Samples of patients infected with SARS-CoV-2 and healthy controls (HCs) were collected between March 19^th and June 7^th, 2020 in the University Hospitals Leuven as part of the CONTAGIOUS consortium. SARS-CoV-2 samples were classified in 36 WARD patients and 61 ICU patients, of which we collected samples at different time points during disease. Follow-up samples were collected from 8 WARD patients and 20 ICU patients, six weeks after discharge from the hospital (post-COVID-19 patients). COVID-19 was defined as a positive qRT-PCR (n=87) on a respiratory sample and/or CT imaging highly suggestive of COVID-19 (n=8, no qRT-PCR was performed) compatible with COVID-19. Control samples were derived from hospital staff with negative COVID-19 serology (n=18). The demographic and disease characteristics are listed in Supplementary Table 1 . Patients comorbidities, non-COVID-related treatments and ethnicity are shown in Supplementary Tables 2–4 . Demographic, clinical, treatment, and outcome data were obtained from the patient electronic medical records stored in the Research Electronic Data Capture Software REDCAP, Vanderbilt University.

Blood samples were collected in EDTA tubes, stored at 15°C, and processed within 3-6 hours. Separation of plasma and PBMCs was performed using a lymphocyte separation medium (LSM, MP Biomedicals), after which PBMCs were frozen in 10% dimethyl sulfoxide (Sigma) and stored in liquid nitrogen, while plasma was kept at -80°C until processing. Due to ethical constraints, a limited amount of blood was collected. Considering some patients had low blood cell counts, we have not always been able to perform a full flow cytometric analysis or cytotoxicity assay on all patient samples. In each analysis, the number of samples is indicated in the figure legends.

Frozen PBMCs were thawed, washed with medium, and used for staining. Cells were incubated with FcR-block (Miltenyi Biotec) and stained with live/dead marker (Fixable Viability Stain 620, BD Biosciences). Next, cells were extracellularly stained with fluorochrome-conjugated antibodies against surface markers ( Supplementary Table 5 ) and fixed with 4% paraformaldehyde. Staining for intracellular protein was performed with the Cytofix/Cytoperm kit (BD biosciences) according to the manufacturer’s instructions. Flow cytometry was performed on the BD LSR Fortessa X20, and data were analysed with FlowJo (LLC, V10). A representative gating strategy is shown in Supplementary Figure 1 . NK-platelet aggregate complexes were studied using an ImageStream Imaging Flow Cytometer (Amnis Corporation). Flow cytometric data are presented in the three groups of patients (WARD, ICU and post-COVID-19) along with HCs. In addition, WARD and ICU patients were grouped according to the days post-onset of COVID-19 symptoms, i.e. day 0-10 (group 1), day 11-20 (group 2), day 21-30 (group 3), and >30 days (group 4).

Levels of IFN-γ and TNF-α in NK cells were determined either ex vivo or upon in vitro stimulation with IL-12+IL-18. PBMCs used for ex vivo staining of cytokines were plated for 2 hours (without any in vitro stimulation) in complete RPMI containing Monensin (BD GolgiStop, 1:1500 dilution) and Brefeldin (BD GolgiPlug, 1:1000) at 37°C with 5% CO₂. To measure the production of IFN-γ and TNF-α upon in vitro stimulation, PBMCs were cultured in medium alone or in presence of IL-12 (2 ng/ml, PeproTech) plus IL-18 (100 ng/ml, MBL) for 18 hours (with GolgiStop and GolgiPlug added during the last 4 hours of stimulation). Intracellular expression of IFN-γ and TNF-α, along with extracellular NK cell markers, was analysed by flow cytometry as described above.

scRNA-seq was performed on PBMCs from 26 COVID-19 patients. 14 WARD patients and 12 ICU patients (all upon admission except for one at discharge) were included. Single-cell suspensions were converted to barcoded scRNA-seq libraries by using the Chromium Single Cell 5’ library and Gel Bead & Multiplex Kit from 10x Genomics. Captured cells were lysed and the released RNA was barcoded. Libraries were sequenced on an Illumina NovaSeq 6000 sequencer according to a paired-end reading strategy. Raw gene expression matrices were generated per sample by the Cell Ranger pipeline using the human reference version GRCh38. Additionally, we processed scRNA-seq data from Wilk et al. (17), including 6 HCs, 4 severe, and 4 moderate COVID patients. All datasets were analysed by the Seurat package and merged using Harmony. The R package Slingshot was used to explore pseudotime trajectories (34).

