B.R.E.A.S.T.: AN INTERPRETABLE MACHINE LEARNING FRAMEWORK FOR PROTEOMICS-DRIVEN THERAPY OUTCOME PREDICTION IN BREAST CANCER AND TRIPLE-NEGATIVE BREAST CANCER

Breast cancer (BC) is a highly heterogeneous disease characterized by marked molecular diversity and variable therapeutic response. Among its subtypes, Triple-Negative Breast Cancer (TNBC) is one of the most aggressive and clinically challenging, lacking targeted therapies and associated with poor prognosis. Predicting treatment outcome remains a major unmet need in precision oncology [1]. In this study, we developed B.R.E.A.S.T. (Breast cancer Enhanced AI Supported Therapy), an interpretable Artificial Intelligence (AI) framework designed to predict therapeutic response by integrating tumor proteomic profiles with clinical and treatment-related data. Proteomic data from The Cancer Genome Atlas (TCGA) [2] and The Cancer Proteome Atlas (TCPA) [3] were integrated with clinical outcomes and pharmacological annotations from the cBIOPortal [4]. Two parallel analytical settings were investigated: the full BC population and the TNBC subgroup. Thirteen supervised machine learning algorithms were benchmarked using stratified 5-fold cross-validation. Model performance was assessed using accuracy, precision, recall, F1-score, and ROC-AUC. Feature importance was evaluated through a two-step consensus ranking strategy combined with SHAP (SHapley Additive exPlanations). Differential expression between responders and non-responders was assessed using Mann–Whitney U test with FDR correction and effect size estimation (Cliff’s delta) and visualized through volcano plots. Across both TCGA and TCPA BC cohorts, ensemble-based models consistently achieved the highest predictive performance, with Random Forest and ExtraTrees reaching near-perfect discrimination (ROC-AUC up to 0.998). These results were reproduced in TNBC subset, with ROC-AUC values up to 0.999 despite reduced sample size and increased biological heterogeneity, suggesting that performance is driven by biologically meaningful proteomic patterns. Learning curve analysis further mitigated overfitting risk. Feature selection identified key biomarkers mapped to key oncogenic pathways, including PI3K/AKT/mTOR signaling (e.g., AKT_pS473, RAPTOR), DNA damage response (e.g., CHK1, MRE11, PCNA), apoptosis (e.g., BCLXL), metabolism (e.g., FASN), and invasion (e.g., SNAIL, MMP14) [5]. Volcano plots revealed significant differences between responders and non-responders, with proteins such as AKT_pS473 and CHK1 enriched in non-responders [6], and RAPTOR and MRE11 associated with responder phenotypes, highlighting mechanisms related to survival signaling, DNA repair, and pathway-specific regulation of therapeutic sensitivity. Importantly, retraining models using only the top 30 features preserved high predictive performance across both cohorts, demonstrating that a parsimonious proteomic signature captures key determinants of therapeutic response. B.R.E.A.S.T. provides a robust and interpretable framework for therapy outcome prediction in TNBC, supporting the identification of clinically relevant biomarkers and advancing precision oncology strategies, particularly in aggressive subtypes such as TNBC.