Mutational escape from the polyclonal antibody response to SARS-CoV-2 infection is largely shaped by a single class of antibodies

Monoclonal antibodies targeting a variety of epitopes have been isolated from individuals previously infected with SARS-CoV-2, but the relative contributions of these different antibody classes to the polyclonal response remains unclear. Here we use a yeast-display system to map all mutations to the viral spike receptor-binding domain (RBD) that escape binding by representatives of three potently neutralizing classes of anti-RBD antibodies with high-resolution structures. We compare the antibody-escape maps to similar maps for convalescent polyclonal plasma, including plasma from individuals from whom some of the antibodies were isolated. The plasma-escape maps most closely resemble those of a single class of antibodies that target an epitope on the RBD that includes site E484. Therefore, although the human immune system can produce antibodies that target diverse RBD epitopes, in practice the polyclonal response to infection is dominated by a single class of antibodies targeting an epitope that is already undergoing rapid evolution.

have previously determined escape maps 30,33,35,36 , including class 4 antibodies, which bind to the RBD outside the receptor-binding motif and are generally less potently neutralizing 17,19,26,31 . Antibodies with similar escape mutations are located close to one another in the multidimensional scaling projection, and antibodies with very distinct escape mutations are far apart ( Figure 1C ). The projection shows that while there is some clustering by structural class in space of escape, the antibodies are continuously distributed. For instance, the class 3 antibody C110 also selects escape at some "class 2 sites," such as F490 and S494, and so C110 is located somewhat between the class 2 and 3 antibodies in the multidimensional scaling projection ( Figure 1C ).

RBD mutations reduce antibody binding at only a subset of contact sites
We took advantage of the availability of high-resolution structures to compare the sites of escape mutations to the structural contacts between the antibodies and RBD. Most mutations that escape antibody binding are at sites in the RBD that directly contact the antibody; these sites are highlighted in gray in Figure 1B , with a structural contact defined as any non-hydrogen atom within 4Å. To visualize the escape mutations in a structural context, we mapped the extent of escape at each site to the structure of the antibody-bound RBD 25,28,29 ( Figure 2A, S5A; interactive, zoomable versions at https://jbloomlab.github.io/SARS-CoV-2-RBD_MAP_Rockefeller ). All sites at which mutations strongly escape binding are in direct (<4Å) or proximal (4-8Å) contact with antibody in the resolved structures ( Figure 2B ).
However, not all antibody-contact sites had mutations that strongly escaped antibody binding ( Figure 2B, S5A ). There are several explanations. First, our approach maps functional antibody escape mutants that retain proper RBD folding and bind to ACE2 with ≥1% the affinity of the unmutated RBD 32 (see Methods ). Only 2,304 of the 3,819 possible amino-acid mutations to the RBD meet these criteria, and some sites have no tolerated mutations. For instance, G416 and R457 are both in the structural epitope of C105, but these sites have no tolerated mutations and thus do not appear in the escape map (sites with no tolerated mutations are indicated in dark gray in Figure 2A, S5 ). Second, sometimes mutations at antibody-contact sites simply do not strongly disrupt antibody binding 37 . For instance, site F486 is in structural contact with both C105 and LY-CoV016 and has many well-tolerated mutations, but mutations at this site more strongly affect the binding of LY-CoV016 than C105 ( Figure   1B, 2A, S5A ). Other examples include site R346, where nearly all mutations escape C135 but only charge-reversal mutations escape C110 ( Figure 1B, 2A, S5A ). Similarly, at site Q493, C144 and C002 are escaped by many mutations, but C121 is only escaped by Q493K/R ( Figure 1B, 2A, S5A ).
For some of the class 2 antibodies, the antibody makes a quaternary contact with an adjacent RBD in the context of spike trimer 25 ( Figure S5B ). Our yeast-display system assays antibody binding to isolated RBD, and so does not map escape mutations to quaternary contact sites and cannot inform on their importance for antibody binding.
The escape maps of polyclonal plasma often differ from those of monoclonal antibodies isolated from the same individual 4 . CC-BY 4.0 International license available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint this version posted March 18, 2021. The six antibodies newly mapped in this study were isolated from four different SARS-CoV-2 convalescent individuals ( Figure 3A ). Plasma was collected from these individuals at the same time that blood was collected for antibody isolation 21 . Because our escape-mutant mapping approach works for polyclonal sera or plasma in addition to monoclonal antibodies 18 , we mapped mutations that reduced binding by each of the four plasma plus one plasma sample without corresponding antibodies ( Figure 3B ; the plasma are prefixed with "COV-" to distinguish them from the antibodies which are prefixed with "C").
The plasma-escape maps shared many commonalities across all five individuals ( Figure 3B ), and generally resembled those from our prior study of a larger cohort of convalescent individuals 18 . In particular, mutations to sites F456 and E484 reduced binding for all five plasma samples ( Figure 3B ).
Mutations to site E484 are of special note as the E484K mutation is present in the emerging B.1.351, P.1, P.2, and B.1.526 SARS-CoV-2 lineages 1,2,4-6 and can reduce the neutralization titer of convalescent plasma by 3-fold or more 8,[10][11][12]18 . Mutations to other sites, such as G446-N450 and F486 also reduce binding for some of the plasma samples profiled here ( Figure 3B ), consistent with our prior study of a larger cohort 18 . These findings suggest that while there is some heterogeneity among which mutations reduce binding of different individuals' polyclonal plasma antibodies, there are also sites that are commonly targeted and should be monitored for antigenic evolution.
While the escape maps for the different plasma samples shared broad similarities, they often starkly differed from the escape maps of monoclonal antibodies isolated from the same individuals ( Figure 3B , compare plasma escape maps in logo plots with overlay bars showing sites of escape for antibodies from the same individual). For instance, mutations to R346 had the largest effects on binding by the class 3 antibody C135, but had little effect on the same individual's plasma (COV-72). Similarly, mutations at K417 had the largest effects on binding by the class 1 antibody C105, but had little effect on the corresponding plasma (COV-107). Conversely, mutations to site G496 reduced binding by the COV-21 plasma, but did not strongly affect any of the monoclonal antibodies in this study ( Figure 3B ).
Overall, the correlations between the sites at which mutations escaped binding for the monoclonal antibodies and their corresponding polyclonal plasma were highest for the class 2 antibodies, and lower for the other antibody classes ( Figure 3C ).

