key: cord-1045612-lh8i4h0o
authors: Atyeo, Caroline; Fischinger, Stephanie; Zohar, Tomer; Slein, Matthew D.; Burke, John; Loos, Carolin; McCulloch, Denise J.; Newman, Kira L.; Wolf, Caitlin; Yu, Jingyou; Shuey, Kiel; Feldman, Jared; Hauser, Blake Marie; Caradonna, Tim; Schmidt, Aaron; Suscovich, Todd J.; Linde, Caitlyn; Cai, Yongfei; Barouch, Dan; Ryan, Edward T.; Charles, Richelle C.; Lauffenburger, Douglas; Chu, Helen; Alter, Galit
title: Distinct early serological signatures track with SARS-CoV-2 survival
date: 2020-07-30
journal: Immunity
DOI: 10.1016/j.immuni.2020.07.020
sha: 29e23bfdf6e364aff85562a828b912257735a8c0
doc_id: 1045612
cord_uid: lh8i4h0o

Summary As SARS-CoV-2 infections and death counts continue to rise, it remains unclear why some individuals recover from infection whereas others rapidly progress and die. While the immunological mechanisms that underlie different clinical trajectories remain poorly defined, pathogen-specific antibodies often point to immunological mechanisms of protection. Here, we profiled SARS-CoV-2–specific humoral responses on a cohort of 22 hospitalized individuals. Despite inter-individual heterogeneity, distinct antibody signatures resolved individuals with different outcomes. While no differences in SARS-CoV-2-specific IgG levels were observed, spike–specific humoral responses were enriched among convalescent individuals, whereas functional antibody responses to the nucleocapsid were elevated in deceased individuals. Furthermore, this enriched immunodominant S-specific antibody profile in convalescents was confirmed in a larger validation cohort. These results demonstrate that early antigen-specific and qualitative features of SARS-CoV-2-specific antibodies, point to differences in disease trajectory, highlighting the potential importance of functional antigen-specific humoral immunity to guide patient care and vaccine development.

SARS-CoV-2 is the newest coronavirus to cross into the human population (Wu et al., 2020b; Zhu et al., 2020) . Millions of infections have been diagnosed (WHO, 2020) ; however, the number of asymptomatic carriers is likely to far exceed these numbers . While the rapid spread of SARS-CoV-2, even during the asymptomatic phase of this infection, is alarming, more harrowing is our inability to predict disease trajectories among symptomatic individuals. In the absence of therapeutics and vaccines as countermeasures for this infection, there is an urgent need to begin to map the evolution of immunity to the pathogen to guide patient care and future immune interventions.

Although both antibody responses and T cells have been linked to disease resolution , and neutralizing antibodies have been demonstrated to block infection in small animal models (Quinlan et al., 2020) , little is known about the antibody features that are important for protection.

Neutralizing antibodies develop in the majority of SARS-and MERS-infected individuals (Chang et al., 2005; de Wit et al., 2016) ; however, the virus can mutate to overcome these antibody responses ter Meulen et al., 2006) . Passive immunization studies with both neutralizing and poorly neutralizing antibodies have shown protection in lethal MERS infection in mice (Zhao et al., 2015; Zhao et al., 2017) , suggesting that both the neutralizing and extra-neutralizing functions of antibodies may play a critical role in control and resolution of disease. Moreover, recent studies have found lower neutralization titers in younger individuals and higher neutralization among individuals with severe disease (Wu et al., 2020a , suggesting that antibodies may depend on additional mechanisms to clear the virus.

Antibody dynamics during the acute window of infection have been linked to differential outcomes across infections, including HIV (Tomaras and Haynes, 2009) , influenza (Cobey and Hensley, 2017) , and Ebola virus infection (Saphire et al., 2018) . Specifically, the selection of specific antibody subclasses and functional profiles is heavily influenced by inflammatory cascades and may not only forecast disease outcomes but also point to antibody mechanisms of action vital in early pathogen control and clearance. However, whether identifiable antibody functional profiles across SARS-CoV-2 antigen-specificities evolve early following infection that track differentially with disease outcome is unknown. In this study, we assembled two cross-sectional sample sets of SARS-CoV-2-infected individuals at the time of hospital admission to begin to comprehensively profile the evolution of the early SARS-CoV-2 S-specific response and to define antibody features that are predictive of disease outcome. Through this analysis, we found that deceased and convalescent individuals present different humoral profiles, with a more S-focused response in individuals who convalesced and a stronger Nspecific response in individuals who succumbed to disease.

