Integrative multi-omics and machine learning identify mitochondrial biomarkers for pathogen-specific sepsis stratification and translational prioritization

BACKGROUND: Sepsis is a leading cause of critical illness and mortality, yet substantial heterogeneity limits risk stratification and biomarker translation. Mitochondrial dysfunction is widely implicated in sepsis, but genetically supported, multi-layer regulatory features and their clinical relevance remain incompletely characterized. METHODS: We integrated publicly available sepsis GWAS summary statistics (general sepsis: 1634 cases/454,714 controls; gram-positive sepsis: 168/456,180; gram-negative sepsis: 383/455,965) with blood-based molecular QTL resources (including GTEx v8 whole blood, n = 670) to prioritize mitochondrial genes and infer regulatory cascades. Independent whole-blood transcriptomic cohorts (the GAinS cohort, GSE65682, n = 802; GSE54514, n = 163) were used for clinical and pathogen-specific expression characterization. We developed machine learning models using mitochondrial gene features and evaluated performance by internal tenfold cross-validation. RESULTS: We identified mitochondrial genes with convergent genetic, epigenetic, and transcriptional regulatory evidence, showing stronger effects in inner membrane and matrix compartments. Transcriptomic analyses supported clinically relevant dysregulation and pathogen-associated patterns. In predictive modeling, aggregating mitochondrial gene features improved discrimination, with the best-performing random forest model achieving an AUC of 0.91 under internal cross-validation. These results require validation in independent external cohorts. CONCLUSIONS: This study provides a genetically supported, multi-omics framework linking compartment-specific mitochondrial dysregulation to sepsis heterogeneity and nominates candidate biomarkers for prioritization. The reported model performance reflects internal resampling and requires validation in independent clinical cohorts and future multi-omics profiling (including metabolomics) before translational implementation.