Researchers have developed COMPASS, a pan-cancer AI foundation model that predicts immunotherapy outcomes across diverse cancer types and treatments with significantly higher accuracy than existing methods.
Published in Nature Medicine, COMPASS uses a concept bottleneck transformer to analyze bulk tumor transcriptomes through 44 biologically grounded immune concepts. These include immune cell states, tumor-microenvironment interactions, and signaling pathways.
Trained on 10,184 tumors across 33 cancer types, the model outperformed 22 existing methods across 16 clinical cohorts spanning seven cancers and six different immune checkpoint inhibitors. It improved accuracy by 8.5% and area under the precision-recall curve by 15.7% on average.
Critically, COMPASS generalized to cancer types and treatments it was not fine-tuned on, suggesting it could inform indication selection and patient stratification in real-world settings. In survival analyses, patients classified as responders had significantly longer overall survival (hazard ratio 4.7).
The model also generates personalized response maps connecting gene expression to immune concepts, identifying resistance programs such as TGF-beta signaling and CD4+ T cell dysfunction in immune-inflamed non-responders. This mechanistic insight could guide clinical trial design and translational studies.
