S2613
Radiobiology - Translational radiobiology
ESTRO 2026
upregulated genes across the three training sets. Prognostic utility was benchmarked against original tumour-specific signatures via Kaplan–Meier overall survival curves and multivariate Cox proportional hazards models adjusted for age, sex, and stage. Sequential multivariate models (clinical variables alone vs clinical and hypoxia signature) were used to quantify incremental prognostic value in the TCGA pooled validation cohort. Independent external validation was performed across five GEO datasets (GSE68465, GSE30219, GSE31210, GSE65858, GSE41613). Robustness was assessed through permutation testing of both survival labels and gene- set composition. Results: A total of 1,748 patients were included (LUAD/LUSC n=1,050; HNSC n=520; PAAD n=178). A robust 20-gene pan-cancer hypoxia signature was derived, comprising PLOD2, EGLN3, SLC2A1, ADM, ANGPTL4, ERO1L, LOX, GAPDH, SERPINE1, SLC16A3, STC2, PFKP, DDIT4, CA9, NDRG1, HK2, BNIP3, KCTD11, P4HA1, and VEGFA. Across all TCGA training cohorts, the signature demonstrated significantly superior prognostic performance compared with tumour-specific signatures; Table1. In pooled TCGA validation (n=478), incorporation of the signature provided a statistically significant but modest improvement in prognostic discrimination beyond clinical variables (HR 1.22, 95% CI 1.04–1.43; p=0.015), with a C-index increase from 0.612 to 0.623; figure 1A,B. External validation across all five GEO datasets (n= 1,303) confirmed consistent prognostic ability; Table1. Permutation testing demonstrated high robustness, outperforming randomised label permutations (empirical P ≤ 0.003) and randomly generated gene sets (P=0.02).
Made with FlippingBook - Share PDF online