Construction of a sepsis recovery-failure axis and single-cell anchoring using public multi-cohort whole-blood and single-cell transcriptomic data.

Sepsis exhibits marked heterogeneity in host responses. Although prior time-course transcriptomic studies have provided important information about sepsis diagnosis and mortality-associated gene-expression dynamics, many bioinformatic signatures remain optimized for cross-sectional classification or endpoint labels. Using public multi-cohort whole-blood transcriptomic and single-cell datasets, we constructed a sepsis recovery-failure axis constrained by longitudinal trajectory rules and evaluated its cross-cohort transferability, clinical relevance, and single-cell anchoring. In GSE54514, we identified 20 recovery-related genes and 20 failure-related genes and defined a recovery-failure score. Survivors showed a non-monotonic but ultimately upward movement by day 5, whereas non-survivors progressively shifted toward the failure end of the axis, with clear separation by day 5. Linear mixed-effects models using all repeated measurements from day 1 to day 5 showed a significant interaction between time and outcome that remained stable after adjustment for neutrophil proportion and baseline APACHE II. GSE57065 and GSE65682 supported transfer of the axis for host-state stratification, while also showing that cohort-wise standardization limits interpretation as a universal absolute cutoff. Additional null and cross-cohort structure analyses showed that external-cohort patient-level score variance exceeded matched random-gene expectations and that high-score patients across external cohorts shared non-axis transcriptomic effects. The score showed only weak correlations with bedside severity scores, suggesting that it may serve as a complementary host-state indicator rather than a replacement for clinical severity scores. At the single-cell level, broad PBMC annotation followed by monocyte refinement localized the strongest donor-level axis signal to classical monocytes; additional T-cell cluster screening did not identify a robust hidden T-cell cluster that superseded monocyte anchoring. In non-survivors, classical monocytes exhibited reduced interferon responses and oxidative phosphorylation together with enhanced inflammatory, IL6-JAK-STAT3, and coagulation programs. Axis displacement within interferon-activated monocytes was supported by limited cellular coverage and should be treated as exploratory. Taken together, the recovery-failure axis is a longitudinally derived host-state score that may support sepsis state stratification, disease-course monitoring, and hypothesis generation, but