Main

Coronavirus SARS-CoV-2 is the etiologic agent for COVID-19 (ref. 1), a disease from which over 500,000 persons have died in the past 6 months (https://www.who.int/emergencies/diseases/novel-coronavirus-2019). Nonetheless, most SARS-CoV-2-infected individuals recover without hospitalization, with many remaining pauci- or asymptomatic2, while making strong immune responses3. Hence, the threat posed by SARS-CoV-2 might be mitigated by better understanding and management of COVID-19. This particularly applies to its potential for phasic progress from an initial syndrome typified by fever and respiratory symptoms, to one or more severe complications, including acute respiratory distress syndrome (ARDS), marked coagulopathy, hyperinflammation and multi-organ failure4. Better understanding should also increase our preparedness for any future zoonotic coronavirus transmissions.

To this end, it is important to characterize the host–pathogen relationship, including correlates of immunoprotection, such as virus-specific antibodies that limit disease5 and correlates of immune dysregulation, such as cytokine over-production that may promote disease6,7. Collectively, those correlates can compose a core disease-associated immune signature, various elements of which may exist in other infections and/or immunopathologies, including but not limited to those caused by other human and animal coronaviruses8.

However, the prospect of a consensus COVID-19 immune signature is challenged a priori by the patient heterogeneity spanning ethnicity, age, sex, clinical presentation and underlying illness9,10. Thus, the development of SARS-CoV-2-specific antibodies and T cells3,5,11,12 have been variably set against descriptions of lymphopenia and immunodeficiency13,14, aligned with the transfusion of patients with antibody-rich convalescent plasma15; and descriptions of neutrophilia and cytokine storms6,16,17, for which immunosuppressants have been prescribed18.

Acknowledging this complexity, we nonetheless considered that a core immune signature may exist, akin to the capacity of highly diverse individuals to make signature responses to vaccination19 or to display similar immunological dysregulation in sepsis20. Thus, our study sought to identify immunological traits common to most individuals among a highly heterogeneous cohort of 63 patients with COVID-19 treated at Guy’s and St Thomas’ Hospitals, London. The study, termed COVID-IP (COVID-immunophenotyping), also included 55 healthy controls (HCs), including 23 previously exposed seropositive individuals and 10 patients hospitalized for non-COVID-19 lower respiratory tract infections (LRTIs).

Although COVID-19 is mostly defined by pneumonitis and organ dysfunctions reflecting local as well as systemic inflammation6, COVID-IP focused on peripheral blood analyses for several practical reasons: their applicability in multiple settings and locations thereby facilitating meta-analyses; their ready contextualization with clinical blood measurements; their suitability for longitudinal assessments; and their established capacity to reflect, albeit incompletely, immune cells trafficking to and from tissues21,22.

The data obtained clearly identify core immunological traits common to most patients with COVID-19. Several were shared with the non-COVID-19 LRTI cohort and have previously been observed in COVID-19, in vaccine responses and immunopathologies such as sepsis and severe influenza. Other traits have been less-well described, with some being overtly severity-related, thereby offering parameters for tracking disease progression and for better understanding COVID-19 pathobiology. Furthermore, some traits measured at initial bleed anticipated rapid patient recovery and the duration of hospitalization: prognoses that may be particularly beneficial in managing a clinically heterogeneous disease in the context of limited resources. To facilitate independent validation, the signature COVID-IP dataset provided here is supported by a comprehensive online portal (www.immunophenotype.org).

Results

A COVID-19 immune signature

In seeking a COVID-19 immune signature, we chose easily transferable, high-throughput standard-operating protocols that increased opportunities for global meta-analysis. Thus, we employed eight multiparameter flow cytometry panels (P1–P8), measuring: (1) broad lymphocyte composition; (2) effector/memory T cell status; (3) γδ T cell status; (4) B cells; (5) cell cycling; (6) leukocyte counts; (7) lymphocyte activation and exhaustion; and (8) innate immune cells. P1–P5 were applied to peripheral blood mononuclear cells and P6–P8 to whole blood. Gating strategies are described and illustrated in Supplementary Materials. Also measured were 22 cytokines and antibodies against SARS-CoV-2 nucleocapsid (N), spike (S) and receptor binding domain (RBD), and against autoantigens. Because COVID-IP included longitudinal analyses, we use n to refer to sample numbers and n′ for patient numbers.

The World Health Organization’s (WHO) eight-point scale for COVID-19 trial endpoints (http://www.who.int/blueprint/priority-diseases/key-action/novel-coronavirus/en/) was employed to classify ordinal severity: ‘low’ (WHO scores 1–2) (n′ = 6) reflected mild COVID-19 symptoms of patients scoring SARS-CoV-2(+) who were hospital-treated for other reasons; ‘moderate’ (WHO scores 3–4) (n′ = 26) reflected COVID-19 but with little or no requirement for supplemental oxygen; and ‘severe’(n′ = 31) reflected a disease course, including any or all of: high-flow oxygen requirement (score 5; n′ = 14), mechanical ventilation (6; n′ = 9), multi-organ support (7; n′ = 3) and death (8; n′ = 5). Patient classification reflected peak severity, although clinical scores were recorded with every sampling (www.immunophenotype.org), permitting identification of immunological parameters which tracked and/or anticipated changes in clinical status (Extended Data Fig. 1).

Patients with COVID-19 were compared with HCs (n′ = 55), including seropositive individuals (n′ = 23) who had previously experienced either asymptomatic (n′ = 15) or mild (score 1–2; n′ = 8) infection, without requiring hospitalization (Extended Data Fig. 1; www.immunophenotype.org). Seropositivity classification required RBD-specific and/or S-specific antibodies, as N-specific antibodies might have been induced by seasonal coronaviruses23.

Additionally, the impact of LRTI was controlled for by analyzing ten patients hospitalized for LRTI but repeatedly scoring as SARS-CoV-2-negative. They matched the COVID-19 cohort for age, comorbidities and hospitalization time, with five having clinically equivalent to moderate COVID-19 and five having severe infection (Extended Data Fig. 1). For immune cell types, we recorded changes in compartment size (cell numbers) and subset composition (cell percentages), which may change even when compartment size does not. All phenotypes apparently distinguishing COVID-19 and HCs were evaluated for possible impacts of age and/or sex and the significance of some were discounted as a consequence (Supplementary Table 1; see Methods).

Because the timing, anatomical route, dose and multiplicity of infection were indeterminate, two reference time points were employed according to their appropriateness: (1) declared symptom onset, which bears an approximate relationship to time of infection; and (2) day 1 bleed, as a point of certainty within 24 h of patient recruitment as SARS-CoV-2(+), albeit being from 1–15 d following declared symptom onset. In eight cases, day 1 bleed occurred on ward admission following improvement after initial intensive care. Following day 1 bleeds, samples were taken from approximately two-thirds of patients on day 3; from 15 still hospitalized on day 9; and from rare cases at occasional time points thereafter. HC bleeds were drawn concurrently and flow cytometry was performed on fresh cells.

To test the hypothesis that patients with COVID-19 would display an over-arching immune signature, despite their heterogeneity, we undertook principal-component analysis (PCA) of 176 phenotypes for all patients and controls for whom full datasets were available, as is required for PCA. In seeking a signature of active COVID-19 disease we began by considering only flow-cytometric parameters. Strikingly, these alone segregated patients with COVID-19 from HCs, whereas seropositive HCs were indistinguishable from seronegative HCs (Fig. 1a), making them appropriate to combine as a control cohort in most assays that followed.

