Longitudinal analysis reveals that delayed bystander CD8+ T cell activation and early immune pathology distinguish severe COVID-19 from mild disease

Summary The kinetics of the immune changes in COVID-19 across severity groups have not been rigorously assessed. Using immunophenotyping, RNA sequencing, and serum cytokine analysis, we analyzed serial samples from 207 SARS-CoV2-infected individuals with a range of disease severities over 12 weeks from symptom onset. An early robust bystander CD8+ T cell immune response, without systemic inflammation, characterized asymptomatic or mild disease. Hospitalized individuals had delayed bystander responses and systemic inflammation that was already evident near symptom onset, indicating that immunopathology may be inevitable in some individuals. Viral load did not correlate with this early pathological response but did correlate with subsequent disease severity. Immune recovery is complex, with profound persistent cellular abnormalities in severe disease correlating with altered inflammatory responses, with signatures associated with increased oxidative phosphorylation replacing those driven by cytokines tumor necrosis factor (TNF) and interleukin (IL)-6. These late immunometabolic and immune defects may have clinical implications.


Introduction
The immune pathology associated with COVID-19 is complex Zhou et al., 2020).
Most infected individuals mount a successful anti-viral response, resulting in few if any symptoms. In a minority of patients there is evidence that ongoing cytokine production develops, associated with persistent systemic inflammation, end-organ damage and often death (Del Valle et al., 2020;Lucas et al., 2020). The relationship between the initial immune response to SARS-CoV-2, viral clearance, and development of the ongoing inflammatory disease which drives severe COVID-19 is not clearly established, nor have the kinetics of the immune changes seen in COVID-19 been fully assessed as disease progresses. Defective immune recovery might drive ongoing disease, and perhaps contribute to secondary immunodeficiency with an increased risk of subsequent infection.
J o u r n a l P r e -p r o o f 4 Severe COVID-19 is associated with profound abnormalities in circulating immune cell subsets. There is a decrease in many peripheral blood subsets of both CD4 + and CD8 + T cells, but an increase in activated and differentiated effector cells (Arunachalam et al., 2020;Hadjadj et al., 2020;Kuri-Cervantes et al., 2020;Laing et al., 2020;Mann et al., 2020;Mathew et al., 2020;Su et al., 2020).
Cells expressing programmed cell death protein-1 (PD1) and other inhibitory molecules are increased, though whether these reflect genuine T cell exhaustion or changes accompanying T cell activation, has not been firmly established. There is, nonetheless, evidence of functional impairment in both CD8 + and CD4 + T cells in a number of studies (Chen and Wherry, 2020). Data on T helper (Th) cell subsets is variable, but there is evidence of increased Th17 cells and markedly reduced T follicular helper cells (Tfh) (Chen and Wherry, 2020;Mathew et al., 2020;Su et al., 2020). There have been conflicting reports regarding B cell immunity, but increased circulating plasmablasts (Arunachalam et al., 2020;Hadjadj et al., 2020;Laing et al., 2020;Mathew et al., 2020) and reduced germinal centre responses (Su et al., 2020) are consistently observed in severe COVID-19. Innate T cell subsets, including gamma-delta (γδ) T cells and mucosal-associated invariant T (MAIT) cells, are also reduced, as are non-classical monocytes and both plasmacytoid and myeloid dendritic cells (pDCs and mDCs) Laing et al., 2020) By analysing transcriptome, serum cytokine and immunophenotyping data of longitudinal samples from COVID-19 patients with a range of disease severities for up to 3 months from symptom onset, we were able to address two important questions regarding the immune response to SARS-CoV-2: (i) How does the very early immune response in patients who cleared virus and recovered from disease with few or no symptoms compare with those who progressed to severe inflammatory disease? This provided insight into the features of the initial immune response that correlate with severe inflammatory outcomes, and whether systemic inflammation is an early or later development in those who progress to severe disease. (ii) How rapidly do the profound immune defects that J o u r n a l P r e -p r o o f 5 accompany severe COVID-19 recover, and do the kinetics of recovery relate to ongoing inflammation and clinical status?
We recruited 207 patients with COVID-19, ranging from asymptomatic healthcare workers in whom SARS-CoV-2 was detected on routine screening, through to patients requiring assisted ventilation, and compared their results to 45 healthy controls. We performed detailed immune phenotyping at multiple time points up to 90 days from symptom onset, reporting absolute cell counts rather than proportions. All data has been made available at https://www.covid19cellatlas.org/patient/citiid/.
We found that the immune response in patients with progressive COVID-19 was delayed compared to those with mild disease, and was inflammatory in nature from the outset. Early immune cellular changes predicted severe disease course, and the variable recovery of these cells over 3 months is associated with marked changes in the nature of the systemic inflammation seen in severe COVID-

Patient cohorts
SARS-CoV-2 PCR positive subjects were recruited for this study between 31 st March and 20 th July 2020. Those without symptoms, or with mild symptoms, were recruited from routine screening of healthcare workers (HCW) (Rivett et al., 2020). COVID-19 patients were recruited at or soon after admission to Addenbrooke's or Royal Papworth hospitals.
Study participants were divided into five categories of clinical severity, used throughout this paper unless otherwise stated (Figures 1A and Data S1). These were: A) asymptomatic HCWs; B) HCWs who either were still working with mild symptoms insufficient to meet the criteria for self-isolation (Rivett et al., 2020), or who were symptomatic and self-isolating; C) patients who presented to hospital but never required oxygen supplementation; D) patients who were admitted to hospital and J o u r n a l P r e -p r o o f 6 whose maximal respiratory support was supplemental oxygen; E) patients who at some point required assisted ventilation. Three patients who died without admission to intensive care were also included in this severe group. In addition, 45 healthy controls (HCs) were recruited, distributed across a range of age and gender.
We analysed the immune phenotype of 605 blood samples from 246 participants out to 90 days from the onset of symptoms ( Figure 1A, Tables S1 and S2). As the clinical severity category increased, patients were more likely to be older and to be male (Figures 1B and 1C), as expected . A high-sensitivity C-reactive protein (CRP) assay demonstrated that classifying disease severity on the basis of maximal respiratory support is reflected in the CRP ( Figure 1D).
Patient courses are measured in time since symptom onset for groups B through to E. As they are asymptomatic, those in group A are measured from the date of their first positive swab -they are therefore likely to have been sampled later post-infection than patients in groups B-E, and are therefore not directly comparable to them in terms of time course. CRP (and, later, other variables) is compared to the interquartile range of 45 healthy controls. Nasopharyngeal swabs were assessed for SARS-CoV-2, allowing inclusion in this study, and were repeated in some patients. Initial viral titres, reflected by low PCR cycle threshold (CT) values, were higher in group E. With only occasional exceptions, patients in all severity groups had cleared virus by 24 days after symptom onset ( Figures   1D and S1A). Of the 6 patients with positive swabs after 30 days, four were overtly immunosuppressed (3 solid organ transplants with recent induction/rejection treatment, 1 myeloma on B-cell depletion therapy) and one was a peritoneal dialysis patient admitted with peritonitis.
We will first outline the major datasets collected in this study, before integrating them to study early and recovering disease.

