HCAM-BRCA: an interpretable multi-omics framework for breast cancer subtype classification and biomarker discovery

Integrating multi-omics data can improve breast cancer subtype classification by capturing complementary molecular information from mRNA, miRNA, and DNA methylation profiles. However, many existing approaches rely on single- or pairwise-omics integration and do not explicitly model hierarchical regulatory interactions across molecular layers, limiting biological interpretability and predictive performance. In this study, we propose HCAM-BRCA, a Hierarchical Cross-Attention Multi-omics framework for breast cancer subtype classification and biomarker discovery. The proposed architecture integrates mRNA, miRNA, and DNA methylation data using modality-specific self-attention and cross-attention mechanisms to capture both intra-omics feature dependencies and inter-omics regulatory relationships. The learned low-dimensional embeddings were subsequently evaluated using multiple machine learning classifiers. Across all classifiers, HCAM-BRCA consistently demonstrated improved performance pairwise RNA-miRNA and RNA-CpG integration strategies. Among the evaluated models, CatBoost achieved the highest macro-average MCC (0.7628 ± 0.0130), while Extra Trees attained the highest AUPRC (0.8925 ± 0.0032). The proposed framework demonstrated strong and stable subtype discrimination, particularly for the clinically challenging Luminal B and Her2-enriched subtypes. Performance metrics were reported as mean ± standard deviation across five-fold cross-validation. Attention-based feature prioritization identified biologically relevant genes associated with transcriptional regulation, chromatin organization, immune signaling, and cancer-related pathways. Survival analysis further revealed several subtype-specific prognostic associations, including CUX1, MYZAP, and CEBPA in Basal-like tumors, HOXA10 and HLA-DQA2 in Luminal B, and GDF10 and ZNF879 in Her2-enriched tumors. Overall, HCAM-BRCA provides an interpretable and biologically informed multi-omics integration framework that combines robust subtype classification with clinically relevant biomarker prioritization, demonstrating its potential utility for computational oncology and precision breast cancer research.