Classification of patients with sepsis according to blood genomic endotype: a prospective cohort study

Background Host responses during sepsis are highly heterogeneous, which hampers the identification of patients at high risk of mortality and their selection for targeted therapies. In this study, we aimed to identify biologically relevant molecular endotypes in patients with sepsis. Methods This was a prospective observational cohort study that included consecutive patients admitted for sepsis to two intensive care units (ICUs) in the Netherlands between Jan 1, 2011, and July 20, 2012 (discovery and first validation cohorts) and patients admitted with sepsis due to community-acquired pneumonia to 29 ICUs in the UK (second validation cohort). We generated genome-wide blood gene expression profiles from admission samples and analysed them by unsupervised consensus clustering and machine learning. The primary objective of this study was to establish endotypes for patients with sepsis, and assess the association of these endotypes with clinical traits and survival outcomes. We also established candidate biomarkers for the endotypes to allow identification of patient endotypes in clinical practice. Findings The discovery cohort had 306 patients, the first validation cohort had 216, and the second validation cohort had 265 patients. Four molecular endotypes for sepsis, designated Mars1-4, were identified in the discovery cohort, and were associated with 28-day mortality (log-rank p=0.022). In the discovery cohort, the worst outcome was found for patients classified as having a Mars1 endotype, and at 28 days, 35 (39%) of 90 people with a Mars1 endotype had died (hazard ratio [HR] vs all other endotypes 1.86 [95% CI 1.21-2.86]; p=0.0045), compared with 23 (22%) of 105 people with a Mars2 endotype (HR 0.64 [0.40-1.04]; p=0.061), 16 (23%) of 71 people with a Mars3 endotype (HR 0.71 [0.41-1.22]; p=0.19), and 13 (33%) of 40 patients with a Mars4 endotype (HR 1.13 [0.63-2.04]; p=0.69). Analysis of the net reclassification improvement using a combined clinical and endotype model significantly improved risk prediction to 0.33 (0.09-0.58; p=0.008). A 140-gene expression signature reliably stratified patients with sepsis to the four endotypes in both the first and second validation cohorts. Only Mars1 was consistently significantly associated with 28-day mortality across the cohorts. To facilitate possible clinical use, a biomarker was derived for each endotype; BPGM and TAP2 reliably identified patients with a Mars1 endotype. Interpretation This study provides a method for the molecular classification of patients with sepsis to four different endotypes upon ICU admission. Detection of sepsis endotypes might assist in providing personalised patient management and in selection for trials