Unbiased characterization of COVID-19 endotypes leads to prognostication of high-risk individuals using routine blood tests

Abstract

Background

Only a subset of individuals infected with SARS‑CoV‑2 develop severe COVID‑19. Improved tools for early diagnosis and prognostication are needed. We hypothesized that unsupervised analysis of detailed circulating proteomes could reveal biologically meaningful patient endotypes and help identify individuals at elevated risk of severe outcomes.

Methods

We performed unsupervised stratification of the circulating proteome in 731 SARS‑CoV‑2 PCR‑positive participants from the Biobanque québécoise de la COVID‑19 (BQC19), representing a range of disease severities. We also developed a prognostic model based solely on clinical laboratory measurements and applied it to 903 patients recruited across early pandemic waves (2020–2022) to generalize identified endotypes.

Results

Six proteomic endotypes (EPs) emerged. Endotype EP6 showed the highest frequencies of ICU admission, ARDS, and mortality. EP6 was marked by elevated C‑reactive protein, D‑dimer, interleukin‑6, ferritin, soluble fms‑like tyrosine kinase‑1, increased neutrophils, and reduced lymphocyte counts. SHC4 emerged as a protein quantitative trait locus associated with EP6. Among EP6 patients requiring mechanical ventilation, we observed alterations in lipoprotein metabolism, and alpha‑L‑iduronidase levels inversely correlated with duration of ventilation.

Conclusions

Unsupervised proteomic analysis identified biologically coherent endotypes that advance understanding of acute lung injury in COVID‑19 and support improved diagnostic and prognostic strategies.

Plain language summary

COVID‑19 affects people in very different ways. Some become only mildly ill, while others develop severe breathing problems. Our study aimed to understand why these differences occur by looking closely at proteins found in the blood of people with COVID‑19. We analyzed blood samples from many patients and used computational methods to group them based on their protein patterns. This information was helpful in identifying the potential clinical trajectories of new patients. One group showed signs of high inflammation and a greater risk of needing intensive care. In the future, this knowledge could support earlier treatment, improve care decisions, and help protect those most at risk during respiratory infections like COVID‑19.

Ma et al., stratify SARS-CoV-2–positive participants using unsupervised analysis of circulating proteomes. This reveals six endotypes, with EP6 showing the highest risk of severe outcomes and distinct inflammatory and hematologic signatures.

Introduction

An important challenge facing respirologists and critical care physicians is the heterogeneity in outcomes following SARS-CoV-2 infection. A minority of people infected with SARS-CoV-2 will develop a severe form of coronavirus disease 2019 (COVID-19) requiring hospitalization and respiratory support. During the first waves of the COVID-19 pandemic, 19% of symptomatic individuals had severe or critical illness associated with high mortality¹. Moreover, almost half of patients undergoing invasive mechanical ventilation (MV) died based on case fatality rate reports². Since vaccination, improvement in clinical management and the appearance of different variant of concerns, the case fatality rate has dropped by approximately 5 times between the ancestral and omicron variant of SARS-CoV-2³. Nevertheless, characterizing molecular features related to acute respiratory failure in COVID-19 that require invasive mechanical ventilation is important to identify treatable traits and improve the survival of critically ill patients⁴.

Successfully reaching this precision medicine goal requires a more granular definition of the underlying pathophysiology. A symptom-based method to discover molecular mechanisms of the disease is inherently confounded by the fact that the same higher-level condition, such as severe COVID-19 disease, can be produced by several different molecular mechanisms, a phenomenon termed the “many-one” limitation⁵. Recent advances in computing strategies can help overcome this limitation by, instead of using symptoms, starting from molecular profiles to define endotypes (i.e., subgroups of individuals who are inapparent to traditional classification methods, but share a common set of molecular factors that can lead to identification of treatable traits^6–8). Current investigations of endotypes in COVID-19 have mainly relied on supervised approaches using fixed outcomes (such as disease severity) and integrating clinical variables at the onset⁹. We hypothesize that using an unsupervised approach exploiting detailed circulating proteomes can provide mechanistic insights into the pathobiology of severe COVID-19 that can help physicians and improve diagnosis, prognosis, and clinical management.

This study identifies six proteomic endotypes (EPs) linked to diverse clinical trajectories of COVID-19 using the extensive molecular phenotyping of a cohort of 731 SARS-CoV-2 positive patients from the Biobanque Québécoise de la COVID-19 (BQC19, www.bqc19.ca)¹⁰. BQC19 is a prospective observational cohort of SARS-CoV-2-positive and negative participants recruited in the province of Québec, Canada, to improve our understanding of COVID-19 pathobiology and our capacity to alter disease outcomes. A prognostic model of disease severity to generalize the EPs to new patients is built using the molecular signature of each endotype. This prognostic model solely utilizes clinical laboratory measurements, creating the possibility of automated identification of high-risk groups in the clinic. Moreover, we leverage the molecular signature of the endotypes to obtain an overview of their underlying pathologic features.

Methods

Study approval

The study was approved by the Institutional Ethics Review Board of the Centre intégré universitaire de santé et de services sociaux du Saguenay-Lac-Saint-Jean (CIUSSS-SLSJ) affiliated to the Université de Sherbrooke [protocol #2021-369, 2021-014 CMDO – COVID19].

Inclusion and ethics

The research included local researchers throughout the research process. The research is locally relevant and has been determined in collaboration with local partners. The study has been approved by a local ethics review committee. Participants to BQC19 provided informed consent in accordance with the Declaration of Helsinki as reported in ref. ¹⁰. Participants’ data is protected and coded to minimize the risk of re-identification. We have taken local and regional research relevant to our study into account in citations.

Datasets and preprocessing

The Biobanque Québécoise de la COVID-19 (BQC19; www.bqc19.ca) organizes the data collected from ten sites across the province of Québec (Canada)¹⁰, including clinical information and multi-omics experimental data, before making it available. There was only one inclusion criteria for BQC19: that a diagnostic test for SARS-CoV-2 was performed (negative or positive). Our main corpus of analysis consisted of n = 1634 SARS-CoV-2 positive patients (based on qRT-PCR) of BQC19 who were hospitalized or visited the ER. This included n = 731 patients for which both clinical and proteomic data was available as well as n = 903 patients for whom proteomic data was not available but whose clinical data contained measurements for more than half (at least 11 out of 21) of the blood markers that we used as a validation set for the prognostic model developed in this study. The clinical outcomes were extracted from medical charts and incorporated in BQC19 case report forms completed at each visit performed for the study. They also included a compilation of clinical features upon discharge from the hospital (or death).

We also obtained data (n = 731) corresponding to the circulating proteome measured between April 2, 2020, and April 20, 2021 by a multiplex SOMAmer affinity array (SomaLogic, 4985 aptamers) from BQC19, using the first blood draw upon admission to the hospital. When measurements of the same patients but at different time points were available, we used the one corresponding to the first time point. SomaScan is a biotechnological protocol commercialized by the SomaLogic company¹¹. It relies on a set of artificial aptamers linked to a fluorophore and each designed to bind a single protein. Once added to the sample, the activity of each aptamer is measured through fluorescence and used to approximate the expression level of the targeted protein. SomaScan protocol comprises several levels of calibration and normalization to correct technical biases. Log2 and Z-score normalization (a.k.a. standard score, i.e., z = (x-μ)/σ, where μ is the mean of the cohort and σ is the standard deviation of cohort) were performed on each aptamer separately in addition to the manufacturer’s provided normalized data (hybridization control normalization, intraplate median signal normalization, and median signal normalization). Since the data was analyzed by SomaLogic in two separate batches, we applied the z-score transformation separately to each batch to reduce batch effects. These additional transformations ensure that the measured values of different aptamers are comparable and can be used in cluster analysis.

