The role of NK cells during SARS-CoV-2 infection remains unknown. We wondered whether we could find differences in the decline in SARS-CoV-2 load between patients admitted to hospital with normal (>40 NK cells per μl) or low (≤40 NK cells per μl) NK cell counts. There was an overall pattern of a faster decline in viral load in patients with normal NK cell counts (Fig. 1a, b) across patient groups with different overall clinical statuses (Extended Data Fig. 1a, Supplementary Table 3). Of note, such a negative correlation between NK cell counts (early during the infection) and viral load was not found for T cells or B cells (Extended Data Fig. 1a). Similarly, a more rapid decline in viral load over time was associated with a more rapid increase in NK cells and vice versa (Extended Data Fig. 1b). Overall, the availability of NK cells early during the course of COVID-19 correlated with a lower abundance of SARS-CoV-2 viral RNA.

Fig. 1: NK cells from patients with severe COVID-19 fail to control SARS-CoV-2 replication.
figure 1

a, b, Temporal viral load regression for patients with >40 NK cells per µl (a; n = 183) or ≤40 NK cells per µl (b; n = 23) at first count (Methods). The x-axis indicates time since the first measurement. The analysis included 206 patients in the intensive care unit (ICU) (Extended Data Fig. 1), of which 47 patients had severe COVID-19 (highlighted in red; Supplementary Table 3). c, d, Vero E6 cells (c) or Calu-3 cells (d) were infected with SARS-CoV-2 (B.1 lineage). At 1 h after infection, NK cells from healthy donors activated for 24 h in vitro as indicated with interleukins were added. Viral replication was measured 12 h later as genome equivalents (GE) per ml (GE ml–1; target cells co-cultured with NK cells versus cultured alone (c) or versus NK cell:target cell ratio 1:3 (d), n = 4 donors, no NK cells n = 5 samples). e, f, Vero E6 cells (e) or Calu-3 cells (f) were infected with SARS-CoV-2 and co-cultured as described above with NK cells from either healthy donors (n = 8 (e) or 6 (f)) or patients with severe COVID-19 (n = 6 (e) or 4 (f)). Viral load as the fold-change of Vero E6 cells cultured alone versus co-cultured with NK cells was determined (e; pooled data from two independent experiments). For f, a 1:1 NK cell:target cell ratio was used. g, Calu-3 cells were infected and co-cultured as above with IL-12- and IL-15-activated NK cells in a 3:1 NK cell:target cell ratio. Before co-culture, NK cells were incubated with the indicated neutralizing receptor antibodies. Each filled dot represents viral replication of target cells co-cultured with NK cells from an individual donor (2B4 n = 4, all others n = 6, no NK cells n = 8 samples, indicated group versus NK cells only). Graphs display mean ± s.e.m. Two-sided Mann–Whitney U-test (f, P = 0.038; g, P = 0.041 and P = 0.0043). *P ≤ 0.05, **P ≤ 0.01, ***P ≤ 0.001, ****P ≤ 0.0001.

We investigated whether NK cells can directly control SARS-CoV-2 (B.1 lineage) replication in an infected human lung epithelial cell line (Calu-3) or in kidney epithelial cells (Vero E6). At the time of infection, highly purified NK cells from healthy donors were added to infected Calu-3 and Vero E6 cells and the intracellular viral load was measured.NK cells reduced viral replication in a dose-dependent manner (Fig. 1c, d), a finding that was confirmed with a second virus variant (B.1.351 lineage; Extended Data Fig. 1c). NK cells are often activated during viral infections5, but NK cells isolated from patients admitted to hospital with COVID-19 were significantly less effective in reducing the viral load compared with NK cells from healthy donors (Fig. 1e, f).

NK cell recognition of virus-infected cells is determined by interactions between activating and inhibitory NK cell receptors and their ligands on target cells6. The large reduction in viral replication induced by NK cells could not be further enhanced by HLA blockade, which suggests that infected Calu-3 cells do not appreciably inhibit NK cells through HLA-I-specific inhibitory receptors (Fig. 1g). Uninfected Calu-3 cells were poor targets of NK cells (Extended Data Fig. 1d). Blockade of single, activating NK cell receptors did not impair virus control, whereas simultaneous blockade of all three natural cytotoxicity receptors (NKp30, NKp44 and NKp46) or of 2B4, NKG2D and DNAM-1 led to a significant increase in virus replication (Fig. 1g). Collectively, our data demonstrate that NK-cell-mediated control of SARS-CoV-2 replication in infected target cells requires redundant recognition by activating NK cell receptors. This process is impaired in infected cells treated with NK cells isolated from patients admitted to hospital with COVID-19.

Impaired NK cell function during COVID-19

We set out to study NK cell effector functions in detail in patients with COVID-19 across the disease spectrum and time (Fig. 2a). Patients with non-COVID-19 flu-like illness (FLI), ambulant patients with COVID-19 and patients with moderate COVID-19 disease severity had normal frequencies of CD56dim NK cells. By contrast, patients with a severe course of COVID-19 had reduced frequencies of both CD56dim and CD56bright NK cells and of innate lymphoid cells during the first weeks after symptom onset (Extended Data Fig. 2).

Fig. 2: Impaired NK-cell-mediated cytotoxicity during severe COVID-19.
figure 2

a, Overview of the study design. b, c, Mean fluorescence intensity (MFI, median) of perforin (b) and granzyme B (c) in CD56dim NK cells. Independent measurements from 44 (b) or 50 (c) healthy donors, 9 patients with FLI, and patients with COVID-19 (24 ambulant, n = 56; 17 moderate, n = 19; 30 severe, n = 73) obtained 1–8 weeks after the onset of symptoms. d, Specific lysis (mean ± s.e.m.) of K562 target cells by NK cells from the indicated donors obtained within the first 2 weeks after symptom onset was determined in a 51Chromium (51Cr) release assay. Data represent pooled data of 12 independent experiments using 18 healthy donors (n = 38 independent measurements), 8 patients with FLI (n = 8) and 28 patients with COVID-19 (7 ambulant, n = 7; 14 moderate, n = 15; 10 severe, n = 10). e, PBMCs from the indicated donor groups were co-cultured for 4 h with K562 cells and the percentage (median) of CD107a+ NK cells was measured. Patients receiving corticosteroid treatment were excluded from the analysis. Data depict independent measurements across the disease course (29 healthy, n = 29; 11 FLI, n = 11; 21 ambulant COVID-19, n = 28; 13 moderate COVID-19, n = 13; 16 severe COVID-19, n = 22). f, g, Quantification (f, mean) and representative flow cytometry plots (g) of conjugation of NK cells from healthy donors (n = 8) and from patients with severe COVID-19 (n = 5) with target cells. Numbers indicate the frequency in the percentage of NK cells conjugated to target cells out of all NK cells (pooled data from two independent experiments). For b, c and e, statistical analysis (indicated group versus healthy) was performed using one-way analysis of variance (ANOVA) followed by a two-sided Mann–Whitney U-test. The dashed line indicates median frequency or MFI of NK cells from healthy donors. For d and f, two-sided Mann–Whitney U-test was used (f, P = 0.03).

