to three mutations per chain. Potential 143 liabilities were informatically filtered from the library process and an N-linked glycan on LCDR1 of 144 CC6.30 was removed by an N28S mutation, reverting that position to the original amino acid found 145 in the germline VK1-39 gene segment so that any improved CC6.30 variant would not contain that 146 glycan. The HC and LC libraries were displayed on the surface of yeast and iterative rounds of 147 selections were used to enrich for clones with higher affinity for SARS-CoV-2 RBD (for CC12.1 and 148

CC6.30 libraries) or S protein (for the CC6.33 library). The sort process also included a round of 149 negative selection, where clones with low binding to a polyclonal preparation of detergent-150 solubilized HEK293 cell membrane proteins were enriched to remove polyreactive variants. The 151 enriched clones were then combined into a heavy/light combinatorial library and screened again 152

with the same four-round selection strategy to identify the optimal heavy/light pairs 33 . At the 153 conclusion of the selection process, sequences of the antibodies were recovered and 12 improved 154 variants of each antibody were selected to be reformatted and expressed as IgG for 155 characterization. 156

All enhanced (e) eCC12.1, eCC6.30 and eCC6.33 variants bound SARS-CoV-2 RBD with 157 higher affinity than the parental antibodies, with an average 45-fold (5-to 267-fold) increase in 158 monovalent equilibrium dissociation constants (Fig.3a ) measured by surface plasmon resonance 159 (SPR). In nearly all cases, the affinity gains came through a reduction in the dissociation rate 160 (Supplementary Table 3 ). The binding affinity for monomeric RBD is notably lower for CC6.33 161 compared to CC12.1, CC6.30 (Fig.3a ) and the majority of other antibodies isolated from our COVID-162 19 cohort 23 . The binding affinity of CC6.33 Fab for S protein is approximately 10-fold higher than 163 for RBD, suggesting that the CC6.33 epitope is poorly formed on monomeric RBD and/or differential 164 processing of the N343 glycan affects mAb binding. ELISA binding to SARS-CoV-2 RBD and S by 165 CC12.1 and CC6.30 parental and engineered nAbs (enAbs) was comparable, however, a large 166 improvement in neutralization EC50 and the maximum neutralization plateau was observed for 167 eCC6.33 variants compared to the CC6.33 parental ( Supplementary Fig.3 ). The enAb variants were 168 evaluated by analytical size exclusion chromatography and found to be monodispersed with similar 169 column retention time to our clinical controls ( Supplementary Fig.4) concentrations of 50 µg/mL ( Fig.3c-e) . As a control, we also tested our enAbs against authentic 191 SARS-CoV-2 and observed similar neutralization activity to the pseudotyped virus ( Fig.3f ), 192 consistent with our previous observations. Taken together, this data suggest that, in this system, 193

increases in binding affinity translate to increases in the in vitro neutralization potency until a 194 "threshold" IC50 around 10 ng/mL is reached, at which point further increases in binding affinity do 195 not appear to affect the in vitro neutralization function of the antibody. Furthermore, this apparent 196 affinity required to reach this neutralization threshold is lowered by the bivalent binding of an IgG. 197 However, a Fab can neutralize the virus provided the monovalent affinity is sufficiently high, 198

indicating inter-or intra-spike cross linking may help but is not necessary for these nAbs to 199 neutralize ( Fig.3c and Fig.3e ). 200

We next sought to investigate how the evolving viral diversity of SARS-CoV-2 variants 201 impacts the binding and the neutralization function of our parental and select engineered nAbs. We 202 first measured the neutralization potency of our nAbs against VOCs with full-spike mutations 203

including Alpha (B. neutralized all the VOCs, including Beta and Delta containing the K417N/T mutations (Fig.4c) . We 214 measured neutralization of all 12 eCC12.1 antibodies against Beta and Gamma as well as the single 215 mutation variants, where we found 11 out of 12 mAbs neutralized the Gamma lineage and 9 out of 216 12 neutralized the Beta lineage ( Supplementary Fig.6a ). Mutational analysis of the 9 enAbs that 217 reacted against both VOCs found a broad assortment of mutations that had been selected across 218 the different antibodies ( Supplementary Fig.6a Fig.1c,d) . 228

The parental CC6.33 and eCC6.33.8 were effective at neutralizing all VOCs tested with 229 similar IC50s to the original Wuhan-1 SARS-CoV-2, consistent with the observation that CC6.33 230 recognizes the conserved class 3 epitope 5,28 distal from the mutations in these viruses (Fig.2b , 231 Fig.4b ,c, and Supplementary Fig.6b ). REGN10987, another class 3 antibody 5,27 retained similar 232 potency for all VOCs tested ( Supplementary Fig.6b ). 233

To systematically investigate the relationship between binding affinity and neutralization for 234

