Cancer Discovery
AI predicts cancer survival from single-cell tumor data
Clinical takeaway: Single-cell risk models may eventually help identify high-risk patients and the tumor or immune cell states driving that risk, but the approach still needs broader validation across cancers and datasets before it can be used routinely.
Survival models in oncology usually rely on bulk tumor data, which can obscure rare but important cell populations. This study tested whether a new tool, scSurvival, could work directly from single-cell RNA sequencing data, preserving cellular heterogeneity while also linking specific cell states to patient outcomes.
The findings suggest that tumor behavior may be shaped by small, but clinically important, cell subsets that are missed when tumor data are averaged across whole samples. The model not only predicted survival but also traced that risk back to biologically meaningful cell populations.
Investigators developed an attention-based machine learning model that weights individual cells by how strongly they relate to survival, then applied it to single-cell cancer cohorts with linked clinical outcomes. The main real-world analyses included a melanoma immunotherapy dataset with 32 patients and a liver cancer cohort with survival data from 121 patients.
In melanoma, lower-risk patterns were linked to B cells and plasma cells, while higher-risk myeloid cells showed a more protumor profile marked by SPP1 and reduced CXCL9. In T-cell analyses, lower-risk cells were enriched for stem-like and memory-associated programs, while higher-risk cells showed exhaustion and stress-response features associated with poorer immunotherapy response.
While in liver cancer, the model identified higher-risk tumor cells with hypoxia, epithelial-mesenchymal transition, and invasive signaling, while lower-risk tumor cells showed a more differentiated, metabolically active hepatocyte-like profile. These findings help explain why tumors with similar broad diagnoses can behave very differently.
At the patient level, the model significantly stratified survival in both cancers and outperformed comparison approaches based on cell-type fractions or pseudobulk expression in the reported analyses.
“By taking a fine-tooth comb to single-cell data, scSurvival is able to consider the varying influence that individual cells have on disease progression and survival outcomes,” said corresponding author Zheng Xia, PhD, associate professor of biomedical engineering at Oregon Health & Science University.
Source: Ren T. Cancer Discovery. 2026 Apr 21. scSurvival: Single-Cell Survival Analysis of Clinical Cancer Cohort Data at Cellular Resolution