Composite pathway-guided Cox modeling for interpretable survival prediction in breast cancer.

Accurate survival prediction in breast cancer is essential for patient risk stratification and personalized treatment planning. Although transcriptomic data offer valuable insights into tumor biology, existing predictive models often suffer from poor interpretability and limited integration of biological knowledge. Standard gene-level feature selection methods, such as variance filtering and regularized regression, are agnostic to functional relationships between genes and may overlook biologically meaningful patterns. In this study, we propose a pathway-guided Cox modeling framework that integrates curated biological knowledge to enhance the interpretability and performance of survival prediction. Using TCGA-BRCA transcriptomic and clinical data, we develop and compare three models: a baseline Cox model using high-variance genes, per-pathway Cox models built from curated KEGG and Reactome gene sets, and a composite model that aggregates top-performing pathways. Our results demonstrate that pathway-informed models consistently outperform the baseline in terms of cross-validated concordance index and survival stratification. Furthermore, the composite model highlights biologically plausible gene signatures and pathways associated with breast cancer prognosis, offering interpretable and clinically relevant insights. This work provides a modular, reproducible, and interpretable framework for survival modeling in high-dimensional genomics. By combining classical statistical models with biological structure, it offers a transparent approach for biomarker discovery and precision oncology.