An interpretable AI model predicts immunotherapy outcomes across multiple cancer types

While immune checkpoint inhibitors—cancer treatments designed to help the immune system recognize and attack tumors—have transformed oncology, the majority of patients still do not experience durable responses. Furthermore, standard clinical biomarkers like PD-L1 protein expression or tumor mutational burden often fall short in reliably identifying who will benefit across different tumor types. To help close this gap, researchers have developed an artificial intelligence platform called COMPASS that predicts patient responses to immunotherapy by translating complex tumor gene expression data into interpretable biological concepts, scientists report today in Nature Medicine. The work, led by senior author Marinka Zitnik at Harvard Medical School and Harvard University, with first author Wanxiang Shen as lead researcher, demonstrates a transparent machine learning approach capable of generalizing across diverse cancer types and drug regimens.

To build the model, the research team trained COMPASS using bulk RNA sequencing data from over 10,000 tumors across 33 cancer types. Rather than making direct predictions from raw genomic data using a "black box" approach, the system routes a patient's gene expression profile through 44 biologically grounded concepts, such as cytotoxic T cell activity, interferon-gamma signaling, and specific tumor microenvironment interactions. When evaluated on 1,133 patients across 16 independent clinical cohorts spanning seven cancer types and six different checkpoint inhibitor therapies, COMPASS outperformed 22 existing predictive methods. On average across the tested cohorts, the model improved prediction accuracy by 8.5% and area under the precision-recall curve by 15.7% compared to alternative approaches.

In survival analyses of a held-out clinical trial involving patients with metastatic urothelial carcinoma, individuals classified by the model as responders demonstrated significantly longer overall survival, achieving a one-year survival rate of 86% compared to 40% for predicted non-responders. Beyond predicting clinical outcomes, the platform generates personalized response maps that link a patient's unique gene expression to immune activation or failure. Notably, the model revealed why certain patients with "immune-inflamed" tumors—who would conventionally be expected to respond to checkpoint inhibitors—still experience treatment resistance. In these cases, COMPASS highlighted alternative immunological roadblocks, including active TGFβ pathway signaling, blood vessel remodeling that physically blocks immune cell infiltration, and B cell deficiency.

Despite its strong predictive performance across datasets, the study's authors highlight several important limitations. Because COMPASS relies on bulk RNA sequencing, the data averages molecular signals across an entire tissue sample, which can obscure critical interactions from rare immune cell populations. Additionally, due to incomplete or heterogeneous records across the historical clinical cohorts analyzed, the model could not uniformly adjust for clinical variables like patient age or tumor stage. The analyzed datasets also lacked non-immunotherapy control arms, meaning the model's predictions may reflect a combination of general survival advantages and treatment-specific responses. Consequently, the authors emphasize that COMPASS is currently an exploratory and hypothesis-generating tool that requires prospective validation in controlled clinical trials before it can be deployed in patient care.

As artificial intelligence increasingly intersects with precision medicine, tools like COMPASS point toward a future of more transparent, biomarker-driven trial design. Future work could integrate single-cell and spatial transcriptomic data into the framework to further refine how investigational cancer therapies are tailored to individual patients.

Reference:
Shen, W., Moon, I., Nguyen, T.H. et al. Generalizable AI predicts immunotherapy outcomes across cancers and treatments. Nat Med (2026). https://doi.org/10.1038/s41591-026-04502-7