Across infectious diseases, pathogen-specific antibodies can serve as biomarkers of infection and aid in the early control and clearance of infection by blocking host-pathogen interactions and/or recruiting innate immune functions (Gunn and Alter, 2016) . In order to investigate whether early SARS-CoV-2-specific humoral immune responses differ across individuals that ultimately recover or die from infection, a cohort of 22 hospitalized SARS-CoV-2-infected individuals, of whom 12 recovered and 10 died, was profiled. Samples were collected at hospital admission, all recruited within the first 20 days following symptom onset (Table 1 and Figure S1 ) at the University of Washington, Seattle, one of the earlier epicenters in the US (Holshue et al., 2020) . Population demographics largely resemble those previously reported (Bhatraju et al., 2020) , including elevated numbers of elderly men in the subset that died.

To profile the SARS-CoV-2-specific humoral immune response, we performed Systems Serology to determine the biophysical and functional characteristics of SARS-CoV-2-specific antibodies that recognize the SARS-CoV-2 spike (S), the S-derived receptor binding domain (RBD), and the nucleocapsid (N). The titers of SARS-CoV-2-specific isotypes and subclasses, the Fcγ-receptor binding profiles, neutralization, as well as antigen-specific innate effector functions were measured. Heterogeneous responses were observed across both populations ( Figure 1A and Figure S2 ), and convalescents did not appear to possess quantitatively superior immune responses that could explain their different later disease course. Univariate analyses further confirmed that no significant differences were observed in SARS-CoV-2-specific IgG1 or IgA1 titers across S, RBD, and N ( Figure 1B -C, and Figure S2 ). Conversely, subtle distribution differences were observed for SARS-CoV-2-specific IgM responses, with a slight shift towards higher S-specific IgM among survivors, and a trend towards increased N-specific IgM responses among individuals who died ( Figure 1C ). Functional antibody profiles displayed similar distributions across the cohorts for antibody dependent cellular phagocytosis (ADCP) ( Figure 1D ) and neutralization ( Figure 1G) . Surprisingly, RBD-specific antibody-mediated NK cell degranulation (NKD) and antibody dependent neutrophil phagocytosis (ADNP), both driven through related Fcγ-receptors FcγR3A and FcγR3B, respectively, trended towards increases among individuals who died (Figure 1D-F) . Antibody measurements were minimally influenced by time since symptom onset ( Figure S1 ), suggesting equivalent evolution of humoral immune responses across groups. However, no single antibody feature could discriminate between the groups.

Beyond univariate differences, emerging data point to a critical role for humoral immune response coordination as a predictor of protection in some infections (Ackerman et al., 2018; Barouch et al., 2015) . Given the polyclonal nature of the early humoral immune response, multiple functions or features may simultaneously contribute to differential control and clearance of infection. Correlation matrices split by group were used to examine the relationships between antibody isotypes or subclasses and antibody-dependent effector functions across the groups (Figure 2A) . Within both groups, isotypes and subclasses were highly correlated. Conversely, the relationship between isotype or subclass and functions differed across the two populations. Stronger correlations between titers and functions were observed in convalescent individuals (Figure 2A ). Disparities were observed in both NK cell and neutralizing antibody coordination between the two groups. Though not significant, individuals who died exhibited correlated isotype or subclass responses with monocyte and neutrophil phagocytosis, but negative and generally poorer correlations of NK cell activating and complement recruiting antibody responses with all other functions (Figure 2A) , suggesting that individuals who pass away develop a functionally biased humoral immune response. While IgG1 responses were associated with all functions across the individuals that later went on to die, diversified isotype and subclass responses were largely inversely correlated with antibody-dependent complement deposition (ADCD) and natural killer cell (NK) NK functions. This observation suggests that these individuals leverage isotype and subclass diversification in a manner that may preclude the full deployment of the humoral immune response.

Conversely, convalescents overall displayed a more uniform correlation profile across subclass and isotype responses and antibody effector function. However, while neutralizing antibody responses were co-induced with isotype and subclass and effector functions among individuals who died, neutralizing antibody responses were largely inversely correlated to all antibody responses among individuals that went on to recover, suggesting a divergent evolution of the antigen-binding and constant domain of the antibody across these populations. These data highlight multiple early functional differences in SARS-CoV-2 specific humoral immunity between the groups.