Class 2 antibodies contribute the most to the RBD escape maps of polyclonal plasma
To more broadly compare how antibodies of different classes contribute to the convalescent plasma escape maps, we used multidimensional scaling to project 22 antibodies and 28 polyclonal plasma into a two-dimensional space of binding-escape mutations ( Figure 4A ; the projection shows the 22 antibodies in Figure 1C , the 5 plasmas from Figure 3 , and 23 plasmas from a previously characterized larger cohort 18 ). The plasmas from both cohorts cluster together in the space of binding escape, but far from some of the antibodies ( Figure 4A ). In particular, most plasmas are positioned closest to the class 2 antibodies in the space of binding escape ( Figure 4A ).
To visualize the escape maps in terms of the RBD's three-dimensional structure, we projected onto the surface of the RBD the total escape at each site averaged across all antibodies in a class, or all convalescent plasma in a cohort ( Figure 4B ). Again, the polyclonal plasma most closely resembled the 5 . CC-BY 4.0 International license available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint this version posted March 18, 2021. ; https://doi.org/10.1101/2021.03.17.435863 doi: bioRxiv preprint class 2 antibodies. For instance, mutations to site E484 greatly reduced binding of both class 2 antibodies and polyclonal plasma ( Figure 4B ). Most of the class 1 antibody contributions to the plasma binding-escape maps came from sites shared with class 2 antibodies, such as F456 ( Figure 4B ), although note that mutations to F456 often do not strongly reduce plasma neutralization 18 . Consistent with the   lesser contributions of class 1 antibodies, mutations at the class 1 site K417 had little effect on plasma binding, and others have found that the K417N mutation alone has a minimal-to-modest effect on plasma neutralization 10,13,15,16 . Sites of escape from class 3 antibodies (e.g., G446) had visible effects on the polyclonal plasma, but again less so than for class 2 antibody sites ( Figure 4B ). However, in a prior larger study 18 , we found that mutations to the 443-450 loop strongly reduced binding of plasma from a minority of individuals, consistent with a few plasma falling closer to class 3 than class 2 antibodies in the space of binding-escape mutations ( Figure 4A ). Overall, these results show that class 2 antibodies usually dominate convalescent polyclonal plasma-although once a virus has accumulated mutations in class 2 epitopes (as has already occurred in some emerging lineages 1,2,4,6 ), then class 1 or 3 antibodies might dominate for the remaining anti-RBD antibody activity.