Consensus agglomerative clustering

Patients were clustered using agglomerative clustering with Euclidean distance and Ward’s linkage. To identify the number of clusters k, we used the elbow method based on the AIC and BIC. More specifically, we calculated both the AIC and BIC values separately for clustering using k = 2, 3, …, 20, using cluster assignments at each k to model the data. To find the points of diminishing return for each criterion with respect to k, we used KneeLocator from the kneed python package¹² (v0.70, curve = ‘convex’, direction = ‘decreasing’, default settings otherwise). Kneed is a Python implementation of the Kneedle algorithm¹³, an algorithm for finding the point of maximum curvature in discrete data to identify the value of k corresponding to the “elbow” (or “knee”), where increasing the value of k does not provide much better modeling of the data. Briefly, after normalizing the data to a unit square, a line segment across a diagonal, depending on concavity and direction of the data, is drawn. Next, the distance of each point to that line segment is calculated and each point with a larger distance than both the preceding point and following point is selected as a candidate point—these can be thought of as local maxima of curvature. Starting from the first candidate point, if the distances of the following points fall below a threshold—the distance of the candidate point minus the average difference between normalized points—before reaching the next candidate point, the search is stopped, and that candidate point is selected. For both AIC and BIC, Kneed identified k = 6 as the number of clusters based on both AIC and BIC; as such, we fixed the number of clusters at 6 for the remainder of our work.

Given the number of clusters in the data, we then used consensus clustering with sub-sampling to obtain robust endotypes based on proteomics (EPs). We randomly sampled 80% of the patients 1000 times. Each time, we used the agglomerative clustering above with k = 6 to identify clusters. Given these 1000 individual clusterings, we calculated the frequency of two patients appearing in the same cluster, when both were present in the randomly formed dataset. We then performed one final agglomerative clustering of these frequency scores to identify the six endotypes. To assess how sensitive the final consensus clustering is to the exclusion of one of the individual sub-sampled clustering above, we systematically calculated the adjusted rand-index (ARI) of the final clustering and a clustering obtained from 999 sub-sampled clusterings. The mean ARI was equal to 0.997 with a standard deviation of 0.029, reflecting a very high degree of robustness. (ARI is a measure of concordance between two clusterings ranging between −1 and 1, with 1 indicating perfect agreement, 0 indicating a random agreement and −1 indicating complete difference). When performing enrichment analysis or correlation, any missing value was not included in calculations.

Nearest-centroid predictor based on blood markers

To predict endotypes from blood tests, we developed a missing-value resilient nearest-centroid classifier. For each EP and each marker, only the set of patients in which that marker was measured was used to form the centroid. This was repeated for each marker separately. We used the dataset of patients that were used to form the original EPs (n = 731) as the training set and the dataset of patients that did not have proteome data as the validation set (n = 903). First, we z-score normalized each of the 43 markers across all the patients in the training set, one marker at a time. We then formed a “marker signature” (a vector of length 43) for each EP. Each element of an EP’s signature corresponds to the mean of the corresponding marker across all patients of that EP.

To predict the EP label of each patient in the test set, we first z-score normalized their blood marker measurements using the mean and standard deviation of the markers in the training set. Then, we calculated the cosine distance between each test patient’s blood marker profile and the centroids and identified the nearest EP as the predicted EP (PEP) label of the patient. Note that distances are calculated only from the test patient’s available markers. To put it more simply, each test patient’s data availability implies a subspace of the 43 dimensions space in which we can always calculate the test patient’s distance to each EP. This resiliency to missing values was a key factor in the design of this model.

To validate the model in the absence of a labelled test set, we performed leave-one-out cross validation on the training set: for each patient p, a centroid predictor was trained using all other patients but p and used to predict an EP label for p.

Western university cohort validation

Validation was carried out in data obtained from the Western University (London, Canada) as previously reported¹⁴. The Western University cohort used the Olink technology. The proteomic measurements made by Olink are not directly comparable to those made by SomaScan. Fortunately, 1868 blood samples of BQC19 from 1218 unique patients had their circulating proteins measured with both SomaScan and Olink (Target 48) technologies and could be used as matched points to link the two.

While Olink relies on antibodies to measure protein expression, each antibody having a distinct protein target, multiple aptamers used in SomaScan may share the same protein target. Consequently, while the two methods share 44 protein targets, two of those proteins were targeted by two different SomaScan aptamers, resulting in 46 (antibody, aptamer) pairs to consider. For each of these 46 pairs, we calculated the spearman rank correlation score between Olink and SomaScan measurements in the 1868 matched samples. We then selected the 15 pairs that had a spearman rank correlation score greater than 0.5 for further analysis (corresponding to 15 different proteins). For each of the 15 selected proteins, we performed polynomial fitting (nympy.polyfit), producing second order polynomials p to transform an Olink protein measurement to a value corresponding to its SomaScan measurement. This was achieved by minimizing the squared distance between p(x) and y, where x and y are the values measured in the 1868 matched samples with Olink (z-score normalised) and SomaScan (log2 and z-score normalized), respectively. We then trained a centroid EP predictor, similar to the approach used for blood markers, but based on the SomaScan-measured values of the 15 proteins of the participants in the BQC19 cohort. Finally, for each patient of the Western University cohort (n = 66), we applied the fitted second order polynomials to their Olink measurements of the 15 proteins and used the trained centroid predictor to assign them EP labels. As the Western University cohort does not contain a direct equivalent to the WHO severity scale, we used the closest available equivalent: the recruitment groups of the Western University cohort as originally described¹⁴. Note that we excluded long covid participants of this cohort due to a lack of relevance and their vastly different sampling timings.

Analyses of the genotype dataset

Annotation of SNVs were done using the biomaRt package¹⁵ from R¹⁶, and all analyses were done using R version 4.1.3. Quality control steps were derived in majority from a 2017 QC tutorial article¹⁷. At the beginning, we had 867,450 markers and 682 samples with genotype and proteomic information available. We imported Plink format data into R using the “read_plink” function from genio R package¹⁸. We removed 103,592 non ACGT bi-allelic markers. We calculated the predicted sex by looking at the rate of homozygote markers on chromosome 23. We removed 3588 markers with call rates <98%, 448,932 monomorphic markers and markers with MAF < 0.05, and 28,092 markers with Hardy–Weinberg equilibrium <1E-6 (calculated by the “HWE.exact” function from the genetics package¹⁹ from R. We additionally removed 8 participants with a sex discrepancy (based on predicted sex calculated earlier and reported sex), and 3 participants with a heterozygosity rate > 3 standard deviations. We finally removed a pair of samples which had approximately the same genome, likely due to an error of manipulation. As we could not know which one was correct, we removed both. We also found 2 pairs of individuals with a pi-hat of ~0.5 (meaning first degree relatives), so we kept the one sample per pair with the higher call rate. All other pairs of individuals had a pi-hat <0.21 that is judged acceptable considering our population. Pi-hats were calculated with the “snpgdsIBDMoM” function from SNPRelate R package²⁰. At the end of quality controls, 283,246 markers on 655 samples (EP6 n = 106 vs EP1-5 n = 549) had been used to perform association analyses.

