2122 - Multi-Omics Prediction of Radiosensitivity in Pancreatic Adenocarcinoma: A Machine Learning Radiogenomics Model

Purpose/Objective(s): Radiotherapy (RT) outcomes for pancreatic adenocarcinoma (PAAD) vary widely, with no validated genomic biomarkers to predict radiosensitivity or guide SBRT patient selection. We hypothesized that integrating somatic mutations, transcriptomic hypoxia signatures, and tumor mutational burden could predict RT response. We developed a machine learning radiogenomics model using TCGA multi-omics data.

Materials/Methods: We queried TCGA-PAAD (n=185) extracting clinical data, somatic mutations, RNA-seq (TPM), and copy number segments. Of 41 RT-treated patients, 26 had documented response (14 sensitive, 12 resistant). We engineered 19 features: somatic mutations (KRAS, TP53, CDKN2A, SMAD4, BRCA1/2, ATM), a composite hypoxia score from 10 genes (HIF1A, EPAS1, VEGFA, SLC2A1, LDHA, CA9, LOX, ADM, PGK1, BNIP3), log mutation burden, and clinical covariates. Gradient Boosting and Random Forest were trained using LOO-CV with class-balanced weighting. Feature importance was assessed by SHAP.

Results: Gradient Boosting achieved AUC 0.702 (accuracy 73%, sensitive precision 82%, resistant 67%); Random Forest AUC 0.607. SHAP identified the hypoxia score as the most impactful predictor, followed by log mutation burden, KRAS status, HIF1A, and SLC2A1. Higher hypoxia scores associated with radioresistance. KRAS-mutant tumors trended toward resistance. DNA repair mutations (BRCA1/2, ATM) had lower SHAP impact due to low mutation frequency.

Conclusion: A machine learning model integrating transcriptomic hypoxia signatures, somatic mutations, and mutational burden achieved meaningful discrimination (AUC 0.70) for PAAD radiosensitivity. The hypoxia score was the dominant predictor, reinforcing tumor hypoxia as a driver of radioresistance. These findings support developing radiogenomic biomarkers for SBRT patient selection in pancreatic cancer. Integration of computational pathology features from matched histopathology may further improve performance.

Model Performance, Classification Metrics, and SHAP Feature Importance (N=26)

GB - Gradient Boosting

RF - Random Forest

Sens - Number of sensitive patients

Resist - Number of resistant patients

Prec(S) - Precision for the sensitive class

Prec(R) - Precision for the resistant class

Hypoxia score - Composite z-score from 10 canonical hypoxia genes

Log mut burden - Log-transformed total somatic mutation count

HIF1A - Hypoxia-Inducible Factor 1 Alpha

SLC2A1 - Solute Carrier Family 2 Member 1

KRAS - Kirsten Rat Sarcoma viral oncogene

LOO-CV - Leave-One-Out Cross-Validation

PAAD - Pancreatic Adenocarcinoma

SBRT - Stereotactic Body Radiation Therapy

Category	Metric	GB	RF	Sens	Resist	N	Prec(S)	Prec(R)	SHAP Rank
Performance	AUC	0.702	0.607	14	12	26	0.82	0.67	-
	Accuracy	73%	65%	-	-	-	-	-	-
SHAP Top 5	Hypoxia score	-	-	-	-	-	-	-	1st
	Log mut burden	-	-	-	-	-	-	-	2nd
	KRAS mutation	-	-	-	-	-	-	-	3rd
	HIF1A	-	-	-	-	-	-	-	4th
	SLC2A1	-	-	-	-	-	-	-	5th
Features	Hypoxia genes	10	10	-	-	-	-	-	Dominant
	Total features	19	19	-	-	-	-	-	-