Fig. 1: Immunophenotyping reveals distinct features of the immune system in patients with COVID-19.
figure 1

a, PCA of 176 cell type frequencies analyzed in COVID-19 and control donors. PC1 explains 17.3% of the variation, while PC2 explains 13.7% of the variation; color denotes disease status (sero control, n = 21, n′ = 21; sero+ control, n = 20, n′ = 17; COVID-19, n = 54, n′ = 37). b, Volcano plot of 366 nonredundant immune parameters analyzed in COVID-19 relative to control samples. Parameters statistically significantly affected in patients with COVID-19 (P < 0.01, difference of means at least 3 × s.d. relative to the control group) are shown in red (control, n = 78, n′ = 55; COVID-19, n = 114, n′ = 63). Pooled sero+ and sero controls compared to COVID-19 for all parameters except when sero+ and sero were statistically significantly different from each other (see Methods). c, Correlations between immune parameters in control (y axis) and patients with COVID-19 (x axis). Color corresponds to the number of overlapping points. Note that all points outside the dashed vertical and horizontal lines at r = −0.33 and r = 0.33 denote significant correlations (P < 0.01). d, Correlations between a subset of parameters with statistically significantly different correlation coefficients in COVID-19 relative to control (as highlighted in c). Plots display results from Spearman correlation tests and a linear regression line with 95% confidence interval shading. Control, n = 78, n′ = 55; COVID-19, n = 114, n′ = 63 (bd). e, Spearman correlation network in COVID-19 samples (n = 114, n′ = 63). Nodes are manually clustered into functional groups of immune parameters. Node size corresponds to the degree of relatedness. Edge color denotes direction of correlation: orange, positive correlation; navy, negative correlation. Only Spearman correlations in which P < 0.01, r > 0.3 or <−0.3 and a delta r of greater than ± 0.3 relative to control (n = 78, n = 55) values are represented. n, samples; n′, individuals; n/n′ may vary slightly between graphs due to data filtering or experimental dropouts (see Methods).

Source data

To identify the major statistically significant discriminators between COVID-19 and HCs, we analyzed effect sizes for all data, now including antibodies and analytes (Fig. 1b). This depicted a collective COVID-19 immunophenotype prominently enriched in SARS-CoV-2-specific antibodies, plasmablasts, cytokines IL-8, IL-6 and IL-10, chemokine IP-10, highly cycling T cells and CD8+ T cells coexpressing exhaustion-associated markers, PD-1 and TIM3. Concomitantly, αβ and γδ T cells were depleted and the composition of the B cell compartment was altered (Fig. 1b and Supplementary Table 2).

In addition to single parameters, immune systems in different settings are distinguishable by structures imposed by significant positive and negative correlations between non-co-dependent parameters24. Whereas many immune correlations were conserved in COVID-19 versus HCs (rCOVID = rcontrol) (Fig. 1c), others were highly exaggerated; others were new (for example, IP-10 versus CXCR3+CCR6negCD8+ cells); and others were inverted (for example, CD4+ T effector memory (TEM) cells versus PD-1+Vδ1+ T cells; basophils versus CD4+ T cells) (Fig. 1c,d). Correlations for which rCOVID differed from rcontrol by greater than or equal to ±0.3 composed an aggregate immune structure that emphasizes the inter-connectedness of many traits composing the COVID-19 immunophenotype (Fig. 1e). We then investigated which of those traits reflected consensus COVID-19 responses, which were severity-related and which if any might anticipate disease progression.

Consensus COVID-19 B cell responses

First we clarified the potential impact of virus burden as assessed by clinical nasopharyngeal swab test. Predictably, viral PCR with reverse transcription (RT–PCR) measures, controlled for sample adequacy, trended inversely with time from symptom onset, but they did not correlate with peak severity or admission severity (Extended Data Fig. 2a,b), in line with some other studies25,26. Nor did they correlate with most immune parameters, two exceptions being CD45RA+CD27negVδ1+ T cells and natural killer (NK) cells, both of which cell types have been shown to respond directly to virus-infected cells27,28 (Extended Data Fig. 2b).

Seeking severity correlates within the COVID-19 signature, we first examined antibodies detecting recombinant viral antigens (measured by ELISA) and/or native antigens (measured by luciferase-based immunoprecipitation (LIPS) assay). The assays showed good correspondence (Extended Data Fig. 2c), each detecting individuals with high-titer antibodies to RBD, S and N in all severity classes and in recovered seropositive HC (Fig. 2a and Extended Data Fig. 2d; note, although some seropositive HCs scored negative in the assays shown for S-reactive or RBD-reactive IgG/IgM, they will have been defined by scoring positive in at least one other assay). Furthermore, the antibodies’ host-protective potentials were inferred from strong correlations of ELISA-detected RBD-specific IgG with virus-entry neutralization assayed using a subset of samples29 (K.D., unpublished data).

Fig. 2: Coexistent anti-SARS-CoV-2 antibody responses and disruption to the B cell compartment in patients with COVID-19.
figure 2

a, Peak antibody titers against SARS-CoV-2 SPIKE and RBD antigens (sero control, n′ = 32; sero+ control, n′ = 23; low, n′ = 6; moderate, n′ = 26; severe, n′ = 31). Analysis performed by Kruskal–Wallis test with Dunn’s post hoc test. Horizontal line denotes the threshold for positivity. b, B cell numbers. c, CD19 mean fluorescence intensity (MFI) within B cells. d, Frequency of CD5+ B cells. e, Plasmablast numbers. Sero control, n = 33, n′ = 26; sero+ control, n = 34, n′ = 22; low, n = 10 n′ = 6; moderate, n = 34, n′ = 24; severe, n = 53, n′ = 28; LRTI, n = 16, n′ = 8 (be). f, Plasmablast frequencies and antibody titers over time in control and individuals with COVID-19 with multiple sampling dates (in samples <50 d after symptoms/first bleed). Repeat samples from the same individual are linked. COVID-19 samples are colored by CRP measurement, if clinical test performed within 48 h of sampling, otherwise they are white. Sero control, n = 17, n = 8; sero+ control, n = 24, n′ = 10; COVID-19, n = 71, n′ = 31. Box plots denote median and 25th to 75th percentiles (boxes) and 10th to 90th percentiles (whiskers) and were statistically evaluated by a linear mixed model grouped by severity (be), with patient as a random variable, corrected for age- and sex-dependency. LRTIs were compared with pooled COVID-19 samples. Sero and sero+ controls were pooled for comparison with COVID-19 samples for most parameters. Where sero and sero+ were statistically significantly different from each other, sero controls were instead compared with COVID-19. All tests were two sided and without multiple testing adjustment.

Source data

As illustrated by RBD-specific and N-specific IgGs, antibody titers against different specificities generally correlated well (Extended Data Fig. 2e). Additionally, IgM titers correlated with IgG titers for single specificities (Extended Data Fig. 2f). There was a clear, albeit not significant, trend toward higher titers in severe patients (Fig. 2a), but this might reflect increased time windows for antibody maturation, as indicated by the color coding; indeed, three patients lacking anti-SARS-CoV-2 IgG died within 12 d of symptom onset (Fig. 2a).

Although lymphopenia has been widely described for COVID-19, patients displayed considerable inter-individual variation in B cell numbers, ranging from overt cytopenia (<104 B cells ml−1) to atypically high counts (2–3 × 105 ml−1) (Fig. 2b). Notably, CD19 expression levels were substantially lower on B cells from most patients with COVID-19 and from some non-COVID-19 LRTI samples (Fig. 2c; note that because seropositive and seronegative HCs were comparably distributed, significance values were calculated against aggregate HCs, as indicated). Additionally, B cell composition was altered with most patients with COVID-19 and LRTIs displaying highly significant reductions in the frequencies of CD5+ B cells, which often display antigen poly-reactivity and ordinarily account for ~25% of B cells (Fig. 2d; note, given that non-COVID-19 LRTIs consisted of moderate and severe patients, significance values for LRTIs were calculated versus moderate and severe COVID-19). Additionally, ‘natural effector’ B cells, were overtly reduced in numbers and frequency (Extended Data Fig. 2g,h).

Given the near-universal production of antibodies, patients with COVID-19 unsurprisingly showed increased CD38+CD27+ plasmablast frequencies, particularly in the severe subcohort (Fig. 2e and Extended Data Fig. 2g,h). High inter-individual variation seemingly reflected the transience of human plasmablast expansion in COVID-19 samples tracked from symptom onset (Fig. 2f). These kinetics together with accumulating antibody levels (Fig. 2f) bore no relationship to concomitant C-reactive protein (CRP) measurements, but they conspicuously resembled counterpart measurements made 2–3 weeks post-vaccination19, in contrast to the overt stability of plasmablasts and antibodies tracked in a subset of HCs (Fig. 2e).

