A cell arrives at an AI model as a record of measurements. Its apparent differences from another cell may reflect biology, but they may also reflect how the experiment captured that biology. Learning across laboratories means learning to navigate both kinds of variation.[1][5]

That problem gives China's single-cell foundation models a more interesting story than a race to collect the largest training set. Two releases illustrate the direction: scFoundation, developed by Tsinghua University researchers and BioMap and published in June 2024, and CellFM, from a collaboration including Sun Yat-sen University and Chongqing University, published in May 2025.[1][3] They seek reusable knowledge from gene-expression measurements. Their designs and public implementations show what must happen before that knowledge becomes useful in another experiment.

The experiment is part of the input

Single-cell RNA sequencing records gene activity at the level of individual cells. These records are sparse and noisy, and measurements collected in different batches can carry technical differences.[3] A model asked to organise cells therefore faces an awkward possibility: it could produce an orderly result by learning the laboratory's habits.

Tsinghua's Chinese-language announcement describes scFoundation's response. The model handles approximately 20,000 genes, has 100 million parameters, and was trained on roughly 50 million human cell profiles. Its design preserves continuous expression values, while a pretraining task teaches it to recover higher-quality information from lower-quality input.[1]

The released model documentation makes that idea tangible. Users supply measurements in a specified gene order, and a target-resolution setting controls read-depth enhancement. The same package offers cell-level representations and gene-level outputs.[2] A representation, often called an embedding, is a numerical description that another analytical method can use.

My reading is that the important advance begins with an admission: the record is incomplete. A model's estimate of what a deeper measurement might reveal remains an estimate. It should travel alongside the observed data, with its origin clear, rather than quietly acquiring the status of another laboratory observation.

Scale still needs a specific job

CellFM enlarged the ambition to a training collection of 100 million human cells and a model with 800 million parameters. It used a modified RetNet architecture implemented in MindSpore.[3] These are measures of the training effort, not a count of biological questions answered.

The paper's perturbation experiment shows why the distinction matters. A perturbation changes a gene's activity; the prediction task asks how other expression measurements respond. The researchers inserted CellFM's gene representations into GEARS, an existing perturbation-prediction method, and evaluated the combination on the Adamson and Norman datasets. Reported improvements concerned that defined task and pipeline.[3]

That is a productive form of reuse. A pretrained model can supply better inputs to a specialist method without absorbing every part of its job. It also gives a reader somewhere concrete to inspect the claim: which intervention data, which downstream method, and which comparison?

Even the parameter headline needs this care. The CellFM paper used an 80-million-parameter variant for cell annotation and batch-effect correction.[3] Attaching the largest model's size to every result would erase a meaningful experimental choice.

A release becomes something another lab can inspect

scFoundation's repository separates model inference from applications including expression enhancement, drug-response prediction and perturbation prediction. Its model instructions expose the choices needed to obtain different representations.[2] That separation helps a researcher locate where a result comes from: the pretrained model, its input preparation, or the method built on top.

CellFM likewise publishes an implementation, links to datasets and checkpoints, and provides tutorials for annotation, gene-function prediction, batch integration and perturbation. Its documented environment specifies MindSpore 2.2.10; training examples distinguish loading pretrained weights and using LoRA, a method that adapts a model by updating a smaller set of parameters.[4]

These details turn a paper into a research object that can be challenged. Public code does not establish that a result will reproduce everywhere. It makes the necessary questions more answerable: which weights were loaded, what processing preceded them, and how much additional learning occurred?

What survives the move to unfamiliar data?

An independent April 2025 study supplies a useful caution. Kasia Kedzierska and colleagues evaluated scGPT and Geneformer without further task-specific training, a setting called zero-shot. Across five datasets, they compared the resulting cell representations with established approaches, including selecting highly variable genes, Harmony and scVI. The foundation models did not consistently outperform those alternatives.[5]

This was not a test of scFoundation or CellFM. It establishes a question to ask of them, rather than an adverse result to borrow. Fine-tuning can be valuable when suitable examples exist; exploratory work may begin before researchers know the categories they want to find. The study also identified overlap between some evaluation and pretraining data, reinforcing the need to document what a model has already encountered.[5]

Taken together, these signals suggest a practical way to follow China's cell-model research. Look for measurement-aware inputs, releases that expose the complete analytical procedure, and tests that distinguish adaptation from performance without additional training. A particularly persuasive next result would preserve its advantage across an unfamiliar experimental setting and against a strong simpler method.

The promise is that one laboratory can benefit from patterns learned across many others. The proof comes when that borrowed experience helps explain a new measurement without mistaking the experiment's fingerprints for the cell's identity.

Sources

1. Tsinghua University, “Department of Automation's life foundation model laboratory collaborates on an AI cell model,” June 12, 2024; Chinese-language primary announcement, collaborators, scale and measurement-quality design.
2. BioMap Research, scFoundation repository and model documentation; released inference, representation and downstream-analysis workflows, accessed October 20, 2026.
3. Yuansong Zeng et al., “CellFM: a large-scale foundation model pre-trained on transcriptomics of 100 million human cells,” Nature Communications, May 20, 2025; model design, task-specific variants and GEARS evaluation.
4. biomed-AI, CellFM official implementation; environment, data, checkpoints and application tutorials, accessed October 20, 2026.
5. Kasia Z. Kedzierska et al., “Zero-shot evaluation reveals limitations of single-cell foundation models,” Genome Biology, April 18, 2025; evaluation of scGPT and Geneformer, baseline comparisons and data-overlap limitations.
6. N509FZ, “Main building of Tsinghua University (20190709103617).jpg,” July 9, 2019; original campus photograph, Wikimedia Commons, CC BY-SA 4.0.