Overview
| Name | Lingshu-Cell · aliases: Lingshu-Cell, 灵枢·细胞 |
|---|---|
| Full name | The paper gives no expanded name |
| Ambiguous names (not auto-linked) | Lingshu |
| Family | Not a family entry. A single model with no later version |
| Organization | Alibaba DAMO Academy |
| Organization type | Big-tech lab |
| Country | China |
| Model class | Single-cell generative / virtual cell |
| Domain | Single-cell & perturbation |
| First release | · paper revision Basis: arXiv preprint v1 source |
| Other milestones | |
| Latest update | · no later event Basis: Only arXiv v1 exists No later sourced event found (as of 2026-11-16); no new version, code or weights source The most recent sourced event, not a claim that nothing changed since. |
| Status | Active · First released 2026-03, within 12 months |
| Openness | Code: Closed Weights: Closed Neither the project page nor the paper gives a code repository or weights; parameter count not reported |
| Benchmarks reported | 1 counted (retrospective 1) |
| Wet-lab validation | None recorded |
| Unverified fields | None |
| As of | 2026-11-17 |
What it does
Discretises expression into tokens and learns the distribution of cell states over about 18k genes by masked discrete diffusion, generating unconditionally or conditioned on cell type, donor and perturbation; gene tokens are grouped and compressed to cut attention cost. Unlike STATE, which regresses pre- to post-perturbation states, it targets distributions rather than mean responses, and unlike CellFM its main output is not a cell embedding.
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 | Retrospective dataset use | — | Third-party | A post-hoc evaluation on the VCC H1 leaderboard rather than participation. · evidence |
Other evaluations (outside the 44)
- Cytokine perturbation prediction (PARSE PBMC) (Own task set) Self-reported: 12 donors × 90 cytokine conditions against PertMean, STATE, scGPT and scVI; bar charts only, no per-metric table
- Transcriptome distribution reproduction (Own task set) Self-reported: Reproduces distributions, marker genes and subtype proportions across nine tissues and several species; author-reported
Developers report: generative modelling of cell-state distributions, the best mean rank in a post-hoc comparison on the VCC H1 genetic-perturbation board, and the best average on cytokine perturbation prediction.
Independent evidence: No independent evidence found. Code and weights are not public; no third-party reproduction found
Training data
Group in Figure 7: Public data only.
| Class | Value | Detail |
|---|---|---|
| a. Public databases | Yes | PARSE 10M PBMC, CZ CELLxGENE, scBaseCount, VCC H1 training data and public perturbation datasets source |
| b. Private or licensed data | Not disclosed | Not disclosed in the paper; no use found source |
| c. Own wet-lab data used for training | No | Perturbation data were produced by third parties such as Arc and Parse source |
| d. Wet lab used only for validation | No |
Wet-lab validation
None. The developers report no wet-lab validation.
Background
- Organization: DAMO Academy is Alibaba's research arm, announced in 2017-10 with a plan for RMB 100 billion over three years and seven labs in several countries; a big-tech lab, not separately funded.
- Funding: Funded from Alibaba's internal R&D budget; the team's own budget is not public.
- Products and partnerships: DAMO has other recent science projects such as a materials-discovery agent; the team discloses no pharma partnerships and relies on Arc Institute's VCC data.
- Related models: STATE, CellFM, Stack, Tahoe-x1, xTrimoSCPerturb / scFoundation
Reading notes
- The best mean rank is driven by a few metrics while another sits near the bottom, so metric choice changes the conclusion.
- Evaluation was post hoc on public data and did not go through the blind challenge process.
- Independent VCC-era evaluations generally find simple baselines hard to beat, so its edge needs third-party reproduction.
- Without code or weights it cannot be cross-checked on other benchmarks.
Caveats
- The VCC relation is post-hoc data reuse, not counted as participation.
- Parameter count and licence are not published.
- All results are author-reported.
Sources
- arXiv preprint (preprint)
- Project page (official site)
- VCC leaderboard (leaderboard)
- arXiv version record (preprint)
- DAMO Academy launch coverage (press)
Corrections
None recorded.
Part of the AI for Bio benchmarks survey · all 112 models · open in the explorer. Record as of 2026-11-17.