Development of an AI-based transcriptional surrogate for epigenetic aging in breast cancer.

Abstract

e12575 Background: In our previous omics studies of breast cancer in Mexican patients, we observed an unexpected enrichment of mutational signature 1, a clock-like process positively correlated with age in both normal and malignant tissues, particularly in hormone receptor–positive tumors. This finding was independently replicated in a cohort of Mexican women residing in the United States. Given that breast cancer in Mexico is typically diagnosed at least ten years earlier than in other populations, we hypothesized that hormone receptor–positive tumors in Mexican women may exhibit features of accelerated biological aging relative to patients’ chronological age. Methods: To address this, we developed a machine learning–based transcriptional surrogate of Horvath’s epigenetic clock using gradient boosting (XGBoost), using large-scale public transcriptomic and DNA methylation datasets (N = 1,514 samples). The model was trained to predict DNA methylation age directly from gene expression profiles and demonstrated high predictive performance. Feature selection was based on both algorithmic importance and biological relevance, allowing prioritization of genes linked to aging biology and facilitating functional interpretation of molecular changes across the replicative and tumor evolutionary timeline. To further characterize biological processes associated with methylation age, we analyzed DNA methylation profiles from public breast cancer cohorts (N = 3,901). Samples were stratified based on chronological age and calculated epigenetic age using multiple clocks, including Horvath, PhenoAge, and GrimAge. Differential methylation and pathway enrichment analyses were performed. Results: This approach enabled the identification of transcriptional patterns consistent with age acceleration in Mexican patients, as well as higher estimated transcriptomic age. Integrative analyses identified methylation-associated pathways and molecular features that differed according to aging metrics, revealing biological processes associated with distinct measures of chronological and epigenetic aging in breast tumors. Conclusions: These findings support the value of combining artificial intelligence–based transcriptional surrogates of epigenetic clocks with large-scale methylation profiling and biological prioritization to provide a novel framework for studying accelerated tumor aging in breast cancer, with potential relevance for biomarker development and clinical risk stratification.