Chemokine and cytokine levels in plasma were assessed by Meso Scale Discovery (MSD) using the V-plex human cytokine 30-plex kit (Eotaxin, Eotaxin-3, GM-CSF, IFN-γ, IL-1α, IL-1β, IL-2, IL-4, IL-5, IL-6, IL-7, IL-8, IL-8 (HA), IL-10, IL-12/IL-23p40, IL-12p70, IL-13, IL-15, IL-16, IL-17A, IP-10 (CXCL10), MCP-1 (CCL2), MCP-4 (CCL13), MDC (CCL22), MIP-1α (CCL3), MIP-1β (CCL4), TARC (CCL17), TNF-α, TNF-β, VEGF-A), complemented with Human IL-1RA (V-plex), Human IL-18 (U-plex), and Human CXCL9 (R-plex) kits.

Purified NK cells from PBMCs of COVID-19 patients were used in a chromium release assay to assess cytotoxicity. Briefly, NK cells from HCs and COVID-19 patients were enriched from PBMCs by magnetic negative selection using the human NK cell isolation kit (Stemcell Technologies), according to the manufacturer’s protocol. The purity of enriched cells was between 85-95% (data not shown). NK cells were stimulated overnight with IL-15 (20 ng/ml, PeproTech) before using them in the assay. NK cells were plated with Raji cells loaded with chromium 51 (⁵¹Cr) (Perkin Elmer) at effector:target (E:T) ratio’s of 3:1, 1:1, 0.3:1, and 0.1:1, respectively, in V-bottomed 96-well plates for 4h at 37°C. To determine ADCC, cells were incubated with rituximab (RTX), a monoclonal antibody that binds to CD20. All conditions were plated in duplicate. After incubation, the cell culture supernatant was transferred to a LumaPlate (Perkin Elmer) coated with a solid scintillator. The release of ⁵¹Cr was measured using the microplate scintillation counter (Perkin Elmer, Microbeta, 2450 Microplate counter). The percentage of specific lysis was calculated as [(experimental release – spontaneous release)/(maximal release – spontaneous release)] x 100. The maximal release was determined by adding a mixture of medium and soap to the target cells, while the spontaneous release was measured by adding medium alone without effector cells.

Statistical analysis and plots were performed using R. Boxplots represent the median and 25^th to 75^th percentiles, and the whiskers denote the lowest and highest values. For the comparison of groups, P-values were obtained using Wilcoxon signed-rank tests, Bonferroni corrected for multiple comparisons. * Represent differences between indicated groups, # represent differences with HCs. */#p<0.05; **/##p<0.01; ***/###p<0.001; ****/####p<0.0001. For correlation statistics, Pearson correlation was applied.

All procedures were approved by the UZ Leuven Ethical Committee (protocol study number S63881). Informed consent was obtained from all individuals or their legal guardians according to the Declaration of Helsinki.

95 patients with COVID-19 were prospectively recruited after admission to the University Hospital Leuven, as part of the CONTAGIOUS trial (NCT04327570). Demographics and laboratory results at the time of blood sampling of COVID-19 patients are summarised in Supplementary Tables 1–4 . As schematically presented in Figure 1A , recruited patients consisted of COVID-19 patients hospitalised at the WARD (n = 36) or ICU (n = 61). WARD patients had a mild to moderate clinical condition either receiving no respiratory support or oxygen via nasal cannula. ICU patients had a critical clinical condition receiving high flow oxygen support or mechanical ventilation. Blood samples were also collected from COVID-19 patients six weeks after discharge from the WARD (n = 8) or ICU (n = 20). These are referred to as post-COVID-19 patients. To study NK cells, PBMCs were isolated and analysed either by flow cytometry or scRNA-seq or were used to purify for NK cells to study their cytolytic activity ( Figure 1A ). In the flow cytometric analysis and NK cell activity assays, PBMCs from 18 HCs, with no prior diagnosis or symptoms of COVID-19 (see M&M), were included. In plasma from patients and HCs, cytokines were analysed by MSD multiplex. For each patient, a detailed timeline was generated displaying the onset of symptoms, time of hospital admission and discharge, time points of blood sampling, and glucocorticoids (CS) treatment ( Figure 1B ). Figure 1C additionally summarises gender, BMI, and provides details on respiratory levels, and additional treatment.

