“The human brain can’t.” That was Stephen Quake’s answer to a question he put himself: how do you think about tens of thousands of genes across hundreds of millions of cells at once? Quake, a professor of bioengineering at Stanford University, and his colleagues built AI models to do it for us.
Stanford Medicine described the work in a news article dated September 14, 2026. The team made two models of the biological cell. The first, universal cell embedding, led to a second-generation model called TranscriptFormer, trained on data from 112 million cells from 12 species, from single-celled yeast to humans. Two papers on the work appeared in Nature and Science. Quake shares senior authorship with Jure Leskovec, a professor of computer science, and Theo Karaletsos of the Chan Zuckerberg Initiative.
“We’re trying to open up a whole new way to think about cell biology,” Quake said. “These papers mark the beginning of, we hope, a whole new field and a decade of work.”
Reading genes the way a language model reads words
Many biologists now look at gene expression, meaning which genes a given cell is using. A pancreatic beta cell expresses the insulin gene plus genes for storing and releasing the hormone. A B cell, a white blood cell, makes antibodies. A skin cell might switch on genes for hair or pigment. Healthy and sick cells differ in their patterns, and so do cells from different species. Atlases that catalog these patterns have produced more data than anyone can take in.
TranscriptFormer learned from that data the way a large language model such as ChatGPT learns from text. A language model repeatedly sees passages with words missing and adjusts its parameters until it predicts them correctly. TranscriptFormer did the same with gene expression values in place of words.
Besides humans and yeast, the atlases covered mice, rabbits, chickens, zebrafish, fruit flies, the African clawed frog, the parasite that causes malaria, a sea urchin, a sponge and the roundworm Caenorhabditis elegans.
Quake calls the result a “universal space,” a mathematical space for measuring how alike or different things are. “All the cells of every organism on Earth can sit in that space, and you can look at their relationship to each other,” he said.
Sponges, new species and disease
Researchers who study sponges have long asked which sponge cell is closest to a neuron in other animals. The model found that the choanocyte looked quite similar to neurons in the roundworm and the frog. The neuroid cell, named for some neuron-like features, had an expression profile closer to a frog gland, which suggests it may be involved in digestion.
Given expression data from a species it never trained on, TranscriptFormer can identify the cell types. The team also showed it can separate healthy cells from diseased ones. Stanford Medicine said the model might one day help design cell-based treatments, because it can imagine functional cells that do not exist now.
“This is a tool to help skilled biologists think about complex problems,” Quake said.
Researchers from the Chan Zuckerberg Initiative and BioHub contributed. The co-first authors of the Nature paper are Yanay Rosen, a Stanford doctoral student in computer science, and Yusuf Roohani, now an associate director of machine learning at the Arc Institute. Funders include the Chan Zuckerberg Initiative, the Defense Advanced Research Projects Agency and the National Science Foundation.
