Written consent was obtained from the participants or, in some cases their legally authorized representatives. Participation in these studies was voluntary. All research protocols were approved by the Duke University Health System Institutional Review Board (IRB) and performed in accordance with the Declaration of Helsinki. The IRB-approved protocol number for the Molecular and Epidemiological Study of Suspected Infection (MESSI) is Pro00100241. In addition, we utilized previously biobanked samples, which have the following study identifiers and IRB-approved protocol numbers: Validation of Clinical Molecular Predictors in Pre-symptomatic Infectious Diseases (Pro00001176) and the associated Biorepository protocol (Pro00001698).

We enrolled members of the community and hospital patients into the Molecular and Epidemiological Study of Suspected Infection (MESSI), an open prospective cohort study of infections in North Carolina in people aged 2 or older, as previously described (36, 37). The purpose of MESSI is to collect samples and data for research from individuals who have been exposed to someone with an infection, suspected to have an infection, or have been diagnosed with an infection, including COVID-19, to study biomarkers of infectious diseases and better diagnose infection. In this study, we included participants enrolled during the introductory pandemic of SARS-CoV-2 in North Carolina between March 2020 and November 2021. We used samples and clinical data collected during the pre-vaccination era in this observational study of natural infections. To our knowledge, the timepoints included in this study represent one infection event per person. We did not ascertain any SARS-CoV-2 infections that may have occurred prior to enrollment. Given that the study period was during the introductory pandemic of SARS-CoV-2 in North Carolina, with alpha and delta variants, it is likely that we sampled the first infection for most of our study participants, though this cannot be confirmed. Community-based participants were identified through recruitment at homes, a homeless shelter, skilled nursing facilities, hospice care facilities, and long-term care centers, whereas hospitalized participants were identified and recruited through the Duke University Health System (DUHS) or the Durham Veterans Affairs Health System (DVAHS). Upon written consent, participants provided symptom and demographic information, a nasal swab, and blood sample at enrollment. Subsequently, we followed up with participants for symptom information and sampling every few days in the first week, then weekly through the first month, and then at 2, 6, and 12 months post enrollment, where possible. This approach provided longitudinal blood samples and clinical information in the participants who consented to our follow up and were able to be reached. Time of symptom onset was ascertained by participant recall. In addition to this main study group, we used previously biobanked samples from other studies collected before 2020 (pre-SARS-CoV-2 pandemic) to generate comparator groups of cytokine responses in healthy individuals with no apparent infection (n = 18).

The presence of SARS-CoV-2 natural infection was defined by the following: (A) PCR positive in nasopharyngeal swab or plasma within 14 days of symptom onset or first visit and (B) positive result on blood test for viral antigen-binding IgM or IgG commercial tests (EUROIMMUN or Abbott IgG/IgM) within 21 days or symptom onset. Data on symptoms, severity of illness, and medications were collected from case report forms and review of the medical record. SARS-CoV-2 vaccinees were censored from this study at the date of first vaccine dose, and their data were excluded from that point onward. This allowed for controlled study of natural SARS-CoV-2 immunity. If a COVID-19 case was treated in the hospital at any point during a participant’s illness, we assigned that case to the Hospitalized analytic group, whereas those COVID-19 cases not treated in the hospital at any point during their illness were assigned to the Community analytic group.

Viral load in nasal swab was assessed as previously described (38, 39). Briefly, nasal swab viral transport medium (VTM) was aliquoted and cryopreserved from study subjects to determine SARS-CoV-2 N1 gene copy number by RT-PCR to stratify subjects such as COVID PCR positive or negative. Viral RNA was extracted from 140 µl of VTM according to manufacturer’s instructions (QiaAmp Viral RNA minikit). SARS-CoV-2 nucleocapsid (N1) and human RNase P (RPP30) RNA copies were determined using 5 µl of isolated RNA in the CDC-designed kit (CDC-006-00019, Revision: 03, Integrated DNA Technologies 2019-nCoV kit). Standard quantitative RT-PCR (TaqPath 1-step RT qPCR Master Mix, Thermofisher) was run with test RNA and gene-specific standard curves (2 x 10⁵–20 copy/ml). Regression analysis was used to determine gene copy number and corrected to report copies/ml of VTM. Samples with a Ct value greater than 35 PCR cycles were called SARS-CoV-2 PCR negative and samples less than or equal to 35 were called SARS-CoV-2 PCR positive.

