Machine learning-driven integration of multi-omics data uncovers methylation-dependent molecular signatures for luminal A and B breast cancer subtypes.

BACKGROUND: Breast cancer is the most common malignancy in women and includes molecular subtypes with distinct clinical outcomes, such as luminal A and luminal B. Although luminal tumors account for most cases, the epigenetic mechanisms differentiating luminal A from the more aggressive luminal B subtype remain unclear. We developed an interpretable machine learning pipeline to reduce genomic noise and identify methylation-dependent regulatory signatures by distinguishing these subtypes.

METHODS: The INTEND algorithm was trained on 4,441 paired RNA-seq and DNA methylation samples from 14 cancer types, excluding breast cancer, to prevent data leakage. This training generated an epigenetic filter identifying genes whose expression is predictable from methylation. The filter was applied to 537 TCGA-BRCA luminal samples to predict expression and extract CpG-level regulatory signals.

RESULTS: The model identified 2,670 high-confidence genes (R2 ≥ 0.6) and achieved strong predictive accuracy (average R2 = 0.76). Subtype-specific genes were identified through differential expression of the filtered dataset and removal of shared genes, yielding 30 methylation-driven biomarkers unified by strong methylation-expression coupling and distinct subtype-specific expression patterns.

CONCLUSION: This framework identifies potential methylation-controlled genes underlying luminal breast cancer differentiation, reducing transcriptional noise while revealing CpG regulatory mechanisms.