NK cells were investigated by flow cytometry in our cohort of active and post-COVID-19 patients and further compared to HCs. Considering the differences in timing relative to the onset of COVID-19 symptoms in WARD and ICU patients, data are also presented after grouping blood samples from hospitalised patients (WARD + ICU) according to the days post onset of COVID-19 symptoms, i.e. day 0-10, day 11-20, day 21-30 and >30 days. SARS-CoV-2 infection is known to be associated with a dramatic drop in almost all blood leukocytes (except for neutrophils and plasmablast cells), as we and others previously reported (17, 35, 36). However, when the relative percentages of NK cells within total PBMCs were plotted, we found significantly increased frequencies of NK cells in WARD patients compared to HCs and ICU patients ( Figure 2A upper panels). When NK cells were divided into CD56^bright and CD56^dim subsets, the CD56^dim subset was increased in WARD as well as in ICU patients, while consequently CD56^bright NK cells were decreased. Grouping of data according to days post onset of COVID-19 symptoms showed an increase in frequency of NK cells during the early phase of COVID-19 disease more specifically in samples taken from patients of 0-10 and 11-20 days post onset symptoms and revealed a significant decrease in NK cell numbers in blood of patients taken at >30 days post symptoms as compared to HCs ( Figure 2A lower panels). NK cell numbers remained significantly lower in post-COVID-19 patients ( Figure 2A upper panels).

NK cell-mediated activities are tightly regulated by a balance of activating and inhibitory receptors to kill infected and malignant cells while leaving healthy cells intact. When we analysed the expression of activating (NKp30, NKp46, NKG2D, and activating KIRs) and inhibitory receptors (inhibitory KIRs and NKG2A) on total NK cells, we observed an increased frequency of NKp30+ and NKG2D+ NK cells in WARD patients ( Figure 2B upper panels, MFI for NKp30 was significantly increased as shown in Supplementary Figure 2 ). In ICU patients, expansion of NKp30+ and NKG2D+ NK cells was not seen and levels were significantly lower than in WARD patients. NKp46 was decreased in ICU patients and partially restored in post-COVID patients. No differences could be detected in activating KIR and inhibitory KIR receptor expression. Analysis of the inhibitory NKG2A receptor showed a decreased percentage of NKG2A+ NK cells both in WARD and ICU patients, which was restored in the post-COVID group ( Figure 2B upper panels, MFI of NKG2A was significantly lower in WARD patients only, Supplementary Figure 2 ). When days post onset of symptoms were taken into account, substantial differences between the four groups of patients were seen for NKp30 and NKG2D-expressing NK cells and were in general increased during early time point post onset of COVID-19 disease and decreased at later time points ( Figure 2B lower panels). The decrease in frequency of NKG2A expressing NK cells was most evident during the early phase of COVID-19 disease. The same pattern of expression (increased frequency of NKp30+ or NKG2D+ NK cells in WARD and decreased percentage of NKG2A+ NK cells in WARD and ICU) was seen in both CD56^bright (NKG2D only) and CD56^dim NK cell subsets ( Supplementary Figure 3 ).

In summary, the balance in activating and inhibitory NK cell receptors is shifted toward activation in WARD patients and in samples taken during the first 10 days post COVID-19 symptoms, but not in the other hospitalised patient groups. In post-COVID-19 patients, this balance is restored.

