Modality-aware external attention with explicit cross-omics interaction modelling for breast cancer subtype classification

Breast cancer is molecularly heterogeneous, and multi-omics measurements provide complementary information for molecular subtype classification. Their high dimensionality and modality-specific characteristics, however, make effective integration challenging. We developed a modality-aware external attention (MAEA) framework for classifying five breast cancer subtypes using mRNA expression, DNA methylation, and miRNA expression data. Boruta feature selection is performed separately for each modality within each training fold. Modality-specific encoders generate latent representations, while pairwise Hadamard interaction tokens explicitly represent cross-omics relationships. These structured tokens query shared external key and value memories and are combined through a learned token gate. In stratified 10-fold cross-validation on TCGA-BRCA, MAEA achieved 86.78 ± 2.46% accuracy, 83.47 ± 4.38% macro-recall, 81.36 ± 3.48% macro-F1, and 0.9694 ± 0.0115 macro-AUC. A matched modality-aware self-attention model achieved 85.65 ± 3.76% accuracy and 81.26 ± 3.61% macro-F1, with no statistically significant fold-wise difference. MAEA improved accuracy by 3.40 percentage points and macro-F1 by 3.54 percentage points compared with a concatenation-based external-attention baseline. Ablation experiments further supported the contribution of feature selection, interaction modelling, and external-memory attention. MAEA provides a structured multi-omics fusion strategy that preserves modality identity while explicitly modelling pairwise cross-omics interactions. Independent external validation remains necessary before clinical application.