Machine learning identifies eRNA signatures for breast cancer prognosis and early detection
The study reports that machine learning applied to enhancer RNA expression in 1,073 TCGA breast cancer samples and independent public cohorts identifies distinct noncoding signatures for prognosis and early detection. A 10-eRNA model stratified patients into high- and low-risk groups with a 3-year AUC of 0.743 and an independent hazard ratio of 3.18, while a 19-eRNA panel paired with a multilayer perceptron classifier achieved an AUC of 0.959 for early-stage detection. These signatures converge on ESR1, FOXA1, GATA3, and AR regulatory networks and are accessible via a public scoring tool. While eRNAs offer a complementary noncoding layer to conventional mRNA diagnostics, prospective clinical validation and standardized assay development remain necessary before laboratory implementation.
The original study
Integrative multi-omics profiling and machine learning reveal enhancer RNA signatures for early detection and prognostic stratification in breast cancer.
- Authors
- Xia J, Tian J, Cao Y, Zhang S, Qian W, Qi Y, et al.
- Journal
- NAR cancer
- Type
- Journal Article
- PMID
- 42798840
Original abstract
Current breast cancer biomarkers rely predominantly on protein-coding transcriptomes, leaving the regulatory information encoded by enhancer RNAs (eRNAs) largely unexploited. Here, we profiled eRNA expression across 1073 The Cancer Genome Atlas Breast Cancer (TCGA-BRCA) samples integrated with The Cancer eRNA Atlas annotations, using GTEx healthy breast tissue (n = 179) as a curated normal reference. For prognostic stratification, least absolute shrinkage and selection operator (LASSO)-Cox regression identified a 10-eRNA signature that stratified patients into high- and low-risk groups in both training (P <.0001) and independent testing cohorts (P = 0.0022), with a 3-year time-dependent area under the curve (AUC) of 0.743. Multivariate Cox analysis confirmed the signature as an independent predictor [hazard ratio = 3.176, 95% confidence interval (CI): 2.075-4.864, P <.001]. Multi-omics network integration linked these prognostic eRNAs to regulatory hubs centered on the ESR1/FOXA1/GATA3/AR axis. For early-stage detection, we developed a 19-eRNA panel via bootstrap LASSO feature selection and benchmarked ten classifiers; a Multilayer Perceptron achieved the best generalization (AUC = 0.959, 95% CI: 0.925-0.986) on an independent cohort (GSE225846). Both modules were deployed in eRNACare, a publicly accessible web application for individualized survival and tumor probability scoring. These findings establish eRNAs as a complementary noncoding layer to conventional messenger RNA classifiers, with clinical potential for breast cancer risk stratification, therapeutic guidance, and early detection.