TRAPS benchmark reveals no single AI model wins all cancer therapy predictions

A new paper on arXiv introduces TRAPS (Therapeutic Response Analysis via Pathway-informed Stratification), the first unified benchmark designed to fairly compare pathway-informed deep learning models for cancer treatment planning. The study evaluates three biologically informed architectures—BINN, GraphPath, and PATH—across five cancer cohorts from The Cancer Genome Atlas (TCGA), representing 2,622 patients encoded using Reactome pathway activity scores. Crucially, the authors trained each model jointly on three clinical outcomes simultaneously: targeted molecular therapy response, radiation therapy response, and six-month survival prediction, under identical data and evaluation conditions. This is the first time pathway-structured deep learning has been treated as a combined therapy and survival prediction problem.

The results reveal that no single architecture dominates all tasks. PATH performed best overall for targeted molecular therapy prediction, BINN was most reliable for survival prediction, and surprisingly, none of the models produced useful predictions for radiation therapy—suggesting that key clinical variables not captured in gene expression data drive those decisions. Most strikingly, GraphPath achieved an AUROC of 0.92 on prostate targeted molecular therapy prediction, the highest score in the entire benchmark, despite an extreme class imbalance of only 11% positive prevalence. This demonstrates that the model's lateral co-regulation structure provides exceptional discriminative power when matched to a cohort with a narrow targetable driver programme.