Overview
| Name | Stack · aliases: Stack-Large |
|---|---|
| Full name | The preprint gives no expanded name |
| Ambiguous names (not auto-linked) | Stack |
| Family | Not a family entry. Only one public size exists, and the companion prediction set is not a separate model |
| Organization | Arc Institute |
| Organization type | Nonprofit institute |
| Country | United States |
| Model class | Single-cell foundation model (in-context learning) |
| Domain | Single-cell & perturbation |
| First release | · paper revision Basis: Preprint v1 posted (10.64898/2026.01.09.698608) Note: Code v0.1.1 followed the next day (2026-01-10) source |
| Other milestones | |
| Latest update | · paper revision Basis: Preprint v2 posted Preprint v2 posted; the earlier event is GitHub v0.1.3 (2026-03-16, dependency fixes), the last release. No journal version found Note: The latest update was recorded as maintenance-only; corrected to the later paper revision source The most recent sourced event, not a claim that nothing changed since. |
| Status | Active · Paper revision and a version release within 12 months |
| Openness | Code: Open Weights: Non-commercial · non-commercial licence The largest size's weights are on the model hub under a non-commercial licence; code is public and the licence text was not retrieved this round |
| Benchmarks reported | 1 counted (modified protocol 1) |
| Wet-lab validation | None recorded |
| Unverified fields | † release_milestones.weights_release, open_status.code_license, other_evaluations[1] |
| As of | 2026-11-16 |
What it does
After pretraining, reference cells in the prompt express a condition (a perturbation, a disease state) that the model transfers to target cells without updating weights; a companion release contains about twenty thousand predicted responses across cell type, tissue and perturbation combinations, with tissue lineages taken from a public human atlas. Unlike the institute's STATE, it is a general in-context model rather than a response predictor trained on perturbation data.
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 | Own perturbation sets; its VCC claim is rejected on the model page. · evidence |
Other evaluations (outside the 44)
- Large predicted-response set (Case study) Self-reported: About twenty thousand predicted combinations, all prediction with no wet-lab validation: a hypothesis library, not facts
- Conflicting perturbation counts (Own task set) Self-reported: Different versions of the material give two perturbation totals while the launch page gives only combination counts; cite the source version
Developers report: in-context learning transfers conditions without fine-tuning and consistently beats classical methods trained from scratch and other foundation models on perturbation prediction.
Independent evidence: No independent evidence found. No third-party evaluation or replication was found, and the independent literature's findings about linear and mean baselines on this task have not been tested on this model
Training data
Group in Figure 7: Public data only.
| Class | Value | Detail |
|---|---|---|
| a. Public databases | Yes | Pretrained on 149 million cells from a public count collection; post-trained on 55 million cells from a public cell atlas and peripheral-blood perturbation data produced by a reagent company source |
| b. Private or licensed data | Not disclosed | |
| c. Own wet-lab data used for training | No | None by the developer; the perturbation data used in post-training was produced by a third-party company source |
| d. Wet lab used only for validation | No | The companion prediction set has no wet-lab validation source |
Wet-lab validation
None. The developers report no wet-lab validation.
Background
- Organization: A California non-profit research institute founded in 2021 with over $650M committed at launch; core investigators hold multi-year renewable terms alongside university appointments. No citable official headcount, so none is given.
- Funding: Non-profit and endowment-funded; 2024 budget figures differ between sources, so no single number is given.
- Products and partnerships: The pretraining dataset is institute-built as part of its cell-atlas initiative; the post-training perturbation data comes from a third-party reagent company. Sibling models are STATE and Evo 2 under the same virtual-cell initiative.
- Related models: STATE, Evo 2, Tahoe-x1
Reading notes
- Comparisons are mainly against two classical methods and other foundation models; the independent literature finds linear and mean baselines hard to beat on perturbation prediction, and those baselines are absent here.
- The roughly twenty thousand predicted responses have no wet-lab validation and are not established biology.
- No public score exists on the challenge, so there is no standing to speak of.
- The perturbation data came from a third-party company and must not be attributed to the developer.
Caveats
- The only site pairing is self-reported under a modified protocol with no citable numbers.
- No journal version; only two preprint versions.
- Perturbation totals conflict between materials.
- Weights carry a non-commercial licence.
Sources
- Preprint (preprint)
- Preprint version record (preprint)
- Developer launch page (official site)
- Code repository releases (repository)
Corrections
2 corrections
| Field | Was | Now | Basis | Found |
|---|---|---|---|---|
| latest_update | 2026-03-16, code maintenance only | 2026-06-08, preprint v2 (paper revision) | source | 2026-11-16 |
| benchmarks_reported | Claims related to the challenge | The release page does not mention the challenge and gives no score; only the modified-protocol self-report on the developers' own perturbation split is kept | 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.