Cross-reactive antibodies against human coronaviruses and the animal coronavirome suggest diagnostics for future zoonotic spillovers

A library of 12,924 hCoV and aCoV peptides

We designed a PhIP-Seq library (experimental outline is shown in Fig. 1a) covering all open reading frames of human and animal CoVs as 64 amino acid (aa) sections with 20 aa overlaps between adjacent peptides (Fig. 1b, Data file S1). The sequences of 48 aCoVs were obtained from the NCBI reference genome (RefSeq, April 2020) database and the sequence of another bat CoV related to SARS-CoV-2 (24) was included in addition. These strains broadly covered all groups of α, β, γ, and δ CoVs with their phylogeny illustrated in Fig. S1. The CoV antigen library consisted in total of 12,924 peptides, with hCoVs representing ~20% of peptides and aCoVs ~80% of CoV antigens (Fig. 1b, Data file S2). The 49 aCoVs contained 11 strains infecting birds, 19 strains infecting bats, and 19 more CoVs infecting other mammals such as rabbits, rodents, bovines etc. (see Data file S1 for a full list of aCoVs and hosts). For SARS-CoV-2, in addition to the reference genome, variants deposited in the NCBI database as of mid-April of 2020 were included. The antigen library also includes positive controls which confirmed detection of antibody responses against viruses previously reported to elicit population-wide immunity (21) and negative controls, that did not show substantial binding (Fig. S2).

We tested IgG antibody binding against this CoV library with 260 pre-pandemic serum samples of individuals unexposed to SARS-CoV-2, that had been collected in 2013 to 2016 (25, 26) (Fig. 1c). These antibody repertoires were compared to 269 samples of recovered COVID-19 patients obtained in April and May 2020. The serum samples were mixed individually with the phage library displaying the CoV antigens (Fig. 1a). Phages bound by antibodies were immunoprecipitated and unbound phages washed away. The bound phages were PCR amplified and sequenced. Thereby we obtained for each library variant in each sample a read count after immunoprecipitation. These read counts were compared to the ‘input’ read counts of the phage material before mixing with serum samples. We employed a Generalized Poisson distribution approach previously reported (27) to calculate p-values for significance of the enrichment of each library variant in each sample. These p-values were filtered in each sample by strict Bonferroni correction (<0.05) to counteract the problem of multiple hypothesis testing (see materials and methods for details). In total we have assayed for approximately five million antibody-peptide interactions (12,924 hCoV peptides in each of the 529 individuals). On average of 114 CoV peptides were significantly bound per unexposed individual and 189 CoV peptides per recovered COVID-19 patient. Most analyses were based on antibody responses against 579 CoV peptides from different CoV strains and proteins, shared by more than five percent of either group (Data file S3). Out of these, 190 peptides showed significantly different abundances between the groups (Data file S3). Bound antigens included peptides originating from hCoVs (Fig. 2, Fig. 3) as well as aCoVs (Fig. 4), which are discussed sequentially in the following sections.

Antibody repertoires against hCoVs and cross-reactivities

As expected, sera of recovered COVID-19 patients bound significantly more peptides of SARS-CoV-2 than sera of unexposed individuals (Fig. 2a). Also many peptides of SARS-CoV, MERS-CoV, as well as the seasonal hCoVs OC43 and HKU1 were significantly more frequently bound in recovered patients than unexposed individuals, in line with cross-reactivity previously reported (13).

Unexposed individuals showed abundant antibody responses against all seasonal hCoVs (Fig. 1e-h): Peptides of hCoV-NL63 were significantly bound in up to 88% of unexposed individuals, peptides of hCoV-HKU1 in up to 87%, peptides of hCoV-229E in up to 83%, and peptides of hCoV-OC43 in up to 46% (Fig. 1e-h, Data file S2). The same peptides were bound at similar frequencies in recovered COVID-19 patients and originated mostly from S- or N-proteins (Fig. 3, Data file S3) with these epitope resolved results being in agreement with previous studies on the seroprevalence of seasonal hCoVs using ELISAs (28).

