BPS2026 – Leveraging transcriptomic signatures and machine learning to predict high risk progression in HPV-related cervical lesions

Progression risk in HPV-positive patients is difficult to judge from cytology alone, and a noninvasive molecular adjunct could strengthen triage. Published models often lose accuracy when scored on truly independent cohorts because label construction, platform differences, and probe-content shifts distort distributions. This study evaluates whether gene-expression profiles from exfoliative cervical cells can support transportable risk stratification once distributions are explicitly aligned. A labeled discovery matrix was curated from public GEO series and mapped to HGNC gene symbols. Features were standardized and duplicates were collapsed. Supervised classifiers were trained with stratified validation, and model explanations were derived from SHAP to connect predictions to biology. Before external scoring, five alignment strategies were compared: intersection of shared genes, Z -score transfer, within-sample rank transform, gene-wise quantile matching, and empirical-Bayes harmonization with ComBat. External discrimination was tested on two independent cohorts that differed by platform and probe content (Affymetrix GSE75132 and Illumina GSE56303). Performance was summarized by ROC-AUC, operating points were chosen on the discovery set, and calibration was assessed. Tree-based ensembles produced compact, interpretable panels whose attributions consistently emphasized immune and cell-cycle genes. On GSE75132, ComBat achieved the highest discrimination (ROC-AUC ≈ 0.69) with improved calibration, while rank and quantile methods were stable but mid-range. On GSE56303, results plateaued at AUC ≈ 0.54–0.55 across tags, consistent with limits imposed by label construction and probe-content differences. These findings indicate that exfoliative-cell transcriptomes contain transferable signal for risk stratification, but careful harmonization is required for reliable external discrimination and probability calibration. The workflow is reproducible end-to-end and supports the feasibility of a transcriptomic adjunct to cytology, with prospective, platform-matched validation as the next step toward clinical translation.