QIAGEN QIAsymphony DSP Virus/Pathogen Midi Kit (96-well QIAgility) was used for viral RNA extraction from 0.4 or 0.8 ml of plasma. The envelope gene of SARS-CoV-2 was reverse transcribed and amplified by PCR on an Applied Biosystems QuantStudio 3 Real-Tim PCR System. A standard curve was run with each batch to extrapolate RNA copies/ml. The lower limit of quantification (LLOQ) for this assay was 62 RNA copies per mL (1.79 log₁₀) with 0.8 ml of plasma input or 128 RNA copies per ml with 0.4 ml of plasma diluted 1:1. Values below the LLOQ may be outside of the 95% confidence interval for reproducibility. Any positive SARS-CoV-2 viral load detected was considered “positive” and not detected was considered “negative.”

Over the course of the early pandemic and this study, multiple kits were used to detect spike-binding IgG and IgM antibodies in blood samples. Human EDTA-plasma or serum were used for all assays.

The magnitude of anti-spike and anti-nucleoprotein IgG, IgA, and IgM antibodies was assessed via ELISA. A total of 384-well high-binding plates (Corning) were coated for 2.5 h at room temperature (RT) with 2 μg/ml of recombinant SARS-CoV2 spike S2 Extracellular domain (Sino Biological 40590-v08b) or nucleocapsid protein (Sino Biological 40588-v08b) in CBC buffer (15 mM Na₂CO₃, 35 mM NaHCO₃, pH 9.5; all Sigma). Antigen was removed by aspiration, and then plates were blocked overnight at 4°C in CBC, 5% goat serum, 0.01% Kathon, 0.05% Tween-20 (all Sigma), and washed in PBS (ScyTek) 0.1% Tween-20 (Sigma). Heat inactivated sera (1 h, 60°C) were diluted 1:30 in PBS, 5% goat serum, 1% milk, 0.05% Tween-20, 0.01% Kathon, threefold serially diluted to 1:143,489,070 in the same buffer and then applied to the antigen-coated, blocked plates. Plates were incubated overnight at 4°C, washed, and then horseradish peroxidase (HRP)-conjugated detection antibodies (goat anti-IgG, goat anti-IgM both Jackson Labs; goat anti-IgA from Southern Biotech) were applied at optimized dilutions in PBS, 10% goat serum, 0.05% Tween-20. Plates were incubated for 105 min at RT, washed, and then TMB-Ultra HRP substrate (Pierce) added. Plates were incubated for 14 min at room temperature and then the HRP reaction stopped by the addition of 2 N H₂SO₄ (Sigma). Plate absorbance at 450 nm was immediately read using a SpectraMax Plus plate reader (Molecular Devices). Area under the titration curve was determined using the trapezoid method using Microsoft Excel. All steps, except transfer to cold storage, were performed by a BioMek i7 liquid handler.