Previous data regarding the expression profile of cytotoxic molecules in COVID-19 were inconclusive3,4,7,8. We found a significant and early upregulation of perforin and granzyme B both in ambulant patients and in hospitalized patients with COVID-19 (Fig. 2b, c, Extended Data Fig. 3a–c), which is an early sign of NK cell activation that is observed in the context of various other viral infections5,9,10,11. NK cells isolated during the first two weeks after symptom onset from hospitalized and, to a lesser extent, from ambulant patients with COVID-19 showed impairments in cell-mediated cytotoxicity despite the high levels of perforin and granzyme B expression (Fig. 2d). Such a reduced cytotoxic activity of NK cells was not found in patients with FLI. Given the apparent paradox of high levels of cytotoxic molecule expression and low cytotoxic function, we analysed the release of cytotoxic granules. NK cells from healthy donors, patients with FLI and ambulant patients with COVID-19 did not show differences in degranulation. By contrast, NK cells from hospitalized patients with COVID-19 showed impaired degranulation (Fig. 2e, Extended Data Fig. 3d). One of the first steps during interactions between cognate NK cells and target cells is the formation of cellular conjugates, and the establishment of such conjugates was reduced for NK cells from patients with severe COVID-19 (Fig. 2f, g). Of note, reduced cell–cell interactions and degranulation were not a consequence of a reduced expression of activating NK cell receptors in severe COVID-19 (Extended Data Fig. 3e, f).

NK cells from ambulant patients with COVID-19 showed an increased production of interferon-γ (IFNγ), whereas NK cells from patients with severe COVID-19 produced only low levels of IFNγ and tumour necrosis factor (TNF)12 (Extended Data Fig. 3g, h). The T-box transcription factor T-bet coordinates NK cell effector programmes, including the expression of granzyme B, IFNγ and perforin13. T-bet was upregulated in NK cells from patients with FLI and its expression was maintained in ambulant patients with COVID-19. By contrast, T-bet was substantially suppressed at all time points in NK cells from hospitalized patients with COVID-19 (Extended Data Fig. 3a, i). These data cannot be easily explained by differences in age because we did not find strong negative correlations between age and any of the NK cell readouts (Extended Data Fig. 4). Reductions in NK-cell-mediated cytotoxicity and effector programmes were also not caused by dexamethasone treatment, as comparable data were obtained for samples taken during the first wave of COVID-19 (March to April 2020), a period when dexamethasone was not administered to patients (Fig. 2e).

Single-cell RNA sequencing atlas of NK cells

We used single-cell RNA sequencing (scRNA-seq) to generate a time-resolved, droplet-based single-cell transcriptomics atlas of peripheral blood NK cells from the following individuals: patients with severe COVID-19; outpatients with oligosymptomatic SARS-CoV-2 infection; and healthy donors (Fig. 3a). Using the gating strategy depicted in Extended Data Fig. 5a, 80,325 single NK cell transcriptomes were captured. Using uniform manifold approximation and projection for dimension reduction (UMAP), we identified seven transcriptionally distinct clusters of cells14 (Fig. 3a). A small number (1,375 cells) of contaminating non-NK cells was found in cluster 6; therefore this cluster was excluded from further analysis. Cells in the remaining clusters (0–5) expressed genes that define conventional NK cells, including surface markers (Extended Data Fig. 5b), effector molecules, such as PRF1, GNLY, GZMB, GZMH and GZMM, and TBX21 (which encodes T-bet)15,16 (Extended Data Fig. 5b–f). Cluster 2 represented CD56bright NK cells characterized by high expression of IL7R, SELL, XCL1, LTB and GZMK and low expression of CD160 (refs. 16,17) (Extended Data Fig. 5b, c). Cluster 0 contained cells with high expression of NK cell effector molecules and of TBX21, but with low expression of NCAM1, and we therefore identified these as a subset of CD56dim NK cells. Cluster 1 was closely related to cluster 0. Among the few differentially expressed genes were CD96 and KLRG1, both of which have been linked to NK cell maturation and functional exhaustion18,19. In addition, they showed reduced expression of most NK effector genes, thus they demarcate late effector NK cells that may be reduced in function (Extended Data Fig. 5d). Within CD56dim NK cells, a third cluster (cluster 3) was discriminated with low expression of effector molecules (GZMB, GZMH and GZMM; Extended Data Fig. 5d). This NK cell subset may correspond to the previously described 'terminally differentiated' CD56dim NK cells that are in a post-activation state16,20.

Fig. 3: A TGFβ response signature is a hallmark of NK cells in severe COVID-19 but not in severe influenza.
figure 3

ag, scRNA-seq of highly purified peripheral blood NK cells of 5 healthy donors (n = 5) and patients with COVID-19 from day 2 to day 68 after symptom onset (11 ambulant patients (n = 11) and 13 patients with severe COVID-19 (n = 52) (Supplementary Tables 13). The gating strategy is depicted in Extended Data Fig. 5a. a, UMAP representation of 80,325 sorted NK cells from all samples (n = 68). Colours indicate unsupervised clustering. b, Percentage (median) of cells allocated to each cluster in the indicated groups. Each dot represents one sample subjected to scRNA-seq. P values determined by one-way ANOVA followed by two-sided Mann–Whitney U-test. c, d, Pseudotime trajectories (c) and representation of the individual NK cell clusters in trajectories (d) of 11,613 randomly selected NK cell transcriptomes from all groups. eg, Single-cell GSEA of indicated gene sets was projected on the UMAP analysis (clusters 0, 1 and 3 only). Single cells with enriched gene expression are displayed as red dots (left), cells with depletion of the genes are displayed as blue dots (right). All sequenced NK cells per group are displayed as grey dots in the background. e, Hallmark IFNα response GSEA. f, g, Enrichment of TGFβ-induced genes (f, left) and depletion of TGFβ-suppressed genes (f, right) in the indicated groups and across the course of severe COVID-19 (g; early: ≤day 14 after symptom onset; intermediate: days 15–28, late: >day 28). h, NK cells were extracted from a publicly available single-cell dataset of PBMCs from healthy donors and from patients with COVID-19 or with severe influenza30 and a single-cell GSEA of the NK-cell-specific TGFβ response gene set was performed as described in e. GSEA P values were calculated using two-sided Fisher’s exact test comparing the indicated groups with the left-sided (in e, early was compared to healthy). *P value describes a reduction in enrichment or depletion.

