ARTICLE https://doi.org/10.1038/s41467-022-28621-0 OPEN Stratification of hospitalized COVID-19 patients into clinical severity progression groups by immuno-phenotyping and machine learning Yvonne M. Mueller 1, Thijs J. Schrama 1, Rik Ruijten1, Marco W. J. Schreurs1, Dwin G. B. Grashof 1, Harmen J. G. van de Werken 1,2, Giovanna Jona Lasinio 3, Daniel Álvarez-Sierra 4, Caoimhe H. Kiernan1, Melisa D. Castro Eiro 1, Marjan van Meurs1, Inge Brouwers-Haspels 1, Manzhi Zhao1, Ling Li1, Harm de Wit1, 1234567890():,; Christos A. Ouzounis5,6, Merel E. P. Wilmsen1, Tessa M. Alofs1, Danique A. Laport1, Tamara van Wees1, Geoffrey Kraker7, Maria C. Jaimes7, Sebastiaan Van Bockstael7, Manuel Hernández-González4,8,9, Casper Rokx 10, Bart J. A. Rijnders10, Ricardo Pujol-Borrell4,8,9,11 & Peter D. Katsikis 1 ✉ Quantitative or qualitative differences in immunity may drive clinical severity in COVID-19. Although longitudinal studies to record the course of immunological changes are ample, they do not necessarily predict clinical progression at the time of hospital admission. Here we show, by a machine learning approach using serum pro-inflammatory, anti-inflammatory and anti-viral cytokine and anti-SARS-CoV-2 antibody measurements as input data, that COVID- 19 patients cluster into three distinct immune phenotype groups. These immune-types, determined by unsupervised hierarchical clustering that is agnostic to severity, predict clinical course. The identified immune-types do not associate with disease duration at hospital admittance, but rather reflect variations in the nature and kinetics of individual patient’s immune response. Thus, our work provides an immune-type based scheme to stratify COVID-19 patients at hospital admittance into high and low risk clinical categories with distinct cytokine and antibody profiles that may guide personalized therapy. 1 Department of Immunology, Erasmus University Medical Center, Rotterdam, The Netherlands. 2 Cancer Computational Biology Center, Erasmus MC Cancer Institute, Erasmus University Medical Center, Rotterdam, The Netherlands. 3 Department of Statistical Sciences, University of Rome “La Sapienza”, Roma, Italy. 4 Immunology Division, Hospital Universitari Vall d’Hebron, Campus Vall d’Hebron, Barcelona, Spain. 5 School of Informatics, Faculty of Sciences, Aristotle University of Thessaloniki, Thessalonica, Greece. 6 Chemical Process & Energy Resources Institute, Centre for Research & Technology Hellas, Thermi, Thessalonica, Greece. 7 Cytek Biosciences, Fremont, CA, USA. 8 Cell Biology, Physiology and Immunology Department, Universitat Autònoma de Barcelona, Barcelona, Spain. 9 Translational Immunology Research Group, Vall d’Hebron Institut de Recerca (VHIR), Campus Vall d’Hebron, Barcelona, Spain. 10 Department of Internal Medicine, Section Infectious Diseases, and Department of Medical Microbiology and Infectious Diseases, Erasmus University Medical Center, Rotterdam, The Netherlands. 11Present address: Vall d’Hebron Institute of Oncology (VHIO), Barcelona, Spain. ✉email: p.katsikis@erasmusmc.nl NATURE COMMUNICATIONS | (2022)13:915 | https://doi.org/10.1038/s41467-022-28621-0 | www.nature.com/naturecommunications 1 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-28621-0 T he newly emerged SARS-CoV-2 virus has caused three distinct immunotypes BRI, EXI, and LAI, who exhibited the COVID-19 pandemic and infected >120 million people similar cytokine and antibody characteristics as those discovered over the world, resulting in >2.8 million deaths1. In the in the Rotterdam cohort (Fig. 1A, B). absence of a highly effective therapy against COVID-19, there We further tested the robustness of our immunotypes using remains an urgent need to understand both the pathological both cohorts and a multinomial model. The model showed a high mechanisms that lead to severe disease but to also identify clear level of accuracy (100%) in predicting the immunotypes using the phenotypes that predict disease severity progression and outcome input data. Next, we estimated the stratification accuracy from a as this may instruct a more personalized therapy. In an attempt to predictive perspective by cross-validation of the dataset into a test understand the features of COVID-19 that associate with disease and training set. We used increasing test percentages from 5%, severity, studies have aimed at capturing the perturbation of the 10%, 15%, 20%, and 25% and using 150 iterations. The latter immune system and the associated inflammatory syndrome resulted in a median accuracy of 0.83, 0.80, 0.80, 0.80, 0.81, observed. Some of these studies have applied high-dimensional respectively, which shows that the model prediction is marginally analysis using multiplex cytokines, flow or mass cytometry, or depending on the training set size. The above confirmed the scRNAseq to identify changes in cytokine profiles, peripheral robustness of our immunotypes. blood immune cell composition and/or gene expression related to The three immunotypes were characterized by distinct serum COVID-19 severity. Universally, however, these studies have cytokine profiles and anti-SARS-CoV-2 antibody responses employed disease severity classification to identify immunotypes (Fig. 2). Anti-SARS-CoV- nucleocapsid protein (NC) antibodies that characterize mild, moderate or severe disease2–8. Although, were measured since this CE-certified ELISA system allowed these studies have identified specific changes present in COVID- detection of NC-specific IgM, IgG and IgA isotypes and this could 19 patients compared with healthy individuals, identifying clear reveal isotype-specific differences. Compared to healthy controls, immunotypes that strongly associate with or predict disease all three immunotypes had increased pro-inflammatory cytokines severity has proven more challenging2–5. Defining, however, (Fig. 2A and Supplementary Fig. S1), displaying further immunotypes based on clinical severity is based on the assump- significant differences between them. Immunotype BRI was tion that a single mechanism underlies all patients and that characterized by low pro-inflammatory, anti-viral and anti- kinetics are exclusively driven by days of infection. This approach inflammatory cytokines and normal TGFβ1 levels (Fig. 2A and is, thus, hampered by the dynamic nature of the immune and Supplementary Fig. S1). BRI exhibited robust IgM, IgG, and IgA inflammatory response to SARS-CoV-2 virus, the very different anti-SARS-CoV-2 NC antibodies (Figs. 1A, B and 2B). In kinetics that individual patients may exhibit, and the likelihood contrast, EXI had a much more pro-inflammatory profile, low that very different immune mechanisms underlie the same clin- IFNα and normal TGFβ1 (Fig. 2A). Immunotype EXI was also ical severity. associated with IgM, IgG and IgA anti-SARS-CoV-2 antibodies By applying machine learning to a discovery and a validation (Figs. 1A, B and 2B). Immunotype LAI exhibited a distinct cohort, here we show that COVID-19 patients can be classified, at profile from the previous two and was characterized by the hospital admittance, into distinct immune-phenotypes. These presence of a strong IFNα response, reduced TGFβ1 (Fig. 2A), immunotypes predict subsequent clinical progression and out- and very low antibody immunity (Figs. 1A, B and 2B). Pro- come. Such immunotypes can guide the development of practical inflammatory cytokines were significantly higher in EXI com- biomarkers but may also instruct more personalized treatments. pared to both other immunotypes (Fig. 2A). IL-17A was increased in some LAI and EXI patients, whereas IL-5 was decreased in some LAI patients (Supplementary Fig. S1). Very Results few patients had IL-2 or IL-12 in serum (Supplementary Fig. S1) Distinct immunotypes are identified by machine learning in while IL-4 was undetectable (data not shown). To verify that anti- acute COVID-19 disease. In this study, we chose to take an NC responses is an appropriate surrogate for the overall anti- unbiased approach in terms of clinical severity to identify SARS-Cov-2 antibody response, the anti-NC IgG response was immunotypes by first defining immunotypes in COVID-19 correlated with anti-Spike receptor-binding domain (RBD) Ig patients and then examining if these relate to clinical severity response (Supplementary Fig. S2A, total anti-RBD Ig) and serum and progression. At time of hospital entry, we measured in the neutralizing anti-SARS-Cov-2 titers (Supplementary Fig. S2B, serum of COVID-19 patients (Rotterdam discovery cohort; PRNT50) in the Rotterdam cohort. Anti-NC IgG responses n = 50, Table 1) modules of specific cytokines with pro-inflam- strongly correlated with both (Supplementary Fig. S2A, B). matory, anti-inflammatory or anti-viral activities. We combined As expected, we found that anti-SARS-CoV-2 IgM, IgG and these serum cytokines with the host adaptive antibody response IgA antibodies correlated strongly with each other (Fig. 2C). The and