Changes in cytokines and complement components with time and disease severity.
J o u r n a l P r e -p r o o f 7 Asymptomatic HCWs in group A had no evidence of cytokine or complement dysregulation, while those with mild symptoms (group B) showed an early, transient increase in C3c and the terminal complement complex (TCC), but not in CRP or cytokine concentrations (Figures 1E and S1B). Once patients developed symptoms severe enough to attend hospital (group C or above), a different picture was apparent. Both IL-6 and TNF-α were raised, along with other cytokines, as were all complement components measured. These abnormalities were maximal at the first bleed, and largely persisted in group E. Abnormal IL-6 and TNF-α persisted in groups C and D despite clinical improvement (all had been discharged by days 49-60). Interferon-gamma (IFN-γ) was briefly raised in only a subset of patients. Increased C3c was prominent in group B, while C3a became the dominant complement component elevated in more severe disease (groups C -E).
IL-6, TNF-α, IL-10 and IL-1β rose in those with more severe disease (groups C-E) but, in contrast, there was no increase in inflammatory cytokines in groups A and B, pointing to marked differences in very early inflammatory responses between resolving and progressive disease. In addition, the persistence of cytokine abnormalities even beyond 60 days from symptom onset could have implications for resolution of clinical disease.

Both onset and recovery of immune cellular abnormalities vary with disease severity
Using standardised flow cytometry panels, we explored the size of 64 cell populations over time across the five clinical strata. Trucount analysis enabled calculation of absolute cell numbers. Cellular changes with time were assessed (examples in Figure 2A), and outcomes for 30 cell types has been shown relative to the median for HCs ( Figure 2B). CyTOF, which uses whole blood rather than peripheral blood mononuclear cells (PBMCs), was also used in a subset of patients, to allow quantification of granulocytes and non-classical and intermediate monocytes ( Figure S2, and Methods). See Data S2 for details of all cell types, including data beyond 48 days of symptom onset and time as a continuous variable.
J o u r n a l P r e -p r o o f 8 Few changes in the immune phenotypes were seen in patients with asymptomatic (A) and very mild (B) disease, but once symptoms warranted presentation to hospital, the picture changed.
Widespread immune abnormalities in groups C-E were most marked at the first bleed ( Figure 2B), even when this was within 0-2 days of symptom onset (Data S2). Almost all CD4 + T cells subsets were reduced, as were many CD8 + T cell subsets and both naive and memory B cells -in contrast plasmablast numbers rose in all groups. Innate lymphoid subsets, including MAIT cells, various γδ T cell subsets, and natural killer (NK) cells, were also reduced, as were mDCs, and both non-classical and intermediate monocytes. These changes were correlated with, and were predictive of, severity, as discussed below.
We also calculated leucocyte number as a proportion of "parent" populations, either of total PBMC ( Figure S3A) or of major lymphoid compartments ( Figure 2B and S3A), to allow comparison with most published literature (and all single cell studies). Considering proportion underestimates COVID-19-associated pathology, for example missing the early severity-correlated reduction in lymphocyte subsets, and the late persistence of low numbers of T cell subsets in severe disease. In agreement with this, analysis of CyTOF and CITE-seq data, which also examine cell proportions, did not identify most alterations in cell populations between severity groups that were observed by flow cytometry ( Figure S3B and C).

