Quantification and Statistical Analysis

Metabolite Panel for Identifying COVID-19
To generate a plasma metabolite panel for differentiating COVID-19 patients from healthy individuals, variables with p < 0.05 between healthy controls and COVIDP19 patients after adjustment for age, sex and BMI were sieved out to form a starting pool. From this pool, a starting variable with lowest p value was added to Set 1, and remaining variables from the starting pool significantly correlated (p < 0.05) with this starting variable were added together to form Set 1. The process then was repeated in an iterative process using starting variable with the second lowest p value, and so on, finally generating a total of ten established sets. Representative metabolite from each established set was chosen based on (1) smallest p value and (2) reported biological function through a PubMed search. The selection process finally created a panel of ten plasma metabolites, and its performance was evaluated in a logistic regression model with leave-one-out (LOO) cross-validation, which distinguished between COVID-19 patients and healthy controls with an area under curve (AUC) = 0.955.

Logistic Regression Analysis
Logistic regression model with covariates BMI, age and sex were built with each variable (lipid/metabolite) to search for significant variables that can predict the conditions (i.e., healthy control, mild, moderate, severe COVID-19) of subjects. The p value of the variable estimate in the model was extracted and those with false discovery rate (fdr) smaller than 0.05 were shortlisted. Forest plots of these significant variables were constructed and the boxplots of individual significant variables were generated. A list of 61 lipids and 89 metabolites were found to be significant (fdr < 0.05) in their respective logistic regression models. Forest plots illustrate the magnitude of estimates on signed log x axis, with indicator of significance of the estimate in the model, with ∗∗∗ representing p < 0.001, ∗∗ representing p < 0.01 and ∗ representing p < 0.05. For non-significant species, the estimates were plotted as zeros.

Spearman Correlation Analysis
Spearman correlation between variables and clinical indicators were performed with data from COVID-19 patients. For each pair of variable and clinical indicator, samples with missing clinical data were omitted from the calculations. Correlation plots for all variables grouped by metabolite class were presented. Only correlations with p < 0.05 were indicated with colored circles. Negative correlations were shown in red and positive correlations were shown in blue, with sizes of circles representing the magnitude of the correlations.

Differential Correlation Analysis
MEGENA R package was used to build correlation networks from differentially correlated lipid pairs in mild COVID-19 relative to healthy controls to reveal changes in lipid co-regulation upon early infection. Differential correlation was calculated using R package DGCA. Only lipid pairs with differential correlation (empirical p < 0.05) were included for analyses.