Coronavirus disease 2019 subphenotypes and differential treatment response to convalescent plasma in critically ill adults: secondary analyses of a randomized clinical trial.

PURPOSE: Benefit from convalescent plasma therapy for coronavirus disease 2019 (COVID-19) has been inconsistent in randomized clinical trials (RCTs) involving critically ill patients. As COVID-19 patients are immunologically heterogeneous, we hypothesized that immunologically similar COVID-19 subphenotypes may differ in their treatment responses to convalescent plasma and explain inconsistent findings between RCTs .
METHODS: We tested this hypothesis in a substudy involving 1239 patients, by measuring 26 biomarkers (cytokines, chemokines, endothelial biomarkers) within the randomized, embedded, multifactorial, adaptive platform trial for community-acquired pneumonia (REMAP-CAP) that assigned 2097 critically ill COVID-19 patients to either high-titer convalescent plasma or usual care. Primary outcome was organ support free days at 21 days (OSFD-21) .
RESULTS: Unsupervised analyses identified three subphenotypes/endotypes. In contrast to the more homogeneous subphenotype-2 (N = 128 patients, 10.3%; with elevated type i and type ii effector immune responses) and subphenotype-3 (N = 241, 19.5%; with exaggerated inflammation), the subphenotype-1 had variable biomarker patterns (N = 870 patients, 70.2%). Subphenotypes-2, and -3 had worse outcomes, and subphenotype-1 had better outcomes with convalescent plasma therapy compared with usual care (median (IQR). OSFD-21 in convalescent plasma vs usual care was 0 (- 1, 21) vs 10 (- 1, to 21) in subphenotype-2; 1.5 (- 1, 21) vs 12 (- 1, to 21) in suphenotype-3, and 0 (- 1, 21) vs 0 (- 1, to 21) in subphenotype-1 (test for between-subphenotype differences in treatment effects p = 0.008).
CONCLUSIONS: We reported three COVID-19 subphenotypes, among critically ill adults, with differential treatment effects to ABO-compatible convalescent plasma therapy. Differences in subphenotype prevalence between RCT populations probably explain inconsistent results with COVID-19 immunotherapies.

Take-home message

Introduction

The coronavirus SARS-CoV-2 (severe acute respiratory syndrome coronavirus 2) causes COVID-19 (coronavirus disease 2019), an acute illness affecting pulmonary and extrapulmonary organs [1]. COVID-19 patients requiring hospitalization (moderate-to-severe disease) have significant viral load in the respiratory tract [2] and/or detectable viral ribonucleic acid (RNA) in blood [3]. Therefore, in moderate-to-severe COVID-19, antiviral therapies (either passive immunotherapy or antiviral medications) are considered potential treatment options [4]. The benefit of passive immunotherapy with convalescent plasma (blood product containing SARS-CoV-2–specific polyclonal antibodies) reported in cohort studies have not been observed in randomised clinical trials (RCTs), with guidelines recommending against the use of convalescent plasma outside of RCTs [5].

It is well recognized that in hospitalized patients, COVID-19 is an immunologically heterogenous illness [6-13]. It is also recognized that immunological heterogeneity in COVID-19 patients is observable at protein level, i.e., differences in cytokine, chemokines, and other biomarker profiles [8, 13]. Abnormal immune responses persist throughout critical illness in COVID-19 patients [13]. Thus, we hypothesized that differences in immune responses will manifest as subphenotypes and may be associated with subphenotype differences in treatment effect to convalescent plasma therapy (also known as heterogeneity of treatment effect (HTE) [14]) on OSFD-21 (primary outcome of pandemic appendix to REMAP-CAP) [15] and on hospital mortality (outcome of interest).