In order to kill target cells, NK cells must upregulate cytotoxic molecules which we investigated by intracellular flow cytometric analysis. We found a significantly increased expression (% and/or MFI) of perforin (PRF), granzyme A (GZMA), granzyme B (GZMB), and granzyme K (GZMK) in NK cells of COVID-19 WARD patients, whereas these were not different in ICU patients compared to HCs, except for GZMA that was decreased and GWMK that was increased ( Figure 3A upper panels). When data are plotted according to the days of symptoms, we observed increased expression of PRF, GZMA and GZMB at early time points of COVID-19 disease (day 0-10 and day 11-20 post onset of symptoms) as compared to HCs, whereas decreased expression in samples taken at later time point of COVID-19 disease (day 21-30 and >30 days post onset of symptoms) ( Figure 3A middle panels and Supplementary Figure 4A ). PRF and GZMB expression was significantly decreased in post-COVID patients. The profile of these cytotoxic molecules was seen in both CD56^dim and CD56^bright NK cell subsets wherein, as expected, levels were highest in CD56^dim NK cells ( Supplementary Figures 4B, C ). GZMK showed a different expression pattern as its expression was predominantly increased in ICU patients ( Figure 3A , upper panel), a finding that was mainly due to increased GZMK levels in CD56^dim NK cells ( Supplementary Figures 4B, C ). Even more strikingly, post-COVID-19 patients were seen to have increased GZMK expression compared to HCs, WARD, and ICU patients ( Figure 3A upper panel).

To study if the differences in cytotoxic effector molecule expression in NK cells of WARD vs ICU patients is related to disease duration, we performed a Spearman correlation analysis. This revealed a negative correlation of PRF, GZMA, and GZMB expression (and a positive correlation for GZMK) with disease duration ( Figure 3A lower panels, line graphs).

In addition to cytotoxic molecules, NK cells may also produce cytokines. To detect intracellular cytokine production by flow cytometry, NK cells require in vitro stimulation with either cytokines or target cells. However, considering the inflammatory nature of the disease, we sought to analyse direct IFN-γ and TNF-α production, without prior stimulation. Importantly, we found expression of IFN-γ and TNF-α in NK cells of most of COVID-19 patients, and some of the post-COVID-19 patients. The frequency of cytokine-producing NK cells as well as the levels of IFN-γ and TNF-α were significantly higher in ICU than in WARD patients ( Figure 3B , MFI data in Supplementary Figure 5A upper panels). Levels of IFN-γ in NK cells significantly increased in samples taken at later time points post-onset of symptoms whereas levels of TNF-α remained high amongst the 3 groups of patients when sampling was done at >10 days post onset of COVID-19 symptoms (MFI and percentages in Supplementary Figures 5A, B lower panels). Supplementary Figures 5C, D shows that in addition to CD56^bright NK cells, CD56^dim subset are also important cytokine-producing cells in COVID-19 patients. Furthermore, as shown in a representative flow cytometric staining ( Figure 3C ), a large proportion of NK cells of ICU patients were double-positive for IFN-γ and TNF-α.

To investigate the correlation of cytotoxic moledules and cytokines with NK cell differentiation and activation, cells were stained with an additional antibody panel including CD107a, CD69, CD49a, CD16, CD57, and HLA-DR. The expression of the tested differentiation and activation markers on NK cells of HCs versus COVID-19 patients in different stages of their disease are shown in Figure 3D and Supplementary Figures 5A, B and the correlation of these markers with cytotoxic molecules or with IFN-γ and TNF-α are shown in Figure 3E . Expression of CD107a was significantly increased in NK cells from WARD patients, and even higher CD107a levels were noted in ICU patients ( Figure 3D ). Interestingly, both IFN-γ, TNF-α, and CD107a expression showed a positive correlation with the activation marker CD69 and the adhesion molecule CD49a (red box in Figure 3E ), and negatively correlated with CD16, CD57, GZMA, and/or GZMB (green box in Figure 3E ). Figure 3F plots a representative flow cytometric image of the CD16 and CD49a expression on NK cells of HCs, WARD, and ICU patients. It shows the appearance of a CD49a⁺ NK cell population in COVID patients and a general low expression of CD16 in NK cells from ICU patients. HLA-DR is upregulated on activated NK cells and HLA-DR+ cells are characterised by a more intensive cytokine-induced IFN-γ expression (37). In this study, HLA-DR expression was significantly upregulated in ICU and post-COVID-19 patients ( Figure 3D ). Note that HLA-DR was positively correlating with CD49a, CD69, and GZMK, whereas inversely with GZMA, CD16, CD57, and PRF ( Figure 3E ).

