AI-driven biomarker discovery for early diagnosis and prognosis in oral oncology

This study presents an AI-powered multi-omics framework for early detection and prognosis of oral squamous cell carcinoma (OSCC), integrating genomic, transcriptomic, and proteomic data through advanced deep learning architectures. Analysing 1,527 OSCC samples from TCGA and GEO databases, we developed a novel multimodal pipeline combining: (1) graph neural networks for heterogeneous data fusion, (2) LASSO regression for robust feature selection, and (3) explainable AI (SHAP, attention mechanisms) for clinical transparency. Our CNN-based diagnostic model demonstrated exceptional performance (accuracy: 93.2%, 95% CI: 91.4-94.7; sensitivity: 91.5% for Stage I tumors; AUC: 0.96), significantly surpassing conventional histopathology (p<0.001). Three clinically validated biomarker panels were established: (i) a diagnostic panel (TP53/CDKN2A/EGFR, 94.1% specificity), (ii) an HPV-associated prognostic panel (P16/RB1/E2F1), and (iii) a metastasis prediction panel (TWIST1/VIM/CDH1, C-index=0.82). Prospective validation in 412 patients showed 43% reduction in false negatives (15.2% to 8.7%) with 82% pathologist concordance. The modular platform addresses critical clinical needs: high-risk screening, therapeutic decision support, and intraoperative margin assessment. IRB-approved implementation confirms real-world viability, positioning this framework as a transformative tool for OSCC precision oncology. • This study leverages AI-driven models for early oral cancer detection by integrating multi-omics data to identify potential biomarkers and enhance clinical outcomes. • AI models achieve high accuracy (92%) and sensitivity (90%) in early-stage oral cancer detection. • Key biomarkers, including TP53, P16, and KRAS, are identified for diagnosis and personalized treatment. • Kaplan-Meier survival analysis stratifies patients into low-, moderate-, and high-risk groups. • ROC curve analysis validates diagnostic models with an AUC of 0.92, ensuring clinical relevance.