In summary, a dynamic B cell biology contributed to the immunophenotype of essentially all patients with COVID-19. Within this, only elevated plasmablast numbers were evidently severity-related. Indeed, whereas anti-thyroglobulin and thyroid peroxidase antibodies correlated with adverse clinical events following swine flu vaccination19, they too showed no severity correlation when assessed in a subcohort (Extended Data Fig. 2i).

Severity-related innate immune traits

We next investigated the cytokines commonly associated with infections. Elevated IL-8 (CXCL8) levels were a consensus trait of COVID-19 and of most LRTI samples (Fig. 3a). Albeit not in all patients, IL-6 and IL-10 levels were also highly elevated in COVID-19 and LRTIs, but unlike IL-8, the increases were severity-related (Fig. 3b,c). IP-10, a chemokine rapidly and transiently induced following vaccination and other virus infections19,30,31,32, was almost invariably increased in COVID-19 and was severity-related, but was conspicuously not increased in LRTIs (Fig. 3d). Thus, many patients with COVID-19, but not non-COVID-19 LRTIs, were described by a severity-related triad of IP-10, IL-6 and IL-10 (Fig. 3e).

Fig. 3: Patients with COVID-19 display dysregulated cytokine responses.
figure 3

ad, Plasma concentrations of IL-8 (a), IL-6 (b), IL-10 (c) and IP-10 (d). e, Three-way correlation between plasma concentrations of IP-10, IL-6 and IL-10 in COVID-19 and LRTI samples. Plane depicts three-dimensional linear regression of all three parameters and coloring depicts IP-10 concentration. Individual Spearman correlation results are reported. f, IP-10, IFN-γ and IFN-α concentration over time in individuals with COVID-19. Repeat samples from the same individual are linked and are colored by CRP measurement, if clinical test performed within 48 h of sampling, otherwise they are white. Box plots of controls and LRTIs shown for range reference. All cytokines: sero control, n = 41, n′ = 32; sero+ control, n = 37, n′ = 23; low, n = 10, n′ = 6; moderate, n = 41, n′ = 26; severe, n = 60, n′ = 31; LRTI, n = 18, n′ = 10 (ae). f, IP-10, IFN-γ, sero control n = 41, n′ = 32; sero+ control n = 37 n′ = 23; COVID n = 82 n′ = 32; IFN-α, COVID n = 31, n′ = 12. Box plots denote median and 25th to 75th percentiles (boxes) and 10th to 90th percentiles (whiskers) and were statistically evaluated by a linear mixed model grouped by severity, with patient as random variable, corrected for age- and sex-dependency. LRTIs were compared with pooled COVID-19 samples. Sero and Sero+ controls were pooled for comparison with COVID-19 samples for most parameters. Where Sero and Sero+ were statistically significantly different from each other, Sero controls were instead compared with COVID-19. All tests were two sided and without multiple testing adjustment.

Source data

Usually, IP-10 is transiently induced by type-I (α/β) or type-II (γ) interferon (IFN)19. Indeed, IFN-γ levels were correlated with IP-10, but in several patients were not elevated and increases were only significant in the severe group, partly because HC IFN-γ coefficients of variation were themselves high (Extended Data Fig. 3a,b and Fig. 3f). Likewise, whereas type-I IFN responses were reportedly low in COVID-1913, a highly sensitive assay detected markedly increased IFN-α in some patients, although it was strikingly transient, particularly in relation to largely sustained IP-10 levels (Extended Data Fig. 3a and Fig. 3f). As was true for plasmablast dynamics (above), neither IP-10 nor IFN dynamics bore any obvious relationship to concomitant CRP measurements (Fig. 3f). Changes in other chemokines and cytokines, including IL-1β and tumor necrosis factor were only occasionally observed (Extended Data Fig. 3c), excluding broad-based cytokine storms from a consensus COVID-19 signature.

We hypothesized that selective cytokine overexpression might be reflected in the cellular composition of a COVID-19 immune signature. For example, IL-8 activates neutrophils and as was reported6, their counts were higher in COVID-19, albeit moderately and in LRTI (Fig. 4a). Conversely, eosinophil counts were comparably variable across HCs, COVID-19 and LRTIs (Fig. 4b). Most striking in severe COVID-19 were dramatic depletions of plasmacytoid dendritic cells (pDCs), prime sources of type-I IFN and of basophils (Fig. 4c,d). Neither was evident in LRTI (Fig. 4c,d) and nor was basophil depletion attributable to potentially confounding factors such as asthma, allergy or antihistamines. Blood DC composition also changed markedly across COVID-19 in relation to HCs and LRTIs, with the CD11c+CD1cneg subset becoming more predominant, accompanied by rarely reported cell cycling33 (Fig. 4e,f and Extended Data Fig. 4a).

Fig. 4: Patients with COVID-19 display a disrupted monocyte and dendritic cell phenotype.
figure 4

ad, Quantification of neutrophils (a), eosinophils (b), pDCs (c) and basophils (d) from whole blood. e,f, Frequencies of CD1cneg (e) and CD1cpos mDCs (f). gi, MFI of CD86 (g) and HLA-DR (h) in intermediate monocytes and HLA-DR in classical monocytes (i). Sero control, n = 29, n′ = 27; sero+ control, n = 25, n′ = 19; low, n = 9, n′ = 5; moderate, n = 24, n′ = 15; severe, n = 18, n′ = 9; LRTI, n = 18, n′ = 10. jl, Correlation between plasma IL-6 concentration and HLA-DR MFI in classical monocytes (j), plasma IP-10 concentration and total basophil counts (k) and plasma IP-10 concentration and CD1cneg myeloid DC (mDC) frequency (l). COVID-19, n = 51, n′ = 29. Correlation plots display results from Spearman correlation tests and a linear regression line with 95% confidence interval shading. Box plots denote median and 25th to 75th percentiles (boxes) and 10th to 90th percentiles (whiskers) and were statistically evaluated by a linear mixed model grouped by severity, with patient as random variable, corrected for age- and sex-dependency. LRTIs were compared with pooled COVID-19 samples. Sero and sero+ controls were pooled for comparison with COVID-19 samples for most parameters. Where sero and sero+ were statistically significantly different from each other, sero controls were instead compared with COVID-19. All tests were two sided and without multiple testing adjustment.

Source data

As reported for other healthy cohorts34, HC monocyte numbers displayed a high coefficient of variation, but nonetheless their counts trended downward in COVID-19 versus upward in LRTI (Extended Data Fig. 4b), and their composition changed, with depletions of classical monocytes versus increased frequencies of CD16+CD14+ intermediate monocytes (Extended Data Fig. 4b,c), as reported for virus infections35. However, the most overt monocyte phenotype was near-universal diminution of CD86 and HLA-DR expression, particularly in intermediate monocytes, which was likewise observed in most LRTIs (Fig. 4g–i) and in sepsis20,36.

As expected, many COVID-19 innate immune cell phenotypes were related. Thus, IL-6 upregulation and HLA-DR downregulation were correlated (Fig. 4j), as were sustained IP-10 and basophil loss and increased CD11c+CD1cneg DC frequencies (Fig. 4k,l). In summary, COVID-19 patients displayed innate immune traits, of which some were shared with LRTIs, but among which some correlated strikingly with severity and with each other.

Coexisting suppression and activation of T lymphocytes

We next sought to characterize the COVID-19 T cell signature, given that several T cell phenotypes were among the most significant discriminators of COVID-19 from HCs (see Fig. 1b and Supplementary Table 2). Indeed, whereas pan-lymphopenia has been frequently cited in severe COVID-19 (refs. 6,37,38), we observed primarily T cytopenia, notwithstanding some samples being within the normal range (Fig. 5a). T lymphopenia, which was reported in other severe infections39 and sepsis20, most overtly affected CD8+ cells and γδ cells (Fig. 5a). Indeed, severe losses of Vγ9Vδ2 cells, which ordinarily dominate blood γδ cells, substantially shifted compartment composition toward Vδ1+ cells, which can respond to virus infection27 (Extended Data Fig. 5a). Significant NK cell cytopenia was also evident in COVID-19 and LRTI, but was less pronounced than T cytopenia (Extended Data Fig. 5b) and gross NK cell subset composition, demarcated by CD56 and CD16, was mostly unaltered (Extended Data Fig. 5b).

