Benchmarking single-cell foundation models for aging biology

Inside every living thing, from a tiny moss to a human, there are trillions of tiny units called cells. Each cell is a specialized worker, and while they all carry the same set of instructions, they read different parts of those instructions to become a heart cell, a skin cell, or a brain cell. Scientists have long wanted to understand how these cells change as an organism gets older, a process we call aging. To do this, they look at the activity of genes, the chemical switches that tell a cell what to do. In recent years, computers have become powerful enough to read the activity of millions of these genes at once, creating a massive map of cellular life. However, making sense of this overwhelming amount of data requires a new kind of tool: a computer program trained to recognize patterns in cell behavior, much like a librarian who has read every book in a vast library and can instantly tell you which books belong together.

A team of researchers set out to test a new generation of these computer programs, known as single-cell foundation models, to see if they could help answer difficult questions about aging. They gathered a massive collection of data, more than 2.5 million snapshots of individual cells from various studies, to create a rigorous test. The goal was to see which computer program could best translate raw genetic data into useful insights about how cells age, how they change during disease, and how they differ from one another. The researchers did not just look at one type of question; they tested the programs on five different biological puzzles, ranging from predicting how old a cell is to finding rare types of cells that might be hiding in the data.

The results showed that no single computer program was perfect for every task, but several stood out for specific jobs. When the goal was to predict the chronological age of a cell or to track its progress through time, one program called Geneformer performed better than the others. It was so effective that it could often match the performance of methods that required scientists to manually select the most important genes, a task that usually takes significant human effort. However, the study also found that for some tasks, a simpler approach using just a few thousand key genes still worked better than the complex computer models, suggesting that these new tools are powerful but not yet a replacement for all traditional methods.

The researchers also looked at how well these programs could spot changes in cells caused by disease. Several of the models successfully identified molecular shifts that matched what scientists already knew about aging in the context of illness, confirming that these tools can recognize real biological patterns. One program, called SCimilarity, proved particularly skilled at finding rare cellular states, such as unusual cell types that appear in very small numbers. It outperformed both the models built specifically for aging and the standard methods used in the past, showing that these general-purpose tools can be surprisingly good at spotting the unusual. At the level of individual genes, another program named scGPT did the best job of reconstructing the known relationships between genes, effectively mapping out the regulatory hubs that control aging.

Ultimately, the study demonstrates that these advanced computer models are becoming useful tools for aging research, but their value depends entirely on the specific question being asked. They are not a magic key that solves every problem, but they do offer a new way to identify rare cell types and understand the complex rules that govern how cells change over time. By testing these tools against a massive, shared dataset, the researchers provided a clear guide for other scientists on which program to use for which job, helping to ensure that future discoveries in aging biology are built on the most reliable foundations available.