Cytokines were assayed in duplicate using non-heat-inactivated undiluted serum with a commercial 47-plex bead array (Millipore HCYTA-60K-47C) performed according to the manufacturer’s instructions and read using a Flex-MAP 3D bead reader (Luminex). Data were analyzed using Bio-Plex manager v6.2. The 47 measured cytokines and chemokines included the following: interferon alpha (INF-α), interferon gamma (INF-γ), interleukin 1 alpha (IL-1α), interleukin 1 beta (IL-1β), interleukin 1 receptor antagonist (IL-1Ra), interleukin 2 (IL-2), interleukin 3 (IL-3), interleukin 4 (IL-4), interleukin 5 (IL-5), interleukin 6 (IL-6), interleukin 7 (IL-7), interleukin 8 (IL-8), interleukin 9 (IL-9), interleukin 10 (IL-10), interleukin 12 heterodimer p40 and p35 (IL-12p70), interleukin 12 homodimer p40 (IL-12p40), interleukin 13 (IL-13), interleukin 15 (IL-15), interleukin 17 alpha (IL-17α), interleukin 22 (IL-22), interleukin 27 (IL-27), interleukin 18 (IL-18), interleukin 17E (IL-17E), interleukin 17F (IL-17F), macrophage colony-stimulating factor (M-CSF), monokine induced by gamma (MIG), tumor necrosis factor alpha (TNF-α), epidermal growth factor (EGF), fibroblast growth factor (FGF), granulocyte colony-stimulating factor (GCSF), granulocyte–macrophage colony-stimulating factor (GM-CSF), monocyte chemotactic protein 1 (MCP-1), interferon gamma-induced protein 10 (IP-10), macrophage inflammatory protein 1 alpha (MIP-1α), macrophage inflammatory protein 1 beta (MIP-1β), tumor necrosis factor beta (TNF-β), vascular endothelial growth factor (VEGF), FMS-like tyrosine kinase 3 ligand (Flt-3L), fractalkine, growth-related protein (GROα), monocyte chemotactic protein-3 (MCP-3), macrophage derived chemokine (MDC), platelet derived growth factor AA dimer (PDGFa), platelet derived growth factor AB and BBdimer (PDGFab), soluble CD40 ligand (sCD40L), RANTES (CCL5), eotaxin, and transforming growth factor alpha (TGF-α).

Vero E6 cells (ATCC CRL-1586) were seeded at 4 x 10⁴ cells/well in black 96-well plates in growth medium made of MEM + Earl’s salts + L-glutamine (Gibco 11095), 100 U/ml of penicillin and 100 µg/ml of streptomycin (Gibco 15140), 1 mM sodium pyruvate (Gibco 11360), 1× non-essential amino acids (Gibco 11140), and 10% fetal bovine serum (Gemini 100-106). Cells were allowed to adhere overnight. Serum samples were heat-inactivated for 1 h and diluted twofold in virus diluent, which is the same as the growth medium except with 2% fetal bovine serum instead of 10%. Each serum sample was incubated with 100 TCID50 fluorescent SARS-CoV-2 (40), strain USA/WA1/2020, for 1 h at 37°C. Following incubation, the serum/virus solution was added to Vero E6 cells and incubated at 37°C for 24 h. Then, cells were fixed using 10% neutral-buffered formalin and washed with PBS. Fluorescent signal was measured using a BioTek Synergy H1 microplate reader. Percent inhibition was calculated and used to determine IC₅₀ values with non-linear regression analysis tools from GraphPad Prism. In each assay, samples were measured in duplicate alongside established positive and negative controls.

To identify a group of key cytokines with independent contributions toward hospitalization, each cytokine was evaluated in a stepwise manner to produce a minimal unbiased cytokine set related to outcome for each <6 DPOS and 8–12 DPOS periods. For all analyses, each cytokine (C₁) was centered to the mean of log₂(pg/ml) for the healthy group, and the reference was baseline odds for a 45-year-old male with mean level of cytokines as per the healthy group. First, each cytokine (C₁) was regressed onto the outcome of hospitalization in a univariable regression model (Equation [Eq1] in Supplementary Table S3 ). Second, each cytokine (C₁) with potential for independent contribution toward hospitalized from Eq 1 moved forward into Eq 2 and assessed in multivariable logistic regression model for significant contribution to hospitalization with adjustment for known confounders of age and sex (Eq 2; Supplementary Table S3 ). Third, Eq 2 was expanded to include the potential of age and sex as effect modifiers for each cytokine (Eq 3; Supplementary Table S3 ). Based on these relationships, independently contributing cytokines and their effect modifiers (p < 0.05 in any of the models of Eq 2 or 3) were tested together in a multivariable logistic regression to identify a minimal combination at each <6 and 8–12 DPOS, which optimally explains COVID-19 hospitalization. A top–down strategy was applied to build three models for each period, and the optimal model was selected on the basis of goodness of fit assessed with the lowest Akaike information criterion. All models included age and sex as known confounders. The final multivariable logistic regression model of cytokines for <6 DPOS was Equation 4 ( Supplementary Table S3 ). The final multivariable logistic regression model of cytokines for 8–12 DPOS was Equation 5 ( Supplementary Table S3 ).

