Metabolomic profiling and prognostication in COVID-19 acute respiratory distress syndrome

To the Editor,

Coronavirus disease 2019 (COVID-19)-related acute respiratory distress syndrome (ARDS) is biologically and clinically heterogeneous with a broad spectrum of organ dysfunction and illness severity.^[1] In COVID-19-related ARDS (C-ARDS), two phenotypes have been identified with latent class analysis (LCA) that are associated with divergent clinical outcomes and response to corticosteroids, and correspond to previously identified hypoinflammatory and hyperinflammatory phenotypes.^[2,3] Despite insights gained from this work, understanding the biological processes that determine mortality risk in C-ARDS remains imprecise. Here, we report that metabolomic profiling of plasma at the onset of C-ARDS provides insights into pathogenesis and may predict clinical outcomes.

This was a retrospective, matched cohort study. Participants were adults with COVID-19 who met Berlin criteria for ARDS on the initial day of mechanical ventilation. Patients were cared for at Columbia University hospitals from March 2, 2020 to April 30, 2020. Exclusion criteria included death before intensive care unit admission, pre-existing tracheostomy, or endotracheal intubation before transfer from an outside hospital. Of 558 consecutively admitted patients screened, 483 met inclusion for the broader cohort study. For this specific pilot study of metabolomics profiling, 25 survivors to 90 days were matched on age, sex, and ethnicity – factors known to influence the metabolomic profile – to 25 patients who died within 28 days of intubation (Supplementary Figure S1). Ethnicities were self-identified by the patient and classified as Hispanic, black, white, or other. All participants had prospectively banked plasma samples collected within 1 week of endotracheal intubation; patients without banked samples in this timeframe were excluded.

Untargeted and targeted metabolomic analysis was performed using mass spectrometry and compared between survivors and non-survivors. Data from metabolomics assays were normalized using systematic error removal with random forest for eliminating the unwanted systematic variations in large sample sets. Metabolites with levels below the detection limit in more than 50% of the study population were eliminated from further analyses. Levels of each metabolite were then log-transformed and divided by the standard deviation within the survivor group. Please see additional details on methods in the supplementary materials.

Statistical analyses were performed with conditional logistic regression models in which the binary outcome (survival vs. non-survival at 90 days) was fitted as the dependent variable and the levels of each metabolite were fitted as independent variable adjusting for body mass index (BMI) and severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) cycle threshold (CT) value. Bayesian analysis was implemented to calculate the Bayes factor (BF) and the 95% highest density credible interval. A biological clustering analysis was performed using Chemical Similarity Enrichment Analysis (ChemRICH) to categorize the metabolites. Competitive modeling by four machine learning models – Least Absolute Shrinkage and Selection Operator (LASSO), adaptive LASSO, Random Forest, and Extreme Gradient Boosting (XGBoost) – were used to predict mortality. In all models, we performed random-sampling cross-validation with 1000 iterations. In each iteration, a randomly selected 80% of the cohort samples were used as a training set, and the remaining 20% of the set was used as a test set. Three sets of predictors were explored: all metabolites, metabolites with BF >1, and metabolites with BF >3.

The 25 patients who died had a median survival time from intubation of 16 days (interquartile range: 14–21). The surviving and deceased cohorts were matched on age ([63±14] years vs. [63±14] years), sex (40% in both cohorts), and self-reported ethnicity (P=0.470). The cohorts had similar BMIs ([31.2±5.7] kg/m² vs. [33.9±10.1] kg/m², P=0.250), plasma SARS-CoV-2 cycle thresholds (28.2±4.5 vs. 25.8±4.0, P=0.250), and Sequential Organ Failure Assessment (SOFA) scores (9.5±3.2 vs. 10.1±2.9, P=0.530). The cohorts had similar rates of pre-existing hypertension (56% vs. 68 %, P=0.380), diabetes (48% vs. 40%, P=0.570), cardiac disease (28% vs. 28%, P >0.990), and smoking history (20% vs. 8%, P=0.420). Compared to LCA-derived subphenotypes, the patients who died by 90 days were more likely to be in C-ARDS class 2, which corresponded to the hyper-inflammatory phenotype (24% vs. 4%, P=0.042).^[3] The average time from intubation to the collection of plasma samples used for the metabolomic profiling was (0±3.7) days.

