To derive a plasma metabolite panel for distinguishing between COVID-19 patients and healthy controls, significant variables (p < 0.05) after adjustment of age, sex, and BMI were included in a starting pool. In an iterative process, ten sets of variables were established from the starting pool (Figure 1 ), and one representative variable was finally selected from each set based on (1) smallest p value and (2) reported biological function from a PubMed search. A final panel of ten metabolites was generated, which separated healthy controls from COVID-19 patients with AUC = 0.975 in a logistic regression model with leave-one-out (LOO) cross validation. Among these metabolites, sphingosine-1-phosphate (S1P) was reduced (p < 0.001) in COVID-19, and its level was raised (p = 0.0065) at hospital discharge relative to admission in a small subset of patients followed longitudinally (Figure S2). S1P generation via sphingosine kinase-2 in monocyte-derived macrophages was recently shown to promote the resolution of inflammation by alveolar macrophages in acute lung injury (Joshi et al., 2020). Biliverdin, the oxidized form of bilirubin, is part of the redox cycle constituting the primary physiologic function of bilirubin as a cytoprotective antioxidant (Baranano et al., 2002). Increases in biliverdin (p = 0.0077) in COVID-19 probably indicated enhanced oxidative stress in disease state, and its level was reduced longitudinally at hospital discharge with marginal significance (p = 0.0558) (Figure S2). Plasma 5-hydroxy-tryptophan was elevated in COVID-19 (p = 0.0203), and its depletion via induction of the indolamine 2, 3-dioxygenase pathway in human alveolar carcinoma type II-like cells was previously reported to suppress the growth of parainfluenza virus type 3 (Rabbani and Barik, 2017). As for lipids, increases in lysophopholipids including lysophosphatidic acid (LPA) 18:1 (p = 0.0249) and lysophosphatidylcholine (LysoPC) 18:1 (p = 0.0350) were observed in COVID-19, while neutral lipids including medium-chain TAG 48:1(18:0) (p = 0.0190), long-chain TAG 60:3(18:1) (p = 0.0390), and DAG 34:1(16:1/18:0) (p = 0.0010) were generally reduced. On the other hand, sphingolipids such as SM d18:1/18:1 (p = 0.002) and GM3 d18:1/25:0 (p = 0.0036) were generally increased with disease.
Figure 1 Plasma Panel for Differentiating COVID-19 Patients from Heathy Controls
Overview of selection scheme for plasma metabolite panel to differentiate COVID-19 (n = 50) patients from healthy controls (n = 26).
(A) From a total of 1,002 variables measured (598 lipids and 404 polar metabolites), variables with p < 0.05 between healthy controls and COVID-19 patients after adjustment for age, sex, and BMI were sieved out to form a starting pool comprising 322 variables.
(B) A starting variable with the lowest p value was selected, and variables with significant correlations (p < 0.05) to the starting variable selected were removed from consideration. From remaining variables in the starting pool, the next starting variable with the second lowest p value was identified, and the process was repeated in an iterative fashion until all variables in the starting pool were exhausted. This process generated a list of ten variables.
(C) Variables with significant correlations (p < 0.05) to each of the selected variables were added together to form ten established sets. To select a representative variable from each established set, the variable with the smallest p value and with reported biological function from a Pubmed search was chosen. A final panel of ten plasma metabolites, including S1P d18:1, SM d18:1/18:1, TAG60:3(18:1), LPA 18:1, biliverdin, TAG 48:1(18:0), DAG34:1(16:1/18:0), GM3 d18:1/25:0, lysoPC18:1, and 5-hydroxy-L-tryptophan, was generated, which distinguished between healthy controls and COVID-19 patients with an area under the curve (AUC) = 0.975 in a logistic regression model with leave-one-out (LOO) cross-validation. Boxplots for the ten selected metabolites in the final panel were illustrated and p values were indicated on top of each boxplot. Levels of polar metabolites measured using untargeted metabolomics were presented as corrected intensities, and lipids quantitated using targeted lipidomics were presented in nanomoles of lipids per liter (nmol/L) plasma.
See also Figures S2–S4.