Recovered COVID-19 patients’ sera showed an overrepresentation of several peptides of SARS-CoV-2 that showed no binding or binding at very low percentages in unexposed individuals (Fig. 3b). Twelve peptides passed FDR (Benjamini Hochberg false discovery rate (29), a method to correct for multiple hypothesis testing) correction for being significantly different between the two groups of individuals (for the SARS-CoV-2 reference genome, additional peptides of SARS-CoV-2 variants passed significance thresholds and are listed in Data file S3). While nearly all COVID-19 patients showed binding against at least one peptide in S- or N-proteins and some peptides being bound in up to 81.4% of recovered patients (Fig. 1b), no convergence of antibody responses against the same peptide were detected in all individuals (which is limited to the interactions detectable with this PhIP-Seq library, see the discussion section). This finding differs from near universal recognition of some viral epitopes previously observed by PhIP-Seq/VirScan for other human viruses (12) and replicated with controls in this study (Fig. S2), suggesting that the antibody response against SARS-CoV-2 can exhibit substantial inter-individual variability.

COVID-19 serum samples also showed common binding against SARS-CoV (Fig. 2c), to which it is unlikely that these individuals had been exposed, indicating detection of cross-reactivity of antibodies targeting SARS-CoV-2 (14, 15). Most notably three SARS-CoV spike peptides had significantly enriched binding in up to 88.1% of COVID-19 recovered individuals compared to up to 6.9% of unexposed individuals (Fig. 3c, Data file S3). A non-structural protein (NSP2) of SARS-CoV was even bound in 30% of unexposed individuals and 36% of recovered, possibly owing to higher conservation of such NSPs underlying less selective pressure than S- and N-proteins mostly responsible for infectivity and targeted by neutralizing immune responses. Also three peptides from the MERS-CoV (Fig. 3d) S-protein were differentially enriched in the COVID-19 recovered cohort, passing FDR correction for significance of this difference (Fig. 3a, Data file S3).

Strikingly, cross-reactive responses from SARS-CoV-2 also extended to the spike proteins all four seasonal hCoVs (Fig. 3a) related to specific motifs (Fig. 3c,d). Three peptides from each hCoV-OC43 and hCoV-HKU1 were significantly differentially bound between unexposed and recovered individuals. One of these peptides arose from similar regions of the S protein (FP site, Fig. 3a) and also hCoV-229E and hCoV-NL63 peptides around this position were significantly differentially bound. Such cross-reactivity has been primarily reported for hCoV-OC43 (13). Our data suggests, that cross-reactivity arising from SARS-CoV-2 infection targets a similar motif in all human hCoVs (RSXIEDLLFXK, Fig. 3c, see Fig. S3a for an alignment of the complete bound peptides and Fig. S4 illustrating that the motif is exposed on the surface of SARS-CoV-2). Toward the C terminus of the spike protein, another region was bound at significantly different frequencies between recovered and unexposed individuals in SARS-CoV-2, SARS-CoV, MERS-CoV, hCoV-OC43, and hCoV-HKU1 (Fig. 3a). This longer region appeared to contain two distinct motifs (Fig. 3d, Fig. S3b, and Fig. S4 illustrating surface exposure on the SARS-CoV-2 spike protein). In the nucleocapsid protein, less clear motifs were apparent in regions bound by antibodies at similar levels between hCoVs (Fig. 3b and Fig. S3c,d) potentially owing to separate binding of smaller motifs by different antibodies (which may explain why some of the peptides were bound at different levels and others not Fig. 3b).

Antibody cross-reactivity extends toward animal coronaviruses

In addition to antibody binding to hCoV peptides, we detected also dozens of aCoV peptides significantly bound per individual (Fig. 4a). These peptides originated from aCoVs with diverse hosts including all three major groups: bats, other mammals (such as rodents), and birds. Recovered COVID-19 patients’ sera bound significantly more aCoV peptides than unexposed individuals when scoring these groups (Fig. 4a). Scoring differences on a strain level (Fig. S5) showed varying antibody responses, as observed for hCoVs (Fig. 2a): Peptides of some bat CoVs closely related to SARS-CoV-2 (Fig. S1) were highly significant for being more frequently bound in recovered COVID-19 patients, than in unexposed individuals (p-value <10⁻⁴). In contrast, peptides of bat CoVs such as a NL-63 related strain were not bound to a different extent between the two groups (p-value >0.05, Fig. S5). This finding is in agreement with the seasonal hCoV-NL63 also not exhibiting a significant difference (Fig. 1a). Similarly, for aCoVs of birds or other mammalian hosts, we observed some strains being bound by antibodies at similar levels in unexposed and recovered COVID-19 patients, and other strains bound differentially. For example aCoVs from rabbit, mouse rat, pig and cow as well as the bird night heron showed highly significant differences between the two groups, whereas aCoVs from felines, ferrets and camel did not show significant differences (Fig. S5). We also analyzed differences in antibody binding based on genetic differences of the aCoVs (Fig. S6). These results suggest that antibody responses against beta aCoVs show the greatest statistical difference between unexposed and recovered individuals (as expected from their closer relationship to the beta hCoV SARS-CoV-2). Differential responses were not limited to beta aCoVs with some antigens of alpha, gamma, and delta aCoVs also being significantly more frequently bound by antibodies in recovered COVID-19 patients than unexposed individuals (Fig. S6).

Relating these antibody responses to hCoVs, aCoV peptides amounted for approximately 80% of the initial library content while hCoVs represented about 20% (Fig. 4b, Fig. 1b). In recovered COVID-19 patients, antibody responses against SARS-CoV-2 were overrepresented, especially regarding peptides bound in larger fractions of the cohort (Fig. 4b). In unexposed individuals, aCoV peptides took up a larger fraction of bound peptides with seasonal hCoVs still being overrepresented. These results suggest that peptides of hCoVs are dominantly bound by human antibodies and aCoVs contribute to the detected antibody repertoire by cross-reactivity.

To pinpoint the potential cross-reactivity underlying the binding signal within aCoVs, we mapped the bound aCoV peptides to the SARS-CoV-2 proteome (summarized in Fig. 4c,d, see Data file S3 for alignment data for all peptides). Two clusters in the S protein near the FP site and the C terminus were apparent (Fig. 4c, Fig. S7a,b), which are similar to the cross-reactive regions observed in hCoVs (Fig. 3a,c,d; (13), accessible on the surface of the S protein Fig. S4). Near the FP site, the RSXIEDLLFXK motif observed from hCoVs (Fig. 3c) appeared in near identical form in aCoVs (Fig. 4e). The N-terminal cluster yielded for aCoVs even a clearer motif (FKEELDXXFKN, Fig. 4f) than the same region in hCoVs (Fig. 3d), possibly owing to the larger number of peptides aligned. In the N-protein, bound aCoV peptides clustered mostly in the RNA binding domain (Fig. 4d), exhibiting also a shared motif (Fig. 4 g and see Fig. S7c,d for full alignments). While similar peptides in the RNA binding domain of the N-protein of hCoVs are bound (Fig. 3c), the exact motif identified in aCoVs (Fig. 4 g) was not apparent. Furthermore, a C-terminal motif found in aCoVs (Fig. 4 g) did not have a direct equivalent in hCoVs (Fig. 3b). The two main spike motifs detected from aCoV peptides (Fig. 4e,f) also match sequences previously reported in a different cohort (13) suggesting that human antibodies raised upon SARS-CoV-2 infection can also cross-react with aCoVs. Additionally, cross-reactivity of human antibodies against aCoVs appears to arise also from seasonal hCoVs, as some S- and N- peptides are bound at similar levels in unexposed controls and recovered COVID-19 patients. These include rat/mouse CoV peptides around amino acid 600 in the S-protein (Fig. 4c) or bat peptides around amino acid 100 in the N-protein (Fig. 4d).

Patient-derived immunoglobulins with cross-reactivity potential

While we observed pronounced antibody cross-reactivity between SARS-CoV-2 and hCoVs as well as aCoVs (Fig. 2 to Fig. 4), the underlying mechanisms are unclear. Cross-reactivity is most likely to arise from polyclonal antibodies, yet, the occurrence of shared motifs (Fig. 3c,d, Fig. 4e-h) suggests potential cross-reactive recognition by a single antibody. To examine this possibility, we sequenced B cell immunoglobulin genes derived from recovered patients and generated monoclonal antibodies (30) for testing in PhIP-Seq. Since cross reactive binding to the receptor binding domain (RBD) is less likely to occur, we sequenced patient-derived B cells that show reactivity with the intact spike trimer that carries more conserved domains with cross-reactive potential. Spike-specific single memory B cells (CD19+, CC27+, IgG1+, Igκ+) were sorted from peripheral blood mononuclear cells of convalescent patients and subjected to PCR amplification of their immunoglobulin genes followed by gene sequencing (Fig. 5a). Analysis of the emerging immunoglobulin sequences revealed significant enrichment for VH3 and VK3 and these types of V genes were the most abundant paired chains (Fig. 5b,c). Based on the variable region (VDJ) sequences we clustered the clones to determine the clonal expansion of spike-specific memory B cells. This analysis revealed that most of the spike-binding memory B cells were not significantly expanded, and the majority of the recovered immunoglobulin sequences appeared only once (Fig. 5b). The average number of somatic mutations per cell was 16 and 10 for the heavy and light chain, respectively (Fig. 5d), which is slightly higher than the average mutation load of memory cells in healthy subjects (31). As expected, the number of mutations in the heavy chains was in correlation to the mutation load in the light chains (Fig. 5e) and the CDR3 lengths of the light and heavy chains (Fig. 5f) were similar to typical immunoglobulins in naive B cells (31). Comparing our findings to previous published data shows that the frequency of mutations was slightly higher in our cohort, however, the observed clonal expansion was lower (Table S 1). It is important to note, that most studies focused on the RBD target whereas we used the full spike trimer as a bait which carry multiple domains. We conclude that although robust clonal expansion was not detected in recovered patients, some of the mutated spike-specific memory cells are the product of antibody affinity maturation and may have broad binding reactivity.