Blood transcriptomic inflammation-related signatures vary with severity and time.
Whole blood RNA was isolated, and transcriptomes were generated by RNA-Sequencing (RNA-Seq) and analysed (in two time "bins" -0 to 24 days and 25 to 48 days) using PLIER, which performs matrix factorization to identify interpretable latent factors. The contribution to each latent factor by immune cell subsets was then calculated (Figure S4A and B). These RNA-Seq-derived latent factors were broadly aligned with the pattern observed in the cell count data ( Figure 2B). An exception was the pronounced neutrophil signature seen early across groups C to E, and persisting at day 25-48 in group E, demonstrating more pronounced neutrophil dysregulation across severity categories than J o u r n a l P r e -p r o o f 9 suggested by increasing neutrophil number alone. An erythrocyte gene expression-driven latent factor was also prominent in group E at late times, may be associated with heme metabolism, and discussed later.
We then used unbiased weighted gene correlation network analysis (WGCNA) in the whole blood transcriptome data to identify modules of co-regulated genes, where each could be summarised as an "eigengene". Prominent modules correlated with both disease severity and time (Figures 3A and   B, S4C, S4D and Table S3). The module enriched for TNF-α /IL-6 genes correlates well with the cytokine concentrations determined in Figure 1 -rising early in groups C-E and then largely resolving by 25-48 days. A neutrophil activation module was also prominent early across groups C to E, as was one associated with glycolysis. Thus, there was clear transcriptional evidence of activation of broad inflammatory pathways at early time points, and these largely recovered in most patient groups (with the exception of group E, in which many patients had persistent disease). In contrast, an interferon-related module was upregulated in groups B-E at day 0 to 24 from symptom onset, before declining ( Figures 3A). The relative contributions of Type I, II and III interferons to this interferonstimulated gene (ISG)-associated module cannot be easily distinguished (Banchereau et al., 2017), but a more detailed analysis of its kinetics showed that, while expression peaked at different eigengene values in each severity group, it then declined in all of them by around 30 days ( Figure   3C), coincident with viral clearance and occurring irrespective of clinical and inflammatory state ( Figure 3D).
Finally, a supervised gene set enrichment analysis (GSEA) performed using Hallmark gene signatures ( Figure 3E) (Liberzon et al., 2015) showed results largely consistent with the unbiased approaches just described, including demonstrating a late upregulation of genes associated with reactive oxygen species (ROS) and oxidative phosphorylation (OXPHOS) discussed below in the context of immune recovery.
J o u r n a l P r e -p r o o f Immune phenotype at presentation correlates with severity and may predict outcome.
To determine if the immune phenotype at presentation correlated with, or indeed could predict, subsequent disease course, we first performed a Principal Component Analysis using cell numbers across 24 primary immune cell populations from the 84 participants with blood draws taken between 0 and 10 days after the development of symptoms. Those in groups A and B clustered with HCs and away from those in groups C-E ( Figure S5A). Hierarchical clustering of absolute cell counts identified two clusters (Figure 4A), one almost entirely comprised of HCWs from groups A and B (cluster 2), and the other containing all patients who progressed to ventilation and/or death, and most who required supplementary oxygen support (cluster 1). Clustering using RNA-Seq data obtained from 1-10 days after onset largely recapitulated that seen using cell number, and was driven by ISG, TNF-α and IL-6 associated gene pathways ( Figure S5B). The severe cluster 1 was associated with increased age, CRP, TNF-α and IL-6 (Figures 4B and Data S3A). Early differences between cell types drive this clustering, irrespective of their different subsequent trajectories (Data S3B). We used a sparse partial least squares discriminant analysis (sPLS-DA) to determine which cell subsets were most informative for cluster prediction: clusters 1 and 2 could be discriminated with a minimum classification error rate of 0.07 ± 0.02 (93% accuracy) based on 13 key cell populations selected by the model (Figures 4C, S5C-E). The area under the receiver operator characteristic curve (AUROC) for patient cluster classification based on these 13 cell types was 0.98 (98% chance of accurate cluster prediction) ( Figure S5F). CRP alone was inferior to the cell types in classifying patients, and nor did it improve their performance when added ( Figure S5G). These cell types were often the most profoundly and persistently affected by severe COVID-19 (particularly MAIT, γδ T cells, Tfh and CD4 + Temra cells).
The ability to cluster patients and separate those with mild or no symptoms (groups A and B) from those presenting to hospital (groups C-E) underlines the profound association of immune subset abnormalities with disease severity. It does not, however, predict outcome in a clinically useful way, J o u r n a l P r e -p r o o f 11 as such predictions are only of practical value in those who present for medical attention. We therefore incorporated additional data related to inflammation, including cytokines, age, CRP and complement components; re-clustering with this combined dataset led to a smaller severityassociated group (Figure 4D). Input from both immune cell subsets and inflammation-related markers was required for optimal classification ( Figure 4E). This clustering, when considering only the subset of patients recruited at hospital presentation (groups C-E) and bled within 10 days of symptom onset, predicted disease progression, as defined by the subsequent need for increased respiratory support, or death, after blood sampling ( Figure 4F). This analysis requires validation with larger patient numbers, but is comparable to similar observations made by others (e.g. (Laing et al., 2020;Mathew et al., 2020) and indicates that a combination of immune phenotype data and inflammatory markers could provide potentially clinically useful prediction of disease progression.

Early immune responses correlate with COVID-19 severity.
HCW in categories A and B did not progress to severe COVID-19 disease, and they also clustered apart from those with more severe disease when using either early immune cell counts or transcriptome ( Figure 4A and S5B). We therefore compared the early features of immune responses in groups A and B to those in patients with more severe COVID-19, binned in 7-day intervals to provide finer definition, to seek early immune and inflammatory correlates of disease severity and brief compared to severe disease, and was consistent with tissue localisation and local interferon production to support the antiviral response (Cella et al., 1999).
In group B cytotoxic CD8 + T cells rose earlier than in groups C to E, apparent by day 7 and peaking up to 2 weeks after symptom onset, in contrast to the later and more sustained rise seen in severe COVID-19 (Figures 5A, 5B and S6A). Early enrichment of a CD8 + cytotoxic RNA signature was also seen in group B compared to C to E ( Figure S6C). Consistent with these findings, spontaneous generation of IFN-γ by T cells was more pronounced in group B samples taken two weeks post symptom onset ( Figure 5D).
Antigen-specific CD8 + T cell responses were determined after simulation with anti-SARS-CoV-2 peptides and subtraction of background spontaneous IFN-γ-producing T cells. They were similar in all severity groups early ( Figure 5D and S6D), and so were not responsible for the increased effector CD8 + T cells seen in mild disease, suggesting instead a role for "bystander" activation. Such bystander-activated cells are important in early anti-viral defence (Maurice et al., 2021), and characteristically express NKG2D, involved in the killing of infected cells, and both IL-7 receptor (IL7R) and CD8, which are down-regulated in a T-cell receptor (TCR)-dependent fashion (Kaech et al., 2003;Slifka and Whitton, 2000). CITE-Seq protein expression analysis of activated CD8 + T cells demonstrated an increased in surface NKG2D, IL7R and CD8 in group B compared to more severe groups ( Figure 5E), observations confirmed by RNA analysis (Figure 5E). Bystander CD8 + T cells expressing CXCR3 rapidly home to sites of inflammation (Maurice et al., 2021), consistent with enrichment for CXCR3 RNA but not surface protein ( Figure 5E). Transcriptional signatures derived bystander-activated CD8 + T cells were enriched in patients with mild disease, while those from TCRactivated cells were associated with severe COVID-19 ( Figure 5F), again consistent with widespread early bystander activation in the CD8 + T cell population in patients destined to have good disease outcomes.
J o u r n a l P r e -p r o o f 13 An early increase in plasmablasts was seen in groups A and B, occurring up to a week before a larger rise in more severe disease (Figures 5C and S6B). We therefore measured total immunoglobulin concentrations and anti-spike IgG and anti-SARS-CoV-2 neutralising antibody titers ( Figure 5G, H and S6E, F). Group B patients maintained their serum IgM concentrations, which fell rapidly in those with more severe disease (Figure S6G), and their titres of anti-spike IgG and early neutralisation responses were comparable to patients progressing to more severe COVID-19 ( Figure 5G and H).
This suggests that the B cell difference between patients with mild and more severe disease might lie in more robust non-antigen-specific B cell activation, perhaps impacting via "natural" antibody and/or non-antibody-dependent mechanisms.
Virus at first swab, as assessed by PCR CT value, was comparable in groups B, C and D (and the lower titre in group A samples was not comparable as they will have been taken later after infections, as described above). Initial viral titre was therefore not associated with an increased risk of hospital admission (being similar in groups B, C and D), but was higher in group E ( Figure 5I). These viral titres were reflected in interferon-related transcription signatures, which are prominent in groups B-E ( Figures 3A, C and S6H). Despite the fact that high IFN signatures correlated with severe disease, they were not necessarily driving that severity. The fact that the subgroup with the highest IFN within group E do best ( Figure S6I), suggests that the inflammatory pathways that create severe disease are distinct from IFN (which is also consistent with comparable kinetics of reduction of IFN signature regardless of severity - Figure 3C), and indeed that a robust IFN response may be beneficial in this context, though this needs confirmation in a larger dataset.
Taken together, these data suggest that an early adaptive immune response is prominent in individuals who are asymptomatic or have mild disease, characterised by a rapid production of activated bystander CD8 + T cells, plasmablasts and likely pDC tissue localisation before antigen-J o u r n a l P r e -p r o o f 14 specific responses become apparent. This appeared a more important correlate of severity than viral titre, which only became relevant in those progressing to ventilation or death ( Figure 5J).
In those with more severe disease (groups C-E), evidence of systemic inflammation was present from the first blood test (Figure 5). If we focused on the 16 patients in groups C-E sampled between 2 days before and 4 days after symptom onset, 15 had a CRP > 10mg/L and/or neutrophil activation eigengene > 0. All 5 patients sampled between 2 days before and 2 days after symptom onset met these criteria. In contrast, this was not seen in groups A or B despite the fact that they mounted a more prominent early cellular response. It was not clear whether inflammation seen in C-E was a causally related to the poor early cellular response seen in these patients. Such inflammation clearly did not, however, develop later from the progression of a non-inflammatory immune response or as a result of failure to clear virus, and suggested that the inflammatory die is cast by the time symptoms appear, and thus strategies to prevent it and reduce its clinical impact would need to be established very early ( Figure 5J).

