key: cord-1025757-7c4db7eg authors: Castro Dopico, X.; Hanke, L.; Sheward, D. J.; Christian, M.; Muschiol, S.; Grinberg, N. F.; Adori, M.; Perez Vidakovics, L.; Chang Il, K.; Khoenkhoen, S.; Pushparaj, P.; Moliner Morro, A.; Mandolesi, M.; Forsell, M.; Coquet, J.; Corcoran, M.; Rorbach, J.; Aleman, S.; Bogdanovic, G.; Mcinerney, G.; Allander, T.; Wallace, C.; Murrell, B.; Albert, J.; Karlsson Hedestam, G. B. title: Disease-associated antibody phenotypes and probabilistic seroprevalence estimates during the emergence of SARS-CoV-2 date: 2020-07-17 journal: nan DOI: 10.1101/2020.07.17.20155937 sha: baeb78543ba2cde8f74db3f8918bad128661e6cd doc_id: 1025757 cord_uid: 7c4db7eg Serological studies are critical for understanding pathogen-specific immune responses and informing public health measures (1,2). By developing highly sensitive and specific trimeric spike (S)-based antibody tests, we report IgM, IgG and IgA responses to SARS-CoV-2 in COVID-19 patients (n=105) representing different categories of disease severity. All patients surveyed were IgG positive against S. Elevated anti-SARS-CoV-2 antibody levels were associated with hospitalization, with IgA titers, increased circulating IL-6 and strong neutralizing responses indicative of intensive care status. Antibody-positive blood donors and pregnant women sampled during the pandemic in Stockholm, Sweden (weeks 14-25), displayed on average lower titers and weaker neutralizing responses compared to patients; however, inter-individual anti-viral IgG titers differed up to 1,000-fold. To provide more accurate estimates of seroprevalence, given the frequency of weak responders and the limitations associated with the dichotomization of a continuous variable (3,4), we used a Bayesian approach to assign likelihood of past infection without setting an assay cut-off. Analysis of blood donors (n=1,000) and pregnant women (n=900) sampled weekly demonstrated SARS-CoV-2-specific IgG in 7.2% (95% Bayesian CI [5.1-9.5]) of individuals two months after the peak of spring 2020 COVID-19 deaths. Seroprevalence in these otherwise healthy cohorts increased steeply before beginning to level-off, following the same trajectory as the Stockholm region deaths over this time period. As SARS-CoV-2 only recently crossed the species barrier 5 , populations globally were 53 immunologically naïve to the virus. Characterizing the antibody response to nascent 54 outbreaks is, therefore, central to optimizing approaches to tackle COVID-19 and 55 future pandemics, furthering our basic understanding of human immunology 6-8 . 56 To date, numerous SARS-CoV-2 studies have reported seroprevalence and disease-57 associated antibody phenotypes 9-11 , isolated virus-neutralizing monoclonal 58 antibodies 12-15 and used convalescent individuals to define metrics for plasma 59 therapy 16 . However, consensus on several key issues remains outstanding. For 60 instance, the majority of serology data are derived from commercial kits utilizing 61 SARS-CoV-2 spike derivatives (e.g. RBD, S1 or S2 domains) or the nucleocapsid to 62 detect pathogen-specific antibodies 9,17,18 . Several of these assays suffer from epitope 63 loss 19 , increased cross-reactivity 9,20 and lower sensitivity 21 , highlighting the need for 64 comparative studies to identify optimal assay formats for individual and population 65 level analysis 22-24 . Here, we developed a highly sensitive and specific ELISA assay 66 based on native-like prefusion-stabilized spike trimers 25 , as well as the receptor-67 binding domain (RBD), to accurately evaluate anti-viral antibody levels in COVID-19 68 patients and key community groups. To identify disease-associate antibody 69 phenotypes, anti-viral IgM, IgG and IgA levels were analyzed alongside in vitro virus 70 neutralisation capacity and a descriptive set of clinical features, including intensive 71 care status. 72 A major consideration for antibody testing concerns setting the cut-off for positivity 28 , which significantly affects seroprevalence estimates and individual clinical 74 management. Currently employed approaches to the problem are severely limited by 75 their high dependency on the representative nature of negative control values and the 76 dichotomization of a continuous variable. Therefore, to obtain more accurate 77 seroprevalence estimates in blood donors and pregnant women, we strictly controlled 78 our assay with a large number of historical controls (n=295) and used our patient data 79 to train probabilistic algorithms to handle ELISA data. We used our cut-off-80 independent approach to model population changes in seropositivity over time. We sampled blood donors and pregnant women weekly throughout the outbreak. 