VOCs across our collection of parental and engineered nAbs, a panel of monomeric RBD variants 235 were expressed for SPR analysis. Overall, the RBD binding affinities and the off-rate correlated well 236

with the in vitro neutralization data ( Fig.4d-f ). Mutations that completely abrogated neutralization 237 usually showed a complete loss of binding by SPR or weak reactivity that could not be fit to a simple 238 kinetics model. Of particular note were the Beta and Gamma RBDs binding to the CC12.1 variants. 239

Parental CC12.1 bound Wuhan-1 RBD with an affinity of 6 nM, and had a complete loss of both 240

binding and neutralization to both Beta and Gamma RBDs. In contrast, affinity matured eCC12.1.4 241 bound Wuhan-1 with an affinity of 286 pM, and although the mutations in Beta and Gamma reduced 242 binding affinity by 182-fold and 43-fold (Supplementary Table 4 ), respectively, eCC12.1.4 was still 243 able to neutralize both VOCs (Fig.4c) . These data indicate that enhanced nAb affinity for the target 244 antigen helps to offset the affinity losses resulting from viral mutations within the nAb paratope, 245 allowing the nAb to maintain sufficient affinity for neutralization. Table  253 5). Consistent with previous reports 36 and our neutralization screening, CC12.1 is vulnerable to 254 multiple mutations at K417 with a false discovery rate (FDR) below 0.1 for K417N/T but eCC12.1.4 255 is able to accommodate all mutations at K417 (Fig.5a ,b, and Supplementary Fig.7b ). We also 256 detected multiple mutations at position D420 and N460 that confer escape from the parental CC12.1 257 ( Fig.5a,b) . Alanine scanning and pseudovirus escape mutations had identified these D420 and 258 N460 residues as important for public VH3-53 SARS-CoV-2 antibodies 35, 37, 38 , but structural analysis 259

shows these two positions on the periphery of the CC12.1 epitope and making relatively insignificant 260 contacts to the antibody (Fig.5c ,d). 261

Pseudoviruses containing the individual D420K, N460H, N460P, and N460A mutations, 262 identified as potential escape mutations in the deep mutational scanning, were produced and 263 evaluated to determine if parental CC12.1 and several eCC12.1 variants were sensitive to these 264 mutations in a neutralization assay. In agreement with RBD library screening, a D420K substitution 265 completely disrupted CC12.1 neutralization, while substitutions at the N460 residue significantly 266 decreased its neutralization potency by 12-to 246-fold ( Fig.5e-g) . By contrast, although D420K and 267 N460H were identified as potential escape mutations against eCC12.1.4, neutralization potency 268 was reduced by a more modest 8-fold against D420K and remained insensitive to a N460H 269 substitution ( Fig.5d ,f). These data suggest that increasing the affinity of SARS-CoV-2 nAbs restricts 270 the potential escape mutations that can arise in RBD, rather than just altering the critical nAb 271 contacts and shifting the escape mutations to a comparable number of different positions and/or 272 mutations. This is particularly important in the context of developing antiviral antibodies where viral 273 escape is a serious and constant threat. 274

In vitro analysis of the enAbs suggested that the increased affinity provided functional 276 improvements for two of the three candidates. eCC12.1 variants were better able to overcome 277 mutations that are emerging in the VOCs and eCC6.33 variants had increased neutralization 278 potency and higher MPN compared to CC6.33. The improvements to CC6.30 were less clear cut 279 and the parental antibody was found to have poor pharmacokinetics in hamsters and so was not 280 pursued further in vivo (data not shown). A series of experiments were designed to compare 281 parental and engineered nAbs in the Golden hamster model of COVID-19 infection. The 282 experimental design for passive transfer studies is shown in Fig.6a . Groups of six hamsters were 283

prophylactically treated with serially diluted doses of antibody starting at 2 mg per animal to 8 µg 284 per animal via intraperitoneal (i.p.) injection 72 hours before intranasal challenge with SARS-CoV-285 2 at a dose of 1 x 10 5 plaque-forming units (PFU). A group receiving 2 mg doses of an irrelevant 286 human mAb against dengue virus (Den3) was used as a control for each experiment. All hamsters 287

were monitored daily for weight loss as a measure of disease 39 and serum was collected from each 288 animal to determine antibody titer at the time of viral challenge (D0) compared to the time of sacrifice 289 (D7) (Fig.6b) . Hamsters have been shown to clear SARS-CoV-2 infection after 7 days, so a replicate 290 of the original experiment was performed in which the groups of hamsters were euthanized four 291 days post infection (D4) and lung tissue was collected to quantify lung viral titers. In addition to 292 collecting serum at D0 and D7 to measure nAb titers, the half-life of the parental and several 293 engineered versions were assessed to try to find engineered nAbs that closely matched the bioavailability of the parental versions to allow a comparison of the two. Ultimately, eCC6.33.3 and 295 eCC12.1.6 were selected to compare to the parental nAbs. 296 We first tested the ability of CC12.1 and eCC12.1.6 to protect against challenge from the 297 Beta VOC, as the in vitro data showed that eCC12.1.6 effectively neutralized the variant 298