Distinct patterns of immune recovery in COVID-19.
In contrast to groups A and B, cellular changes in groups C -E were profound and usually most prominent at the first bleed (Figure 2), so determination of change in immune cell subsets over time was likely to be most informative in these groups. We therefore explored cell kinetics in groups C-E, assigning patients to two categories based on whether their CRP concentrations remained elevated above 10mg/L ("Persisting CRP") or fell below 10mg/L ("Resolving CRP") by their final bleed within 3 months post symptom onset ( Figure 6A). The latter group included both individuals with early high CRP that then fell, together with those for which CRP remained low (10mg/L) throughout. Changes in CRP over time differed between these two groups when assessed using a mixed-effects model, with time modelled as a continuous variable ( Figure 6B).

J o u r n a l P r e -p r o o f
15 To compare cellular changes over time between persisting and resolving CRP patient groups, a "rate of change" for each cell population was calculated over 60 days post symptom onset. In brief, this rate captured both the initial deviation in cell counts from normal within a window of 0-12 days, and the time taken for cells counts to stabilise within a normal range if cellular recovery did occur (see Methods). Five predominant trajectories were observed; populations that did not deviate from heathy numbers over the duration of study (e.g. NKT cells), those which increased progressively from normal over time (e.g. effector CD8 + T cells), those which fell progressively from normal over time (e.g. transitional B cells), those which trended toward recovery after an initial rise in numbers (plasmablasts) and those which tended toward recovery after an initial drop in numbers (e.g. naïve CD4 + T cells) ( Figure 6C).
The absolute number of most cell populations fell precipitously early, and then showed variable recovery. For descriptive purposes these were arranged into a group of cell subtypes that failed to recover, or recovered, in the persisting CRP group ( showing an early drop in counts (with the exception of naïve CD8 + T cells) recovered in those with resolving CRP (II and IV), and at rates more rapid than seen in those with persisting high CRP. In the persisting CRP group, a number of cell types remained markedly abnormal (including memory B cells and various CD4 + T cell subsets: quadrant I), whereas a second group of cell types recovered despite persisting inflammation (including NK cells and some CD8 + T-cell subsets: quadrant III).
We then explored the relationship between cell recovery and the inflammatory response. It might be expected that where CRP remained persistently elevated, immune defects might persist, if these defects were secondary to the inflammatory state. Consistent with this, the cohort with persistently J o u r n a l P r e -p r o o f 16 raised CRP also has raised TNF-α and IL-6 protein ( Figure 1D). Likewise, transcriptional signatures of TNF-α/IL-6 and neutrophil activation were increased in severe disease (Figure 3), particularly in the persistent CRP group ( Figure 6D). This ongoing inflammation may contribute to the sustained reduction in cell numbers at late times seen in quadrant I, together with persistently raised HLA-DR + /CD38 + effector T cells and plasmablasts ( Figure 6C). It was also therefore not unexpected that most cell types reduced in acute disease recover over a few weeks as the CRP falls, as was seen for most cells in quadrants II and IV ( Figure 6C).
More intriguing were the cells that recovered rapidly in the face of ongoing inflammation (quadrant III). While the reasons for this are likely to differ between cell types, and to be multifactorial, these cell reductions might be driven in part by the viral infection per se and/or virus-induced interferon. It was notable that, after initially rising, IFN-γ (Figures 1E and S1B) and Interferon-stimulated gene (ISG) signatures fell to normal values independent of both disease severity group ( Figure 3C) and CRP ( Figure 6D), but correlating with declining virus titre (Figures 3D and S4E). Thus, cell types known to leave the circulation due to interferon stimulation (Kamphuis et al., 2006), such as T and NK cells (Hirsch and Johnson, 1986;Zafranskaya et al., 2007), may recover as viral infection is controlled and interferon-dependent inflammation falls, independent of ongoing CRP-associated inflammation.
Finally, a small number of cell types remained statistically abnormal after 60 days, even in the resolving CRP group. Thus effector CD4 + and CD8 + T cells and plasmablasts remained elevated, and pDCs, Tfh, Vg9Vd2 expressing γδ T cells and MAIT cells remained reduced ( Figure 6C) and were among those cells most predictive of poor prognosis ( Figure 4C). These abnormalities persist despite resolution not only of CRP but of neutrophil degranulation, TNF-α /IL-6 and glycolysis-related signatures (e.g. Figures 1D and 6D). Possible mechanisms behind these sustained abnormalities are discussed below.

