Single-cell foundation models transfer across deep evolutionary distances

Single-cell foundation models, typically trained on human and vertebrate cells, have demonstrated the ability to transfer knowledge across vast evolutionary distances. When tested on developmental atlases of Dictyostelium discoideum and Caenorhabditis elegans, which diverged from humans by 1.2 billion and 600 million years respectively, these models, including Geneformer, scGPT, and UCE, successfully recovered developmental order and preserved organism-specific cell identities without any retraining.

Among the three models, UCE (Universal Cell Embedding) stood out for its remarkable performance. It encodes genes using protein sequence information, enabling the utilization of up to 97% of the measured genes. Despite employing only a fraction of the available genes compared to Geneformer and scGPT, UCE consistently outperformed native species baselines across most evaluation tasks.

Moreover, even when restricted to ortholog-mappable genes, UCE maintained its leading position, often surpassing vocabulary-based models on five out of six annotation tasks. Some of these tasks were won by a narrow margin.

The findings suggest that single-cell foundation models, with their protein-sequence gene representations, can yield informative cell embeddings that span over a billion years of evolutionary history. This capability could prove invaluable for understanding the fundamental principles of cellular development and function across diverse species.