Escape maps are consistent with the RBD mutations that arise when virus is grown in the presence of monoclonal antibodies
We assessed how well our escape maps predicted the actual antibody-escape mutations that arose when virus was grown in the presence of the antibodies. Prior work selected viral escape mutants by passaging chimeric VSV encoding the SARS-CoV-2 spike in the presence of several of the monoclonal antibodies we mapped in this study 7 . We hypothesized that the mutations selected during viral passage would reduce antibody binding without impairing ACE2 binding affinity. Accordingly, we examined how all of the selected mutations affected both antibody binding (as measured in the current study) and ACE2 affinity (as measured in our prior deep mutational scanning 32 ). Figure 5A,B shows that in every case, the antibody-escape mutations selected in the virus were indeed among the single-nucleotide-change accessible amino-acid mutations that mediated the strongest escape from antibody binding without strongly impairing ACE2 affinity. Conversely, mutations that escaped antibody binding but were deleterious for ACE2 binding or RBD expression (e.g., E484V/A and mutations to sites 455 and 456) were not selected in the viral passaging. Therefore, our escape maps can be used in conjunction with prior data on the functional effects of RBD mutations to largely predict which escape mutations will arise when virus is grown in the presence of antibodies.
Weisblum and colleagues also tested many RBD point mutations for their effects on neutralization of chimeric VSV or lentiviral particles by the antibodies studied here 7 . There was generally good agreement between our escape maps and these previously measured effects of mutations on viral neutralization. Nearly all mutations with a >100-fold reduction in neutralization also had large effects in our escape maps, although in a few cases mutations with more moderate effects on neutralization were not prominent in the escape maps ( Figure 5C, S6A ). Previously, we and others have reported that single point mutations can reduce the neutralization of some plasma by >10-fold, although other plasma are largely unaffected by any single mutation 7,8,18 . For the plasma in this study, prior work found that no tested mutation had such large effects on neutralization 7 . However, the class 2 6 . CC-BY 4.0 International license available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made antibody-escape mutation E484K did reduce neutralization by COV47 plasma by approximately 5-fold 7 , concordant with the prominence of site 484 in that plasma's escape map ( Figure 3B, S6B ).

Mutations that reduce binding by class 1 and 2 antibodies are present in emerging viral lineages
To assess the extent that SARS-CoV-2 has already acquired mutations that reduce binding by each antibody class, we compared the total escape at each site averaged across all antibodies of that class to the frequency of mutations at the site among sequenced SARS-CoV-2 isolates in GISAID as of March 15, 2021 38 . Figure 6A shows that mutations at sites targeted by each antibody class are present at appreciable frequencies among sequenced SARS-CoV-2 isolates. In particular, sites K417 and E484 are the strongest sites of escape for class 1 and class 2 antibodies, respectively-and mutations at both these sites are present in a substantial number of sequenced viral isolates ( Figure 6A ). In contrast, mutations that escape class 3 antibodies are currently not as prevalent among sequenced isolates ( Figure 6A ).
While mutations have been observed at site K444 (the strongest site of escape from class 3 antibodies), these are at lower frequency than site 417 or 484 mutations ( Figure 6A ). Mutations at the class 3 site L452 are present at higher frequency, but L452 is only a moderate site of escape for this antibody class.
As expected from the fact that class 2 antibodies dominate convalescent polyclonal plasma, the natural frequency versus escape plots for the plasma closely resemble those for class 2 antibodies, with site E484 clearly the most concerning mutation ( Figure 6B ).
We also examined the presence of escape mutations for each antibody class in some key emerging viral lineages ( Figure 6C ). All these emerging viral lineages except B.  Figure 6C ). No viral lineages currently combine mutations that escape all three antibody classes-but the future emergence of a class 3 escape mutation in one of the lineages that already escape class 1 and 2 antibodies (B.1.351 or P.1) would be a worrying development, and should be monitored for closely.

