Multi-Omics ML Models Predict Breast Cancer ER Status

Predicting Estrogen Receptor (ER) status is crucial for breast cancer diagnosis, prognosis, and treatment. This study conducted a rigorous benchmarking analysis of various classical machine learning models to predict ER status using multi-omics data from the TCGA-BRCA cohort. The datasets included transcriptomic (RNA expression), genomic (copy number variation), and proteomic (RPPA) information. The research employed a robust experimental framework, incorporating stratified train-test splitting, cross-validation, class imbalance handling, and fold-specific feature selection to ensure reliable evaluation and prevent data leakage. Models such as Random Forest, XGBoost, LightGBM, SVM, and Logistic Regression were tested in both single-omic and multi-omic configurations. Findings indicated that RNA expression data provided the most potent predictive signal. While multi-omic integration yielded only modest improvements, these gains were consistent across models. Random Forest emerged as the top performer in the integrated multi-omic setting, achieving a balanced accuracy of 90.3% and an ROC-AUC of 97.1%. The recurrent selection of known biologically relevant genes further validated the models' findings.