Tabula: A Privacy-Preserving Foundation Model for Single-Cell Genomics and Cell Rejuvenation Factor Discovery

Researchers have introduced Tabula, a foundation model for single-cell analysis that trains on data from multiple institutes without moving it off-site. It was used to identify candidate rejuvenation factors for human cells. The preprint was published on arXiv in July 2026.

What is the Tabula model

Tabula is a privacy-preserving foundation model for single-cell genomics that explicitly accounts for the tabular structure of single-cell data. The authors note that, unlike text, single-cell data is unordered and has a distinctive tabular form that existing models ignore. Tabula is built on federated learning, an approach in which several organizations jointly train one model without releasing their raw data to a shared pool.

• Model: Tabula, a privacy-preserving foundation model for single cells
• Method: federated learning plus explicit modeling of the data's tabular structure
• Platform: Chiron, a decentralized AI-agent-based system for collaborative training across institutes
• Data: a new scRNA-seq dataset of paired young and old human fibroblasts
• Biological systems: hematopoiesis, pancreatic endogenesis, neurogenesis, cardiogenesis

How to train without leaking data

Chiron handles collaborative training without exchanging raw data. It is a decentralized AI-agent-based platform that the authors developed specifically for Tabula. Institutes train the shared model on their own closed single-cell datasets without uploading them anywhere. This removes a major barrier to scaling genomic models, since patient data is protected. According to the preprint, in addition to strong results on downstream benchmarks, Tabula uncovers combinatorial regulatory logic across four biological systems at once: hematopoiesis, pancreatic endogenesis, neurogenesis, and cardiogenesis.

What the model found about aging

Tabula proposed candidate rejuvenation factors on a new scRNA-seq dataset of paired young and old human fibroblasts. The model ranked candidates in silico, guided by age and cell-identity scores (age- and identity score-guided prioritization), and, as the authors report, outperformed classical methods for this kind of selection. The output is a nomination of candidates: genes and factors worth testing in experiments that return old cells to a younger state. It is not a finished therapy.

"Tabula represents an important step in single-cell modeling,

combining tabular learning with federated learning and paving the way toward private virtual cells for human health," the preprint's arXiv abstract states.

What this means

Tabula shows how to combine two hot topics, foundation models and data privacy, in biomedicine. Organizations can train a shared model on closed genomic data, and its output of rejuvenation candidates renews interest in "virtual cells" as a tool for finding aging therapies.

Frequently asked questions

How does Tabula differ from other single-cell models?

Tabula explicitly models the tabular structure of single-cell data, which, according to the authors, existing foundation models ignore. It also trains through federated learning, without transferring raw data between institutes.

What are the rejuvenation factors Tabula found?

They are candidate genes that the model selected from data on young and old human fibroblasts, guided by age and cell-identity scores. They still have to be tested experimentally. This is a prioritization for further research, not a finished therapy.