Discussion
We comprehensively mapped all mutations that escape binding by three major classes of antibodies targeting the SARS-CoV-2 RBD, and compared these escape maps to those for convalescent polyclonal plasma. We find that a single antibody class (class 2) largely shapes how RBD mutations affect binding by polyclonal plasma, even for individuals from whom potent neutralizing antibodies of other classes were isolated. The dominance of class 2 antibodies in RBD-targeting portion of polyclonal plasma could be due in part to the twin facts that their epitope is exposed in both "up" and "down" conformations of the RBD, and that such antibodies are often generated from frequently observed germline genes (including VH1-2 , VH3-53 , VH1-69 ) 21,42-47 21,42-45,48 . Consistent with our work here showing that class 2 antibodies shape how mutations affect polyclonal plasma binding, mutations at the site that most strongly affects binding by this antibody class (E484) have arisen multiple times in emerging viral 7 . CC-BY 4.0 International license available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint this version posted March 18, 2021. ; https://doi.org/10.1101/2021.03.17.435863 doi: bioRxiv preprint lineages 1,2,4-6 . Therefore, our results show the importance of thinking about antigenic evolution in the context of different classes of antibodies that recognize different epitopes on the RBD.
Our work also sheds light on the extent to which it is functionally meaningful to subdivide anti-RBD antibodies into distinct classes based on their structurally defined epitopes. These structurally defined classes are inherently approximate groupings, since even antibodies with superficially similar structural epitopes bind to their antigens in subtly distinct ways 25 . Our comprehensive maps of binding escape mutations capture these subtle differences, and show that antibodies in the same structurally defined class can be differentially affected by the same mutation. Using the escape maps, we can visualize how the antibodies are related in terms of how their binding is functionally impacted by mutations at different RBD sites. These visualizations show that the arrangement of antibodies in the space of "viral escape" is indeed continuous, but that the class definitions based on structural analyses capture the high-level features of this arrangement since the structural footprint of an antibody largely determines which mutations most impact its binding. However, we suggest that in some cases, more Fortunately, no viral lineages currently combine these class 1 and 2 escape mutations with a class 3 escape mutation, although a moderate class 3 escape mutation (L452R) is present in the B.1.427/B.1.429 viral lineage that lacks other RBD escape mutations 3,49 . However, we suggest the appearance of a class 3 escape mutation on the background of a lineage that already has class 1 and 2 escape mutations such as K417N/T and E484K would be a worrying development, and should be monitored closely.

8
. CC-BY 4.0 International license available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

RBD deep mutational scanning library
Monoclonal antibody and polyclonal clonal plasma selection experiments were performed in biological duplicate using a deep mutational scanning approach 30  >95% present as single mutants on at least one barcode in the libraries. We previously used these libraries to measure the effect of all RBD mutations on yeast-surface RBD expression and ACE2 affinity 32 . As previously described, these libraries were sorted to eliminate variants that lose ACE2 binding prior to mapping the antibody-escape variants 30 .

9
. CC-BY 4.0 International license available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

FACS sorting of yeast libraries to select mutants with reduced binding by polyclonal plasma
Antibody labeling and selection was performed essentially as described in Greaney, et al. (2020) 30 . Specifically, 9 OD aliquots of RBD libraries were thawed and grown overnight at 30°C 275 rpm in 45mL SD-CAA (6.7 g/L Yeast Nitrogen Base, 5.0 g/L Casamino acids, 1.065 g/L MES, and 2% w/v dextrose). Libraries were diluted to an OD of 0.67 in SG-CAA+0.1% dextrose (SD-CAA with 2% w/v galactose and 0.1% w/v dextrose in place of 2% dextrose), and incubated for 16-18 hours at room temperature with mild agitation to induce RBD surface expression. Antibody-escaped cells collected per sample into SD-CAA supplemented with 1% w/v BSA and grown overnight in 1.5mL SD-CAA + 100 U/mL penicillin + 100 µg/mL streptomycin at 30°C 275 rpm. (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made Escape fractions were computed as described in   30 , with minor modifications as noted below.
For each antibody selection, we then computed the "escape fraction" for each barcoded variant using the deep sequencing counts for each variant in the original and antibody-escape populations and the total fraction of the library that escaped antibody binding via the formula provided in   30 . These escape fractions represent the estimated fraction of cells expressing that specific variant that falls in the antibody escape bin, so a value of 0 means the variant is always bound by antibody and a value of 1 means that it always escapes antibody binding. We then applied a computational filter to remove variants with low sequencing counts or highly deleterious mutations that might cause antibody escape simply by leading to poor expression of properly folded RBD on the yeast cell surface. Specifically, we ignored all variants with pre-selection sequencing counts that were lower than the counts for the 99th percentile of the stop-codon containing variants because stop codon variants are largely purged by the earlier sorts for RBD expressing and ACE2-binding variants and so any residual presence provides an indication of low-count "noise." Next, we removed any variants that had poor RBD expression or ACE2 binding, or contained mutations that individually cause poor RBD expression and ACE2 binding to eliminate misfolded or non-expressing RBDs. Specifically, we removed variants that had (or contained mutations with) ACE2 binding scores < -2.35 or expression scores < -1, using the variant-and mutation-level deep mutational scanning scores 32 . Note that these filtering criteria are slightly more stringent than those used in binding cutoff of -2.35 is used to represent the binding of RaTG13 to human ACE2 32 , which possesses the lowest known affinity capable of mediating cell entry 53 . The RBD expression cutoff of -1 is used to eliminate mutations that have as large an expression deficit as mutations to core disulfide residues. 2,034 of the 3,819 possible RBD amino acid mutations passed these filtering steps and were included in our escape maps. All previously reported escape mapping data 18,30,33,36 were reanalyzed in this study with the parameters listed above. A markdown rendering of the computation of the variant-level escape fractions and the variant filtering is at https://github.com/jbloomlab/SARS-CoV-2-RBD_MAP_Rockefeller/blob/main/results/summary/counts_to_scores.

