TRAPS: Treatment-Assignment Prediction via Pathway-informed Stratification

Cancer treatment planning requires decisions across multiple clinical dimensions at once. Clinicians must determine whether a patient should receive targeted molecular therapy, radiation therapy, and whether they are likely to survive beyond six months. Existing pathway-informed deep learning models have been developed and tested in isolation, so it is unclear how they compare on a common footing. We present a harmonized benchmark for pathway-guided prediction of treatment assignment and short-term survival, evaluating three biologically informed architectures, BINN, GraphPath, and PATH, across five cancer cohorts drawn from The Cancer Genome Atlas, representing 2,622 patients encoded using Reactome pathway activity scores. We emphasise that the labels reflect the treatment a patient was recorded as receiving in TCGA, i.e. treatment exposure, rather than a measured response to therapy. Each model is trained jointly on all three clinical outcomes over a shared Reactome pathway representation, with each architecture retaining its original reference training protocol; to our knowledge this is the first study to treat pathway-structured deep learning as a combined treatment-assignment and short-term survival prediction problem. Our results show that, once every architecture is evaluated on identical stratified folds with matched uncertainty quantification (five repeated splits and paired-bootstrap tests), most inter-model differences fall within their 95% confidence intervals: the architecture rankings reported when these models are studied in isolation are largely not statistically resolved. The one consistent and significant effect is that the sparse-hierarchy model BINN is the strongest survival (OS) predictor, significantly outperforming both graph models on breast-cancer OS (ΔAUROC up to 0.14, 𝑝 ≤0.01) and leading on lung and prostate OS. For treatment-assignment prediction, targeted molecular therapy is best discriminated in the prostate cohort (AUROC ≈ 0.80 for all three models), yet no architecture significantly outperforms the others on any TMT cohort; radiation-therapy assignment is weakly predicted throughout, consistent with the hypothesis that its drivers are largely clinical rather than transcriptomic. Under a fair unified protocol, then, the choice among pathway-informed architectures matters far less than prior isolated results suggest, with short-term survival prediction the clearest point of differentiation.