J o u r n a l P r e -p r o o f
17 The late appearance of OXPHOS and ROS pathways correlates with differential immune recovery.
In late, severe COVID-19, whole blood transcriptome analysis showed prominent inflammationrelated signatures that were distinct from those seen early in disease. These signatures were related to OXPHOS, ROS and heme metabolism. These were demonstrated in an un-biased fashion using WGCNA, where modules characterised by OXPHOS and heme metabolism were prominent at day 25 to 48 post symptoms, with OXPHOS most prominent in group E, and heme metabolism in C-E ( Figures 3B and 3C). Enrichment of Hallmark signatures confirmed the association of OXPHOS and heme metabolism in groups C-E, and also found association of a ROS signature (Figures 3E and 7A).
Consistent with this, the expression of the genes driving the enrichment of each signature was upregulated in the three most severe clinical groups ( Figure 7B and Table S3). The late rise in these three correlated signatures occurred irrespective of persisting or resolving CRP-associated inflammation ( Figure 6D), and appeared independent of specific cell population recovery ( Figure   S7A).
We then correlated cellular and transcript signature changes in COVID-19. In the first 24 days after symptom onset, there was a strong association between TNF-α /IL-6, neutrophil degranulation and interferon signatures with most cell subsets reduced in severe disease ( Figure S7B). Later, between 24 and 48 days after symptom onset, these associations changed ( Figure 7C and S7C). While TNF-α /IL-6 and neutrophil degranulation signatures were still associated with many cell subsets that continue to be reduced, the interferon signature was no longer a significant player. The persistent increase seen in activated effector CD4 + and CD8 + T cells and plasmablasts was now particularly associated with the OXPHOS signature which, having become more prominent later in disease ( Figure 3B), had a much more restricted and specific association with immune dysfunction than other inflammatory signatures.

J o u r n a l P r e -p r o o f
18 It was thus clear that, for some cell types, the association between their number and the inflammatory milieu changed over time, but for others it was more consistent. It was the inflammatory signatures which appeared late in disease, in particular OXPHOS, that were specifically associated with persistent derangement of cell types of potential pathological importance, such as increased HLA-DR + CD38 + T cells and plasmablasts, and reduced pDCs ( Figure 7D).

