Associations of the plasma lipidome with mortality in the acute respiratory distress syndrome: a longitudinal cohort study

The Institutional Review Board at the University of Michigan approved this study. Samples were previously collected as part of a multi-centered prospective clinical trial examining the utility of GM-CSF for critically ill adult patients with ARDS. Research funds allowed for the analysis of 60 samples. Therefore, 30 subjects with samples available from two time points were selected to maximize sequential organ failure assessment (SOFA) scores. Two time points were analyzed in order to compare the importance of the change in lipid concentrations over time with the absolute lipid concentrations measured early in the course of disease. The details of the original study, which randomized patients with acute lung injury or ARDS to receive GM-CSF or placebo, have been published [11]. The definition of ARDS was consistent with guidelines that were available at the time of the original study [12]. Exclusion criteria for the original study included evidence of preexisting chronic respiratory failure, neutropenia (absolute neutrophil count < 1000 cells/mm³), and a history of hematologic malignancy or bone marrow transplantation. Ventilator management followed a protocol that focused on minimizing tidal volumes and plateau pressures [13].

Plasma samples were stored at − 80 °C until the time of analysis, which was conducted by the Michigan Regional Comprehensive Metabolomics Resource Core. Lipids were extracted using a modified Bligh-Dyer Method [14]. The extraction was performed using water/methanol/dichloromethane (2:2:2 v/v/v) at room temperature after the addition of internal standards. The organic layer was then collected and dried under a stream of nitrogen before being re-suspended in 100 μL of Buffer B (5%H₂O:10%ACN:85%IPA containing 10 mM NH₄OAc) and analyzed using a liquid chromatography tandem mass spectrometry (LC/MS/MS) lipidomics assay [15].

The lipid extract was injected onto a 1.8 μm particle 50 × 2.1 mm internal diameter Waters Acquity HSS T3 column (Waters, Milford, MA) that was heated to 55 °C. For chromatographic elution, we used a linear gradient over a 20 min total run time. A 60% Solvent A and 40% Solvent B gradient was used for the first 10 min. Then the gradient was ramped in a linear fashion to 100% Solvent B which was maintained for 7 min. Thereafter, the system was switched back to 60% Solvent B and 40% Solvent A for 3 min. The flow rate used for these experiments was 0.4 mL/min and the injection volume was 5 μL. The column was equilibrated for 3 min before the next injection and run at a flow rate of 0.400uL/min for a total run time of 20 min. Data were acquired in positive and negative modes using data-dependent MS/MS with dynamic mass exclusion. Pooled human plasma samples and pooled experimental samples (prepared by combining small aliquots of each experimental sample) were used to control for the quality of sample preparation and analysis [16]. A randomization scheme was used to distribute pooled samples within the set and a mixture of pure authentic standards was used to monitor instrument performance on a regular basis.

Lipids were identified using the LIPIDBLAST computer-generated tandem MS library [17]. This database contains 212,516 spectra covering 119,200 compounds representing 26 lipid classes, including phospholipids, glycerolipids, bacterial lipoglycan, and plant glycolipids. Quantification of lipids was completed using AB-SCIEX MultiQuant software. After excluding compounds with a relative standard deviation greater than 30% in the pooled samples, the raw mass spectrometry data were normalized using the internal standards and adjusted for batch effects using loess smoothing. The nomenclature used for individual lipids begins with the abbreviation of the lipid class (Table 1) followed by the number of carbon atoms in the molecule and then by the number of double bonds. If the same lipid was detected in both the positive and negative modes, an underscore after the name of the lipid was used to denote the second measurement.

We removed lipids with coefficients of variation > 30% to focus on more precisely measured compounds. The remaining concentrations were log transformed and auto scaled (mean-centered and divided by the standard deviation of each variable) to create normally distributed lipid concentrations and make the lipid concentrations more comparable to each other, respectively. To ensure that the administration of GM-CSF did not impact lipid concentrations, the lipid levels were compared between those did and did not receive the study drug. Next, the differences in lipid concentrations between survivors and non-survivors were compared using repeated measures analysis of variance. Lipid concentrations were the independent variable and mortality was the dependent variable. False discovery rate (FDR) adjusted p-values (q-values) were used to account for multiple comparisons [18]. A q-value less than 5% was considered statistically significant. Heatmaps were used to display both the relative concentrations of significant lipids and the Pearson product-moment correlation coefficient (PPMCC) between significant lipids. The PPMCC was calculated separately for each combination of lipids analyzed. Scatterplots of each lipid relationship were examined to identify the presence of nonlinear relationships. Hierarchical clustering was performed using Pearson distance and the complete linkage clustering algorithm. The overall differences in the variability and distribution of lipid concentrations between survivors and non-survivors was evaluated using principal component analysis (PCA). Finally, in order to describe associations between lipid concentrations and outcomes, the area under the receiver operating characteristic (AUROC) for the lipid concentrations at the first time point were compared with the Acute Physiology Score (APS) from the Acute Physiology and Chronic Health Evaluation III [19] at randomization and the SOFA [20] score obtained at the time of sample collection. Software used in the analysis of these data were MetaboAnalyst 3.0 (Xia Lab, McGill University, Montreal, Canada) [21], SAS software (Version 9.4, SAS System for Windows, SAS Institute Inc., Cary, NC, USA), and R (version 3.2.1; R Foundation, Vienna, Austria).