Universal Cell Embedding (UCE)
Summary¶
Rosen, Roohani, Agrawal, et al. (2026) present the Universal Cell Embedding (UCE), a 650-million-parameter, 33-layer transformer foundation model that maps any cell — from any tissue or species — into a single shared, zero-shot embedding space, requiring no fine-tuning, cell-type labels, or dataset-specific retraining. UCE was trained on the Integrated Mega-scale Atlas (36 million cells, >1,000 uniquely named cell types, hundreds of experiments, dozens of tissues, 8 species) and, in benchmarking against Geneformer and scGPT on the held-out Tabula Sapiens v.2 atlas, outperformed the next-best method (Geneformer) by 13.9% overall, 16.2% on biological conservation, and 10.1% on batch correction — while its zero-shot embeddings even slightly outperformed fine-tuned, dataset-specific methods like scVI and scArches.
Method¶
- Gene-as-protein tokenization: rather than treating genes as arbitrary categorical tokens, UCE represents each gene by the ESM2 protein-language-model embedding (a 15-billion-parameter protein LM operating on amino acid sequence) of the protein it encodes — this lets UCE meaningfully represent any protein-coding gene from any species based on sequence alone, even species never seen during training.
- Cell representation: a cell's expressed genes are sampled with replacement, weighted by expression level, converted to their ESM2 protein-embedding tokens, grouped by chromosome (bounded by special start/end tokens) and sorted by genomic location, then passed through the transformer with a CLS token whose final-layer output is taken as the cell embedding.
- Scale: 33 transformer layers, >650M parameters, trained on >300 datasets (mostly from CellXGene) totaling >36M cells, for 40 days across 24 A100 80GB GPUs.
Performance¶
- Held-out benchmark (Tabula Sapiens v.2): 581,430 cells, 27 tissues, 167 batches, 162 cell types — UCE outperformed Geneformer by 13.9% (overall score), 16.2% (biological conservation), 10.1% (batch correction), using standard single-cell integration benchmark metrics; UCE's zero-shot embeddings also modestly outperformed fine-tuned methods (scVI, scArches) that require dataset-specific retraining.
- Cross-technology batch correction: on ovary tissue profiled by both 10x 3' v3.1 (45,757 cells) and Smart-seq2.5 (3,610 cells), UCE zero-shot embeddings matched fine-tuned methods' batch-correction scores while achieving higher biological-conservation scores.
- Zero-shot cross-species generalization: embedding datasets from species never included in training (naked mole rat, chicken, green monkey, western clawed frog), UCE's nearest-cross-species-centroid cell-type matching succeeded for 17/24 naked mole rat cell types and 12/15 chicken heart cell type centroids (within the nearest two centroids); UCE outperformed the supervised cross-species label-transfer methods SATURN and SAMap on 3 of 4 novel-species datasets, despite UCE being fully zero-shot/unsupervised and the comparators being fine-tuned specifically for cross-species mapping. UCE produced coherent (if less fine-grained) embeddings even for highly evolutionarily divergent species like Drosophila.
- Emergent structure: UCE-derived cell clusters showed greater similarity to the expert-curated Cell Ontology hierarchy (measured by adjusted Rand index) than other zero-shot embedding methods, without being explicitly trained on ontology structure.
- Novel cell-type discovery case study: zero-shot embedding of mouse renal cells recovered a distinct cluster corresponding to the recently identified "Norn cell" (the kidney's erythropoietin-producing fibroblast-like cell), illustrating a workflow where a novel cell type identified in one dataset can be searched for across any other tissue/species dataset using only UCE embeddings, without retraining.
Limitations (as reported)¶
- Benchmarks primarily evaluate recovery of expert-annotated cell types, which may not capture UCE's ability to represent finer processes like perturbation responses or cross-modality integration.
- Like other large foundation models, UCE is largely a black box, complicating mechanistic interpretation of its embeddings.
- Training data are biased toward mammalian species (particularly human and mouse) and certain tissues (e.g., brain), raising generalizability concerns for under-represented species/contexts.
- UCE does not replace traditional batch-effect correction for known experimental confounders, and expression-weighted gene sampling discards some fine-grained quantitative variation.
See Also¶
- Geneformer and scGPT — the two prior single-cell foundation models UCE is directly benchmarked against.
- Biologically Informed Neural Networks — a related but architecturally distinct approach (sparse biology-constrained networks rather than protein-language-model-based tokenization) to injecting external biological knowledge into a deep learning model.