Universal cell embedding provides a foundation model for cell biology

2026-07-08 · Source: Machine learning : nature.com subject feeds · Field: Science & Research — Life Sciences & Biology, Mathematics & Computational Sciences, Research Methodology & Innovation · Depth: Expert, extended

Summary

The Universal Cell Embedding (UCE) foundation model provides a universal representation space for cells across diverse tissues and species. Trained self-supervised on a large corpus of over 36 million cells from hundreds of experiments, dozens of tissues, and eight species, UCE creates a unified biological latent space. This 33-layer model, with over 650 million parameters, was trained for 40 days across 24 A100 80 GB GPUs. UCE enables zero-shot embedding of new cells without labeling or fine-tuning, outperforming methods like Geneformer and scGPT by 13.9% in overall score on Tabula Sapiens v.2. It successfully integrates data from unseen species, including green monkey and chicken, and reveals emergent organization of cell types consistent with developmental lineages and ontological similarity. UCE also facilitates unbiased analysis, such as identifying Norn-like cells across tissues and exploring their role in lung diseases like IPF and COPD.

Key takeaway

For research scientists integrating diverse single-cell RNA-seq datasets, UCE provides a powerful, universal foundation model. You can embed new data, even from unseen species, without costly retraining or manual cell labeling. This enables efficient cross-species cell type annotation, unbiased discovery of novel cell functions, and hypothesis generation across various biological contexts and disease states. Consider UCE to streamline your single-cell data analysis workflows.

Key insights

UCE is a self-supervised foundation model creating universal cell representations for zero-shot analysis across species and tissues.

Principles

• Universal cell representations emerge from self-supervised training.
• Protein language models enable genome-agnostic cell embedding.
• Ontological similarity correlates with embedding space proximity.

Method

UCE converts scRNA-seq data into expression-weighted gene samples, represents genes via ESM2 protein embeddings, and processes them with a 33-layer transformer, trained via masked gene expression reconstruction.

In practice

• Embed new single-cell datasets without retraining.
• Transfer cell type annotations across species.
• Identify novel cell types and functions in disease.

Topics

• Universal Cell Embedding
• Single-cell RNA-seq
• Foundation Models
• Protein Language Models
• Cell Biology
• Zero-shot Learning

Code references

• snap-stanford/uce
• yhr91/uce_reproduce

Best for: AI Scientist, Research Scientist

Related on AIssential

• scLong: a billion-parameter foundation model for capturing long-range gene context in single-cell transcriptomics — scLong is a billion-parameter foundation model for scRNA-seq, capturing long-range gene dependencies and integrating external gene knowledge.
• Viral Proteins Reveal Geometry of Protein Language Models — Protein language models organize sequence embeddings by a "nativeness axis" while preserving specific biological group signals.
• CellNiche represents cellular microenvironments in atlas-scale spatial omics data with contrastive learning — CellNiche uses contrastive learning on spatial omics data to map cellular microenvironments across diverse biological contexts.
• Identifying spatial single-cell-level interactions with graph transformer — A self-supervised graph transformer identifies spatial single-cell interactions without known ligand-receptor pairs.
• SegTME-UNI2: A Foundation Model-Based Framework for Generalisable Multiclass Cell Segmentation and LLM-Driven Tumour Microenvironment Characterisation in Histopathology — SegTME-UNI2 integrates foundation models and pseudo-labeling for robust, generalizable multiclass cell segmentation and TME characterization.
• U-CECE: A Universal Multi-Resolution Framework for Conceptual Counterfactual Explanations — U-CECE offers a multi-resolution framework for conceptual counterfactual explanations, balancing expressivity and efficiency across different data complexities.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine learning : nature.com subject feeds.