applied machine learning using unsupervised hierarchical strongest cytokine correlation with antibodies, was a negative clustering to identify immunophenotypes that capture both innate correlation with IFNα (Fig. 2C). This could indicate that and adaptive responses to SARS-CoV-2 infection. Investigators antibodies reduce viral loads and thus IFNα or conversely, high were not involved in clustering or grouping of patients. Impor- IFNα levels delay or inhibit antibody production. Pro- tantly, we did not use clinical severity as a clustering variable. inflammatory cytokines correlated with each other (Fig. 2C). Using this approach, we identified three distinct immunotypes, Anti-viral cytokines IFNα and IFNγ correlated with each other, (labeled: balanced response immunotype: BRI, excessive inflam- while IFNα negatively correlated with TGFβ1 (Fig. 2C). Thus, mation immunotype: EXI, and low antibody immunotype: LAI) in characteristics of the three immunotypes could be driven by acutely infected COVID-19 patients (Fig. 1A). To validate these distinct cytokine networks in action. immunotypes, we applied the same machine learning approach on a second independent cohort of patients from a hospital in Bar- celona (Barcelona validation cohort; n = 88, Table 1) (Fig. 1B). COVID-19 immunotypes predict clinical severity progression. Principal component analysis (PCA) showed that measurements We next investigated whether these distinct immunotypes asso- of the Barcelona cohort data matched very well with the Rotter- ciated with clinical parameters (Table 2). As mentioned, clinical dam data (Fig. 1C). Independent unsupervised hierarchical clus- severity was not used as a variable for the unsupervised hier- tering of the measurements from the Barcelona cohort archical clustering. At study entry, BRI and LAI did not differ in consistently revealed a very similar classification of patients into WHO clinical severity scores9 while EXI was significantly higher 2 NATURE COMMUNICATIONS | (2022)13:915 | https://doi.org/10.1038/s41467-022-28621-0 | www.nature.com/naturecommunications NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-28621-0 ARTICLE Table 1 Clinical and laboratory characteristics of Rotterdam discovery and Barcelona validation cohorts. Level Overall Barcelona Rotterdam N 138 88 50 Immunotype BRI 33 (23.9%) 19 (21.6%) 14 (28.0%) EXI 46 (33.3%) 33 (37.5%) 13 (26.0%) LAI 59 (42.8%) 36 (40.9%) 23 (46.0%) Gender Male 91 (65.9%) 58 (65.9%) 33 (66.0%) Age Years 62 (54–70) 61 (50–70) 63 (57.25–69) Days from symptom onset 8 (6–12) 8 (6–10) 9 (6–14.75) [n = 134] [n = 84] WHO 8-point score at study entry 3 38 (27.5%) 34 (38.6%) 4 (8.0%) 4 60 (43.5%) 31 (35.2%) 29 (58.0%) 5 20 (14.5%) 12 (13.6%) 8 (16.0%) 6 12 (8.7%) 10 (11.4%) 2 (4.0%) 7 8 (5.8%) 1 (1.1%) 7 (14.0%) Obesity Yes 43 (31.2%) 20 (22.7%) 23 (46.0%) NA 6 (4.3%) 2 (2.3%) 4 (8.0%) Diabetes mellitus Yes 32 (23.2%) 18 (20.5%) 14 (28.0%) NA 2 (1.4%) 2 (2.3%) 0 (0.0%) Heart disease Yes 48 (34.8%) 35 (39.8%) 13 (26.0%) NA 2 (1.4%) 2 (2.3%) 0 (0.0%) Lung disease Yes 30 (21.7%) 15 (17.0%) 15 (30.0%) NA 2 (1.4%) 2 (2.3%) 0 (0.0%) Kidney disease Yes 5 (3.6%) 3 (3.4%) 2 (4.0%) NA 2 (1.4%) 2 (2.3%) 0 (0.0%) Liver disease Yes 6 (4.3%) 5 (5.7%) 1 (2.0%) NA 2 (1.4%) 2 (2.3%) 0 (0.0%) Cancer Yes 7 (5.1%) 3 (3.4%) 4 (8.0%) NA 2 (1.4%) 2 (2.3%) 0 (0.0%) Fever (>38 °C) Yes 91 (65.9%) 68 (77.3%) 23 (46.0%) NA 2 (1.4%) 2 (2.3%) 0 (0.0%) C-reactive protein (mg/L) 102 (68–181) 116 (73–192) 96 (52–151) [n = 127] [n = 78] [n = 49] Ferritin (µg/L) 666 (382–1101) 598 (314–1034) [n = 72] 738 (487–1171) [n = 116] [n = 44] Lactate dehydrogenase (U/L) 369 (293–472) 369 (296–474) 373 (286–461) [n = 121] [n = 72] [n = 49] Alanine aminotransferase (U/L) 29 (19–54) 32 (18–54) 25 (21–54) [n = 49] [n = 137] [n = 88] Bilirubin (μmol/L) 9.8 (7.1–13.0) 10.1 (7.8–13.2) 8.5 (6.0–12.0) [n = 46] [n = 122] [n = 76] D-dimer (mg/L) 0.40 (0.24–0.98) [n = 120] 0.30 (0.19–0.48) [n = 79] 0.97 (0.60–1.49) [n = 41] Hemoglobin (mmol/L) 8.30 (7.51–9.06) [n = 136] 8.66 (7.88–9.18) [n = 86] 7.65 (6.60–8.30) [n = 50] Leukocytes (x109/L) 7.45 (5.90–9.84) [n = 136] 7.62 (6.11–9.87) [n = 86] 7.00 (5.20–9.78) [n = 50] Neutrophils (x109/L) 5.70 (4.16–8.01) [n = 133] 6.02 (4.60–8.03) [n = 86] 5.10 (3.70–7.70) [n = 47] Lymphocytes (x109/L) 1.02 (0.83–1.41) [n = 133] 1.02 (0.83–1.38) [n = 86] 1.00 (0.80–1.50) [n = 47] Monocytes (x109/L) 0.45 (0.30–0.64) [n = 133] 0.48 (0.33–0.66) [n = 86] 0.40 (0.29–0.60) [n = 47] Thrombocytes (x109/L) 216 (163–294) 209 (166–289) [n = 86] 229 (151–298) [n = 49] [n = 135] Tocilizumab before study entry Yes 7 (5.1%) 0 (0.0%) 7 (14.0%) NA 0 (0.0%) 0 (0.0%) 0 (0.0%) Tocilizumab after study entry Yes 2 (1.4%) 0 (0.0%) 2 (4.0%) NA 88 (63.8%) 88 (100.0%) 0 (0.0%) Corticosteroids after study entry Yes 10 (7.2%) 0 (0.0%) 10 (20.0%) NA 88 (63.8%) 88 (100.0%) 0 (0.0%) Convalescent plasma after study entry Yes 29 (21.0%) 0 (0.0%) 29 (58.0%) NA 88 (63.8%) 88 (100.0%) 0 (0.0%) Data are n (%) or median (Q1–Q3). Complete data was available for the continuous values shown if not stated otherwise [n = ]; NA not available. The clinical characteristics and laboratory measurements of Rotterdam discovery cohort, Barcelona validation cohort and the combination of both shown. (Fig. 3A). Thus at study entry, the immunotypes are not deter- improved after entry and clinical scores declined (the median mined by disease severity. To assess clinical severity progression peak score of 3, was the score of entry) (Fig. 3A). Reflecting the during hospitalization, we examined the highest/worst clinical more severe disease scores during hospitalization, all mortality score that patients exhibited within 30 days of admittance. During occurred in EXI and LAI patients, while no patients died in BRI hospitalization, EXI and LAI were characterized by clinical (Figs. 1A, B and 3B). Furthermore, total days in hospital (Fig. 3C) deterioration and higher peak WHO clinical severity scores and total days in ICU (Fig. 3D) also differ significantly between (median peak score of 6 for both) (Fig. 3A). In contrast, BRI patients with the BRI immunotype and patients either in the EXI NATURE COMMUNICATIONS | (2022)13:915 | https://doi.org/10.1038/s41467-022-28621-0 | www.nature.com/naturecommunications 3 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-28621-0 A B Immunotype Immunotype Outcome Outcome Entry WHO score Entry WHO score Peak WHO score Peak WHO score DFSO DFSO IgM 3 Immunotype IgM Immunotype BRI 3 BRI 2 EXI EXI 2 IgG 1 LAI IgG 1 LAI 0 Outcome 0 Outcome IgA −1 Deceased IgA −1 Deceased −2 Discharged −2 Discharged −3 Entry WHO score TGFβ1 −3 Entry WHO score TGFβ1 7 7 IL-5 IL-5 3 3 Peak WHO score IFNγ Peak WHO score IFNγ 8 8 IFNα IFNα 3 3 DFSO CCL2 DFSO CCL2 25 35 IL-6 IL-6 0 5 TNFα TNFα IL-1β IL-1β IL-8 IL-8 IL-18 IL-18 IL-10 IL-10 C 0.2 0.1 PC2 (21%) Origin Barcelona Rotterdam 0.0 Healthy Controls −0.1 −0.2 −0.1 0.0 0.1 PC1 (27%) Fig. 1 Unsupervised hierarchical clustering identifies three distinct immunotypes in acute COVID-19 patients. Applying machine learning by unsupervised hierarchical clustering solely to serum cytokines and anti-SARS-CoV-2 antibodies identifies three distinct immunotypes. Analysis was performed on samples collected at study entry and without clinical data input. The three immunotypes, identified independently in two patient cohorts, are depicted in A the heat map of the Rotterdam discovery cohort (n = 50) and B the heat map of the Barcelona validation cohort (n = 88) using row-based log-transformed z-scores. C Principal component analysis (PCA) of serum cytokines and anti-SARS-CoV-2 antibodies shows that both Rotterdam and Barcelona cohorts cluster together while they lie apart from healthy controls. The first two components with their percentage of variance are shown in parentheses. Red-blue color depicts z-scores. Top banners of heatmaps show: immunotypes BRI, EXI and LAI, WHO clinical score at entry and the peak during hospitalization, patient death and discharge, and days from symptoms onset (DFSO). or LAI immunotype group. These differences in severity between duration shortest in LAI, differences in DFSO did not explain immunotypes could not be attributed to age as these did not vary these phenotypes (Fig. 4A–C). Cytokine and antibody levels did significantly (Fig. 3E). Although sex distribution significantly not significantly correlate with DFSO, suggesting that immuno- differs between the BRI and EXI group (Fig. 3F), no significant types are not determined by disease duration (Fig. 4B, C). More difference was detected between BRI and LAI and EXI and LAI, importantly, the kinetics of antibodies and cytokines differed indicating that gender is not the predictive determinant of the between the three immunotypes with LAI even after DFSO of three immunotypes. To assess and emphasize the importance of 10 days still having high levels of IFNα while the antibody the distinct molecular markers and other variables, these variables