Retrospective multi-cohort validation of a real-world transcriptomics-guided machine learning model for treatment response prediction in breast cancer

Selection of systemic therapy for breast cancer remains largely empirical, particularly for chemotherapy, due to the lack of robust biomarkers that predict treatment response at the individual patient level. We developed Oncology CoPilot, a real-world, transcriptomics-guided machine learning (ML) decision-support model designed to integrate heterogeneous tumor gene expression data and treatment response annotations to support treatment response stratification across therapeutic classes. Oncology CoPilot was trained on a pan-cancer cohort comprising 11,414 patients across 15 cancer types and 150 systemic drug regimens derived from publicly available and published datasets. Retrospective external validation was performed using five independent breast cancer cohorts comprising 503 patients, spanning multiple molecular subtypes, transcriptomic platforms, and six commonly used treatment settings, including chemotherapy, endocrine therapy, and targeted therapy. Across the external validation cohort, the model demonstrated an overall accuracy of 72.8%, with balanced sensitivity (71.5%) and specificity (73.4%), and an ROC-AUC of 0.783. Regimen-specific analyses demonstrated stable performance for chemotherapy-based regimens (accuracy 70.1%-79.1%), highlighting the potential of transcriptomics-guided modeling to inform treatment response stratification in clinical settings where therapy selection is often empirical. For endocrine therapy, the model achieved 95.0% accuracy for tamoxifen, suggesting that transcriptomic features may capture biologically relevant estrogen-responsive and resistance-associated programs beyond receptor status alone, although this result is exploratory and based on a small sample size. In contrast, HER2-targeted therapies showed lower and more variable predictive performance, with accuracies of 66.0% for trastuzumab monotherapy and 61.9% for anthracycline-taxane chemotherapy combined with trastuzumab, likely reflecting smaller cohort sizes and the biological heterogeneity characteristic of HER2-positive disease. Overall, these findings demonstrate the feasibility of leveraging real-world transcriptomic data and machine learning to achieve generalizable treatment response stratification across diverse cohorts, platforms, and therapeutic classes, supporting the potential role of transcriptomics-guided models as complementary decision-support tools in oncology.