To further probe the overall humoral profile between groups, the mean percentile of each antibody metric was determined across SARS-CoV-2-antigen specificities for both populations ( Figure   2B ). The nightingale rose plots reveal that deceased individuals exhibited a more N-focused humoral immune response compared to the S-centric response elicited among convalescents. In particular, higher S-specific ADCD, ADNP, ADCP and enhanced IgG1, IgA1, and IgM responses were observed among survivors. In contrast, S-specific NK cell activating responses were enriched in deceased.

Unexpectedly, RBD-specific responses were largely enriched among individuals that went on to pass away, with the exception of RBD-specific monocyte phagocytosis which was enriched among individuals that went on to survive. These data point to both antigen-specific and antibody-effector differences early in infection that differ by clinical trajectory.

Given the unique correlation and immunodominance profiles across the groups (Figure 2A and B), we next aimed to define whether a minimal set of features could be identified that could segregate individuals with different clinical outcomes. To this end, feature down-selection was performed to avoid overfitting, followed by partial least squares discriminant analysis (PLSDA) to visualize differences (Lau et al., 2011) . Despite the small numbers, separation was observed across the groups ( Figure 3A ). All antibody features as well as sex and interventions ( Table 1) were included in the analysis, and as few as 5 features were sufficient to drive separation across the subjects (Figure 3A and B). S-specific IgM and IgA1 responses were enriched in survivors, whereas N-specific complement activity (ADCD), IgM, and IgA1 titers were enriched in individuals who died. These data likely relate to the immunodominant shift towards S in convalescent individuals and towards N in deceased individuals ( Figure 2B ). Model performance was evaluated using leave-one-out cross validation, to test that significance of the model using different sets of subjects and to test outlier effects. The model clearly outperformed (Cliff's ∆) permuted and size-matched random controls ( Figure 3C) . Moreover, sensitivity analysis, evaluating model performance with the removal of individual outliers, highlighted the minimal impact of any given individual ( Figure S3A) . Furthermore, individual model features only possessed modest predictive power in resolving the groups, but collectively, combining all 5 features -in Latent Variable 1 (LV1)exhibited improved predictive accuracy ( Figure 3D ). Confounding features, such as days since symptom onset, sex, age, and viral load were also over-layed on the PLSDA scores plot ( Figure S3B -F), highlighting the limited capacity of any of these features to distinguish individuals into those who convalesced or died. Furthermore, at individual levels, these demographic factors were poorly predictive of disease outcome, underperforming classification compared to the LV1 classification model ( Figure S3G ). Thus, a minimal set of SARS-CoV-2 humoral profiles, rather than demographic information, appear to significantly resolve individuals who later went on to die from those who recover.

Given that the feature down-selection algorithm selects a minimal set of features to avoid overfitting, a co-correlates network was used to explore additional features that may distinguish these two groups ( Figure 3F) . A larger set of co-correlates can help provide mechanistic clues related to the immunologic mechanisms by which antibodies contribute to control and clearance of infection. Thus, a co-correlate network was built highlighting the relationship of model-selected features (large nodes) with additional highly correlated features (smaller nodes). Features enriched among individuals who later died, included N-specific IgM and IgA2, that were linked to a large number of additional N-and RBD-specific poorly functional antibody features. For example, correlates of risk were linked to the induction of less-functional IgG subclasses, IgG2 and IgG4, pointing to the early rise of dysregulated or less functional humoral immune responses as biomarkers or even drivers of ineffective control or clearance of infection. Conversely, S-specific IgM titers, enriched in convalescent individuals, were correlated with functional S-specific IgG3 responses, RBD-specific IgM, and S-specific monocyte and neutrophil phagocytosis. Moreover, S-specific IgA1 responses, also enriched among convalescents, were linked to RBD-specific complement activation (ADCD) and S-, RBD-, N-specific FcγR2A binding, the Fcγreceptor involved in phagocytosis. Given our emerging appreciation for the role of complement and phagocytosis in vaccine mediated protection against SARS-CoV-2 (Yasui et al., 2014) , these data potentially argue for a similar role for these functions in natural protection against disease. Moreover, the data also highlight the potential importance of a less N-focused, but more functional S-specific phagocytic response as an early correlate of recovery from infection.

Collectively, the data point to a shift in immunodominance of spike versus nucleocapsid functional antibody responses. To test this hypothesis, we next compared the overall ratio of Spike(S):Nucleocapsid(N)-specific antibody isotype, subclasses, and functions across the groups ( Figure   4A and Figure S4A ). As expected, several antibody features were selectively biased towards S-immunity in the convalescents compared to individuals that later died, including IgM, ADCP, ADNP, and ADCD.

