ABSTRACT Background Breast cancer is characterized by substantial molecular heterogeneity, and clinical features alone are insufficient for personalized management. This study aimed to develop a machine learning–based prognostic signature for breast cancer. Methods Ten independent breast cancer cohorts and 76 combinations of machine learning algorithms were used to identify the consensus machine learning‐derived prognosis signature (CMDPS). We then collected 62 published transcriptome signatures for comparison with CMDPS. The associations between CMDPS and the immune cell profile, multi‐omics alterations, and pharmacological landscape were further investigated. Results Nineteen genes consistently linked to survival across the 10 cohorts were identified through univariate analysis, with 76 algorithm combinations employed to determine the most reliable model. CMDPS was best at forecasting overall survival. In the cancer genome atlas breast invasive carcinoma (TCGA‐BRCA) cohort, CMDPS displayed a C‐index value of 0.696 (3‐year survival area under the curve: 0.769; hazard ratio: 5.065 3.233–7.936). Patients with high CMDPS scores had a dismal prognosis. Compared with clinical features and 62 published signatures, CMDPS displayed stronger robustness. Furthermore, an in‐depth examination of the CTRP and PRISM drug collections revealed that individuals with elevated CMDPS levels exhibited increased responsiveness to various widely used chemotherapy medications. In parallel, patients with a low CMDPS score exhibited more abundant immune cell infiltration. Conclusion CMDPS is a promising tool that has profound implications for optimizing the clinical management and personalized treatment of breast cancer.
Cheng et al. (Wed,) studied this question.