Similar to CD56bright NK cells, cells in cluster 4 were characterized by high expression of IL7R and GZMK, but expressed low levels of XCL1 and SELL (Extended Data Fig. 5c) and corresponded to “transitional NK cells”16. Cluster 4 also contained cells that express the activating NK cell receptor KLRC2 (which encodes NKG2C; Extended Data Fig. 5e). Low expression levels of NKG2A (encoded by KLRC1) and of the signalling adaptor molecule FcεRγ (encoded by FCER1G) corroborated the enrichment of adaptive NK cells in cluster 4 (ref. 21). Thus, cluster 4 is heterogenous and represents transitional and adaptive NK cells. Cluster 5 represented proliferating NK cells identified by the high expression of MKI67 and of genes controlling the cell cycle (Extended Data Fig. 5e, g).

An analysis of the representation of each NK cell cluster across COVID-19 disease states revealed dynamics in NK cell differentiation. In severe COVID-19, we observed a significant increase in proliferating (cluster 5) NK cells (Fig. 3b), and this result was in line with flow cytometry data (Extended Data Fig. 5h). Although the frequency of cells belonging to cluster 1 (late effector) NK cells was reduced in patients with severe COVID-19, these patients had increased levels of terminally differentiated cluster 3 NK cells (Fig. 3b). Although not statistically significant, adaptive NK cells (cluster 4) were reproducibly increased in severe COVID-19, results that are in line with previous data3.

To investigate whether COVID-19 introduces changes in the dynamics of NK cell differentiation, we used monocle pseudotime trajectory analysis22, which connects related clusters to construct differentiation trajectories (Fig. 3c, d, Extended Data Fig. 6a–d). We defined cluster 2 CD56bright NK cells as the root of the progression trajectory. The pseudotime model predicted that CD56bright NK cells (cluster 2)—through a proliferative state in the left leaf (cluster 5)—differentiate into effector NK cells (cluster 0: CD56dim NK effector I cells). Next, by gradually decreasing effector molecule expression, the effector NK cells differentiate into cluster 1 CD56dim effector II cells until reaching the terminally differentiated state (cluster 3), which dominates the distal right leaf of the differentiation trajectory (Fig. 3c, d, Extended Data Fig. 6a, b). Along the pseudotime trajectory, major differentiation states were CD56bright NK cells (state 12: high expression of IL7RA and GZMK), proliferating NK cells (state 1: MKI67 high), effector NK cells (states 2–7; with graded levels of effector genes such as PRF1) and finally terminally differentiated NK cells (states 8 and 9) (Extended Data Fig. 6c, d). This differentiation trajectory was generally conserved in patients with COVID-19. However, patients with particularly severe COVID-19 had increased levels of proliferating NK cells that fed into one root of the differentiation trajectory and had an accumulation of terminally differentiated NK cells at the distal end of the trajectory (Extended Data Fig. 6a–d).

A TGFβ signature is a hallmark of severe COVID-19

We interrogated our dataset for enrichment of genes that may affect NK cell effector functions and applied gene set enrichment analysis (GSEA) to the effector NK cell clusters 0, 1 and 3. In patients with COVID-19, there was a significant enrichment of ‘Hallmark IFNα response genes’ (Fig. 3e) and ‘Hallmark IFNγ response genes’ (Extended Data Fig. 6e) predominantly in the NK effector cluster 0. This result was in line with increased virus-induced type I and type II interferon levels observed in the serum of patients with COVID-19 (Extended Data Fig. 6f). We also confirmed3 an increased expression of markers linked to NK cell activation and differentiation (for example, CD69 and CD57) in hospitalized patients with COVID-19 (Extended Data Fig. 6g–i). Among the enriched functional networks in NK cells isolated from patients with COVID-19 were several gene sets related to cellular metabolism and translation. Together, these results demonstrate that there are substantial changes in cellular activation in the context of SARS-CoV-2 infections (Extended Data Fig. 7a).

We next performed a Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analysis. We noted that among the transcriptional networks significantly enriched in effector NK cells from ambulant patients with COVID-19 and in particular from patients with severe COVID-19, was the pathway ‘intestinal immune network for IgA production’ (Extended Data Fig. 7b). IgA class switching is strictly controlled by TGFβ23,24, and many genes in this KEGG pathway are direct targets of TGFβ signalling, a cytokine that suppresses NK cell function25,26. Even terminally differentiated (cluster 3) NK cells, which were substantially depleted in gene sets regulated by IFNα or cellular metabolism (Extended Data Fig. 7a), were enriched for genes of this KEGG pathway (Extended Data Fig. 7b). Because the transcriptional changes instructed by TGFβ in NK cells are not known at single-cell resolution, we generated a scRNA-seq dataset of highly purified peripheral blood NK cells from healthy donors cultured in the presence or absence of TGFβ. UMAP analysis of the two treatment groups showed a large effect of TGFβ on the transcriptional state of NK cells that was highly reproducible across individual donors (Extended Data Fig. 8a). An analysis of differentially expressed genes revealed that the majority of genes were negatively regulated by TGFβ in their expression, including TBX21 and STAT1 and several effector genes such as PRF1 and granzyme family members (Extended Data Fig. 8b). The expression of a small group of genes was induced by exposure to TGFβ, including genes that encode the chemokine receptor CXCR4 or the transcription factor EOMES, which are both known TGFβ target genes (Extended Data Fig. 8b). Within the cells that were exposed to TGFβ (Extended Data Fig. 8c, red), we found that various pathways that are central to granule exocytosis27 were negatively regulated by TGFβ (Extended Data Fig. 8d).

Using this established TGFβ signature, we interrogated the scRNA-seq dataset from patients with COVID-19 for changes in the expression of TGFβ-controlled genes. TGFβ-controlled genes were expressed at a low level in NK cells from healthy donors, but a substantial enrichment in TGFβ response genes was noted in effector NK cell clusters from patients with severe COVID-19 (Fig. 3f). Of note, changes in the expression of TGFβ-controlled genes were already detectable during the very early stages of COVID-19 and increased during the course of infection (Fig. 3g). Indeed, the expression of canonical TGFβ target genes such as EOMES and ITGAE was upregulated in NK cells from patients with severe COVID-19 (Extended Data Fig. 8e, f). Our data demonstrate that there is a progressive and long-lasting reprogramming of NK cells by TGFβ during COVID-19 that starts within the first days after symptom onset and is commensurate with disease severity.

