An information-theoretic framework identifies tumor dormancy–associated gene signatures for very late recurrence in ER-positive breast cancer

Late recurrence beyond 10 years after diagnosis remains a major unresolved challenge in luminal A breast cancer, and current multigene assays have limited predictive power for long-latency events. We developed MI-PER (Mutual Information-based Person-year-anchored Extraction of Risk markers), an information-theoretic framework that integrates normalized mutual information with time-resolved recurrence rates, to identify gene expression features associated with very late recurrence. MI-PER was applied to 461 endocrine therapy-treated patients with luminal A breast cancer in the METABRIC cohort. Candidate genes were selected by combining MI-based feature selection with genome-wide person-year recurrence screening. Biological signatures derived from enrichment analysis were evaluated in an independent cohort (GSE17705), and propensity score matching was applied in METABRIC to assess robustness against clinical confounders. Mechanistic relevance was explored using single-cell RNA sequencing (scRNA-seq) from a breast cancer dormancy model and functional assays in luminal A breast cancer cells under dormancy-inducing conditions. MI-PER identified 81 genes associated with recurrence beyond 10 years. These genes showed no overlap with the prognostic content genes of widely adopted multigene assays, suggesting that MI-PER captured late recurrence-associated features largely distinct from established prognostic gene sets. Functional enrichment analysis revealed pathways related to cytoskeletal dynamics, cell adhesion, and lipid metabolism—processes implicated in tumor cell dormancy—and refined the genes into four biologically coherent signatures. In GSE17705, elevated signature scores at diagnosis were associated with increased late recurrence risk by Kaplan–Meier and person-year analyses. In METABRIC, these associations remained consistent after propensity score matching. scRNA-seq analysis revealed preferential expression of late recurrence-associated genes, most notably SH3KBP1 , in dormant disseminated tumor cell-like populations. SH3KBP1 knockdown significantly impaired cell viability under dormancy-inducing conditions, supporting its contribution to cell viability in this model. MI-PER links diagnostic gene expression profiles to very late recurrence risk in luminal A breast cancer. Our findings connect dormancy-associated transcriptomic features at diagnosis to prolonged disease latency and identify SH3KBP1 as a candidate contributor to maintaining cell viability under dormancy-inducing conditions, with potential implications for understanding late metastatic recurrence.