Discussion.
In asymptomatic or mildly symptomatic, non-progressive SARS-CoV-2 infection (groups A and B) there was evidence of an early robust adaptive immune response. Circulating plasmablasts and CD8 + HLA-DR + CD38 + activated T cells expanded earlier and in higher numbers than in more severe COVID-19 groups, most notably in the first week after symptom onset. Both of these cell populations then contracted in A and B, as they continued to rise in groups C-E. Despite this, and a prominent early interferon transcription signature, both the antibody and T cell SARS-CoV-2-specific immune responses were comparable between early mild and severe disease. This suggested the increased plasmablasts and effector CD8 + T cells could reflect an enhanced bystander response in mild disease, something then supported by single cell CITE-Seq analysis.
Bystander CD8 + T cell activation was first described in the context of LCMV infection in mice (Tough et al., 1996). It involves activation of memory CD8 + T cells independent of TCR stimulation (whether by the initiating or cross-reactive antigen) (Maurice et al., 2021). It is driven by type I IFN in many inflammatory contexts, in particular viral infection. Bystander CD8 + T cells are activated well before the antigen-specific response is seen, being detectable within a day of viral infection, and often peaking within a week (Berg et al., 2003). Bystander cells can migrate to infected organs using CXCR3, and kill virally infected cells through NKG2D and granzyme B-dependent mechanisms (Maurice et al., 2019). They have both been shown to be critically important in the early defence against viral infections, in part by direct anti-viral effects, and in part by producing IFN-γ and other cytokines which then activate antigen presenting cells (Soudja et al., 2014) and control the memory/effector balance of subsequent antigen-specific responses (Krummel et al., 2018). Similar, though less well defined, bystander phenomena occur in CD4 + , MAIT and γδ T cell responses (Holzapfel et al., 2014;Ribeiro et al., 2015;Ussher et al., 2018). The early and pronounced production of bystander-activated CD8 + T cells in patients groups A and B could be directly related to disease resolution and prevention of progression to more severe disease, a process likely to begin before symptom onset. Understanding the factors which impede this early bystander response could be important for developing strategies to prevent severe disease.
In contrast, persistent bystander CD8 + T cell activation has been associated with inflammatory pathology in the context of both chronic infection and autoimmunity. NKG2D-dependent killing of non-virally infected hepatocytes in hepatitis A exacerbates liver damage (Kim et al., 2018) and may also play a role in inflammatory lung disease (Borchers et al., 2009), and there is evidence that NKG2D ligands are expressed in the lungs in COVID-19 (Wauters et al., 2021), so it could be that a late and persistent expansion of bystander-activated effector T cells in severe COVID-19 could be a driver of lung pathology. The well-documented links between bystander CD8 + T cell activation and autoimmunity (Groh et al., 2003;Meresse et al., 2004) also raises the prospect they may play a role in the autoimmune manifestations observed in COVID-19.
The early increase in circulating plasmablasts seen in mild COVID-19 is likely to be comprised of SARS-CoV-2 specific cells, perhaps to an extent not yet fully reflected in serum antibody, together with non-SARS-CoV-2-specific bystander-activated cells. Bystander B cell activation is known to involve memory (Horns et al., 2020) as well as innate-like B cells (including B1 cells in the mouse, and marginal zone cells in both mouse and human). Their function is not fully defined, but includes an increase in production of "natural antibodies" (Antin et al., 1986) alongside non-antibody-associated functions, which may include antigen transport to secondary lymphoid organs, antigen presentation J o u r n a l P r e -p r o o f 20 to T cells, cytokine production and immune regulation. These "bystander" B cell functions are known to play a role in early defence against bacterial infection (Boes et al., 1998;Haas et al., 2005;Ochsenbein et al., 1999), and may be involved in autoimmunity (Espeli et al., 2019;Sanderson et al., 2017). Their role in viral infection is less clear, but it may be that early bystander B cell activation helps determine the outcome of COVID-19, in a manner analogous to the role proposed for bystander CD8 + T cells. A closer investigation of the early B cell response in COVID-19 will be required to confirm this.
At the same time as this pronounced early immune response is seen in groups A and B, there is no evidence of systemic inflammation, apart from early, transient complement activation. CRP, circulating TNF-α and IL-6, and transcriptional signatures of a number of inflammatory pathways are not raised in groups A and B, but are already prominent in groups C-E. The severe and widespread leucocyte subset depletion seen at the initial bleed in the more severe COVID-19 patient groups C-E, and observed by many others (Carissimo et al., 2020;Laing et al., 2020;Mathew et al., 2020;Su et al., 2020), is not apparent in those with asymptomatic or mildly symptomatic disease, suggesting this too is a feature of a pathological immune response. Immune cell subset numbers correlated strongly with severe and progressive disease. Coupled with evidence of early systemic inflammation in severity groups C-E, our findings suggest that the immune pathology associated with severe COVID-19 is either established immediately post-infection or, if there is a transition point from an effective to a pathological response, this is likely to occur before the time of symptom onset. This finding may have major implications as to how disease needs to be managed, since intervention to prevent immune pathology would need to be targeted very early in the disease course, and perhaps preemptively in high risk groups screened and diagnosed before symptoms develop.
The reason for the failure to mount a robust early B and T cell response in the context of severe COVID-19 was likely to be multifactorial. There was no evidence for a relationship with viral load and J o u r n a l P r e -p r o o f 21 progression to inflammatory disease, as initial viral titres were comparable between groups B, C and D. Once inflammatory disease is established, however, viral titre may be associated with subsequent outcome, as increased viral titre was seen in group E, consistent with reports that high initial viral titre might be associated with mortality (Pujadas et al., 2020). Genetic association studies in severe COVID-19 point to genes that are implicated in driving antiviral responses. The most prominent are associations with genes involved in the type I interferon pathway (Pairo-Castineira et al., 2020;Zhang et al., 2020), known to be the key driver of bystander T cell activation (Maurice et al., 2021).
Increasing age and comorbidity such as diabetes and chronic inflammatory disease are known to suppress early CD8 + T and B cell responses (Shen-Orr et al., 2016;Weiskopf et al., 2009). An in-depth understanding of these risk factors may instruct strategies to assess risk of progression before inflammatory responses become self-sustained.
While clear distinctions in immune responsiveness were apparent between groups A and B versus groups C, D and E, differences between the more severe groups themselves were less obvious.
Those with symptomatic disease warranting admission to hospital clustered together using the size of just 13 key cell populations: these clusters correlated strongly with clinical severity, and immune cell subset numbers together with inflammatory mediator concentrations provided prediction of subsequent progression, as well as COVID-19 associated death. Similar observations have been made by others (Chen and Wherry, 2020). Predicting progression after presentation to hospital with moderate/severe disease could be of limited clinical use, and it would be of more benefit to predict progression in cases with milder COVID-19 -but this may not be possible in practice, given inflammatory immunopathology is likely to be present at symptom onset. A study to address this issue would need to be conducted in particularly high-risk patient groups to ensure an adequate event rate, and require diagnosis through asymptomatic screening to detect changes before symptoms develop. probably as a result of migration to the lungs. Finally, plasmablast expansion has also been seen, particularly in dengue fever (Wrammert et al., 2012). The nature of these studies makes a direct comparison with the situation in SARS-CoV-2 difficult, but it would seem likely that many immunological changes seen in COVID-19 mirror those seen in responses to other infections, though in general the cellular changes we observe appear to be more profound, peak earlier in disease course, and are more persistent in COVID-19 than in studies performed in other infectious contexts.
The recovery of the profound immune dysregulation seen in severe COVID-19 is potentially of major clinical relevance, as such recovery may be required for the resolution of inflammatory disease or to prevent secondary infection or SARS-CoV-2 reinfection. We found immune cell abnormalities often persist for weeks to months after SARS-CoV-2 infection, and different cell populations exhibited quite different patterns of resolution. Some recovered as systemic inflammation resolved, others did not, and a third group resolved even in the face of persistent systemic inflammation. Understanding the inflammatory drivers or associations of this differential recovery could provide insight into the immune pathology of COVID-19, and potentially of other infections. To begin to explore this, we correlated immune changes with measurements of systemic inflammation throughout the disease course. Patients with severe COVID are characterised by high CRP, and this correlates with evidence of TNF-α and IL-6 driven processes both at the protein and transcriptome expression, as well as with both neutrophil activation and glycolytic metabolism. The fact that many cellular abnormalities persisted while these biologic processes were apparent -while others appeared to resolve alongside Finally, some cell populations remained markedly abnormal, or showed a limited recovery, even once CRP-associated inflammation has resolved, and indeed after patients have been discharged from hospital. These persistent changes may reflect a slow intrinsic regenerative capacity of the cell type concerned, but in other situations, such as the continued elevation of effector T and B cells, it is tempting to speculate that there is ongoing abnormal signalling driving such changes. For that reason, we explored late changes that are seen in the inflammatory response in COVID-19.
Transcriptional signatures associated with OXPHOS-, ROS-and heme-related metabolic pathways arose late in those with severe COVID-19, but were not prominent in early disease. Activation of immune cells results in metabolic reprogramming that supports cell growth, proliferation and differentiation. Disruption of metabolic pathways can, through many machanisms, result in immune dysfunction (Bantug et al., 2018). It is unlikely that the metabolic signatures observed here simply reflect heightened bioenergetic requirements of activated immune cells, as one would expect similar requirements to be also present early in disease. OXPHOS can drive inflammation (Mills et al., 2017), and we note that COVID-19 patients treated with metformin, which inhibits Complex I of the respiratory chain, had lower amounts of circulating inflammatory cytokines .The ROS transcriptional signature may reflect more abundant production of ROS-species inevitably accompanying increased OXPHOS. Alternatively, it may reflect specific mitochondrial pathology, and thus per se contribute to immune cell dysfunction (Nathan and Cunningham-Bussel, 2013).
Mitochondria are also critically involved in heme biosynthesis. Heme serves as a prosthetic group for haemoglobin as well as many other proteins -including several that constitute the respiratory chain of mitochondria. While free heme can act as damage-associated molecular pattern and promote ROS formation, the role of heme biosynthesis vs. catabolism in balancing cellular sensitivity to oxidants is complex and context dependent (Prestes et al., 2020). Here, given correlated regulation J o u r n a l P r e -p r o o f 24 of heme and OXPHOS pathways in the clinical categories C-E, activity of these modules may be interrelated and possibly together reflect dysfunctional mitochondria. How heme and OXPHOS transcriptional programmes are linked on a molecular level cannot be inferred from our data.
Erythroid cell activation has recently been detected in severe COVID-19 (Bernardes et al., 2020) and could also contribute to a heme transcriptional signature. However, the increase in heme metabolism in our cohort correlates strongly with falling haemoglobin, and reticulocytes in patients in groups C, D and E are low -suggesting suppression rather than activation of erythropoiesis.
Understanding the mechanism linking metabolic dysregulation to persistent immune pathology in COVID-19 will require further study over longer disease courses.