responses remained muted (Fig. 4B, C). IL-6 and TNFα, on the were directly compared within the three immunotypes using other hand, were high in EXI patients already at DFSO of 5 days Wilcoxon rank-sum/Mann–Whitney U-test. We additionally (Fig. 4B). For the individual immunotypes, anti-SARS-CoV-2 applied a logistic regression model to each biomarker using both antibody responses had different trajectories in terms of DFSO. In a Uni- and Multivariate approach with gender, age, and DFSO as BRI and EXI patients, antibodies come up early and stay up additional variables for the multivariate analysis. Overall, no (Fig. 4C). The frequency of patients positive for anti-SARS-CoV-2 striking difference between the non-parametric and the para- antibodies was increased in BRI and EXI compared to LAI irre- metric tests were detected (Supplementary Table S1) and further spective of time since symptoms onset (Fig. 4D). Thus, duration confirmed that gender, age, and DFSO were not defining para- of infection could not explain LAI’s high IFNα or low antibodies. meters of the immunotypes. DFSO also failed to correlate with days of hospitalization and Thus, despite immunotypes having a relatively mild disease at days in ICU (Rs = −0.17, p = 0.08 and Rs = −0.18, p = 0.07, study entry, the EXI and LAI phenotypes captured patients that respectively). This supports the notion that immunotypes reflect would clinically deteriorate after hospitalization with higher the individual patient’s nature of the response and the rate of mortality, longer hospital stays and increased ICU days, while development of immunity rather than a strictly linear chron- BRI identified patients that would improve clinically. ological relationship to duration of infection. COVID-19 immunotypes are not defined by disease duration. Immunotypes differ in their clinical laboratory characteristics. Although the three immunotypes differ at time of entry in the The three immunotypes differed significantly in terms of plasma number of days from symptoms onset (DFSO), with disease inflammation markers and blood cell numbers (Table 2). Plasma 4 NATURE COMMUNICATIONS | (2022)13:915 | https://doi.org/10.1038/s41467-022-28621-0 | www.nature.com/naturecommunications NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-28621-0 ARTICLE A IL-6 TNFα CCL2 IL-8 **** * 10000 HC **** **** **** **** **** 10000 **** * ** **** Immunotype **** **** 30000 **** **** BRI **** 75 **** **** * *** **** EXI * 1000 1000 10000 LAI 50 3000 100 pg/mL 91 102 54 100 20 60 853 1000 10 600 38 26 25 420 19 280 300 16 1.5 10 8.9 12 1 Healthy control BRI EXI LAI Healthy control BRI EXI LAI Healthy control BRI EXI LAI Healthy control BRI EXI LAI Immunotype Immunotype Immunotype Immunotype IL-10 IFNα IFNγ TGFβ1 1000 **** **** 1000 ** 300000 **** 300 * **** **** **** * ** **** **** *** ** **** 100 100000 ** 100 100 38153 39893 36156 30 29 pg/mL 25524 8.3 10 30000 6.7 18 16 10 2.6 10 5.9 1 10000 0.6 3 2.3 1.5 1.5 1.5 1 1 1 Healthy control BRI EXI LAI Healthy control BRI EXI LAI Healthy control BRI EXI LAI Healthy control BRI EXI LAI Immunotype Immunotype Immunotype Immunotype B IgM IgG IgA 80 HC ** **** **** **** **** 300 **** Immunotype **** **** **** **** BRI 75 **** EXI 60 LAI 100 U/mL 50 40 47 43 33 30 23 25 20 18 11 11 11 11 11 11 11 1 10 1 1 1 1 Healthy control BRI EXI LAI Healthy control BRI EXI LAI Healthy control BRI EXI LAI Immunotype Immunotype Immunotype C Corr IL-10 −0.2−0.2−0.2−0.3−0.2 0.5 0.5 0.5 0.7 0.3 0.3 0.3 0.3 1.0 IL-18 0.2 0.2 0.2 0.4 0.4 0.4 0.2 0.2 0.3 0.5 IL-8 0.2 0.3 0.3 0.5 0.5 0.5 0.2 0.3 0.0 −0.5 IL-1β 0.2 0.4 0.5 0.5 0.2 0.3 −1.0 TNFα 0.2 0.3 0.1 0.5 0.4 0.5 0.5 0.4 0.3 IL-6 −0.1 0.3 0.2 0.6 0.4 0.4 0.5 0.4 0.7 CCL2 0.4 0.2 0.6 0.5 0.2 0.3 0.4 0.5 IFNα −0.4−0.4−0.4−0.3−0.4 0.5 0.2 0.2 0.1 0.2 0.5 IFNγ 0.5 0.4 0.3 0.3 0.2 0.5 IL-5 0.2 0.3 0.3 0.2 −0.4 0.2 −0.2 TGFβ1 0.2 0.3 0.3 0.2 −0.3 −0.1 −0.3 IgA 0.6 0.8 0.3 0.3 −0.4 0.2 0.3 −0.2 IgG 0.6 0.8 0.3 0.3 −0.4 0.2 −0.2 IgM 0.6 0.6 0.2 0.2 −0.4 −0.2 1 L2 Fα β 8 0 α γ -5 TN 6 -8 M G TG A Fβ -1 -1 N -1 - N Ig Ig Ig IL IL IL C IF IF IL IL IL C Fig. 2 COVID-19 immunotypes have distinct cytokine and anti-SARS-CoV-2 antibody levels. Immunotypes BRI, EXI and LAI have distinct patterns of cytokines and anti-SARS-CoV-2 antibodies. A Serum cytokines and B IgM, IgG, and IgA anti-SARS-CoV-2 antibodies are shown for BRI (n = 33), EXI (n = 46) and LAI patients (n = 59) and healthy controls (n = 14). Violin plots with single data points depict the range of levels and number indicates medians. Whiskers indicate the 1.5x interquartile range (IQR) value. A two-sided Wilcoxon rank-sum statistical test with Bonferroni correction was applied on all six comparisons but only shown for significant differences. Red dashed line indicates lower limit of detection. C Correlograms depicting correlations between cytokines and antibodies (n = 138 patients). Red indicates positive and blue negative correlation. Only statistically significant association (p < 0.05) are shown. Spearman’s rho (Rs) is shown in circles. *p < 0.05, **p < 0.01, ***p < 0.001, and ****p < 0.0001. levels of markers of inflammation C-reactive protein (CRP), Deep phenotyping of immunotypes reveals distinct cellular ferritin, d-dimer, Lactate dehydrogenase (LDH) were all sig- changes in blood of COVID-19 patients. To acquire insight into nificantly higher in EXI (Fig. 5A). Strikingly, these were not the potential mechanisms that are behind the described three increased in LAI compared to BRI. Leukocytes and neutrophils immunotypes we used 40-color spectral flow cytometry and iden- were highest in EXI, while lymphocyte levels trended to be lower tified the immune cell subset changes associated with these in EXI and LAI but were not significant (Fig. 5B). Thrombocytes, immunotypes (Fig. 6). Although COVID-19 patients as a whole however, were significantly lower in LAI patients (Fig. 5B and differed clearly from healthy individuals, comparing immunotypes Table 2). Viral loads in nasopharyngeal swabs were not sig- at study entry found remarkably few differences between them. nificantly different between immunotypes but trended to be COVID-19 patients compared to healthy individuals had increased higher in LAI (Fig. 5C), while within LAI patients, viral loads did plasmablasts (Fig. 6B–D and Supplementary Table S2). The pro- not correlate with IFNα levels (Fig. 5D). Finally, we found no inflammatory intermediate monocytes 1 and 210 and pro- correlation between viral load and DFSO (Rs = –0.18, p = 0.34). inflammatory IgD-negative non-conventional memory B cells NATURE COMMUNICATIONS | (2022)13:915 | https://doi.org/10.1038/s41467-022-28621-0 | www.nature.com/naturecommunications 5 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-28621-0 Table 2 Clinical and laboratory characteristics of BRI, EXI and LAI immunotypes. Level BRI EXI LAI p-value N 33 46 59 Cohort Barcelona 19 (57.6%) 33 (71.7%) 36 (61.0%) 0.367 Rotterdam 14 (42.4%) 13 (28.3%) 23 (39.0%) Gender Male 15 (45.5%) 36 (78.3%) 40 (67.8%) 0.009 Age Years 57 (49–64) 63 (57–69.75) 65 (52.50–71) 0.059 Days from symptom onset 10 (8–13) 10 (6.50–14) [n = 43] 7 (5–10) [n = 58] 0.002 WHO 8-point score at study entry 3 17 (51.5%) 5 (10.9%) 16 (27.1%) 0.002 4 11 (33.3%) 20 (43.5%) 29 (49.2%) 5 3 (9.1%) 8 (17.4%) 9 (15.3%) 6 0 (0.0%) 8 (17.4%) 4 (6.8%) 7 2 (6.1%) 5 (10.9%) 1 (1.7%) Outcome Deceased 0 (0.0%) 13 (28.3%) 14 (23.7%) 0.004 Discharged 33 (100%) 33 (71.7%) 45 (76.3%) Obesity Yes 6 (18.2%) 13 (28.3%) 24 (40.7%) 0.192 NA 1 (3.0%) 3 (6.5%) 2 (3.4%) Diabetes mellitus Yes 4 (12.1%) 11 (23.9%) 17 (28.8%) 0.113 NA 0 (0.0%) 2 (4.3%) 0 (0.0%) Heart disease Yes 6 (18.2%) 20 (43.5%) 22 (37.3%) 0.034 NA 0 (0.0%) 2 (4.3%) 0 (0.0%) Lung disease Yes 8 (24.2%) 11 (23.9%) 11 (18.6%) 0.311 NA 0 (0.0%) 2 (4.3%) 0 (0.0%) Kidney disease Yes 0 (0.0%) 4 (8.7%) 1 (1.7%) 0.049 NA 0 (0.0%) 2 (4.3%) 0 (0.0%) Liver disease Yes 1 (3.0%) 3 (6.5%) 2 (3.4%) 0.293 NA 0 (0.0%) 2 (4.3%) 0 (0.0%) Cancer Yes 2 (6.1%) 1 (2.2%) 4 (6.8%) 0.271 NA 0 (0.0%) 2 (4.3%) 0 (0.0%) Fever (>38 °C) Yes 15 (45.5%) 29 (63.0%) 47 (79.7%) 0.004 NA 0 (0.0%) 2 (4.3%) 0 (0.0%) C-reactive protein (mg/L) 82 (39–125) [n = 32] 159 (86–245) [n = 44] 94 (59–152) [n = 51] <0.001 Ferritin (µg/L) 434 (278–676) [n = 32] 887 (648–1883) [n = 38] 605 (313–1052) [n = 46] <0.001 Lactate dehydrogenase (U/L) 320 (257–381) [n = 32] 453 (362–599) [n = 39] 341 (296–454) [n = 50] <0.001 Alanine aminotransferase (U/L) 25 (18–54) 31 (18–57) [n = 46] 29 (21–53) [n = 58] 0.725 Bilirubin (μmol/L) 8.9 (6.9–13.0) [n = 31] 11.4 (8.0–14.8) [n = 44] 9.0 (6.0–12.2) [n = 47] 0.1 D-dimer (mg/L) 0.39 (0.30–0.61) 0.89 (0.32–2.22) [n = 39] 0.30 (0.18–0.62) <0.001 [n = 29] [n = 52] Hemoglobin (mmol/L) 8.01 (7.45–8.69) 8.30 (7.60–9.05) [n = 44] 8.38 (7.70–9.12) 0.352 Leukocytes (x109/L) 7.35 (6.10–9.30) 9.70 (7.10–12.33) [n = 44] 6.50 (4.84–8.06) <0.001 Neutrophils (x109/L) 5.11 (4.16–7.69) 7.78 (5.69–10.05) 4.78 (3.61–6.47) [n = 56] <0.001 [n = 