NIH-funded AI model predicts cancer survival from single-cell tumor data
Research news: NIH-funded investigators report that scSurvival, a machine-learning model designed for single-cell tumor datasets, predicted survival outcomes and highlighted cell populations associated with risk. In testing on clinical single-cell data from >150 patients (including melanoma and liver cancer), the model reportedly outperformed traditional approaches that average signals across tumors or cell types. The study also traced predictions to immune and tumor cell groups, including melanoma populations associated with immunotherapy response. The work is not a regulatory approval.
NIH-funded AI model predicts cancer survival from single-cell tumor data (research news)
An NIH-funded team reports development of scSurvival, a machine-learning framework intended to predict cancer survival outcomes from single-cell gene expression data while retaining cellular-level detail (Ren et al., Cancer Discovery, 2026; DOI: 10.1158/2159-8290.CD-25-0965).
According to the report, Oregon Health & Science University (OHSU) evaluated the model using clinical single-cell datasets, including data from more than 150 patients, and found that scSurvival predicted survival outcomes and linked specific tumor cell populations to higher risk.
How the model works
Rather than averaging expression across entire tumors or broad cell types, scSurvival reportedly:
- Assigns each cell a weight based on how strongly that cell is related to survival,
- Filters less-informative cells, and
- Aggregates weighted signals to generate survival predictions.
Reported performance and biological interpretability
The authors state the model was trained on single-cell datasets paired with survival outcomes from hundreds of patients, then tested in melanoma and liver cancer, where it predicted outcomes more accurately than traditional methods.
They also report tracing predictions to specific immune and tumor cell groups associated with better or worse survival; in melanoma, scSurvival highlighted cell populations associated with immunotherapy response.
This work was supported in part by NCI grants R01CA283171, U01CA253472, U01CA281902, and U24CA264128.