COAT: A Cross-Omics Attention Transformer for PAM50 Breast Cancer Subtype Classification

Breast cancer is the most frequently diagnosed cancer in women worldwide, with approximately 2.3 million new cases annually. Accurate molecular subtyping is essential for guiding treatment decisions; however, existing PAM50 classifiers rely solely on mRNA expression and remain susceptible to normalization artifacts and platform-specific biases. To overcome these limitations, we propose COAT (Cross-Omics Attention Transformer), a novel deep learning framework that integrates mRNA, miRNA, and DNA methylation data to robustly classify PAM50 breast cancer subtypes. The model projects each omics modality into a shared latent space using modality-specific multilayer perceptrons and leverages a directed inter-omics attention mechanism to capture complementary interactions across modalities. The merged representations are processed by a classification head trained with class-weighted cross-entropy to correct for class imbalance. The model was evaluated on the TCGA-BRCA dataset (824 PAM50-tagged samples) using 5-fold stratified cross-validation, achieving an accuracy of 0.822 ± 0.020, a macro F1 score of 0.817 ± 0.033, and a macro area under the ROC curve (AUC) of 0.954 ± 0.011. These results demonstrate high performance compared to mono-omics approaches and traditional machine learning methods, while remaining competitive with recent multi-omics models. An additional 10-fold cross-validation experiment with Bayesian hyperparameter optimization further improved performance (accuracy = 0.852 ± 0.030, macro F1 score = 0.836 ± 0.041, macro AUC-ROC = 0.957 ± 0.013), indicating stable performance across different validation conditions. GradientSHAP interpretability analysis revealed that COAT identified biologically relevant biomarkers, including ERBB2 and GRB7 for the HER2-enriched subtype, ESR1 and PGR for the Luminal A subtype, and KRT5 and FOXM1 for the Basal subtype. Overall, COAT demonstrates that directed inter-omics cross-attention effectively integrates complementary multi-omics signals, achieving strong predictive performance while preserving biological interpretability and providing a generalizable framework for multi-omics-based cancer classification.