To perform the principal component analyses (PCA), we took a subsample of independent markers (pruning) with a maximum sliding window of 500,000 base pairs and a linkage disequilibrium (LD) threshold of 0.2 using the “snpgdsLDpruning” function from the SNPRelate R package²⁰. We ran the PCA with the “snpgdsPCA” function from SNPRelate package from R. The first 2 principal components (PCs) were considered significant.

For the analyses of the 283,246 remaining markers between EP6 cluster compared to all others, we used a logistic regression model, with the dichotomous variable indicating if the participants belong to the EP6 cluster as the outcome variable. We used the additive model for markers as an independent variable, and we adjusted the models with the first two PCs. Odds ratio and p-values were calculated on each model. Quantile–quantile (Q–Q) plots have been performed as quality control of the models; p-values were plotted using “qqplot.pvalues” function from gaston R package²¹ (data not shown). We selected SNVs with a GWAS p-value < 1E-04 and looked if the related gene has a corresponding aptamer (n = 6). We then compared the aptamers’ normalized level of expression between the three groups of genotypes for each studied gene (n = 6) by performing standard analysis of variance (ANOVA) analyses followed by Tukey post hoc tests (referred to in this study as pQTL analysis). Finally, to identify if EP6 may be characterized by the aptamers and the SNVs, we performed multiple logistic regression analysis including aptamers expression values, SNV genotypes (additive model), and the two principal components as independent variables (n = 6 aptamers tested). Odds ratios are reported with 95% confidence intervals.

Statistics and reproducibility

In this study, Mann–Whitney U test was used when comparing a continuous variable between two groups. We used this test since it is a non-parametric test that does not assume normal distribution for the tested variables. When comparing enrichment and depletion of EPs or PEPs in a particular manifestation of a clinical variable (e.g., sex at birth being female), we used Fisher’s exact test. To assess correlation between two continuous variables, we used Spearman’s rank correlation as a non-parametric correlation test. Benjamini–Hochberg false discovery rate (FDR) was used throughout the manuscript to adjust the p-values when performing multiple tests. Details of each statistical test are provided when describing the results. In statistical tests, missing values were not imputed computationally but were instead excluded from the analysis. The sample sizes are described in the main text of the manuscript.

Results

Unsupervised clustering of SARS-CoV-2-positive BQC19 participants reveal endotypes associated with varying disease severity

We sought to identify endotypes of COVID-19, based on the circulating proteome of patients, in a cohort of SARS-CoV-2-positive participants in the BQC19 (Table 1) using an unsupervised approach (Fig. 1). The cohort (n = 731) included participants who came to one of the participating hospitals and visited the hospital emergency room (ER) or where hospitalized and had proteomic data from their samples collected at admission in the BQC19 database (Fig. S1). Fig. S2 shows the distribution of the patient hospital admission dates and the corresponding waves of COVID-19 infection as defined by National Institute of Public Health of Quebec (INSPQ, https://www.inspq.qc.ca/covid-19). It is important to note that 1) only 8 participants had been vaccinated prior to their recruitment and first blood sampling and 2) the great majority of the participants were recruited prior to the publication of the results from RECOVERY on corticosteroid used for severe COVID-19²². Consensus agglomerative clustering was performed on all participants for whom the circulating proteome was measured (n = 731, Table 1) using a multiplex SOMAmer affinity array (SomaLogic, ~5000 aptamers)¹¹. The optimal number of clusters (k = 6) was identified first using two criteria: Akaike’s Information Criterion (AIC) and Bayesian Information Criterion (BIC) (Fig. 2A). Then, consensus agglomerative clustering (Euclidean distance and Ward linkage)^23,24 using 1000 bootstrap subsamples of the participants was performed to obtain six robust clusters (Figs. 2 and  S3 and S4).

Table 1.

Clinical and pathological characteristics of the BQC19’s participants used to identify endotypes (EPs) in this study

		Cohort (n = 731) No. (%)	EP1 (n = 189) No. (%)	EP2 (n = 130) No. (%)	EP3 (n = 160) No. (%)	EP4 (n = 75) No. (%)	EP5 (n = 59) No. (%)	EP6 (n = 118) No. (%)
Age (years)	<45	17.0	18.0	21.5	21.2	16.0	6.8	10.2
	45–65	34.9	31.7	34.6	43.1	36.0	18.6	36.4
	>65	48.2	50.3	43.8	35.6	48.0	74.6	53.4
	Unknown	0.0	0.0	0.0	0.0	0.0	0.0	0.0
Body mass index in kg/m2	<20	2.2	2.6	3.8	0.6	0.0	5.1	1.7
	20–25	11.8	11.6	14.6	12.5	4.0	18.6	9.3
	25–35	27.2	33.9	26.9	18.8	21.3	28.8	31.4
	>35	7.0	5.8	3.1	6.9	10.7	10.2	9.3
	Unknown	51.8	46.0	51.5	61.3	64.0	37.3	48.3
Sex at birth	Female	43.8	52.4	50.0	34.4	54.7	39.0	31.4
	Male	56.2	47.6	50.0	65.6	45.3	61.0	68.6
Severity	Deceased	3.1	0.5	1.5	3.1	0.0	5.1	10.2
	Severe	31.5	12.7	19.2	33.8	37.3	39.0	64.4
	Moderate	51.2	65.1	56.9	47.5	58.7	52.5	22.0
	Mild	11.5	19.6	17.7	13.1	1.3	1.7	0.8
	Unknown	2.7	2.1	4.6	2.5	2.7	1.7	2.5
Oxygen therapy	Yes	66.6	45.0	48.5	68.1	92.0	78.0	97.5
	No	22.8	45.0	38.5	11.2	5.3	13.6	1.7
	Unknown	10.5	10.1	13.1	20.6	2.7	8.5	0.8
Admission to intensive care unit	Yes	32.0	12.2	20.0	30.6	30.7	37.3	77.1
	No	65.8	84.7	77.7	66.2	68.0	62.7	22.0
	Unknown	2.2	3.2	2.3	3.1	1.3	0.0	0.8

Fig. 1. Summary of the unsupervised consensus clustering pipeline.

Starting with the proteomic data of all EP candidates (cf. consort diagram in Fig. S1), we first determine the optimal number of clusters k (Step 1) by calculating the Akaike and Bayesian information criteria (AIC and BIC) of multiple candidate values for k ([2,…,20]). These scores are then passed to the kneedle algorithm which locates the “knee”, i.e. the point where the benefits of adding a cluster starts decreasing. This point is reached when k = 6, which is thus selected as the optimal number of clusters. We then perform the consensus agglomerative clustering (Step 2). First, we randomly sample 80% of the participants from the cohort and cluster them based on their protein expression (hierarchical clustering, k = 6). We repeat this process 1000 times and use the resulting clusterings to calculate a score reflecting the tendency of any two participants to be clustered together. Finally, we cluster all participants based on these scores (hierarchical clustering, k = 6) resulting in 6 endotypes (EPs). Please note that the 2D datapoints illustrating this part of the process (the ones with dotted border) do not correspond to real participant data and are only provided to help visualize the computational pipeline.

Fig. 2. Unsupervised consensus clustering of SARS-CoV-2 positive patients.

A AIC and BIC by the number of clusters, used to identify the elbow (shown in red) using kneedle. B The heatmap shows the expression of aptamers (rows) in each sample (columns). The dendogram shows the identified endotypes. C Characterization of samples based on sex at birth, highest world health organization (WHO) severity level achieved, intensive care unit (ICU) admission, and oxygen therapy. For the last two rows, a sample colored “black” reflects a label of “yes”.

