A professor in Beijing spent years trying to turn a cell's chemical instructions into something a machine could read the way a person reads a sentence.
This spring, the work was recognized. The project was named among the year's ten most important advances in bioinformatics in China — a rare crossover moment for a large model (大模型) pointed not at language, but at the machinery of life.
The scientist leading it is Rui Jiang (江瑞) of Tsinghua University, and the model is called EpiAgent.
A language for the genome's switches
To understand what EpiAgent does, start with a distinction. DNA is the text of a cell, but epigenomics is about the annotations in the margins: chemical marks that decide which genes are switched on, in which cells, and when. Those marks help explain how a liver cell and a neuron can carry the same genome and behave completely differently.
The problem is scale. A single cell can carry millions of regulatory elements, and reading them one by one is slow and fragmentary. Jiang's team approached it as a language problem. EpiAgent compresses the regulatory elements of one cell into short sequences the researchers call "cell sentences" (细胞语句), then feeds those sequences to a model that learns the grammar of gene control.
The resulting foundation model (基础模型) has 1.4 billion parameters. It was trained on a dataset the team assembled called Human-scATAC-Corpus, covering roughly five million cells and 35 billion regulatory elements. Tsinghua describes EpiAgent as the first foundation model built for single-cell epigenomics (单细胞表观基因组) — a claim made by the university, not an independent judgment.
The paper behind the recognition
The work was published in Nature Methods in 2025, in volume 22, pages 2316 to 2327. The paper is titled "EpiAgent: foundation model for single-cell epigenomics" and carries the DOI 10.1038/s41592-025-02822-z.
The Nature Methods record is clear about who did what. Rui Jiang conceived the study and supervised the project, and is the corresponding author. Xiaoyang Chen, the first author, collected and processed the data and designed, implemented and validated the model. The affiliations place the work inside Tsinghua's Department of Automation and its bioinformatics divisions, including the Beijing National Research Center for Information Science and Technology.
On March 23, 2026, the journal Genomics, Proteomics & Bioinformatics announced its selection of the 2025 China Top Ten Bioinformatics Advances (中国生物信息学十大进展). The EpiAgent paper made the list. That is the event that brought a fairly technical paper to a broader audience — an annual recognition, not a new result.
What a "virtual cell" could change
The interesting part of EpiAgent is not its parameter count but what the team says it can do with that compressed representation. The list of capabilities points to a shift in how biology might be done:
- Integrating large, messy epigenomic datasets that previously had to be analyzed in fragments.
- Predicting how tumor cells respond to internal and external perturbations, without running every experiment in a lab.
- Simulating the virtual knockout (虚拟敲除) of a regulatory element to see what changes downstream.
- Supporting what the researchers call a virtual cell (虚拟细胞) paradigm — a model that can be probed before a wet-lab test is designed.
None of that removes the need for experiments. It changes which experiments get prioritized. A model that can suggest which of a hundred regulatory elements is worth studying could save months of trial and error, which is why this kind of work attracts attention well outside machine learning.
The data and code have been made available. The training data sit in the ENSEMBL database, built from 27 published datasets plus an integrated set of 30 PBMC samples from 10x Genomics. The code is on GitHub under an open repository. That matters because a "foundation model" claim is only useful to the field if other researchers can inspect and reproduce it.
The broader shift
Jiang's project is part of a wider pattern in Chinese research: foundation models originally developed for text and images being turned toward scientific data. The same idea shows up in protein structure, weather and materials. What is distinctive here is the target — not predicting a sequence, but modeling the control layer that sits above it.
For ordinary readers, the relevance is indirect but real. Epigenomics sits behind questions about how cancers become resistant to drugs and how cells change with age. Better models of gene regulation do not deliver therapies on their own, but they shape the questions researchers ask next.
Honest limitations
The key facts here come from the Nature Methods paper and official Tsinghua announcements of the award; both are authoritative, but they describe different things. The paper establishes what the model was built to do and who built it. The award is a recognition chosen by a journal's review process, not a measure of real-world impact. The "first foundation model for single-cell epigenomics" framing is Tsinghua's own, and should be read as such. The claimed abilities — perturbation prediction, virtual knockout, virtual cell — are research results on specific datasets, not validated clinical tools. No independent replication of the full system was available in the sources reviewed. Readers should treat EpiAgent as a strong research contribution, not a proven route to new treatments.
What readers can do now
- If you work in computational biology, the open Nature Methods paper and the GitHub repository are the place to start, not the award announcement.
- If you are evaluating AI-for-science claims, separate the model's published benchmark results from the broader "virtual cell" ambition.
- If you follow Chinese research, watch whether other groups reproduce EpiAgent on independent data; that is the test that matters.
