Identification of diagnostic biomarkers for neonatal necrotizing enterocolitis via machine learning screening and single-cell virtual gene knockout validation.

Neonatal necrotizing enterocolitis (NEC) continues to be the most severe gastrointestinal emergency affecting preterm infants, with reported mortality rates ranging from 20% to 30%. The absence of distinct early biomarkers results in delayed intervention and unfavorable outcomes, and the molecular mechanisms underlying the dysregulated immune response in NEC remain incompletely defined. We integrated five gene expression omnibus datasets (36 NEC cases and 34 controls) using weighted gene co-expression network analysis and systematically evaluated 113 machine learning algorithm combinations. The best-performing model (RF + XGBoost) identified candidate diagnostic genes, which were validated through independent bulk RNA-sequencing cohorts, single-cell RNA sequencing (11,308 intestinal cells), single-cell virtual gene knockout (scTenifoldKnk), and experimental models including mice and the intestinal epithelial cells cell line. Combinatorial in silico perturbation was further performed using Geneformer, and a ferroptosis gene panel was validated by qRT-polymerase chain reaction in the mouse NEC model. The RF + XGBoost model achieved high diagnostic accuracy (area under the curve = 0.979, 95% CI: 0.940-1.000). Three key biomarkers were identified: SLC26A3, CCL20, and CXCL5. Multi-platform validation showed consistent downregulation of SLC26A3 and upregulation of CCL20 and CXCL5 in NEC (all