The clinical and pathological characteristics of patients within each endotype are provided in Table 1. The aptamers ranked by one-vs-rest differential measurements for each endotype are listed in Supplementary Data 1. To characterize the identified endotypes (EPs) with respect to disease severity, we performed two-sided Fisher’s exact tests to assess their enrichment (or depletion) in either of “severe” or “deceased” outcomes. Severity was defined using the World Health Organization definition²⁵ for the peak level attained during hospitalization. EP6 was significantly enriched in the severe/deceased outcomes (Benjamini–Hochberg false discovery rate (FDR) = 1.74E–21), with either of these outcomes observed in 74.6% of EP6 patients. Meanwhile, EP1 was significantly depleted in severe/deceased outcomes (FDR = 1.89E–13) (Fig. 3A, Table 1, Supplementary Data 2), with either of these outcomes observed in only 13.2% of EP1 patients. In addition, EP6 was enriched (two-sided Fisher’s exact test) in participants receiving oxygen therapy (97.5%, FDR = 4.23E–18) and being admitted to intensive care unit (ICU) (77.1%, FDR = 9.51E–28) (Fig. 3A, Table 1, Supplementary Data 2). Kaplan–Meier analysis²⁶ also confirmed that the identified EPs have a distinct temporal pattern of admission to ICU since their first symptoms (Fig. 3B, multivariate logrank test P = 1.58E–38) or since their arrival at hospital (Fig. 3C, multivariate logrank test P = 2.17E–40), with EP1 and EP6, respectively, having the lowest and highest probability of being admitted to ICU (or dying prior to that) in a 40-day span. A similar pattern was also observed when patients that died before admission to ICU were excluded (Fig. S5).

Fig. 3. Characterization of endotypes (EPs).

A Enrichment or depletion of each EP in clinical variables. Two-sided Fisher’s exact tests are used to calculate the p-values, which are corrected for multiple tests using Benjamini–Hochberg false discovery rate (FDR). Gradients of blue show depletion, while gradients of red show enrichment. FDR values above 0.05 are depicted as white. The Kaplan–Meier analysis of the time between participants’ first symptom (B) or arrival at hospital (C) and their admission to intensive care unit (ICU) (or death if earlier) for each EP (Delta). “n” represents number of participants with available data that were included in the analysis. Boxplot displays of age at arrival (D) and BMI (E) in each EP. In D, the numbers of each group are: EP1 (n = 189), EP2 (n = 130), EP3 (n = 160), EP4 (n = 75), EP5 (n = 59), and EP6 (n = 118). In E, the numbers of each group are: EP1 (n = 102), EP2 (n = 63), EP3 (n = 62), EP4 (n = 27), EP5 (n = 37), and EP6 (n = 60). F Stacked barplot displays of COVID-19 severity in each EP.

Notably, EP6 (which had the most severe patients) did not show enrichment in older patients or individuals with high body mass index (BMI) (two-sided Mann–Whitney U (MWU) test FDR > 0.05) (Fig. 3D–F, Table 1, Supplementary Data 2). On the other hand, patients in EP5 were generally older than other EPs (FDR = 7.73E–5), while EP3 included younger patients (FDR = 1.53E–4). Interestingly, EP6 was not enriched in any comorbidities (total of 27, Supplementary Data 3), reflecting that the severity of the disease observed in this group (and captured using proteomic data) could not be simply explained based on these factors. On the other hand, EP5 was enriched in several comorbidities such as chronic kidney diseases (two-sided Fisher’s exact test FDR = 3.09E-18) and previous heart failure (FDR = 1.05E–4). Finally, the distribution of dead outcome was similar in EP6 (10.2%) and in all its three subclusters of patients (ranging from 8.0% to 12.1%), which supports that the proximity to death as an outcome is not a major explaining factor for the clustering.

To validate our identified EPs in an independent study, we used data corresponding to n = 66 participants from the Western University (London, Canada) cohort as previously reported¹⁴ labeled as healthy controls, Ward COVID-19 or ICU COVID-19. This study used Olink technology for proteomic measurements, which are not directly comparable with those obtained from SomaScan. To overcome this, we used a subset of BQC19 participants that had measurements from both technologies and developed a method to predict the EP labels of this cohort (see Methods). These participants showed similar outcome distribution, with most healthy controls appearing in those predicted as EP1 (followed by those predicted as EP3), whereas the majority of severe ICU COVID-19 were found in predicted EP6 (followed by EP5) (Fig. S6, Supplementary Data 4).

Our hypothesis was that the unsupervised approach would identify groups of participants that shared distinct proteome leading to similar outcomes. Interestingly, when we compared each EP’s proteomic profile to that of a supervised approach that stratified participants based on mild/moderate versus severe/deceased during hospitalization, we observed a decent correlation between EP6 and this supervised approach (Spearman r = 0.61; Table 2, Fig. S7). However, other EPs showed a much lower correlation. Interestingly, the ranking of aptamers associated with EP5, the group with the second most severe disease based on time to ICU admission, showed almost no correlation (Spearman r = 0.022) with the supervised approach above (Table 2, Fig. S7).

Table 2.

Comparison of aptamer rankings based on severity vs. aptamer rankings based on endotypes

Endotype	Spearman correlation coefficient a
EP1	0.166
EP2	0.055
EP3	–0.276
EP4	–0.018
EP5	0.022
EP6	0.608

^aThe Spearman correlation coefficients are obtained as follows. We first performed two-sided MWU tests comparing the aptamer measurements of mild/moderate participants vs. severe/deceased participants and used the resulting p-values (adjusted) to obtain a “severity-based” aptamer ranking. We then repeated this procedure, this time comparing aptamer measurements of participants belonging to a given EP with participants of all other EPs (gathered as a single group, e. g., EP1 vs. EP2-6). We thus obtained 6 “EP-based” aptamer rankings, one for each EP. The table contains the Spearman’s rank correlation coefficient between the first “severity-based” aptamer ranking and each of the 6 “EP-based” rankings.

These analyses revealed that the unsupervised approach using the circulating proteome of BQC19’s participants was able to identify generalized endotypes with distinct disease characteristics and outcomes that were not evident simply based on BMI, age, or comorbidities. We identified EP6 as a group of participants with an increase in key measures of COVID-19 disease severity, including admission to ICU and the need for oxygen therapy.

Clinical laboratories reveal that members of EP6 have increased levels of C-reactive protein, D-dimer, elevated neutrophils, and decreased levels of lymphocytes

To further characterize each EP, we assessed the clinical laboratory results obtained from blood draws and compared them between the groups. We focused on 21 markers and used the summary value reported in the BQC19 database corresponding to the most extreme measurement among multiple blood draws (Fig. 4A and Supplementary Data 5), as well as the measurements from the first blood draw (Supplementary Data 5, Fig. S8). The analyses were performed both separately and together for male and female participants (Supplementary Data 5, Fig. S8). Figure 4A shows the elevation and reduction of the median of these markers in the identified EPs. EP6 is characterized by abnormal values in markers of inflammation (lymphopenia, total white blood cell count, neutrophilia, C-reactive protein (CRP)), liver damage (alanine aminotransferase (ALT)), coagulopathy (D-dimer, low hemoglobin, International Normalized Ratio (INR)), and hyperglycemia (glucose). These patterns were generally consistent among both male and female participants, and a similar pattern was also observed based on the first blood draw (Supplementary Data 5). Three notable differences were bilirubin and creatine kinase, which were enriched in males only, and vitamin D, which was depleted in males only (Fig. 4).

Fig. 4. Patterns of blood markers and Roche diagnostic markers in the endotypes (EPs), separately for male (left) and female (right) participants.

