Single-cell foundation models transfer across deep evolutionary distances

Zero-shot embeddings from Geneformer, scGPT, and UCE recover developmental order and preserve cell identities in Dictyostelium discoideum and Caenorhabditis elegans despite 1.2B and 600M years from humans.

• Why it matters: It is unclear whether single-cell foundation models trained on vertebrates transfer to deeply divergent species, and which representation types remain informative without retraining.
• What they did: The study compared zero-shot embeddings against PCA and scVI in developmental atlases, using ortholog mapping for Geneformer and scGPT and protein-sequence encoding for UCE. UCE covered 97% of measured genes, while Geneformer and scGPT used under one-fifth after mapping.
• The result: UCE most consistently matched or exceeded species-native baselines and, even restricted to ortholog-mappable genes, scored above vocabulary-based models on five of six annotation tasks, supporting cross-billion-year transfer.

AI summary of the abstract · may contain errors · not medical advice · read the paper.