Abstract
BACKGROUND: Breast cancer (BC) remains a significant threat to human health, with substantial variations in prognosis and treatment responses. Intra-tumor heterogeneity (ITH) presents a critical challenge in developing reliable prognostic models. METHODS: This study integrated multi-region RNA sequencing data from BC patients with the TCGA BC dataset. Genes resistant to sampling bias were identified by evaluating inter-patient heterogeneity (IPH) and ITH. A machine learning framework incorporating ten algorithms was used to construct a prognostic signature.The expression levels and oncogenic function of the prognostic genes were validated through RT-qPCR and in vitro experiments. RESULTS: The signature, comprising CFL2 and SPNS2, demonstrated stable predictive performance in both training and validation cohorts (C-index > 0.6). High-risk patients exhibited enriched immune infiltration, particularly CD8+ T cells, and higher expression of immune checkpoint molecules, suggesting sensitivity to immunotherapy. A nomogram integrating risk score with clinical variables further improved prognostic accuracy. The dysregulation of signature genes was confirmed in BC cell lines. CONCLUSION: By minimizing ITH interference, this study developed a robust prognostic signature for BC, offering insights into the tumor immune microenvironment and potential therapeutic strategies.