A generalized estimating equation (GEE) model was generated for SARS-CoV-2 neutralizing antibodies, and each binding antibody population was measured (n = 6): anti-spike IgG, IgM, and IgA; anti-nucleoprotein IgG, IgM, and IgA ( Supplementary Figure S3 ). The outcome for each model was the level of antibody, where the magnitude of antibody was measured in terms of log₂ of area under the curve, and neutralizing antibodies in terms of log₁₀ of 50% maximal virus neutralizing titer (IC₅₀). All models were adjusted for intercorrelation due to inclusion of multiple longitudinal measures from the same individual over time. Time bins for comparison were 0–20 (bin #1) and >20–40 (bin #2) DPOS, where interval 2 denotes the intercept for time bin 2. The reference group was <40 years old males among the community COVID-19 cases (Eq 6; Supplementary Table S3 ). To assess rates of change in antibody level over time among community COVID-19 cases, a generalized linear mixed effects model was applied by time bins: 0–20 (bin #1), >20–40 (bin #2), >40–100 (bin #3), and >100 (bin #4). Int1, 2, 3, and 4 define the intercepts for each of the aforementioned time bins. The reference group was 45-year-old males in the community ( Supplementary Figure S4 ; Eq 7 in Supplementary Table S3 ). Data were analyzed in Stata and visualized with either Stata, GraphPad Prism, or Microsoft Excel.

In our observational cohort, we enrolled 639 participants from the community and hospital settings in the state of North Carolina, USA, and collected data on 2,736 person-visits from March 2020 through November 2021. Our main objective was to understand immune responses to natural infection in the community. Thus, 84% of the participants’ first visit occurred in the community: 67% at home and 17% in communal living conditions including hospice care, long-term acute care, skilled nursing facilities, or homeless shelters ( Supplementary Table S1 ). During our observational study, the alpha, beta, and delta lineages of SARS-CoV-2 circulated in North Carolina, and we included timepoints before vaccine availability (41, 42). We found that 406 individuals were SARS-CoV-2 test positive at the time of enrollment (COVID-19 group), 67 had proven non-COVID-19 related illnesses, and 166 had an indeterminate diagnosis at the time of analysis (defined as clinical symptoms of possible infection without positive confirmatory testing, Supplementary Table S1 ; Supplementary Figure S1 ). The COVID-19 group includes 72 individuals whose diagnosis was confirmed through self-reported molecular or antibody tests. During our study period, more COVID-19 cases (29%) were treated in the hospital than in other etiologies (4%–10%; Supplementary Table S1 ).