44] Lymphocytes (x109/L) 1.40 (0.90–1.50) 0.99 (0.82–1.26) [n = 44] 1.00 (0.80–1.28) [n = 56] 0.067 Monocytes (x109/L) 0.48 (0.40–0.68) 0.50 (0.30–0.72) [n = 44] 0.38 (0.27–0.58) 0.008 [n = 56] Thrombocytes (x109/L) 280 (208–343) 274 (176–320) [n = 44] 182 (152–229) [n = 58] <0.001 Tocilizumab before study entry Yes 0 (0.0%) 6 (13.0%) 1 (1.7%) 0.01 NA 0 (0.0%) 0 (0.0%) 0 (0.0%) Tocilizumab after study entry Yes 1 (3.0%) 0 (0.0%) 1 (1.7%) 0.582 NA 19 (57.6%) 33 (71.7%) 36 (61.0%) Corticosteroids after study entry Yes 3 (9.1%) 4 (8.7%) 3 (5.1%) 0.476 NA 19 (57.6%) 33 (71.7%) 36 (61.0%) Convalescent plasma after study entry Yes 9 (27.3%) 5 (10.9%) 15 (25.4%) 0.356 NA 19 (57.6%) 33 (71.7%) 36 (61.0%) Data are n (%) or median (Q1–Q3). Complete data was available for the continuous values shown if not stated otherwise [n = ]; NA not available. Significance between immunotypes for categorical variables determined by Chi-squared test and for continuous variables by Kruskal–Wallis Rank-Sum Test. The clinical characteristics and laboratory measurements of patients in each immunotype group are presented. (IgD- CD27- B cells)11 were both increased in blood of patients strongly increased in EXI compared to LAI (Fig. 6F, G). Overall, (Fig. 6C; Supplementary Fig. S3 and Supplementary Table S2). In despite the large differences between COVID-19 patients and contrast, non-classical monocytes were reduced in blood of patients healthy controls, the differences between immunotypes were subtle, (Fig. 6C; Supplementary Fig. S3, Supplementary Table S2), and loss underscoring a disconnect between peripheral blood cells and sys- of these anti-inflammatory cells may further contribute to immune temic plasma cytokines in the immunotypes. activation12,13. Plasmacytoid dendritic cells (DC) and conventional DC were also both reduced in patients (Fig. 6B–D and Supple- mentary Table S2). When the three immunotypes were compared Discussion with each other, remarkably few differences existed (Fig. 6E, F and Our study does not aim to determine whether COVID-19 differs Supplementary Table S2). Only intermediate monocytes 2 were from other respiratory infections but seeks to identify groups of 6 NATURE COMMUNICATIONS | (2022)13:915 | https://doi.org/10.1038/s41467-022-28621-0 | www.nature.com/naturecommunications NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-28621-0 ARTICLE A Entry WHO score Peak WHO score B Survival C Total days in hospital **** **** 100 Outcome *** Immunotype Immunotype Deceased 150 **** BRI **** Discharged 8 **** BRI EXI EXI LAI % of patients ** 75 *** LAI 7 100 WHO score 6 6 6 50 Days 50 5 25 26 18 4 4 4 5 0 0 3 3 3 BRI EXI LAI BRI EXI LAI BRI EXI LAI Immunotype Immunotype Immunotype D Total days in ICU E Age F Sex *** Immunotype Immunotype ** 120 **** BRI BRI Female 100 EXI EXI Male LAI 80 LAI % of patients 75 80 Years 65 Days 63 60 57 50 40 40 25 5 0 0 0 BRI EXI LAI BRI EXI LAI 0 Immunotype Immunotype BRI EXI LAI Immunotype Fig. 3 Immunotypes predict clinical improvement or deterioration in acute COVID-19. COVID-19 immunotype BRI identifies at study entry patients that will clinically improve while immunotypes EXI and LAI predict clinical deterioration as peak severity scores during hospitalization are higher than entry scores. Violin plots and stacked bar graphs show per immunotype A WHO clinical score at study entry and peak WHO clinical score during hospitalization, B Percentage of deceased patients, C total days in the hospital, D total days in the ICU, E Age of patients and F Percentage of females and males. Data shown are from BRI (n = 33), EXI (n = 46), and LAI (n = 59) patients. Medians indicated by number in plots. Whiskers depict the 1.5x IQR value. Significance between the three immunotypes for the WHO scores was tested using a two-sided Wilcoxon rank-sum test with Bonferroni correction. Significance for each immunotype between entry WHO score and peak WHO score was tested using a paired Wilcoxon rank-sum test. Significant differences in total days in hospital and ICU were tested using two-sided Wilcoxon rank-sum tests with Bonferroni correction, significant differences in age were tested using two-sided Student’s t-tests with Bonferroni correction, significance difference in sex distribution was tested using a chi-square test; *p < 0.05, **p < 0.01, ***p < 0.001, and ****p < 0.0001. patients within COVID-19 patients that present a different clin- combined signature of multiple cytokines and antibodies has ical course. This study was conducted to find phenotypes or more power to predict disease course and outcomes. identify patient groups and characteristics that can facilitate The immunotypes described herein cannot be attributed to biomarker discovery, at hospital admission of COVID-19 strictly chronological differences of the duration of infection in patients, that predict clinical outcome. These phenotypes or patients. Instead of time since infection, the identified immuno- biomarkers could also suggest evidence-based treatment types reflect the variation in individual patient’s kinetics of approaches for the patients that will become severely ill. The mounting innate and adaptive immunity to the virus. This notion phenotypes we identified could thus help discover practical bio- is supported by the large variation in kinetics with which different markers. This focus on hospital entry is a very different approach hosts mount anti-SARS-CoV-2 immunity14. This variation could compared to other studies performed so far. Previous studies use be due to host factors such as different host genetics15, epigenetic the approach of grouping COVID-19 patients on the basis of differences imposed by varying host immunological experiences severity to understand mechanisms of disease, which is based on such as previous pathogen exposures or vaccination16, or the assumption that one pathophysiological mechanism underlies potential microbiome differences17. All these factors could lead to COVID-19 severity, and that patients respond uniformly to varying kinetics of immunity and control of viral loads, which can infection and display a similar linear disease progression with associate with disease severity and clinical course18,19. Non-host time. Using an unbiased analysis that only includes immune related factors such as dose or route of infection may also play a parameters can thus provide a more comprehensive mapping of significant role in immunotype development20. These elements the interplay between the immune response and SARS-CoV-2 could all influence the rapidity of mounting innate or adaptive virus in patients. Indeed, using this approach, we were able to immunity or even the duration of different phases of innate clearly demarcate immunotypes that predict clinical severity immunity. The question arises, whether there are more immu- changes and disease outcome. Importantly, the three immuno- notypes that can be revealed in larger cohorts of patients. Our types are able to predict disease severity progression, recovery or finding that small subgroups of patients in our cohorts have persistent severe disease. A previous study reported IL-6, IP-10 increased IL-17A or IL-5 in serum, points to additional immu- and IL-10 of first blood draw as predictors of COVID-19 clinical notypes that may have distinct clinical characteristics in acute deterioration while antibody responses negatively correlated with infection or associate with post-recovery sequelae of long deterioration7. However, in that study, 25% of patients that COVID-19. Our study of 138 patients did not have the power to clinically improve still have high levels of IP-10 and would be reveal such rarer immunotypes, but larger cohorts may well be classified as patients that will worsen7. Similarly, in our study, IL- able to do so. 6 and IL-10 on their own, do not separate well the three The three identified immunotypes reflect different pathophy- immunotypes as there is considerable overlap in serum levels. siological mechanisms for COVID-19. Based on serum markers Such overlap is also apparent in antibody responses where BRI of inflammation such as CRP and d-dimer, BRI and LAI are less and EXI have similar anti-SARS-CoV-2 antibody levels but inflammatory than EXI. However, LAI exhibits more thrombo- clearly very different clinical courses. Thus, we argue that the cytopenia compared to BRI and EXI. EXI and LAI had a much NATURE COMMUNICATIONS | (2022)13:915 | https://doi.org/10.1038/s41467-022-28621-0 | www.nature.com/naturecommunications 7 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-28621-0 DFSO A 40 Immunotype ** * BRI EXI 30 LAI Days 20 10 10 10 7 0 BRI EXI LAI Immunotype IFNα IL-6 TNFα B Immunotype Immunotype Immunotype BRI;rS: 0.089 BRI; rS: -0.045 BRI; rS: -0.018 100 EXI; rS: -0.071 1000 EXI; rS: -0.071 EXI; rS: -0.221 LAI; rS: -0.131 50 LAI; rS: 0.05 LAI; rS: -0.214 n 30 1 2 30 pg/ml pg/ml pg/ml 3 100 4 10 5 3 10 10 0 5 10 15 0 5 10 15 0 5 10 15 DFSO DFSO