AI Model Creates a Universal Map of Cell Biology, Works Across Species Without Retraining

A new artificial intelligence model can instantly place any cell—from human, mouse, chicken, or even fruit fly—into a shared biological map, without needing to be retrained on new data. The advance, reported this week in Nature, offers a powerful tool to compare cells across tissues, species, and diseases in a way that has been extremely difficult until now.

The work was led by computer scientist Jure Leskovec at Stanford University, with first author Yanay Rosen. The team built a foundation model called Universal Cell Embedding (UCE), which learned a unified representation of cell states by analyzing the gene expression of over 36 million cells from eight species.

What sets UCE apart is its “zero-shot” ability: it can take a completely new single-cell RNA sequencing dataset—even from a species it never saw during training—and map those cells into the same universal space alongside all previously seen cells. No manual labeling, model fine-tuning, or gene selection is required. When tested on a high-quality human cell atlas not used in training, UCE’s embeddings outperformed other self-supervised methods by a substantial margin and even matched or beat some fine-tuned approaches that rely on cell-type labels.

The model represents cells in a way that automatically organizes them by biological identity. Immune cells cluster with immune cells, epithelial cells with epithelial cells, despite originating from different tissues or labs. This organization extends to developmental lineages: for example, in a held-out human dataset, mesoderm-derived cell types were almost perfectly neighbored by other mesoderm-derived cells. The universal space also captures evolutionary relationships, successfully aligning cell types between human and green monkey, naked mole rat, or chicken—species absent from training.

These properties enable practical, large-scale analyses that were previously cumbersome. The researchers illustrate this by hunting for “Norn cells”—a recently discovered fibroblast-like cell in the kidney that produces the hormone erythropoietin. Using a simple classifier trained on UCE embeddings of mouse kidney, they scanned the entire 36-million-cell atlas and found Norn-like cells not only in kidney but also in lung and heart across many datasets. The predicted cells expressed canonical Norn markers. In a lung disease dataset, Norn-like cells were present in patients with idiopathic pulmonary fibrosis (IPF) and chronic obstructive pulmonary disease (COPD), and showed distinct gene expression patterns, including lower levels of an oxygen-sensing enzyme in IPF. The finding hints that these Norn-like cells could have a previously unrecognized role in disease.

UCE’s broad utility comes with caveats. The model remains largely a black box, making it hard to interpret why certain cells end up where they do. Its training data, while vast, is skewed toward mammalian species and specific tissues like brain. The approach also loses some fine-grained quantitative expression information because it relies on sampling genes based on their expression level. And while UCE reduces batch effects, it does not replace dedicated correction methods when experimental confounders are known.

Nevertheless, the model represents a significant step toward a universal reference for cell biology. By mapping any cell into a shared coordinate system, UCE could accelerate the annotation of new cell types, enable cross-species comparisons, and generate hypotheses about cell function in health and disease, all without the need to retrain models for each new experiment.

Reference: Rosen, Y., Roohani, Y., Agrawal, A. et al. Universal cell embedding provides a foundation model for cell biology. Nature (2026). https://doi.org/10.1038/s41586-026-10689-z