Integrative transcriptomic analysis identifies a four‑gene autophagy‑related diagnostic classifier and an associated candidate transcript, FOS, in breast cancer

Breast cancer exhibits substantial transcriptional and microenvironmental heterogeneity. Autophagy‑related transcripts may provide diagnostic information and cell‑type‑specific biological context, but their relationship with immune features requires cautious evaluation. We analysed three GEO breast cancer expression datasets (GSE10810, GSE22820, GSE42568) and TCGA‑BRCA data to identify consistently dysregulated autophagy‑related genes. Candidate diagnostic models were fitted in GSE10810 and evaluated across GSE22820 and GSE42568. Because the evaluation cohorts contributed to final model ranking, performance is reported as cross‑cohort evaluation rather than independent validation. Tumor‑only non‑negative matrix factorization (NMF) was performed in 31 GSE10810 tumours using five preselected genes (BIRC5, CXCR4, EGFR, FOXO1, FOS). Immune scores were compared with effect sizes and Benjamini‑Hochberg correction. GSE198745 single‑cell data were reannotated and used for FOXO1‑associated T/NK‑cell pathway analysis and BayesPrism deconvolution of GSE10810. The fitted glmBoost plus plsRglm classifier contained four genes—BIRC5, CXCR4, EGFR, and FOXO1—and achieved cross‑cohort AUCs of 0.985 in GSE22820 and 0.947 in GSE42568. The fifth candidate gene, FOS, was not retained in this classifier but was analysed separately as a transcript with exploratory biological associations. Tumor‑only NMF generated two exploratory groups of 15 and 16 tumours (cophenetic coefficient 0.933, consensus silhouette 0.807). Cluster 1 was enriched for DNA‑replication and cell‑cycle programs, whereas Cluster 2 was enriched for TGF‑beta and complement/coagulation pathways. No immune‑cell comparison remained significant after FDR correction. Corrected single‑cell analysis identified 1,216 T/NK cells; FOXO1‑high cells showed enrichment of T‑cell activation, antigen processing and presentation, and Th17‑cell differentiation programs. BayesPrism deconvolution did not support a tumour‑internal association between EGFR and CAF abundance. The identified transcripts provide reproducible diagnostic and biological associations in breast cancer. The four‑gene classifier and the candidate FOS are exploratory tools that require independent prospective validation and functional testing. These results do not establish direct autophagic activity, causal immune regulation, or clinical utility.