Development and validation of a machine learning-driven mitochondrial gene signature for the diagnosis of breast cancer.

Background: Breast cancer (BC) ranks among the most prevalent malignant tumors in women globally, with mitochondrial dysfunction constituting one of its pathogenic mechanisms. Objectives: To investigate the relationship between mitochondrial function-related genes and BC progression. Methods: We identified BC differentially expressed genes via the GEO database, constructed a weighted co-expression network to determine BC pathogenesis-related key modules. Using 113 machine learning algorithms and MitoCarta mitochondrial genetics data, we developed a mitochondrial gene-based diagnostic model. GO/KEGG enrichment analyses delineated BC-related biological processes of mitochondrial genes, offering clues for understanding BC mechanism. High-throughput tissue chip and Immunohistochemistry (IHC) validated key genes' local expression in tissues. CiberSort immune infiltration analysis highlighted NK and T cells' role in BC; single-cell analysis identified gene expression patterns across tumor microenvironment cell types. Computational drug prediction and molecular docking explored targeted therapeutic candidates. Additionally, we conducted molecular dynamics simulations. Results: The glmBoost+LDA model had the highest C-index (0.947) in the validated cohort, including 18 potential BC biomarkers (e.g., ACADS, AUC = 0.810; AIFM2, AUC = 0.806). The results of experimental validation showed that the expression score of ACADS in cancerous tissues was significantly lower than that in adjacent non-cancerous tissues. Immune infiltration and single-cell analyses emphasized the crucial roles of NK cells and T cells in BC. Disulfiram and eugenol were predicted as potential therapeutics and validated by docking. Molecular dynamics simulations validated that Eugenol exhibits strong binding interactions with the target proteins AIFM2 and ACADS. Conclusions: This study identifies mitochondrial gene signatures associated with BC and proposes a computational model distinguishing tumor from normal tissue. These findings offer potential leads for future biomarker development but require additional clinical and functional validation.