A tri-omics and machine learning framework identifies prognostic biomarkers and metabolic signatures in sepsis

Sepsis is a complex systemic inflammatory syndrome that currently lacks stable and specific biomarkers. Multi-omics integration combined with machine learning and single-cell analysis offers new approaches for elucidating molecular mechanisms and nominating candidate regulators for further validation. Establish an integrated analytical framework combining transcriptomics, proteomics, metabolomics, and single-cell transcriptomics to identify sepsis-associated candidates and generate testable hypotheses for downstream mechanistic and translational studies. A multi-omics integrative framework was established by combining transcriptomic, proteomic, and metabolomic data. Weighted gene co-expression network analysis (WGCNA) was first applied to identify modules significantly associated with sepsis, and key candidate genes were further refined using LASSO regression and SVM-RFE algorithms. Compare differentially expressed transcripts and proteins through quadrant analysis. Further leverage open-access proteome-phenome resources to assess the association between priority target proteins and sepsis-related phenotypes/disease risk. Metabolomic profiling was conducted to identify metabolites significantly correlated with the core genes and their enriched pathways. A joint gene-metabolite classification model was then constructed using logistic regression, support vector machines, and random forests, with predictive performance evaluated using ROC curve analysis. LASSO regression was used to visualize coefficient shrinkage and effect direction, while random forest-based feature importance quantified the relative contribution of each variable. Finally, leveraging the ITCM database, an in silico screening was performed to identify natural compounds with potential regulatory effects on TPR and ERN1 expression. TPR and ERN1 were prioritized as key candidate biomarkers associated with sepsis. External validation in GEO cohorts and analyses of protein-disease risk association databases further supported their reproducible disease association in external datasets, motivating further validation of their prognostic/diagnostic potential. Non-targeted metabolomics screening identified 136 differentially expressed metabolites associated with them, primarily enriched in glycerophospholipid metabolism and fatty acid-related pathways. Metabolite-gene correlation integration linked these metabolic alterations to TPR/ERN1, suggesting their potential involvement in immune-metabolic pathway remodeling associated with sepsis heterogeneity. Single-cell transcriptomics provided cellular origin-level support, showing enrichment of TPR and ERN1 expression in immune subpopulations, including monocyte-macrophages and NK cells. A joint gene-metabolite model constructed using a fixed feature set (2 genes + 5 metabolites) demonstrated strong discriminatory potential in an internally derived cohort, though external validation in independent, multi-omics matched cohorts remains urgently needed. Finally, exploratory computational screening based on the ITCM database identified several potential active monomers in traditional Chinese medicine that may influence TPR or ERN1 expression, providing hypothesis-generating clues for subsequent mechanistic studies and drug translation. By integrating multi-omics data, TPR and ERN1 were identified as sepsis-associated candidate biomarkers and linked to immune-metabolic alterations associated with immune cells. The combined gene-metabolite signature panel and exploratory compound screening provide a hypothesis-generating framework for downstream experimental validation and clinical study design.