Abstract
Prognostic transcriptomic signatures for prostate cancer (PCa) often overlook the cellular origin of expression changes, an important consideration given the heterogeneity of the disorder. Current clinicopathological factors inadequately predict biochemical recurrence, a critical indicator guiding post-treatment strategies following radical prostatectomy. To address this, we conducted a meta-analysis of four large-scale PCa datasets and found 33 previously reported PCa-associated genes to be consistently up-regulated in prostate tumours. By analysing single-cell RNA-sequencing data, we found these genes predominantly as markers in epithelial cells. Subsequently, we applied 97 advanced machine-learning algorithms across five PCa cohorts and developed an 11-gene epithelial expression signature. This signature robustly predicted biochemical recurrence-free survival (BCRFS) and stratified patients into distinct risk categories, with high-risk patients showing worse survival and altered immune cell populations. The signature outperformed traditional clinical parameters in larger cohorts and was overall superior to published PCa signatures for BCRFS. By analysing peripheral blood data, four of our signature genes showed potential as biomarkers for radiation response in patients with localised cancer and effectively stratified castration-resistant patients for overall survival. In conclusion, this study developed a novel epithelial gene-expression signature that enhanced BCRFS prediction and enabled effective risk stratification compared to existing clinical- and gene-expression-derived prognostic tools. Furthermore, a set of genes from the signature demonstrated potential utility in peripheral blood, a tissue amenable to minimally invasive sampling in a primary care setting, offering significant prognostic value for PCa patients without requiring a tumour biopsy.