Heatmaps show the false discovery rate (FDR) values for two-sided one-vs-rest Mann–Whitney U tests for 21 blood markers (most extreme value during hospitalization. Please refer to Fig. S8 and Supplementary Data 5 for the data obtained using the first blood draw) (A) and 22 Roche diagnostic markers (B) for each EP (first blood draw). Gradients of red imply that the median of the variable was higher in the EP (compared to other EPs), while gradients of blue show that the median was lower. FDR values below 0.05 are shown as white. The numerical values show the median value of the maker in each EP. WBC white blood cells, LDH lactate dehydrogenase, ALT alanine aminotransferase, aPTT activated partial thromboplastin time, INR International Normalized Ratio.

We also used 22 markers from the Elecsys diagnostic panel (Roche Diagnostic) to further characterize EP6 (only available for first blood draw) (Fig. 4B and Supplementary Data 5). This led to the identification of additional elevated and highly significant markers: (a) alpha-1 antitrypsin, an acute phase reactant elevated during inflammatory conditions; (b) Interleukin 6 (IL-6), a pleiotropic cytokine associated with systemic inflammatory response syndrome²⁷, shown to be elevated in severe COVID-19²⁸ and linked to endothelial damage and liver injury²⁹; (c) ferritin, an iron-storage protein and acute phase reactant, elevated in COVID-19, and like other hyperferritinemic syndromes, associated with coagulopathy^30,31; and (d) soluble vascular endothelial growth factor (VEGF) receptor sFLT1 (soluble fms-like tyrosine kinase-1), previously shown to be associated with endothelial damage and COVID-19 severity³².

A computational prognostic model based on blood biomarkers predicts EPs in a separate cohort

Since each EP showed a clear and distinct clinical laboratory measurement signature based on 21 blood markers and 22 Roche Elecsys diagnostic markers, we sought to develop a computational prognostic model of disease severity based on these signatures. We focused on data from the first blood draw for all markers (Fig. 4B and S8, Supplementary Data 5) and developed a nearest-centroid classifier, resilient to missing values, to predict EPs based on these 43 markers (see Methods for details).

Internal cross-validation showed a good performance for this model, both for the binary classification between EP1 and EP6 (Area Under the Receiver Operating Characteristic curve or AUROC = 0.90), as well as for the muti-class classification (weighted average AUROC = 0.75) (Fig. S9). To further test the prognostic ability of this model on an independent yet similar dataset, we analyzed 903 SARS-CoV-2 positive BQC19 participants that did not have circulating proteome data and had not been used to identify the endotypes (see Fig. S10 for the distribution of the patient hospital admission dates). These patients were recruited across waves of the pandemic between March 2020 and October 2022 and included hospitalized participants or those who visited the hospital ER. The clinical and pathological characteristics of patients in each predicted endotype (PEP) are provided in Fig. 5 and Supplementary Data 6.

Fig. 5. Characterization of predicted endotypes (PEPs) based on the prognostic model using 43 blood diagnostic markers.

The Kaplan–Meier analysis of the time between participants’ first symptom (A) or arrival at hospital (B) and their admission to intensive care unit (ICU) (or death if earlier) for each PEP (Delta). “n” represents number of participants with available data that were included in the analysis. Boxplot displays of age at arrival (C) and BMI (D) of participants for each PEP. In C, the numbers of each group are: PEP1 (n = 29), PEP2 (n = 35), PEP3 (n = 13), PEP4 (n = 13), PEP5 (n = 21), and PEP6 (n = 27). In D, the numbers of each group are: PEP1 (n = 16), PEP2 (n = 24), PEP3 (n = 10), PEP4 (n = 6), PEP5 (n = 10), and PEP6 (n = 10). E Stacked barplot displays of COVID19 severity for each PEP.

Our prognostic model identified 167 of these 903 patients as belonging to predicted EP6 (PEP6). Two-sided Fisher’s exact tests showed significant enrichments of PEP6 in severe/deceased outcomes (59.9%, FDR = 5.62E–21), while PEP1 and PEP2 were significantly depleted in these outcomes (10.9%, FDR = 5.29E–8 for PEP1; 13%, FDR = 1.19E–8 for PEP2), as shown in Supplementary Data 6. Like EP6, PEP6 was also significantly enriched in participants receiving oxygen therapy (83.2%, FDR = 2.87E–13) and being admitted to ICU (56.9%, FDR = 1.34E–19) (Supplementary Data 6), all based on two-sided Fisher’s exact tests. Kaplan–Meier analysis also confirmed that these PEPs have a distinct temporal pattern of admission to ICU (similar to those of EPs) since their first symptoms (multivariate logrank test P = 6.78E–19) or since their arrival at hospital (multivariate logrank test P = 6.75E–27) in a 40-day span (Fig. 5B). Additionally, see Supplementary Figs. S11-S12 in which different thresholds on the number of available blood variables were used to form the PEP cohort. A similar pattern using Kaplan–Meier analysis can be observed in these cases, showcasing the ability of the model to predict PEPs in the presence of data sparsity. These results suggest that our prognostic model based on 43 blood biomarkers can be used to generalize the definition of endotypes to patients for whom proteomic data is not available, with different degrees of available blood variables.

Since the 21 blood markers are more commonly available, we also developed a prognostic model based only on these markers, which also showed strong prognostic capabilities (Fig. S13). Put together, these results suggest that it is possible to leverage detailed molecular information on a smaller number of participants to predict clinical outcomes on a larger population using routinely available information collected during hospitalization.

EP6 is enriched with BQC19 participants with acute respiratory distress syndrome

To better characterize all EPs with regards to different complications, we performed two-sided Fisher’s exact tests to assess their enrichment in different binary variables. In accordance with increased COVID-19 disease severity, EP6 was enriched in several medical complications including ARDS (50.0%, FDR = 1.12E–11), acute kidney injury (50.0%, FDR = 5.73E–8), secondary bacterial pneumonia (38.1%, FDR = 2.25E–5), liver dysfunction (22.0%, FDR = 1.37E–3), cardiovascular complications (39.0%, FDR = 1.37E–3), and bacteremia (12.7%, FDR = 4.28E–3) (for the full list see Supplementary Data 7). Notably, the frequency of ARDS, defined according to the Global ARDS definition³³ by the attending physician, was 9% in EP1 compared to 50.0% in EP6, making this complication a key feature of this endotype (Fig. 6, Supplementary Data 7). To determine whether the molecular features of ARDS were similar between the EPs, we compared ARDS+ and ARDS- patients in each EP and obtained a ranking for the aptamers based on their differential expression (MWU test, Supplementary Data 8). Results show that the aptamer profiles of ARDS are different between EPs, once again supporting the many-one hypothesis, with distinct proteome profiles leading to ARDS across the EPs.

Fig. 6. Frequency and statistical significance of complications in different EPs.

The value in each cell shows the percentage of patients of that EP (column) that suffered from the complication (row). The colors represent two-sided Fisher’s exact test false discovery rate (FDR, corrected for multiple tests). Red represents enrichment, while blue represents depletion. FDR values below 0.05 are shown as white.

