Bringing Cancer Survival Modeling to Single-Cell Resolution
Researchers have developed a first-of-its-kind tool capable of linking the survival outcomes of patients with cancer to molecular features detected in individual tumor cells. The findings, published in Cancer Discovery, suggest that single-cell analyses may improve prognosis and help guide more personalized treatment.
Survival analysis is central to clinical oncology, yet existing approaches do not directly model survival outcomes from single-cell data, according to the researchers. To address this gap, they developed scSurvival, an attention-based multiple-instance learning framework designed to predict patient risk while identifying outcome-associated cell subpopulations.
“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 senior author Zheng Xia, PhD, of Oregon Health & Science University, Portland, in a news release from the National Institutes of Health (NIH).
How the Model Works
scSurvival was designed to model patient survival directly from cohort-level single-cell RNA sequencing data while preserving cellular resolution, Dr. Xia told ASCO AI in Oncology. Unlike earlier approaches that “lose intra-cell-type heterogeneity” by importing survival labels from bulk data, flattening cell-level gene expression into pseudobulk profiles, or using cell-type proportions as patient-level covariates, he noted, scSurvival treats each tumor as a collection of multiple cellular “instances,” or individual data points, to identify subpopulations most strongly associated with clinical outcomes.
scSurvival combines two core components: a zero-inflated Gaussian variational autoencoder (ZIG-VAE)–based cell feature extraction module and an attention-based multiple-instance Cox regression (AMICR) module. According to Dr. Xia, the ZIG-VAE generates stable, batch-corrected low-dimensional cell embeddings, “handling [the] intrinsic challenges of single-cell data,” whereas the AMICR module uses a multihead attention mechanism to aggregate cell-level features into a patient-level representation for survival modeling.
Diving deeper, Dr. Xia outlined the workflow in six steps for ASCO AI in Oncology:
Input: The model receives per-patient single-cell expression matrices together with survival event status and event time, with optional batch labels and clinical covariates.
Feature extraction: The VAE is pretrained on all cells from all patients to learn batch-corrected low-dimensional cell embeddings.
Aggregation: A multihead attention mechanism assigns each cell a weight reflecting its contribution to patient risk, and weighted cell embeddings are summed into a patient-level representation.
Scoring: A hazard-scoring network converts the patient-level representation into a relative hazard estimate for Cox partial likelihood training.
Fine-tuning: All three loss components (Cox loss, VAE reconstruction, and entropy regularization) are jointly optimized to maintain sparse, interpretable attention patterns.
Output: Patient-level risk and attention-adjusted hazard scores are generated for individual cells, enabling downstream identification of risk-associated cellular subpopulations.
According to Dr. Xia, “scSurvival is the first to learn directly from true single-cell survival data while simultaneously predicting patient outcomes and identifying the cells driving them,” bringing cancer survival analysis into what he described as the “cellular resolution era.”
Survival Signals at Cellular Resolution
Among the study’s key findings, Dr. Xia highlighted that scSurvival demonstrated the feasibility of survival modeling at single-cell resolution while preserving cellular heterogeneity—the first framework to do so. He also emphasized its scalability to real-world cohort sizes. scSurvival processed approximately 1 million cells across 100 patients in about 17.5 minutes on a single H100 graphics processing unit with linear runtime and memory scaling, “making it practical for the cohort-level single-cell RNA sequencing atlases [that are] now becoming standard.”
Simulation studies were conducted to test the robustness and performance of scSurvival across diverse data conditions. In the most challenging setting, the framework accurately identified risk-driving cells embedded within a large cluster lacking clear boundaries, achieving 90% precision, 97% recall, and an F1 performance score of 0.931.
In a melanoma cohort comprising 48 samples from 32 patients who were treated with immunotherapy, scSurvival independently identified higher-hazard SPP1-positive/CXCL9-negative vs lower-hazard SPP1-negative/CXCL9-positive macrophages and stress-response HSPA1A/B-positive T-cell states associated with immunotherapy resistance, which was consistent with prior studies. Spatial transcriptomic analyses further showed that lower-hazard regions co-localized with tertiary lymphoid structures, Dr. Xia noted.
The researchers also applied scSurvival to a single-cell RNA sequencing atlas comprising approximately 1.09 million cells from 121 patients with liver cancer and available survival data. Higher-hazard tumor cells were enriched for epithelial-mesenchymal transition, hypoxia, and transforming growth factor-β signaling programs, Dr. Xia observed, whereas lower-hazard tumor cells retained hepatocyte-like metabolic features. From these analyses, the researchers derived a 200-gene single-cell liver cancer unfavorable survival signature that stratified overall and progression-free survival across multiple independent bulk cohorts, including the liver hepatocellular carcinoma data set from The Cancer Genome Atlas.
Considering these data, Anthony Letai, MD, PhD, Director of the NIH’s National Cancer Institute (NCI), stated, “A risk assessment tool that not only tells you who may be at higher risk but also provides clues as to why could really help in these difficult cancers.”
“We believe that scSurvival will accelerate the broad adoption of cohort-level single-cell profiling for survival analysis and deepen our understanding of how specific cell subpopulations in the tumor microenvironment influence cancer outcomes, positioning it as a valuable tool for translational cancer research,” the researchers concluded.
DISCLOSURES: The study was funded in part by the NIH. For full disclosures of the study authors, further funding information, and data and code availability, visit aacrjournals.org.
ASCO AI in Oncology is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (ASCO®). The ideas and opinions expressed in ASCO AI in Oncology do not necessarily reflect those of Conexiant or ASCO. For more information, see Policies.