From the COVID-19-positive participants (n = 406), we formed comparator groups of severe versus mild disease using hospitalization as a proxy for severity. We compared individuals treated in the hospital (n = 114, hospital group) versus not treated in the hospital (n = 258, community group) during their illness ( Table 1 ; Supplementary Figure S1 ). Of note, 34 participants were excluded from analysis due to missing virologic or immunologic data. Mean age was significantly higher (54–60 years) in the hospitalized group compared to the community group (40-44 years; Wilcoxon signed-rank test, p < 0.05; Table 1 ). Sex was not significantly different by hospitalization (Fishers exact test; Table 1 ). The races of participants were significantly different across groups with fewer White participants and more Black participants in the hospitalized group than in the community group (Fishers exact test, p < 0.05; Table 1 ). Other races of participants included in this study were Asian (n = 20), Native Hawaiian-Pacific Islander (n = 1), “More than one” race (n = 22), American Indian-Alaska Native (n = 6), and unknown/unspecified (n = 6). Hispanic ethnicity was comparable across groups (14% versus 15%; Fishers exact test; Table 1 ). The median day of first visit was 11 and 8 days post symptom onset (DPOS) in the community and hospital groups, respectively ( Table 1 ). At the first visit, 63% of all COVID-19 participants demonstrated detectable viral RNA (vRNA) in nasopharyngeal (NP) swab ( Table 1 ). The NP swab viral load at first visit and peak within 20 DPOS (4.5–4.7 log₁₀ vRNA copies/ml) were comparable across groups, as well as the timing of peak viral load on the ninth DPOS (Wilcoxon signed-rank test; Table 1 ). However, in line with our prior findings, a significantly greater portion of the hospital group demonstrated viremia in blood than the community group at first visit (27% versus 2%; Fishers exact test, p < 0.05; Table 1 ) (43). In the hospital group, 25% of the participants were admitted to the intensive care unit (ICU), 14% were intubated, and 9% died. Among our study participants, older age, viremia, and Black race were associated with severe disease as measured by hospitalization.

Using a 47-plex cytokine assay, we characterized early serum cytokines upon SARS-CoV-2 infection in different individuals from the first (<6 DPOS; n = 49) versus second week (8–12 DPOS; n = 76) of illness ( Supplementary Figure S1 ). These time periods distinguish the virologic versus immunologic phases of pathology (13). As a reference group, we selected serum samples from 18 healthy (i.e., no acute illness or infection) North Carolina individuals collected before 2020 (i.e., pre-pandemic). Of the 47 cytokines measured, 24 were at significantly different serum concentrations (log₂ pg/ml) in early SARS-CoV-2 infections compared to those of the healthy (Wilcoxon signed-rank test; p < 0.05). Of these, 19 were different at both time periods (FGF, GM-CSF, IL-1Ra, IL-2, IL-3, IL-4, IL-6, IL-8, IL-12p70, IL-13, IL-17, IL-18, IL-22, IP-10, MCP-1, M-CSF, MIG, MIP-1α, TNF-α), four were different only at <6 DPOS (EGF, PDGFABBB, IL-10, and GROα), and TNF-α was different only at 8–12 DPOS ( Figure 1 and Table 2 ). Thus, we used an unbiased approach to identify subsets of cytokines that were differentially expressed in COVID-19 in the first versus second week of illness.

We then asked whether the directionality of cytokine dysregulation was related to severe COVID-19 by comparing community versus hospitalized groups ( Supplementary Figure S1 ). We calculated the mean log₁₀ fold difference in cytokine level relative to healthy individuals, where above the mean of healthy indicated an upregulation and below the mean indicated a downregulation ( Figure 1 ). Some hospitalized patients received antiviral or immunomodulatory medications for COVID-19 or participated in therapeutic clinical trials ( Supplementary Table S2 ), whereas the community group did not receive any recorded clinical treatments. We found that IL-8, IL-6, and IP-10 were the most highly upregulated in all COVID-19 community and hospital groups relative to healthy participants, indicating that their upregulation is a hallmark of all SARS-CoV-2 infections ( Figure 2 ). However, IL-2 and GM-CSF were upregulated in the community and downregulated in the hospital group at <6 DPOS, suggesting that their early upregulation may be involved in milder disease trajectories ( Figure 1 ). Also, IL-17 was upregulated in the hospital group but downregulated in the community group at 8–12 DPOS, suggesting that it may have a role in severe pathology ( Figure 1 ).