83 Blood donors serve as an important clinical resource, including for potential COVID-84 19 plasma therapies, while pregnant women require close clinical monitoring with 85 respect to fetal-maternal health and are known to employ unique, yet poorly 86 characterized immunological mechanisms that impact infectious pathology 29-31 . As 87 Sweden did not impose a strict lockdown in response to the pandemic and has 88 reported relatively high per-capita morbidity and mortality (ECDC, EU), understanding 89 4 SARS-CoV2 seroprevalence in these cohorts helps plan clinical need and understand 90 the development of immunity in the population. 91 92 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 17, 2020. . https://doi.org/10.1101/2020.07.17.20155937 doi: medRxiv preprint 6 group varied greatly, with some persons displaying levels similar to those measured 135 in COVID-19 patients while others had titers similar to those observed at the higher 136 end of the historical negative control range. This group of weak responders highlighted 137 the challenge of using a pre-determined assay cut-off, a critical point that we address 138 and present potential solutions for in Figure 4 . 139 140 When examining IgM and IgA responses against S and RBD, we found them to be 141 less potent and more variable between individuals than the IgG response ( Fig 1B and 142 S3B). Given their unique immunological roles and what is known about COVID-19 143 pathology, we sought to investigate whether isotype responses segregated with 144 clinical features. Therefore, the COVID-19 patient cohort was classified according to 145 clinical disease severity: Category 1 -non-hospitalized; Category 2 -hospitalized; 146 Category 3 -intensive care (on mechanical ventilation) ( Table 1) . Within-patient anti-S and anti-RBD responses were highly correlated for all three 149 isotypes ( Fig S3C) . Furthermore, multivariate analyses showed that increased anti-150 viral IgM, IgG and IgA levels significantly correlated with disease severity (Fig 1C, S3D 151 and S Table 1 ), in line with the lower titers observed in blood donors and pregnant 152 women. This was most pronounced for pathogen-specific IgA, suggestive of 153 advancing mucosal disease 33 . A more severe clinical picture was also strongly 154 associated with elevated serum IL-6 ( Fig 1C) , a cytokine that feeds antibody 155 production 34-37 . IL-6 is dysregulated in polygenic metabolic diseases 38-40 and acute 156 respiratory distress syndrome (ARDS) 41 , which are risk factors for COVID-19-157 associated mortality 42,43 . Notably, under univariate analysis, female patients showed 158 lower anti-viral IgA levels than males in non-hospitalized and hospitalized groups, 159 which was also true for anti-RBD IgA in multivariate analysis (Fig S4A) 44 . Overall, 160 severity showed the most consistent relationship with any measure and was the 161 primary driver of Ig levels. To summarize patient antibody phenotypes, we generated a Spearman's rank 164 correlation matrix (Fig 2A) , which highlights a negative association between patient 165 date of birth and all dataset features, in-keeping with the worse prognosis for elderly 166 COVID-19 patients, as well as a lack of association between serum IL-6 and anti-viral 167 IgM levels, further supporting that levels of the cytokine and IgA mark a protracted, 168 more severe clinical course. After accounting for the effect of age, sex and days since 169 PCR+ test, anti-viral IgA titers were approximately 3-fold higher in intensive care vs. 170 non-hospitalized patients, while IL-6 levels were ca. 10-fold increased ( Fig 2B) . Notably, we also detected low levels of anti-viral IgA when analyzing a small subset 172 (n=100) of healthy donors from this year ( Fig S4B) . Critical future studies are required to establish the longevity of immunity to SARS-175 CoV-2, including the efficacy of cellular re-call responses, as detectable antibody titers 176 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 17, 2020. . 