MRI-derived radiogenomic surrogate of a hypoxia–angiogenesis–EMT transcriptional state in nasopharyngeal carcinoma: multicenter external validation

Background This study aimed to develop a pretreatment MRI-derived radiogenomic surrogate of a prespecified hypoxia–angiogenesis–epithelial–mesenchymal transition transcriptional state and to evaluate its associations with treatment response and survival in nasopharyngeal carcinoma. Methods This multicenter observational study integrated patient-derived public transcriptomic datasets, experimental CNE2/CNE2-IR radioresistance models, a prospective RNA-sequencing cohort of 49 patients with paired contrast-enhanced T1-weighted MRI and RNA data, and two independent clinical cohorts (n = 532 and n = 118). HAE-CAS was the equal-weight mean of three within-cohort standardized module scores. Treatment response was assessed approximately 3 months after chemoradiotherapy according to RECIST version 1.1 using nasopharyngeal MRI, nasopharyngoscopy, and clinical evaluation. Rad-HAE used a five-feature Gaussian ridge model and was evaluated with 20 repetitions of five-fold nested cross-validation; response and survival outcomes were excluded from model development. Results Among 49 response-evaluable RNA-sequenced patients, 17 had complete response (CR) and 32 had partial response (PR); no stable or progressive disease occurred. HAE-CAS was higher in PR than CR (rank-biserial correlation = 0.365, P = 0.035). In GSE102349, continuous HAE-CAS was not significantly associated with progression-free survival (HR per 1-SD increase = 1.20, 95% CI 0.74–1.95; P = 0.465), whereas the median-dichotomized comparison was exploratory (log-rank P = 0.024). Aggregated out-of-fold Rad-HAE correlated with HAE-CAS (Spearman rho = 0.637, P < 0.001) and discriminated PR from CR among 49 paired response-evaluable patients (AUC = 0.846, 95% CI 0.732–0.943). In two independent clinical cohorts, higher Rad-HAE was associated with shorter progression-free and overall survival. Conclusion Pretreatment CE-T1WI radiomics was associated with interpatient variation in a prespecified composite HAE transcriptional state. Rad-HAE represents an imaging-based surrogate associated with the composite HAE transcriptional state and does not directly quantify individual biological processes.