SHC4 genotype and protein abundance levels within EP6

To discover molecular factors associated with EP6, we first sought to determine whether we could narrow down a small number of single-nucleotide variants (SNVs) that were enriched in EP6 compared to the other endotypes (analysis visually summarized in Fig. S14). We identified 25 SNVs distributed in 13 annotated genes that distinguish EP6 at a threshold p-value of 1E-4 (Table 3). We then investigated each of those SNVs to which we could assign an aptamer from the proteomic profiling (n = 6 SNVs) to assess whether their protein product in circulation was differentially regulated by the genotype (Table 4). We discovered two genes, SHC4 (encoding SHC adaptor protein 4) and CACNA2D3 (encoding calcium voltage-gated channel auxiliary subunit alpha2 delta3), for which there was a significant association between genotype and protein expression levels (raw p-values < 0.05). While CACNA2D3 may have mild impact on EP6 membership (odds ratio = 0.61, CI95%: 0.46–0.79, Table 4), SHC4 was one of the top expressed aptamers (position 32 out of 4,985) in EP6, with an odds ratio of 11.98 for the protein product (CI95%: 7.61–18.87) and 2.00 for the alternate allele (CI95%: 1.30–3.09), respectively, in a multiple logistic regression model. Therefore, this analysis revealed that the signaling adaptor protein SHC4 could play a role in severe disease pathology and should be investigated further.

Table 3.

SNVs differentiating EP6 against all other endotype clusters

SNV ID	Position	Gene Symbol	SNVa	MAF	HWE	ORb	p-value
rs1394671	chr5:7348567	-	G > A	0.240	0.095	2.28	2.41E–06
rs12186698	chr5:168880668	SLIT3	T > C	0.087	0.247	2.93	3.72E–06
rs11625406	chr14:50049474	-	C > A	0.438	0.137	0.47	1.03E–05
rs11862889	chr16:83828886	-	T > C	0.104	0.446	2.62	1.07E–05
rs2995918	chr4:37905775	TBC1D1	C > T	0.494	0.776	0.49	1.20E–05
rs16897810	chr5:68415296	-	G > A	0.093	0.001	2.67	1.43E–05
rs2376263	chr17:35436659	SLFN13	G > A	0.261	0.073	2.01	2.48E–05
rs4790712	chr17:1614620	SLC43A2	G > A	0.441	0.003	1.97	2.50E–05
rs2294566	chr20:41472939	CHD6	C > A	0.239	0.265	2.04	2.78E–05
rs7164451	chr15:48921859	SHC4	G > A	0.386	0.053	1.89	3.01E–05
rs57664621	chr8:22808443	PEBP4	G > C	0.174	0.016	2.16	3.07E–05
rs28482919	chr3:14903094	FGD5	T > C	0.187	0.053	0.32	3.20E–05
rs6559283	chr9:89712560	-	T > C	0.375	<0.001	1.96	3.27E–05
rs657075	chr5:132094425	-	A > G	0.104	0.328	2.49	4.28E–05
rs56235109	chr15:62424001	TLN2	A > G	0.235	0.366	0.38	4.67E–05
rs7620057	chr3:179473377	GNB4	T > C	0.099	1.000	2.51	4.67E–05
rs10466868	chr12:131455375	-	T > G	0.108	0.401	2.51	5.94E–05
rs2236798	chr1:18735127	PAX7	A > G	0.054	0.114	2.90	6.11E–05
rs3774814	chr4:5464702	STK32B	C > G	0.205	<0.001	0.36	6.75E–05
rs4497815	chr19:22903215	-	G > A	0.216	0.205	2.07	7.52E–05
rs6765694	chr3:54601810	CACNA2D3	G > A	0.404	0.833	0.51	7.66E–05
rs2797773	chr6:37559045	-	C > T	0.422	0.000	0.53	8.06E–05
rs17014760	chr4:129419725	-	A > A	0.316	0.000	1.89	9.33E–05
rs10948260	chr6:45835559	-	G > A	0.366	0.189	1.81	9.52E–05
rs12035677	chr1:232391209	-	A > G	0.063	0.000	3.02	9.95E–05

SNV Single nucleotide variation, HWE Hardy Weinberg Equilibrium, MAF Minor allele frequency, OR Odd ratio.

^aSNV are described following GWAS annotations: refence allele > alternative allele (e.g., G > A).

^bLogistic regression analyses using additive model adjusted for the 2 principal components.

Table 4.

Association between genotypes and aptamer expression levels

SNV	Gene Symbol	Nearest gene	Aptamers normalized expressiona				Multiple logistic regression analysesb
			HMref	HTZ	HMalt	p-valc	Aptamers		SNV
							OR	p-val	OR	p-val
rs6765694	CACNA2D3	-	–0.09 ( ± 0.89)	0.01 ( ± 1.10)	0.24 * ( ± 0.99)	0.011	0.61 (0.46–0.79)	2.5E–4	0.52 (0.37–0.74)	2.4E–4
rs10948260	-	CLIC5	0.02 ( ± 0.92)	0.02 ( ± 1.09)	–0.02 ( ± 0.97)	0.796	0.68 (0.55–0.83)	1.7E–4	1.83 (1.35–2.48)	8.8E–5
rs657075	-	IL3	–0.01 ( ± 1.01)	0.05 ( ± 0.99)	0.32 ( ± 0.46)	0.328	1.37 (1.14–11.64)	8.2E–4	2.49 (1.60–3.88)	5.2E–5
rs7164451	SHC4	-	–0.07 ( ± 0.90)	0.01 ( ± 1.05)	0.23 * ( ± 1.16)	0.017	11.98 (7.61–18.87)	8.4E–27	2.00 (1.30–3.09)	1.8E–3
rs12186698	SLIT3	-	0.00 ( ± 1.02)	0.07 ( ± 1.06)	0.15 ( ± 0.89)	0.509	0.82 (0.64–1.06)	0.135	2.99 (1.89–4.73)	2.7E–6
rs56235109	TLN2	-	–0.03 ( ± 0.98)	0.06 ( ± 1.03)	0.04 ( ± 1.27)	0.353	0.58 (0.45–0.75)	2.7E–5	0.38 (0.23–0.60)	4.7E–5

SNV Single nucleotide variation, HM_ref Homozygotes for the reference allele, HM_alt Homozygotes for the alternative allele, HTZ Heterozygotes, OR Odd ratio.

^aAptamers’ normalized levels of expression are reported as mean ( ± standard deviation). Normalization steps for aptamer expressions are described in Methods.

^bMultiple logistic regression analyses models include aptamers expression values, SNV genotypes (additive model) and the two principal components. All tests were two-sided. OR are reported with 95% confidence intervals in parentheses.

^cp-value, standard ANOVA analyses followed by Tukey post hoc analyses. Asterisks (*) identify difference between HM_ref and HM_alt genotypes.

Mechanical ventilation in EP6 is associated with lipoprotein metabolism

As noted in the introduction, mechanical ventilation (MV) in COVID-19 is associated with a high mortality rate. Within EP6, 55 out of the 118 individuals received mechanical ventilation (MV + ) and 63 did not (MV-) (Supplementary Data 9). There were no significant differences between these MV+ and MV- groups in terms of sex, age, or body mass index (BMI) (Supplementary Data 9 and Fig. 7A). The small number of participants belonging to each group limits our power to draw definitive conclusions about the impact of these variables. Therefore, we decided to leverage proteomic and metabolomic data available within BQC19 to see whether circulating molecular features can distinguish the two groups. We identified only one significant molecular feature associated with MV duration: alpha-L-iduronidase (IDUA), which was inversely correlated with duration of ventilation in EP6 (Spearman’s rank correlation = –0.60, FDR = 2.57E–2, Fig. 7B and Supplementary Data 9). Analysis of the metabolomic data revealed alterations in lipid metabolism that distinguish patients undergoing mechanical ventilation (Fig. 7C and Supplementary Data 9). Finally, within all EP6 members that share the overall similar molecular signature, we identified a subset of aptamers that were significantly different between MV+ and MV- patients: two aptamers corresponding to the proteins G antigen 2D (GGE2D) and angiopoietin-like 3 (ANGPTL3) were over-expressed, while four aptamers corresponding to Luteinizing hormone (LH), Beta-crystallin S (CRYGS), Neural cell adhesion molecule 1 120 kDa isoform (NCAM1, CD56), and Type 2 lactosamine alpha-2;3-sialyltransferase (ST3GAL6) were under-expressed in the MV+ group (two-sided MWU test FDR < 0.01) (Fig. 7D). Overall, these results point to an association between altered lipoprotein metabolism and mechanical ventilation in EP6.