( Supplementary Fig.6b ). The prophylactic protection experiment described above was done using 299 Beta (20H/501Y.V2) SARS-CoV-2. Consistent with our in vitro neutralization data, eCC12.1.6 300 exhibited a dose-dependent protective response both in terms of weight loss and lung viral titers 301 ( Fig.6c-e) . Parental CC12.1 showed no protection compared to the Den3 control group. We also 302 assessed eCC12.1.6 against the original SARS-CoV-2 (USA/WA1/2020) using the same groups 303 ( Supplementary Fig.8 ) and the weight loss trend was nearly identical to that of the Beta variant, 304 albeit the Beta variant showed a lower overall percentage of weight loss ( Supplementary Fig.8e ). 305

The second protection experiment was designed to test whether the increased in vitro 306 neutralization potency and maximum neutralization percentage of eCC6.33.3 relative to the 307 parental nAb provided enhanced protection. CC6.33 (in vitro IC50 = 0.228 µg/mL and MPN of 81%) 308 was compared with eCC6.33.3 (in vitro IC50 = 0.008 µg/mL and MPN of 100%) for protective efficacy 309 against challenge from the original SARS-CoV-2 (USA/WA1/2020) ( Supplementary Fig.9 ). 310

Following viral challenge, animals that received either the parental or engineered nAb, including the 311 groups that received the 8 µg dose, showed a statistically significant reduction in weight loss 312 compared to the group receiving the Den3 (Fig.6f,g) . The protection from weight loss correlated 313 with the amount of nAb the animals received, and there was no apparent difference in efficacy 314 between the parental and engineered nAbs. However, in contrast to the weight loss results, the 315 enAbs showed superior ability to control viremia in the lung (Fig.6h) . Animals that received the 2 316 mg dose of eCC6.33.3 had sterilizing immunity and the animals that received 500 µg dose of 317 eCC6.33.3 had lung viral titers 5 logs lower than the Den3 control group. Animals that received the 318 parental CC6.33 had viral titers only 2 logs lower than the Den3 group, protection that was 319 comparable to the group receiving 125 µg of eCC6.33.3. It is unclear why there is a disconnect 320 between lung viral titers and weight loss, however, eCC6.33.3 was clearly superior at controlling 321 lung viral load. Broadly, the two experiments confirm that the affinity engineering of these nAbs 322 provide a superior in vivo benefit predicted from in vitro analysis. 323

During the preparation of this manuscript the Omicron VOC was reported. The VOCs 325 characterized above contained 7 to 12 mutations in S compared to the original SARS-CoV-2, and 326 at most contained 3 mutations in RBD (Beta and Gamma). In contrast, Omicron contains 30 327 mutations, 3 deletions and an insertion in S, with 15 of these mutations located in the RBD. The 328 mutations in RBD are heavily concentrated across the class I, class II and class III neutralizing 329 antibody epitopes and have been shown to reduce the neutralization efficacy of plasma from vaccinated and/or infected donors 40-43 . They also confer resistance or complete escape from the 331 majority of clinical antibody candidates 43 . 332

Omicron completely escaped from the parental CC6.30 and all the eCC6.30 variants 333

( Supplementary Fig.10 ). This was largely consistent with data from other VOCs demonstrating the 334 importance of the E484 interaction for this antibody family. Omicron not only contains an E484A 335 mutation, but also contains a Q493K mutation that removes a hydrogen bond between Q493 and 336 R97H on CC6.30 variants. Similarly, the parental CC6.33 and all eCC6.33 variants were also unable 337 to neutralize Omicron ( Supplementary Fig.10 ). This observation was more unexpected. The only 338

Omicron mutation immediately within the CC6.33 epitope is G339D that may introduce a clash with 339 the CC6.33 CDRH3 backbone. Omicron also contains mutations S371L, S373P and S375F that 340 could alter the conformation of an adjacent loop and prevent the N343 glycan from adopting the 341 conformation observed in the CC6.33 bound structure. Finally, we observed that Omicron was 342 resistant to parental CC12.1 and 11 of 12 eCC12.1 variants, however, eCC12.1.6 retained 343 neutralizing activity against Omicron (IC50 of 0.20 µg/mL), albeit with approximately 25-fold reduced 344 potency (Fig.7a , and Supplementary Fig.10 ). eCC12.1.6 was also the most effective antibody 345 against Beta and Gamma VOCs, with comparable potency to the Wuhan-1 strain (Fig.7a ). Analysis 346 of the selected mutations in eCC12.1.6 compared to other eCC12.1 variants did find a unique S31W 347 mutation in LCDR2 that is located in close proximity to the N501Y mutation in all VOCs as well as 348 G466S, G496S, Q498R and Y505H that are present in Omicron (Fig.7b) . It is also possible that the 349 other mutations in the antibody that do not interact directly with RBD stabilize the CDR loops in a 350

conformation that happened to be slightly more compatible with Omicron compared to the other 351 antigen through directed evolution and then investigated the relationship between binding affinity, 362

in vitro neutralization potency and in vivo efficacy. Broadly, we found that monovalent binding affinity 363 and in vitro neutralization potency are correlated, until the in vitro neutralization IC50 reaches a 364

