Metabolic septic shock sub-phenotypes, stability over time and association with clinical outcome

Abstract

Purpose: Machine learning has shown promise to detect useful subgroups of patients with sepsis from gene expression and protein data. This approach has rarely been deployed in metabolomic datasets. Metabolomic data are of interest as they capture effects from the genome, proteome, and environmental. We aimed to discover metabolic sub-phenotypes of septic shock, examine their temporal stability and association with clinical outcome.

Methods: Analysis was performed in two double-blind randomized trials in septic shock (LeoPARDS (1402 samples from 470 patients) and VANISH (493 samples from 173 patients)). Patients were included soon after the onset of shock and had serum collected at up to four time points. Metabolic clusters were identified from 474 metabolites using k-means clustering in LeoPARDS and predicted in VANISH with an elastic net classifier.

Results: Three sub-phenotypes were found. The main determinants of cluster membership were lipid species, especially lysophospholipids. Low lysophospholipid sub-phenotypes were associated with higher circulating cytokine levels. Persistence of low lysophospholipid sub-phenotypes was associated with higher mortality compared to the high lysophospholipid sub-phenotype (LeoPARDS: cluster 2 odds ratio 3.66 (95% CI 1.88-7.20), p = 0.0001, cluster 3 2.49 (1.29-4.81), p = 0.006; VANISH: cluster 2 4.13 (1.17-15.61), p = 0.03), cluster 3 3.22 (1.09-9.92), p = 0.04, vs cluster 1). We found no heterogeneity of treatment effect for any of the trial interventions by baseline metabolic sub-phenotype.

Conclusion: Three metabolic subgroups exist in septic shock which evolve over time. Persistence of low lysophospholipid sub-phenotypes is associated with mortality. Monitoring these subgroups could help identify patients at risk of poor outcome and direct novel therapies such as lysophospholipid supplementation.

Registration: Clinicaltirals.gov Identifiers, VANISH: ISRCTN 20769191, LeoPARDS: ISRCTN12776039.

Keywords: Clustering; Lysophospholipids; Metabolomics; Sepsis; Triglycerides.

Fig. 1

Heat map (A) showing the distribution of metabolite classes between the three metabolomic clusters (cluster 1 in red, cluster 2 in blue, cluster 3 in yellow) defined by k-means clustering. B Volcano plots comparing metabolite differences in the baseline sample between clusters 1 vs 2, clusters 1 vs 3 and clusters 2 vs 3 in the LeoPARDS dataset. Metabolites were considered significant if they had an adjusted p value of < 0·05 and a fold change > 1·5 (the dashed lines represent these thresholds). Significant metabolites are colored by class as shown in the key. (A1 apolipoprotein A-1, AB apolipoprotein B, CAR carnitine, CE cholesterol ester, CH cholesterol, FC free cholesterol, H high density lipoprotein subspecies, HDFC high density lipoprotein free cholesterol, IDCH intermediate density cholesterol, IDFC intermediate density lipoprotein free cholesterol, IDPL intermediate density lipoprotein phospholipids, IDTG intermediate density lipoprotein triglycerides, L low density lipoprotein subspecies, LPC Lysophosphatidylcholine, LPE Lysophosphatidylethanolamine, LPG Lysophosphatidyl glycerol, PL phospholipids, SM Sphingomyelins, TG Triglycerides, V very low density lipoprotein subfractions

Fig. 2

Principal component analysis (PCA) scores plot (A) combining samples from LeoPARDS (open circles) and VANISH (closed circles) showing the similarity in cluster structure between k-means clustering derived clusters in both datasets (cluster 1 in red, cluster 2 blue, cluster 3 yellow). Receiver operating characteristic curve (B) comparing false negative rate (1-specificity) against sensitivity for the elastic net classifier to predict VANISH k-means clusters (cluster 1 vs others in red, cluster 2 vs others in blue, cluster 3 vs others in yellow). Correlation plots (C) comparing the fold change of all metabolites measured in baseline samples in cluster 1 vs 2, cluster 1 vs 3 and cluster 2 vs 3 between clusters predicted by the elastic net classifier and by unsupervised k-means clustering in VANISH. R = Spearman’s correlation coefficient, points are colored by significance in both (pale blue), k-means only (yellow), elastic net only (green) or neither (grey) models with significance being defined as a corrected p-value < 0·05 and a fold change > 1·5

Fig. 3

A Alluvial plots showing the transition between clusters over time for LeoPARDS and VANISH (red = cluster 1, blue = cluster 2, yellow = cluster 3, white = no sample). Outcome on intensive care discharge after the final sample was collected is shown with the final column representing the outcome at day-28 (green = alive, grey = dead). B Proportion of non-survivors by cluster transition from baseline to the final sample collected, analysis was restricted to patients who had a sample collected at baseline and at least one other sample. X-axis shows the cluster transition ranked by proportion of deaths in LeoPARDS with first cluster—final cluster, e.g., 3–1 represents patients who moved from cluster 3 to cluster 1, n gives the number of patients with each transition. C) Kaplan–Meier curves for 28-day mortality from the time of final sample collection (time 0), crosses represent censored data, p value given by the log-rank test (red lines = cluster 1, blue lines = cluster 2, yellow lines = cluster 3)