Ontology-Guided Pathway Activity Identifies a Cell-Intrinsic Defense Response Program Associated with MEK Inhibitor Sensitivity
Predicting cancer drug response from gene expression requires models that expose which biological pathways drive cell-line-specific sensitivity. We introduce Gene-Ontology Pathway Attention (GOPA), whose core module computes deterministic attention weights softmax(xc {middle dot}[B] ) from expression and the column-normalized gene-term annotation matrix with no learned parameters. Applying GOPA to 542 drugs across the GDSC panel under leave-cell-line-out evaluation, we find that defense response pathways predict sensitivity to kinase inhibitors in GDSC, with the strongest and most confound-resistant signal in MEK/MAPK inhibitors. This association shows cross-assay support in PRISM (11 of 11 overlapping drugs; all p < 0.002) and retains 72% signal strength after controlling for five confounds, but was not reproduced in the gCSI panel (different response metric, smaller sample, MEK inhibitors absent), indicating the finding may be MAPK-pathway-specific and assay-dependent. On the 16-drug benchmark, XGBoost achieves the lowest RMSE (1.185); GOPA (1.226) is the strongest neural model. On the full 542-drug panel, target-encoded XGBoost matches GOPA on RMSE (1.322 vs. 1.327); GOPA achieves higher residual Pearson correlation (0.469; 95% CI [0.453, 0.485]) than target-encoded XGBoost (0.391; [0.379, 0.404]); paired difference +0.078 [0.062, 0.093]; GOPA wins on 363 of 539 drugs (67.3%); Wilcoxon p = 1.1 x 10-23). When pathway representations are evaluated with matched downstream learners, simple gene-set projections achieve equivalent prediction, indicating that GOPAs value lies in its deterministic, population-comparable pathway summaries rather than representational superiority. A controlled geometry comparison shows Poincare-ball embeddings preserve GO graph distances better than Euclidean ({rho} = 0.732 vs. 0.474) while Euclidean embeddings achieve stronger ancestor retrieval; neither geometry improves prediction ({Delta}RMSE = +0.008; p = 0.18).