Generalizable AI predicts immunotherapy outcomes across cancers and treatments
9 affiliations
- 1College of Pharmaceutical Sciences, Zhejiang University, Hangzhou, China
- 2Department of Biomedical Informatics, Harvard Medical School, Boston, MA, USA
- 3Division of Immunology, Boston Children's Hospital, Harvard Medical School, Boston, MA, USA
- 4Roche Pharma Research and Early Development, Oncology Early Clinical Development, Roche Innovation Center Basel, F. Hoffmann-La Roche Ltd., Basel, Switzerland
- 5Computational Sciences Center of Excellence, F. Hoffmann-La Roche Ltd., Basel, Switzerland. [email protected]
- 6Broad Institute of MIT and Harvard, Cambridge, MA, USA. [email protected]
- 7Department of Biomedical Informatics, Harvard Medical School, Boston, MA, USA. [email protected]
- 8Harvard Data Science Initiative, Cambridge, MA, USA. [email protected]
- 9Kempner Institute for the Study of Natural and Artificial Intelligence, Harvard University, Allston, MA, USA. [email protected]
Abstract
Immune checkpoint inhibitors are standard across cancers, yet most patients do not respond and existing biomarkers generalize poorly across tumor types, drugs and clinical settings. We present C ompass , a pan-cancer foundation model that predicts immunotherapy response from bulk tumor transcriptomes using a concept-bottleneck transformer. C ompass encodes gene expression through 44 biologically grounded immune concepts representing immune cell states, tumor-microenvironment interactions, and signaling pathways. Trained on 10,184 tumors across 33 cancer types, C ompass outperforms 22 baseline methods in 16 independent clinical cohorts spanning seven cancers and six immune checkpoint inhibitors, increasing accuracy by 8.5% and area under the precision-recall curve by 15.7%, with minimal additional training. The model generalizes to unseen cancer types and treatments, supporting indication selection and patient stratification in early-phase clinical trials. In survival analyses, C ompass -stratified responders have longer overall survival (hazard ratio = 4.7, p < 0.0001). Personalized response maps connect gene expression to immune concepts, revealing mechanisms of response and resistance; in immune-inflamed non-responders, C ompass highlights programs including TGF- β signaling, endothelial exclusion, CD4+ T cell dysfunction, and B cell deficiency. By combining interpretability with transfer learning, C ompass enables robust prediction and mechanistic insight to inform trial design and translational studies.
Related
- What Predicts Psychotherapy Outcomes Across Disorders and Treatments: an Umbrella Review
- Metabolic signature of phosphoethanolamine and arginine pathways predicts prognosis and immunotherapy resistance in lung adenocarcinoma
- Immune Cell Engagers and “Masking Immunotherapy” for Solid Tumors: Focus on Head and Neck Cancers
- Functional single-cell omics in gynecological cancers immunotherapy: from heterogeneity dissection to clinical translation
- Spatial architecture of the tumor–stroma interface and transcriptional programs shape immunotherapy response in MSI-H gastrointestinal cancers
BibTeX
@article{shen2026generalizable,
title = {{Generalizable AI predicts immunotherapy outcomes across cancers and treatments}},
author = {Wan Xiang Shen and Intae Moon and Thinh H. Nguyen and Michelle M. Li and Yepeng Huang and Nitya Nair and Daniel Marbach and Marinka Žitnik},
journal = {Nature Medicine},
year = {2026},
month = jul,
doi = {10.1038/s41591-026-04502-7},
url = {https://doi.org/10.1038/s41591-026-04502-7},
note = {CC BY-NC-ND},
}
RIS
TY - JOUR
TI - Generalizable AI predicts immunotherapy outcomes across cancers and treatments
AU - Wan Xiang Shen
AU - Intae Moon
AU - Thinh H. Nguyen
AU - Michelle M. Li
AU - Yepeng Huang
AU - Nitya Nair
AU - Daniel Marbach
AU - Marinka Žitnik
PY - 2026
DA - 2026/07/03
JO - Nature Medicine
DO - 10.1038/s41591-026-04502-7
UR - https://doi.org/10.1038/s41591-026-04502-7
AB - Immune checkpoint inhibitors are standard across cancers, yet most patients do not respond and existing biomarkers generalize poorly across tumor types, drugs and clinical settings. We present C ompass , a pan-cancer foundation model that predicts immunotherapy response from bulk tumor transcriptomes using a concept-bottleneck transformer. C ompass encodes gene expression through 44 biologically grounded immune concepts representing immune cell states, tumor-microenvironment interactions, and signaling pathways. Trained on 10,184 tumors across 33 cancer types, C ompass outperforms 22 baseline methods in 16 independent clinical cohorts spanning seven cancers and six immune checkpoint inhibitors, increasing accuracy by 8.5% and area under the precision-recall curve by 15.7%, with minimal additional training. The model generalizes to unseen cancer types and treatments, supporting indication selection and patient stratification in early-phase clinical trials. In survival analyses, C ompass -stratified responders have longer overall survival (hazard ratio = 4.7, p < 0.0001). Personalized response maps connect gene expression to immune concepts, revealing mechanisms of response and resistance; in immune-inflamed non-responders, C ompass highlights programs including TGF- β signaling, endothelial exclusion, CD4+ T cell dysfunction, and B cell deficiency. By combining interpretability with transfer learning, C ompass enables robust prediction and mechanistic insight to inform trial design and translational studies.
SN - 1078-8956
SN - 1546-170X
SN - 1744-7933
ER -