7 decline in the absence of antigen 45 . Here, anti-viral IgG levels were maintained two 177 months post-disease onset/positive PCR test, while IgM and IgA declined (Fig S4C) , 178 in agreement with their circulating t1/2 46 and reflecting their diagnostic utility. Patient 179 serum was collected 6-61 days post-PCR (Fig S4D) , with the median time from 180 symptom onset to PCR test being 5 days. In longitudinal patient samples where we 181 observed seroconversion, IgM, IgG and IgA peaked with similar kinetics when all three 182 isotypes developed, although anti-viral IgA was not always generated in Category 1 183 and 2 patients (Fig 1D and S4E ). We note that anti-viral IgM could be maintained two 184 months post-PCR+/symptom onset (Fig S4E) Fig 3A) , with binding and 193 neutralisation titers being highly correlated (Fig 2A) 3C ). Across the two antigens and three isotypes, anti-RBD IgG was most strongly 201 correlated with neutralization ( Fig 3D) . Community seroprevalence estimations 204 205 As the Stockholm region is a busy metropolitan area and Sweden did not impose strict 206 lockdown in response to the emergence of SARS-CoV-2, we sought to better 207 understand the frequency and nature of anti-SARS-CoV-2 antibody responses in 208 healthy blood donors and pregnant women during weeks 14-25 (March 30 to June 22 209 2020) (Fig 4A) . By surveying a large number of historical controls (n=295) during assay development, 212 we identified a considerable number of samples with weak reactor-phenotypes, which 213 must be taken into account when setting the assay cut-off. When using a small set of 214 historical controls, such values may be missed, resulting in an incorrect cut-off that 215 significantly increases the uncertainty of individual and population seroprevalence 216 estimates. To illustrate this, we randomly sampled groups of 20 negative control 217 samples (from 890 measures in 295 individuals) and calculated seroprevalence in 218 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 17, 2020. . 8 blood and pregnant women from weeks 17-19 based on a 6SD cut-off. Depending on 219 which 20 negative controls were sampled, the seroprevalence ranged from 5.7 to 220 8.7%, a 35% difference ( Fig 4B) . Weak responder status is likely influenced by many 221 factors, including genetic background, health status, total serum Ig and protein levels, 222 assay variability. Critically, test samples with low anti-viral titers may also fall into this 223 range, highlighting the need to better understand the assay boundary. Taking a one-224 dimensional 6SD cut-off for anti-S IgG responses based on all 890 values from 295 225 negative donors, 7.7% of healthy donors tested positive two months after the peak of 226 deaths in the country (Fig 4C) . To exploit individual antibody titers and improve our estimates, we modelled the 229 probability that a sample is positive by training two parallel learners using our patient 230 and historical control anti-S and -RBD data, rather than setting a one-dimensional cut-231 off. Our Bayesian approach for inferring seroprevalence without thresholds is 232 described in Christian et al 47 . Briefly, we exploit a class-balance corrected Bayesian 233 logistic regression to infer seroprevalence, using a modified Gaussian Process to 234 construct a prior over seroprevalence trajectories, sharing information between weeks. 235 Using this approach, which allowed us to model population changes over time, we 236 found the steep increase in positivity at the start of the pandemic to slow between 237 weeks 17-25, approximating 7.2% (95% equal tailed Bayesian credible intervals [5.1 238 -9.5]) at the last time point and suggesting that humoral immunity to the virus 239 develops slowly in these populations despite considerable virus spread in the 240 community. To confirm our novel Bayesian approach, we compared it to various machine learning 243 algorithms, creating an ensemble SVM-LDA learner to maximize sensitivity, specificity 244 and consistency across different cross-validation strategies, where the proportion of 245 positives in the held-out test set was deliberately varied from the training set (see 246 Materials and Methods). We obtained convergent results between these tools and our 247 framework ( Fig S5A) . Strikingly, seroprevalence inferred using our Bayesian approach 248 exhibited the same trajectory as Stockholm County deaths when lagged by two and 249 half weeks (Fig S5B) , allowing for the calculation of a seroprevalence case fatality rate 250 in appropriately powered cohorts. Furthermore, such probabilistic approaches that annotate uncharacterized uncertainty 253 pave the way towards greater clinical utility for antibody measures. For example, 254 individuals 50% likely to be antibody-positive according to the ensemble learner ( Fig 255 4E) can targeted for further investigation (e.g. re-testing). These tools have the 256 potential, therefore, to provide more qualitative individual antibody measurements. 257 258 259 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 17, 2020. . Serology remains the gold standard for estimating previous exposure to pathogens 262 and benefits from a large historical literature 48 . When robust, it indicates whether an 263 individual has mounted an adaptive immune response against a specific agent and is 264 a strong predictor of an anamnestic response and quicker recovery upon re-infection. Anti-viral antibody responses are central for protection against re-infection and for the 266 protective efficacy of successful vaccines 49 . Although the concept of herd immunity is based upon the study of antibodies, 269 worryingly, there is little-to-no standardization (or validation) of many of the available 270 SARS-CoV-2 antibody tests. would allow for more specific targeting and control of immunity, with IL-6 being a case-292 in-point (NCT04322773, NCT04359667), although the cytokine's role in orchestrating 293 adaptive responses 52 may make its modulation a double-edged sword in some 294 patients 53 . Outside of the severe disease setting, it is important to determine how many people 297 have seroconverted to SARS-CoV-2. Blood donors and pregnant women are both 298 good sentinels for population health, although they are not enriched for high-risk 299 groups, such as public transportation employees, where rates of infection may be 300 higher. Blood donors are generally working age, active and mobile members of society 301 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 17, 2020. . https://doi.org/10.1101/2020.07.17.20155937 doi: medRxiv preprint with a good understanding of health, and the majority of pregnant women in Sweden 302 would have been advised to take precautions against infectious diseases. However, 303 interestingly, both groups showed a similar seroprevalence. By tracking these cohorts over time, we modelled seroprevalence changes at the 306 population level. We found a steep increase in antibody positivity at the start of the 307 pandemic, which increased, although at a slower rate, in blood donors and pregnant 308 women during subsequent weeks, in-line with a decreasing caseload and fatalities 54 . 309 Indeed, ICU occupancy and deaths are a better proxy for viral spread than PCR+ 310 diagnoses, which are highly dependent on the number of tests carried out, and we 311 note the close approximation of our seroprevalence data to Stockholm County deaths 312 (Fig S5B) , illustrating the power of mortality data to infer seroprevalence and vice 313 versa. Together, our study defines key features of the humoral immune responses to 316 emerging beta-coronaviruses, delineating disease features, while our seroprevalence 317 data indicate that population immunity to SARS-CoV-2 develops slowly even in the 318 absence of lockdowns, suggesting that classical serological herd immunity following 319 natural infection will require a larger outbreak and higher clinical toll; notwithstanding 320 yet unknown contributions from T lymphocytes and other lineages. Given the 321 uniqueness of the approach in Sweden 55 , these data may inform the management of 322 future pandemics. More widely, given the high inter-individual variability in antibody 323 responses to SARS-CoV-2, the genetic and environmental factors influencing 324 individual antibody responses will be important to elucidate. 325 326 327 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 17, 2020. . Human samples and ethical declaration 330 331 Samples from COVID-19 patients (n=105) were collected through the Departments of 332 Medicine and Clinical Microbiology at the Karolinska University Hospital and were 333 handled and analyzed in accordance with approval by the Swedish Ethical Review 334 Authority (registration no. 2020-02811). All personal identifiers were pseudo-335 anonymized, and clinical feature data were blinded to the researchers carrying out 336 experiments until data generation was complete. 