Whether these effects were exclusive to this group of individuals from Seattle or could be generalized was next addressed in a second larger cohort of acutely infected individuals from Boston, of which 20 individuals convalesced and 20 died. Similarly, to the Seattle cohort, Boston samples were profiled in the first 20 days following symptom onset ( Table 2) . Similar to the Seattle discovery cohort, though differences were observed in S-and N-specific immune responses at a univariate level none passed multiple hypothesis correction ( Figure S4B ). Yet, when S:N ratios were compared across features, convalescent individuals exhibited a bias towards elevated S-specific humoral immunity compared to Nspecific immunity in contrast to individuals who went on to later pass away ( Figure 4B and Figure S4C ).

Thus, to ultimately capture the extent of S:N skewing across the groups, the number of features that had greater S than N responses were summed across convalescents and deceased individuals and compared within each cohort (Figure 4A and B) . In both cohorts, a significant enrichment of S:N immunity was observed in convalescents ( Figure 4C) . Therefore, these findings suggest that a consistent overall shift in S:N immunity early in SARS-CoV-2 infection may have a protective role and aid in recovery from severe disease.

Both cellular and humoral immune responses have been linked to protection against several coronaviruses . Importantly, antibodies represent pathogen-specific markers of exposure, serve as powerful biomarkers of disease activity, and often point to immunological mechanisms of protection able to guide therapeutic or vaccine development (Gunn and Alter, 2016) . By deeply profiling the SARS-CoV-2 humoral immune response early in infection, here, we defined a unique SARS-CoV-2specific humoral signature associated with later disease outcomes. A combination of five SARS-CoV-2specific antibody measurements were sufficient to distinguish individuals with different disease trajectories in a cohort from Seattle, including antibody measurements to S and N, with an overall enhanced S-centric response in individuals who recovered from infection. S-specific phagocytic and complement activity were enriched early in individuals that recovered from infection. This signature was confirmed in a second, larger SARS-CoV-2 infection cohort from Boston, where convalescent individuals exhibited a higher S:N ratio in their humoral immune response. These data point to early diverging humoral immune responses that may mark more effective immunity and suggest that functional antibodies directed against S might be beneficial for SARS-CoV-2 disease trajectory.

In SARS-CoV-1 and SARS-CoV-2 infection, N is highly immunogenic, with N-specific humoral immune responses arising concurrently with S-specific humoral immunity Shi et al., 2004; Timani et al., 2004) . However, immunization of hamsters with a vector expressing N offered no protection against SARS-CoV-2 challenge despite a strong anti-N response, whereas immunization with the same vector expressing S protected hamsters against challenge (Buchholz et al., 2004) . It is estimated that 100 copies of S and 1000 copies of N are incorporated into each virion (Bar-On et al., 2020) , suggesting that 10-fold more N may be produced compared to S during infection to effectively generate viral progeny. Due to the high amounts of nucleocapsid, N-directed responses may be indicative of higher disease burden and increased antigen exposure. However, the similarity in viral loads between the individuals who recovered and those who died does not support this hypothesis.

Rather, the data point to compromised evolution of S-immunity in individuals that later pass away. The potential beneficial role of S-targeted immunity in viral control is reinforced in new studies in nonhuman primates (NHP), demonstrating elevated and robust functional humoral immune responses to S, rather than RBD and N, following primary infection that were associated with protection upon reexposure to the virus .

It is well known that timing of sampling may influence humoral profiles, where sampling time could result in the comparison of immature versus mature immune responses. Despite the sampling differences in the group, comparable titers were observed across the convalescents and individuals that ultimately passed away. Moreover, similar overall functional profiles were also observed, suggesting that the humoral immune responses were comparable in magnitude across the two groups. Additional analysis of the influence of sampling time on the spread of the antibody profiles in the PLSDA highlighted a minimal influence of time from symptoms on overall antibody profile variation and the time of sampling exhibited a minimal predictive power in classifying individuals into convalescents or deceased. Yet, longitudinal analyses will be illuminating providing further information on the evolving humoral immune response that tracks with protection from infection.

