Universal cell embedding provides a foundation model for cell biology

Developing a universal representation space for cells that encompasses the tremendous molecular diversity of cell types across species would be transformative for cell biology. Recent work using single-cell transcriptomic approaches to create molecular definitions of cell types in the form of cell atlases has provided the necessary data for such an endeavour1–3. Here we present the universal cell embedding (UCE) foundation model. UCE was trained on a large corpus of cell data using self-supervision, creating a unified biological latent space that can represent cells across diverse tissues and species. This latent space captures important biological variation despite the presence of experimental noise. UCE’s universality means that new cells can be embedded with no data labelling, model training or fine-tuning. We used UCE to create the Integrated Mega-scale Atlas, embedding 36 million cells, with more than 1,000 uniquely named cell types, from hundreds of experiments, dozens of tissues and eight species. We gain insights into the organization of cell types and tissues within the space. UCE’s embedding space exhibits emergent behaviour, identifying biology that it was never trained for, such as identifying developmental lineages and embedding data from species that were not included in the training set. Overall, by enabling a universal representation for every cell state and type, UCE is a valuable tool for analysis, annotation and hypothesis generation over single-cell data. The universal cell embedding foundation model learns to capture the organization and variation of cells by training on 36 million cells from hundreds of experiments, dozens of tissues and eight species.