Feature-Selection Stability and Functional Characterization of Sepsis-Related Transcriptomic Candidates Across Heterogeneous Cohorts

Sepsis-related transcriptomic signatures may differ across tissues and clinical settings, complicating their reproducibility and biological interpretation. We analyzed three public human microarray datasets spanning blood and skeletal muscle (GSE12624, GSE13205, and GSE69063; 26, 21, and 30 participants, respectively) using regularized log-odds least-squares feature screening and repeated participant-level cross-validation, with feature selection restricted to training folds. Stable candidates were defined by selection in at least 80% of training partitions and characterized using exploratory within-cohort expression comparisons and separate functional annotations. Mean internally cross-validated areas under the receiver operating characteristic curve (AUCs) were 0.735 ± 0.080, 1.000 ± 0.000, and 0.974 ± 0.020, respectively. Eighteen genes met the stability criterion, six per dataset, but none did so in more than one dataset. In GSE13205, AQP4, CA14, CHI3L1, LGI1, MYH7B, and SLPI showed complete apparent univariate separation (AUC = 1.000; Benjamini–Hochberg-adjusted p = 2.78 × 10−5), although these comparisons were not independent of candidate selection. Functional characterization retrieved 211 distinct Gene Ontology terms, 41 gene–pathway associations, and five database-derived functional associations. These findings support context-dependent candidate characterization, rather than independent biomarker validation, and motivate evaluation in distinct clinical populations.