Skip to content

UCE

Family: sc-fm Modality: Single-cell RNA-seq (cross-species; ESM2 protein-embedding gene tokens) Released: 2023 (bioRxiv) License: MIT (code) Code/checkpoint: github.com/snap-stanford/UCE (4-layer default auto-downloads; 33-layer flagship via Figshare); model card on the CZI Virtual Cells Platform

UCE (Universal Cell Embedding) is a single-cell foundation model from the Leskovec and Quake groups at Stanford (with collaborators at Genentech) that learns a species-agnostic embedding space by tokenizing genes through their ESM2 protein representations rather than as integer IDs. Trained self-supervised on cell atlas data spanning eight species, UCE produces zero-shot cell embeddings that transfer across tissues and to previously unseen species, and is the engine behind the Integrated Mega-scale Atlas (IMA) of ~36M cells.

What it does

UCE turns any single-cell transcriptome — human, mouse, or a species the model has never seen — into a fixed-dimensional embedding without any task-specific fine-tuning. Because gene tokens are derived from the underlying protein sequence via ESM2, orthologs across species map to similar positions in the input space, so cell-type structure transfers zero-shot from well-studied to under-sampled organisms. The released checkpoints are typically used as a frozen feature extractor for clustering, label transfer, and atlas integration.

How it works

Architecture. UCE is a transformer encoder. The flagship paper checkpoint has 33 layers and over 650 million parameters; a smaller 4-layer variant ships as the GitHub default for users without multi-GPU resources (Rosen et al., bioRxiv 2023; snap-stanford/UCE README).

Tokenization. Each gene is represented by the ESM2 protein language-model embedding of the protein it encodes (ESM2 is a 15B-parameter protein LM operating on amino-acid sequences), with a chromosome positional token and a species token concatenated. A cell is then encoded as a sampled bag of expressed genes — UCE draws genes with replacement, weighted by their expression value — rather than as a ranked list or fixed-length vector (Rosen et al., bioRxiv 2023). This design lets the model accept any gene set from any species without retraining the vocabulary.

Pretraining corpus + objective. UCE is trained self-supervised, without any cell-type labels, on cell atlases from human and seven other species (eight species total) (Rosen et al., bioRxiv 2023). The objective is a binary masked-gene-presence task: given a sampled subset of a cell's expressed genes, predict which other genes from a candidate set are also expressed in that cell. This contrastive-style objective avoids quantitative reconstruction of expression values and emphasizes which genes co-occur in a cellular context.

Scale. 33 layers, ~650M parameters for the flagship; pretraining GPU-hours are not stated in the preprint (unverified — paper does not report wall-clock training cost). Inference of the 33-layer model requires a multi-GPU node; the 4-layer model fits on a single GPU (snap-stanford/UCE README).

Fine-tuning recipe. UCE is positioned as a zero-shot model: the standard workflow is to run a single forward pass over a query AnnData and use the cell embeddings directly for downstream clustering, label transfer (e.g. k-NN against a labeled reference), or visualization. No fine-tuning code is provided in the official repository, and the paper's headline results are zero-shot.

Downstream tasks evaluated. Cell-type annotation by label transfer, cross-species cell-type matching, batch integration on the scIB benchmark, and construction of the Integrated Mega-scale Atlas — a joint embedding of ~36M cells, >1,000 named cell types, spanning dozens of tissues and eight species (Rosen et al., bioRxiv 2023).

Headline benchmarks

  • scIB cell-embedding benchmark (zero-shot). UCE reportedly outperforms the next-best foundation model (Geneformer) by 13.9% on overall score, 16.2% on biological-conservation score, and 10.1% on batch-correction score (Rosen et al., bioRxiv 2023).
  • Per-cell-type win rate (zero-shot). UCE beats Geneformer on 80% of cell types, tGPT on 73%, and scGPT on 83% in the original paper's evaluation (Rosen et al., bioRxiv 2023).
  • Perturbation prediction (third-party). When evaluated outside its design envelope on the Adamson/Norman/Replogle perturbation benchmarks, UCE — like scGPT (0.641/0.554/0.327/0.596 Pearson-Delta) and scFoundation (0.552/0.459/0.269/0.471) — fails to beat a simple "predict the training-set mean" baseline that scores 0.711/0.557/0.373/0.628 (Ahlmann-Eltze et al., Nature Methods 2025). This is a class-wide limitation of single-cell FMs, not unique to UCE, but worth flagging for cancer-perturbation use cases.

Why it matters for AACR / virtual cells

UCE's species-agnostic tokenization is the cleanest published answer to a problem AACR audiences face every day: mouse models do not share gene IDs with human tumors, and PDX or syngeneic-model scRNA-seq is hard to project into a human reference. UCE's ESM2-grounded gene embeddings sidestep ortholog-mapping pipelines entirely, making it a natural backbone for cross-species cancer cell-of-origin questions and for harmonizing mouse perturbation screens with human atlases. It is also one of two published single-cell FMs (with scGPT) that the CZI Virtual Cells Platform packages as a hosted model card for the "virtual cell" research program (CZI Virtual Cells Platform).

Known limitations / open questions

  • No native fine-tuning workflow. The model is designed for zero-shot use; users who want task-specific adaptation must wire up their own head and risk catastrophic forgetting of the cross-species embedding geometry.
  • Inference cost. The 33-layer flagship requires multi-GPU inference and is impractical for sample-level workflows without caching embeddings (snap-stanford/UCE README).
  • Perturbation prediction is out of scope. Third-party benchmarks show UCE (alongside scGPT, scFoundation, Geneformer) loses to linear baselines on perturb-seq response prediction (Ahlmann-Eltze et al., Nature Methods 2025).
  • Tumor-cell coverage is underrepresented. IMA is biased toward normal cell atlases; the paper does not report a dedicated cancer evaluation (unverified — paper claims general transfer, but no pan-cancer benchmark is reported).
  • Still a preprint at time of writing. The work has circulated as a bioRxiv preprint since November 2023; a peer-reviewed venue is not confirmed in the public record we surveyed (unverified — paper remains on bioRxiv).

Sources