Abstract / Summary
Prostate cancer treatment outcomes vary substantially between patients receiving radiation therapy alone versus combined radiation and androgen deprivation therapy (ADT), yet the molecular basis of these differences remains uncharacterised. Understanding how therapeutic interventions alter tumour transcriptomes is critical for personalising treatment. This study identifies treatment-associated transcriptomic differences, derives a molecular signature, and evaluates its prognostic and clinical utility. RNA-seq data and clinical annotations from 537 TCGA-PRAD patients were analysed, comparing radiation therapy alone versus combined radiation and ADT. Consensus differentially expressed genes (DEGs) were identified using a three-method framework (DESeq2, edgeR, limma-voom), with a separate tumour purity adjusted DESeq2 model used for functional enrichment. A machine learning framework combining LASSO feature selection and balanced random forest classification derived a compact molecular signature. Prognostic performance was assessed using multivariable Cox regression, bootstrap validation, and decision curve analysis. A consensus set of 765 DEGs was identified, with CDK1, BUB1, and AURKB as central hub genes. Distinct transcriptional programmes were observed, particularly in cell cycle regulation, DNA damage response, and immune processes. An 8 gene signature (H3C2, KRT222, AKR1D1, PLEKHD1, IBSP, PNMA6E, ADCYAP1R1, IL36RN) distinguished treatment groups with AUC = 0.803. The molecular risk score retained independent prognostic significance after adjustment for Gleason grade (HR = 1.39; p = 0.004). Exploratory external evaluation using a reduced 5-gene score showed no statistically significant associations with BRFS or MFS and limited discriminatory performance. Thus, the findings provide internal support for the prognostic relevance of the 8-gene signature, while its external generalisability and clinical utility require further independent validation.