Generalizable AI predicts immunotherapy outcomes across cancers and treatments.

Immune checkpoint inhibitors have become standard care across many cancers, but most patients do not respond. Predicting response remains challenging due to complex tumor-immune interactions and the poor generalizability of current biomarkers and models. Predictors such as tumor mutational burden, PD-L1 expression, and transcriptomic signatures often fail across cancer types, therapies, and clinical settings. There is a clear need for a robust, interpretable model that captures shared immune response principles and adapts to diverse clinical contexts. We present Compass, a foundation model for predicting immunotherapy response from pan-cancer transcriptomic data using a concept bottleneck architecture. Compass encodes tumor gene expression through 44 biologically grounded immune concepts representing immune cell states, tumor-microenvironment interactions, and signaling pathways. Trained on 10,184 tumors across 33 cancer types, Compass outperforms 22 baseline methods in 16 independent clinical cohorts spanning seven cancers and six immune checkpoint inhibitors, increasing precision by 8.5%, Matthews correlation coefficient by 12.3%, and area under the precision-recall curve by 15.7%, with minimal or no additional training. The model generalizes to unseen cancer types and treatments, supporting indication selection and patient stratification in early-phase clinical trials. Survival analysis shows that Compass-stratified responders have significantly longer overall survival (hazard ratio = 4.7,