Overview
| Name | CellFM · aliases: CellFM |
|---|---|
| Full name | The name is an acronym compound with no expanded form in the paper |
| Ambiguous names (not auto-linked) | None |
| Family | Not a family entry. A single model with no family of versions |
| Organization | Sun Yat-sen University (computer science), Chongqing University, Huawei, Singleron Biotechnologies |
| Organization type | Mixed |
| Country | China |
| Model class | Single-cell foundation model (linear-complexity backbone) |
| Domain | Single-cell & perturbation |
| First release | · paper revision Basis: Preprint posted (registry record) source |
| Other milestones | |
| Latest update | † unverified · author correction Basis: Author correction statement in the repository README An author correction in the repository README: miscalculated averages in several figure panels plus evaluation bugs affecting two baselines on two datasets, with the authors stating the conclusions are unchanged. The earlier event is the journal publication on 20 May 2025. Note: No independent permalink for the commit day was obtained, so the date is unverified, and no formal journal erratum was found source The most recent sourced event, not a claim that nothing changed since. |
| Status | Active · An author correction statement within 12 months |
| Openness | Code: Open Weights: Open Code in a public repository with a long-term archive and weights in a public model hub; licence text not retrieved this round |
| Benchmarks reported | 1 counted (modified protocol 1) |
| Wet-lab validation | None recorded |
| Unverified fields | † latest_update, open_status.code_license, open_status.weights_license |
| As of | 2026-11-16 |
What it does
Pretrained on human single-cell data only, so that mixing species does not drag down performance on human cells, with a linear-complexity sequence backbone and downstream tasks spanning cell-type annotation, perturbation prediction, gene function prediction and gene relationship capture; training ran on a domestic AI framework and accelerators. It follows the same route as scGPT/Geneformer, distinguished by human-only data and domestic compute.
Benchmarks reported
Every row is one pairing from the survey's relation file; a score belongs to the benchmark page's ranking, where it is shown with metric, eval set, source type and date. 1 of 1 pairings are counted.
| Benchmark | Relation | Eval set / protocol | Source type | Note and evidence |
|---|---|---|---|---|
| 24 Virtual Cell Challenge and perturbation benchmarks | Developer report, modified protocol | — | Self-reported | Adamson and Norman data through GEARS, not VCC. · evidence |
Other evaluations (outside the 44)
- Cell annotation and gene function prediction (Public benchmark not among the 44) Self-reported: Compared with two single-cell foundation models on public datasets, reporting relative gains in mean accuracy
- Third-party perturbation evaluations (Public benchmark not among the 44) Third-party: Two independent evaluations find simple baselines competitive with such foundation models; neither includes this model
Developers report: human-only pretraining on about a hundred million cells with a linear-complexity backbone, beating comparable models on annotation, perturbation prediction and gene function prediction, mostly as single-digit relative gains.
Independent evidence: No independent evidence found. Two third-party evaluations do not include this model, and no independent evaluation of it was found
Training data
Group in Figure 7: Public data only.
| Class | Value | Detail |
|---|---|---|
| a. Public databases | Yes | 19,914 samples and about 102.3 million human cells from public databases source |
| b. Private or licensed data | No | |
| c. Own wet-lab data used for training | No | |
| d. Wet lab used only for validation | No |
Wet-lab validation
None. The developers report no wet-lab validation.
Background
- Organization: An academic team in a Chinese university's computer science school, with collaborators at another university, a laboratory, a telecommunications and compute company, an affiliated hospital and a single-cell sequencing company.
- Funding: The acknowledgements list national key R&D programme, national natural science foundation, central university research funds, postdoctoral funds and a provincial natural science foundation.
- Products and partnerships: Code sits under a public repository organisation with a weights card in a public model hub; comparable models include scFoundation and scGPT/Geneformer.
- Related models: scGPT / Geneformer, xTrimoSCPerturb / scFoundation, STATE
Reading notes
- All performance claims are self-reported as relative gains over the runner-up, so the absolute gaps are small.
- Independent work finds simple baselines competitive with such foundation models on perturbation prediction, and that doubt applies here too.
- The authors posted their own correction of figure and evaluation bugs in the repository, which is commendable, though no formal journal erratum was found.
- Its value lies more in human-only data and the feasibility of domestic compute than in a verified predictive edge.
Caveats
- No independent permalink was obtained for the correction commit's day.
- No formal journal erratum was found; the correction appears only in the repository README.
- It has no result on the public blind-test challenge.
- Licence text was not retrieved this round.
- The perturbation evaluation uses early public datasets, which differ from current blind-test protocols.
Sources
- Journal paper 2025 (journal)
- Preprint 2024 (preprint)
- Code repository and correction statement (repository)
- Third-party perturbation evaluation (third-party paper)
Corrections
1 correction
| Field | Was | Now | Basis | Found |
|---|---|---|---|---|
| latest_update | 2025-05-20 (journal publication) | 2026-09-07 (author correction notice in the repository); event kind changed to author correction | source | 2026-11-16 |
Part of the AI for Bio benchmarks survey · all 112 models · open in the explorer. Record as of 2026-11-16.