Main Session
Sep 28
PQA 03 - Digital Health Innovation and Informatics, Patient Safety & Quality, and Radiation and Cancer Biology

2513 - Development and Evaluation of a Pan-Cancer Transcriptomic Hypoxia Signature for Radiation Therapy Response Prediction across Tumor Types: A TCGA Multi-Omics Analysis

10:45am - 12:00pm ET
Poster Hall - Exhibit Hall A
Screen: 4
POSTER

Presenter(s)

Jeril Lasington, MD, MS, MBBS - New York Medical College at St.Mary's and St. Clare's, Denville, NJ

J. Lasington1, L. S. Mathew Lasington2, and S. Umamaheshwaran3; 1New York Medical College at St. Mary's and St. Clare's, Denville, NJ, 2Rutger's University, East Hanover, NJ, 3Personio, Manhattan, NY

Purpose/Objective(s): Tumor hypoxia drives radioresistance, yet no validated pan-cancer transcriptomic hypoxia biomarker exists for predicting RT response. We developed a hypoxia gene signature as a universal predictor of radiosensitivity using multi-omics data from TCGA across 33 cancer types, to determine whether a single hypoxia-based biomarker can identify patients likely to benefit from RT irrespective of histology.

Materials/Methods: We queried the TCGA pan-cancer cohort (n=11,301; 33 cancer types) via ISB-CGC BigQuery, extracting clinical data, RNA-seq expression (TPM), somatic mutations, and RT response. Of 3,117 RT-treated patients, 1,414 had documented response (1,027 sensitive, 387 resistant). A composite hypoxia score was constructed from 10 canonical hypoxia genes (HIF1A, EPAS1, VEGFA, SLC2A1, LDHA, CA9, LOX, ADM, PGK1, BNIP3). The feature set was restricted to 13 hypoxia-centric variables. Gradient Boosting with stratified 5-fold CV and class-balanced weighting was used. SHAP values assessed feature importance. Per-cancer AUC subgroup analysis was performed for tumor types with n=10.

Results: The pan-cancer model achieved AUC 0.726 (accuracy 77%). SHAP analysis identified LDHA as the most impactful feature, followed by CA9, SLC2A1, VEGFA, and ADM - all canonical HIF-downstream targets. Per-cancer analysis showed strongest performance in GBM (AUC 0.750, n=11), melanoma (0.708, n=41), breast (0.658, n=191), and sarcoma (0.656, n=59). The signature underperformed in HNSC (0.458, n=160) and LGG (0.482, n=165), suggesting alternative radioresistance mechanisms. Hypoxia gene co-expression confirmed coherent pathway activation (LDHA-PGK1 r=0.61).

Conclusion: A 10-gene hypoxia signature achieves robust pan-cancer RT response prediction (AUC 0.73) across 1,414 patients from 33 tumor types. The signature performed best in hypoxia-driven tumors (GBM, melanoma, sarcoma), with LDHA and CA9 as top predictors. Cancer types where the signature failed (HNSC, LGG) may require immune- or DNA-repair-based biomarkers, motivating multi-pathway radiogenomic models.

Pan-Cancer Hypoxia Model Performance, Cancer-Type Subgroup AUCs, and SHAP Feature Importance (N=1,414)

Acc - Accuracy

Sens N - Number of sensitive patients

Res N - Number of resistant patients

Prec - Precision

Recall - Recall

SHAP Rank - Feature importance ranking by SHAP (SHapley Additive exPlanations) values

LDHA - Lactate Dehydrogenase A

CA9 - Carbonic Anhydrase 9

GBM - Glioblastoma Multiforme

HNSC - Head and Neck Squamous Cell Carcinoma

LGG - Low-Grade Glioma

CV - Cross-Validation (stratified 5-fold)

TCGA - The Cancer Genome Atlas

HIF - Hypoxia-Inducible Factor

Category

Metric

AUC

Acc

Sens N

Res N

Total

Prec

Recall

SHAP Rank

Pan-Cancer

Overall

0.726

77%

1027

387

1414

0.75

0.77

-

Top Cancers

GBM

0.750

-

-

-

11

-

-

Hypothesis

Melanoma

0.708

-

-

-

41

-

-

Strong

Breast

0.658

-

-

-

191

-

-

Moderate

Sarcoma

0.656

-

-

-

59

-

-

Moderate

Weak

HNSC

0.458

-

-

-

160

-

-

Failed

LGG

0.482

-

-

-

165

-

-

Failed

SHAP Top 3

LDHA

-

-

-

-

-

-

-

1st

CA9

-

-

-

-

-

-

-

2nd