Decoding Breast Cancer Heterogeneity via Multi-Omics Integration and Language Model-Based Interpretation

We present a novel pipeline combining Multi-Omics Factor Analysis (MOFA) and fine-tuned Large Language Models (LLMs) to predict breast cancer subtypes using proteomics, DNA methylation, and RNA-Seq data. Breast cancer is a heterogeneous disease characterized by diverse molecular alterations across multiple biological layers, necessitating integrative approaches for accurate subtype classification. Our methodology leverages MOFA for dimensionality reduction to identify key latent factors driving het-erogeneity, followed by LLM fine-tuning on these multi-omics signatures to enhance prediction accuracy. MOFA analysis identified five key latent factors capturing distinct biological processes: immune response, cell cycle regulation, metabolic reprogramming, tumor microenvironment interactions, and DNA repair mechanisms. We extracted the top features per omics layer for each factor and performed Gene Set Enrichment Analysis (GSEA) to characterize their biological significance. Our LLM, trained on curated multi-omics signatures and clinical metadata encoded as structured text prompts, significantly outperformed conventional statistical models in subtype classification, achieving AUC=0.93 and accuracy=0.89, compared to Random Forest (AUC=0.87, accuracy=0.82) and SVM (AUC=0.85, accuracy=0.80). The superior performance of our approach is attributed to the LLMs ability to capture complex, non-linear relationships and hierarchical feature interactions across omics layers. This integrative pipeline provides both improved predictive performance and interpretable biological insights, offering potential for enhanced clinical decision-making in breast cancer management.