Multi-cohort consensus clustering identifies three distinct transcriptomic endotypes in sepsis

Sepsis is a heterogeneous syndrome in which patients with the same clinical diagnosis may harbour different host immune responses. Existing transcriptomic endotyping frameworks differ in gene selection, clustering, and cohort composition, limiting comparability. Few studies have distinguished endotype-intrinsic prognostic signal from baseline severity, and most prior frameworks relied on curated immune-gene panels or single-cohort discovery.

We performed consensus clustering on 1002 sepsis patients from 5 discovery cohorts after quantile normalisation, ComBat batch correction, and data-driven feature selection (122 genes with median absolute deviation [Formula: see text] from a 2456-gene set common to nine platforms), using PAM clustering with Pearson correlation distance ([Formula: see text]-10, 1000 bootstrap iterations) and composite metric-based k selection. Endotype-mortality associations were assessed by logistic regression and, where severity scores were available, by SOFA-adjusted models; a LASSO multinomial classifier evaluated generalisability across 4 external cohorts.

Three endotypes were resolved at optimal [Formula: see text]: C1 (Immune Activation, [Formula: see text], 30.5%), C2 (Interferon Response, [Formula: see text], 44.6%), and C3 (Erythroid Dysregulation, [Formula: see text], 24.9%). Mortality differed across endotypes (logistic regression likelihood ratio [Formula: see text]), with C2 showing lower 28-day mortality than C1 (15.4% vs. 26.1%; OR [Formula: see text], 95% CI [Formula: see text], [Formula: see text]). Crucially, SOFA-adjusted analysis in GSE185263 ([Formula: see text], the only discovery cohort with SOFA scores) fully attenuated the C2 advantage (adjusted OR [Formula: see text], [Formula: see text]), indicating that in this cohort endotype-mortality associations are primarily explained by baseline severity. A LASSO-regularised multinomial classifier achieved >98% per-class sensitivity and generalised to 4 external cohorts ([Formula: see text]) with 100% gene coverage and mean classification confidence 0.86-0.94. Longitudinal analysis of 129 patients across four multi-timepoint datasets demonstrated 66.7% endotype stability (Cohen's [Formula: see text]).

Three severity-aligned transcriptomic patterns-immune activation, interferon response, and an underrecognised erythroid/metabolic axis-recur across heterogeneous sepsis cohorts. In the one cohort with organ-dysfunction scores, the C2 mortality advantage was fully attenuated after SOFA adjustment, indicating that-in this cohort-endotype-mortality associations are largely explained by baseline severity rather than independent prognostic biology; whether this generalises requires cohorts with harmonised severity data. Future endotyping studies should incorporate systematic severity adjustment before drawing prognostic or treatment-stratification conclusions.