Overview
| Name | xTrimoSCPerturb / scFoundation · aliases: xTrimoSCPerturb, scFoundation, xTrimoscFoundationα, BM_xTVC |
|---|---|
| Full name | Neither name has an expanded form |
| Ambiguous names (not auto-linked) | xTrimo |
| Family | scFoundation is the public foundation model; xTrimoSCPerturb is an unreleased challenge solution built on a modified version
|
| Organization | BioMap, Tsinghua University |
| Organization type | Mixed |
| Country | China |
| Model class | Single-cell FM / perturbation |
| Domain | Single-cell & perturbation |
| First release | · paper revision Basis: scFoundation bioRxiv preprint v1 source |
| Other milestones | |
| Latest update | † partly verified · results Basis: Official VCC 2025 wrap-up names xTrimoSCPerturb overall first The organiser published the ranking but no scores, paper or weights; the scFoundation repository had only README edits in 2025-11 Note: The wrap-up's exact day was not retrieved; month-level evidence only source The most recent sourced event, not a claim that nothing changed since. |
| Status | Active · Challenge result published within 12 months |
| Openness | Code: Partly open · Apache-2.0 Weights: Open · Apache-2.0 scFoundation code and weights are public (Apache-2.0); xTrimoSCPerturb has no public repository or weights (checked 2026-11-16) |
| Benchmarks reported | 1 counted (board 1) |
| Wet-lab validation | None recorded |
| Unverified fields | † latest_update, open_status.weights_license, training_data.b_private_licensed, background.funding |
| As of | 2026-11-17 |
What it does
scFoundation is an asymmetric Transformer of about 100M parameters pretrained on expression profiles over about 20k genes; the paper reports downstream tasks such as expression enhancement, drug response, perturbation prediction and cell-type annotation. xTrimoSCPerturb is the team's entry for the Arc Virtual Cell Challenge 2025, described as deep learning plus classical statistical features on a re-pretrained modified scFoundation; unlike general baselines such as scGPT/Geneformer it is a task-specific solution for one challenge.
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 | Official leaderboard or results | — | Board | Overall winner of VCC 2025 per the organizer wrap-up; scores unpublished. · evidence |
Other evaluations (outside the 44)
- Independent perturbation-prediction study (third party evaluation) Third-party: Five foundation models plus two deep models against deliberately simple linear baselines on single and double perturbations; none beat the baseline, scFoundation included
- scFoundation paper downstream tasks (Own task set) Self-reported: Expression enhancement, drug response, cell-type annotation, gene-module inference and others; author-reported
Developers report: scFoundation leads on many single-cell downstream tasks; xTrimoSCPerturb ranked first overall in VCC 2025 (the latter published by the organiser).
Independent evidence: The organiser confirmed the VCC 2025 ranking but its wrap-up notes pure AI methods did not consistently beat statistical baselines; an independent study found scFoundation did not beat simple linear baselines.
Training data
Group in Figure 7: Public data only.
| Class | Value | Detail |
|---|---|---|
| a. Public databases | Yes | scFoundation pretraining used over 50 million human single-cell transcriptomes source |
| b. Private or licensed data | Not disclosed | |
| c. Own wet-lab data used for training | No | The H1 perturbation data were produced by Arc and issued to teams for fine-tuning, not generated by the developer source |
| d. Wet lab used only for validation | No |
Wet-lab validation
None. The developers report no wet-lab validation.
Background
- Organization: BioMap was founded in 2020, is headquartered in Beijing with R&D sites in Suzhou, Hong Kong and Silicon Valley, and builds the xTrimo life-science foundation models; scFoundation also has Tsinghua University authors.
- Funding: A Series A above $100M in 2021; a 2024 strategic partnership and a further round reported with undisclosed amounts; a 2026-03 media report of a confidential Hong Kong IPO filing, not announced by the company.
- Products and partnerships: A 2023-10 Sanofi deal for biologics AI modules ($10M upfront, potential milestones over $1B); the protein language model xTrimoPGLM is a sibling.
- Related models: xTrimoPGLM, scGPT / Geneformer, STATE
Reading notes
- The VCC first place is a genuine official result, but the organiser itself notes pure AI methods did not consistently beat statistical baselines.
- With no public scores or weights, the rank cannot be turned into comparable accuracy and cannot be reproduced.
- Independent evaluations of scFoundation are generally unfavourable, a problem shared across single-cell foundation models.
Caveats
- xTrimoSCPerturb is unreleased with no paper; architecture and scores are not public.
- The site's Nature Methods 2024-08 is the issue month; online publication was 2024-06-06.
- Training data are observational only, with no self-generated perturbation set.
Sources
- Official VCC 2025 wrap-up (official site)
- scFoundation journal paper (journal)
- scFoundation preprint (preprint)
- scFoundation repository (repository)
- Independent perturbation-prediction evaluation paper (third-party paper)
Corrections
1 correction
| Field | Was | Now | Basis | Found |
|---|---|---|---|---|
| release_milestones.journal_online | 2024-08 (volume issue) | 2024-06-06 (published online) | source | 2026-11-16 |
Part of the AI for Bio benchmarks survey · all 112 models · open in the explorer. Record as of 2026-11-17.