We tested these hypotheses in a biological sampling substudy conducted in the United Kingdom (UK) within the immunoglobulin domain of the REMAP-CAP, which randomized 2097 patients with severe COVID-19 into two units of high-titer convalescent plasma or usual care, and found no overall benefit with convalescent plasma therapy in critically ill COVID-19 patients [15]. Informed by previous work on protein biomarkers [6–13, 16], we explored whether differences in treatment effect (HTE [14] on OSFD-21 and mortality outcomes) was detectable between COVID-19 subphenotypes that were identified based on unsupervised analyses of changes in the CXC family of chemokines (CXCL1, CXCL5, CXCL8, CXCL9, CXCL10, CXCL11), the CC family of chemokines (CCL3, CCL4, CCL11, CCL17, CCL20), transforming growth factor-beta 1(TGF-β1), vascular endothelial growth factor (VEGF), interleukins (IL-6, IL-2, IL-4, IL-5, IL-10), interferons (IFN-α2, IFN-β, IFN-λ1, IFN-y), granulocyte monocyte colony stimulating factor (GM-CSF), soluble tumor necrosis factor receptor-1 (sTNF-R1), angiopoietin-2, intercellular adhesion molecule-1 (ICAM-1), A proliferation-inducing ligand (APRIL, TNFSF13) and B cell-activating factor (BAFF, TNFSF13B) [17]. We selected these biomarkers a priori, i.e., before our primary RCT results were available.

Methods

Study design

Briefly, REMAP-CAP is an international, multicentre, open-label adaptive platform designed to determine the best treatment strategies for patients with severe pneumonia in both settings during the pandemic and outside the pandemic [18]. This trial’s design, eligibility criteria, and results regarding glucocorticoids [19], anticoagulants [20, 21], antivirals [22], interleukin-6 (IL-6) receptor antagonists [23], antiplatelet therapy, and immunoglobulin domain convalescent plasma [15] for treatment of COVID-19 have been reported previously.

Study population

Our study population consisted of critically ill adult patients (> 18 years old) with microbiologically confirmed COVID-19, randomized to receive 2 units of high-titer, ABO-compatible convalescent plasma (total volume approximately 550 mL ± 150 mL) within 48 h of randomization or no convalescent plasma, between March 9, 2020 and January 18, 2021 [15]. These patients had a baseline blood sample collected after consent and before administration of the allocated intervention (convalescent plasma vs no convalescent plasma (usual care)).

Biomarker measurements

Serum was separated from whole blood by centrifugation (1300 g for 10 min at room temperature) and stored in 200 µl aliquots at − 80 °C until analyses. Two custom 14-plex Legendplex™ (BioLegend) bead-based multiplex assays were used to measure a priori selected biomarkers described in “Introduction”, as per manufacturer’s instructions (eMethods-1). SARS-CoV-2 immunoglobulin G (IgG) antibody against spike was measured using enzyme-linked immunosorbent assay (ELISA), as reported previously [2, 24]. Viral loads and strains in the respiratory tract were measured as described previously [2] (eMethods-2).

Data analyses

We described the study cohort characteristics (overall and by randomized allocation status) in terms of age, sex, body mass index (BMI), pre-existing chronic health conditions defined using the Acute Physiology and Chronic Health Evaluation II (APACHE II) score, SARS-CoV-2 antibody status, viral loads in respiratory tract, respiratory support status, concomitant COVID-19 therapy use, and allocation status.

Before biomarker analyses, two proteins (IL-2 and TGF-β1) with more than 20% missing data were excluded from the dataset (eFigure-1a, b). We then used Gibbs sampler-based left-censored missing value imputation approach (GSIMP) [25], which considers the lower limit of detection calculated in the LEGENDplex™ data analysis software suite. We checked data for batch effects and did not observe any batch effects (eFigure-1c, d). Thus, the final analytic dataset consisted of 26 biomarkers. The following analyses were performed in R statistical environment [26]. First, we assessed biomarker differences by SARS-CoV-2 antibody status and by hospital mortality, as differences in immune responses may be associated with antibody status and clinical outcomes. Second, we assessed biomarker differences by allocation status (convalescent plasma vs usual care), to check for baseline biomarker imbalances by randomization status, as any imbalances will need to be accounted for in the subsequent unsupervised analyses. Third, we used the agglomerative hierarchical clustering method with WARD2 linkage function on log10 transformed data. [27]. Additional details are reported in online supplement (eMethods-3).