337 338 All patients in the study were confirmed PCR+ for SARS-CoV-2 by nasopharyngeal 339 swab or upper respiratory tract sampling after being admitted to Karolinska University 340 Hospital. As viral RNA CT values were determined using different qPCR platforms 341 between patients, we did not analyze these alongside other available features. Patients were questioned about the date of symptom onset at their initial consultation 343 and followed-up for serology during their care, up to 2 months post-diagnosis. In 344 addition, longitudinal samples from 10 of these patients were collected to monitor 345 seroconversion and isotype persistence. 346 347 Anonymized samples from blood donors (100/week) and pregnant women (100/week) 348 were randomly selected and obtained from the department of Clinical Microbiology, 349 Karolinska University Hospital. No metadata, such as age or sex information was 350 available for these samples. Pregnant women were sampled as part of routine for 351 infectious diseases screening during the first trimester of pregnancy. Blood donors 352 (n=295) collected through the same channels a year previously were randomly 353 selected for use as negative controls. Serum samples from individuals testing PCR+ 354 for endemic coronaviruses, 229E, HKU1, NL63, OC43 (n=20, ECV+) in the prior 2-6 355 months, were used as additional negative controls. The use of these anonymized 356 samples was approved by the Swedish Ethical Review Authority (registration no. 357 2020-01807). Stockholm County death and Swedish mortality data was sourced from the ECDC and 360 Swedish Public Health Agency, respectively. 361 362 Study samples are defined in Table 1 . Serum sample processing 365 366 Blood samples were collected from patients by the attending clinical team and serum 367 isolated by the department of Clinical Microbiology. Samples were barcoded and 368 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 17, 2020. . 12 stored at -20 o C until use. Serum samples were not heat-inactivated for ELISA 369 protocols but were heat-inactivated at 56 o C for 60 min for neutralization experiments. 370 371 SARS-CoV-2 antigen generation 372 373 The plasmid for expression of the SARS-CoV-2 prefusion-stabilized spike ectodomain 374 with a C-terminal T4 fibritin trimerization motif was obtained from 25 . The plasmid was 375 used to transiently transfect FreeStyle 293F cells using FreeStyle MAX reagent 376 (Thermo Fisher Scientific). The ectodomain was purified from filtered supernatant on 377 Streptactin XT resin (IBA Lifesciences), followed by size-exclusion chromatography 378 on a Superdex 200 in 5 mM Tris pH 8, 200 mM NaCl. 379 380 The RBD domain (RVQ -QFG) of SARS-CoV-2 was cloned upstream of a Sortase A 381 recognition site (LPETG) and a 6xHIS tag, and expressed in 293F cells as described 382 above. RBD-HIS was purified from filtered supernatant on His-Pur Ni-NTA resin 383 (Thermo Fisher Scientific), followed by size-exclusion chromatography on a Superdex 384 200. The nucleocapsid was purchased from Sino Biological. Anti-SARS-CoV2 ELISA 387 388 96-well ELISA plates (Nunc MaxiSorp) were coated with SARS-CoV-2 S, RBD or 389 nucleocapsid (100 μl of 1 ng/μl) in PBS overnight at 4 o C. Plates were washed six times 390 with 300 ml PBS-Tween-20 (0.05%) and blocked using PBS-5% no-fat milk (Sigma). Human serum samples were thawed at room temperature, diluted (1:100 unless 392 otherwise indicated), vortexed and incubated in blocking buffer for 1h before plating. 393 Serum samples were incubated overnight at 4 o C before washing, as before. 394 Secondary HRP-conjugated anti-human antibodies were diluted in blocking buffer and 395 incubated with samples for 1 hour at room temperature. Plates were washed a final 396 time before development with TMB Stabilized Chromogen (Invitrogen). The reaction 397 was stopped using 1M sulphuric acid and OD values were measured at 450 nm using 398 an Asys Expert 96 ELISA reader (Biochrom Ltd.). Secondary antibodies (all from 399 Southern Biotech) and dilutions used: goat anti-human IgG (2014-05) at 1:10,000; 400 goat anti-human IgM (2020-05) at 1:1000; goat anti-human IgA (2050-05) at 1:6,000. 