Every living thing is built from cells, yet biology has long lacked a common coordinate system for comparing them. A cell studied in a mouse liver and one studied in a human tumour are usually described in separate datasets, each with its own labels and quirks, which makes direct comparison hard. A new foundation model aims to change that by giving researchers a single map on which any cell can be placed.
Described in Nature and developed by a team at Stanford University together with the Chan Zuckerberg Biohub and Chan Zuckerberg Initiative — among them Yanay Rosen, Yusuf Roohani, Stephen Quake and Jure Leskovec — the Universal Cell Embedding (UCE) learns a shared "latent space" for cells. Trained by self-supervision on large amounts of single-cell gene-expression data, without any manual labels, it captures genuine biological signal even when experiments are noisy or gathered in different laboratories.
To demonstrate the approach, the researchers assembled an Integrated Mega-scale Atlas of 36 million cells spanning more than 1,000 named cell types, dozens of tissues and eight species: human, mouse, zebrafish, mouse lemur, crab-eating macaque, rhesus macaque, tropical clawed frog and pig. Most of the data — more than 33 million cells — came from the open CZ CELLxGENE repository.
Why a universal map helps
UCE's central advantage is that it needs no retraining to handle new data. Existing tools usually require researchers to relabel cells and tune the model for every fresh experiment, which is slow and resource-intensive. Because UCE represents genes through the ESM2 protein language model rather than by name, it can place cells from any set of protein-coding genes, including species it has never seen and genes with no counterpart in its training data.
That design produced capabilities the model was never explicitly taught. In tests, its map organised cells into developmental lineages and correctly embedded species that had been left out of training. In standard zero-shot benchmarks it outperformed earlier single-cell models such as scGPT and Geneformer.
For working biologists, the practical payoff is straightforward: new measurements can be annotated and compared against a shared reference in a consistent way, and the likely function of newly discovered cell types can be inferred from their neighbours on the map. The team has released the model openly — an early step toward richer "virtual cell" models that could help researchers reason about health and disease across the whole tree of life.