Finally, we explored the associations between sub-populations, allocation status and outcomes, with regression models incorporating robust standard errors using Stata 15.1 [28]. We report the association between the outcome of OSFD-21 and subphenotypes and by allocation to convalescent plasma using ordered logistic regression models to relate our substudy results to the primary outcome in our original publication [15]. OSFD-21 is an ordinal scale outcome, where mortality is given a score of −1 and among survivors OSFD is calculated up to day 21, such that a higher number represents faster recovery [15]. We also report the associations between hospital mortality and SARS-CoV-2 antibody status in the overall cohort, by subphenotypes and by allocation to convalescent plasma using logistic regression models. We report unadjusted analyses, as testing associations between subphenotypes and treatment effects of convalescent plasma is equivalent to performing a subgroup analyses with moderate-sized RCT data, where the additional value of baseline prognostic covariates adjustment needs careful consideration, due to risk of alpha error [29]. After the regression models, we used post-estimation commands and test of heterogeneity for differences in treatment effects.

Results

Study cohort and clinical characteristics

Amongst the 2097 participants randomized to a COVID-19 immunoglobulin domain within the REMAP-CAP, 1023 were assigned to convalescent plasma and 868 to usual care in the UK [15]. Our report is based on data from 1239 participants (737/1023 (72%) assigned to convalescent plasma and 502/868 (57.8%) assigned to usual care) from the UK, who had baseline blood samples after consent but before intervention (Table 1). Clinical characteristics of our substudy cohort were similar to the overall trial population [15]. Importantly, clinical characteristics were similar between patients assigned to convalescent plasma and usual care, enabling us to analyse this UK sub-population as a cohort (eTable-1).

Table 1

Clinical characteristics of subphenotypes

Characteristic	Overall (n = 1,239)	Phenotype-1 (n = 870)	Phenotype-2 (n = 128; 10.3%)	Phenotype-3 (n = 241; 19.5%)
Allocation n (%)
Convalescent plasma  Usual care	737 (59.5%) 502 (40.5%)	534 (61.4%) 336 (38.6%)	67 (52.3%) 61 (47.7%)	136 (56.4%) 105 (43.6%)
Age, median (IQR) y	61 (52, 70)	62 (53, 70)	58 (48, 65.5)	61 (52, 70)
Female n (%)	408 (32.9%)	286 (32.9%)	39 (30.5%)	83 (34.4%)
BMI (kg/BSA m2)	30.9 (26.7, 36.3) [n = 1,111]	30.5 (26.3, 36) [n = 776]	33.4 (28.1, 37.1) [n = 116]	30.9 (27.4, 36.1) [n = 219]
Pre-existing conditions
Diabetes	358 (28.9%)	247 (28.4%)	35 (27.3%)	76 (31.5%)
Respiratory disease	292 (23.6%)	204 (23.5%)	37 (28.9%)	51 (21.2%)
Severe CVS disease	103 (8.3%)	73 (8.4%) [n = 848]	14 (10.9%) [n = 124]	16 (6.6%) [n = 236]
Immunosuppression treatment/disease	83 (6.7%) [n = 1,237]	61 (7%)	6 (4.7%) [n = 127]	16 (6.6%) [n = 240]
SARS-CoV-2 type
Wild type	424 (34.2%)	301 (34.6%)	45 (35.2%)	78 (32.3%)
B.1.1.7	270 (21.8%)	183 (21%)	27 (21.1%)	60 (24.9%)
Inconclusive	391 (31.6%)	265 (30.5%)	39 (30.5%)	87 (36.1%)
Not available	154 (12.4%)	121 (13.9%)	17 (13.3%)	16 (6.6%)
SARS-CoV-2 viral load, median (IQR) (105 IU /ml)
Wild type	7.88 (0.62—96.28)	8.21 (0.63 – 115.68)	14.27 (2.35 – 88)	3.03 (0.34 – 47.28)
B.1.1.7	24.38 (1.39 – 248.65)	19.82 (1.32 – 185.69)	57.91 (0.38 – 420.74)	88.01 (2.60 – 307.7)
Inconclusive	0.01 (0.00001 – 0.047)	0.013 (0.0016 – 0.050)	0.0085 (0.00010 – 0.031)	0.01 (0.00010 – 0.05)
SARS-CoV-2 antibody, n (%)
Detected	846 (68.3%)	582 (66.9%)	95 (74.2%)	169 (70.1%)
Not detected	348 (28.1%)	259 (29.8%)	29 (22.7%)	60 (24.9%)
Not available	45 (3.6%)	29 (3.3%)	4 (3.1%)	12 (5%)
SARS-CoV-2 antibody positive n (%)
Wild type	262 (64.8%)	184 (64.1%)	29 (67.4%)	49 (66.2%)
B.1.1.7	176 (66.7%)	115 (64.2%)	24 (88.9%)	37 (63.8%)
Inconclusive	318 (83.2%)	212 (80.9%)	32 (84.2%)	74 (90.2%)
Not available	90 (62.5%)	71 (62.8%)	10 (64.1%)	0 (60%)
APACHE II score, median (IQR)	13 (8, 19) [n = 1,196]	13 (8, 19) [n = 841]	11.5 (8, 17) [n = 122]	13 (8, 18) [n = 233]
Use and type of acute respiratory support, n (%)
Non-invasive mechanical ventilation	562 (45.3%)	397 (45.6%)	61 (47.7%)	104 (43.2%)
Invasive mechanical ventilation	418 (33.7%)	290 (33.3%)	43 (33.6%)	85 (35.3%)
High-flow nasal cannula	243 (19.6%)	172 (19.8%)	23 (18%)	48 (19.9%)
None or supplemental oxygen only	16 (1.3%)
COVID-19 therapy use
Glucocorticoids	1,166 (94.1%)	820 (94.3%)	121 (94.5%)	225 (93.4%)
Remdesivir	442 (35.7%)	300 (34.5%)	48 (37.5%)	94 (39%)
Il-6 receptor antagonists	41 (3.3%)
Outcomes
Overall
Number of OSFD at D21* (median (IQR))	1 (− 1, 21)	0 (− 1, 16)	6 (− 1, 17)	8 (− 1, 17)
Hospital mortality
Overall	444 (35.9%) [N = 1214]	331 (38.7%) [n = 855]	37 (30.1%) [N = 123]	76 (32.2%) [n = 236]
Seropositive	260/827 (31.4%)	195/571 (34.2%)	21/91 (23.1%)	44/165 (26.7%)
Seronegative	166/343 (48.4%)	123/255 (48.2%)	14/28 (50%)	29/60 (48.3%)

