From machine learning to causal insight: a robust 12-gene signature to distinguish sepsis risk from systemic inflammatory response syndrome.

BACKGROUND: Sepsis, a life-threatening condition driven by a dysregulated host response, poses significant challenges for early diagnosis and early intervention. This study aimed to identify robust biomarkers capable of distinguishing sepsis from non-infectious systemic inflammation (SIRS). METHODS: We analyzed public Gene Expression Omnibus (GEO) data sets using a multi-modal workflow. This included differential expression analysis, weighted gene co-expression network analysis (WGCNA), and an evaluation of 113 machine learning (ML) models to build a predictive signature. We further investigated hub genes using immune infiltration, single-cell RNA sequencing (scRNA-seq), and two-sample Mendelian randomization (MR) to infer causality. Key findings were validated experimentally (in vitro and in vivo). RESULTS: The optimal LASSO + RF model identified a 12-gene signature strongly associated with the septic state, achieving a high area under the curve (AUC) of 0.998 and robust external validation (AUCs > 0.93). This high discriminatory power highlights its potential for stratifying patients. Notably, MR analysis provided causal evidence linking elevated EIF4G3 expression to increased sepsis risk. ScRNA-seq located the expression of hub genes, including DNAJC5, to myeloid cells. Experimental validation confirmed the upregulation of EIF4G3 and DNAJC5 in sepsis models. CONCLUSIONS: This study identifies and validates a highly accurate 12-gene signature for distinguishing sepsis from SIRS. The causal evidence for EIF4G3 and robust validation of DNAJC5 underscore their potential as biomarkers for early diagnosis and differentiation from SIRS. This signature holds significant promise for the early identification of high-risk patients, enabling timely intervention and improving outcomes.