Аннотация
Breast cancer is a heterogeneous disease with diverse molecular subtypes that exhibit distinct clinical outcomes and therapeutic responses. Accurate subtype classification is crucial for diagnosis, treatment stratification, and prognosis. With advances in sequencing technologies, multi-omics data integration has emerged as a promising approach for characterizing the biological complexity of tumor heterogeneity. This study evaluates feature selection strategies, data fusion architectures, and classification algorithms for breast cancer subtype identification using the METABRIC dataset. Three omics modalities, mRNA expression (EXP), copy-number alterations (CNA), and somatic mutations (MUT), were systematically assessed across single-, two-, and three-modality configurations. Feature selection was performed with ElasticNet-regularized cross-validated logistic regression, XGBoost importance ranking, and Binary Particle Swarm Optimization (BPSO) with CatBoost as the fitness evaluator. Fusion was analyzed through direct concatenation and cross-modal Transformer architectures that enrich expression embeddings with contextual information from CNA and MUT modalities. The evaluated classifiers were support vector machine (SVM), multilayer perceptron (MLP), Tabular Prior-Data Fitted Network (TabPFN), and convolutional neural network (CNN). The proposed pipeline achieved performance comparable to state-of-the-art multi-omics models while maintaining interpretability and reproducibility across different fusion and feature selection paradigms. The results underscore the complementary value of multi-omics integration and the importance of pipeline optimization for robust breast cancer subtype classification.