Overview
| Name | X-Cell · aliases: X-Cell, X-Cell-Ultra, X-Cell Mini |
|---|---|
| Full name | A product name with no expanded form |
| Ambiguous names (not auto-linked) | None |
| Family | Two sizes from one release: the largest at about 4.9 billion parameters and a small version initialised from a public single-cell model and not yet released
|
| Organization | Xaira Therapeutics |
| Organization type | AI-native biotech |
| Country | United States |
| Model class | Single-cell perturbation prediction (diffusion language model) |
| Domain | Single-cell & perturbation |
| First release | † partly verified · product update Basis: Developer press release, with the repository public a day earlier Note: Both dates rest on a company release and the repository, with no third-party corroboration source |
| Other milestones | |
| Latest update | † partly verified · no later event Basis: The release event itself No substantive event after the release was found as of 2026-11-16; the project page still lists weights and inference code as coming soon. Note: Same-day latest update means only that no later event was found source The most recent sourced event, not a claim that nothing changed since. |
| Status | Active · Released within 12 months, with no event found since |
| Openness | Code: Closed Weights: Closed Weights and inference code are marked as coming soon under a planned non-commercial share-alike licence; the test perturbation set is already public |
| Benchmarks reported | 1 counted (modified protocol 1) |
| Wet-lab validation | None recorded |
| Unverified fields | † first_release, release_milestones.paper_v1, release_milestones.code_release, latest_update, training_data.a_public, training_data.c_own_wetlab_for_training, other_evaluations |
| As of | 2026-11-16 |
What it does
Iteratively replaces control expression values with post-perturbation values using a diffusion language model, bringing in gene and protein interaction priors through cross-attention; the preprint reports a power law in perturbation prediction with scale and zero-shot transfer to two new cell types (second-hand). It differs from State and Tahoe-x1 in architecture and in training on data the developer produced itself.
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 held-out iPSC set; VCC participation unconfirmed. · evidence |
Other evaluations (outside the 44)
- Held-out iPSC perturbation set (Own task set) Self-reported: A held-out perturbation set built by the developer, with training and evaluation from the same source
- Zero-shot transfer to new cell types (Case study) Third-party: Zero-shot evaluation on two unseen cell types, per second-hand reports
Developers report: training on an in-house genome-wide perturbation dataset, a power law in perturbation prediction with model size, and a large margin over an unnamed runner-up on a held-out perturbation set.
Independent evidence: No independent evidence found. Weights are not public, no third-party evaluation was found, and it has no result on the public blind-test challenge
Training data
Group in Figure 7: Own wet-lab training data.
| Class | Value | Detail |
|---|---|---|
| a. Public databases | Not disclosed | The project page lists only the in-house perturbation set; the small version is initialised from a public single-cell model and so inherits its public data indirectly, with other public data use undisclosed source |
| b. Private or licensed data | No | |
| c. Own wet-lab data used for training | Yes | An in-house CRISPRi perturbation set of 25.6 million cells; the context count conflicts, with the project page and press release giving seven screens (9,735-18,924 perturbations each) and most reports giving 16 biological contexts source |
| d. Wet lab used only for validation | No |
Wet-lab validation
None. The developers report no wet-lab validation.
Background
- Organization: The developer is a US AI-first drug discovery company launched in April 2024, working on protein and antibody drugs and continuing a university lab's protein design lineage; headcount is not officially disclosed, with a third-party data provider estimating about 193 in March 2026.
- Funding: It launched with over $1B in Series A commitments co-led by two venture firms; the valuation is undisclosed.
- Products and partnerships: Its data products are two generations of perturbation atlas (about 8 million cells, then 25.6 million), and its models are this one at about 4.9 billion parameters plus a planned small version; it competes in perturbation prediction with State and Tahoe-x1.
- Related models: STATE, Tahoe-x1, scGPT / Geneformer
Reading notes
- The held-out result is self-reported and the project page names no competitor, so it is not evidence of a lead.
- The in-house dataset is a moat, but training and evaluation share a source and cross-cell-type generalisation has no independent reproduction.
- With no public blind-test result it cannot be compared with peers under one protocol.
- The context count appears as both seven and 16, so any citation must state which basis it uses.
Caveats
- Weights and inference code are unreleased, so nothing can be rechecked.
- Both first release and latest update rest on a company release and the repository.
- Sources disagree on the number of contexts in the training set.
- Its only pairing here is a self-built held-out set, not a challenge result.
- No event found since release is not the same as the project having stalled.
Sources
- Project page (official site)
- Company press release (press)
- Preprint 2026 (preprint)
- Code repository (repository)
Corrections
None recorded.
Part of the AI for Bio benchmarks survey · all 112 models · open in the explorer. Record as of 2026-11-16.