Fig. 5: Patients with COVID-19 display selective cytopenia in particular T cell subsets and coexistent active cycling and exhausted phenotype of residual T cells.
figure 5

a,b, Cytopenia effect size and quantification of total T cells, CD4+, CD8+ and γδ T cells (a) and of naive and memory CD8+ subsets (b) (sero control, n = 39, n′ = 32; sero+ control n = 36, n′ = 23; low n = 10, n′ = 6; moderate n = 40, n′ = 24; severe n = 58, n′ = 31; LRTI n = 17, n′ = 9). NKT, NK T cells. NS, not significant. c, Hierarchical clustering of activation- and effector-associated gene expression in CD4+ and CD8+ effector memory cells from NanoString analysis. Control n′ = 3, COVID-19 n′ = 3 (CD4+) or 2 (CD8+). d, Representative flow cytometry and quantification of cell cycle status of CD4+ and CD8+ TEM cells and γδ T cells (sero n = 36 n′ = 31; sero+ n = 35 n′ = 23; low n = 9, n′ = 5; moderate, n = 33, n′ = 21; severe, n = 52, n′ = 29; LRTI, n = 17, n′ = 10). e, Correlation between PD-1+TIM3+ CD8+ TEM cells and CD8+ TEM cell G1 frequencies in moderate and severe patients with COVID-19 (moderate n = 25, n′ = 15; severe, n = 33, n′ = 18): plots display results from Spearman correlation tests and a linear regression line with 95% confidence interval shading. f, Representative flow cytometry and quantification of PD-1+TIM3+ CD8+ TEM cell frequency (sero control, n = 35, n′ = 32; sero+ control, n = 34, n′ = 23; low, n = 10, n′ = 6; moderate, n = 37, n′ = 24; severe, n = 48, n′ = 25; LRTI, n = 18, n′ = 10). Box plots denote median and 25th to 75th percentiles (boxes) and 10th to 90th percentiles (whiskers) and were statistically evaluated by a linear mixed model grouped by severity (excluding low), with patient as random variable, corrected for age- and sex-dependency. LRTIs were compared with pooled COVID-19 samples. Sero and sero+ controls were pooled for comparison with COVID-19 samples for most parameters. Where sero and sero+ were statistically significantly different from each other, sero controls were instead compared with COVID-19. All tests were two sided and without multiple testing adjustment. Effect size is expressed as the difference in mean values divided by the control mean value.

Source data

The αβ T cytopenia was strikingly subset selective. Thus, classifying TH subsets by chemokine receptor expression revealed disproportionate depletions of CD4+ TH17.1 and CD4+ TH1 cells, which both produce IFN-γ, versus relative preservation of TH2 cells and only modest Treg cell losses (Extended Data Fig. 5c). A similar hierarchy was evident in LRTI (Extended Data Fig. 5c), in which T cytopenia was less overt, and in a small set of sepsis samples available for analysis, the relative frequencies (as opposed to counts) of TH2 cells greatly increased relative to TH1 cells, with Treg cells again less affected (Extended Data Fig. 5d).

For CD4+ cells, cytopenia in severe COVID-19, LRTI and sepsis greatly affected TEM cells (Extended Data Fig. 5d,e). Whereas CD8+ TEM cells were also markedly depleted in severe COVID-19 and LRTI, there were also some losses of terminally differentiated TEMRA cells, but particularly of naive (TN) and central memory (TCM) CD8+ cells, especially in severe COVID-19 (Fig. 5b and Extended Data Fig. 5d,e). Note that depletions of some subsets, for example, TN, may in part reflect their ongoing differentiation into other states. This notwithstanding, the many changes across many T cell subsets (Extended Data Fig. 6a,b) resulted in a signature COVID-19-associated correlation matrix (Extended Data Fig. 6c).

In COVID-19, but not in LRTI, frequencies of activated HLA-DR+CD38+ T cells were increased, particularly among CD8+ cells and especially in severe patients who also showed the greatest CD8+ cell depletions (above) (Extended Data Fig. 7a). Additionally, COVID-19 showed higher frequencies of activated CD25+CD4+ and CD25+CD8+ cells, as did LRTI (Extended Data Fig. 7b). When compared to CD4+ TEM cells of three HCs, CD4+ TEM cell activation in COVID-19 was suggested by increased gene expression of HLADR, CD38 and MKI67 (encodes the cell-cycling marker, Ki67) in CD4+ TEM cells from three patients with COVID-19 who each displayed actively cycling T cells (see below) but who collectively spanned a range of lymphopenia (Fig. 5c; Methods). Similarly, effector activation of CD8+ TEM cells, relative to concurrently sampled HC TEM cells, was strongly suggested by overexpression of cytolysis-associated genes NCR1 (encoding NKp46), FASLG, several GZM genes and PRF1 (encoding perforin) (Fig. 5c). Overexpression of TNFRSF10B (encoding TRAIL-receptor) and CASP3 (encoding Caspase 3) additionally suggested that COVID-19 CD8+ TEM cells were more prone to apoptosis than HC TEM cells (Fig. 5c). Furthermore, COVID-19 CD4+ TEM cells showed striking downregulation of genes encoding T cell receptor signaling components and overexpression of HIF1-α (Extended Data Fig. 7c), suggesting an adaptation to hypoxia or dysoxia. In summary, molecular and flow-cytometric criteria jointly revealed greatly altered biologies of CD4+ and CD8+ TEM cells in COVID-19 relative to HCs. Nonetheless, COVID-19 TEM cells could not be segregated from HC TEM cells by expression of chemokine receptors, which might have indicated a propensity to home to the lungs or other tissues (Extended Data Fig. 7d), in the same way that blood-borne skin-homing T cells express cutaneous lymphocyte-associated antigen40.

Ordinarily ≥97% of blood T cells are G0 cells in transit. Thus, the most overt severity-related read-outs of CD4+ and CD8+ T cell activation in COVID-19 relative to HC or LRTI subjects were ~tenfold increases in T cells in G1 or S-G2/M. These changes, illustrated for TEM cells (Fig. 5d) applied to all CD4+ and CD8+ differentiation states, albeit much less so for TN cells, of which >96% remained in G0 (Extended Data Fig. 7e). Likewise, frequencies of γδ cells (mostly Vδ1+) in G1 increased >tenfold, although few transitioned into S-G2/M (Fig. 5d). The frequencies of cycling CD8+ cells mostly correlated with frequencies of CD8+ cells expressing activation markers (Extended Data Fig. 6c, arrowed), and given that T cell activation and cycling are primarily induced in tissues and draining lymph nodes, cell cycling measurements offer a systemic portal onto local immune dynamics21,22.

Conspicuously, G1 CD8+ cell frequencies also strongly correlated with CD8+ cells coexpressing two terminally differentiated/exhaustion markers, PD-1 and TIM3 (Fig. 5e and Extended Data Fig. 6c, arrowed). Such cells are usually barely detectable, but they emerged in COVID-19 and to a lesser extent in LRTI (Fig. 5e,f). Additional evidence of chronic activation of COVID-19 versus HC CD8+ TEM cells was provided by increased expression of genes encoding PD-1 (PDCD1), LAG3 and CTLA4 (Extended Data Fig. 7f), albeit there was also downregulation of the gene encoding TIGIT, another inhibitory receptor, whose ligand PVR may directly interact with SARS-CoV-2 proteins41.

In summary, T cells in most patients with COVID-19 displayed coexisting alterations in numbers, subset composition, cycling, activation and gene expression. Whereas some such changes occurred in LRTI and sepsis, and were observed in other infections, others did not, with some traits, particularly cell-cycling, being strongly related to severity. Moreover, the scale of the T cell phenotypes, for example, ~20% of CD8+ TEM cells in G1 concurrent with an 80% depletion of CD8+ TEM cells, suggests impacts well beyond SARS-CoV-2-specific T cells, for which varying estimates exist3,11,12. Finally, we noted that T cell cycling and increased representation of PD-1+TIM3+CD8+ cells were two parameters that were statistically significantly, albeit slightly, increased in seropositive versus seronegative HCs (Fig. 5d,f), suggesting that they may be legacies of previous, mild SARS-CoV-2 infection.