"threshold" (around 10 ng/mL for IgG) after which further affinity improvements did not translate to 365 improvements in neutralization potency, at least for CC12.1 and CC6.30 nAbs that directly compete 366 with ACE2 binding to the RBD. These affinity improvements do help to expand the breadth of antibody reactivity, allowing them to better neutralize VOCs that contain mutations in and around 368 the antibody epitope. This was particularly evident with the eCC12.1 variants that are part of the 369 shared VH3-53 lineage nAbs and are broadly susceptible to the K417T/N mutations found in Beta 370

and Gamma VOCs 14,36 . These mutations abrogated binding of the parental CC12.1 nAb, much like 371 other reported VH3-53 nAbs; however, the higher affinity of the eCC12.1 variants for wild-type S-372 protein was able to compensate for lost the K417 contact allowing the nAb to maintain sufficient 373 affinity that it could still potently neutralize the VOCs. Of note, eCC12.1 variants were affinity 374 matured against the original SARS-CoV-2 RBD sequence and had no specific selective pressure 375

to accommodate mutations in the VOCs. Importantly, our saturated mutagenesis screening showed 376 that the affinity maturation restricted the number of potential escape mutations rather than just 377 altering them to different positions. While increasing the affinity could restrict escape mutations, it 378 did not abrogate them entirely, as eCC12.1 variants showed modest sensitivity to the D420K 379 mutation, and all eCC6.30 variants were still unable to bind or neutralize Beta and Gamma VOCs 380 with the E484K mutation. In the extreme case of the Omicron variant that contains so many 381 mutations on the RBD, especially within the footprint of class 1 antibodies, there was still one 382 CC12.1 variant that had significant neutralizing activity. This finding illustrates the value of affinity 383 maturation in the context of natural infection in that the generation of a diverse set of related 384 antibodies, as was done here in vitro, will likely generate some antibodies able to bind to and act 385 against many different viral variants, including those with multiple mutations as for Omicron (and 386 see below). It is possible that part of the reason so many clinical antibody candidates failed against 387

Omicron is that most were selected shortly after COVID infection before much affinity maturation 388 had occurred 46 . 389

The engineering of higher affinity enAbs also improved protective efficacy in vivo. The 390 protection data also show that eCC12. CoV-2 RBD and RBD mutant proteins. NB: no binding. NF: Not fit to a simple kinetics model. 616

Antibodies were captured to SPR sensors via a Fc-capture, multi-cycle method. Association and 617 dissociation rate constants were calculated through a 1:1 Langmuir binding model using the 618 BIAevaluation software. e-f, Pearson correlation analysis between antibody e off-rate constant (Kd) 619

or f equilibrium dissociation constant (KD) binding kinetics against RBD variants and antibody 620 neutralization potency against mutant pseudoviruses. 621 variants that did not disrupt ACE2 interaction but evaded nAb recognition were sorted and 625 sequenced. A control with no ACE2 labeling was also sorted and served as an empirical false 626 discovery rate (FDR). The enrichment ratio for each mutation relative to the reference population 627 was colored according to the key. The heatmap of RBD residues 405 -470 was shown, whereas 628 the full map of residues from 333 to 527 was shown in Supplementary Fig.7 . c-d, Crystal structure 629 of CC12.1 interacting with SARS-CoV-2 RBD modified from PDB: 6XC2 24 . CC12.1 heavy chain 630 and light chain were colored in dark green and light green respectively. Key binding residues N501, 631

K417, N460 and D420 on RBD were highlighted in orange. e-f, Representative neutralization curves 632 of CC12.1, eCC12.1, eCC12.1.6, and eCC12.1.7 against SARS-CoV-2 e D420K and f N460H. 633

Antibodies were colored according to the key. Error bars represent standard deviation. Data were 634

representative of at least two independent experiments. g, Neutralization potency of parental 635 CC12.1 and eCC12.1 variants against wild type SARS-CoV-2 virus and D420K, N460H, N460P, 636 N460A mutant viruses. 637 Representative neutralization curves of eCC12.1.6 against SARS-CoV-2 VOCs. Wildtype SARS-655

CoV-2 and VOCs were colored according to the key. Error bars represent standard deviation. b, 656

Structure of eCC12.1.6 interacting with Omicron RBD modified from PDB: 6XC2 24 . eCC12.1.6 657 heavy chain and light chain were colored in dark green and light green respectively. Antibody 658 mutant residues were highlighted in yellow spheres while Omicron RBD mutant residues were 659 highlighted in navy blue spheres. 660

