UCE Single-Cell Foundation Model Published in Nature

The Universal Cell Embedding (UCE) paper, a single-cell foundation model by Jure Leskovec and Stephen Quake's teams, has been published in Nature. This work addresses the long-standing fragmentation in single-cell RNA sequencing analysis, where data from different diseases, tissues, or species are often confined to isolated atlases. UCE aims to map new data directly into a unified representation space without requiring fine-tuning or retraining.

Key Details

According to Stephen Quake, UCE is a representation learning model trained on 36 million cells from various species and tissues. Its core goal is to enable researchers to analyze and utilize single-cell data on new samples or entirely new organisms outside the training set.

Technical Approach

Jure Leskovec explains that UCE bridges molecular and cellular scales, moving beyond treating genes merely as columns in an expression matrix. The model represents genes based on the proteins they encode and uses ESM for encoding. The authors note that this cross-species protein language modeling design allows the model to generalize to unseen species, supporting cross-species single-cell representation.

2026-07-14 ~ 2026-07-15 · 5 related posts

• [source] UCE Paper Published in Nature — jure · 2026-07-14
• UCE: Universal Cell Embeddings for Single-Cell Data — jure · 2026-07-14
• Details of UCE's Cross-Species Representation — jure · 2026-07-14
• UCE: Cross-Species Protein Language Modeling — jure · 2026-07-14
• [source] Universal Cell Embedding Model Released — StephenQuake · 2026-07-15