<p>Breast cancer, notable for its extensive heterogeneity, remains a serious threat to public health. Accurate stratification of breast cancer cases based on somatic mutation profiles offers precise guidance for personalized treatment. However, this area has been challenging due to intrinsic sparsity of somatic mutation profiles. In this study, a smooth network propagation-based gene profile is constructed from the sparse mutation profile of patients with breast cancer using information derived from a protein interaction network. By integrating the network propagation-based gene profile with an immune-related background network, an immune-related delta rank matrix (IRDRM) is developed. A deep clustering algorithm is then executed on prognosis-related gene pairs from IRDRM, categorizing patients into distinct clustering subtypes characterized by biological and clinical relevance. We find that our deep clustering-based subtypes are associated with prognosis, clinicopathological features, immune infiltration levels, and response to chemotherapy and immunotherapy. Furthermore, a predictive model for subtyping is constructed using the XGBoost algorithm, which achieves favorable prediction results. SHAP is subsequently employed to identify the gene pairs that contribute most significantly to the prediction accuracy of the XGBoost algorithm. Our study establishes IRDRM that integrates somatic mutation data with gene pairs within a network to aid in the identification of cancer subtypes, which could potentially advance personalized treatment of breast cancer.</p> Graphical Abstract <p>Core aspects of this study:&#xa0;(a) discovery of immune-related subtypes of breast cancer patients from the delta rank matrix; (b) biological and immunogenomic characteristics of clustering subtypes; (c) construction of XGBoost classifier for breast cancer patients; (d) validating the clustering model in TCGA cohort, METABRIC cohort and 13 pan-cancer cohorts.</p> <p></p>