Golden Syrian hamsters were provided by Charles River Laboratories (CRL:LVG(SYR)) and 663

housed at the Scripps Research Institute. Male 12-13-week-old hamsters were infused with 664 antibodies intraperitoneally as described previously 23 . The Scripps Research Institutional Animal 665

Care and Use Committee (IACUC) approved all experimental procedures involving all the animals 666

in accordance with Protocol #20-0003. 667

Cell lines 669

Saccharomyces cerevisiae YVH10 cells (ATCC) were used in antibody library generation and FACS 670 sort. Hela-hACE2 cells 23 were used in pseudovirus neutralization assay. Saccharomyces cerevisiae 671 EBY100 cells (ATCC) were used in RBD mutant library generation and FACS sort. Human 672

HEK293T cells (ATCC) were used for pseudovirus production. FreeStyle HEK293 cells 673 (ThermoFisher) were used for recombinant S protein production. Expi293F cells (ThermoFisher) 674

were used for monoclonal antibody and recombinant RBD production. Vero-E6 cells (ATCC) were 675 used for live virus plaque assay. 676 677

Recombinant S and RBD production 678 SARS-CoV-1 (Genbank AAP13567) or SARS-CoV-2 (Genbank MN908947) S proteins were 679 transiently expressed in Freestyle 293F system (ThermoFisher) whereas RBD proteins were 680 expressed in the Expi293 system (ThermoFisher). In brief, S expression plasmids were 681 cotransfected with 40K PEI (1 mg/mL) at a ratio of 1:3. After incubation for 30 min at RT, transfection 682 mixture was added to Freestyle 293F cells at a density of approximately 1 million cells/mL. RBD 683 plasmids with His-Avitag were cotransfected with FectoPRO (Polyplus 116-010). SARS-CoV-2 684 RBD mutant plasmids were generated by Quikchange site-directed mutagenesis according to 685 manufacturer's instructions (Agilent, 210513) . Biotinylated proteins were made by co-transfecting 686

Avitagged RBD plasmids with a BirA expression plasmid and into Expi293 cells using FectoPRO. 687

After incubation for 10 min at RT, transfection mixture was added to Expi293 cells at a density of ~ 688 3 million cells/mL. After 24h of transfection, cells were fed with D-(+)-glucose solution and 300 mM 689 of sterile sodium valproic acid solution. After 5 days of transfection, supernatants were harvested 690 and filtered with a 0.22 µm membrane.The His-tagged proteins were purified with the HisPur Ni-691 NTA Resin (Thermo Fisher, 88222). After three columns of washing with 25 mM Imidazole (pH 7.4), 692

proteins were eluted with an elution buffer (250 mM Imidazole, pH 7.4) at slow gravity speed (~4 693 sec/drop). Eluted proteins were buffer exchanged and concentrated with PBS using Amicon tubes 694 (Millipore). The proteins were further purified by size exclusion chromatography (SEC) using 695

Superdex 200 (GE Healthcare). The selected fractions were pooled and concentrated. 696 697

For cryoEM, a stabilized version of CoV-2 S protein, HPM7 (hexaproline mutant 7) was used. The 698 design, expression and purification has been described previously 14 . Briefly, the S protein is 699 stabilized with 6 engineered proline residues 29 and an interprotomer disulfide between residues 705 700 and 883 of S2. HPM7 was expressed in HEK293F, and purified using a C-terminal 2x StrepTag 701 followed by size exclusion chromatography to isolate trimers. 702

Antibody production and purification 704

Monoclonal antibody was transiently expressed in the Expi293 system (ThermoFisher, A14635). In 705 brief, antibody HC and LC plasmids were co-transfected at a ratio of 1:2.5 with transfection reagent 706

FectoPRO (Polyplus 116-010). After 24 h of transfection, 300 mM of sterile sodium valproic acid 707 solution (Sigma-Aldrich, P4543) and 45% D-(+)-glucose solution (Sigma Aldrich, G8769-100ML) 708

were added to feed cells. using the Patch CTF application in cryoSPARC. Automated particle picking, particle extraction, and 733 initial 2D classifications were performed in cryoSPARC. Particles belonging to selected 2D classes 734

were then imported into Relion 3.1 51 . Interactive rounds of 3D classification and refinement, and 735 CTF refinements were performed for each dataset. To improve resolution of the antibody epitope and paratope, the best refinement from each dataset (Spike with 2 Fabs bound for both CC6.30 737 and CC6.33) was subjected to C3 symmetry expansion and focused classifications, using a 738 spherical mask around the expected Fab/RBD region of a single protomer and Relion 3D 739 classification without alignments. Particles containing density for Fab and RBD in the region of 740 interest were imported into cryoSPARC. Signal outside of the RBD and Fab Fv was subtracted, and 741 the subtracted particles were subjected to cryoSPARC local refinement ( Supplementary Fig.1-2) . 742