When plasma levels of cytokines and chemokines were analysed using a 30-plex kit, several of these (IFN-γ, TNF-α, IL-18, IL-7, IL-10, IL-15, CXCL8, CXCL10, CCL2, CCL3, and CCL4) were significantly increased in all COVID-19 patient groups as compared to HCs ( Figure 4 ). Levels were most pronounced in samples taken at early time points post onset of COVID-19 symptoms with the exception of TNF-α and IL-18 which remained high in all hospitalised patients or even increased (IL-18) in samples taken at day 21-30 post symptoms ( Figure 4 ). In post-COVID-19 patients, some plasma cytokines (i.e. TNF-α, IL-18, CXCL8, CXCL10, CCL2, CCL3, and CCL4) remained higher than in HCs, but levels were much lower as compared to hospitalised patients. Interestingly, cytokines and chemokines were found to correlate with CD69 expression on NK cells and to some extent to CD107a ( Supplementary Figure 6A ). All these cytokines and chemokines are associated with NK cell biology, either as an activating factor for – or as a product of – NK cells. Of special interest is the correlation of CD49a, CD69, and CD107a expression (corresponding with the cytokine-producing NK cells, Figure 3E ) with plasma IL-15, a key cytokine in NK cell development and activation (38) that was previously linked to an attenuated NK cell inflammation signature in fatal COVID-19 (27). In our study, the emergence of this CD49a⁺CD69⁺CD107a⁺ NK cell subset was also correlated to the number of neutrophils, which is a surrogate disease severity marker ( Supplementary Figure 6B ) (35).

In summary, while NK cells from COVID-19 WARD patients have increased expression of cytotoxic molecules, NK cells from ICU patients have higher CD107a expression and have greatly increased production of cytokines in the absence of deliberated in vitro simulation. Throughout the progression of COVID-19, the cytotoxic effector molecules in NK cells significantly decrease whereas those of IFN-γ and TNF-α increase. While cytokine-producing NK cells are still present in the circulation of post-COVID-19 patients, their frequencies were decreased.

The elevated cytokine levels of IFN-γ and TNF-α in NK cells of COVID-19 patients as we found by our ex vivo analysis, contradict with the reported decreased cytokine production when cells of COVID-19 patients are stimulated in vitro with either cytokines or tumour target cells (31, 33). To understand this discrepancy, we stimulated PBMCs from part of our patient cohort for 18h with IL-12+IL-18, and subsequently analysed NK cell-produced IFN-γ and TNF-α by intracellular flow cytometry. As shown in Supplementary Figure 7 , NK cells of COVID-19 ICU patients failed to produce IFN-γ upon stimulation with IL-12+IL-18 while those of WARD patients are diminished in their production as compared to the HCs. TNF-α levels in NK cells presented no differences between HCs and COVID-19 patients.

To investigate the broader cytotoxic capacity of NK cells from COVID-19 patients against tumour target cells and their capacity to perform antibody cell-mediated cytotoxicity (ADCC), we performed ⁵¹chromium-release assays. We used Raji B cells (derived from Burkitt’s lymphoma) as NK cell targets because they express CD20, which provides a way to additionally analyse ADCC of NK cells by adding rituximab (i.e. anti-CD20 antibody) to the assay. Direct cytolysis of Raji cells by NK cells was measured in the absence of rituximab. The cytolytic assay is schematically presented in Figure 5A and was performed with NK cell-enriched samples from 4 WARD patients and 26 ICU patients. 7 HC samples were included. The percentage of target cell lysis was plotted for four different effector:target (E:T) ratios and averages for ratio 3:1 are depicted in a bar graphic ( Figures 5B, C ). NK cells from WARD and ICU patients were capable to kill Raji target cells equally well as HCs, both in the absence or presence of rituximab. NK cells from 7 out of 26 ICU patients were poorly cytolytic as seen by a low level of direct cell lysis. In the presence of rituximab, the complete defect in cytolysis remained in 4 of these 7 ICU patients.