Emerging data point to higher mortality among the elderly and across the genders (Hauser et al., 2020) . Along these lines, individuals who passed away were on average older than those who convalesced. Age can have a profound effect on immune function, and though this study was not suited to explore the relationship between age, outcome, and humoral responses, future larger studies across age groups could provide insights on the differential susceptibility among the elderly. However, the impact of age, sex, and viral loads illustrated a minimal influence of each of these variables on the overall variability of the humoral immune responses. Additionally, the individual predictive power of these demographic variables were lower than the predictive power of the model selected antibody features (LV1). While S-specific antibodies able to recruit NK cell activity were expanded in individuals that went on to pass away, pointing to a potentially negative influence of NK cells, coordination of NK and phagocytic activity was enriched among convalescents. These seemingly contradictory data point to the potential importance of synergy between innate immune effector functions. While NK cells have been implicated both in protection (Lu et al., 2016 , Jegaskanda et al., 2016 , van Erp et al., 2019 and pathology (Cong and Wei, 2019) , it is possible that the evolution of antibodies able to harness the cytotoxic power of NK cells to eliminate infected or phagocytic cells may play a critical role in elimination and clearance of the infection. Interestingly, this coordination was associated with the synergistic evolution of a broader isotype and subclass specific response among convalescents. However, whether additional changes in antibody-Fc-glycosylation also contribute to this unique functionalization of antibody isotypes and subclasses, enabling coordination, remains unclear but could point to promising target immune profiles that may confer the greatest level of protection against the virus.

There are a number of limitations in this study. First, because these samples were collected early during the COVID-19 pandemic in the US, the Seattle study included a small number of participants, and the groups were not age or sex matched. Confounding factors such as timing of sampling, sex and age are all known to influence SARS-CoV-2 infection and disease trajectory. While antibody profiles clearly segregated individuals that survived compared to that did not survive, more limited variation in antibody profiles were observed across age, sex, viral load and days from symptom onset. However, among the co-morbidities, age was the second major driver of variation in antibody profiles, pointing a potentially critical role for age-associated defects in Fc-variation that may contribute to altered antiviral immunity to SARS-CoV-2 and beyond. The larger validation cohort from Boston identified a similar humoral signature that discriminated survivors from non-survivors, highlighting the conserved nature of this immunological signature, independent of demographic characteristics.

Whereas this study only attempted to understand the humoral disparities between convalescent and deceased individuals in a cohort of severly infected individuals, further studies may attempt to define humoral profiles able to further classify individuals aross the clinical trajectory spectrum ranging from asymptomatic to severe disease.

Collectively, the data presented here argue for the evolution of distinct antigen-specific and functional humoral immune responses early in SARS-CoV-2 disease. While further analysis on longitudinal cohorts may provide further mechanistic insights on the specific role of antibodies in control and clearance of infection, here we validated an early functional humoral immune signature that appears to predict disease progression across two distinct cohorts. Linked to emerging animal model experiments, the correlates defined here may provide key mechanistic insights to guide therapeutic and vaccine design efforts. 

Correspondence. Correspondence and requests for materials should be addressed to G.A.

(galter@partners.org) and Helen Chu (helenchu@uw.edu). 

Acute respiratory distress syndrome 5/12 (41.7) 6/10 (60) Non-ST-elevation myocardial infarction 1/12 (41.7) 5/10 (50) * For 4 of the deceased individuals' days from symptom onset was unknown **For half the recovered individuals viral load measurements were not available 

Further information and requests for resources and reagents should be directed to and will be fulfilled by the Lead Contact, Galit Alter (galter@partners.org).

This study did not generate new unique reagents.

The dataset generated during and/or analyzed during the current study have been made available in the supplemental material.

Plasma samples from 22 SARS-CoV-2 patients from Seattle were profiled for anti-SARS-CoV-2 antibody responses ( Table 1) . Patients who tested positive for SARS-CoV-2 by real-time reverse-transcriptasepolymerase-chain-reaction (RT-PCR) of a nasopharyngeal swab were enrolled in the study upon hospital admission, and samples after admission were included in this study ( Figure S1 ). All enrolled participants gave written, informed consent. The enrolled hospitalized 22 individuals were monitored over the course of their stay, and final outcomes were reported. 12 individuals convalesced and were healthy enough to be discharged, whereas 10 individuals died. Demographic information including age, race, and interventions are summarized across the two groups ( Table 1 and Data S1).

As a validation cohort, a cohort of 40 individuals from MGH in Boston were enrolled, all participants tested positive for SARS-CoV-2 by RT-PCR and they were monitored over their hospital stay. Samples at time of hospitalization were included in this study. Outcomes were reported as deceased or discharged.

Demographics and clinical data for the validation cohort are summarized in Table 2 .

