1286 - Comparative Evaluation of Immune, Hypoxia, and DNA Repair Transcriptomic Signatures for Predicting Radiation Therapy Response in Head and Neck Squamous Cell Carcinoma

Purpose/Objective(s): Molecular predictors of RT response in head and neck squamous cell carcinoma (HNSCC) remain poorly defined. Pan-cancer hypoxia signatures predict RT response across tumor types but fail in HNSCC (AUC 0.458). We hypothesized that HNSCC's unique biology, HPV-driven immune activation and frequent TP53 mutations, requires pathway-specific biomarkers. We conducted a systematic comparison of DNA repair, hypoxia, and immune transcriptomic signatures, with and without HPV status, for predicting RT response.

Materials/Methods: We queried the TCGA-HNSC cohort (n=528) via ISB-CGC BigQuery, extracting clinical data, somatic mutations, RNA-seq, RT response, and HPV status. Of 294 RT-treated patients, 160 had documented response (140 sensitive, 20 resistant). Four pathway signatures were constructed: DNA repair (10 genes + TP53 mutation), hypoxia (HIF1A, CA9, VEGFA, LDHA), immune (PD-L1, CD8A, IFNG), and HPV status. Six models were evaluated, each pathway alone and combined with HPV, using Gradient Boosting with stratified 5-fold Computer Vision and class-balanced weighting. Feature importance was assessed by SHAP analysis.

Results: Immune + HPV achieved the highest AUC (0.633), followed by HPV alone (0.593) and immune-only (0.581). DNA repair + HPV (0.493), hypoxia + HPV (0.433), and the combined model (0.466) performed at or below chance. HPV status alone outperformed every non-immune molecular signature. SHAP analysis identified HPV status as the top feature alongside IFNG and CD8A. TP53 mutation lacked independent discriminatory power. DNA repair and hypoxia expression showed no association with RT outcome.

Conclusion: Immune features combined with HPV status (AUC 0.633) outperform DNA repair and hypoxia signatures for predicting HNSCC RT response. Hypoxia biomarkers validated in pan-cancer settings are ineffective in HNSCC, highlighting the need for tumor-type-specific radiogenomic approaches. These findings support immune-based stratification over DNA repair or hypoxia markers and motivate investigation of combined immunotherapy-RT predictive models incorporating PD-L1, CD8 infiltration, and IFNG signaling alongside HPV status.

Comparative Model Performance and SHAP Feature Importance for RT Response Prediction (N=160)

Sens - Sensitive patients (complete or partial response to RT)

Resist - Resistant patients (stable or progressive disease after RT)

IFNG - Interferon-gamma gene expression

CD8A - CD8+ T-cell marker gene expression

CV - Cross-Validation

Top SHAP - Most important features by SHAP (SHapley Additive exPlanations)

Category	Model	AUC	Genes	+HPV	Sens	Resist	N	Top SHAP	Rank
Best	Immune+HPV	0.633	3	Yes	140	20	160	HPV, IFNG	1
Viral	HPV Alone	0.593	0	Yes	140	20	160	HPV	2
Immune	Immune Only	0.581	3	No	140	20	160	IFNG, CD8A	3
Repair	DNA Rep+HPV	0.493	10	Yes	140	20	160	-	4
Combined	All+HPV	0.466	17	Yes	140	20	160	-	5
Hypoxia	Hypoxia+HPV	0.433	4	Yes	140	20	160	-	6
SHAP Features	Immune+HPV	-	-	-	-	-	-	hpv_pos	1st
		-	-	-	-	-	-	IFNG	2nd
		-	-	-	-	-	-	CD8A	3rd