Integrating DNA methylation biomarkers for breast cancer risk prediction using artificial intelligence

Abstract

Abstract Breast cancer remains a major cause of morbidity and mortality among women worldwide, highlighting the necessity for predictive tools to identify at-risk individuals prior to the manifestation of symptoms. Standard imaging, like mammography, has limited sensitivity in dense breasts and relies on visible morphological changes. In contrast, circulating DNA methylation biomarkers provide a minimally invasive, stable alternative that detects systemic epigenetic changes before tumors develop. This study analyzes DNA methylation profiles from the GSE51032 EPIC-Italy cohort (Illumina HumanMethylation450K platform), involving 845 participants, including only 658 cancer-free and breast cancer samples. After quality control (QC), 224 pre-diagnostic breast cancer cases and 418 controls are included, totaling 642 samples across 483,848 CpG sites. Differential methylation analysis identifies 4621 CpG sites with significant methylation differences (FDR < 0.1, |Δβ| >= 0.03). These features are subsequently utilized to train and evaluate a variety of machine learning (ML) and deep learning (DL) classifiers. Among these, the random forest demonstrated the highest overall performance, attaining an area under the curve (AUC) of 0.849 and an accuracy of 0.798. These findings highlight the potential of integrating high-dimensional epigenomic biomarkers with artificial intelligence (AI) to enable early prediction of breast cancer risk, thereby promoting minimally invasive screening and personalized prevention strategies.