Next, we compared shifts from the first (<6 DPOS) to second week (8–12 DPOS) of illness. In the community group, mean levels (log₂ pg/mL) of GROα, IL-1R antagonist, IP10, and MCP1 were significantly lower in the second week (n = 43) compared to the first week (n = 22; Mann–Whitney test, unadjusted p < 0.05; Table 2 ). In contrast, in the hospital group, GROα, IL-1R antagonist, and MCP1 were at similar levels, and only IP10 was significantly lower in the second week (n = 33) compared to those of the first week (n = 27; Mann–Whitney test, unadjusted p < 0.05; Table 2 ). This suggests that prolonged GROα, IL-1R antagonist, and MCP1 may be involved in COVID-19 pathology.

To define whether the magnitude of cytokines during COVID-19 differed with severity, we tested differences in mean cytokine levels in community versus hospital groups (post hoc Mann–Whitney test with Bonferroni multiple test correction, α < 0.001; Supplementary Figure S1 ). We found that IL-6 was significantly higher in the hospitalized group at both <6 DPOS and 8–12 DPOS than in the community group (p < 0.001; Figures 2A, B ). Also, IL-1R antagonist, IP-10, and MIG were significantly higher in the hospitalized group at 8–12 DPOS than in the community group (p < 0.001; Figures 2C–E ). In contrast, the community group elicited 2- to 10-fold lower levels of the same cytokines at <6 DPOS, which further decreased by 8–12 DPOS ( Table 2 ). Thus, we found that an early and high level of IL-6 as well as the maintenance of high levels of IL-6, IL-1Ra, IP10, and MIG through 8–12 were related to hospitalized outcome.

Since cytokines operate in concert with each other to shape downstream immune responses, we defined minimum sets of cytokines that contribute to increased odds of hospitalization at each <6 and 8–12 DPOS using logistic regression models ( Figure 3 ). We directly compared the hospital versus community groups, by testing whether a twofold increase in serum cytokine concentrations was associated with a significantly greater likelihood of being in the hospital group than the community group using odds ratio (OR). We also assessed for interaction effects to test whether the relationship of each cytokine with hospitalization depended on either sex or age. At <6 DPOS, a model with FLT3, IL-17, IL-6, IL-15 emerged as the minimal cytokine set that explained 62% of variability in COVID-19 hospitalization (Akaike information criterion or AIC = 41.4; n = 49; Figure 3A ). Specifically, a twofold increase in the level of IL-6 was significantly associated with two-times greater odds of hospitalization relative to the community group [OR = 2.07 (95% CI = 1.04–4.09), p < 0.05]. At <6 DPOS, higher age, female sex, and higher levels of FLT3, IL-17, and IL-15 were not significantly associated with hospitalization in our study participants ( Figure 3A ). While IL-15 may be slightly related to reduced odds of hospitalization with increasing age, this association was not significant.

In the 8–12 DPOS period, a larger combination of cytokines as well as age- or sex-dependent effects were required to optimally explain 60% of variability in hospitalization (AIC = 72, n = 76; Figure 3B ). We identified that a twofold increase in IL-10 was significantly associated with a 63% reduction in odds of hospitalization, relative to the community group [OR = 0.37 (95% CI = 0.16–0.81), p < 0.05; Figure 3B ], whereas a twofold increase in IL-6 [OR = 2.26 (95% CI = 1.2–4.2)] and IP10 [OR = 1.82 (95% CI = 0.87–3.8)] was significantly associated with 1.8- to 2.2-fold increased odds of hospitalization (p < 0.05; Figure 3B ). A twofold increase in IL-1R antagonist led to a significant 10-fold increase in the odds of hospitalization relative to the community group [OR = 9.7 (95% CI = 1.7–54], p < 0.05; Figure 3B ). However, IL-1R antagonist levels also demonstrated a small but significant association with reduced odds of hospitalization in higher age groups [OR = −0.09 (95% CI = −0.16 to −0.02), p < 0.05; Figure 3B ]. Meanwhile, an increase in age significantly raised odds of hospitalization in the second week of illness by 16% in comparison to the community group [OR = 1.25 (95% CI = 1.07–1.46), p < 0.05; Figure 3B ]. Female sex was not significantly associated with the hospital group. While it is possible that MDC and IL-6 may slightly reduce the odds of hospitalization in older age groups and females, and GROα may slightly increase the odds of hospitalization in older age groups, these associations were not significant. Both models revealed that distinct groups of cytokines are related to odds of hospitalization in the first versus second week of COVID-19 and explain more variability in the odds of hospitalization than age and sex only (AIC = 137–161; n = 129–145).

