Breast cancer age-linked genes fall from 80 to four: only IL7R, SQSTM1, EGR1 and IGF1 survive replication and receptor robustness

Age-associated expression signals in breast tumors are compressed into a tiered candidate list that separates cross-cohort replication, receptor dependence, and network or functional convergence alone. Earlier age-related gene screens largely rested on statistical association within a single cohort, without cross-platform replication or receptor-status stratification. Among 1,079 TCGA-BRCA primary tumors, 292 of 307 GenAge genes were expression-testable; continuous-age screening yielded 126 FDR-significant genes, and adjustment for ER, PR, HER2 and pathological stage left 80.

Cross-cohort replication cut the 80 adjusted signals to eight, and estrogen-receptor stratification separated four genes with stable direction and no significant between-stratum heterogeneity. Before replication, the 80 signals could not be separated into those dependent on estrogen-receptor status and those with concordant direction across cohorts. In 104 GSE42568 tumors, EGR1, FOXO1, IGF1, EIF5A2, A2M, IL7R, SQSTM1 and ESR1 replicated directionally at BH-adjusted one-sided P<0.05; ESR1 showed age correlations of ρ=0.426 in ER-positive and ρ=0.062 in ER-negative tumors, with heterogeneity FDR of 3.9×10^-5.

The same molecular pipeline is represented as a bounded symbolic controller that executes retrieval, identifier mapping, replication, network and enrichment steps within a predefined action space, deciding to complete, block or fail safe according to evidence state. Prior workflow evaluation looked only at whether the final artifact was correct, although two trajectories can reach the same artifact while differing in safety, evidentiary adequacy and provenance. On 760 prespecified scenarios whose reference policy comprised 629 COMPLETE, 60 BLOCK and 71 FAIL_SAFE outcomes, the bounded controller reached policy conformance of 1.000, versus 0.075 for a static workflow, 0.703 for a goal planner and 0.782 for an adaptive planner.

Single-component ablation shows each governance module maps to a distinct failure mode, so endpoint correctness and process quality can diverge. Judged by endpoint conformance alone, removing cost awareness leaves the controller at 1.000 and the process degradation stays invisible. Removing the evidence verifier dropped evidence-quality family conformance from 1.00 to 0.00, removing the safety gate dropped the safe-block rate from 1.00 to 0.00, and removing cost awareness dropped action appropriateness from 1.000 to 0.845 and resource stewardship from 0.988 to 0.883.

The analysis uses public data: 1,079 primary tumors from TCGA-BRCA and 104 pretreatment biopsies from GSE42568, with age at diagnosis treated as a continuous variable. Replication tests cross-platform directional agreement of expression only and is not prospective or clinical validation. The orchestration experiment runs within a predefined action space, and benchmark and controller come from the same study, making it an internal execution-policy evaluation rather than an external test of open-domain agent generalization.

Chronological age at diagnosis is not biological age, and age-expression associations are observational, so they do not establish causality or an ageing mechanism. TCGA-BRCA and GSE42568 differ in assay platform, cohort composition and clinical annotation. STRING aggregates known and predicted protein associations rather than breast-tumor-specific physical interactions. No wet-lab validation, survival model, diagnostic classifier or treatment-response model was performed, and no compound screening, docking, pharmacokinetic or pharmacodynamic modeling was included.