Abstract
BACKGROUND: Breast cancer is an invasive disease with complex molecular mechanisms. Prognosis-related biomarkers are still urgently needed to predict outcomes of breast cancer patients. METHODS: Original data were download from The Cancer Genome Atlas (TCGA) and the Gene Expression Omnibus (GEO). The analyses were performed using perl-5.32 and R-x64-4.1.1. RESULTS: In this study, 1086 differentially expressed genes (DEGs) were identified in the TCGA cohort; 523 shared DEGs were identified in the TCGA and GSE10886 cohorts. Eight subtypes were estimated using non-negative matrix factorization clustering with significant differences seen in overall survival (OS) and progression-free survival (PFS) (P < 0.01). Univariate Cox analysis and least absolute shrinkage and selection operator (LASSO) regression analysis were performed to develop a related risk score related to the 17 DEGs; this score separated breast cancer into low- and high-risk groups with significant differences in survival (P < 0.01) and showed powerful effectiveness (TCGA all group: 1-year area under the curve [AUC] = 0.729, 3-year AUC = 0.778, 5-year AUC = 0.781). A nomogram prediction model was constructed using non-negative matrix factorization clustering, the risk score, and clinical characteristics. Our model was confirmed to be related with tumor microenvironment. Furthermore, DEGs in high-risk breast cancer were enriched in histidine metabolism (normalized enrichment score [NES] = 1.49, P < 0.05), protein export (NES = 1.58, P < 0.05), and steroid hormone biosynthesis signaling pathways (NES = 1.56, P < 0.05). CONCLUSIONS: We established a comprehensive model that can predict prognosis and guide treatment.