Scientists Build a Universal 'Map' of Cell Types Across Species

Researchers have developed an artificial intelligence system that can represent and compare cells from any tissue or species within a single, unified framework, without needing to retrain the model for each new dataset, according to a study published in Nature.

The system, called the universal cell embedding (UCE) foundation model, was trained on a large collection of single-cell data using self-supervision. This process allowed it to build what the researchers describe as a unified biological latent space, a kind of common coordinate system that can position cells according to their molecular characteristics regardless of where or in what species they were measured.

Using UCE, the team built what they call the Integrated Mega-scale Atlas, embedding 36 million cells spanning more than 1,000 uniquely named cell types. The atlas draws on hundreds of experiments across dozens of tissues and eight species.

How it works

Single-cell RNA sequencing has made it possible to capture detailed molecular snapshots of individual cells from many tissues, donors, timepoints and species. But according to the study, existing computational methods have struggled to analyze these diverse datasets together. Differences between experiments, often called batch effects, can obscure real biological signals, and many existing tools require dedicated retraining or data labeling for every new dataset, making analysis slow and often limited to small, isolated collections of data.

UCE addresses this by treating each cell's RNA expression as what the authors call a "bag of RNA." The model converts a cell's gene expression into a sample weighted by how strongly each gene is expressed, then represents each gene by its protein product using a separate protein language model called ESM2. Because this protein-based representation relies on amino acid sequences rather than dataset-specific gene lists, UCE can process protein-coding genes from any species, even ones absent from its original training data. Additional information about gene locations on chromosomes is then incorporated, and the combined representation is processed by a large transformer model, the same type of architecture used in general-purpose AI systems.

The result, according to the researchers, is a model that can map new cells into the shared embedding space with no additional data labeling, model training or fine-tuning, and with no need to preselect genes beforehand.

Emergent behavior

The study reports that UCE's learned representations organize cell types and states in ways consistent with known biology, and can be used to predict cell types without retraining the model, performing better than existing large-scale data integration methods in some tests. The researchers also observed what they describe as emergent behavior: UCE identified developmental lineages and meaningfully embedded data from species it was never trained on.

The authors suggest UCE could serve as a general-purpose tool for analyzing, annotating and generating hypotheses from single-cell data, allowing insights from one dataset, such as a classifier trained to recognize specific immune cell types, to be applied directly to entirely new datasets without rebuilding the model each time.