In ambulant patients with COVID-19, gene sets associated with cell adhesion and NK cell cytotoxicity were slightly upregulated in some NK cells. However, substantial downregulation was found in most effector NK cells in severe COVID-19 (Extended Data Fig. 8g). Of note, among the genes substantially downregulated in patients with severe COVID-19 was ITGB2, which encodes the β2-integrin (also known as CD18). This integrin associates with the α-integrins CD11a to CD11d to generate functional integrin receptors (Extended Data Fig. 8h) and is involved in NK-cell-mediated cytotoxicity28,29. This may provide a molecular explanation for the substantial failure of NK cells from patients with severe COVID-19 to form conjugates with target cells. Collectively, NK cells from patients with severe COVID-19 show a gene signature characterized by IFN-controlled cell activation programmes (for example, upregulation of perforin, granzyme B and CD69) superimposed by strong and long-lasting TGFβ-controlled transcriptional reprogramming, in particular the downregulation of genes linked to NK-cell-mediated cytotoxicity.

We wondered whether reprogramming of NK cells by TGFβ is a common event in pneumonia. We extracted NK cell data from a previously published scRNA-seq dataset of peripheral blood mononuclear cells (PBMCs) from patients with COVID-19 and from patients with influenza pneumonia30. This analysis revealed a strong enrichment of TGFβ-controlled genes in NK cells from patients with COVID-19 but not from patients with influenza pneumonia (Fig. 3h). To explore whether TGFβ signatures can also be found in lung NK cells, we analysed a single-nucleus RNA-seq dataset of post-mortem lung tissue samples from patients with severe COVID-19 or with SARS-CoV-2-negative pneumonia31. The dataset enabled the analysis of single-cell transcriptomes of lung NK cells (Extended Data Fig. 9a–c). GSEA revealed that lung NK cells from patients with severe COVID-19 pneumonia but not those from patients with SARS-CoV-2-negative pneumonia were significantly enriched in TGFβ-controlled genes (Extended Data Fig. 9d).

Untimely peak of serum TGFβ in severe COVID-19

Our transcriptional and functional data suggested that TGFβ is produced early in the phase of SARS-CoV-2 infection, particularly in the context of severe COVID-19. Patients with FLI and ambulant patients with COVID-19 did not show increased serum TGFβ within the first week after symptom onset, whereas hospitalized patients with COVID-19 had significantly increased TGFβ serum levels at week 1 that peaked at week 2 (Fig. 4a). Ambulant patients with COVID-19 showed only a small increment in serum TGFβ at later time points of the infection (>3 weeks after symptom onset). To obtain insights into the cellular sources of TGFβ, we analysed the lung single-nucleus RNA-seq dataset for TGFB transcripts and for transcripts associated with TGFβ bioactivity. In comparison to non-COVID-19 pneumonia illnesses, we observed a large increase in TGFB1 and TGFB2 expression in type 1 alveolar epithelial cells, fibroblasts, myofibroblasts, endothelial cells and myeloid cells (Extended Data Fig. 9e, g). Expression levels of various genes required for the bioactivity of TGFβ (such as LTBP3, LAP3 and MMP2) were also increased in patients with COVID-19 (Extended Data Fig. 9f). Of note, some of the clusters with the highest expression of TGFB1, LAP3 and MMP28 contained SARS-CoV-2 transcripts (that is, were infected) (Extended Data Fig. 9h). Thus, the early TGFβ peak in hospitalized patients with COVID-19 is closely correlated to the early impairment of NK cell effector programmes.

Fig. 4: Serum of patients with severe COVID-19 inhibits NK cell function in a TGFβ-dependent manner.
figure 4

a, Serum levels of active TGFβ. Independent measurements (mean) from 13 patients with FLI (n = 13) and 66 patients with COVID-19 (30 severe, n = 74; 7 moderate, n = 12; 39 ambulant, n = 53) at indicated time points after symptom onset (group >3 weeks contains samples from weeks 4 and 5). The dashed line indicates the median TGFβ serum level of 34 healthy donors. Patients receiving corticosteroids were excluded. b, NK cells from healthy donors were cultured in medium containing IL-12 and IL-15 with (green) or without TGFβ (blue) and co-cultured with SARS-CoV-2-infected Vero E6 cells. Viral load (mean ± s.e.m.) was measured 12 h later. Each data point represents NK cells from one individual donor (E:T 9:1 n = 3, all others n = 4, no NK cells n = 11 samples). c, d, Sorted NK cells from 3–9 healthy donors were cultured in medium containing either serum from a healthy donor or serum from a patient with severe COVID-19 (n = 9 (c) or 7 (d)) and the frequency of CD107a+ NK cells (mean) after co-culture with K562 (c) or the MFI of T-bet (d, mean) was determined. Each dot represents NK cells of one healthy donor cultured with serum of another healthy donor (blue; n = 17 (c) or 6 (d)) or of one patient with severe COVID-19 (red; n = 24 (c) or 14 (d)). e, f, Sorted NK cells from 3–8 healthy donors were cultured in medium alone or with serum from patients with severe COVID-19. In a second condition, patient sera were pre-incubated with anti-TGFβ and then added as described above. The frequency of CD107a+ NK cells (e, mean, n = 14) and the MFI of T-bet (f, mean, n = 32) were determined. Shown is the fold-change frequency or MFI between NK cells cultured in sera from patients (with or without prior anti-TGFβ treatment) and NK cells cultured in medium alone. For e, f, two-sided Wilcoxon matched-pairs rank test was used. For ad, two-sided Mann–Whitney U-test was used.

We wondered whether exposure of NK cells to TGFβ has any effect on their capacity to control viral replication. TGFβ completely abrogated the NK cell effector programme, that is, NK-cell-mediated control of SARS-CoV-2 replication in vitro (Fig. 4b), cell-mediated cytotoxicity, degranulation in response to target cells and cytokine production (Extended Data Fig. 10a–d). TGFβ treatment of stimulated NK cells also led to the downregulation of T-bet, whereas the canonical TGFβ target EOMES was upregulated (Extended Data Fig. 10e, f).

NK cells from healthy donors that were exposed to serum from patients with severe COVID-19, obtained during the first weeks after symptom onset, significantly inhibited NK cell degranulation and T-bet expression (Fig. 4c, d). This inhibitory effect was lost when sera was pretreated with a TGFβ-blocking antibody with activity against TGFβ1, TGFβ2 and TGFβ3 (Fig. 4e, f, Extended Data Fig. 10g–i). By contrast, neutralization of interleukin-6 (IL-6), IL-10 or IL-15 did not restore NK cell degranulation or T-bet expression (Extended Data Fig. 10j–p), even though these cytokines have been previously linked to severe COVID-19 (refs. 4,32). Thus, the untimely expression of TGFβ leads to suppressed NK cell function, which may reduce virus control and be detrimental in severe COVID-19.

Although TGFβ is thought to curtail excessive immune responsiveness and to restore or maintain immune homeostasis33, we now demonstrate that an untimely early production of TGFβ and associated NK cell dysfunction is a hallmark of severe COVID-19. Multilevel proteomics data support a specific dysregulation of TGFβ signalling by the SARS-CoV-2 ORF8 protein34 and we have previously shown that TGFβ impairs B-cell responses in the context of COVID-19 (ref. 35). Two matrix metalloproteinase inhibitors (prinomastat and marimastat) that diminish TGFβ bioactivity strongly inhibited the replication of SARS-CoV-2 but not of SARS-CoV34. TGFβ-mediated impairment of NK cell function may also have an impact on other aspects of COVID-19. A dysregulated myeloid response is another hallmark of severe COVID-19 in which there is an inadequate production of inflammatory cytokines36,37. Lessons learnt from genetic syndromes of NK cell cytotoxic defects have implicated NK cells in the quality control of innate immune responses by curtailing excessive myeloid responses, thereby preventing immunopathology27,38. Another intriguing link is the role of NK cells in the control of fibrotic reprogramming and the elimination of pre-fibrotic cells that undergo a senescence programme39. Thus, the inhibition of untimely TGFβ production and the promotion of NK cell function may positively affect SARS-CoV-2 control on multiple levels40.


Human participants

The recruitment of study participants was approved by the Institutional Review Board of Charité (EA2/066/20, EA2/072/20, EAEA4/014/20 and EA2/092/20). Written informed consent was provided by all patients or legal representatives for participation in the study. Forty-five ambulant patients with COVID-19 (World Health Organization (WHO) disease severity 1 and 2 according to the WHO clinical ordinal scale), 21 hospitalized patients (WHO 3–4) with moderate COVID-19 and 79 patients with severe COVID-19 who required ventilation (WHO 5–7, 52 of which fulfilled acute respiratory distress syndrome criteria according to the Berlin definition of this condition41) were enrolled in this study. All patients with COVID-19 were tested positive for SARS-CoV-2 RNA via nasopharyngeal swabs. Twenty patients who presented with influenza-like symptoms but were tested negative for SARS-CoV-2 (FLI) and a total of 96 healthy donors who did not present any clinical sign of viral infection were enrolled as controls. The clinical characteristics of all participants are summarized in Supplementary Tables 13. For autopsy, informed consent was given by the next of kin, and autopsies were performed on the legal basis of section 1 SRegG BE of the autopsy act of Berlin and section 25(4) of the German Infection Protection Act. The sequencing of the post-mortem tissue was approved by the Ethics Committee of the Charité (EA2/066/20, EA1/144/13 and EA1/075/19 to H.R.) and by the Charité-BIH COVID-19 research board and complied with the Declaration of Helsinki. Additional use of anonymized clinical data is covered by section 25 of the Berlin Hospital Law and did not require further ethical or legal clearance.

Isolation of PBMCs and serum

Peripheral blood was drawn from each donor into EDTA collection tubes and, for selected samples (CD107a assay), into heparin tubes (BD Biosciences). PBMCs were separated from peripheral blood by Pancoll human (PAN-Biotech) density gradient centrifugation at room temperature. Cells were either used directly for analysis or stored in heat-inactivated fetal bovine serum (FCS; Pan-Biotech, P30-3602) with 10% dimethylsulfoxide at −80 °C before analysis. Serum samples were drawn from each donor into Vacutainer SSTTM tubes (BD Biosciences), centrifuged for 10 min at 2,000g and stored at −20 °C before analysis.

Flow cytometry analysis

PBMCs were incubated with Fc blocking reagent (Miltenyi Biotec) according to the manufacturer’s instructions. To exclude dead cells, the cells were stained using a Live/Dead (LD) Fixable Aqua Dead Cell staining kit (ThermoFisher, L34965). For surface antigen staining, the cells were incubated with monoclonal anti-human antibodies (Supplementary Table 4) for 20 min at 4 °C. The Foxp3 Transcription Factor Staining Buffer Set (eBioscience, 00-5523-00) was applied before intracellular staining of transcription factors, cytotoxic molecules and cytokines. The samples were analysed using a FACS Fortessa X20 (BD Biosciences). Data were analysed using FlowJo Software v.10.3 (Treestar). Mean fluorescent intensity (MFI) values of NK cell populations were normalized to the MFI of lineage-negative marker-negative cells42. For intracellular cytokine staining of IFNγ (BioLegend, 502527) and TNF (BioLegend, 502908), NK cells were stimulated for 4 h at 37 °C in the presence of brefeldin A (Sigma-Aldrich) with phorbol 12-myristate 13-acetate (25 ng ml–1, Sigma-Aldrich) and ionomycin (500 ng ml–1, Sigma-Aldrich). Alternatively, NK cells were co-cultured with K562 target cells (American Type Culture Collection (ATCC) CCL-243, verified by ATCC) for 4 h at 37 °C in the presence of brefeldin A. In Supplementary Table 4, all antibodies are listed and assigned to staining panels A, B, C, D and E. Patient information and time point after onset of symptoms when the peripheral blood was obtained for each flow cytometry analysis are listed Supplementary Tables 13.

In vitro culture of sorted NK cells

Frozen PBMCs were gradually thawed at 37 °C and resuspended in RPMI medium (Gibco, 31870074) supplemented with 20% heat-inactivated FCS (Pan-Biotech P30-3602), l-glutamine (200 mM, Gibco, 25030081), penicillin–streptomycin (10,000 U ml–1, Gibco, 5140122) and gentamicin (Lonza BioWhittaker, BW17-519L). Live cells were discriminated using a Live/Dead Fixable Aqua Dead Cell staining kit (ThermoFisher, L34965) and incubated in human Fc blocking reagent (Miltenyi Biotec) according to the manufacturer’s instructions. The cells were stained with the following anti-human antibodies for 20 min at 4 °C: CD3 (eBioscience, 11-0039-42), CD4 (eBioscience, 11-0048-42), CD14 (BioLegend, 325604), CD19 (BioLegend, 363008), CD45 (BioLegend, 393409), CD7 (BioLegend, 343119) and CD56 (BioLegend, 362511). NK cells were sorted as LDLin (CD3, CD4, CD14 and CD19) CD45+CD7+CD56+ using a FACS Aria II Cell Sorter (BD Biosciences).

Sorted NK cells were cultured in RPMI containing rh-IL-12 (20 ng ml–1; PeproTech, 200-12H), rh-IL-15 (20 ng ml–1; PeproTech, 200-15), rh-TGFβ (10 ng ml–1; PeproTech, 100-21), rh-IFNα (10,000 U ml–1; R&D Systems, 11100-1) or rh-IFNβ (20 ng ml; R&D Systems, MAB1835-SP) and cultured for 2–4 days at 37 °C with 5% CO2 as indicated.

Chromium release assay

A chromium release assay was performed as previously described43. In brief, full PBMCs were cultured overnight in RPMI supplemented with rh-IL-12 and rh-IL-15. For selected analyses, rh-TGFβ was added. NK cell frequencies in the PBMC fraction were determined using flow cytometry. K562 target cells were radioactively labelled by incubating 2 × 106 K562 cells in 450 µl RPMI with 50 µl of 51Cr (CR-RA-8, Cr51, 185 MBq, 5 mCi ml–1) for 2 h at 37 °C on a rotator. After labelling, target cells were washed twice with RPMI before adding PBMCs at the indicated NK cell:target cell ratios (9:1, 3:1, 1:1 or 1:3). Following a co-culture incubation of 4 h (37 °C with 5% CO2), the supernatant was collected and the 51Cr released was quantified using a Wallac Wizard 1470 gamma counter. To quantify the maximal 51Cr release for each experiment, the gamma count of the supernatant of target cells cultured without effector cells (spontaneous 51Cr release) was subtracted from the gamma signal of labelled target cells only. The percentage of specific lysis or the percentage of 51Cr release induced by NK cells from healthy donors in a 3:1 NK cell:target cell ratio (per cent of maximum 51Cr release) was calculated for each sample.

Degranulation assay of NK cells

CD107a expression on NK cells of the PBMC fraction was measured as previously described44 (protocol 2 of the consensus protocol). In brief, PBMCs were cultured at 2 × 106 ml–1 in RPMI overnight. NK cells were subsequently co-cultured for 2 h with K562 cells and stained for CD107a (eBioscience, 11-1079-42), CD56 (Beckman Coulter, A82943), CD8 (Beckman Coulter, IM2469), CD3 (Beckman Coulter, A94680) and CD45 (Beckman Coulter, B36294). The samples were analysed on a Navios-EX FACS (Beckman Coulter). Data were analysed using the Navios 2.0 software. CD107a mobilization of sorted NK cells (as described in the section ‘In vitro culture of sorted NK cells’) cultured in RPMI, supplemented with rh-IL-12, rh-IL-15 and rh-TGFβ as indicated, was analysed after co-culture with K562 cells at a 2:1 NK cell:target cell ratio for 4 h at 37 °C. Subsequently, cells were re-stained using a Live/Dead Fixable Aqua Dead Cell staining kit, human Fc blocking reagent and the following antibodies: CD3 (eBioscience, 11-0039-42), CD4 (eBioscience, 11-0048-42), CD14 (BioLegend, 325604), CD19 (BioLegend, 363008), CD45 (BioLegend, 393409), CD7 (BioLegend, 343119), CD107a (BD Bioscience, 562622) and CD56 (BioLegend, 362511). The frequency of CD107a expression was determined in LDLin (CD3, CD4, CD14 and CD19) CD45+CD7+CD56+ NK cells using a FACS Fortessa X20 (BD Biosciences).

Conjugation assay

NK cell:target cell adhesion was assessed as previously described45. In brief, PBMCs were separated from peripheral blood by a density gradient as described above. Cells were enriched by negative selection using a NK cell isolation kit (Miltenyi Biotec) according to the manufacturer’s instructions. After isolation, NK cells were labelled for 20 min using a Celltrace Far Red Cell Proliferation kit (Invitrogen, C34564) according to the manufacturer’s instructions. K562 target cells were labelled for 10 min at 37 °C with cell proliferation dye (Invitrogen, 65-0842-85). After labelling, cells were washed with RPMI, mixed in a 1:4 NK cell: target cell ratio and centrifuged at 300g for 1 s. Cells were co-cultured for 1 h at 37 °C before conjugates were quantified by flow cytometry.

Co-culture of NK cells with SARS-CoV-2-infected cells

PBMCs were separated from freshly drawn peripheral blood by a density gradient as described above. Cells were enriched by negative selection using a NK cell isolation kit (Miltenyi Biotec) according to the manufacturer’s instructions. Enrichment was ensured by flow cytometry. After isolation, NK cells were stimulated for 24–48 h in RPMI 1640 supplemented with rh-IL-12 (PeproTech, 200-12H), rh-IL-15 (PeproTech, 200-15) and rh-TGFβ (PeproTech, 100-21) as indicated at 37 °C with 5% CO2. NK cells were washed with DMEM and co-cultured with SARS-CoV-2-infected Vero E6 cells (Cercopithecus aethiops kidney epithelial cells; ATCC CRL-1586, verified by ATCC) or Calu-3 cells (human bronchial epithelial cells; ATCC HTB-55, verified by ATCC) as indicated. Approximately 175,000 Vero E6 cells and 300,000 Calu-3 cells per well were seeded in 24-well plates 24 h before infection as indicated.

For masking experiments, NK cells were incubated with the neutralizing antibodies anti-NKp30 (clone F252), anti-NKp44 (clone KS38), anti-NKp46 (clone KL247), anti-NKG2D (clone BAT221 and clone ON72), anti-DNAM-1 (clone F5) and anti-2B4 (clone CO54) at room temperature 45 min before co-culture as indicated. For all blocking experiments, target cells were incubated with anti-HLA-I (clone A6/136). Monoclonal antibodies were provided by E. Marcenaro (University of Genova, Italy).

