Cross-Modal Self-Attention Fusion for Breast Cancer Subtype Classification Using Multi-Omics Data

Background: Accurate classification of breast cancer subtypes is essential for personalized therapy and prognosis. Traditional subtype classification basically relies on gene expression profiling, usually overlooking other genomic signals like copy-number alterations (CNA) and mutations. At the same time most of the multi-omics models often rely on early or late fusion strategies, which do not capture complex inter-modality interactions. Methods: This study proposes a cross-modal transformer-based approach that integrates gene expression, copy number alterations, and mutation data for robust breast cancer subtype classification. Each omics modality is encoded as a separate sequence and projected into a shared embedding space. Gene expression is treated as the primary modality and enriched through cross-modal self-attention mechanisms with CNA and mutation features. The final enriched embeddings are flattened and passed through a residual-connected MLP classifier. We evaluate performance on the METABRIC dataset using ElasticNet-selected top-K features (K = 300, 500, 1000, 1500) and mostly focus on macro F1-score, weighted F1-score, and ROC AUC due to class imbalance. Results: Integrating copy-number and mutation data with expression features improved subtype classification across most feature set sizes. The tri-omic model (EXP+CNA+MUT) achieved the best performance for smaller feature sets (K = 300–500), whereas for larger feature sets (K = 1000) the highest scores were obtained by the bi-omic model (EXP+CNA) with macro-F1 = 0.859, weighted F1 = 0.868, accuracy = 0.866 and ROC AUC = 0.969. Paired statistical tests across five folds showed that differences between modality configurations did not reach significance at any K (all p > 0.09), whereas feature-set size did. Within the EXP+CNA configuration alone, macro-F1 increased significantly from K = 300 to K = 500 (paired t-test, p = 0.012) and from K = 300 to K = 1000 (p = 0.036); and in the higher-powered pooled analysis across all three modality configurations (n = 15 paired folds), K = 1000 also outperformed K = 300 (p = 0.030). Conclusion: This pipeline demonstrates an application of cross-modal attention for omics integration in subtype classification task, offering a scalable and biologically grounded alternative to traditional fusion approaches.