Shared Depletion of p53, Apoptosis and TNF-α/NF-κB Gene-Expression Programmes in Prostate Cancer Cell Lines Is Also Observed in Patient Tumours, and a Sulforaphane-Derived Signature Reflects NRF2-Associated Activity: An Integrative Re-Analysis of Gen

Background: RNA-sequencing series GSE48812, including PrEC, LNCaP and PC-3 cells ±15 μM sulforaphane (SFN) for 6 and 24 h, has been used to characterise SFN responses. We separated baseline cancer–normal differences from treatment, time, and cell-line effects and assessed relevance in tissue. Methods: NCBI counts were re-analysed using DESeq2. After reproducing the GEO2R contrast, vehicle-only, time-adjusted models were fitted for the pooled and individual cancer lines. Pre-ranked GSEA identified shared gene sets. TCGA-PRAD (497 tumours, 52 normals, 52 matched pairs) was analysed using composition-adjusted paired regression. SFN responses were assessed by variance partitioning and line × treatment interactions, followed by leave-one-cell-line-out machine learning with nested feature selection. The signature was tested in seven datasets (12 contrasts) against an NRF2 panel. Results: The published contrast pooled treated and vehicle samples; our pipeline reproduced it (log2 fold-change r = 1.000) and identified 4901 differentially expressed genes between cancer and normal cells. Sixteen Hallmark sets were depleted in both cancer lines relative to PrEC. Shared genes separated tumours from normal tissue (AUC = 0.93); p53, apoptosis, and TNF-α/NF-κB depletion retained 67–87% of effects after adjustment (FDR < 0.05). In a gene-level ANOVA, cell line explained a median 94% of expression variance versus 2.5% for SFN. PrEC showed stronger NRF2 induction. A 17-gene signature trained on SFN exposure transferred between lines (AUC 1.00) but did not outperform an untrained 16-gene NRF2 panel (AUC 1.00), so machine learning added no discrimination. Externally, the signature rose with KEAP1 knockdown and NRF2 overexpression and fell with NRF2 knockdown; six canonical NRF2 genes carried the transferable signal. Conclusions: Baseline cell-line biology dominates the transcriptomic signal, whereas SFN contributes relatively little. Shared depletion of p53, apoptosis and TNF-α/NF-κB gene-expression programmes is also observed in tumours but does not establish functional pathway loss, and the learned signature reflects NRF2-associated activity rather than serving as a specific predictor of SFN exposure or response.