To explain the weak cell-mediated cytotoxicity in these specific ICU patients, we examined whether their NK cells showed any phenotypic abnormalities. While we could not find a link with any of the NK cell-activating and inhibitory receptors, nor with the expression of PRF, GZMA, and GZMB, we noticed a positive correlation between target cell lysis and the expression of CD57 (R=0.47, p=0.02) by NK cells. Additionally, a negative correlation between target cell lysis and the expression of CD107a (R=-0.43, p=0.02) by NK cells was found ( Figure 5D ). Remark that the defective NK cells were not restricted to blood samples taken at the late phase of COVID-19 disease ( Figure 5E ).

Thus, while 4 ICU patients failed to kill target cells either directly or via ADCC, the majority of COVID-19 patients we analysed in this study displayed normal NK cell cytotoxic activity against Raji target cells.

We previously described heterogeneity in PBMCs isolated from COVID-19 WARD (n=14) and ICU (n=12) patients, based on scRNA-seq (36). NK cells formed a separate cluster, representing 7.9% and 7.4% of total PBMCs in WARD and ICU patients respectively, and were identified by expression of marker genes NCR1 (NKp46), NCAM1 (CD56), KLRB1 (CD161), NCR3 (NKp30), KLRD1 (CD94), KLRF1 (NKp80), FCGR3A (CD16), natural killer cell granule protein 7 (NKG7) and the TYRO binding protein (TYROBP) (36). Here, we additionally dissected the NK cells, exploring NK cell heterogeneity. To compare NK cells from COVID patients with HCs, we merged our single-cell data with the PBMC dataset from Wilk et al. (17), adding 6 HCs, 4 WARD, and 4 ICU patients (using similar clinical criteria as in our cohort).

In our analysis, unsupervised clustering identified 10 subclusters (C0 – C9; Figures 6A, B and Supplementary Figure 8 ). These comprised of CD56^dim mature NK cells (C0), CD56^dim immature NK cells (C1), CD56^dim adaptive NK cells (C2), CD56^bright NK cells (C3), and NK-platelets (C4). The other remaining clusters (C5-9) were defined as NK-T cells (C5), proliferating NK cells (C6 and C8), monocytes (C7), and B cells (C9) and are described in the Supplementary M&M . Cells from the two merged datasets ( Supplementary Figure 9A ) and cells from the different patients and HCs clustered together (individual patient stratification is shown in Supplementary Figure 9B ).

The maturational state of the different NK cell-related clusters (C0-4) is presented in Figure 6C . C3 NK cells were the least mature and were annotated as CD56^bright NK cells based on the expression of NCAM1 (CD56), XCL1, XCL2, SELL (CD62L), CD44, granzyme K (GZMK), high expression of KLRC1 (NKG2A), and low expression of FCGR3A (CD16) ( Figure 6D ). A comparison with the gene expression published by Collins et al. (39) further supported the identity of the CD56^bright NK cells ( Figure 6D and Supplementary Figure 10A ). C3 cells represented 7.9% of the NK cells in HCs and were decreased in WARD (5.4%) and ICU (4.7%) patients, albeit non-significantly ( Figure 6B ).

Based on the maturation score, we noticed a gradual maturation from the CD56^bright NK cells (C3), into CD56^dim mature NK cells (C0) and CD56^dim adaptive NK cells (C2) via a CD56^dim immature NK cell subcluster (C1). The NK-platelets (C4) mainly showed a mature phenotype. We used pseudotime trajectory to confirm the differentiation trajectory of the 4 main NK cells populations (C0-C3) in which we defined C3 CD56^bright NK cells as the root of the trajectory ( Supplementary Figure 9C ).