md .
Because some library variants contain multiple amino acid mutations, we next deconvolved variant-level escape scores into escape fraction estimates for single mutations using global epistasis models 54 implemented in the dms_variants package, as detailed at ( https://jbloomlab.github.io/dms_variants/dms_variants.globalepistasis.html ). In this fitting, we excluded variants that contained mutations that were not seen as either single mutants or in at least two multiple-mutant variants.
We then computed the estimated effect of each mutation as the impact of that mutation on the "observed phenotype" scale transformation of its "latent phenotype" as computed using the global epistasis models, and applied a floor of zero and a ceiling of 1 to these escape fractions. All of the above analysis steps were performed separately for each of the duplicate mutant libraries. We then only retained mutations that passed all of the above filtering and were measured in both libraries or had at least two-single mutant variant measurements in one library. The reported scores throughout the paper are the average across the libraries; these scores are also in 11 . CC-BY 4.0 International license available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made  Figure S2 . A markdown rendering of the computation that computes these mutation-level escape fractions is at https://github.com/jbloomlab/SARS-CoV-2-RBD_MAP_Rockefeller/blob/main/results/summary/scores_to_frac_es cape.md .
For plotting and analyses that required identifying RBD sites of "strong escape" (e.g., choosing which sites to show in logo plots in Figure 1A or Figure 3B or label in Figure 2B ), we considered a site to mediate strong escape if the total escape (sum of mutation-level escape fractions) for that site exceeded the median across sites by >10 fold, and was at least 10% of the maximum for any site. A markdown rendering of the identification of these sites of strong escape is at https://github.com/jbloomlab/SARS-CoV-2-RBD_MAP_Rockefeller/blob/main/results/summary/call_strong_escap e_sites.md .

Comparison of mutation escape fractions to previously measured neutralization concentrations
In Figure 5C and Figure S6 , mutation-level antibody-escape fractions measured in this study are compared to previously measured neutralization titers (inhibitory concentration 50%, IC50) of the same monoclonal antibodies and polyclonal plasma against some RBD point-mutants 7

Analysis of mutations in circulating human SARS-CoV-2 strains
For the analysis in Figure 6 , all 765,455 spike sequences on GISAID 38 as of March 15, 2021 were downloaded and aligned via mafft 55 . Sequences from non-human origins and sequences containing gap or ambiguous characters were removed, as were sequences with extremely high numbers of RBD mutations relative to other sequences, leaving 679,502 retained sequences. All RBD amino-acid mutations were enumerated compared to the reference Wuhan-Hu-1 SARS-CoV-2 RBD sequence (Genbank MN908947, residues N331-T531). We acknowledge all contributors to the GISAID EpiCoV database for their sharing of sequence data (all contributors listed at: https://github.com/jbloomlab/SARS-CoV-2-RBD_MAP_Rockefeller/blob/main/data/gisaid_hcov-19_acknowledge ment_table_2021_03_15.pdf ).

