MGP_HNSC32: A H&E Morphology-Derived 32-Gene Prognostic Signature in Head and Neck Squamous Cell Carcinoma

Head and neck squamous cell carcinoma (HNSC) exhibits substantial molecular and morphological heterogeneity, yet current morphology-based prognostic models often lack biological interpretability. We developed a transformer-based computational framework (CTransPath) to virtually predict gene expression directly from hematoxylin and eosin (H&E) whole-slide images. Morphology-associated genes were identified from the TCGA-HNSC cohort, and a 32-gene morphology-guided prognostic signature (MGP_HNSC32) based on H&E-derived predicted gene expression was developed using Random Survival Forest and multivariable Cox regression modeling. In TCGA-HNSC, the MGP_HNSC32 risk score significantly stratified patients by progression-free interval (PFI; HR = 4.13, 95% CI: 2.84-6.02), overall survival (OS; HR = 2.42, 95% CI: 1.78-3.29), and disease-specific survival (DSS; HR = 4.88, 95% CI: 2.95-8.06; all p < 0.001), and remained independently associated with prognosis after multivariable adjustment. High-risk tumors were enriched for cell cycle, E2F, MYC, and epithelial-mesenchymal transition pathways and displayed poorly differentiated histology with nuclear atypia and stromal desmoplasia. In contrast, low-risk tumors showed keratinizing morphology and immune-rich stroma, corresponding to differentiation- and metabolism-related transcriptional programs. In the independent HANCOCK cohort (n = 701), risk scores calculated entirely from H&E-derived predicted gene expression using the TCGA-derived coefficients and prespecified cutoff were associated with progression-free survival (HR = 1.40, 95% CI: 1.03-1.90, p = 0.034). However, the continuous risk score was not significantly associated with progression-free survival (HR = 1.001, 95% CI: 0.999-1.003, p = 0.328). MGP_HNSC32 links histomorphology to transcriptional programs and demonstrates preliminary prognostic associations using routine H&E images. The H&E-based gene-expression prediction component should be considered proof-of-concept, and further independent and prospective validation is required before clinical implementation.