The non-symmetry expanded particles from the best Relion global refinements were imported into 743 cryoSPARC and subjected to a final round of non-uniform refinement 52 . Additionally, the CC6.33 744 dataset contained thousands of ligand-free HMP7 Spike particles which were also imported into 745 cryoSPARC for final non-uniform refinement, with or without symmetry ( Figure S2C and S2D ABodyBuilder 53 and fitted into the respective local refinement maps using UCSF Chimera 54 . 752

Coordinates for RBD with complete ridge were taken from PDB 7byr. The RBD:Fv models were 753 subjected to interactive cycles of manual and automated refinement using Coot 0.9 55 and Rosetta 56 . 754

Once a high map-to-model agreement was reaching, as measured by EMRinger 57 , and geometries 755 were optimized, as judged by MolProbity 58 , the models were fit into the non-uniform refinement full 756 trimer maps and combined with a Spike model refined into the ligand-free map (PDB 6vxx was used 757 as the initial model and HPM7 mutations were added manually in Coot following iterative rounds of 758

Rosetta relaxed refinement and Coot manual editing). The resulting Fv:trimer models were refined 759 in Rosetta. The Phenix software suite 59 was used for structure validation, and for editing and 760 preparation of PDB files for deposition. Final refinement statistics and PDB deposition codes for 761 generated models can be found in Supplementary Table 1 loops and the CDR1/2/3 mini-libraries were assembled into combinatorial heavy chain and light 767 chain libraries as previously described 33 . The libraries were displayed on the surface of yeast as 768 molecular Fab using the yeast display vector pYDSI containing the bidirectional Gal1-10 promoter. 769

The heavy chain contains a C-terminal V5 epitope tag and the light chain contains a C-terminal C-770 myc epitope tag to assess the amount of Fab displayed on the surface of the yeast. The HC library 771 was generated by cloning the HC CDR1/2/3 library into the vector containing the wildtype light chain 772 by homologous recombination, and the LC library was generated by doing the inverse. The 773

combinatorial H/L library was generated by amplifying the HC and LC sequences with primers 774 overlapping in the Gal1-10 promoter. The recovered Gal-HC and Gal-LC fragments were ligated via Gibson assembly and amplified. The resulting LC-Gal1-10-HC product was cloned into an empty 776 pYDSI vector by homologous recombination 33 . 777 778

Yeast library labeling and sorting 779

After yeast transformation, yeast cells were passaged 1:20 the following day. Cells were then 780 induced at OD600 = 1.0 overnight at 30°C in SGCAA induction medium (20 g galactose, 1 g glucose, 781

6.7 g yeast nitrogen base without amino acid, 5 g bacto casamino acids, 5.4 g Na2HPO4, 8.56 g 782 NaH2PO4!"2O, 8.56 mg uracil in 1 L deionized water, pH 6.5, and sterilize by filtration). For each 783 library, in the first round of selection, 5 x 10 7 of yeast cells were stained per sample. In the second 784 to final round of selection, 1 x 10 7 cells were stained. Yeast cells were firstly spun down and washed 785

with PBSA (PBS + 1% BSA), then incubated with biotinylated SARS-CoV-2 RBD or S or HEK cell 786 membrane protein at several non-depleting concentrations respectively for at least 30 min at 4°C. 787

After washing, yeast cells were stained with FITC-conjugated chicken anti-C-Myc antibody 788 (Immunology Consultants Laboratory, CMYC-45F), AF405-conjugated anti-V5 antibody (made in 789 house), and streptavidin-APC (Invitrogen, SA1005) in 1:100 dilution for 20 min at 4 °C. After 790 washing, yeast cells were resuspended in 1 mL of PBSA and loaded on BD FACSMelody cell sorter. 791

Top 5-10% of cells with high binding activity to a certain SARS-CoV-2 RBD labeling concentration 792

were sorted and spun down. Sorted cells were expanded in 2 mL of synthetic drop-out medium 793 without tryptophan (Sunrise, 1709-500) supplemented with 1% Penicillin/Streptomycin (Corning,  794 30-002-C) at 30°C overnight. 795

Size exclusion chromatography analysis 797

The antibodies were analyzed by size exclusion chromatography using the 1260 Infinity II (Agilent). 798 15 uL of each antibody at 2 mg/mL was injected into the TSKgel SuperSW mAb column (Tosoh) 799 with the flow rate of 1 mL/min. 800 801