Antibodies that bind viruses can target the virus for destruction, whereas the subset of antibodies that neutralize the virus can prevent infection of host cells. So, we assessed how the magnitude of early SARS-CoV-2-binding and -neutralizing antibody responses differed in community (n = 120–153) versus hospitalized (n = 61–73) SARS-CoV-2 infections. Surprisingly, the hospital group demonstrated significantly greater SARS-CoV-2-neutralizing antibody, and binding IgG, IgM, and IgA responses against SARS-CoV-2 spike (S) and nucleoprotein (NP) than the community at enrollment (Wilcoxon signed-rank test, p < 0.05; Figures 4A–D and Supplementary Figures S2A–C ). This persisted through the first 40 days from symptom onset, such that peak antibody level within 40 DPOS was significantly higher in the hospital versus community group for neutralizing antibodies, S-binding IgG and IgA, and NP-binding IgG, IgA, and IgM ( Supplementary Table S2D ). In each community and hospital group, males generally demonstrated higher SARS-CoV-2-binding antibody levels than females, and antibody levels increased with age from <40, to 40–60, and >60 years of age ( Supplementary Figure S3 ). To explain higher antibody levels in more severe cases, we further probed the protective effects of antibodies and asked whether high SARS-CoV-2 antibody responses were related to low early viral burden in the respiratory tract. To test this, we examined the relationship between viral load in the nasopharyngeal swab at the first visit versus serum antibody levels ( Figures 4E–H ). We found that S-binding-IgG, -IgA, and -IgM, as well as neutralizing antibody levels from the first visit were negatively correlated (rho −0.49 to −0.32) with viral load (Spearman rank correlation, p < 0.001). Thus, high early SARS-CoV-2 antibodies were related to lower early viral burden, but not hospitalization.

Due to a longer follow up of 156 community participants, we were able to evaluate SARS-CoV-2 antibody kinetics up to 6 months post symptoms. We modeled the dynamics of antibody responses as linear rates over four time bins and found that each time bin had distinct rates, as characterized by a steep rise (orange; ≤20 DPOS), then a short and steep decline (blue; >20–40 DPOS), then a period of stabilization (red; >40–100 DPSO), and finally a long-term steady state (green; >100 DPOS; Figure 5 and Supplementary Figure S4 ). Rates were calculated as twofold change in antibody level for each day post symptom onset ( Figure 5 and Supplementary Figure S4 ). At <20 DPOS, neutralizing antibodies, anti-S IgG, and anti-NP IgG rose faster (rates = +0.12 to +0.15) than antiviral IgA and IgM antibodies (rates = +0.09 to +0.11; Supplementary Figure S4 ). In the >20–40 DPOS period, neutralizing antibodies, antiviral IgG, and anti-S IgA declined at a slower rate (+0.0008 to −0.03) than antiviral IgM and anti-NP IgA (rate = −0.06 to −0.07; Supplementary Figure S4 ). Most antibody populations underwent a slight increase during the >40–100 DPOS period, including antiviral IgA and IgM, neutralizing antibodies, and anti-S IgG (rates= +0.01 to +0.07; Figure 5 and Supplementary Figure S4 ). Importantly, the longest-term follow-up of >100 DPOS revealed the presence of long-lived antiviral IgM and IgA antibodies, as well as a stable rate for neutralizing and long-lived anti-S IgG antibodies (rate = +0.005 to −0.014; Figure 5 and Supplementary Figure S4 ).