Single-cell foundation models transfer across deep evolutionary distances
- 1Faculty of Computer and Information Science, University of Ljubljana, Ljubljana, Slovenia
Abstract
Single-cell foundation models are trained mostly on human and other vertebrate cells, yet are proposed as general-purpose models. Whether they transfer far beyond their training species is unknown. We evaluated zero-shot embeddings from Geneformer, scGPT and UCE against PCA and scVI in developmental atlases of Dictyostelium discoideum and Caenorhabditis elegans, separated from humans by 1.2 billion and 600 million years. Without retraining, the models recovered developmental order and preserved organism-specific cell identities. UCE encodes genes by protein sequence and could use 97% of measured genes, whereas Geneformer and scGPT, after ortholog mapping, used under one-fifth. UCE performed most consistently, often exceeding species-native baselines. Restricted to ortholog-mappable genes, it still scored above vocabulary-based models on five of six annotation tasks, some narrowly. Foundation models can thus yield informative cell embeddings across more than a billion years of evolution, and protein-sequence gene representations make these embeddings more complete and often more accurate.
Related
- Prediction transfer does not imply reliability transfer in single-cell foundation models
- Inferring evolutionary distances from whole prokaryotic genome alignments, accounting for among-locus substitution rate heterogeneity
- Mental Timing in Locomotion: Walking Across Real and Imagined Distances
- Relaxed evolutionary constraints on mitochondrial electron-transfer-associated variants in human cancers
- Lightweight Semantic EEG Foundation Model for Frozen Cross-Disorder Transfer
BibTeX
@misc{trnovec2026singlecell,
title = {{Single-cell foundation models transfer across deep evolutionary distances}},
author = {L. Trnovec and M. Katoh-Kurasawa and R. N. Arey and G. Shaulsky and B. Zupan},
howpublished = {bioRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.28.762457},
url = {https://www.biorxiv.org/content/10.64898/2026.10.28.762457v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Single-cell foundation models transfer across deep evolutionary distances
AU - L. Trnovec
AU - M. Katoh-Kurasawa
AU - R. N. Arey
AU - G. Shaulsky
AU - B. Zupan
PY - 2026
DA - 2026/10/28
PB - bioRxiv
DO - 10.64898/2026.10.28.762457
UR - https://www.biorxiv.org/content/10.64898/2026.10.28.762457v1
AB - Single-cell foundation models are trained mostly on human and other vertebrate cells, yet are proposed as general-purpose models. Whether they transfer far beyond their training species is unknown. We evaluated zero-shot embeddings from Geneformer, scGPT and UCE against PCA and scVI in developmental atlases of Dictyostelium discoideum and Caenorhabditis elegans, separated from humans by 1.2 billion and 600 million years. Without retraining, the models recovered developmental order and preserved organism-specific cell identities. UCE encodes genes by protein sequence and could use 97% of measured genes, whereas Geneformer and scGPT, after ortholog mapping, used under one-fifth. UCE performed most consistently, often exceeding species-native baselines. Restricted to ortholog-mappable genes, it still scored above vocabulary-based models on five of six annotation tasks, some narrowly. Foundation models can thus yield informative cell embeddings across more than a billion years of evolution, and protein-sequence gene representations make these embeddings more complete and often more accurate.
ER -