401 All assays of the same antigen and isotype were developed for their fixed time and 402 samples were randomized and run together on the same day when comparing binding 403 between patients. All data were log transformed for statistical analyses. In vitro virus neutralisation assay 406 407 Pseudotyped viruses were generated by the co-transfection of HEK293T cells with 408 plasmids encoding the SARS-CoV-2 spike protein harboring an 18 amino acid 409 truncation of the cytoplasmic tail 25 ; a plasmid encoding firefly luciferase; a lentiviral 410 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 17, 2020. . packaging plasmid (Addgene 8455) using Lipofectamine 3000 (Invitrogen RBD and -S training data to model the relationship between the ELISA measurements 447 and the probability that a sample is antibody-positive. We adjust for the training data 448 class proportions and use these adjusted probabilities to inform the seroprevalence 449 estimates for each time point. Given that the population seroprevalence cannot 450 increase dramatically from one week to the next, we construct a prior over 451 seroprevalence trajectories using a transformed Gaussian Process, and combine this 452 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 17, 2020. 2, 5, 10) via an internal 5-fold CV with 5 repeats (with the winning parameter used for 473 the final model for the main CV iteration). We also note that the natural output of SVM 474 are class labels rather than class probabilities, so the latter are obtained via the 475 method of Platt 56 . We considered three strategies for cross-validation: 478 479 • random: individuals were sampled into folds at random 480 • stratified: individuals were sampled into folds at random, subject to ensuring 481 the balance of cases:controls remained fixed 482 • unbalanced : individuals were sampled into folds such that each fold was 483 deliberately skewed to under or over-represent cases compared to the total 484 sample 485 486 We sought a method that worked equally well across all cross-validation schemes, as 487 the true proportion of cases in the test data is unknown and so a good method should 488 not be overly sensitive to the proportion of cases in the training data. 489 490 We found most methods worked well, although logistic regression was sensitive to 491 changes in the case proportion in the training data. We chose to create an ensemble 492 method combining that with the highest specificity (LDA) and the highest sensitivity 493 (SVM), defined as an unweighted average of the probabilities generated under SVM 494 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 17, 2020. We then trained the ensemble learner on all 719 training samples and predicted the 498 probability of anti-SARS-CoV-2 antibodies in blood donors and pregnant volunteers 499 sampled in 2020. We inferred the proportion of the sampled population with positive 500 antibody status each week using multiple imputation. We repeatedly (1,000 times) 501 imputed antibody status for each individual randomly according to the ensemble 502 prediction, and then analyzed each of the 1,000 datasets in parallel, combining 503 inference using Rubin's rules, derived for the Wilson binomial proportion confidence 504 interval 57 . 505 506 Data and code availability statement 507 508 Data generated as part of the study, along with custom code for statistical analyses, is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 17, 2020. . The study authors declare no competing interests related to the work. is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 17, 2020. is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 17, 2020. is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 17, 2020. . . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 17, 2020. . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 17, 2020. . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 17, 2020. . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 17, 2020. . Rolling weekly avg. COVID-19 deaths/per million Study sampling interval . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 17, 2020. . Table 1 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 17, 2020. . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 17, 2020. Figure S2 : Patient binding titrations A subset (n=40) PCR+ COVID-19 patients were titrated in our assay. Shown are these responses against the three antigens and isotypes in the study. Four ECV+ (red lines) were in a similar dilution series alongside. . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 17, 2020. . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 17, 2020. . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 17, 2020. . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 17, 2020. . https://doi.org/10.1101/2020.07.17.20155937 doi: medRxiv preprint Anti-nucleocapsid IgM Cross-reactive individuals Patients