DFSO IgM IgG IgA C Immunotype Immunotype Immunotype BRI;rS: 0.115 BRI; rS: 0.075 BRI; rS: 0.329 EXI; rS: 0.285 100 EXI; rS: 0.059 EXI; rS: 0.208 50 LAI; rS: 0.099 LAI; rS: 0.189 LAI; rS: 0.135 50 n n n Units/ml Units/ml Units/ml 1 1 1 2 30 2 2 30 3 3 3 4 4 30 4 5 20 5 5 10 10 0 5 10 15 0 5 10 15 0 5 10 15 DFSO DFSO DFSO D IgM IgG IgA DFSO (days) 0−7 8−12 13−17 0−7 8−12 13−17 0−7 8−12 13−17 Negative 100 100 100 Positive 75 75 75 % of patients 50 50 50 25 25 25 0 0 0 I I I I I I I I I I I I I I I I I I I I I I I I I I I BR BR BR BR BR BR BR BR BR EX LA EX LA EX LA EX LA EX LA EX LA EX LA EX LA EX LA Immunotype Immunotype Immunotype Fig. 4 Disease duration does not determine immunotypes. Cytokine differences and the absence of antibodies in LAI are not due to lesser disease duration. A Days from symptoms onset (DFSO). Medians shown by number and whiskers indicate the 1.5x IQR value. Two-sided Wilcoxon rank-sum tests with Bonferroni correction were applied. *p < 0.05, **p < 0.01. B Scatterplots of IFNα, IL-6, and TNFα with DFSO per immunotype shown. C Scatterplots of anti-SARS-CoV-2 antibodies with DFSO is shown per immunotype. For B and C, Spearman’s rho is shown per immunotype and Bonferroni adjusted p-value is added when significant. Non-parametric local polynomial regression lines plotted using the locally estimated scatterplot smoothing (LOESS) method are shown for individual immunotypes. D The percentage of antibody positivity for the three immunotypes is depicted in relation to their DFSO. Red depict percent of patients positive for antibodies, blue depicts percent negative. Data shown are from BRI (n = 33), EXI (n = 46), and LAI (n = 59) patients. worse clinical course compared to BRI yet the signatures of EXI analysis to exclude that the different immunotypes we identified, and LAI differed in terms of their anti-SARS-CoV-2 antibody despite being sampled at hospital entry, could be due to some response, the level of pro-inflammatory cytokines and IFNα while patients being very early in the infection and therefore have not TGFβ1 was uniquely downregulated in LAI. Strikingly, these mounted an antibody response yet while other were later in the immunotypes, identified at hospital entry, predicted which course of infection. Furthermore, we also confirmed that the patients were less likely to survive, would have clinical severity immunotypes we identified remain stable over a range of DFSO. scores increase, would have longer hospitalizations and would Applying a number of different statistical approaches, we can spend more days in the ICU. exclude that duration of infection determines immunotypes at Gender is a known factor that can affect clinical severity in hospital admission. COVID-19 with females generally having milder disease. Analysis of the salient association between select pro- Although females were increased in BRI patients compared to inflammatory cytokines, interferons and anti-SARS-CoV-2 anti- EXI patients, BRI and LAI immunotypes were not statistically bodies, revealed clear associations of immunotypes with the different in terms of gender. Univariate and multivariate analysis adaptive immune response, inflammation and the anti-viral further excluded that the immunotype differences could be response (Fig. 7). BRI appears to mount an early anti-SARS-CoV- attributed to gender, age or DFSO. We included DFSO in our 2 antibody response that controls viral replication, and dampens 8 NATURE COMMUNICATIONS | (2022)13:915 | https://doi.org/10.1038/s41467-022-28621-0 | www.nature.com/naturecommunications NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-28621-0 ARTICLE A CRP D-dimer Ferritin LDH 500 ** * *** *** Immunotype ** *** * BRI * EXI 400 3000 LAI 10.0 4000 U/L mg/L 300 mg/L μg/L 1000 1.0 0.9 200 2000 160 0.4 453 0.3 340.5 319.4 100 887 94 300 82 605 0.1 434 0 0 BRI EXI LAI BRI EXI LAI BRI EXI LAI BRI EXI LAI Immunotype Immunotype Immunotype Immunotype B Neutrophiles Lymphocyes Thrombocytes C Viral load * *** Immunotype 30 **** *** BRI 3 100000000 EXI 750 LAI Copies/mL 20 x10⁹/L 2 100000 19648 x10⁹/L x10⁹/L 500 3661 1.4 10 280 274 100 1 1 1 7.8 250 182 6.7 5.1 4.8 BRI EXI LAI BRI EXI LAI BRI EXI LAI BRI EXI LAI Immunotype Immunotype Immunotype Immunotype D Immunotype LAI; rS: 0.372 100000000 Viral load (copies/mL) 1000000 10000 20 40 60 80 IFNα (pg/ml) Fig. 5 Immunotypes differ in their clinical laboratory characteristics. A Violin plots of CRP, d-dimer, ferritin and LDH measurements shown. B Violin plots of blood neutrophil, lymphocyte, and thrombocyte counts. Data from BRI (n = 33), EXI (n = 46), and LAI (n = 59) patients are shown. C Viral loads of immunotypes at entry shown (BRI n = 11, EXI n = 8, and LAI n = 16). Medians shown by number in plots. Whiskers indicate the 1.5x IQR value. D The scatterplot of IFNα and viral loads in LAI patients shown. Line depicts the non-parametric local polynomial regression line plotted using the locally estimated scatterplot smoothing (LOESS) method. Spearman’s rho (rs) is indicated. Correlation was not significant. A subgroup of Rotterdam patients (n = 35) with nasopharyngeal viral load measurements are shown in C and D. Two-sided Wilcoxon rank-sum tests with Bonferroni correction were applied for each measurement in A–C. *p < 0.05, **p < 0.01, ***p < 0.001, and ****p < 0.0001. the IFN response and subsequent hyper-inflammation. EXI This can set up a positive feedback loop between viral replication mounts a vigorous inflammatory response. LAI arises as a result and type I IFN in LAI patients and together with the anti-viral of a delayed adaptive immune response against SARS-CoV-2 that effect of type I IFN may set up a complex non-linear interplay leads to sustained high viral loads accompanied by a type I IFN between them. response. Clearly an impaired type I IFN response is detrimental We used 40-color spectral flow cytometry to identify the to the clinical course of COVID-19 as indicated by individuals immune cell subset changes associated with COVID-19 infection. with defects of IFN signaling and antibodies against IFN21–23. We find increased plasmablasts as previously described, but also However, increased type I IFN in serum has also been reported in increases in pro-inflammatory cell types such intermediate severe COVID-19 patients3,24 although others have reported low monocytes 1 and 2, and pro-inflammatory IgD- CD27- non- responses25. Such differences may be due to varying patient conventional memory B cells10,11, while anti-inflammatory cells populations and grouping of different immunotypes. Employing such as non-classical monocytes were reduced12,13. This under- single cell sequencing on just 21 patients, one study reported that scores the tipped balance towards inflammation in patients. anti- SARS-CoV-2 antibodies can hamper type I IFN responses in Plasmacytoid DC and conventional DC were both reduced in severe COVID-19 patients26. Although this could be the case in patients and this could impair both anti-viral and adaptive EXI patients, such results should be interpreted cautiously, as immune responses. Plasmablasts were increased in all three others have reported that antibodies protect from severe COVID- immunotypes irrespective if they had antibodies, arguing they are 197. Additionally, we clearly show that high type I IFN and not a good predictor of anti-SARS-CoV-2 immunity. We found absence of antibodies in LAI patients associates with disease only one difference between immunotypes, an increase in EXI deterioration and markedly worse outcomes, while BRI patients patients of pro-inflammatory intermediate monocytes 2 com- with high antibodies but no type I IFN do well clinically. The pared to LAI patients. Thus, COVID-19 patients had significant increased IFNα and viral loads in LAI, but their failure to cor- cellular variations compared to healthy controls, but such dif- relate (Fig. 5D) indicates that in this immunotype their interac- ferences between immunotypes were not apparent, indicating that tion is potentially deregulated. Such correlations can be seen the blood cellular compartment does not directly reflect plasma in vivo in viral infections such as SIV infection, where uncon- cytokines that may be produced locally in tissue such as lung. trolled viremia positively correlates with type I IFN at the chronic The classification of COVID-19 patients by immunotype may phase, yet negatively correlates at early SIV infection27. On the guide personalized therapeutic strategies. Pro-inflammatory other hand, type I IFN can induce ACE2 receptor expression28. cytokines do not distinguish clearly BRI, EXI and LAI from NATURE COMMUNICATIONS | (2022)13:915 | https://doi.org/10.1038/s41467-022-28621-0 | www.nature.com/naturecommunications 9 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-28621-0 A Overlay (n=44) Healthy controls (n=12) Patients (n=32) B umap_2 umap_2 umap_2 umap_1 umap_1 umap_1 CD123+ pDCs Plasmablasts CD123+ pDCs Plasmablasts C COVID-19 