All experimental data was performed in two technical and two biological (for primary cell assays)

replicates and the average value was used throughout the study. This study was approved by the University of Washington Human Subjects Division Institutional Review Board.

Primary immune cells were isolated from fresh peripheral blood from healthy human volunteers collected by the MGH Blood bank or the Ragon institute. The study was approved by the MGH Institutional Review Board. All subjects were over 18 years of age and provided informed consent. All samples were completely de-identified prior to use. Human NK cells and neutrophils isolated from fresh peripheral blood were cultured in RPMI supplemented with with 10% fetal bovine serum, L-glutamine, penicillin/streptomycin and maintained at 37˚C, 5% CO 2 .

THP-1 cells (ATCC) were grown at 37˚C, 5% CO 2 in RPMI supplemented with 10% fetal bovine serum, Lglutamine, penicillin/streptomycin and 0.01% b-mercaptoethanol.

Antigen-specific antibody subclass, isotype, sialic acid, galactose and Fcγ-receptor (FcγR) binding levels were assessed using a 384-well based customized multiplexed Luminex assay, as previously described (Brown et al., 2017) Relative antibody concentration was measured against a panel of SARS-CoV-2 antigens (Data S1). SARS-CoV-2 RBD (kindly provided by Aaron Schmidt), SARS-CoV-2 nucleocapsid (N) protein (Aalto Bio Reagents), and SARS-CoV-2 spike protein (S) (kindly provided by Bing Chen) were used to profile the SARS-CoV-2-specific humoral immune response. Briefly, antigens were coupled by covalent NHS-ester linkages via EDC and NHS (Thermo Scientific) to fluorescent carboxyl-modified microspheres (Luminex). Antigen-coupled microspheres were then washed with an automated plate washer (Tecan) and incubated with plasma samples at an appropriate sample dilution (1:500 for IgG1 and all Fcγ-receptors, and 1:100 for all other readouts). Detection of antigen-specific antibody titers occurred using a PE-coupled detection antibody for each subclass and isotype (IgG, IgG1, IgG2, IgG3, IgG4, IgA1 and IgM, Southern Biotech) , and Fc-receptors were fluorescently labeled with PE before addition to immune complexes (FcγR2A, 2B, 3A, Duke Protein Production facility). For detection of sialic acid and galactose, fluorescein-labeled plant-based lectin detects, SNA and RCA (Vectorlabs) were added as detection reagents at a 1:100 (SNA) and 1:500 dilution (RCA). Plasma samples were acquired via flow cytometry, using an iQue (Intellicyt) and S-Lab robot (PAA). Analysis was done using ForeCyt software by gating on fluorescent bead regions and PE median fluorescent intensity (MFI) is reported as readout for antigen-specific antibody titers.

For the functional analysis of plasma samples, bead-based assays were used to quantify antibodydependent cellular phagocytosis (ADCP), antibody-dependent neutrophil phagocytosis (ADNP) and antibody-dependent complement deposition (ADCD), as previously described (Fischinger et al., 2019) (Data S1). Fluorescent streptavidin beads (Thermo Fisher) were coupled to biotinylated antigen SARS-CoV-2 RBD, N and S and incubated with diluted plasma (ADCP and ADNP 1:100, ADCD 1:10). For ADCP, THP-1 cells were added to the immune complexes and incubated for 16h at 37˚C. For ADNP, primary neutrophils were isolated via negative selection (Stemcell) from whole blood. After 1h incubation at 37°C, neutrophils were stained with an anti-CD66b PacBlue detection antibody (Biolegend). For the ADCD assay, lyophilized guinea pig complement (Cedarlane) was resuspended according to manufacturer's instructions and diluted in gelatin veronal buffer with calcium and magnesium (Boston BioProducts). Post incubation, C3 was detected with Fluorescein-Conjugated Goat IgG Fraction to Guinea Pig Complement C3 (Mpbio).