Pseudovirus neutralization assay 802

Pseudovirus was generated as described previously 23 . In brief, MLV gag/pol backbone (Addgene,  803 14887), MLV-CMV-Luciferase plasmid (Addgene, 170575), and SARS-CoV-2-d18 (Genbank 804 MN908947) or SARS-CoV-1-d28 (Genbank AAP13567) or SARS-CoV-2 VOC spike plasmid were 805 incubated with transfection reagent Lipofectamine 2000 (Thermo Fisher, 11668027) following 806 manufacturer's instructions for 20 min at RT. Full-spike mutations were introduced by overlapping 807 extension polymerase chain reaction (PCR) to generate mutated spikes of circulating SARS-CoV-808 2 VOC, i.e., B. In the neutralization assay, antibody samples were serially diluted with complete DMEM medium 815 (Corning, 15-013-CV) containing 10% FBS (Omega Scientific, FB-02), 2 mM L-Glutamine (Corning,  816 25-005-Cl), and 100 U/mL of Penicillin/Streptomycin (Corning, 30-002-C). 25 µL/well of diluted 817 samples were then incubated with 25 µL/well of pseudotyped virus for 1 h at 37 °C in 96-well half-818 area plates (Corning, 3688) . After that, 50 µL of Hela-hACE2 cells at 10,000 cells/well with 20 µg/mL 819 of Dextran were added onto each well of the plates. After 48 h of incubation, cell culture medium 820 was removed, luciferase lysis buffer (25 mM Gly-Gly pH 7.8, 15 mM MgSO4, 4 mM EGTA, 1% 821 Triton X-100) was added onto cells. Luciferase activity was measured by BrightGlo substrate 822 (Promega, PR-E2620) according to the manufacturer's instructions. mAbs were tested in duplicate 823 wells and independently repeated at least twice. Neutralization IC50 values were calculated using 824 "One-Site Fit LogIC50" regression in GraphPad Prism 8.0. 825 826 Authentic SARS-CoV-2 neutralization assay 827

Vero-E6 cells were seeded in 96-well half-well plates at approximately 8000 cells/well in a total 828 volume of 50 µL complete DMEM medium the day prior to the addition of antibody and virus mixture. 829

The virus (500 plaque forming units/well) and antibodies were mixed, incubated for 30 minutes and 830

added to the cells. The transduced cells were incubated at 37°C for 24 hours. Each treatment was 831 tested in duplicate. The medium was removed and disposed of appropriately. Cells were fixed by 832

immersing the plate into 4% formaldehyde for 1 hour before washing 3 times with phosphate 833 buffered saline (PBS 

Neutralization of mAbs against VOCs

Antibody started at 20 ug/mL and was tested in duplicates. b, Neutralization curves of REGN10987

CC6.33 and eCC12.1.7 against Wuhan-1 as well as VOCs

Identification of SARS-CoV-2 RBD escape mutants using yeast 1007 screening. Related to Fig.5. a, Gating strategy and gates set for the escape mutant analysis for 1008

Heatmap showing predicted RBD (residues 333-527) escape 1010 mutants for CC12.1 and eCC12.1.4 in yellow to burnt orange with varying levels of confidence 1011 according to the key. 1012 1013 1014 Supplementary Fig.8. Supplemental Animal Protection Studies, CC12.1 and eCC12.1.6. 1015 Related to Fig.6. a, Weight trends of all groups included in the CC12.1 vs eCC12.1.6 prophylactic 1016 protection study. b, Weights of animals at time of challenge (Day 0) compared to weights at time of 1017 sacrifice (Day 7). c, Serum human IgG concentration at time of infection (Day 0) compared to 1018 sacrifice (Day 7). d, Percent weight loss by day compared to weights recorded at time of infection 1019

Supplemental Animal Protection Studies, CC6.33 and eCC6.33.3. 1025 Related to Fig.6. a, Weight trends of all groups included in the CC6.33 vs eCC6.33.3 protection 1026 study. b, Weights of animals at time of challenge (Day 0) compared to weights at time of sacrifice 1027 (Day 7). c, Serum human IgG concentration at time of infection

Percent weight loss by day compared to individual weights recorded at time of infection at 1029 day 0

Supplementary Fig.10. Neutralization of enAbs against the Omicron variant

Neutralization curves of eCC6.33, eCC6.30, and eCC12.1 variants against pseudotyped Omicron 1035 VOC. Parental antibodies were highlighted in black whereas enAbs were in grey

