Abstract / Summary
Abstract Background Cancer prognosis models trained on transcriptomic data face two interconnected problems. Unconstrained deep survival models ignore biological structure, and pathway-constrained graph neural networks assume a patient-invariant interaction topology that is difficult to reconcile with the malignancy-dependent rewiring of tumor regulatory networks. Methods We propose Path-AGNN-Cox, a pathway-constrained adaptive graph neural network for survival prediction. Genes are partitioned into KEGG cancer-core pathway modules; within-pathway attention weights are modulated by a learnable malignancy gate; and a Cox partial-likelihood objective with dual regularization is optimized on prognostic risk. The model was benchmarked against seven survival baselines across 11 TCGA cancer types and validated on 25 independent GEO cohorts. Results On internal cross-validation, Path-AGNN-Cox reached a mean C-index of 0.56 (SD 0.04), significantly below the strongest penalized Cox baseline (Ridge-Cox, P = 0.005) and the deep survival networks; on 25 external GEO cohorts its mean C-index (0.51) was comparable to the deep baselines, though close to chance; no discrimination gain is claimed. Its distinguishing output is the interpretable rewiring: cohort-level pathway-weight tests exceeded permutation nulls in all three cohorts, correlated with clinical indicators of malignancy, and were absent by construction in static pathway models. In LUAD the risk score itself was near chance, so the LUAD between-stratum differences are interpreted as attention variation associated with the model's own stratification rather than with survival. A standard-GAT control detected cohort-dependent reductions in significant pathways, and matched random gene-set controls indicated cohort-dependent selectivity with no signal beyond the matched null in LUAD. Three-seed out-of-fold and platform-matched sensitivity analyses further showed no pathway that was significant in every seed and near-chance portable discrimination, so all rewiring findings remain cohort-specific and hypothesis-generating. Conclusion Path-AGNN-Cox provides a reproducible framework in which patient-specific pathway graphs become objects of formal statistical testing. The model and pipelines are released as an open-source Python package with an archived snapshot.