Cells were infected with the SARS-CoV-2/München 984 virus isolate (B.1 lineage, hCoV-19/Germany/BY-ChVir-984/2020, accession ID: EPI_ISL_406862; Pango lineage v.3.1.1, lineages v.2021-06-15) or the variant B.1.351 (hCoV-19/Germany/BW-ChVir22131/2021, accession ID: EPI_ISL_862149) at cell culture passage 2 (ref. 46) with a multiplicity of infection (MOI) of 0.001 (Vero E6) or a MOI of 0.1 (Calu-3). Virus was diluted in OptiPro serum-free medium (Thermo Fisher). For infection, the supernatant was removed, cells were rinsed once with 0.5 ml PBS (Thermo Fisher) and 200 µl of virus-containing dilution was inoculated on the cells for 1 h at 37 °C. Next, 500 µl of NK cell suspension was added 12 h after infection as indicated. Adherent cells were collected 12 h (Vero E6) or 24 h (Calu-3) after NK cell addition for isolation of viral RNA. For isolation of viral RNA, 35 µl of MagNA Pure 96 external lysis buffer (Roche) was added to the adherent cells. All samples were heat-inactivated for 10 min at 70 °C. Isolation and purification of viral RNA was performed using a MagNA Pure 96 System (Roche) according to the manufacturer’s recommendations. Viral RNA was quantified using real-time PCR with reverse transcription (RT–PCR; E gene assay) as previously described47. All infection experiments were done under biosafety level 3 conditions with enhanced respiratory personal protection equipment.

Viral load data analyses

Viral load (RNA copies per swab) measurements were obtained based on a calibrated curve of viral RNA copies and RT–PCR cycle threshold values as previously described48. For the examination of temporal viral load dynamics, patients with at least two viral load measurements and two immune cell count measurements were included, and a linear regression was calculated. A maximum offset for both parameters was set to 7 days, and only infections with a duration of less than 40 days were considered. Regression analysis was performed using the seaborn (regplot) package v.0.11.1 and the scipy (stats.linregress) package (v.1.6.0) running under Python 3.9.1. For indicated analyses, patients were categorized as having either low or normal absolute immune cell counts according to whether their first immune cell count had a value below or above a threshold value. The following thresholds were used: 40 NK cells per µl; 40 B cells per µl; 360 T cells per µl; 200 CD4+ T cells per µl; 300 CD8+ T cells per µl; 600 lymphocytes (CD45+CD14) per µl. Analysis was performed on hospitalized patients who tested positive (through SARS-CoV-2 RT–PCR), including patients who tested positive in an intensive care unit ward at any point during their infection and including patients with severe COVID-19 of the study cohort (Extended Data Fig. 1, Supplementary Table 3).

NK cell isolation from peripheral blood for single-cell sequencing

Frozen PBMCs from 13 patients with severe COVID-19 (52 samples), 11 ambulant patients with COVID-19 (11 samples) from day 2 to day 68 after onset of symptoms (Supplementary Tables 13) and 5 healthy donors (5 samples) were thawed in RPMI (Gibco, 31870074) supplemented with 20% heat-inactivated FCS and incubated with Fc blocking reagent (Miltenyi Biotec) following the manufacturer’s instructions. Up to 1 × 107 cells per 100 µl were stained with the following anti-human antibodies: CD3 (Miltenyi Biotec, 130-113-133), CD14 (Miltenyi Biotec, 130-113-152), CD19 (Miltenyi Biotec, 130-113-172), CD45 (BioLegend, 304008) and CD56 (Miltenyi Biotec, 130-113-305). To allow cell pooling, each sample was also incubated with one of eight different TotalSeq-C anti-human Hashtags (LNH-94;2M2, Barcoded, BioLegend, 394661, 394663, 394665, 394667, 394669, 394671, 394673 and 394675). 4,6-diamidino-2-phenylindole (DAPI) was added before sorting to enable exclusion of dead cells. NK cells were identified and sorted as DAPICD3CD14CD19CD45+CD56+. All sortings used a MA900 Multi-Application Cell Sorter (Sony Biotechnology). Cell counting was performed using a MACSQuant flow cytometer (Miltenyi Biotec). Sorted NK cells were further processed for scRNA-seq.

For the generation of a NK cell-specific TGFβ-response dataset, live cells in thawed PBMCs were discriminated using a Live/Dead Fixable Aqua Dead Cell staining kit (ThermoFisher, L34965) and incubated in human Fc blocking reagent (Miltenyi Biotec) according to the manufacturer’s instruction. The cells were stained with the following anti-human antibodies for 20 min at 4 °C: CD3 (eBioscience, 11-0039-42), CD4 (eBioscience, 11-0048-42), CD14 (BioLegend, 325604), CD19 (BioLegend, 363008), CD45 (BioLegend, 393409), CD7 (BioLegend, 343119) and CD56 (BioLegend, 362511). NK cells were sorted as LDLin (CD3, CD4, CD14 and CD19) CD45+CD7+CD56+ using a FACS Aria II Cell Sorter (BD Biosciences). Sorted NK cells were cultured in RPMI containing rh-IL-12 and rh-IL-15 with or without additional rh-TGFβ for 4 days at 37 °C with 5% CO2 before scRNA-seq.

scRNA library preparation and sequencing

The 10X Genomics workflow for cell capturing and scRNA gene expression (GEX) was applied to sorted NK cells using a Chromium Single Cell 5′ Library & Gel Bead kit and a Single Cell 5′ Feature Barcode Library kit (10X Genomics). Final GEX was obtained after fragmentation, adapter ligation and final index PCR using the Single Index Kit T Set A. A Qubit HS DNA assay kit (Life Technologies) was used for library quantification, and fragment sizes were determined using the Fragment Analyzer with the HS NGS Fragment kit (1–6,000 bp; Agilent). Sequencing was performed on a NextSeq500 device (Illumina) using High Output v2 kits (150 cycles) with the recommended sequencing conditions for 5′ GEX libraries (read 1: 26 nt, read 2: 98 nt, index 1: 8 nt).

Single-cell transcriptome analysis

Chromium single-cell data were processed using cellranger-3.1.0. The mkfastq and the count pipeline were used in default parameter settings for demultiplexing, alignment of reads to the Refdata-cellranger-hg19-1.2.0 genome, barcode and unique molecular identifier (UMI) counting, and calling of intact cells. The number of expected cells was set to 3,000. Further analysis was performed using the Seurat R-package (v.3.1.1)49. The pooled samples were separated using eight cite-seq hashtags (TotalSeq-C, BioLegend). Cells with more than 30% cite-seq reads for a particular hashtag were assumed as positively stained. Cells with no (undefined origin) or ambivalent assignments (doublets) were removed from further analysis. The resulting transcriptome profiles of NK cells from healthy donors, ambulant patients with COVID-19 and patients with severe COVID-19 were normalized and integrated as previously described50. Variable genes were detected and scaled. Data were scaled and UMAP was performed in default parameter settings using Scale Data, RunPCA and RunUMAP with 30 principal components. Quality control was performed by visual inspection of the fraction of mitochondrial genes, the number of detected genes and UMI counts per cell. No notable abnormalities were observed. Transcriptionally similar cells were clustered using shared nearest neighbour (SNN) modularity optimization with a SNN resolution of 0.2. Marker genes for clusters were identified using FindAllGenes with a log fold-change of >2.5 and a minimum of 0.1 expressing cells. The genes used for manual annotation of the function of the clusters was informed by previous publications16,20,51. One out of six clusters revealed B-cell-specific and monocyte-specific genes (CD19, CD14) and was excluded from further analysis.

