Oregon Health & Science University researchers have developed a first-of-its-kind method to predict cancer patient survival using advanced molecular data from individual cells.

Survival analysis is central to clinical oncology. Modern cancer studies can now measure gene activity in single cells from a patient’s tumor and link this information to how long patients live. However, until now, there has not been a good way to use this detailed cell-level data to directly predict survival.

The study, published today in Cancer Discovery, describes a method called scSurvival that uses single-cell genetic data to identify which cells inside a tumor are most strongly linked to patient survival. Unlike traditional methods that average signals across an entire tumor, the new approach pinpoints harmful and helpful cell populations that can drive disease progression. The research team presented these findings today at the American Association for Cancer Research conference.

Co-lead author Faming Zhao, Ph.D., said the approach helps solve a long-standing problem in cancer research.

“Tumors are very complex, and important signals can be lost when data are averaged across thousands or millions of cells,” said Zhao, a postdoctoral fellow specializing in cancer biology in the OHSU School of Medicine. “By looking at survival at single-cell resolution, we can better understand why patients with the same cancer can have very different outcomes.”

The framework of scSurvival

scSurvival is a new, scalable, and interpretable tool for modeling survival outcomes from single-cell cohort data, with cellular-resolution risk profiling. It first employs a feature extraction module based on a variational autoencoder and generative modeling to learn batch-invariant single-cell representations, and then aggregates cell-level features to the patient level to perform multi-head attention-based multiple instance Cox regression. scSurvival not only enables the integration of single-cell expression data with patient-level clinical variables to build accurate survival risk prediction models, but also identifies key cell subpopulations most associated with survival risk and characterizes their risk tendencies, thereby facilitating more refined downstream analyses. 

Expertise across fields

In tests using melanoma and liver cancer data, the tool more accurately predicted patient outcomes than standard methods. It also uncovered specific immune and tumor cell states tied to better or worse survival. For example, the researchers found certain immune cells that appear to help patients respond better to immunotherapy, while others were linked to poorer outcomes.

“This study was made possible by strong collaboration at the Knight Cancer Institute between computational scientists, cancer biologists and clinicians,” Xia said. “By bringing together expertise from different fields, we were able to use artificial intelligence to develop a new way to study survival using single-cell data.”

Xia said the model goes beyond traditional machine learning approaches by capturing complex biological patterns that were previously difficult to study.

“This work uses artificial intelligence to develop a new way to study survival using single-cell data,” Xia said. “The model is more complex than traditional machine learning approaches, and it allows us to capture information that was not accessible before.”

Understanding these differences matters for patients, researchers said, because tumors are made up of many cell types that behave differently. Treatments that work for one patient may fail in another if harmful cell populations are missed.

While scSurvival is not yet used in clinical care, the researchers say it could eventually help doctors better identify high-risk patients and support the development of more precise, targeted cancer therapies.

Availability – The open-source scSurvival program and its tutorials are freely available at GitHub, Zenodo and Code Ocean.

Source – Oregon Health & Science University