835 100 µl of 3% bovine serum albumin (BSA) was added, followed by room temperature (RT) 836 incubation at 2 hours. 837 838 A mix of primary antibodies consisting of CC6.29, CC6.33, CC6.36, CC12.23, CC12.25 23 in equal 839 amounts for detection. The primary antibody mixture was diluted in PBS/1% BSA to a final 840 concentration of 2 µg/ml. The blocking solution was removed and washed thoroughly with wash 841 buffer (PBS with 0.1% Tween-20). The primary antibody mixture, 50 µl/well, was incubated with the 842 cells for 2 hours at RT. The plates were washed 3 times with wash buffer. 843 844Peroxidase AffiniPure Goat Anti-Human IgG (H+L) secondary antibody (Jackson 845 ImmunoResearch, 109-035-088) diluted to 0.5 µg/mLl in PBS/1% BSA was added at 50 µL/well 846 and incubated for 2 hours at RT. The plates were washed 6 times with wash buffer. HRP substrate 847 (Roche, 11582950001) was freshly prepared as follows: Solution A was added to Solution B in a 848 100:1 ratio and stirred for 15 minutes at RT. The substrate was added at 50 µL/well and 849 chemiluminescence was measured in a microplate luminescence reader (BioTek, Synergy 2). 850 851The following method was used to calculate the percentage neutralization of SARS-CoV-2. First, 852we plotted a standard curve of serially diluted virus (3000, 1000, 333, 111, 37, 12, 4 , 1 PFU) versus 853 RLU using four-parameter logistic regression (GraphPad Prism 8.0). 854 855Recombinant protein ELISA 856 6x-His tag antibodies (Invitrogen, MA1-21315) were coated at 2 µg/mL in PBS onto 96-well half-857 area high binding plates (Corning, 3690) overnight at 4˚C. After washing and blocking, 1 µg/mL of 858 his tagged recombinant SARS-CoV-2 (or SARS-CoV-1) RBD or S proteins were diluted in PBS with 859 1% BSA and incubated for 1 h at RT. After washing, serially diluted antibodies were added in plates 860 and incubated for 1 h at RT. After washing, alkaline phosphatase-conjugated goat anti-human IgG 861Fcγ secondary antibody (Jackson ImmunoResearch, 109-055-008) was added in 1:1000 dilution 862and For conventional kinetic/dose-response, listed antibodies were captured to 50-100 RU via Fc-893 capture on the active flow cell prior to analyte injection. A concentration series of SARS-CoV-2 RBD 894 was injected across the antibody and control surface for 2 min, followed by a 5 min dissociation 895 phase using a multi-cycle method. Regeneration of the surface in between injections of SARS-CoV-896 2 RBD was achieved by a single, 120s injection of 3M MgCl2. Kinetic analysis of each reference 897 subtracted injection series was performed using the BIAEvaluation software (Cytiva). All 898 sensorgram series were fit to a 1:1 (Langmuir) binding model of interaction. 899 900

Yeast display plasmids pJS697 and pJS699 used for surface display of Wuhan-Hu-1 S RBD N343Q 902 were previously described 60 . Using these plasmids, 119 surface exposed positions on the original 903Wuhan-Hu-1 S RBD N343Q (positions 333-537) were mutated to every other amino acid plus stop 904 codon using degenerate NNK primers using comprehensive nicking mutagenesis 61 exactly as 905 previously described 36,62 . Two tiles were constructed for compatibility with 250bp paired end Illumina 906 sequencing (tile 1: positions 333-436; tile 2: positions 437-527). Libraries were transformed into S. 907cerevisiae EBY100 and stocks of 1e8 viable yeast cells in 1 mL were stored in yeast storage buffer 908 (20 w/v % glycerol, 200 mM NaCl, 20 mM HEPES pH 7.5) at -80°C. Library coverage was confirmed 909 by 250 bp paired end Illumina deep sequencing, with statistics reported in Supplementary Table 5 . 910 S RBD escape mutants are identified by a competitive assay between a nAb and soluble ACE2 as containing 1g/L BSA) and then 3x10 7 induced EBY100 yeast cells displaying S RBD were labelled 914 with 10 !g/ml nAb IgG for 30min at room temperature with mixing by pipetting every 10min in PBSA. 915The same cells were labelled with 75 nM chemically biotinylated ACE2, in the same tube, for 30min 916 at room temperature in PBSA with mixing by pipetting every 10 min. of 6 CDR loops, binding affinity against SARS-CoV-2 RBD, SARS-CoV-1/2 neutralization potency 1048 and maximum percentage of neutralization (MPN). Mutations of engineered antibodies at CDR 1049 loops were highlighted in red. Antibodies were captured via Fc-capture to an anti-human IgG Fc 1050 antibody and varying concentrations of SARS-CoV-2 RBD were injected using a multi-cycle 1051 method. Association and dissociation rate constants calculated through a 1:1 Langmuir binding 1052 model using the BIAevaluation software. Neutralization assay was performed using pseudotyped 1053 SARS-CoV-1 and SARS-CoV-2 with Hela-hACE2 cell line. All antibodies were tested in duplicates. Antibodies were captured via Fc-capture to an anti-human IgG Fc antibody and varying 1061 concentrations of Wuhan-1 RBD or SARS-CoV-2 RBD variants were injected using a multi-cycle 1062 method. Association and dissociation rate constants calculated through a 1:1 Langmuir binding 1063 model using the BIAevaluation software. Antibodies that did not bind to RBD variants or their binding 1064 curves did not fit into the model were shown as no values in the table. 1065 S RBD N343Q (Wuhan-1)