APACHE II score measures the severity of illness based on age, medical history, and physiological variables. Scores range from 0 to 71; higher numbers represent greater risk of death. The median score of 12 is typical for critically ill COVID-19 patients [15]. Immunosuppression treatment refers to recent chemotherapy, radiation, high dose, or long-term glucocorticoid treatment

APACHE II, Acute Physiology and Chronic Health Evaluation II score; BMI, body mass index; BSA, body surface area; COVID-19, coronavirus disease 2019; CVS, cardiovascular; IQR, interquartile range; OSFD, organ support free days

The study cohort had a median (interquartile range, IQR) age of 61 (52, 70) years, 408 (32.9%) were females and median (IQR) APACHE II score was 13 (8, 19). Typing of the specific SARS-CoV-2 variant was successful in 56% of cases, of which wild type and B.1.1.7 variants represented 61.1% and 38.9%, respectively. SARS-CoV-2 antibodies (seropositive) were detected at baseline in 846 (68.3%) patients. Nearly all (98.7%) patients required either invasive or non-invasive respiratory support; 94.1% received low dose corticosteroids and 35.7% received remdesivir. The overall cohort had a hospital mortality of 35.9%. Seronegative patients had higher hospital mortality compared to seropositive patients (odds ratio (OR) (95% confidence interval (CI)) of 2.05 (1.58–2.65; p < 0.001), which is consistent with the literature and explained by the deficient or delayed humoral immunity in severe COVID-19 [30].

Unsupervised clustering identified three subphenotypes

We found that a three subphenotype model (Fig. 1a, b; eFigure-2) optimally explained our biomarker data (eFigure-3). Subphenotype-1 was most common (70.2%; n = 870/1239), followed by subphenotype-3 (19.5%; 241/1239) and subphenotype-2 (10.3%;128/1239) (Table 1). The top ten contributing proteins to principal component-1 (PC1) and PC2 were biomarkers determining subphenotype-2 (IL-4, IFN-α2, GM-CSF, IFN-γ, IL-5) and subphenotype-3 (CXCL8, CCL4, IL-6, CCL20, CCL3) allocations (Fig. 1c). We observed striking differences between these subphenotypes on biomarker changes (Fig. 1d), with biologically plausible correlations between biomarkers (Fig. 1e) and differences in average biomarker concentrations (Fig. 1f).

