Interpretable AI Tool Improves Prediction of Immunotherapy Response

Immune checkpoint inhibitors are standard treatment options across many cancers, but only a subset of patients benefit, making patient selection difficult. Because predictive biomarkers remain limited, clinicians and trial teams often rely on trial-and-error approaches that can expose nonresponders to toxicity and treatment delays. Interpretable, biology-based tools could help laboratories better stratify likely responders using transcriptomic data. A new study now shows that an artificial intelligence approach may improve identification of patients most likely to benefit from immunotherapy.

Researchers at Harvard Medical School have developed COMPASS, an artificial intelligence (AI) model designed to predict which patients will respond to immune checkpoint inhibitors (ICIs). The system aims to improve patient selection while providing a rationale for each prediction. It emphasizes interpretability anchored in gene programs related to immune cell states, tumor-microenvironment interactions, and signaling pathways.

Image: Researchers have developed an AI model to predict which patients may respond to immune checkpoint inhibitors (Photo courtesy of Shutterstock)

COMPASS evaluates tumor gene activity across nearly 16,000 genes to generate human-interpretable outputs rather than opaque, black box scores. The model was trained on 10,184 tumors representing 33 cancer types from The Cancer Genome Atlas (TCGA), learning transcriptomic patterns associated with ICI response and nonresponse. By basing decisions on gene expression profiles, the tool links predicted outcomes to underlying biological processes that promote or impede antitumor immunity.

For model adaptation and testing, investigators fine-tuned COMPASS with results from 16 clinical trials spanning seven cancer types and a range of ICI regimens. Performance was assessed using a leave-one-trial-out strategy in which each trial was withheld in turn and the model predicted responders and nonresponders for the omitted cohort. Across these evaluations, COMPASS outperformed the strongest existing method for predicting ICI response by nearly 10% on average, with gains observed across tumor types, checkpoint drugs, gene transcript sequencing platforms, and biopsy sites.

The interpretable outputs also clarified outliers, such as nonresponders with immune‑inflamed tumors whose expression patterns indicated processes that suppress immune activity, and responders with immune‑desert tumors whose signatures suggested alternative immune engagement. Findings were published in Nature Medicine on July 3, 2026. If validated prospectively, the approach could assist clinical decision-making for ICI use, improve trial enrollment by identifying participants most likely to benefit, and generate hypotheses that point to new therapeutic targets.

“ICIs are an exciting therapeutic modality that has transformed cancer treatment over the past decade by engaging the immune system to fight cancer cells and destroy them. By leveraging cutting-edge AI capabilities, we can identify who would be most likely to respond to a particular ICI before that patient receives the drug,” said study senior author Marinka Zitnik, associate professor of biomedical informatics in the Blavatnik Institute at HMS.

Related Links
Harvard Medical School