patients/Healthy controls ratio D Plasmablasts CD123+ pDCs Plasmablasts Significant n=14 FDR=4.2x10-13 FDR=3.1x10-5 Healthy Controls (n=12) 12 Not significant Patients (n=32) Threshold (FDR 0.1) 10 Intermediate % of total 8 Monocytes 1 -Log10(FDR) HLA-DR-CD38+ basophils 6 CD123+ pDCs 4 Non-classical Monocytes ILCs Intermediate 2 Monocytes 2 0 −5 0 5 logFC BRI (n=9) EXI (n=6) LAI (n=17) E umap_2 umap_2 umap_2 umap_1 umap_1 umap_1 F EXI/BRI ratio LAI/BRI ratio Intermediate LAI/EXI ratio 1 1 Monocytes 2 Significant 2.5 Not significant Threshold (FDR 0.1) 0.8 0.8 2 -Log10(FDR) -Log10(FDR) -Log10(FDR) 0.6 0.6 1.5 0.4 0.4 1 0.2 0.5 0.2 0 0 0 −2 0 2 −2 −1 0 1 2 −4 −2 0 2 4 logFC logFC logFC G Intermediate monocytes 2 FDR=0.003 EXI (n=6) LAI (n=17) % of total Fig. 6 High-dimensional flow cytometry reveals distinct immune subset differences between immunotypes. A Population cluster identification in high- dimensional 40-color flow cytometry data using uniform manifold approximation and projection (UMAP) dimensionality reduction. Visualization is performed on combined COVID-19 patients and healthy control data (n = 44). B UMAP plots of healthy controls (n = 12) and all COVID-19 patients (n = 32). Plasmablast and pDC indicated by arrows. C Volcano plots of the comparison COVID-19 patients/Healthy controls. Positive fold-changes indicate cellular population increases in patients while negative fold-change indicate decreases in patients. D Plasmablast and pDC frequencies in healthy controls and COVID-19 patients show. Violin plots depict the percentage of cells in the total population. False-detection rate (FDR) of population comparisons for plasmablasts is 4.2 × 10−13 and for pDC 3.1 × 10−5. E UMAP plots of the three immunotypes (BRI n = 9, EXI n = 6, LAI n = 17). F Volcano plots of the pairwise comparison of immunotypes. Positive fold-change indicates increase in numerator of ratio, while negative fold-change indicates decreases. All patients are from Rotterdam cohort. G Intermediate monocyte 2 frequencies in EXI and LAI patients show. Violin plots depict the percentage of cells in the total population. The violin shapes were defined by kernel density estimation the boxes define the Q1 to Q3 range while box whiskers indicate min/max. The mean and median are indicated by dotted and solid middle lines, respectively. FDR of population comparison is 0.003. 10 NATURE COMMUNICATIONS | (2022)13:915 | https://doi.org/10.1038/s41467-022-28621-0 | www.nature.com/naturecommunications NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-28621-0 ARTICLE muted antibody responses do worse clinically in the future. Such patients may benefit the most from anti-SARS-CoV-2 antibody therapy. Although convalescent plasma therapy has failed to show clear benefit in COVID-1943,44, this could be due to most patients already having antibodies. Treatment with convalescent plasma of specific populations at risk very early in the disease course have suggested benefit in terms of preventing clinical deterioration45. Therefore, patients within the LAI group, that have no or low anti-SARS-CoV-2 antibodies may stand to benefit from high neutralizing antibody titer plasma or monoclonal antibodies. Finally, the observation that very few patients in the BRI immunotype progressed to more severe disease could help hos- pitals overwhelmed with COVID-19 patients to triage which patients could be transferred to step-down units outside the hospital. Immunotypes, therefore, could guide the formulation of personalized therapies for COVID-19 patients based on mechanistic evidence. Our study proposes immunotypes that are free from the pre- sumption that clinical classification should dominate the analysis and could be the basis for prospective studies. From a clinical perspective, it is important to identify at an early stage which patients may progress in disease severity as the clinical condition of COVID-19 patients can rapidly deteriorate within days. Here, we have identified immunotypes that predict disease progression but also shed light on underlying pathways and suggest bio- markers and therapeutic targets before severe COVID-19 devel- ops. Our current study only addressed the immunotypes in the context of acute COVID-19, but it also points towards employing larger cohort studies that may reveal immunotypes that predict long term post-COVID-19 complications. As immunotype identification requires only serum analysis, these immunotypes can be determined rapidly after patient admission and help instruct personalized therapy. Methods Patients. Rotterdam cohort samples were collected from patients (n = 50) parti- cipating in the ConCOVID nationwide multicenter open-label randomized clinical trial in the Netherlands. Inclusion criteria were patient age of at least 18 years, admittance to the hospital for COVID-19 and SARS-CoV-2 genome positive by reverse transcription–polymerase chain reaction (RT-PCR) test in the previous 96 h. All patients entered the study and were sampled at a median of 2 days after hospitalization (Q1, Q3: 1, 3.75 days). No exclusion criteria were applied. Healthy controls were age and sex matched. The study was reviewed and approved by the institutional review board of the Erasmus University Medical Center. Written informed consent was obtained from every patient or legal representative. All samples were processed and frozen within three hours of bleeding. Barcelona cohort samples and data from patients included in this study were Fig. 7 Unsupervised clustering defines three immunotypes at hospital provided by the Hospital Universitari Vall d’Hebron (HUVH) Biobank (PT17/ entry that predict clinical outcome. Anti-SARS-CoV2 nucleocapsid 0015/0047), integrated in the Spanish National Biobanks Network. The collection antibodies (IgM, IgG, IgA) and the most prominent anti- and pro- and secondary use of de-identified diagnostic samples and data for research was approved by the Ethics Committee for Research with Medicines of HUVH and inflammatory cytokines out of n = 14 shown. BRI: balanced response consent was waived. Sample and data transfer to Erasmus MC for research purpose immunotype; EXI: excessive inflammation immunotype; LAI: low antibody was approved by the Ethics Committee for Research with Medicines of HUVH. immunotype; SOC: standard of care. Patient inclusion criteria were an age of at least 18 years, having a confirmed SARS- CoV-2 genome positive by RT-PCR test and having a serum sample on the day of hospitalization. No exclusion criteria were applied. Clinical severity was classified each other and this may explain the relatively small benefit and according to the WHO 8 point COVID-19 disease severity score (at study inclusion mixed results of anti-IL-6 antibody treatment29–31. Using for patients and during hospitalization) in which 0 is no clinical or virological immunotypes to predict who will develop a hyperinflammatory evidence of infection, 1 is no limitation of activities, 2 is limitation of activities, 3 is syndrome32,33 may enable pre-emptively blocking key drivers of hospitalized, no oxygen, 4 is oxygen by mask or nasal prongs, 5 is non-invasive ventilation or high-flow oxygen, 6 is intubation and mechanical ventilation, 7 is the pro-inflammatory cytokine network. This would provide the ventilation and additional organ support (vasopressors, renal replacement therapy, maximal clinical benefit as dampening of the cytokine storm is ECMO), and 8 is death9. Clinical characteristics, laboratory measurements and achieved when anti-cytokine treatment is applied early34–37. The treatments for both cohorts are shown in Table 1. potential benefit of prophylactic anti-cytokine treatment for Days from symptom onset (DFSO) were determined by asking patients at study inclusion when they experienced the first symptoms of COVID-19 infection. None COVID-19 is also suggested from autoimmune patients in which of the patients were vaccinated against SARS-CoV-2 spike protein. anti-cytokine therapy associates with milder COVID-19 disease38–40. Longitudinal studies have shown anti-spike IgA Cytokine measurements. The cytokines of interest (IL-6, TNFα, IL-1β, IL-8, and antibody affinity can correlate with disease severity41,42 and CCL2, IL-18, IL-10, IL-12, IFNγ, IL-5, IL-17A, IL-2, IL-4, IFNα, and TGFβ1) were this could be driven by persistent antigen. In contrast, our study analyzed using the ELLA Simple Plex system (Protein simple, San Jose, CA). After that samples patients at hospital entry, reveals that patients with thawing serum samples on ice, they were centrifuged at 1300 x g for 5 min at room NATURE COMMUNICATIONS | (2022)13:915 | https://doi.org/10.1038/s41467-022-28621-0 | www.nature.com/naturecommunications 11 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-28621-0 temperature. Twofold dilutions were prepared in low-protein-binding plates transformed data using a univariate binary logistic regression of BRI vs. EXI, BRI according to the