Limitation of study
Many of the abnormalities we have observed in COVID-19 might also be features of other severe viral infections. To identify which are COVID-specific will require a comparison with an appropriate disease control group. Sample size is critical when studying heterogeneous disease, and severe COVID-19 falls into this category. While the cohorts described here are comparable in size with large detailed immunophenotyping studies, larger ones would nonetheless increase our power. In particular, validation of the prognosis prediction models described here is required in large independent cohorts. Very early features of the immune response are associated with disease outcome -exploring these during the period between infection and symptom onset will be important to understand disease progression, but presents major practical difficulties. Continued follow-up of patients will be needed to determine the persistent of abnormalities still observed at late time points. Finally, since our patients were recruited during the first pandemic wave, a followup study examining the immune response to new SARS-CoV-2 strains with different virulence could be informative.

Declaration of Interests
The authors declare they have no competing interests. E.J.M. Toonen is an employee of Hycult Biotechnology b.v.

Lead contact
Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact Prof Kenneth Smith (kgcs2@cam.ac.uk).

Materials availability
This study did not generate new unique reagents.

Data and code availability
The datasets generated during this study are available at NIHR CITIID COVID-19 Cohort (https://www.covid19cellatlas.org/patient/citiid/). In addition, whole blood RNA-seq data are available at Europen Genome-phenome Archive (EGA, ID: ), and flow cytometry data are available at Detailed information on the patients (age, gender, and clinical features) can be found in Table S1. Inpatients were sampled at study entry, and then at regular intervals as long as they remained admitted to hospital (approximately weekly up to 4 weeks, and then every 2 weeks up to 12 weeks).
Discharged patients were invited to provide a follow-up sample 4-8 weeks after study enrolment.
Health care workers were sampled at study entry, and subsequently after approximately 2 and 4 J o u r n a l P r e -p r o o f 32 weeks. At each time-point, blood samples were drawn in EDTA, sodium citrate, serum and PAXgene Blood RNA tubes (BD Biosciences) and processed by the CITIID-NIHR COVID BioResource Collaboration group (Document S1).

Cell lines
HEK293 T cells (a female cell line) were a kind gift from Paul Lehner, authenticated by STR profiling (Menzies et al., 2018;Miles et al., 2017). They were cultured in DMEM supplemented with 10% fetal calf serum (FCS), 100 units/ml penicillin, and 0.1 mg/ml streptomycin at 37 C in 5% CO2.

Virus
The virus used in this study was the clinical isolate SARS-CoV428 2/human/Liverpool/REMRQ0001/2020, a kind gift from Ian Goodfellow (University of Cambridge), isolated by Lance Turtle (University of Liverpool), and David Matthews and Andrew Davidson (University of Bristol) (Daly et al., 2020;Patterson et al., 2020).

Clinical data collection
Clinical data were retrospectively collected by review of medical charts and entered into spreadsheets or Castor EDC, a cloud-based clinical data management system. Available laboratory test results and administered in-patient medications were extracted from Epic electronic health records (Addenbrooke's Hospital) and from MetaVision ICU (RPH ITU). Data were merged from the various data sources using R version 3.6 and the R packages readr (1.3.1), openxlsx (4.1.4), dplyr Participants in group A were further sub-grouped according to whether they were completely asymptomatic (n= 8), or had had any of the above COVID-19 symptoms before PCR testing (n=10, median time from symptoms to COVID-19 PCR test 26 days, range 9-42 days).
Group B participants included both staff who were self-isolating because of COVID-19 symptoms (n=30), and staff members who were reporting fit for duty, but had some symptoms that did not reach the threshold for self-isolation at that time (n=10).
Hospital patients were assigned to 3 severity groups, mainly reflecting the maximum intensity of respiratory support for COVID-19 received during their hospital stay: group C: did not receive any supplemental oxygen. Five patients were discharged soon after initial diagnosis and assessment but followed up as part of the study.
group D: received supplemental oxygen using low flow nasal prongs, simple face mask, Venturi mask or non-rebreather face mask.
group E: received any of non-invasive ventilation (NIV), mechanical ventilation or ECMO. Patients who received supplemental oxygen (but no ventilation) and deceased in hospital were also assigned to group E.
In patients who were already established on home NIV for chronic respiratory failure, NIV delivered as per the home prescription (e.g. nocturnal) was not considered for the purpose of classification.
Moreover, oxygen requirements that were clearly not related to COVID-19 were also not considered for classification purposes. In particular, 2 cases who received low flow supplemental oxygen for non-COVID-19 indications (ascitic splinting in decompensated cirrhosis in one case, and recovery J o u r n a l P r e -p r o o f 34 from anesthesia after orthopedic surgery in the other) were assigned to group C. Cases in group C were further sub-classified according to chest radiology results (X-ray and, if available, CT scan), as: abnormal radiology: chest X-ray/ CT scan showed changes compatible with COVID-19 normal radiology: chest X-ray/ CT scan did not show any abnormality compatible with COVID-19 (reported as normal or showing lung changes diagnostic of conditions other than COVID-19).
Immunological parameters were analyzed according to time since onset of symptoms, or otherwise time since positive SARS-CoV-2 NAAT (in group A and in 4 asymptomatic patients in group C). Seven cases admitted to hospital for COVID-19 had no date of onset of symptoms documented in the medical records. In these cases, the date of onset of symptoms was estimated as follows: hospital admission date -median time from symptoms to hospital admission in patients admitted for COVID-

19.
Following clinician review, 6 cases were considered not classifiable, due to complex concomitant pathologies that coexisted with COVID-19 and dominated the clinical picture, confounding the interpretation of clinical outcome. These cases were not included in any analyses; more details are reported in Table S2. Data S1 summarizes the timing of research samples and clinical trajectories for volunteers in severity groups C, D and E included in the analysis.

Peripheral blood mononuclear cell preparation and flow immunophenotyping
Each participant provided 27 mL of peripheral venous blood collected into 9 mL sodium citrate tube.

Peripheral blood mononuclear cells (PBMCs) were isolated using Leucosep tubes (Greiner Bio-One)
with Histopaque 1077 (Sigma) by centrifugation at 800x g for 15 minutes at room temperature.

