Abstract
<title>Abstract</title> <p> Background. Breast cancer is molecularly heterogeneous, and multigene expression signatures have become useful additions to conventional clinicopathological staging for prognostic stratification. This study aimed to derive and externally validate a prognostic gene expression signature using a transparent bioinformatics pipeline that directly addresses several methodological limitations common to retrospective signature studies based on The Cancer Genome Atlas (TCGA), including inadequate control of gene-selection bias and the frequent absence of independent external validation. Methods. RNA sequencing and clinical data from 1,043 TCGA breast invasive carcinoma patients (139 deaths) were analyzed using a multi-step sample and gene selection pipeline. Differential expression analysis using DESeq2 and multi-criteria filtering reduced 56,705 tested genes to 3,691 candidates. Univariate Cox screening was then performed for two independent survival endpoints, overall survival (OS) and progression-free interval (PFI), followed by multivariate modeling using backward elimination. This yielded a parsimonious six-gene signature comprising <italic>XG, FIBCD1, LINC01235, CEL, VGF</italic> and <italic>CEMIP</italic> . A risk score was constructed, and patients were stratified at the median value. External validation was performed in the independent METABRIC cohort (n = 1,979) using platform-standardized expression values. Results. The signature significantly stratified overall survival (log-rank p = 4 × 10⁻⁶) and progression-free interval (p = 7 × 10⁻⁷) in the TCGA cohort, with time-dependent areas under the curve of 72.3% and 70.8% at three and five years, respectively. A combined clinicogenomic nomogram incorporating age, stage, and molecular subtype achieved a concordance index of 0.769, compared with 0.681 for the gene signature alone. This performance is comparable to independently re-derived Oncotype DX and MammaPrint models (concordance index approximately 0.68) and to other recently published TCGA-derived multigene signatures. In METABRIC, the continuous risk score was significantly associated with both overall survival (hazard ratio 1.284, p = 4.65 × 10⁻⁵) and relapse-free survival (hazard ratio 1.249, p = 0.0022), confirming generalisability across an independent, cross-platform cohort. Conclusions. A novel six-gene prognostic signature for breast cancer was identified and externally validated using a transparent, power-appropriate feature-selection strategy adopted in place of conventional LASSO regularisation, which failed to converge on a non-null model at the available event count. Four signature genes ( <italic>CEL, LINC01235, FIBCD1</italic> , and <italic>CEMIP</italic> ) have established prognostic roles in breast cancer, whereas two ( <italic>VGF</italic> and <italic>XG</italic> ) represent candidate novel components that warrant functional validation. </p>