Development and multi-cohort validation of a prognostic risk score model for oral squamous cell carcinoma based on a three-gene signature

Abstract

Background: Oral squamous cell carcinoma (OSCC) carries substantial mortality despite surgery-based management. Reliable biomarkers and practical prognostic tools are needed to guide individualized care.

Methods: Transcriptomic and clinical data from TCGA and multiple GEO cohorts were integrated. Candidate genes identified by differential expression analysis, WGCNA, and PPI were refined using LASSO and Random Forest, then entered a multivariable Cox model to derive a three-gene risk score. Performance was assessed with Kaplan-Meier curves, time-dependent ROC, and calibration, and validated across internal, external, and combined datasets. Expression was examined by RT-qPCR and Western blot in OSCC and normal oral cell lines. Immune infiltration and pathway enrichment analyses were conducted to contextualize biology.

Results: The three-gene signature (CXCL12, PLAU, PXDN) separated risk groups in the training cohort with 1/3/5-year AUCs of 0.767/0.625/0.714. In three independent external cohorts, high-risk patients consistently had worse overall survival (log-rank p = 0.0092, ≤0.0001, ≤0.0001), with time-AUC ranges of 0.581-0.747 (1-year), 0.555-0.795 (3-year), and 0.603-0.812 (5-year). In TCGA, the score remained prognostic across sex, age, smoking, drinking, and stage III/IV subgroups (all p ≤ 0.05), with a consistent trend in stage I/II (p = 0.09). A nomogram integrating clinical variables with the risk score achieved a C-index of 0.63 with good 1-5-year calibration and outperformed TNM staging alone (C-index 0.63 vs 0.58; 95 % CI 0.58-0.70 vs 0.54-0.61). RT-qPCR/Western blot confirmed consistent differential expression of all three biomarkers. Immune-infiltration and pathway analyses revealed distinct microenvironmental and molecular features across risk strata.

Conclusion: We present a robust, externally validated three-gene prognostic model for OSCC, supported by experimental evidence and superior to TNM staging for discrimination, offering a practical nomogram for individualized risk estimation from 1 to 5 years.