Fig. 7. Mechanical ventilation in EP6 and other EPs.

A The distribution of age, BMI, and sex in participants that received mechanical ventilation (MV + ) and those that did not (MV-) in EP6 and other EPs (EP1–5). The p-values are calculated using two-sided Mann–Whitney U (MWU) tests. B The scatter plots show the correlation between the abundance of IDUA aptamer and the number of MV days (PCC: Pearson’s correlation coefficient, SCC: Spearman’s rank correlation coefficient). C The metabolic subpathways associated with MV in EP6 and EP1-5. Within each group, we first identified metabolites whose values were significantly different between MV+ and MV- patients (two-sided MWU test, Benjamini–Hochberg FDR < 0.01). Then, we used those metabolites to perform pathway enrichment analysis (one-sided Fisher’s exact test). D, E The distribution of aptamer abundance for MV+ and MV- patients in EP6 and EP1-5. Two-sided MWU tests followed by Benjamini–Hochberg method for adjusting p-values were used to calculate false discovery rates.

Next, we asked whether the molecular features identified in EP6 are also associated with mechanical ventilation in other EPs (EP1-5) (Fig. 7 and Supplementary Data 10). While, as expected, the percentage of participants receiving MV was much smaller in EP1-5 (Fig. S15), there was not a significant difference between MV+ and MV- participants with regards to age, sex, or BMI (Fig. 7A). Moreover, in contrast to EP6, alpha-L-iduronidase (IDUA) was not found to be correlated with duration of ventilation in EP1-5 (Fig. 7B), whether we exclude fatal outcomes or not (Fig. S16), making this molecular feature specific to participants belonging to EP6. Accordingly, metabolic profiling at the sub-pathway level found largely distinct patterns in EP6 and EP1-5, with only phosphatidinylethalonamine shared between the two groups (Fig. 7C). Finally, except for IDUA, the other aptamers showed a coherent pattern of enrichment/depletion between EP6 and EP1-5 (Fig. 7D, E). It is worth noting, however, that the ranking of aptamers between the two groups is very different, with the highest EP6 depleted aptamer (NCAM120) coming at the 74th position and GGE2D, the highest enriched aptamer ranked in EP6, is ranked 375 in EP1-5 (Supplementary Data 9 and 10). Therefore, we conclude that the molecular pattern (metabolomic and proteomic) of mechanical ventilation is largely distinct between EP6 and EP1-5.

Discussion

In this study, we have bridged the gap between the circulating proteome and routinely available blood diagnostic biomarkers, using machine learning algorithms, to prognosticate COVID-19 outcomes in two independent Canadian cohorts (Fig. 8). The model was applied to data from participants recruited across the pandemic waves from 2020 to 2022, demonstrating that it performs well despite mutations in infecting strains and the development of immunity. This showcases an analytical pipeline that can support physicians in making more informed decisions on potential unfavorable trajectories early during hospitalization and adjusting follow-ups and treatments accordingly.

Fig. 8. Overview of the study design and key findings.

Circulating proteomic, clinical, blood test, metabolomic, and genetic data from the BQC19 cohort were subjected to unsupervised clustering, identifying six distinct proteomic endotypes (EPs). Characterization of these endotypes revealed EP6 as the group most strongly associated with severe COVID 19 outcomes, including ICU admission, ARDS, and mortality. Further analyses showed that EP6 was marked by elevated inflammatory and coagulation biomarkers and, among mechanically ventilated individuals, alterations in lipoprotein metabolism. A protein quantitative trait locus involving SHC4 was linked to EP6. A prognostic model based on routine clinical laboratory measurements was developed and externally validated in an independent cohort, enabling prediction of endotypes in new participants.

One major strength of this study is the use of an unsupervised approach for the analysis of a large and well-phenotyped cohort. This broad approach led to the identification of six COVID-19 disease endotypes with different clinical trajectories and distinguishing characteristics that are summarized in Table 1. Interestingly, our analyses showed that these distinct characteristics could not be explained simply based on age, BMI, or comorbidities. We identified two endotypes with more favorable outcomes (EP1 and EP2), three endotypes with intermediate outcomes in terms of severity (EP3, EP4 and EP5), and one endotype which led to the worst outcomes compared to all others (EP6). EP6 was associated with ARDS, the worst clinical manifestation of COVID-19 that was reflected by a greater proportion of ICU admission, mechanical ventilation, and severe/fatal outcomes (Figs. 3 and 6). Clinical features of this endotype were consistent with published literature, including increased levels of CRP, D-dimer, IL-6, ferritin, sFLT1, elevated circulating neutrophils, and reduced peripheral blood lymphocytes (Figs. 4 and 6, Supplementary Data 5), presenting a profile associated with systemic inflammatory response syndrome and abnormal coagulation. Another endotype (EP5), while leading to unfavorable clinical trajectory during hospitalization, was instead associated with markers of cardiovascular disease, cardiovascular complications during hospitalization, and older age.

The identification of endotypes was done systematically using robust consensus clustering of aptamer expression levels, in which the optimum number of clusters was determined congruently using two well-established measures: AIC and BIC. The consensus clustering using bootstrap sampling (1000 times) ensured identification of robust clusters that are resilient to exclusion of some of the samples (20% randomly selected and excluded at each cycle) or any individual clustering (see Methods for the robustness analysis). Moreover, identifying the best number of clusters using AIC/BIC (both of which agreed with each other) allowed us to reveal the patterns of the EPs directly from the data instead of imposing a pattern onto it through human supervision. This is an important strength of the study that enabled us to identify distinct molecular patterns of patients that could have remained undetected using other traditional approaches. This is supported by our observation that, while a stratified approach on COVID-19 severity shared a good correlation with EP6, the two approaches diverged considerably for the mild/moderate outcomes.

Additionally, to improve the translational applicability of the EPs, we developed a prognostic model based only on measurements of conventional clinical laboratory blood markers to test the generalizability of these endotypes to samples without measured aptamer data. Characteristics of EPs predicted solely based on their blood markers were consistent with the original EPs, suggesting that clinical blood markers could be used as surrogates for assignment of these EPs to new patients and potentially automating identification of high-risk groups in the clinic. This approach takes into account the effect of multiple blood variables simultaneously and incorporates the full distribution of each variable. This contrasts with clinical laboratory results that are automatically flagged as within or outside normal range, one variable at a time, therefore increasing the clinical applicability of our model by leveraging a wider spectrum of information to prognosticate patient outcomes. This suggests that it is worthwhile to include, in addition to standard reports, the information from additional more comprehensive models such as this one, once they are properly validated for clinical use.