Pseudotime trajectory analysis was performed using the monocle2 R package. Cells from each group (healthy, ambulant and severe COVID-19) were randomly downsampled to an equal depth of 3,871 cells. The overall 11,613 cells were used for the calculation in monocle2. The top 1,000 differentially expressed genes across the five clusters (ranked by lowest q-value) were used to order the cells. The BEAM method was used to identify genes that significantly differed in expression level across a branch point22.

Single-cell data derived from NK cells cultured in presence and absence of TGFβ were analysed using the analog workflow except for clustering and trajectory analysis steps. The TGFβ response signature was defined on the basis of differentially expressed genes between samples with and without TGFβ treatment. To compare samples, two-sided Mann–Whitney U-test was used on log-normalized counts. Differentially expressed genes were identified (adjusted P ≤ 0.05 and a log fold-change of >2.2 for enrichment and <2.2 for depletion). The heatmap of differentially expressed genes is based on z-transformed means of log-normalized counts. Hierarchical clustering was performed using Euclidian distances and WARD linkage criterium. All generated sequencing data were deposited into the NCBI Gene Expression Omnibus (accession number GSE184329).

Analysis of peripheral blood NK cells from a publicly available dataset

NK cell data were extracted from a publicly available PBMC single-cell sequencing dataset (GSE149689) consisting of 4 samples from healthy participants, 11 samples from patients with COVID-19 and 5 samples from patients with severe influenza30. In the first step, the individual samples and 59,572 cells were integrated and subdivided into 10 clusters in a UMAP representation with a SNN resolution of 0.1. A heatmap of cell-type-defining genes was used to identify 10,604 NK cells from the PBMCs. GSEA for TGFβ-induced and TGFβ-suppressed NK cell genes was performed as described below.

Analysis of non-haematopoietic and haematopoietic cell populations from lung tissue

The publicly available lung single-nucleus dataset (European Genome–Phenome Archive reference ID: EGAS00001004689) consisting of three patients who died from pneumonia unrelated to COVID-19 (as controls) and seven patients who died from COVID-19-associated pneumonia31 was aligned to a hg19 reference transcriptome that included the SARS-CoV-2 genome (NCBI reference sequence ID: NC_045512) using cellranger v3.0.1 (10X Genomics). Ambient RNA removal was performed with SoupX (v1.4.5)52. First, the 10 samples and 53,709 cells were integrated as previously described50 for the NK cell single-cell sequencing dataset and subdivided into 33 clusters in a UMAP representation with a SNN resolution of 1. A heatmap of 13 genes was used to identify NK cells from the lung, and cluster 25 was defined as the NK-cell-specific cluster of 845 cells. The remaining cell populations were identified according to Gassen et al.31.


GSEA was performed for each cell on the basis of differences in log-normalized counts to the mean of all cells analysed using 1,000 randomizations (false discovery rate < 0.5 and normalized P < 0.2)53. For visualization, the normalized enrichment score per cell within the indicated gene set (upregulation or downregulation) was plotted. Gene sets (Hallmark, Reactome and KEGG) were obtained from the Molecular Signatures Database (MSigDB, v.6.2)54,55. The NK-cell-specific TGFβ response gene set was newly generated in this study. The GSEA-enrichment plot (score curve) of TGFβ-suppressed genes in lung-tissue-resident NK cells was performed on pre-ranked differences between medians of expressing cells in otherwise default parameter settings. Only genes expressed in at least one group were considered. The median was set to 0 if no expressing cells were found.

In vitro NK cell exposure to patient serum

Sorted NK cells from healthy donors were cultured for 48 h in RPMI containing IL-12, IL-15 and IL-2 (25 ng ml–1; PeproTech, 212-12) as indicated and 20% of either serum from another healthy donor (Supplementary Table 1) or from a patient with severe COVID-19 (Supplementary Table 3). Unless stated otherwise, IL-12, IL-15 and serum were used. For indicated experiments, the patient sera were pre-incubated with antibodies directed against TGFβ1, TGFβ2 and TGFβ3 (R&D Systems, MAB1835-SP), anti-IL-6 (5 µg ml–1; R&D Systems, MAB206), anti-IL-10 (30 µg ml; R&D Systems, MAB217) or anti-IL-15 (5 µg ml; eBioscience, 16-0157-82) as indicated for 10 min before being added to the culture and using identical culture conditions. After 48 h, protein expression levels of T-bet were measured, the frequency of CD107a expression in NK cells was analysed after 4 h of co-culture with K562 cells using flow cytometry or viral replication was determined, as described above.

Cytokine measurements

Cytokine levels were measured using a bead-based multiplex cytokine array (Human Cytokine 25-Plex ProcartaPlex Panel 1B, Thermo Fisher Scientific). Before the assay, serum samples were diluted 1:3 in dilution buffer provided with the kit. TGFβ was detected using a Human TGFβ1 Simplex ProcartaPlex kit (Thermo Fisher Scientific). Before measuring serum TGFβ1, the bioactive form of TGFβ1 was generated by incubating the serum with 1 N HCl followed by neutralization with 1.2 N NaOH according to the manufacturer’s instructions. The samples were incubated with antibody-coated magnetic beads for 30 min at room temperature with shaking, then incubated overnight at 4 °C followed by a 1-h incubation period at room temperature. All subsequent incubation steps were performed according to the manufacturer’s instructions. The assay plates were read using a Luminex MAGPIX system and quantified using xPONENT analysis software (Luminex). IFNα serum concentration was analysed using a Simoa IFNα Advantage kit (Quanterix) according to the manufacturer’s instructions.

Statistical analysis and reproducibility

All statistical tests were performed with Graph Pad Prism V7 software as indicated for each analysis (*P ≤ 0.05, **P ≤ 0.01, ***P ≤ 0.001 and ****P ≤ 0.0001; NS, not significant). Representative data of at least three independent experiments are shown in Figs. 1d, e, g, 2f and 4c, f, and Extended Data Figs. 15a, b, c, d. Representative data of two independent experiments are shown in Extended Data Fig. 15j, k, n, o.

Reporting summary

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