Overview
| Name | STATE · aliases: STATE model |
|---|---|
| Full name | State Transition + State Embedding |
| Ambiguous names (not auto-linked) | STATE, State |
| Family | Not a family entry. The developers say a next generation is in development but unreleased; no family entry |
| Organization | Arc Institute |
| Organization type | Nonprofit institute |
| Country | United States |
| Model class | Single-cell perturbation prediction / virtual cell |
| Domain | Single-cell & perturbation |
| First release | † partly verified · paper revision Basis: Preprint (10.1101/2025.06.26.661135) and the developer's launch page Note: The site also records a 2025-06-27 posting; the day-level difference is unresolved source |
| Other milestones |
|
| Latest update | † partly verified · journal Basis: Journal publication (issue is month-level) Journal version published; the repository's v0.11.x releases (2026-06 to 07) are code releases. The developers say a next generation is in development but unreleased Note: No online-first day was found; only month-level evidence source The most recent sourced event, not a claim that nothing changed since. |
| Status | Active · Journal publication and code releases within 12 months |
| Openness | Code: Open Weights: Non-commercial · non-commercial licence Code is public and weights carry a non-commercial licence; the full licence text was not retrieved this round |
| Benchmarks reported | 1 counted (modified protocol 1) |
| Wet-lab validation | 1 recorded campaign |
| Unverified fields | † first_release, latest_update, release_milestones.journal_online, benchmarks_reported[0], open_status.code_license |
| As of | 2026-11-16 |
What it does
Learns state transitions between pre- and post-perturbation cells to predict transcriptional response to drugs or genetic perturbation across cell lines, with a developer-built evaluation tool alongside. Unlike the institute's Stack, which does prompt-style prediction by in-context learning, this model trains on perturbation data for response prediction; it shares the same third-party perturbation corpus as Tahoe-x1.
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 | Self-reported with cell-eval on Tahoe-100M; its VCC blind score is unpublished. · evidence |
Other evaluations (outside the 44)
- Developer-built evaluation tool (Own task set) Self-reported: The official page and the paper abstract frame the same result two ways (about 50% and over 30%) and claim it is the first model to beat linear baselines consistently
- Linear-baseline findings in independent evaluations (Comparison against baselines only) Third-party: Several independent evaluations and a challenge wrap-up find simple baselines hard to beat, cutting against the developer's claim
- scIB (Public benchmark not among the 44) Self-reported: Confirmed not a pairing: neither the paper nor the official page reports that benchmark
Developers report: learning state transitions substantially improves separation of perturbation effects and recovery of truly differential genes, and beats linear baselines consistently for the first time.
Independent evidence: No independent replication of this model was found; related independent evaluations and the perturbation-prediction challenge wrap-up conclude that simple baselines are hard to beat.
Training data
Group in Figure 7: Public data only.
| Class | Value | Detail |
|---|---|---|
| a. Public databases | Yes | About 167 million observational cells (including the institute's atlas and a public count collection), over 100 million perturbed cells across 70 cell lines, plus a public Perturb-seq dataset source |
| b. Private or licensed data | Not disclosed | |
| c. Own wet-lab data used for training | No | None for the developer: the main perturbation corpus was produced by a data company and donated to the institute's atlas (about 100 million cells, 50 cancer cell lines, 1,100 perturbations) and another set came from a reagent company; the institute's own challenge data is used only for evaluation. To avoid leakage, training used 20 of the cell lines with 5 held out source |
| d. Wet lab used only for validation | No | source |
Wet-lab validation
Every rate carries its numerator and denominator. Claims that could not be verified are shown as unverified, without numbers; different paper versions are separate entries.
Denominator basis: The same third-party perturbation corpus; the perturbation count is unified to 1,100 this round (the press release's ~1,200 is not used)
This item exists only to correct the published perturbation count and is not wet-lab validation of the model
Executor: Data produced by a third-party company, not a wet-lab campaign by the developer
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 institute's virtual-cell initiative includes a cell atlas and an annual perturbation-prediction challenge (first launched 2025-06-26 with a $100k grand prize and three sequencing and compute sponsors; a 2026 edition is announced); sibling models are Stack and Evo 2; the main perturbation corpus was donated by a data company that also builds Tahoe-x1.
- Related models: Stack, Tahoe-x1, Evo 2
Reading notes
- The same result is framed two ways on the official page and in the abstract; always say which framing is cited.
- Training and evaluation share one perturbation corpus; even with held-out cell lines this stays a within-distribution comparison.
- The main wet-lab data came from a third-party company and must not be attributed to the developer.
- No blind score for this model is published for the challenge, so its blind-test standing is unknown.
Caveats
- The only site pairing is self-reported under a modified protocol, and the blind-challenge part is unconfirmed.
- The journal online date has month-level evidence only.
- The perturbation count conflicts between sources; the lower framing is adopted this round.
- Weights carry a non-commercial licence.
Sources
- Journal paper 2026 (journal)
- Preprint (preprint)
- Developer launch page (official site)
- Code repository releases (repository)
- Challenge wrap-up (official site)
- Perturbation data donation announcement (official site)
Corrections
2 corrections
| Field | Was | Now | Basis | Found |
|---|---|---|---|---|
| training_data.c_own_wetlab_for_training | Perturbation data on about 1,200 drugs | 1,100 perturbations (about 1,200 appears only in press releases) | source | 2026-11-16 |
| benchmarks_reported | Challenge results (treated as a board) | Modified-protocol self-report; the challenge's blind-test scores are not published, marked unconfirmed | 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.