Metabolomic analyses yielded data for 30 bile acids, 340 biogenic amines, 522 complex lipids, 83 oxylipins, and 133 primary metabolites. Using a cut-off of a BF >3, 25 compounds were identified with significant differences between survivors and non-survivors (Supplementary Table S1). Five compounds had increased levels associated with mortality, and 20 had decreased levels associated with mortality. Biological clustering ChemRICH analysis on these compounds identified four key clusters of compounds – unsaturated and saturated lysophosphatidylcholines, plasmalogens, and saturated ceramides – that were decreased among non-survivors (Figure 1).

Figure 1.

Altered metabolite clusters in COVID-19 ARDS non-survivors vs. survivors showing the ratio of metabolite levels between the cohorts and the statistical significance for each cluster.

ARDS: Acute respiratory distress syndrome; COVID-19: Coronavirus disease 2019.

Machine learning-derived signatures were derived using all metabolites, metabolites with BF >1 (Figure 2), and those with BF >3 (Supplementary Figure S2). These signatures showed excellent discrimination in predicting mortality. The best model utilized Adaptive LASSO using metabolites with BF >1 and demonstrated an area-under-the-receiver-operating-characteristic curve of 0.912 (95% confidence intervals: 0.836 to 0.988, P=2.190×10⁻⁸). Using 0.5 as threshold, this model yielded 76% sensitivity and 76% specificity. It significantly outperformed LASSO and Random Forests using metabolites with BF >1 with P=0.027 and P=0.017, respectively, and the comparison with XGBoost using metabolites with BF >1 was not significant (P=0.323).

Figure 2.

Receiver operator curves for four different machine learning models to predict mortality based on altered metabolites with a BF >1.

AUC: Area under the curve; BF: Bayes factor; ROC: Receiver operating characteristic; LASSO: Least absolute shrinkage and selection operator; XGBoost: Extreme gradient boosting.

Metabolomic analysis identified differential enrichment of lipid metabolites in age, ethnicity, and sex-matched survivors compared to non-survivors with C-ARDS. Machine learning models accurately predicted mortality from C-ARDS based on metabolomic profiles from the day of intubation. The identified metabolite clusters may reflect altered biological pathways of interest within COVID-19 and ARDS. Lipodomic analyses have demonstrated profiles associated with mortality in both ARDS and sepsis, and in patients with ARDS survivors tended to have higher levels of most clinically relevant lipids, consistent with the findings in our study.^[4,5] While the mechanisms behind these associations remain to be fully elucidated, there are multiple proposed impactful pathways. Lysophosphatidylcholine has been shown to mediate inflammation in sepsis by regulating neutrophil activity; ceramide levels, and specifically the proportions of specific ceramides, have been linked to microvascular injury in COVID-19; and plasmalogens are protective antioxidants against cellular oxidative stress.^[[6], [7], [8], [9]^]

Our results are limited by the single-center study design and small sample size. Additionally, it is unclear if the metabolites are markers of biological derangement or play a pathophysiological function in the disease state. Further, it is not possible from these data to establish a causal link between any metabolite or broader profile and the patient outcomes, or to draw conclusions beyond the association observed. Finally, the results are also limited by the variable timing from intubation of the plasma samples used for metabolomic profiling as there may have been shifts in the metabolites analyzed even in a matter of days. Despite these limitations, our exploratory findings suggest that improved characterization of the metabolomic derangements in COVID-19 and ARDS may enhance understanding of drivers or markers of mortality and improve prognostication and precision therapy.

Managing Editor: Jingling Bao/ Zhiyu Wang