Machine learning defines a histone deacetylase-associated transcriptional prognostic signature with single-cell resolution in breast cancer.

BACKGROUND: Breast cancer (BC) is a highly heterogeneous malignancy, and transcriptional programs associated with histone deacetylases (HDACs) may provide clinically relevant prognostic information. However, the downstream prognostic significance of HDAC-associated molecular heterogeneity remains incompletely understood. METHODS: Non-negative matrix factorization (NMF) was performed in the TCGA cohort to identify HDAC-related molecular subtypes, followed by weighted gene co-expression network analysis and differential-expression analysis to derive candidate genes. A prognostic signature was then constructed using multiple machine-learning algorithms and validated across external BC cohorts. Additional analyses included multivariable and continuous-risk Cox regression, correlation analysis between the HDAC-related risk score (HRS) and HDAC family member expression, and single-cell gene-set scoring using Seurat's AddModuleScore function. Key hub genes were further examined by quantitative real-time polymerase chain reaction (qRT-PCR) and western blotting (WB). RESULTS: NMF consensus clustering identified five HDAC-related molecular clusters with distinct survival outcomes and immune-infiltration patterns in TCGA-BRCA. WGCNA further identified subtype-associated co-expression modules, and candidate features were obtained by intersecting HDAC-cluster-associated DEGs with WGCNA-derived hub genes. A 24-gene HRS model based on CoxBoost + SuperPC showed stable prognostic performance across the TCGA training cohort and seven external breast cancer validation cohorts. Multivariable Cox analyses supported the independent prognostic value of HRS in several major cohorts. Single-cell analysis showed cell-type-specific heterogeneity of the model-related signature score, and UTRN, LIMCH1, ARNT2, and PYDC1 showed concordant expression patterns in public datasets and preliminary qRT-PCR/WB validation. CONCLUSION: These findings suggest that the HDAC-associated HRS model may serve as an exploratory prognostic framework for breast cancer, while further mechanistic and prospective validation is required before clinical translation.