Identification of immune cell mitochondrial dysfunction characteristics and clinical predictive biomarkers in sepsis via multi-cohort machine learning and single-cell RNA sequencing

Abstract

Sepsis involves dysregulated host responses with mitochondrial dysfunction as a key mechanism, yet its immune cell landscape and clinical utility are unclear. We integrated five GEO datasets (GSE65682, GSE95233, GSE54514, GSE26440, GSE26378) comprising 1,256 samples. Mitochondrial dysfunction-related genes were identified through differential expression analysis and WGCNA. Ten machine learning algorithms were compared to construct an optimal diagnostic model. NMF clustering defined sepsis subtypes, and a mitochondrial dysfunction score (MAS) was developed. Single-cell RNA sequencing (GSE167363, 59,731 cells) validated cell-type-specific expression patterns. Furthermore, Western blotting was performed to confirm the protein-level expression of selected hub genes in PBMC samples. 438 differentially expressed genes were identified (224 upregulated, 214 downregulated), with upregulated genes enriched in oxidative phosphorylation and downregulated genes in T cell receptor signaling. Elastic net achieved optimal diagnostic performance (AUC = 0.9156), yielding a 12‑gene signature with robust external validation (AUC up to 0.918). NMF stratified patients into two subtypes (Cluster 1, n = 412; Cluster 2, n = 511) with distinct immune profiles; Cluster 2 exhibited elevated NLRP3/IL1B expression and enhanced innate immune infiltration. MAS significantly correlated with sepsis status (p = 7.56e‑27) and showed superior diagnostic performance (AUC up to 0.969). Single‑cell analysis revealed MAS enrichment in monocytes and dendritic cells, with positive correlation to monocyte function (rho = 0.443). Notably, protein-level validation confirmed that TXN, CTSZ, SLPI, PSMC5, and DNM2 were significantly upregulated, while FLI1 was downregulated in sepsis patients, consistent with the transcriptomic findings. Furthermore, clinical correlation analysis revealed that the signature score was significantly positively correlated with SOFA and APACHE II scores, and patients with high signature scores exhibited significantly worse 28-day survival. Pseudotime trajectories showed distinct differentiation patterns (P < 0.001). This comprehensive multi-omics analysis, further supported by experimental protein-level validation and clinical correlation with severity scores and survival outcomes, establishes mitochondrial dysfunction-immunity crosstalk as a central axis in sepsis pathogenesis. The identified 12-gene signature and MAS score provide robust diagnostic tools with translational potential for clinical sepsis management.