Anticipating severe disease progression

COVID-19 can be conspicuously phasic, with major deteriorations in some patients occurring ~7 d post-symptom onset. Hence, there is intense interest in early patient stratification of those more likely to recover rapidly, thereby freeing hospital resources for those at higher risk42. Thus, we compared first-bleed immunological traits in patients segregated according to whether their WHO ordinal scores decreased (that is clinical improvement), increased (worsening) or stabilized over the next 7 d, or at discharge if earlier. As an example, IP-10 levels showed slightly greater capacity than CRP levels to segregate patients who would improve versus worsen (Fig. 6a,b). Moreover, even greater prognostic potential was evident for first-bleed IL-6 and IL-10 measurements, re-emphasizing the prominence of the IL-6–IP-10–IL-10 triad within the COVID-19 signature (Fig. 6b). From a practical standpoint, it is noteworthy that anti-SARS-CoV-2 IgG and flow-cytometric T cell quantification also anticipated improvement better than commonly used clinical lymphocyte count, neutrophil count, ferritin, D-dimer and albumin (Fig. 6b).

Fig. 6: COVID-IP parameters associated with patient prognosis.
figure 6

Selected immunological parameters from the first bleed of patients with COVID-19 admitted as moderate or severe were compared with parameters from contemporaneous clinical blood tests (within 72 h) for association with prognosis. a, First-bleed CRP and plasma IP-10 concentration in patients for whom WHO ordinal score decreased (improve), increased (worsen) or remained constant (stable) within the following 7 d or at discharge if earlier (improved). Two-sided Kruskal–Wallis test with Dunn’s post hoc test. b, Log2 fold change (FC) of clinical and immunological parameters between worsening and improving patients as in a. Two-sided Wilcoxon test. ALT, alanine transaminase; FBC, clinical full blood count; FC, flow cytometry. Improve n′ = 40, stable n′ = 8, worsen n′ = 9 (a,b). c,d, Correlation between length of hospital stay following first bleed and peak WHO ordinal score (c) and first-bleed CRP and IP-10 plasma concentration (d). Patients p049 and p053 were still in hospital at conclusion of study and were assigned an arbitrary discharge date. Plots display results from Spearman correlation tests and a linear regression line with 95% confidence interval shading. e, Summary of correlations between clinical and immunological parameters and length of hospital stay as in c and d. n′ = 57 (ce). f,g, Four patients in our cohort were treated with intravenous methylprednisolone for COVID-19-hyperinflammation; they are highlighted on the graph as HI1-4. Laboratory parameters for ferritin (normal range 30–400 μg L−1), procalcitonin (0.02–0.05 μg L−1), CRP (0–4 mg L−1) and D-dimer (0–0.55 mg L−1) for HI patients versus patients with COVID-19 in the ICU who did not have features of HI (n′ = 11, black lines) (f). Ferritin and plasma IP-10 concentration in HI (n′ = 4) and non-HI patients (n′ = 11, gray) (g). Threshold (dotted gray line) demarcates upper limit of the normal range for ferritin as measured in our clinical laboratory (400 μg L−1) or directly below the lowest value detected for a COVID-19-HI patient for IP-10. Overall, 5 of 11 non-HI patients with COVID-19 fell above the ferritin threshold (labeled). Correlation plots display results from Spearman correlation tests and a linear regression line with 95% confidence interval shading. Box plots denote median and 25th to 75th percentiles (boxes) and 10th to 90th percentiles (whiskers).

Source data

Length of hospital stay is another metric for which early risk-based stratification would be beneficial. It correlated somewhat with peak ordinal severity, as expected, but not well with contemporaneous CRP (Fig. 6c,d). Conversely, it was best anticipated by first-bleed IP-10 and IL-10 and by the reciprocal of T cell counts, together with albumin, although less so by IL-6 (Fig. 6d,e).

Finally, a small subcohort is demarcated in Fig. 6 as hyperinflammatory (HI), evidenced by persistent fevers and hypoxemia, hyperferritinemia and high CRP, together with negative bacterial cultures at diagnosis. Such patients can experience long intensive care unit (ICU) stays, making it important to distinguish them as early as possible from others requiring ICU. Elevated CRP and D-dimers showed little discrimination of HI and other ICU patients; elevated procalcitonin was associated with three of the four HI patients; and highly elevated ferritin was mostly but not exclusively associated with HI (Fig. 6f). By contrast, highly elevated IP-10 (≥350 pg ml−1) almost completely segregated all four HI patients from the others at each time point examined (Fig. 6g; note: patient HI1 is missing from Fig. 6b,d because their day 1 bleed did not include IP-10 quantitation). In summary, specific components of the COVID-19 immune signature showed individual and collective potentials to improve risk-based patient stratification.

Discussion

Our study of hospital-treated patients with COVID-19, including longitudinal analyses, has shown that an ongoing, dynamic, immune response clearly segregates patients from controls, including recently recovered seropositive individuals. The response includes several discrete innate and adaptive traits, including changes in DC and B cell composition and profoundly altered T cell phenotypes that might undermine immunoprotective T cell immunity3. Moreover, striking correlations emerged such as that connecting basophil depletion and elevated IP-10.

Some such traits occur in other settings, including but not limited to immunoprotective vaccination, non-COVID-19 LRTI and sepsis19,20,32. This notwithstanding, those and other traits combine to compose a core COVID-19 immune signature, just as clinical features recognizable in other scenarios combine to compose the core symptoms of COVID-19. The immune signature is provided as a large dataset supported by an online portal, www.immunophenotype.org, thereby facilitating independent validation in various settings. Nonetheless, the signature’s broad applicability is suggested by the description of many of its traits in other COVID-19 cohorts37,43,44,45,46,47,48,49, although some traits—for example, neutrophilia, eosinophil depletions and NK cell composition—may prove to be more variable14,46.

The immune signature offers many potential opportunities. By identifying similarities with other settings, traits such as IL-8 upregulation, CD5+ B cell depletion, plasmablast activation, reduced HLA-DR expression on monocytes and subset-selective T cytopenia20,32,36,39, may expose shared pathogenesis and offer precedents for disease management, as illustrated by IL-7 provision to boost T cell competence in sepsis and now in COVID-19 (https://clinicaltrials.gov/ct2/show/NCT04379076). Other traits, such as pDC, Vδ2+ T cell and basophil depletions that are severity-related and not obviously shared with non-COVID-19 LRTI and/or other diseases, may provide means to track and to better understand unique components of COVID-19 pathobiology. Indeed, some COVID-19 traits, for example, cycling CD1cneg DCs and the scale of peripheral blood CD8+ T cell cycling, are without common precedent.

A third, potentially practical opportunity is offered by the sustained triad of IL-6, IL-10 and IP-10 and some other traits whose early quantification anticipated disease progression and length of hospitalization. Possibly they may be combined into routine clinical tests to aid early risk-based stratification of patients, which can be of major benefit when healthcare provision is over-stretched. Moreover, targeting those inflammatory mediators may be of therapeutic benefit, possibly contributing to the success of dexamethasone treatment18, although beyond this triad, the immune signature lacked a broad-based cytokine storm.

The almost invariable production of SARS-CoV-2 RBD-specific IgG suggests that patients have immunocompetent potential to limit virus infection. Nonetheless, the reportedly high efficiency of SARS-CoV-2 entry into cells may be partially immunoevasive50, increasing the dependence on mechanisms to eradicate virus-infected cells. In that regard, virus burden correlated somewhat with NK and effector Vδ1+ cells, which are both implicated in antiviral responses27,28,51. Moreover, immunoprotective potential was suggested by multifaceted T cell activation, including upregulated HLA-DR and CD38 expression that was reported for SARS-CoV-2-specific T cells12. Ongoing and future studies will characterize virus-specific T cell responses in COVID-19, but their characterization was not required here to establish a profound COVID-19 T cell phenotype whose overt scale (for example, cycling; depletion) almost certainly exceeds frequencies of virus-specific cells3,11,12. Indeed, broad T cell dysregulation might undermine T cell-mediated immunoprotection, particularly as it greatly affected memory T cells upon which adults are probably more dependent than children because of thymic involution. Such cells may include potentially protective, commonly occurring memory T cells seemingly primed against common-cold coronaviruses3. Of note, seropositive individuals who had recently recovered without hospitalization carried lingering traits of T cell activation.