Data visualization
The static logo plots in the paper were created using dmslogo ( https://jbloomlab.github.io/dmslogo/ ) version 0.6.2; a markdown rendering of the code that creates these logo plots is at https://github.com/jbloomlab/SARS-CoV-2-RBD_MAP_Rockefeller/blob/main/results/summary/escape_profiles. md . For each plasma, the y-axis is scaled to be the greatest of (a) the maximum site-wise escape metric observed for that plasma, (b) 20x the median site-wise escape fraction observed across all sites for that plasma, or (c) an absolute value of 1.0 (to appropriately scale plasma that are not "noisy" but for which no mutation has a strong effect on plasma binding).
In Figure S4

12
. CC-BY 4.0 International license available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The multidimensional scaling in Figure 1C and Figure 4A that projects the antibodies into a two-dimensional space of escape mutations was performed using the Python scikit-learn package. We computed the similarity and dissimilarity in the escape maps between each pair of antibodies, then performed metric multidimensional scaling with two components on the dissimilarity matrix exactly as defined in Greaney et al.
(2021) 30 . In Figure 1C , the multidimensional scaling shows antibodies as pie charts colored proportionally to the total squared site escape that falls into that RBD structural region. The code that generates these logo plot  Figure S5B , 7K8T is used instead of 7K8S to illustrate the quaternary antibody epitope. For identifying contact sites to highlight in Figure 1B logo plots or to classify sites in Figure 2B as contact sites (within 4A of antibody) or antibody-proximal sites within 4-8A, the following PDBs were used: 6XCM and 6XCN for C105, 7K8S and 7K8T for C002, 7K8X and 7K8Y for C121, 7K90 for C144, 7K8Z for C135, and 7K8V for C110) 25,29 .
Structural distances were computed using the bio3d package in R 57 . Surface representations of the RBD for non-antibody-bound structures utilize PDB 6M0J 58 .

Declarations of Interests
The Rockefeller University has filed a provisional patent application related to SARS-CoV-2 monoclonal antibodies on which D.F.R. and M.C.N. are inventors. The Rockefeller University has applied for a patent relating to the replication-competent VSV/SARS-CoV-2 chimeric virus on which Y.W, F.S., T.H., and P.B. are inventors (US patent 63/036,124). The other authors declare no competing interests.

18
. CC-BY 4.0 International license available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint this version posted March 18, 2021. The y-axis is scaled separately for each plasma (see Methods ). When there are monoclonal antibodies isolated from the same individual, the total monoclonal antibody escape at each site is shown using the heat maps above the escape maps, with white indicating no effect and black indicating strong escape. (C) Correlation of plasma and monoclonal antibody escape for each plasma / antibody pair from the same individual. Each point in the scatter plots is a site, with the x-axis indicating the total escape at that site for the antibody and the y-axis indicating the total escape at that site for the plasma. Key sites are labeled. Pearson's R shown above each plot.
Colors in B, C reflect antibody classes as in Figure 1 .

19
. CC-BY 4.0 International license available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made   . The x-axis shows the escape fraction measured in the current study, and the y-axis shows the fold-change in IC50 for viral neutralization caused by that mutation, such that larger numbers correspond to greater reductions in neutralization sensitivity. For effects of all antibody-and plasma-binding-escape mutations on ACE2 binding and RBD expression, see Figure S4 . For each mutation's escape fraction compared to fold-change IC50 against each monoclonal antibody or polyclonal plasma tested in Weisblum et al. 7 , see Figure S6 .

21
. CC-BY 4.0 International license available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint this version posted March 18, 2021. ; convalescent plasma (present study, n=5)  23 . CC-BY 4.0 International license available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made  Figure S1 . Approach for mapping RBD mutations that reduce binding by monoclonal antibodies or polyclonal plasma, related to Figure 1. (A) The RBD is expressed on the surface of yeast. Flow cytometry is used to quantify both RBD expression (via a C-terminal MYC tag) and antibody binding to the RBD protein expressed on the surface of each yeast cell. A library of yeast expressing different RBD mutants were incubated with antibodies or plasma and binding was detected using a IgG or IgA+IgG+IgM secondary antibody for monoclonal antibodies or polyclonal plasma, respectively. We then used FACS to enrich for cells expressing RBD that bound reduced levels of antibody, and deep sequencing to quantify the frequency of each mutation in the initial and "antibody escape" cell populations. We quantified the effect of each mutation as the "escape fraction," which represents the fraction of cells expressing RBD with that mutation that fell in the "antibody escape" FACS bin. Escape fractions are represented in logo plots, with the height of each letter proportional to the effect of that amino-acid mutation on antibody binding. The site-level escape metric is the sum of the escape fractions of all mutations at a site. Note that both experimental and computational filtering steps were used to remove RBD mutants that were misfolded or completely unable to bind the ACE2 receptor (see Methods ).