Reticulocyte counts
Reticulocyte numbers were measured using a Sysmex XN-1000 haematology analyser according to manufacturer instruction and as previously described (Akbari et al., 2020) Aldrich)) at 37ºC in a humidified CO 2 atmosphere for 48 hours. The cells and medium were decanted from the plate and the assay developed following the manufacturer's instructions. Developed plates J o u r n a l P r e -p r o o f 38 were read using an AID iSpot reader (Oxford Biosystems, Oxford, UK) and counted using AID EliSpot v7 software (Autoimmun Diagnostika GmbH, Strasberg, Germany). All data were then corrected for background cytokine production and expressed as SFU/Million CD3 T cells.

SARS-CoV-2 serology
Quantification of Spike SARS-CoV-2 specific antibodies was performed by ELISA as described by Xiong X et al . Briefly, serum samples collected at time of enrolment in the study and at the 4-8 week follow-up visit were first screened for positivity and then antibody titres were determined by an end-point analysis. AUC values were calculated in R (3.6.3) using the flux (0.3-0) package. Kruskal-Wallis test was used to calculate p-values among the different disease severities.

SARS-CoV-2 neutralisation assay
The virus used in this study was the clinical isolate SARS-CoV-2/human/Liverpool/REMRQ0001/2020, a kind gift from Ian Goodfellow (University of Cambridge), isolated by Lance Turtle (University of Liverpool) and David Matthews and Andrew Davidson (University of Bristol) (Daly et al., 2020;Patterson et al., 2020).

In brief, HEK293T reporter cells expressing Renilla luciferase (Rluc) and SARS-CoV
All downstream data handling was performed in R (R Core Team, 2015). Counts were filtered using filterByExpr (EdgeR) with a gene count threshold of 10 CPM and the minimum number of samples set at the size of the smallest disease group. Library counts were normalised using calcNormFactors NovaSeq 6000 (Illumina) using S1 flowcells. Droplet libraries were processed using Cellranger v4.0.
Reads were aligned to the GRCh38 human genome concatenated to the SARS-Cov-2 genome (NCBI SARS-CoV-2 isolate Wuhan-Hu-1) using STAR (Dobin et al., 2013) and unique molecular identifiers (UMIs) deduplicated. CITE-seq UMIs were counted for GEX and ADT libraries simultaneously to generate feature X droplet UMI count matrices.

QUANTIFICATION AND STATISTICAL ANALYSIS
All statistical analyses were conducted using custom scripts in R (R Core Team, 2015). Absolute cell counts (cells/uL) were offset by +1 to allow subsequent log2 transformation of zero counts. Where shown, time measures represent time from symptom onset (for severity groups B, C, D and E) or first positive COVID-19 swab (group A). Unless otherwise specified, longitudinally collected data was grouped by bins of 7 or 12 days. Pairwise statistical comparison of absolute cell counts, CRP or serum measures between individuals in a given severity group at a given time bin and HCs, or between severity groups, was conducted by Wilcoxon test unless otherwise specified. For analyses involving repeated measures, false discovery rate corrected (Benjamini & Hochberg) p value were reported.
For individuals sampled more than once within a given time bin, data from the earliest blood collection was used.
J o u r n a l P r e -p r o o f 41 Cell subset deconvolution of the whole blood RNA-Seq dataset was performed using pathway-level information extractor (PLIER) (http://gobie.csb.pitt.edu/PLIER). Latent factors were generated by leveraging off pre-existing knowledge of cell specific pathways. To better understand the relationship between gene expression and clinical severity, weighted gene co-expression network analysis was carried out using the WGCNA package (Langfelder and Horvath, 2008) in R. Briefly, a signed adjacency matrix was generated and a soft thresholding power was chosen to impose approximate scale-free topology. Modules were identified from the resulting topological overlap matrix with a specified minimum module size of n = 30. Modules were summarized using singular value decomposition, and the resulting module eigengene correlated with clinical traits. Significance of the correlation between a given clinical trait and a modular eigengene was assessed using linear regression with Bonferroni adjustment to correct for multiple testing. Modules were annotated using Enrichr (Chen et al., 2013). Longitudinal mixed modelling of gene module changes over time i.e., using the lme formula: module_eigenvalue ~ (time + I(time^2)) * category, random = ~ 1|subject.
Gene set enrichment analysis (GSEA) (Subramanian et al., 2005)  Principal component analysis (PCA) of centred and scaled absolute counts for 24 major cell types was conducted using the pca()function from the package mixOmics (Rohart et al., 2017).
Unsupervised clustering of log2 transformed absolute cell counts, normalised to the median of the corresponding control population, was conducted using the heatmap.2() function from the package gplots (Gregory R. Warnes et al., 2020), with a Euclidean distance function applied to both rows and columns of the data matrix and hierarchical clustering computed using the ward.D method. Partial least squares discriminant analysis (PLS-DA) was conducted using the plsda() function from the package mixOmics (Rohart et al., 2017), a supervised method of sample discrimination whereby sample clustering is informed by group membership (here patient clusters 1 and 2). The classification performance of the PLS-DA model was determined using the perf() function via 10 iterations of 5fold cross-validation, with two components deemed sufficient to minimise the balanced error rate of prediction. Variable selection on components 1 and 2 was conducted using the tune() function, with 13 cell types selected as those most strongly contributing to discrimination of patient clusters. An AUROC curve showing the performance of a predictive model based on these 13 cell types was generated using the auroc() function. To assess whether clinical severity was reflected on a transcriptional level in an unsupervised fashion, K-means clustering was utilised on normalised whole blood RNASeq gene expression counts. Heat maps were created using the ComplexHeatmap package (Gu et al., 2016), with data scaled and centred prior to visualisation.
J o u r n a l P r e -p r o o f 43 Cellular recovery rates over 60 days were calculated for each cell type in patients from groups C, D and E, split into those with persistently elevated (>10mg/L) or resolving CRP (falling below 10mg/L by final bleed), over 60 and 40 days respectively. Using a 12 day sliding window with single day increments, the 'window of recovery' for each given cell population was defined as the window in which absolute cell counts for COVID-19 samples no longer differed from controls when assessed by Wilcoxon test, and remained as such for the subsequent 7 windows, and 80% of all windows remaining. Recovery rate was taken as the log2 normalised ratio of test and control absolute counts for patient samples collected within the first time window (0-12 days), subtracted from the equivalent value calculated within the window of recovery, divided by the upper day boundary of the recovery window.

Highlights
• Longitudinal analysis of COVID-19 patients with a range of disease severity • Early bystander CD8 + T cell and plasmablast responses characterize mild disease • Pronounced systemic inflammation evident at first presentation in more severe COVID-19 • Immune/inflammatory abnormalities persist in severe disease to 60 days post-symptoms