1153 - Novel Counterfactual Machine Learning Reveals Treatment-Specific Radiosensitivity in Gastric Cancer

Purpose/Objective(s):

Radiation therapy (RT) in gastric adenocarcinoma (STAD) is applied heterogeneously in the adjuvant setting without a validated molecular framework to identify patients most likely to benefit. Existing genomic classifiers are prognostic and do not distinguish treatment-specific effects. We developed a machine learning treatment-interaction framework to derive and validate a transcriptomic radiosensitivity signature that predicts differential benefit from RT rather than baseline survival.

Materials/Methods:

TCGA-STAD RNA-seq data (n=415) were integrated with clinical outcomes and adjuvant RT status. Among patients with known RT exposure and survival (n=203; RT=46), radiosensitivity was defined as 24-month survival among RT-treated patients after excluding early censoring (n=40 labeled). Feature reduction used nested univariate Kaplan–Meier screening within cross-validation folds (top 200 log-rank genes from 5,000 most variable genes). Models included logistic regression, random forests, XGBoost, multilayer perceptrons, and generative latent-variable approaches (PCA/LDA, factor analysis). Performance was assessed using 5-fold stratified cross-validated AUC. To separate predictive from prognostic signal, counterfactual treatment-interaction modeling estimated survival risk under RT=1 and RT=0, generating an individualized RT benefit score. Patients were stratified by predicted benefit and tested for differential hazard ratios (HR).

Results:

Using a pre-specified 20-gene panel (RS20), logistic regression achieved mean cross-validated AUC=1.00 for 24-month mortality among RT-treated patients. With de novo nested gene selection, tree-based models performed strongly (Random Forest AUC=0.80; XGBoost AUC=0.76), outperforming neural networks in this small-sample setting. Generative latent-factor models showed comparable discrimination (AUC˜0.72), supporting robustness across paradigms. Unsupervised clustering identified two molecular subtypes with significantly different post-RT survival (log-rank p=2.6×10?4). In the full RT-known cohort, adjuvant RT conferred significant survival benefit in the radiosensitive subtype (median OS 36.9 vs 15.1 months; p=4.4×10?4) but not in the resistant subtype (p=0.71). Counterfactual modeling demonstrated marked heterogeneity of RT effect: machine learning–defined radiosensitive patients exhibited substantial RT-associated hazard reduction (HR˜0.17–0.35), whereas resistant patients derived minimal benefit.

Conclusion:

We introduce a novel counterfactual treatment-interaction machine learning framework for identifying transcriptomic radiosensitivity in gastric cancer. Unlike conventional prognostic models, this approach directly models differential RT benefit, enabling biologically informed treatment stratification and providing a generalizable computational paradigm for treatment-specific biomarker discovery in radiation oncology.