Fig. 1

Unsupervised clustering of 26 protein biomarkers identified three sub-subphenotypes of critically ill COVID-19 patients. a Heatmap displaying the agglomerative hierarchal clustering identified three subphenotypes. Each row is a patient (N = 1239) and each column a biomarker. Each cell is coloured by the scaled log10-transformed protein levels (high = red, low = blue). Rows are annotated by subphenotype (subphenotype-1 = blue, subphenotype-2 = orange, subphenotype-3 = red); allocation of convalescent plasma (yes = dark blue and no = orange); serology (positive = pink and negative = navy) and hospital mortality (alive = blue and deceased = red). b Principal component analysis (PCA) of the same 26 protein biomarkers coloured by subphenotype. Subphenotype-1 = blue, subphenotype-2 = orange and subphenotype-3 = red. Columns are annotated by protein biomarker signature. A = sky blue, B = light green, and C = light red. c Top ten contributing variables to principal component (PC) PC1 and PC2. Arrows are coloured based on their respective protein contribution to variation from low (blue) to high (red). d Box and whisker plots of Log2 fold change of protein biomarkers normalized to median of subphenotype-1 and grouped by protein signature (A–B). Boxes are coloured by subphenotype. The bottom border of the box represents the 25th percentile; line bisecting the box represents the median; upper border of the box is the 75th percentile. The whiskers represent extremes, 1.5 times the 75th (highest) and 25th (lowest) values. e Circos plots of each patient subphenotype represent Spearman correlations between each protein biomarker. Only correlations of an adjusted p value < 0.001 are shown. Positive and negative correlations are coloured by red and blue, respectively. The strength of the correlation is depicted by the strength of the colour. Proteins are grouped into three signatures: A = sky blue (representing biomarkers associated with dysregulated COVID-19 immune responses), B = light green (representing Type ii, Type i and altered interferon responses), C = light red (co-regulated innate immune responses with chemokines and cytokines associated with leukocyte migration and activation). Subphenotype-1 had the weakest positive correlations between the biomarkers evaluated. In subphenotype-2, all 26 biomarkers were positively correlated, consistent with the mixed immune response pattern. In subphenotype-3, CXCL8 was negatively correlated with CXCL9, CXCL10, IFN-γ, and IFN-α2, as previously reported in COVID-19. f Summary radar plot of the 26 protein biomarkers. Medians of the log10-transformed values of each protein by subphenotype are plotted. Lines are coloured by subphenotype: subphenotype-1 = blue, subphenotype-2 = orange, subphenotype-3 = red

Immunologically, subphenotype-1 patients had variable levels of proinflammatory chemokines involved in leukocyte trafficking (CXCL9, CXCL10, CXCL11) [31], immune activating cytokines (IL-10 [32]), interferons [33] (IFN- λ1, IFN-β), TNF family biomarkers (APRIL, BAFF, sTNF-R1), and endothelial biomarkers of COVID-19 severity (ICAM-1 [34], angiopoietin-2 [35]). Taken together, subphenotype-1 represents a dysregulated immune state with biomarkers strongly associated with severe COVID-19 [6, 13], without a dominant effector or co-regulated immune responses.

In contrast, subphenotype-2 appears homogeneous on biomarker patterns, with elevated levels of IL-4 and IL-5 (known to polarize naïve T cells into T-helper 2 (Th2), enable selective B cell class switch, and macrophage activation) [36], elevated levels of interferons (IFN-γ, IFN-α2) dysregulated in severe COVID-19 [33], and high GM-CSF levels with a central role in endothelial injury, Th17 T cell response [37], neutrophil recruitment, and thrombosis seen in COVID-19 [38, 39] (Fig. 1a). Taken together, subphenotype-2 resembles the mixed immune response pattern reported previously in COVID-19 [13].

Subphenotype-3 also appears more homogeneous compared to subphenotype-1 on biomarker patterns, with elevated proinflammatory cytokines such as IL-6 (with prognostic [6, 8, 13]/therapeutic relevance [40]), and elevated chemokines indicating leukocyte recruitment/activation [41] (Fig. 1a). Taken together, subphenotype-3 represents a heightened early innate immune response state [42].