Pan-specific mAbs can mediate cross-reactivity against hCoVs and aCoVs

In order to test the patient-derived immunoglobulins in PhIP-Seq, we cloned the recovered sequences into expression vectors and produced monoclonal antibodies (30). Previous studies found that most of the SARS-CoV-2 neutralizing antibodies are very similar to their germline configuration (32–35). In contrast, broadly reactive antibodies generated during chronic infection, such as in HIV-infected patients, are highly mutated (36). Furthermore, if SARS-CoV-2 activates pre-existing cross-reactive memory cells that were raised against other corona viruses, the emerging antibodies are expected to carry more mutations than those elicited in a primary response. Therefore, to maximize the probability for detection of broadly reactive mAbs, two highly mutated antibodies were chosen for expression as IgG1 (30) followed by spike-binding and PhIP-Seq analyses. In parallel, we also examined additional two antibodies that carried very few somatic mutations, WIS-A7 and WIS-A9 for cross-reactivity comparison. ELISA revealed four antibodies that showed detectable spike-binding activity (Fig. 6a). Among these antibodies, C1 and C3 mAbs carried a significant load of mutations and shared 88.7-90.1% VH sequence identity (Fig. 6b). Using our PhIP-seq assay, we found that both C1 and C3 bound significantly to several peptides of hCoVs (SARS-CoV-2, SARS-CoV, hCoV-OC43, hCoV-HKU1) as well as several aCoVs (including bovine, rodent, thrush (bird), rabbit, and bat as hosts) spike proteins found within our phage display library (Fig. 6c, left and middle panel, see Fig. S8 for a full list of peptides). In contrast a control monoclonal antibody did not show any significantly enriched peptides. C1 and C3 showed nearly identical binding to the same peptides, with slight differences for two peptides (Fig. 6c, right panel), suggesting that the additional mutations of C3 do not strongly impact cross-reactivity against our CoV antigen library. Consistently, a part of the SARS-CoV-2 peptide target was detected in a spike peptide array binding assay (Fig. 6d,e). The sequence of the binding site was very similar among the different strains of corona viruses and was composed of a consensus motif that is located in the S2 domain of the SARS-CoV-2 spike protein (Fig. 6f). This motif lies in a region different from frequently bound peptides identified in hCoVs (Fig. 3a) and aCoVs (Fig. 4c). In agreement with these results, the mAbs also bound the target peptide in ELISA (Fig. 6 g). Since, structurally, part of this peptide sequence is not exposed on the spike trimer surface, it is most likely that the mAbs bind only the N terminus of the peptide (Fig. 6e,h). A7 mAb bound a peptide that was derived from the SARS-CoV-2 spike protein and a very similar peptide that is part of a bat SARS like aCoV spike complex (Fig. S9a). This region of the SARS-CoV-2 spike protein (starting at amino acid 748), was a binding target in 35.3% of our recovered COVID-19 cohort, but not in our healthy control cohort (Fig. 3a). Although A9 was found to bind the spike protein in ELISA, our PhIP-Seq assay did not identify a linear binding target (Fig. S9b). This contradiction can be a result of A9 binding to a non-linear epitope or posttranslational modifications that are inadequately represented within our phage library (see also the discussion section below). Overall, for the highly mutated antibodies showed significant cross-reactivity, providing a proof of concept evidence for the existence that a single antibody can bind multiple corona viruses. Yet, elucidating the frequency of single cross-reactive antibodies in the general population will require further studies that involves pre- and post-exposure to SARS-CoV-2 analyses.

To examine if cross-reactive monoclonal antibodies support protective functional activity, we examined their capacity to neutralize pseudovirus that carried the SARS-CoV-2 spike protein. Both the C1 and C3 displayed poor neutralization activity compared with a control antibody that was previously described (35) (Fig. 6h). Typically, IgG1 antibodies support non-neutralizing activity as well, however, since C1 and C3 mAbs bind an internal epitope that is exposed only in the open conformation of the spike trimer, they might not be able to support these functions. Antibody-dependent cellular phagocytosis (ADCP) is a process wherein immune cells uptake a target through antibody-mediated interaction with Fc receptors (37, 38). To examine ADCP activity, we produced immune complexes by incubating spike-coated beads with monoclonal antibodies and tested the ability of THP-1 monocytes to uptake these targets. Both C1 and C3 monoclonal antibodies promoted the uptake of spike-coated beads by THP-1 cells whereas the control antibody did not contribute to this process (Fig. 6j). To validate if serum antibodies can support ADCP, we examined if patient-derived serum can promote phagocytic activity. We found that all of the patients have antibodies in their serum that can support ADCP (Fig. 6k). This antibody-mediated activity can be supported by few or multiple neutralizing and non-neutralizing antibodies. Collectively, we conclude that cross-reactive monoclonal antibodies derived from COVID-19 patients show functional activity.