To identify biologically relevant and plausible biomarkers, we used a systematic approach to narrow the list of plausible candidates using a combination of matching genomic and proteomics findings. This led us to identify a protein quantitative trait locus in SHC4. We identified SHC4 as a molecular factor associated with EP6. In view of the limited knowledge of this understudied member of the SHC family, we looked at the metabolites associated with SHC4 to gain insights into its possible functions. Interestingly, 6-oxopiperidine-2-carboxylate (the only metabolite with a significant correlation with SHC4 in EP6) was found to be negatively associated with glomerular filtration rate in a GWAS study of kidney disease and hypertension in African Americans³⁴. Acute kidney injury is a frequent complication of acute liver failure³⁵, and liver dysfunction is associated with abnormal coagulation³⁶. Thus, the overall molecular information coming from the multidimensional analysis of EP6 points to a triad between liver function, kidney function, and hemostasis that becomes dysfunctional following ARDS-associated inflammation driven by SARS-CoV-2 infection. Whether the presence of SHC4 in circulation is a marker of dysfunction of this triad or an active mediator remains to be determined. Moreover, the identity of the cells expressing SHC4 leading to its presence in the circulation is not known. It must be noted that change in glomerular filtration may also impact the circulating levels of various molecules. Interestingly, SHC4 and GGE2D (identified as enriched in participants receiving MV in EP6, Fig. 7) were both found as part of an obesity-associated transcriptomic signature using data from The Cancer Genome Atlas³⁷. While we did not detect significant differences in BMI between EP6 and other EPs, it is worth remembering that participants were acutely ill, likely resulting in weight loss that may have impacted BMI measurements. It is intriguing to consider that the circulating proteome may still contain molecular features associated with obesity that can contribute to pathogenesis.

We discovered potential roles for ANGPTL3 and IDUA in mechanical ventilation of EP6 patients. ANGPTL3 is a liver-secreted enzyme that can inhibit lipoprotein and endothelial lipase³⁸. Loss-of-function of ANGPTL3 is associated with hypolipidemia³⁹. Therefore, increased circulating levels of ANGPTL3 would be expected to raise plasma levels of triglycerides and phospholipids, which is consistent with the observed metabolomic profiles (Supplementary Data 9). Whether these alterations are sufficient to increase the likelihood of mechanical ventilation remains speculative. Nevertheless, it is interesting to note that administration of medium and long-chain triglycerides can aggravate gas exchange abnormalities in ARDS⁴⁰. As for the duration of mechanical ventilation, we discovered an inverse correlation with IDUA, a lysosomal enzyme required for the degradation of heparan sulfate and dermatan sulfate⁴¹. In a case report of Hurler syndrome, associated with loss-of-function of IDUA, progressive respiratory failure was reversed with laronidase replacement therapy⁴², supporting the positive association between IDUA activity and respiratory function. Finally, these molecular features play a greater role in EP6 than the other EPs, demonstrating once again, that even for mechanical ventilation there are many molecular paths that can lead to this debilitating complication. Therefore, treatment needs to be tailored to the dominant molecular features within patients’ subgroups.

Subphenotypes of ARDS have been previously recognised, with a more inflammatory phenotype associated with increased mortality⁴³. Interestingly, analogously to how we derived a prognostic model using routine clinical laboratory measurements to assign EPs, a classifier model using clinical variables was able to assign participant of an observational cohort to one of the two subtypes of ARDS⁴⁴. This demonstrates the strength of first performing deep phenotyping on a small subset of participants and then generalizing the insights using routinely available information on larger cohorts.

A recent study in COVID-19 ARDS found that subtypes of severe COVID-19 differed from non-COVID subtypes of ARDS, with the subtype showing evidence of disordered coagulation having higher mortality rate⁴⁵. Our own study supports the idea of the existence of different subtypes of ARDS from a circulating proteome point of view, with the EP6-ARDS characterised by abnormal coagulation associated with worst outcome, in accordance with the aforementioned study. Taken together, these studies highlight the heterogeneity of ARDS, and the importance of identifying underlying features to properly tailor therapeutic interventions^46,47. Accordingly, molecular features such has elevated levels of ANGPTL3 could serve as biomarker for targeted intervention and may warrant the preclinical investigation of ANGTPL3 antagonists like evinacumab or other functionally similar therapeutic agents⁴⁸ in ameliorating outcomes of ARDS.

The data presented in this study comes from individuals participating in the BQC19, a prospective observational cohort built to study COVID-19 in Québec (Canada) with its specific population profile as reported previously¹⁰. A chronological bias may be present, as most of the participants used for endotyping in this study were recruited during the first two waves of the pandemic (Fig. S2), prior to widespread vaccination in Québec and the appearance of the Omicron variant and sub-variants. Therefore, some features of the identified endotypes may have changed over the course of the pandemic. Additionally, the investigation of genotypes associated with EP6 was performed as an exploratory study due to the limited number of participants that had genomic and proteomic information. It will be important to investigate SNV in SHC4 and CACNA2D3 in larger cohorts. For example, SNVs in CACNA2D3 have been associated with type 2 diabetes⁴⁹, a trait that can plausibly influence COVID-19 outcomes. It will be essential to continue longitudinal assessments of the molecular profiles to better understand the dynamic nature of host-pathogen interactions in the face of the future emerging SARS-CoV-2 variants of concerns.

Conclusion

Respiratory infections represent an important challenge for respirologists and critical care physicians due to the heterogeneity of outcomes. Developing better ways to prognosticate poor outcomes is crucial in improving patients’ care and survival. Our work showcases an analytical pipeline that leverages detailed proteomic information yet, in a clinical setting, just requires routinely available information for prognostication to favor translation that may find applications in many other diseases beyond COVID-19. It also demonstrates the complexity of molecular pathogenesis, with distinct proteomes associated with a similar outcome, such as ARDS. Furthermore, implementing data-driven approaches should improve our capacity at delivering personalized medicine.

Author contributions

Study concept and design: A.E., S.R. Acquisition of data: W.M., A.S., S.R., A.E. Analysis and interpretation of data: W.M., A.S., C.A., M.D., K.T., S.R., A.E. Critical revision of the manuscript for important intellectual content: W.M., A.S., M.D., K.T., S.R., A.E. Statistical analysis: W.M., C.A., A.S. Study supervision: A.E, S.R., K.T.

Peer review

Peer review information

Communications Medicine thanks Chao Liang, Namshik Han and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Data availability

BQC19 utilizes a central data management platform, where de-identified clinical and experimental data is stored along with the inventory of samples available for each participant (Secure data 4 health-sd4h cloud-based server). In accordance with the FAIR principles, these data are accessible upon request based on an open-to-public interface policy of BQC19 for content discovery. Therefore, low-sensitivity patients’ data are available through BQC19 (www.bqc19.ca) website or from Bento platform available at https://bqc19.bento.sd4h.ca/en/ or can be requested by email to info@bqc19.ca. Data that has a direct or high risk of re-identification will also go through a tightly controlled access process available at BQC19.ca. The detailed process, available on the biobank’s website (https://www.bqc19.ca/en/access-data), involves completing an application form to explain the research project. Researchers from both academia and private entities are eligible to apply, and the research project must be approved by a research ethics board. A Data Access Committee will then review the application with seven working days, and upon approval, a Data Use Agreement will be signed, which restricts redistribution, re‑identification attempts, and use outside the approved research scope. The data will be made available to the applicant using Globus file transfer technology. The Source Data for main figures are provided as Supplementary Data 11.

Code availability

Since our code is deeply intertwined with BQC19 data and contains protected information regarding BQC19 participants, the code cannot be deposited on public servers in order to respect the participants’ consent and protect their identity. However, we will provide both our code and intermediary data to any requester who has been granted access to BQC19 data. We also paid special attention to providing all information and details necessary to reproduce our tests in their descriptions.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s43856-026-01520-x.