Clinical features of subphenotypes

These subphenotypes were broadly similar in terms of age, sex, prevalence of comorbidities, illness severity, types of respiratory support, and prevalence of immunosuppression. SARS-CoV-2 wild-type infections were detected in a third of patients within each of the three subphenotypes. Compared to other subphenotypes, subphenotype-3 had the highest proportion of SARS-CoV-2 B.1.1.7 variant and the highest B.1.1.7 variant viral load. Compared to seronegative patients, seropositive patients had lower viral loads for all variants, in all three subphenotypes (eFigure-4). There were no differences between subphenotypes in terms of glucocorticoid and remdesivir therapy (Table 1).

Biomarker associations between subphenotypes and serology status

Subphenotype-2 had the highest proportion of seropositive patients (74.2%) (Table 1). In the overall cohort, compared to seronegative, seropositive patients had significantly higher CXCL8, IL-5, CCL3 and CCL4, and lower levels og IFN-λ1, CXCL10, IL-10, IL-6, IFN-α2, CXCL11, and angiopoietin-2 (Fig. 2a; eFigure-5). Serology status did not segregate patients on principal component analyses (PCA; eFigure-6a). In subphenotype-1, the seropositive–seronegative comparison highlighted a pattern similar to the overall cohort (Fig. 2b). There were no significant biomarker differences between seropositive and seronegative patients within subphenotype-2 (Fig. 2c). In addition, seropositive subphenotype-1 patients had higher levels of CCL20, while seropositive subphenotype-3 patients had lower levels of CCL20 (Fig. 2d). Individual biomarker differences by serology status are shown in Fig. 2e–h.

Fig. 2

Biomarker associations between subphenotypes and serology status. Comparison of the overall cohort and subphenotypes by serology status. a Volcano plot of the overall cohort. b Volcano plot of subphenotype-1. c Volcano plot of subphenotype-2. d Volcano plot of subphenotype-3. e–h Box and violin plot of (e) IFN- λ1, (f) IL-6, (g) CCL20 a chemokine increased during microbial insult and required for effective humoral responses [54], and (h) IL-5 by overall and subphenotypes by serology status. For volcano plots, upregulated proteins (higher in serology positive compared to serology negative) are coloured red and defined as log2 fold change > 0.3 and P ≤ 0.05. Downregulated proteins (lower in serology negative compared to serology positive) are coloured blue and defined as log2 fold change < − 0.3 and P ≤ 0.05. For box and whisker plots, the bottom border of the box represents the 25th percentile; line bisecting the box represents the median; upper border of the box is the 75th percentile. The whiskers represent 1.5 times the 75th (highest) and 25th (lowest) values

Biomarker associations between subphenotypes and hospital mortality

Hospital mortality differed between subphenotypes (p = 0.03, Chi-square test), with the highest mortality observed in subphenotype-1 (38.7%), and the lowest hospital mortality in subphenotype-2 (30.1%). In all three subphenotypes, seronegative patients had a higher (and importantly similar) hospital mortality compared to seropositive patients (Table 1).

In the overall cohort, compared to survivors, non-survivors had significantly higher levels of CXCL10, CXCL9, IL-10, sTNF-RI, IL-6, angiopoietin-2, CCL20 and lower levels of CCL3 and CCL4 (Fig. 3a; eFigure-7). Mortality did not segregate patients on PCA (eFigure-6b). In subphenotype-1, and subphenotype-3, the comparison of survivors versus non-survivors had a biomarker pattern similar to the overall cohort (Fig. 3b-d). Although the volcano plot appears to show no biomarker differences between survivors versus non-survivors in subphenotype-2 (Fig. 3c), non-survivors in this cohort had higher IL-6, CXCL-10, and angiopoietin-2 (Fig. 3e–h), which is consistent with the pattern seen in the overall cohort.