For detection of antibody-dependent NK cell activity, an ELISA-based approach was used, as described (Boudreau et al., 2020) . Briefly, plates were coated with 2 μg/mL of antigen (as mentioned above) and samples were added at a 1:50 dilution and incubated for 2h at 37˚C. NK cells were isolated the day prior 

The 2019-nCoV pseudoviruses expressing a luciferase reporter gene were generated as described previously (Data S1) . Briefly, the packaging construct psPAX2 (Cat# 11348, AIDS Reagent), luciferase reporter plasmid pLenti-CMV Puro-Luc (Cat# 17447, Addgene) and Spike protein expressing pcDNA3.1-SARS CoV-2.SΔCT were co-transfected into HEK293T cells at ratio of 1:1:0.5 by Calcium phosphate transfection method. The supernatants containing the pseudotype viruses were collected 48 hours post-transfection and filtered by 0.45-µm filter. The viruses were stored at -80°C freezer till use. To determine the neutralization activity of the antisera from vaccinated animals, HEK293T cells were firstly transfected with pcDNA3.1(-)-hACE2 (Cat# 1786, Addgene). 12 hours post transfection; the HEK293T/hACE2 cells were seeded at 96-well tissue culture plate at density of 2.00E+04 cells/well overnight. Heat (56°C, 30 min) inactivated antisera were twofold serial diluted and mixed with 50µl of pseudoviruses. The mixture was incubated at 37°C incubator for 1 hour before adding into HEK293T/hACE2 cells in 96-well plates. Six hours after infection, the cell culture medium was replenished with fresh DMEM (supplemented with 2% FBS). Forty-eight hours after infection, cells were lysed in Steady-Glo Luciferase Assay (Promega). A standard quantity of cell lysate was used in a luciferase assay with luciferase assay reagent (Promega) according to the manufacturer's protocol.

All analyses were performed using python version 3.6.8 with statistical and machine learning packages (Pedregosa et al., 2011) . Networks were visualized in Cytoscape. Raw data are available in supplementary information.

The classification models were trained to distinguish convalescent and deceased groups with a minimal set of features, to avoid overfitting. PBS controls was subtracted from all features, Fc array features were log transformed, and all data was scaled and centered. Antibody features including sex and interventions ( Table 1) were included the selection process, and covariates were binarized and scaled and center prior to analysis.

The models were built using a backward feature elimination for selection and then classified using the minimal set of features which maximize accuracy (Guyon and Elisseeff, 2003; Pittala et al., 2019) .

Models were trained and tested in a fivefold cross-validation framework using random stratified sampling to ensure that both groups are represented in each group. Within each fold, samples were further subdivided into four sets for each iterative fold-specific elimination. A partial least squares discriminant analysis (PLS) classifier was then trained using the fold-specific selected features to predict the test set. Multiple iterations of fold specific feature selections were performed to obtain a single model. This process was repeated over twenty replicates and convergent correlates were observed (Ackerman et al., 2018) .

Performance and robustness of the model was contrasted with negative control models constructed from permuted data and randomly selected size-matched features, with multiple iterations of fivefold cross-validation used to generate classification accuracies. These control models were generated 100 times. The permuted control was generated in the same process as above shuffling labels randomly for each repetition. Size-matched features were chosen at random for each cross-validation step within each repetition. Predicted and true outcomes were compared to determine accuracy. Robustness was defined as the effect size of the distributions (Cliff's ∆), and the exact P values of the tail probabilities of the true distributions within the control distributions. Reported are the median p-values across twenty independent cross-validation replicates (Ojala and Garriga, 2010) .

Correlation networks were constructed to visualize the additional humoral immune features that were significantly linked to the selected minimal biomarkers, to provide enhanced insights into the biological mechanisms by which antibodies may provide protection following infection. In brief, antibody features that were significantly correlated with a Holms-Bonferroni correction to the final selected PLS model selected-features were defined as co-correlates. Significant spearman correlations above a threshold of |r| > 0.5 were visualized within the networks.

Using the selected features from the original model a new PLSDA model was trained excluding a single outlier at a time in a fivefold cross validation framework. This process was repeated three times, each time generating a unique ROC curve as the top 3 individual outliers were removed. Using these cross validated ROC curves the mean performance and variation were assessed and are summarized as area under curve.

In order to evaluate S vs N ratios, first ratios for each feature were defined separately by simply dividing S-responses over N-responses for every given feature. S:N ratios were visualized by log2 transformation for ease of interpretation. Differences across convalescents and deceased were then tested with a onesided Mann-Whitney U test and a Holm-Bonferroni multiple hypothesis correction criterion.

In order to address whether the overall S-response was enriched over N-responses in the convalescents across all features tested, all data was background corrected and z-score normalized. Then the number of S-features which were greater than their N-counterparts across every feature were summed. This analysis yielded a distribution of individual S greater than N scores for each group and statistical differences were assessed using a one-sided Mann-Whitney U test.

