STANFORD, September 18 — Researchers at Stanford University have built two new artificial intelligence models of the biological cell, the second of which — TranscriptFormer — was trained on gene expression data from 112 million cells spanning 12 species, from single-celled yeast to humans. The work was published recently in the journals Nature and Science, according to a Stanford Medicine news article dated September 14, 2026, written by Helen Fields. The first model, universal cell embedding, paved the way for TranscriptFormer.
The co-senior authors are Stephen Quake, PhD, a professor of bioengineering at Stanford; Jure Leskovec, PhD, a professor of computer science; and Theo Karaletsos, PhD, of the Chan Zuckerberg Initiative.
A new way to think about cell biology
Quake framed the effort plainly. “We’re trying to open up a whole new way to think about cell biology,” he said in the news article, adding, “These papers mark the beginning of, we hope, a whole new field and a decade of work.” The genome contains the instructions for life, but many biologists now focus on gene expression — which genes a particular cell is using. A beta cell in the pancreas expresses the gene for insulin plus genes related to storage and release of the hormone.
B cells, a type of white blood cell, produce antibodies to fight disease. A skin cell might express genes to make hair or pigment.
Those differences in expression help tell cell types apart. Healthy and sick cells also have different gene expression patterns, as do cells from different species. Scientists have built atlases, databases that define cells and tissues based on gene expression patterns, generating staggering amounts of data on tens of thousands of genes for hundreds of millions of cells from many species.
“How do we think about all these genes at once? The human brain can’t,” Quake said.
“These models help us do that. The algorithms are ways for us to understand very complex data that humans can’t wrap our minds around.”
Twelve species, one mathematical space
TranscriptFormer was trained on 12 species: humans and yeast, plus 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. Training used gene expression profiles in a way analogous to large language models such as ChatGPT, which are trained by repeatedly looking at texts with some words missing and adjusting parameters until they can correctly predict the missing words. TranscriptFormer did the same with gene expression values instead of words.
Quake described the resulting model as a “universal space” — not a physical place but a mathematical space that can be used to measure how similar 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, neurons and what comes next
One use is exploring evolutionary relationships between species. Scientists who study sponges have wondered which kind of sponge cell is most closely related to a neuron in other animals. The model showed that a cell known as a choanocyte was quite similar to neurons in the roundworm and the frog, helping scientists understand the evolution of neurons and the role of choanocytes in sponges.
The sponge’s neuroid cell, named for having some neuron-like features, turned out to have a gene expression profile more like a frog gland, according to the article. For a field that has spent years accumulating gene expression data faster than anyone could interpret it, the models offer a way to compare cells across species and across health and disease.
Quake’s own framing points to the scale of what he expects to follow: a decade of work, and a field he hopes is only beginning.
