Researchers have developed COMPASS, a pan-cancer foundation model that predicts immunotherapy response from tumor transcriptomes with greater accuracy than existing methods. Published in Nature Medicine, the model was trained on 10,184 tumors across 33 cancer types.
COMPASS uses a concept bottleneck transformer to encode gene expression through 44 biologically grounded immune concepts, including immune cell states, tumor-microenvironment interactions, and signaling pathways. Across 16 clinical cohorts spanning seven cancers and six immune checkpoint inhibitors, it improved accuracy by 8.5% and area under the precision-recall curve by 15.7% on average.
The model also generalized to cancer types and treatments not represented during fine-tuning. In survival analyses, patients classified as responders had significantly longer overall survival (hazard ratio 4.7).
COMPASS generates personalized response maps that connect gene expression to immune concepts, identifying programs associated with resistance — including TGF-beta signaling and CD4+ T cell dysfunction — offering mechanistic insight for trial design.