manufacturer’s instructions. For TGFβ1, samples were first acti- vs. LAI and EXI vs. LAI. In addition, we applied on the same variables a multi- vated with 1 N HCl and then neutralized with 1.2 N NaOH/0.5 M HEPES. Sub- variate analysis with gender, age, and ln(DFSO + 1) as additional variables. Two sequently, these samples were diluted in a factor 1:15. Diluted samples were loaded patients with no CCL2 measurement were dismissed from analyses while two into ELLA Simple Plex cartridges and analyzed with the ELLA Simple Plex system. patients with no DFSO indication were removed in the multivariate analyses. Wilcoxon rank-sum/Mann–Whitney U-test p-values were adjusted using Bonfer- roni correction on the three comparisons. Anti-SARS-CoV-2 antibody measurements. Anti-SARS-CoV-2 IgM, IgG and IgA antibodies against nucleocapsid protein (N-protein) were measured in serum by ELISA using CE-certified COVID-19 IgG ELISA (Tecan, 30177447), COVID-19 Reporting summary. Further information on research design is available in the Nature IgA ELISA (Tecan, 30177446) and COVID-19 IgM ELISA (Tecan, 30177448) Research Reporting Summary linked to this article. according to the manufacturer’s instructions. Positive cutoff for these ELISAs was 11 units. The anti-nucleocapsid ELISA was chosen as it allows the detection of anti- SARS-CoV-2-specific IgM, IgG, and IgA isotypes and circumvents SARS-CoV-2 Data availability The data used in this study are available on request from the corresponding author PDK. spike vaccination status in future studies. Serum was tested for the presence of SARS-CoV-2 total Ig and IgM antibodies against spike RBD by ELISA test (Wantai The data are not publicly available due to participant privacy/consent. Source data are Biological, Beijing) while neutralizing antibodies were measured by performing a provided with this paper. plaque-reduction neutralization test (PRNT) with the SARS-CoV-2 virus. Both assays are described in Gharbharan et al.43. Code availability The R-code used to cluster and the statistical analysis the cytokines, antibodies and High-dimensional flow cytometry. PBMC from patients and controls of the clinical data can be freely downloaded from https://bitbucket.org/immunology-emc/ Rotterdam cohort were stained with a 40-color antibody panel and collected using covid_severity_publication/src/master/. a 5 laser Aurora spectral flow cytometer (Cytek Biosciences, CA). All samples were stained as described previously46 with the adaptation of including annexin V to Received: 28 September 2021; Accepted: 1 February 2022; exclude dead cells and all buffers contained all buffers contained 2.5 mM CaCl2. Cleaning of flow data was performed in SpectroFlo (version 2.2.0) (Supplementary Fig. S4). The unsupervised, and statistical inference portions of the flow cytometry analysis were performed using OMIQ data analysis software(www.omiq.ai). We used the unsupervised analysis methods based on surface markers without any 2D gating we previously employed46. The workflow included running flowCut to check for changes in channels over acquisition time, UMAP for dimensionality reduction, References flowSOM for clustering, and edgeR for statistical inference. For the statistical 1. World Health Organization. WHO Coronavirus Disease (COVID-19) comparisons of abundance, the Flow Cytometry Standard (FCS) files were sub- Dashboard (WHO, 2021). sampled to ensure the same number of events were included per group (either 2. Mathew, D. et al. Deep immune profiling of COVID-19 patients reveals immunotype or disease state). distinct immunotypes with therapeutic implications. Science 369, eabc8511 (2020). 3. Lucas, C. et al. Longitudinal analyses reveal immunological misfiring in severe Unsupervised hierarchical clustering and principal component analysis. COVID-19. Nature 584, 463–469 (2020). Patient’s clinical and serum data from the Erasmus MC and Barcelona cohort were 4. Arunachalam, P. S. et al. Systems biological assessment of immunity to mild loaded into R (v.4.0.4)47. Only cytokines positive in >20% of patients in both versus severe COVID-19 infection in humans. Science 369, 1210–1220 (2020). cohorts were used in the analysis. Standardized scores (z-scores) for each cohort 5. Wilk, A. J. et al. A single-cell atlas of the peripheral immune response in were calculated based on the log10 transformed cytokine and antibody levels patients with severe COVID-19. Nat. Med. 26, 1070–1076 (2020). measured in serum. Unsupervised hierarchical clustering was performed using 6. Abers, M. S. et al. An immune-based biomarker signature is associated with Ward’s Hierarchical Agglomerative Clustering Method (ward.d2) on the Euclidean mortality in COVID-19 patients. JCI Insight 6, e144455(2021). distances of the z-scores48. The optimal number of clusters for both cohorts was 7. Laing, A. G. et al. A dynamic COVID-19 immune signature includes assigned with the NbClust (v1.0.12) package in R49. NbClust provided 30 indices associations with poor prognosis. Nat. Med. 26, 1623–1635 (2020). that determined the best number of clusters for both datasets and proposed the 8. Szabo, P. A. et al. Longitudinal profiling of respiratory and systemic immune appropriate number of clusters using the majority rule. The majority rule indicated responses reveals myeloid cell-driven lung inflammation in severe COVID-19. three clusters as optimal for both cohorts. Subsequently, heatmaps were plotted using the R package pheatmap (v1.0.12). Principal component analysis (PCA) was Immunity 54, 797–814.e6 (2021). performed and visualized based on the z-scores using the R function prcomp. The 9. World Health Organization. WHO R&D Blueprint novel Coronavirus, grouping of patients in these different clusters was not influenced in any way by the COVID-19 Therapeutic Trial Synopsis (WHO, 2020). investigators. 10. Cormican, S. & Griffin, M. D. Human monocyte subset distinctions and We tested the reliability of our hierarchical clusters by constructing a function: insights from gene expression analysis. Front Immunol. 11, 1070 multinomial logit model50. The model describes ln(Prob(EXI) over Prob(BRI)), (2020). and ln(Prob(LAI) over Prob(BRI)) in relation with the variables used in building 11. Claes, N. et al. Age-associated B cells with proinflammatory characteristics are trees in Fig. 2. For a better coherence with the method used in building the three expanded in a proportion of multiple sclerosis patients. J. Immunol. 197, categories, we log-transformed the 14 cytokines and antibody variables as we used 4576–4583 (2016). to construct the trees in Fig. 2, while removing the patients with no CCL2 data 12. Patel, A. A. et al. The fate and lifespan of human monocyte subsets in steady (BRI = 32, EXI = 46, LAI = 58). The model was 100% accurate in predicting the state and systemic inflammation. J. Exp. Med. 214, 1913–1923 (2017). immunotypes based on the input data. We estimated the stratification accuracy 13. Narasimhan, P. B., Marcovecchio, P., Hamers, A. A. J. & Hedrick, C. C. from a predictive perspective cross-validation by splitting the dataset into a test and Nonclassical monocytes in health and disease. Annu. Rev. Immunol. 37, training set, increasing test percentages from 5%, 10%, 15%, 20%, and 25% using 439–456 (2019). 150 iterations. 14. Huang, A. T. et al. A systematic review of antibody mediated immunity to coronaviruses: kinetics, correlates of protection, and association with severity. Nat. Commun. 11, 4704 (2020). Statistical analysis. Statistical analysis was performed using R (v.4.0.4)47. Nor- 15. Casanova, J. L. & Abel, L. The human genetic determinism of life-threatening mality of the patients’ data was tested using a Shapiro–Wilk normality test. Sta- infectious diseases: genetic heterogeneity and physiological homogeneity? tistically significant differences between immunotypes were calculated using two- Hum. Genet. 139, 681–694 (2020). sided multiple Students t-tests for patient’s age and Wilcoxon rank-sum tests for all the other variables. Subsequently, p-values were adjusted using Bonferroni cor- 16. Bekkering, S., Dominguez-Andres, J., Joosten, L. A. B. Riksen, N. P. & Netea, rection. Adjusted p-values lower than 0.05 were considered statistically significant. M. G. Trained immunity: reprogramming innate immunity in health and The number of asterisks indicate the level of significance of p-adjusted values: disease. Annu. Rev. Immunol. 