AI Model Predicts Cancer Immunotherapy Response Across Tumor Types
Harvard's COMPASS model forecasts which patients will respond to immune checkpoint inhibitors with striking accuracy across seven cancer types.
Summary
A new AI model called COMPASS, developed at Harvard Medical School, can predict whether a cancer patient will respond to immune checkpoint inhibitor therapy by analyzing the gene expression profile of their tumor. Trained on over 10,000 tumors spanning 33 cancer types, it outperformed 22 existing methods across 16 clinical cohorts. Patients the model classified as likely responders had dramatically better survival outcomes, with a hazard ratio of 4.7. Beyond prediction, COMPASS generates personalized maps linking gene activity to immune system behavior, revealing resistance mechanisms like TGF-beta signaling and T cell dysfunction. This could transform how oncologists select treatments and design clinical trials.
Detailed Summary
Immune checkpoint inhibitors have revolutionized cancer treatment, but a persistent challenge remains: most patients do not respond, and there is no reliable way to predict who will benefit before treatment begins. Existing biomarkers like PD-L1 expression or tumor mutational burden perform inconsistently across different cancer types and drug combinations, leaving clinicians without a dependable decision-making tool.
Researchers at Harvard Medical School developed COMPASS, a pan-cancer foundation model that predicts immunotherapy response using bulk tumor gene expression data. The model uses a concept bottleneck transformer architecture, encoding gene expression through 44 biologically grounded immune concepts that represent immune cell states, tumor microenvironment interactions, and signaling pathways. It was trained on 10,184 tumors across 33 cancer types.
COMPASS was benchmarked against 22 competing methods across 16 clinical cohorts covering seven cancer types and six different immune checkpoint inhibitors. It improved prediction accuracy by 8.5% and area under the precision-recall curve by 15.7% on average. Critically, it generalized to cancer types and treatments not seen during fine-tuning. In survival analyses, patients COMPASS classified as responders had a hazard ratio of 4.7 for overall survival — a clinically meaningful separation.
The model also generates personalized response maps that connect individual gene expression patterns to immune concepts. In patients who had inflamed tumors but still did not respond, COMPASS identified resistance programs including TGF-beta signaling, endothelial cell exclusion of immune cells, CD4+ T cell dysfunction, and B cell deficiency — actionable mechanistic hypotheses for future therapeutic targeting.
Caveats include the fact that this summary is based on the abstract only and the full methodology requires review. Industry co-authors from Roche represent a potential conflict of interest. Real-world clinical validation in prospective trials will be essential before COMPASS influences treatment decisions.
Key Findings
- COMPASS outperformed 22 existing immunotherapy prediction methods across 16 clinical cohorts spanning 7 cancers.
- Patients classified as responders had 4.7x better overall survival odds (P < 0.0001) compared to predicted non-responders.
- Model improved prediction accuracy by 8.5% and precision-recall AUC by 15.7% over existing approaches.
- COMPASS generalized to cancer types and therapies not present during model fine-tuning.
- Personalized immune maps identified TGF-beta signaling and T cell dysfunction as key resistance mechanisms.
Methodology
COMPASS is a concept bottleneck transformer trained on 10,184 tumor transcriptomes across 33 cancer types, encoding gene expression via 44 immune concept features. It was evaluated across 16 independent clinical cohorts covering seven cancer types and six immune checkpoint inhibitor drugs. Survival analyses used hazard ratios to compare outcomes between predicted responders and non-responders.
Study Limitations
This summary is based on the abstract only, as the full paper is not open access; methodology and validation details require independent review. Two co-authors are employed by F. Hoffmann-La Roche Ltd., representing a potential industry conflict of interest. Prospective clinical validation is needed before COMPASS can be integrated into routine oncology practice.
Enjoyed this summary?
Get the latest longevity research delivered to your inbox every week.
Enter your email to subscribe: