Integrative Survival Prediction in Breast Cancer Using Extracellular Matrix Protease Transcript Signatures and...

Background/Objectives: Traditional breast cancer prognostic tools relying on clinical staging often miss molecular heterogeneity, leading to divergent patient outcomes. Extracellular matrix (ECM) remodeling, driven by the Matrix Metalloproteinase (MMP), ADAM, and ADAMTS enzyme families, is critical to tumor progression. This study evaluates whether integrating ECM protease transcript abundance with standard clinical variables improves survival prediction accuracy and personalized risk stratification. Methods: Clinical and transcriptomic data from The Cancer Genome Atlas (TCGA) breast cancer cohort were analyzed. We integrated the protein-coding transcripts per million (pTPM) of top-ranked protease genes with standard clinical covariates (age, ordinal stage). Cox Proportional Hazards (CoxPH), penalized Cox (CoxNet), Random Survival Forest (RSF), and Gradient Boosting Survival (GBS) models were evaluated under a stratified 70/30 train-test split, followed by five-fold cross-validation. The locked final RSF model was then externally tested in METABRIC without retraining or risk-cutoff optimization. Results: Univariate screening identified ADAM15, MMP15, and ADAMTSL1 as global risk factors, whereas ADAMTS8 and MMP7 were protective. Prognostic signals were subtype-dependent. Integrated multivariable models outperformed transcript-only approaches in internal testing. The integrative RSF achieved the highest held-out discrimination (C-index = 0.797), outperforming a clinical-only Cox baseline trained on age and stage alone (C-index = 0.742, 95% CI 0.636-0.826). In METABRIC, the external C-index was 0.581 (95% CI 0.562-0.598), with significant survival separation across training-defined risk groups (log-rank p < 0.0001). Conclusions: ECM protease transcript profiles provide complementary prognostic information in TCGA-BRCA and show partial transportability to METABRIC. However, the modest external C-index indicates limited individual-level discrimination across platforms, so these candidate markers should be interpreted as hypothesis-generating and require further validation before clinical implementation.