Early sepsis is characterized by dynamic shifts in immune-cell transcriptional programmes. The authors hypothesised that the trajectory—that is, the direction and magnitude of change over time—of predefined immune transcriptional modules might carry prognostic information for sepsis mortality beyond single baseline measurements. They sought to determine which trajectory associations survive independent testing and clinical-analogue triangulation.
Five immune-related modules were defined and frozen before any testing. These modules were scored in three publicly available longitudinal whole-blood microarray sepsis cohorts. The primary discovery analysis used the two cohorts that recorded mortality (combined 82 patients, 24 deaths). For each module, the authors fitted a logistic model fixed in advance that estimated the change in odds of death per 24-hour change in module score, expressed per one standard deviation, and pooled cohort-specific estimates by inverse-variance fixed-effect meta-analysis. Benjamini–Hochberg correction was applied to control the false discovery rate. No post-hoc re-fitting was reported; sensitivity analyses and alternative scorings were prespecified.
In discovery, none of the module-trajectory associations survived correction for multiple testing. Two directional signals emerged as the leading, uncorrected associations: a rising CD4/NK lymphocyte trajectory associated with lower mortality (pooled odds ratio 0.53, 95% CI 0.31–0.90 per 1 SD) and a rising emergency-granulopoiesis trajectory associated with higher mortality (pooled odds ratio 1.60, 95% CI 0.92–2.79 per 1 SD). Both signals were identified from the a priori-scored modules applied to the discovery cohorts.
The two leading discovery signals were tested in an independent transcriptomic cohort with serial sampling (ArrayExpress/BioStudies accession E-MTAB-5273), consisting of 63 patients with 15 deaths. The same frozen scoring approach was applied. In contrast to discovery, independent testing separated the two signals in the opposite order: the emergency-granulopoiesis trajectory reproduced directionally and with an effect size similar to discovery but did not reach conventional statistical significance on its own (validation OR 1.72, 95% CI 0.92–3.20; p = 0.088). When discovery and validation were pooled, the combined estimate for emergency-granulopoiesis was 1.65 (95% CI 1.09–2.50).
The CD4/NK trajectory did not reproduce in the independent transcriptomic cohort (validation OR 1.06, 95% CI 0.59–1.89) and was null in the temporal window most favorable to it.
To assess whether transcriptomic trajectories had corresponding signals in routine clinical data, the authors derived cell-count analogues and tested them in MIMIC-IV v3.1, a large intensive-care database. The MIMIC-IV cohort included 12,607 adults meeting Sepsis-3 criteria; depending on the analogue, between 744 and 4,206 patients had the serial measurements required. Two of three emergency-granulopoiesis analogues supported the transcriptomic validation signal. Notably, the neutrophil-to-lymphocyte ratio (a clinical analogue) showed a positive association consistent with the emergency-granulopoiesis direction (OR 1.31, 95% CI 1.20–1.42).
By contrast, an analogue that was well powered to detect the discovery CD4/NK effect provided no supporting evidence for that trajectory.
The emergency-granulopoiesis association was positive in all nine sensitivity analyses that were fixed before estimates were examined. Although the transcriptomic validation estimate did not reach conventional statistical significance alone (p = 0.088), the directional consistency across sensitivity analyses and support from clinical analogues strengthened the case for reproducibility. The CD4/NK discovery signal, however, failed independent reproduction and received no support in the prespecified analogue analyses.
All analyses used existing publicly available or credentialed-access datasets. Discovery transcriptomic datasets were from GEO accessions GSE95233, GSE54514, and GSE57065; the validation transcriptomic cohort was E-MTAB-5273. Cross-modality analyses used MIMIC-IV v3.1 via the PhysioNet credentialing process. The authors archived the frozen per-patient exposure tables, module-score tables, analysis code, and derived datasets in Zenodo and GitHub with links and a machine-readable evidence map. The authors declared no competing interests. Because this was a secondary analysis of de-identified data, the authors report that no new ethics approval or exemption was required; MIMIC-IV creation was reviewed by an IRB that granted a waiver.
This cross-cohort transcriptomic analysis demonstrates that not all discovery-stage immune-trajectory signals reproduce when tested independently and when triangulated against routine clinical analogues. The emergency-granulopoiesis trajectory—reflecting granulocytic emergency myelopoiesis—was directionally reproduced and supported by clinical analogues, including the neutrophil-to-lymphocyte ratio, with pooled evidence suggesting an association with higher sepsis mortality. The CD4/NK lymphocyte trajectory, although prominent in discovery, failed independent validation and lacked clinical-analogue support.
Key limitations reported in the source include that no discovery association survived multiple-testing correction, the transcriptomic validation estimate did not reach conventional significance on its own, and analyses rely on secondary use of existing datasets with their inherent heterogeneity. The authors provide complete code and frozen result tables to facilitate replication. These results highlight the importance of independent testing and cross-modality triangulation when evaluating dynamic immune biomarkers in sepsis and suggest emergency-granulopoiesis trajectories merit further prospective evaluation.