Machine Learning-Guided Multi-Cohort Transcriptomic Profiling Identifies SPON1 and ALDH1A2 as Diagnostic and...

Reliable biomarkers for high-grade serous carcinoma (HGSC) with prognostic and microenvironmental relevance remain limited. Here, we developed a machine learning-guided cross-cohort transcriptomic framework to identify stable biomarkers in HGSC. Three GEO cohorts comprising 68 samples (34 HGSC and 34 normal) and 21,355 genes were integrated, and five classifiers were benchmarked under strict Leave-One-Dataset-Out (LODO) validation. Differential expression and random-effects meta-analysis were used to support cross-cohort feature prioritization, and external validation was performed in TCGA-OV tumors (n = 427) versus GTEx normal ovaries (n = 88). This framework identified a robust 22-gene consensus panel with strong cross-cohort discrimination between HGSC and normal tissue. Among these, ALDH1A2 and SPON1 emerged as the only two genes consistently prioritized by all five models. Prognostic analysis showed opposite clinical associations, with higher ALDH1A2 linked to poorer progression-free and overall survival and higher SPON1 linked to better outcomes. Immune-module analysis further demonstrated that predicted HGSC probability was positively associated with T cell, cytotoxic/NK, Treg, checkpoint, and inflammatory programs, indicating an immune-active yet immunoregulatory microenvironment. Together, these findings define a reproducible 22-gene HGSC signature and highlight ALDH1A2 and SPON1 as robust diagnostic and prognostic biomarkers.