COVID-19 postmortems have revealed myelomonocytic lung infiltrates52 that might partially account for the loss from the blood of pDCs and/or basophils, with the latter implicated in tissue repair53 and in regulating coagulation54,55 that may be germane to frequent thrombotic events in COVID-19. Likewise, T cytopenia might reflect trafficking, for example, to the lungs, although only variable lymphocytic infiltrations were revealed by postmortems52,56 and residual TEM cells did not obviously display relevant chemokine receptor expression.

Instead, T cells may be in a hyperactivated state, prone to apoptosis. Again, some postmortems have revealed splenic white pulp and pulmonary and hilar lymph node necrosis with T cell depletion in a context of B cell activation (A.D., personal communication). Whereas, CD8+ TN and CD4+ TN depletions seem challenging to attribute to activation-induced cell death or homing, those might reflect active differentiation toward TEM/TCM cells; fratricide by CD8+ T cells; or hemophagocytosis. It is also possible that the different metabolic profiles of discrete T cell subsets might render them selectively susceptible to hypoxia, dysoxia or specific inflammatory mediators as yet to be characterized.

Near-universal IP-10 upregulation is noteworthy, given the depletions of pDC, TH1 and TH17.1 cells that might ordinarily be sources of type-I and type-II IFNs that are the main inducers of IP-10. Although IP-10 levels showed some correlations with IFN-γ, and although some patients showed high, albeit transient IFN-α levels, IP-10 might be boosted by other, virus-related mechanisms. Indeed, elevated IP-10 characterized SARS30 and MERS31 and MERS-CoV enters T cells via CD26, an ectopeptidase that regulates IP-10 activity57. Conceivably, chemokine dysregulation is a core component of pathogenic coronavirus infections, possibly interfering with ordered immunocyte chemotaxis and contributing to a milieu in which coronaviruses thrive. Possibly germane to this, three chemokine receptor genes lie within a region associated with severe COVID-19 respiratory disease susceptibility58.

In summary, the multifaceted COVID-19 immune signature provides a basis for addressing many clinical and research questions, including whether or not specific immunological traits, considered individually or collectively, are causes or consequences of disease progression. Indeed, our understanding of COVID-19 and of the immune response to SARS-CoV-2 is a rapidly evolving one, reflecting the pace, intensity and global scope of research. That scope provides an appropriate context for swift and independent assessments of the COVID-19 immune signature in myriad settings, enhancing the prospects of consensus understanding and collective actions to better manage disease and thereby to diminish the life-threatening potential of SARS-CoV-2 infection.

Methods

Study design and recruitment

Between 25 March 2020 and 14 May 2020, 63 patients (median age 61 years; interquartile range (IQR) = 47–70) with confirmed SARS-CoV-2 infection by viral PCR (n′ = 62) or serology alone (n′ = 1) were recruited to the COVID-IP study from hospitals within Guy’s and St Thomas’ NHS Trust for an observational cohort study with serial peripheral blood immunophenotyping and analysis of clinical outcomes (Extended Data Fig. 1). Blood sampling was performed within 24 h of recruitment and thereafter, approximately at day 3 and day 9 after recruitment, with two patients having additional sampling time points.

Patients above 18 years of age were approached for informed consent to serial blood sampling by the research nursing team if they met the criteria of a positive PCR result for SARS-CoV-2. An additional ten patients who were confirmed to not have COVID-19, but were admitted with LRTI, were recruited as a control group (community-acquired pneumonia (n′ = 5), bacterial empyema (n′ = 1), infective exacerbation of chronic obstructive pulmonary disease (n′ = 1), viral LRTI (n′ = 3)). For patients in the ICU lacking capacity, this was sought from their next of kin or treating physician under appropriate ethical approval. Informed consent was obtained retrospectively from these patients, where possible.

Thus, the cohort includes patients admitted to the wards and ICU with SARS-CoV-2 infection (n′ = 56), patients admitted for other conditions who acquired SARS-CoV-2 during their stay (n′ = 3) and ambulatory dialysis patients (n′ = 4) diagnosed on screening tests. The majority of patients (n′ = 55) were recruited within 72 h of a positive PCR result or initial clinical suspicion. Due to the prospective nature of our sampling, we were able to capture a heterogenous population of ICU patients recruited before admission (n′ = 3), during admission (n′ = 4) and immediately after ICU discharge (n′ = 8). Patients recruited post-ICU were sampled upon first encounter within a general medical ward.

During this same period, 55 healthy adult volunteers (median age 36 years; IQR = 29–45.5) with no known current malignancy, serious infectious illness, organ transplant or autoimmune disease were recruited as a control cohort for similar serial peripheral blood immunophenotyping. Given the nationwide lockdown during the period of the study recruitment, healthy adult volunteers were largely research or clinical staff employed at Guy’s and St Thomas’ NHS Trust or King’s College London. Potential donors were approached by the clinical research staff within the COVID-IP team for informed consent on a voluntary basis. A number of the HC volunteers (n′ = 23) had experienced previous SARS-CoV-2 infection for which hospitalization had not been required and constitute a cohort of fully recovered, previously mildly infected individuals.

The study protocol for patient recruitment and sampling, out of the intensive care setting, was approved by the committee of the Infectious Diseases Biobank of King’s College London with reference number COV-250320. The protocol for healthy volunteer recruitment and sampling was similarly approved by the same committee as an amendment to an existing approval for healthy volunteer recruitment with reference number MJ1-031218b. Both approvals were granted under the terms of the Infectious Disease Biobank’s ethics permission (reference 19/SC/0232) granted by the South Central Hampshire B Research Ethics Committee in 2019. Patient recruitment from the ICU was undertaken through the ethics for the IMMERSE study approved by the South Central Berkshire Ethics Committee with reference number 19/SC/0187. Patient and control samples and data were anonymized at the point of sample collection by research nursing staff or clinicians involved in the COVID-IP project. We complied with all relevant ethical regulations.

Sample processing and cell isolation

Unfixed patient samples were handled under Biosafety Level 3 containment conditions following risk assessments and code of practice approved by King’s College London. Blood samples in serum separator tubes were centrifuged at 1,500g for 10 min and serum aliquoted and stored at −80 °C. Aliquots of blood from heparin tubes were stained for whole-blood flow cytometry panels (see below) or centrifuged at 2,000g for 10 min and plasma stored at −80 °C. Remaining heparinized blood was diluted with 50% volume PBS, layered over Ficoll (GE Healthcare) in Leucosep tubes (Greiner Bio-One) and centrifuged at 800g for 15 min without brake at room temperature. The peripheral blood mononuclear cell (PBMC) fraction was then washed three times in cold PBS and used for flow cytometry. All flow samples were fixed for 10 min with either Cellfix (BD) or FoxP3 Fix/Perm kit (eBioscience) before removal from the Biosafety Level 3 facility.

Flow cytometry staining and acquisition

All flow cytometry antibodies and concentrations used for analysis can be found in Supplementary Table 3. PBMC samples were stained for viability with BD Horizon Fixable Viability Stain 780 for 10 min at room temperature. PBMC and whole blood cell surface staining was performed in BD Pharmingen Stain buffer (BSA) and BD Horizon Brilliant Stain buffer Plus for 20 min at room temperature. Intracellular staining was performed after permeabilizing cells with Invitrogen permeabilization buffer 10× for 30 min at 4 °C. PBMC samples were stained using staining mix panels 1, 2, 3, 4 and 5 as shown in Supplementary Table 3. PBMCs were spun down and resuspended in 100 μl of staining mix. Cells were then washed in staining buffer and fixed for 10 min and protected from light. Cells were then washed and resuspended in 200 μl of staining buffer for acquisition by flow cytometry for panels 1, 2, 3 and 4. PBMCs stained for panel 5 were resuspended in permeabilization buffer containing intracellular staining antibodies as shown in Supplementary Table 3 and incubated at 4 °C for 30 min and protected from light. Cells were then spun down and washed in Dulbecco’s phosphate-buffered saline (DPBS) before being resuspended in DPBS containing Hoechst 33342 (Thermo Fisher Scientific) and incubated at room temperature for 15 min. Cells were washed, pelleted and resuspended in 200 μl of DPBS for acquisition by flow cytometry. Whole-blood samples were stained using staining mix panels 6, 7 and 8 as shown in Supplementary Table 4. A total 50 μl of whole blood was stained in 50 μl of antibody staining mix, washed in DPBS and then fixed. Red blood cell lysis was then performed using eBioscience RBC lysis buffer (multi-species) 10× diluted in deionized water for 15 min at room temperature. This was repeated up to two times to ensure adequate removal of red blood cells. Samples were then spun down and resuspended in 200 μl of staining buffer for acquisition by flow cytometry. For PBMC panels 1, 2, 3 and 4, 100 μl of sample was analyzed on a five-laser BD LSR Fortessa acquired with a BD high-throughput sampler. For whole-blood panels 6, 7 and 8, 100 μl of sample was analyzed on a four-laser BD LSR Fortessa acquired with a BD high-throughput sampler. For PBMCs stained using panel 5, cells were acquired on a four-laser BD LSR Fortessa in FACS tubes, run on low for 10 min, with samples diluted to achieve an event rate of no more than approximately 200 events s−1.