Route of immunization defines multiple mechanisms of vaccinemediated protection against SIV

A serological assay to detect SARS-CoV-2 seroconversion in humans

How antibodies alter the cell entry pathway of dengue virus particles in macrophages

Protective efficacy of adenovirus/protein vaccines against SIV challenges in rhesus monkeys

Covid-19 in Critically Ill Patients in the Seattle Region -Case Series

Selective induction of antibody effector functional responses using MF59-adjuvanted vaccination

Multiplexed Fc array for evaluation of antigen-specific antibody effector profiles

Contributions of the structural proteins of severe acute respiratory syndrome coronavirus to protective immunity

SARS-CoV-2 infection protects against rechallenge in rhesus macaques

Longitudinal analysis of Severe Acute Respiratory Syndrome (SARS) coronavirus-specific antibody in SARS patients

Clinical and immunological features of severe and moderate coronavirus disease 2019

Polyfunctional Fc-effector profiles mediated by IgG subclass selection distinguish RV144 and VAX003 vaccines

Immune history and influenza virus susceptibility

SARS and MERS: recent insights into emerging coronaviruses

Broadly neutralizing anti-influenza antibodies require Fc receptor engagement for in vivo protection

A highthroughput, bead-based, antigen-specific assay to assess the ability of antibodies to induce complement activation

Modulating Antibody Functionality in Infectious Disease and Vaccination

An Introduction to Variable and Feature Selection

Estimation of SARS-CoV-2 Mortality during the Early Stages of an Epidemic: A Modelling Study in Hubei, China and Northern Italy

A single amino acid substitution (R441A) in the receptor-binding domain of SARS coronavirus spike protein disrupts the antigenic structure and binding activity

First Case of 2019 Novel Coronavirus in the United States

In vivo systems analysis identifies spatial and temporal aspects of the modulation of TNF-alpha-induced apoptosis and proliferation by MAPKs

Substantial undocumented infection facilitates the rapid dissemination of novel coronavirus (SARS-CoV2)

Long-term persistence of robust antibody and cytotoxic T cell responses in recovered patients infected with SARS coronavirus

Evaluation of Nucleocapsid and Spike Protein-based ELISAs for detecting antibodies against SARS-CoV-2

Permutation tests for studying classifier performance

Antibody Fab-Fc properties outperform titer in predictive models of SIV vaccineinduced protection

The SARS-CoV-2 receptor-binding domain elicits a potent neutralizing response without antibody-dependent enhancement. bioRxiv

Antibody-mediated protection against Ebola virus

Antibody responses against SARScoronavirus and its nucleocaspid in SARS patients

Human monoclonal antibody combination against SARS coronavirus: synergy and coverage of escape mutants

Cloning, sequencing, expression, and purification of SARS-associated coronavirus nucleocapsid protein for serodiagnosis of SARS

HIV-1-specific antibody responses during acute and chronic HIV-1 infection

Coronavirus disease (COVID-19) Situtation Report -122

Neutralizing antibody responses to SARS-CoV-2 in a COVID-19 recovered patient cohort and their implications. medRxiv

A new coronavirus associated with human respiratory disease in China

A DNA vaccine induces SARS coronavirus neutralization and protective immunity in mice

Phagocytic cells contribute to the antibody-mediated elimination of pulmonaryinfected SARS coronavirus

DNA vaccine protection against SARS-CoV-2 in rhesus macaques

Passive immunotherapy with dromedary immune serum in an experimental animal model for Middle East respiratory syndrome coronavirus infection

Passive immunotherapy for Middle East Respiratory Syndrome coronavirus infection with equine immunoglobulin or immunoglobulin fragments in a mouse model

A Novel Coronavirus from Patients with Pneumonia in China

Generation and Protective Ability of Influenza Virus-Specific Antibody-Dependent Cellular Cytotoxicity in Humans Elicited by Vaccination, Natural Infection, and Experimental Challenge

Fc-Mediated Antibody Effector Functions During Respiratory Syncytial Virus Infection and Disease

Scikit-Learn: Machine Learning in Python

We thank Bruce Walker, Nancy Zimmerman, Mark and Lisa Schwartz, an Anonymous donor, and Terry and Susan Ragon for their support. We would also like to thank Bing Chen for protein production efforts. We acknowledge support from the Ragon Institute of MGH, MIT, the Massachusetts Consortium on Pathogen Readiness (MassCPR), and the Bill & Melinda Gates Foundation (235730), the NIAID (U19 AI35995) and the U.S. Centers for Disease Control and Prevention (CK000490).