Accurate survival prediction in NSCLC remains a major clinical challenge due to the complexity of multi-omics and clinical data; integrating prior biological knowledge can enhance model interpretability and clinical utility. 

We introduce a machine learning framework that incorporates prior knowledge via Knowledge Graphs, improving survival prediction for NSCLC patients, particularly those receiving immuno-oncology therapies. 

Knowledge graph–based models demonstrated superior hazard ratio performance compared to traditional biomarker models in POPLAR and OAK trials, with a 10-gene mutational signature showing significant survival stratification across both studies.