Harvard's COMPASS AI Model Predicts Immunotherapy Response Across 33 Cancer Types

Harvard Medical School researchers published COMPASS, a pan-cancer AI foundation model that predicts which patients will respond to immune checkpoint inhibitors, in Nature Medicine. COMPASS uses a concept bottleneck transformer that encodes tumor gene-expression data into 44 biologically grounded immune concepts (immune cell states, tumor-microenvironment interactions, signaling pathways), and was pretrained on 10,184 tumors spanning 33 cancer types before fine-tuning on 16 clinical cohorts covering 1,133 ICI-treated patients across seven cancers and six checkpoint inhibitor regimens (anti-PD-1, anti-PD-L1, anti-CTLA-4, and combinations). It outperformed 22 existing prediction methods by roughly 8.5% in accuracy and 15.7% in area-under-precision-recall-curve on average, and generalizes to cancer types and drugs not seen during fine-tuning — a step toward AI-guided patient stratification for immunotherapy.