Deriving consensus sepsis clusters via goal-directed subgroup identification in multi-omics study.

Sepsis, a syndrome of life-threatening organ dysfunction caused by dysregulated host responses to infection, exhibits profound pathobiological heterogeneity, hindering the development of effective therapies. Current subtyping approaches, often reliant on single-omics data or unsupervised clustering, yield poorly reproducible and therapeutically misaligned classifications. Here, we introduce a goal-directed subgroup identification (GD-SI) framework that optimizes patient stratification for differential treatment responses, integrating longitudinal multi-omics data (transcriptomic, proteomic, metabolomic, phenomic) from 1327 subjects across 43 hospitals. While supervised multi-omics integration frameworks (e.g., DIABLO) effectively capture shared biological signals, our approach anchors subgroup discovery directly to treatment-effect optimization. This strategy achieves substantial cross-omic concordance and, crucially, generalizes to predict differential treatment response across international critical care databases. Patients stratified by GD-SI-derived benefit scores for restrictive versus liberal fluid resuscitation exhibited marked survival differences, with similar advantages observed for ulinastatin immunomodulation. External validations in MIMIC-IV and ZiGongDB confirm prognostic generalizability. This framework reconciles biological heterogeneity with clinical actionability, offering a scalable infrastructure for precision trial design and personalized sepsis management. Our findings underscore the translational potential of omics-driven, goal-directed stratification to overcome decades of therapeutic stagnation in critical care.