Abstract
Background: Epithelial-mesenchymal transition (EMT), deemed a pivotal hallmark of tumours, is intricately regulated by DNA methylation and encompasses multiple states along tumour progression. The potential mechanisms that drive the intrinsic heterogeneity of breast cancer (BC) via EMT transformation have not been identified, presenting a significant obstacle in clinical diagnosis and treatment. Methods: A total of 7,602 patients have been included in this study. We leveraged integrated multiomics data (epigenomic, genomic, and transcriptomic data) to delineate the comprehensive landscape of EMT in BC. Subsequently, a subtyping classifier was developed through a machine learning framework proposed by us. Results: We classified the BC samples into three methylation-driven EMT subtypes with distinct features, namely, C1 (the mammary duct development subtype with TP53 activation), C2 (the immune infiltration subtype with high TP53 mutation), and C3 (the ERBB2 amplification subtype with an unfavorable prognosis). Specifically, patients with the C1 subtype might respond to endocrine therapy or the p53-MDM2 antagonist nutlin-3. Patients with the C2 subtype might benefit from combined therapeutic regimens involving radiotherapy, PARP inhibitors, and immune checkpoint blockade therapy. Patients with the C3 subtype might benefit from anti-HER2 agents such as lapatinib. Notably, to increase the clinical applicability of the EMT subtypes, we devised a 96-gene panel-based classifier via a machine learning framework. Conclusions: Our study identified three methylation-driven EMT subtypes with distinct prognoses and biological traits to capture heterogeneity in BC and provided a rationale for the use of this classification as a powerful tool for developing new strategies for clinical trials.