To examine if the monoclonal antibodies were subjected to affinity maturation we examined if amino acid replacements in the complementary determining regions and in the framework regions of the V regions using the Baseline tool (39, 40). Similar to other antigen-experienced B cells, the mutation patterns in these sequences indicate that the B cells producing these IgGs experienced statistically significant negative selection in the FWR (framework region) in both the heavy and the light chain. As for the CDRs, the selection estimations are not significant (not negative and not positive in neither chain) (Fig. S11). This observation is typical when mAb selection is estimated based on a few sequences. These data indicate that according to the selection patterns it seems that these cells were subjected to typical antigen-experienced maturation process that involved negative (purifying) selection in the framework regions and nonnegative in the complementary determining regions.

Animal coronavirome-based high throughput diagnostics

Testing for multiple CoV antigens at high resolution in parallel could yield higher specificity than conventional tests based on single proteins of SARS-CoV-2 by improving discrimination from seasonal hCoVs. Most current SARS-CoV-2 serological tests rely on detection of entire S-/RBD or N-proteins, reporting an aggregate of binding against all epitopes within (10, 15). PhIP-Seq implementations of CoV antigen libraries primarily based on hCoVs have also demonstrated highly accurate diagnostic performance (13, 23). We set out to benchmark our extensive CoV antigen library containing both hCoVs and a large set of aCoVs for diagnostic applications. Interpreting the epitope-resolved antibody binding data reported by our assay extends beyond conventional serological tests, as binding against various epitopes of all hCoV needs to be weighed. We used gradient boosting trees (XGBoost) to build a predictor that accurately separates COVID-19 patients from healthy controls based on antibody signatures (Fig. 7, Fig. S10) when trained on antibody binding data of the complete library (AUC=0.980, Fig. 7a). Depending on the intended application and cutoffs employed, this assay displays with this set of samples 100% specificity at 87% sensitivity (reporting virtually no false-positives for this set of samples) or 92% specificity at 95% sensitivity (Fig. 7a PR curve). Interestingly, even the subset of antibody binding data solely against peptides of aCoVs allowed to accurately separate between exposed and unexposed individuals (AUC=0.96, Fig. 7b).

When looking at the principal components of an analysis (PCA, Fig. 7a,b) performed on the log fold change (number of reads of bound peptides vs. baseline sequencing of phages not undergoing immunoprecipitation) of significantly enriched peptides, unexposed individuals clustered together while recovered COVID-19 patients’ samples showed a greater spread. This illustrates the interindividual variability in hCoV antibody responses elicited by SARS-CoV-2 and the PCA separation based on the entire CoV antigen library (Fig. 7a) or the subset of aCoV antigens (Fig. 7b) yielded similar results. This interindividual variability in COVID-19 recovered patients may represent differences in immune responses or could also be affected by the time passed between infection and blood sampling (although this period is <3 months for our COVID-19 cohort where anti-SARS-CoV-2 antibody responses may show relatively high stability (41)).

As outlined above, aCoVs alone are enough to accurately separate between exposed and unexposed individuals. When further adding other hCoVs except SARS-CoV-2 to the aCoVs (Fig. 7c), the accuracy slightly increased to nearly the same level as using the complete library including SARS-CoV-2 antigens (AUC=0.97, Fig. 7a). Also binding data against hCoVs peptides excluding SARS-CoV-2 allowed for accurate detection of past infection (Fig. 7c). SARS-CoV-2 antigens alone yielded lower predictive power, but including also peptides originating of SARS-CoV-2 variants improved the AUC (Fig. 7c). This difference is possibly owing to the low redundancy in the included SARS-CoV-2 peptides of the reference strain, which improves when adding slightly different variants (that potentially do not strongly affect antibody binding). In this respect, including identical SARS-CoV-2 reference peptides in the input library, could also be a strategy to improve the performance of diagnostic PhIP-Seq assays. However, given the already large redundancy arising from non-SARS-CoV-2 hCoVs and aCoVs in our existing library, it does not appear necessary to also add duplicates of peptides from these groups. Taken together, these results indicate, that is possible to accurately detect SARS-CoV-2 exposure, without including actual SARS-CoV-2 antigens, suggesting a potential diagnostic strategy for the early stage of future pandemics.