Cell sorting by flow cytometry

Selected patient PBMCs (see below) were defrosted and stained as described above. Briefly, PBMCs were stained with BD Horizon Fixable Viability Stain 780 and the following antibodies from the panels shown in Supplementary Table 4; CD3-FITC (P5), TCR γδ-PE-Cy7 (P3), CD4-BV711 (P5), CD8-PerCP-Cy5.5 (P3), CD25-PE (P1: 2A3 & M-A251), CD45RA-BV786 (P3), CCR7-PE-CF594 (P3) and CD127-BV421 (Sort). Live, single CD4+ and CD8+ TEM cells were selected (see Supplementary Fig. 1) and sorted on a BD FACS AriaIII cell sorter equipped with four lasers, a 70-µm nozzle running at 70 p.s.i. and 90 kHz.

Flow cytometry data analysis

FCS files were analyzed using FlowJo (v.10.6.2, Treestar). Gating strategies for all panels are outlined in Supplementary Fig. 1. Event counts for every gate were exported and frequencies to relevant parent populations were calculated in R. Absolute cell counts were back-calculated using the counts per ml of blood for major lineages derived from the whole-blood count panel (panel 6), where the equivalent of 25 μl of whole blood was analyzed per sample. Median fluorescent intensities were calculated using FlowJo for relevant markers on specific populations. For panels 1–4 and 6–7 a minimum threshold of 30 events per parental gate was used to investigate its subpopulations/measure MFI.

Cytokine analysis

The LegendPlex Human Anti-Virus Response Panel (13-plex) (740390, BioLegend) and the LegendPlex Human TH Panel (13-plex) (740721, BioLegend) were used according to manufacturer’s instructions with some modifications. The assay was carried out in V-bottom 96-well plates and plasma was thawed and diluted twofold with assay buffer before being tested. Mixed beads, detection antibodies and streptavidin-PE were diluted twofold in assay buffer and 25 μl of each reagent were used for the following steps. Diluted plasma was added to mixed beads and incubated for 1.5 h, wells were washed twice with wash buffer. Samples and standards were incubated with detection antibodies for 45 min. Streptavidin-PE was then added and the plate was incubated for 20 min. Finally, beads were washed once and resuspended in 200 μl of wash buffer and acquired on a four-laser BD LSR Fortessa X20. All incubation steps were carried out in the dark at room temperature, on an orbital shaker set at 600 r.p.m. Data were analyzed using the LegendPlex data analysis software v.8 for Windows.

Quantification of IFN-α plasma levels using a single molecule array digital ELISA

IFN-α plasma levels were quantified by using a Quanterix Homebrew single molecule array assay on a HD-1 Analyzer instrument. This assay uses two anti-IFN-α autoantibodies isolated and cloned from two APS1/APECED patients59. The 8H1 antibody was used as a capture antibody after coating paramagnetic beads and the 12H5 was biotinylated and used as the detector. The 8H1 and 12H5 antibodies are a property of Servier. Recombinant Human IFN-α17/αI (Hu-IFN-αI; Hu-IFN-α17) (PBL Assay Science) was used as a reference standard.

NanoString gene expression analysis

Approximately 7,500 FACS-sorted, lysed cells were hybridized to a chimeric antigen receptor-T cell panel for profiling 780 human genes at 65 °C overnight (NanoString Technologies). Hybridized samples were processed on an nCounter prep station and data were collected on an nCounter digital analyzer (NanoStringtm), following manufacturer’s instructions. Raw data were imported into nSolver4.0 (NanoStringtm) for data quality checks, background thresholding and normalization. Briefly, imaging performance was above 90% across all samples, with the field of view being 555 for all lanes and the number of fields of view counted observed for all the samples above 500. A binding density in the range of 0.07–0.1 µm−2 was observed across the samples. Background level was determined by mean counts and 2× s.d. of eight negative control probes. Samples that contained fewer than 50% of probes above background or that had imaging or positive control linearity flags were excluded from further analysis. Probes that have raw counts below background in all samples were excluded from differential expression analysis to avoid false-positive results. Data were normalized by geometric mean of housekeeping genes. The heat map was generated using the package ‘ComplexHeatmap’ within R v.3.5.1. Before clustering, data were log10 transformed and mean centered. The heat map data were clustered by selecting ‘clustering_distance_columns = spearman’ and ‘clustering_method_columns = ward.D’ with all the rest of parameters left as default. Samples selected for NanoString analysis are outlined in Supplementary Table 4.

Serology analysis: ELISA

(1) N protein was obtained from L. James and J. Luptak at LMB, Cambridge. The N protein used is a truncated construct of the SARS-CoV-2 N protein consisting of residues 48–365, both ordered domains with the native linker and an N-terminal uncleavable hexahistidine tag. N was expressed in Escherichia coli using autoinducing medium for 7 h at 37 °C and purified using immobilized metal affinity chromatography, size exclusion chromatography and heparin chromatography. (2) The S protein used consists of pre-fusion S ectodomain comprising residues 1–1,138 with proline substitutions at amino acid (aa) positions 986 and 987, a GGGG substitution at the furin cleavage site (aa 682–685) and an N-terminal T4 trimerization domain followed by a Strep-tag II. The plasmid was obtained from P. Brouwer, M. van Gils and R. Sanders at The University of Amsterdam. HEK-293F cells (Invitrogen) grown in suspension at a density of 1.5 million cells ml−1 and transfected with 325 µg of DNA using PEI-Max (1 mg ml−1, Polysciences) at a 1:3 ratio. The supernatant was collected after 7 d and purified using StrepTactinXT Superflow high capacity 50% suspension according to the manufacturer’s protocol by gravity flow (IBA Life Sciences). (3) The RBD plasmid was obtained from F. Krammer at Mount Sinai University. Here the natural N-terminal signal peptide of S is fused to the RBD sequence (aa 319–541) and joined to a C-terminal hexahistidine tag. This protein was expressed in 500 ml of HEK-293F cells (Invitrogen) at a density of 1.5 million cells ml−1. The culture was transfected with 1,000 µg of DNA using PEI-Max (1 mg ml−1, Polysciences) at a 1:3 ratio. Supernatant was collected after 7 d and purified using Ni-NTA agarose beads.

All plasma samples were heat-inactivated at 56 °C for 30 min before use in the in-house ELISA. High-binding ELISA plates (Corning, 3690) were coated with antigen (N, S or RBD) at 3 µg ml−1 (25 µl per well) in PBS, either overnight at 4 °C or 2 h at 37 °C. Wells were washed with PBS-T (PBS with 0.05% Tween-20) and then blocked with 100 µl of 5% milk in PBS-T for 1 h at room temperature. Wells were emptied and serum and plasma were diluted at 1:50 and 1:25 respectively in milk, were added and incubated for 2 h at room temperature. Control reagents included CR3009 (2 µg ml−1), CR3022 (0.2 µg ml−1), negative control plasma (1:25 dilution), positive control plasma (1:50) and blank wells. Wells were washed with PBS-T. Secondary antibody was added and incubated for 1 h at room temperature. IgM was detected using goat-anti-human-IgM-HRP (1:1,000 dilution) (Sigma: A6907) and IgG was detected using goat-anti-human-Fc-AP (1:1,000 dilution) (Jackson: 109-055-043-JIR). Wells were washed with PBS-T and either AP substrate (Sigma) was added and read at 405 nm (AP) or one-step TMB substrate (Thermo Fisher Scientific) was added and quenched with 0.5 M H2S04 before reading at 450 nm (HRP). Data were normalized using a min/max normalization to compare samples across batches. Cutoffs were determined based on data distribution with respect to HCs. Values >0.15 were considered positive.