The C0 CD56^dim mature NK cells ( Figure 6C ) display a low expression of NCAM1 (CD56), a high expression of lytic granule genes such as perforin (PRF1), granzyme A and B (GZMA, GZMB), and a high expression of the cytolytic regulator CST7, which is needed for optimal cytotoxicity (40) ( Figure 6D ). The percentage of C0 NK cells was non-significantly increased in WARD and ICU patients (58.8% and 60.8% respectively) compared to HCs (52.3%). Compared to C0, C1 mainly showed a lower expression of cytotoxic markers including GZMA, GZMB, GZMH, and S100A4 ( Figure 6C ). The C1 subset represented 20.8% of the total NK cells in HCs and was significantly decreased in WARD and ICU patients respectively (10.8% and 5.0% of the total NK cells respectively), demonstrating a shift towards functionally more mature and active NK cells ( Figure 6B ).

Maucourant et al. described adaptive NK cells in COVID-19 patients (26). In our study, cluster C2 represents CD56^dim cells with an adaptive NK cell phenotype. This was based on the gene signature described by Schlums et al. (41), showing low expression of KLRC1 (NKG2A), ZBTB16 (PLZF), FCER1G (FCϵR1γ), and high expression of KLRC2/3/4 (NKG2C/E/F), Granzyme H (GZMH), CCL5, CD3E, and IL-32 ( Figure 6D and Supplementary Figure 10B ). In HCs, 8.0% of all NK cells were annotated as adaptive NK cells, and percentages increased in WARD patients (13.0%), albeit non-significantly. The percentage of C2 subset NK cells was unaltered in ICU patients (7.9%) ( Figure 6B ).

Intriguingly, C4 NK cells were referred to as an NK cell-platelet subset since these cells expressed platelet-related genes PPBP, CLU, F13A1, NRGN, OAZ1, PF4, SELP, SPARC, TAGLN2, amongst others, along with NK cell markers ( Figures 6D, E , and Supplementary Figure 8 ). The NK cell-platelet subset was COVID-19-specific since they were barely present in HCs (0.9%) but significantly increased in WARD (3.5%) and ICU patients (on average 12.1% and with individual percentages up to 34.9%) ( Figure 6B ).

Supplementary Figure 11 describes the activational state of the different NK cell subsets in HCs, WARD and ICU COVID-19 patients. Based on the expression of cytotoxic molecules and cytokines, increased activation was seen in patients ( Supplementary Figures 11A, B ). Interestingly, in agreement with our flow-cytometric analysis, the cytokine expression in both the CD56^dim and CD56^bright NK cells was higher in patients compared HCs. ( Supplementary Figure 11B ). Of note, no exhausted phenotype was found since exhaustion markers were rather decreased in patients when compared to HCs ( Supplementary Figure 11C ).

Taken together, SARS-CoV-2 infection triggers a redistribution of NK cell subsets towards a more mature and activated phenotype. Unexpectedly severe COVID was associated with the emergence of a – so far undescribed – NK cell-platelet subset.

Our single-cell transcriptome analysis revealed a COVID-19 specific cell population in ICU patients with a gene signature of both NK cells and platelets (i.e. cluster 4, Figures 6B, D, E ). We hypothesised that this subset results from platelets strongly adhering to NK cells and therefore performed flow cytometry on PBMCs that were stained both for NK cell markers (CD56 and NKp46) and markers of platelets (CD42a and CD62P, which are expressed on the surface of all platelets and activated platelets, respectively) (42, 43). As shown in a representative dot plot in Figure 7A , a fraction of NK cells from ICU patients stained positive for the CD42a platelet marker and part of these also expressed externalised CD62P, indicating degranulation by activated platelets. By imaging cytometry, we confirmed the adherence of platelets on NK cells of ICU patients. These NK cell-platelet complexes were seen in all investigated ICU patients (n= 18) but not in HCs ( Figures 7B, C ). While on average 12% of total NK cells from ICU patients were covered with platelets, these levels increased in some patients to more than 40%. NK-platelet aggregates showed reduced CD16 expression ( Figure 7D ). NK cell-platelet aggregates were not a freezing artifact as these were also depicted in freshly-isolated samples (coloured data points in Figure 7C ). We found no correlations between the frequency of NK-platelet aggregates and the amount of cytotoxicity, which may be due to the small number of samples on which both assays were performed (n = 16).

These data clearly show that COVID-19 patients are represented with an NK cell-platelet cluster that is partially activated and is absent in HCs. Furthermore, NK-platelet doublets have decreased CD16 expression.