Cross-Cohort Transcriptomic Analysis Identifies an ECM–CAF Stromal Program but Does Not Validate a Seven-Gene Prognostic Score in Head and Neck Squamous Cell Carcinoma

Abstract

Background: Head and neck squamous cell carcinoma (HNSCC) is biologically heterogeneous, and many public-data gene signatures lack cross-platform replication and unbiased evaluation. Objective: We sought reproducible tumor–normal expression programs and tested the transportability of a derived seven-gene score. Methods: TCGA-HNSC RNA-sequencing defined differentially expressed genes (DEGs) between 520 tumors and 44 normal tissues using TMM normalization and voom–limma, with paired sensitivity analysis. Concordant genes were replicated in 22 GSE6631 matched pairs and analyzed for GO and KEGG enrichment. A seven-gene overall-survival score was developed in event-stratified TCGA training data (n=361), evaluated in held-out TCGA (n=156), repeated nested resampling, and examined in GSE65858 (n=253). ESTIMATE and marker scores characterized tumor-microenvironment features. Results: Among 5,657 TCGA and 168 GSE6631 DEGs, 151 replicated. Upregulated genes were enriched for extracellular-matrix organization, ECM–receptor interaction, integrin signaling, focal adhesion, and PI3K–Akt signaling. The score was associated with survival in training (HR per SD=1.68; C-index=0.642) but not held-out TCGA (HR=1.02; C-index=0.546) or unadjusted GSE65858 (HR=1.15; C-index=0.548). Median nested C-index was 0.564. HPV adjustment attenuated the GSE65858 association. The score correlated with CAF/fibroblast (ρ=0.269) and stromal scores (ρ=0.228). Conclusion: Replicated expression changes support an ECM–CAF program, but the seven-gene score did not generalize and is not a validated prognostic model.