Fig. 3

Biomarker associations between subphenotypes and hospital mortality. Comparison of the overall cohort and subphenotypes by hospital mortality. a Volcano plot of the overall cohort. b Volcano plot of subphenotype-1. c Volcano plot of subphenotype-2. d Volcano plot of subphenotype-3. e–h Box and violin plot of (e) angiopoietin-2, (f) CXCL10, (g) IL-6, and (h) CCL4 by overall and subphenotypes by mortality status. For volcano plots, upregulated proteins (higher in deceased patients compared to survivors) are coloured red and defined as log2 fold change > 0.3 and P ≤ 0.05. Downregulated proteins (lower in deceased patients compared to survivors) are coloured blue and defined as log2 fold change < − 0.3 and P ≤ 0.05. For box and whisker plots, the bottom border of the box represents the 25th percentile; the line bisecting the box represents the median; the upper border of the box is the 75th percentile. The whiskers represent 1.5 times the 75th (highest) and 25th (lowest) values

Association between subphenotypes and treatment effect of convalescent plasma

Within each subphenotype, there was no difference in baseline biomarkers between subjects allocated to convalescent plasma or usual care (eFigure-8), and allocation status did not segregate patients based on PCA (eFigure-6c).

The overall treatment effect of convalescent plasma compared to usual care for OSFD-21 was OR (95%CI) of 0.91 (0.74–1.11), which is consistent with our original RCT result [15] (OR (95%CI) 0.97 (0.82 to 1.14)). The overall treatment effect of convalescent plasma compared to usual care for mortality was OR (95%CI) 1.01 (0.80–1.29), which was also consistent with our original RCT result [15] with OR (95%CI) 1.04 (0.85 to 1.27) [15].

There were no major differences in the main baseline prognostic clinical characteristics [43] (age, sex, BMI, comorbidities, and need for mechanical ventilation) by allocation status, within each subphenotype (eTable-2). In subphenotype-1, the median (IQR) OSFD-21 was 0 (−1, 21) in convalescent plasma and 0 (−1, to 21) in the usual care arm. In subphenotype-2 and -3, the usual care had higher OSFD-21 compared to the convalescent plasma arm (subphenotype-2 (OSFD-21 median (IQR) 0 (−1, 21) in convalescent plasma vs 10 (−1, to 21) in usual care) and subphenotype-3 (OSFD-21 median (IQR) 1.5 (−1, 21) in convalescent plasma vs 12 (−1, to 21) in usual care). The corresponding odds ratio differed by subphenotype (test of heterogeneity; p = 0.008 (Fig. 4).

Fig. 4

Treatment effect of convalescent plasma compared to usual care for organ support-free days by subphenotypes. Forest plot comparing organ support-free days at day 21 (OSFD-21) of the overall cohort and by subphenotypes when treated with convalescent plasma, compared to usual care population. Median and inter-quartile ranges (IQR) for OFSD are displayed. Odds ratio was calculated using ordered logistic regression, and 95% confidence intervals are reported. Square dots represent odds ratio of the respective row, and the black line denotes 95% confidence intervals. Odds ratio < 1 favours control. The P value is reported based on the test of heterogeneity estimated post-ordered logistic regression. The odds ratio represents the average odds ratio for each possible cut points of the outcome variable. Proportional odds assumption means that the odds ratios are about the same regardless of the cut point of the ordinal outcome variable

In subphenotype-1, the hospital mortality in the convalescent plasma group was lower than that in the usual care group (37.6% vs 40.5%). In contrast, in subphenotype-2 and subphenotype-3, the hospital mortality in the convalescent plasma group was higher than that in usual care (subphenotype-2 = 35.4% vs 24.1% and subphenotype-3 = 34.1% vs 29.7%). The corresponding odds ratio differed by subphenotype (test of heterogeneity; p = 0.02) (eFigure-9).

Association between serology status of subphenotypes and treatment effect of convalescent plasma

In our main trial publication [15], the treatment effect of convalescent plasma did not meaningfully vary in the prespecified serology status subgroup. Consistent with the main trial result, the treatment effect on mortality did not vary by serology status of the subphenotypes (p = 0.69, for the three-way interaction test between allocation to convalescent plasma, serology status, and subphenotype). It could be that our substudy is underpowered to assess this subgroup within a subgroup effect (i.e., serology status within subphenotypes). Of note, only in subphenotype-1, seronegative patients who received convalescent plasma had lower mortality, compared to seronegative patients who received usual care (eFigure-10). As serology is a prognostic covariate, our sensitivity analyses including serology status as a covariate in two additional regression models (for OSFD-21 and mortality) was consistent with the main findings.