24
. CC-BY 4.0 International license available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

27
. CC-BY 4.0 International license available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made   R346  D405  K417  D420  Y421  N440  K444  G446  G447  N448  Y449  N450  L452  Y453  L455  F456  N460  I472  Y473  A475  G476  V483  E484  G485  F486  N487  Y489  F490  Q493  S494  G496  G504   0 R346  D405  K417  D420  Y421  N440  K444  G446  G447  N448  Y449  N450  L452  Y453  L455  F456  N460  I472  Y473  A475  G476  V483  E484  G485  F486  N487  Y489  F490  Q493  S494  G496  G504   0  28 . CC-BY 4.0 International license available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made CC-BY 4.0 International license available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint this version posted March 18, 2021. ; Figure S5. Visualization of the maximum escape at a site mapped onto cartoon representations of antibody-bound RBD, related to Figure 2. (A) Mapping of maximum antibody escape at a site to the antibody-bound RBD. Antibody-contact sites on the RBD (within 4Å) are shown as spheres. Sites with no escape measurements due to excessive functional constraint on the site are shown in dark gray. Each site is colored according to the maximum escape fraction of any mutation at that site (whereas Figure 2 shows site total escape), scaled from white (no escape) to red (maximum escape for any mutation for that antibody). Inset panels at right indicate key RBD-antibody interactions where mutations to the indicated RBD site disrupt antibody binding. RBD color scale indicates site total escape, as in Figure 2A . (B) Visualization of class 2 antibody quaternary epitopes.
The total escape at each site is mapped onto the surface of the Fab-bound RBD as in Figure 2A , with white indicating no escape, and red indicating the site with the most escape from that antibody. Sites where no mutations are tolerated are indicated in dark gray. Antibody quaternary contact sites are shown in periwinkle.
The C144 antibody binds to spike trimer in the "all RBDs down" conformation and forms a quaternary epitope that bridges across two adjacent RBDs by binding to a hydrophobic RBD cavity at the base of the N343 N-linked glycan. The C002 and C121 antibodies, when bound to a "down" RBD, can form a quaternary epitope with an adjacent "up" RBD. The "up" RBD also contacts another C121 Fab 25 . Our yeast-display system utilizes monomeric RBD and therefore does not map escape mutations to quaternary contact sites. These results thus cannot be used to determine the importance of the quaternary sites for antibody binding. Previous work, however, has shown that the V367F RBD mutation to the C144 quaternary epitope does not affect neutralization of pseudotyped lentivirus by C144 7 . See Methods for PDB accession codes used to generate structural representations.

31
. CC-BY 4.0 International license available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made   7 . The y-axis shows the fold-change in IC50 compared to the Wuhan-Hu-1-like spike, such that larger numbers are greater reductions in neutralization sensitivity. Mutations that had IC50s at or above the limit of detection are indicated as triangles. Points are colored according to their mutation escape fraction (

32
. CC-BY 4.0 International license available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The file gives the "escape fraction" for each mutation, as well as the total escape fraction at each site and the maximum escape fraction for any mutation at the site. The file is also available on GitHub at https://github.com/jbloomlab/SARS-CoV-2-RBD_MAP_Rockefeller/blob/main/results/supp_data/all_sa mples_raw_data.csv . Table 2. Information on FACS sorting to select cells expressing RBD mutants with reduced binding by antibodies or plasma, related to Figure 1 and Figure S1.

Supplementary
The file gives the number of antibody-escaped cells collected per selection for each replicate library and the percent of RBD+ cells in the antibody-escape gate for each selection, and the exact dilution used for each plasma selection. The file is also available on GitHub at https://github.com/jbloomlab/SARS-CoV-2-RBD_MAP_Rockefeller/blob/main/data/SupplementaryTabl e2.xlsx .

33
. CC-BY 4.0 International license available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint this version posted March 18, 2021. ; https://doi.org/10.1101/2021.03.17.435863 doi: bioRxiv preprint . CC-BY 4.0 International license available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made