39, 667–-693 (2021). *p < 0.05, **p < 0.01, ***p < 0.001, and ****p < 0.0001. Non-significant results are 17. Lynch, S. V. & Pedersen, O. The human intestinal microbiome in health and not shown in figures. Correlations coefficients between the variables in each cohort disease. N. Engl. J. Med. 375, 2369–2379 (2016). were calculated using a Spearman’s rank correlation coefficient and presented in 18. Liu, Y. et al. Viral dynamics in mild and severe cases of COVID-19. Lancet correlograms and DFSO correlations. A correlation was considered statistically Infect. Dis. 20, 656–657 (2020). significant if the p-value was lower than 0.05. Non-parametric local polynomial 19. Wolfel, R. et al. Virological assessment of hospitalized patients with COVID- regression lines were plotted using the locally estimated scatterplot smoothing 2019. Nature 581, 465–469 (2020). (LOESS) method. We applied a Uni- and Multivariate analyses next to our Wil- 20. Calisti, R. SARS-CoV-2: exposure to high external doses as determinants of coxon rank-sum/Mann–Whitney U-test that directly compared cytokine, anti- higher viral loads and of increased risk for COVID-19. A systematic review of bodies and other variables in this study. We applied these analyses of the log- the literature. Epidemiol. Prev. 44, 152–159 (2020). 12 NATURE COMMUNICATIONS | (2022)13:915 | https://doi.org/10.1038/s41467-022-28621-0 | www.nature.com/naturecommunications NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-28621-0 ARTICLE 21. Zhang, Q. et al. Inborn errors of type I IFN immunity in patients with life- 47. Team, R.C. R: A language and environment for statistical computing. (ed. R threatening COVID-19. Science 370, eabd4570 (2020). Foundation for Statistical Computing, V., Austria.) (Team, R.C., 2021). 22. Pairo-Castineira, E. et al. Genetic mechanisms of critical illness in COVID-19. 48. Murtagh, F. & Legendre, P. Ward’s hierarchical agglomerative clustering Nature 591, 92–98 (2021). method: which algorithms implement ward’s criterion? J. Classification 31, 23. Bastard, P. et al. Autoantibodies against type I IFNs in patients with life- 274–295 (2014). threatening COVID-19. Science 370, eabd4585 (2020). 49. Charrad, M., Ghazzali, N., Boiteau, V. & Niknafs, A. Nbclust: an R package for 24. Galani, I. E. et al. Untuned antiviral immunity in COVID-19 revealed by determining the relevant number of clusters in a data set. J. Stat. Softw. 61, temporal type I/III interferon patterns and flu comparison. Nat. Immunol. 22, 1–36 (2014). 32–40 (2021). 50. Venables, W. N., Ripley, B. D. & Venables, W. N. Modern Applied Statistics 25. Blanco-Melo, D. et al. Imbalanced host response to SARS-CoV-2 drives With S, (Springer, New York, 2002). development of COVID-19. Cell 181, 1036–1045 e1039 (2020). 26. Combes, A. J. et al. Global absence and targeting of protective immune states in severe COVID-19. Nature 591, 124–130 (2021). Acknowledgements 27. Fraietta, J. A. et al. Type I interferon upregulates Bak and contributes to T cell We would like to thank all the patients who participated in the study. We thank all loss during human immunodeficiency virus (HIV) infection. PLoS Pathog. 9, physicians, medical students, study nurses and trial coordinators in the participating e1003658 (2013). centers. We would like to thank Bart Haagmans and Corine Geurts van Kessel from the 28. Ziegler, C. G. K. et al. SARS-CoV-2 receptor ACE2 is an interferon-stimulated Viroscience department, Erasmus MC for serum anti-RDB and PRNT50 antibody data of gene in human airway epithelial cells and is detected in specific cell subsets CONCOVID trial43. We want to particularly acknowledge the patients and the HUVH across tissues. Cell 181, 1016–1035 e1019 (2020). Biobank (PT17/0015/0047) integrated in the Spanish National Biobanks Network for 29. Stone, J. H. et al. Efficacy of tocilizumab in patients hospitalized with Covid- their collaboration. Thanks to BMW, the Netherlands for providing two free-of-charge 19. N. Engl. J. Med. 383, 2333–2344 (2020). vehicles for sample collection. This work was supported by Health Holland LSHM20056 30. Rosas, I. O. et al. Tocilizumab in hospitalized patients with severe covid-19 grant (PDK), in part from the European Union’s Horizon 2020 research and innovation pneumonia. N. Engl. J. Med. 384, 1503–1516 (2021). program under grant agreement No 779295 (PDK), in part supported by the Erasmus 31. Investigators, R.-C. et al. Interleukin-6 receptor antagonists in critically ill foundation (BJAR), grant PI20/00416 from the Instituto de Salud Carlos III (RPB) and patients with Covid-19. N. Engl. J. Med. 384, 1491–1502 (2021). the European Regional Development Fund (ERDF) (RPB). 32. Jose, R. J. & Manuel, A. COVID-19 cytokine storm: the interplay between inflammation and coagulation. Lancet Respir. Med. 8, e46–e47 (2020). Author contributions 33. Mehta, P. et al. COVID-19: consider cytokine storm syndromes and Y.M.M., T.J.S., R.R., M.W.J.S., D.A.S., C.H.K., M.D.C.E., M.v.M., I.B.H., M.Z., L.L., immunosuppression. Lancet 395, 1033–1034 (2020). H.d.W., C.A.O., M.E.P.W., T.M.A., D.A.L., T.v.W., M.C.J., G.K., S.v.B., C.R., B.J.A.R., 34. Karki, R. et al. Synergism of TNF-alpha and IFN-gamma triggers M.H.G., R.P.B., contributed to data acquisition and analysis. D.G.B.G., G.J.L., and inflammatory cell death, tissue damage, and mortality in SARS-CoV-2 H.J.G.v.d.W. performed machine learning, bioinformatics, and statistical analysis. P.D.K. infection and cytokine shock syndromes. Cell 184, 149–168 e117 (2021). conceived and designed the work, and wrote the manuscript. All authors reviewed and 35. Bengtsson, A. et al. Anti-TNF treatment of baboons with sepsis reduces TNF- edited the manuscript. alpha, IL-6 and IL-8, but not the degree of complement activation. Scand. J. Immunol. 48, 509–514 (1998). 36. Tracey, K. J. et al. Anti-cachectin/TNF monoclonal antibodies prevent septic Competing interests shock during lethal bacteraemia. Nature 330, 662–664 (1987). The authors declare no competing interests. 37. Fong, Y. et al. Antibodies to cachectin/tumor necrosis factor reduce interleukin 1 beta and interleukin 6 appearance during lethal bacteremia. J. Exp. Med. 170, 1627–1633 (1989). Additional information Supplementary information The online version contains supplementary material 38. Gianfrancesco, M. et al. Characteristics associated with hospitalisation for available at https://doi.org/10.1038/s41467-022-28621-0. COVID-19 in people with rheumatic disease: data from the COVID-19 global rheumatology alliance physician-reported registry. Ann. Rheum. Dis. 79, Correspondence and requests for materials should be addressed to Peter D. Katsikis. 859–866 (2020). 39. Akiyama, S., Hamdeh, S., Micic, D. & Sakuraba, A. Prevalence Peer review information Nature Communications thanks the anonymous reviewer(s) for and clinical outcomes of COVID-19 in patients with autoimmune their contribution to the peer review of this work. Peer reviewer reports are available. diseases: a systematic review and meta-analysis. Ann. Rheum. Dis. 80: 384–391 (2021). Reprints and permission information is available at http://www.nature.com/reprints 40. Sarmiento-Monroy, J. C. et al. A multidisciplinary registry of patients with autoimmune and immune-mediated diseases with symptomatic COVID-19 Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in from a single center. J. Autoimmun. 117, 102580 (2021). published maps and institutional affiliations. 41. Ravichandran, S. et al. Longitudinal antibody repertoire in “mild” versus “severe” COVID-19 patients reveals immune markers associated with disease severity and resolution. Sci. Adv. 7, eabf2467 (2021). 42. Tang, J. et al. Antibody affinity maturation and plasma IgA associate with Open Access This article is licensed under a Creative Commons clinical outcome in hospitalized COVID-19 patients. Nat. Commun. 12, 1221 Attribution 4.0 International License, which permits use, sharing, (2021). adaptation, distribution and reproduction in any medium or format, as long as you give 43. Gharbharan, A. et al. Effects of potent neutralizing antibodies from appropriate credit to the original author(s) and the source, provide a link to the Creative convalescent plasma in patients hospitalized for severe SARS-CoV-2 infection. Commons license, and indicate if changes were made. The images or other third party Nat. Commun. 12, 3189 (2021). material in this article are included in the article’s Creative Commons license, unless 44. Kunze, K. L. et al. Mortality in individuals treated with COVID-19 indicated otherwise in a credit line to the material. If material is not included in the convalescent plasma varies with the geographic provenance of donors. Nat. article’s Creative Commons license and your intended use is not permitted by statutory Commun. 12, 4864 (2021). regulation or exceeds the permitted use, you will need to obtain permission directly from 45. Libster, R. et al. Early high-titer plasma therapy to prevent severe Covid-19 in the copyright holder. To view a copy of this license, visit http://creativecommons.org/ older adults. N. Engl. J. Med. 384, 610–618 (2021). licenses/by/4.0/. 46. Park, L. M., Lannigan, J. & Jaimes, M. C. OMIP-069: forty-color full spectrum flow cytometry panel for deep immunophenotyping of major cell subsets in human peripheral blood. Cytom. A 97, 1044–1051 (2020). © The Author(s) 2022 NATURE COMMUNICATIONS | (2022)13:915 | https://doi.org/10.1038/s41467-022-28621-0 | www.nature.com/naturecommunications 13