Serology analysis: LIPS assay

SARS-CoV-2 (NCBI Acc cat. no. NC_045512.2) RBD domain of S (aa 329–538) and N (aa 2–419) gene fragments were cloned into pNanoLuc vector, transfected into HEK293 cells and lysed. The cell lysates containing NanoLuc‐fusion proteins were then probed with plasma samples (0.5–1 × 106 luminescence units (LUs)) for 1 h at room temperature. The Protein G Sepharose beads (25 µl of 4% suspension, Creative BioMart) were used to capture the immune complexes of anti-SARS-CoV-2 antibodies and NanoLuc fusion proteins. After washing, Nano-Glo luciferase substrate (Promega) was added and luminescence was measured in VICTOR X multilabel plate readers (PerkinElmer Life Sciences). LIPS data represent the average of three replicative experiments. Results were given as FC = LU sample/average LUs of HC samples. A fold change >4 was considered as positive.

Criteria for designating a sample as seropositive

A sample was designated as seropositive if IgG/IgM antibody titers against SPIKE or RBD proteins were positive on either the LIPS assay or ELISA by the cutoffs stated above. If samples were positive for antibody titers against N protein alone, they were not considered seropositive on account of some cross-reactivity with seasonal coronaviruses23.

Autoantibody screening

Thyroglobulin and thyroid peroxidase autoantibodies were measured in the serum by the electrochemiluminescence immunoassay on a cobas 6000 analyzer, according to the recommendations of the manufacturer (Roche Diagnostics).

Nucleic acid extraction and PCR for SARS-CoV-2 detection

Nasopharyngeal swabs were collected from patients suspected to have COVID-19. Nucleic acid extraction was carried out using the QIASymphony DSP Virus/Pathogen Mini kit (Qiagen) and an off-board lysis protocol. All swabs received between 12 March and 7 April 2020 were tested using the AusDiagnostics two-step multiplexed-tandem PCR assay (Coronavirus Typing Eight-well Panel; cat. no. 2061901) containing a single SARS-CoV-2 target in the ORF1a gene. All samples received after this date (n′ = 27 of 63) were tested using the AusDiagnostics SARS-CoV-2, Influenza, RSV (eight-well) Panel (cat. no. 80081) containing two SARS-CoV-2 targets in the ORF1a and ORF8 genes and a target within a nontranscribed human gene (NONO; non-POU domain containing, octamer-binding) used as a sample adequacy control. All additional targets in the panels were switched off. AusDiagnostics panels all contain an internal control artificial gene called SPIKE, for which, under correct reaction conditions, the take-off should fall between 10 and 20 cycles and will produce a known concentration of 10,000 copies. The take-off value was recorded for each of ORF1a and ORF8 targets which provided a semiquantitative measure of viral load. No correlation was observed when ORF1a take-off values were plotted against the take-off values for the corresponding sample adequacy controls (unpublished data).

Statistical analysis

Cell subset counts (per ml of blood), LIPS assay values and cytokine concentrations were analyzed after log10 transformation; all other parameters (cell subset frequencies, serology parameters and MFIs) were analyzed without any additional data transformation.

First, we identified parameters that differed between seropositive and seronegative controls by Student’s t-test (P < 0.05). For these parameters, only seronegative controls were taken into account in the downstream testing against COVID-19 samples.

In longitudinal samples, we also identified a set of parameters with much higher intra-individual variation in patients with COVID-19 than in controls, by comparing distributions of within-individuals variation (s.d.) between the two groups with Wilcoxon rank-test (P < 0.01 across two-sample estimate s.d. or three-sample estimate s.d.). We interpreted these parameters as changing. For these parameters we reasoned that samples from an individual with COVID-19 are so variable that they should be treated as independent measures and should not be weighted in any downstream models.

To ensure that the differences observed between COVID-19 and control samples was due to disease status and not due to the differences in age or sex of the cohorts, their effect was tested by comparing nested linear mixed models on healthy control data:

$$\begin{array}{l}{\mathrm{parameter}}\sim 1 + \left( {1|{\mathrm{patient}}} \right) + {\mathrm{weights}}\\ {\mathrm{parameter}}\sim 1 + {\mathrm{age}} + \left( {1|{\mathrm{patient}}} \right)\\ {\mathrm{or parameter}}\sim 1 + {\mathrm{sex}} + \left( {1|{\mathrm{patient}}} \right) + {\mathrm{weights}}\\ {\mathrm{parameter}}\sim 1 + {\mathrm{age}} + {\mathrm{sex}} + \left( {1|{\mathrm{patient}}} \right) + {\mathrm{weights}}\end{array}$$

For the cases when not enough samples were available to estimate patient effects, a linear model was used instead.

Correction with sex and/or age was used whenever a model with additional parameters was better according to Akaike criterion and with P < 0.01 (list of corrected parameters and estimated sex/age effect in Supplementary Table 1). For parameters with significant sex or age influence, estimates of predicted age/sex influence were subtracted from the raw parameter values and residuals were used for downstream statistical testing.

In box plots, testing for differences between COVID-19 (overall status or severity groups), controls and LRTI, raw values (or residuals, where sex/age was significant) between COVID-19 and controls were compared by fitting a linear mixed model:

$$\begin{array}{l}{\mathrm{parameter}}\sim {\mathrm{control}}\_{\mathrm{COVID}}19{\mathrm{status}}\\ + \left( {1|{\mathrm{patient}}} \right) + {\mathrm{weights}}\\ {\mathrm{parameter}}\sim {\mathrm{control}}\_{\mathrm{COVID}}19{\mathrm{status}}\_{\mathrm{severity}}\\ + \left( {1|{\mathrm{patient}}} \right) + {\mathrm{weights}}\end{array}$$

where severity was defined as low for WHO 1–2, moderate for WHO 3–4 and severe for WHO 5–8.

Appropriate comparisons (moderate versus healthy, severe versus healthy, severe versus moderate, COVID-19 versus controls, LRTI versus controls, LRTI versus severity classes) were extracted and the effect size estimated by dividing the difference between estimated population means by the s.d. of controls, unless stated otherwise (for age- and sex-corrected values this was performed on age- and sex-corrected parameters).

Hundreds of hypotheses were tested in parallel (flow cytometry parameters, serology and cytokine levels). These hypotheses are heavily interdependent technically and in some cases biologically, for example, the same subset of cells measured over different panels, complementary subsets of cells and related subsets of cells identified by different markers. Therefore we have provided the raw P values for all comparisons except where stated otherwise. A high, conservative estimate of the number of independent hypotheses tested is 315.

Spearman correlations between parameters in the flow cytometry, serology and cytokine analyses were identified separately in COVID-19 and control cohorts (unless stated otherwise in the text). Correlations with r > 0.3 , r < −0.3 and P < 0.01 were considered significant.

Comparisons between patients with COVID-19 and sepsis (Extended Data Fig. 5d) was undertaken with a two-tailed Student’s t-test.

For NanoString analysis, comparisons per gene between patients with COVID-19 and seronegative controls (Fig. 5c) were performed with a two-tailed Student’s t-test.

Peak antibody titer measurements were compared with a Kruskal–Wallis test and Dunn’s post hoc test. Prognostic parameters were compared as above.

Data structure

Owing to the requirement to establish the experimental pipeline in parallel with patient recruitment and sample processing, the scope of the analysis was expanded over the lifetime of the project. Subsequently, of the 210 samples presented here, the number of samples per panel are as follows: p1 = 196, p2 = 200, p3 = 198, p4 = 180, p5 = 182, p6 = 203, p7 = 182 and p8 = 123. In addition, within a panel, specific parameters may have been omitted based on predefined quality control criteria (see above).

Reporting Summary

Further information on research design is available in the Nature Research Reporting Summary linked to this article.