Discussion

We report the largest biomarker study conducted within an RCT in critically ill COVID-19 patients. We highlight three distinct subphenotypes based on biomarker profiles within critically ill COVID-19 patients who had similar clinical features, but with differences in clinical outcomes and treatment effect estimates for OSFD-21 and hospital mortality. Compared to subphenotype-1, mortality was lower despite higher inflammation in suphenotype-2 and subphenotype-3. Our observations, if validated, favour avoiding convalescent plasma therapy in subphenotype-2 and -3.

Our subphenotypes have biological plausibility. The median (IQR) IL-6 levels in our cohort was 62.2 (23.8–290.6) pg/mL, which is consistent with literature [44]. Negative association between IL-10, CXCL-10, and IL-6 with seropositive status in COVID-19 has been reported previously [6], explaining the prognostic utility of this biomarker triad. Biomarker differences between seropositive and seronegative patients in our study represent altered interferon responses [33], and compromised humoral immunity [12, 13, 30] in critically ill COVID-19 patients. Prognostic associations with many of these biomarkers have been reported previously in acute respiratory distress syndrome [45].

As we are unable to assess potential molecular mechanisms, we propose the following hypotheses as to why convalescent plasma therapy could theoretically worsen outcomes in the more proinflammatory subphenotypes [46-49]. High-affinity antibodies present in convalescent plasma elicit SARS-CoV-2 neutralization [46-48]. However, the low-affinity antibodies present either in donor plasma or formed in recipients following convalescent plasma administration could activate proinflammatory pathways [49], worsening the outcomes. The presence of autoantibodies reported in COVID-19 patients [50, 51] may be present in convalescent plasma, which could worsen outcomes in the more proinflammatory subphenotypes. Although a rare event in our primary trial [15], convalescent plasma is a blood product that can cause transfusion-related adverse events.

Our sampled population appears representative of the overall RCT publication [15]. Our findings also highlight the value of enriching trial populations [52]. Although our findings are likely to be widely applicable to moderate or severe COVID-19 patients, our primary RCT was not powered to detect subgroup effects. We neither have non-COVID controls nor validation cohorts. Research blood sampling was not possible outside the UK, and not every patient enrolled in the UK had sampling. As our RCT was conducted early on in the pandemic, SARS-CoV-2 vaccination may alter the prevalence of reported subphenotypes.

Our results have clinical utility for the following reasons. Our findings support the hypothesis that immunotherapy in COVID-19 could be useful with better patient selection based on host immune response characteristics. It is feasible to determine subphenotype-2 and -3, where we observed a harm signal by measuring a limited biomarker set based on discriminant value (such as IL-6, CCL3, and IL-8 based our data in Fig. 1). Lower overall mortality in the more inflammatory subphenotypes in our cohort supports the notion that strong prognostic association to cytokines such as IL-6 in mild COVID-19 [8] may be masked by complex cytokine networks or hubs in severe inflammatory illnesses [53] observed with severe COVID-19, highlighting the futility of measuring single cytokines as value-added biomarkers for informing treatment decisions.

Our novel findings highlight future research questions. A key next step is to study the molecular mechanisms underpinning these subphenotypes, to consider them as COVID-19 endotypes [52]. A related research question is whether these subphenotypes are associated with HTE to other immunomodulators. It is important to determine whether these subphenotypes are identifiable in non-critically ill COVID-19 patients and whether they have HTE to immunotherapies, as our study focused on critically ill COVID-19. In any viral pandemic, as convalescent plasma will be a potential treatment, understanding the mechanisms for harm may lead to better selection of donor plasma in the future.

Conclusions

We report three COVID-19 subphenotypes with differences in treatment response to ABO-compatible high-titer convalescent plasma therapy among critically ill adults, participating in an RCT. Given the distinct immunological mechanisms, these subphenotypes could be termed endotypes. These findings support the hypothesis that the benefits of immunotherapy in COVID-19 could be enhanced with patient selection based on host immune response characteristics.

Below is the link to the electronic supplementary material.

Supplementary file1 (PDF 3321 KB)

We report three COVID-19 subphenotypes with differences in treatment response to ABO-compatible high-titer convalescent plasma therapy among critically ill adults, participating in a large international multi centre randomized clinical trial. Our findings support the hypothesis that immunotherapies in critically ill adults with COVID-19 could be enhanced with patient selection based on host immune response characteristics.