Overview
| Name | TranscriptFormer · aliases: TranscriptFormer, TF-Metazoa, TF-Exemplar, TF-Sapiens |
|---|---|
| Full name | The name is itself the full name |
| Ambiguous names (not auto-linked) | None |
| Family | Not a family entry. Variants are treated as versions within one record |
| Organization | Chan Zuckerberg Initiative, Chan Zuckerberg Biohub |
| Organization type | Nonprofit institute |
| Country | United States |
| Model class | Single-cell FM (generative) |
| Domain | Single-cell & perturbation |
| First release | · paper revision Basis: bioRxiv preprint posted (registry record) source |
| Other milestones | |
| Latest update | · journal Basis: Science publication Journal version published in Science; GitHub v0.6.1 had been released on 2025-11-06 source The most recent sourced event, not a claim that nothing changed since. |
| Status | Active · Journal paper and code release within 12 months |
| Openness | Code: Open · MIT Weights: Open · MIT Code and weights released via PyPI and a public data registry under MIT (from the registry page and a second-hand source) |
| Benchmarks reported | 0 counted · none of the 44 |
| Wet-lab validation | None recorded |
| Unverified fields | † open_status.weights_license, open_status.code_license |
| As of | 2026-11-17 |
What it does
Models each cell's transcriptome autoregressively as a "cell sentence", trained across 12 species; it supports cross-species cell-type transfer, zero-shot disease-state detection and generation of cell expression, with no perturbation evaluation. Unlike C2S-Scale, which turns expression into text and fine-tunes a general LLM, it is a dedicated single-cell autoregressive model.
Benchmarks reported
None of the 44 benchmarks. This model is on the zero-pairing list.
How it evaluates itself: Cross-species cell-type classification; Spermatogenesis cross-species mapping; Human cell type and disease state; Gene-gene and TF analyses
Other evaluations (outside the 44)
- Cross-species cell-type classification (Own task set) Self-reported: Linear probes on 5 out-of-distribution species against UCE, ESM2-CE, scGPT and Geneformer; author-reported
- Spermatogenesis cross-species mapping (Own task set) Self-reported: Label transfer across 9 vertebrate species; author-reported
- Human cell type and disease state (Own task set) Self-reported: Tabula Sapiens 2.0 across 26 tissues plus infected vs uninfected lung cells; author-reported
- scIB (Public benchmark not among the 44) Self-reported: Not a pairing: scIB-family data are used but no scIB metrics are reported
- Gene-gene and TF analyses (Case study) Self-reported: Qualitative recovery of known interactions against STRING and cell-type TF prediction
Developers report: a generative model pretrained on about 112M cells from 12 species transfers cell-type labels across roughly 685 million years of divergence and detects disease states zero-shot, beating UCE and other baselines.
Independent evidence: No independent evidence found. No independent third-party evaluation of its cross-species results was found
Training data
Group in Figure 7: Public data only.
| Class | Value | Detail |
|---|---|---|
| a. Public databases | Yes | Mainly CELLxGENE, about 112M cells across 12 species source |
| b. Private or licensed data | No | source |
| c. Own wet-lab data used for training | No | source |
| d. Wet lab used only for validation | No |
Wet-lab validation
None. The developers report no wet-lab validation.
Background
- Organization: The Chan Zuckerberg Initiative is a philanthropy that in 2025-11 put Biohub at the centre of its biology and AI work, with CZI as supporting infrastructure; Biohub has sites in San Francisco, New York and Chicago, and the EvolutionaryScale team has joined it.
- Funding: Founders' personal philanthropic funds; Biohub says it plans to reach 10,000 GPUs by 2028, with no AI-specific amount published.
- Products and partnerships: TranscriptFormer is the first step of CZI's virtual-cell strategy; its data source CELLxGENE is shared with STATE, Stack and Tahoe-x1.
- Related models: C2S-Scale, STATE, Stack, Tahoe-x1, scGPT / Geneformer
Reading notes
- With no scIB or perturbation scores, it cannot be compared directly with Geneformer, scGPT or STATE.
- Its strength is cross-species transfer, a task without a shared third-party board.
- Training is all public data, and its value for drug discovery is untested in the wet lab.
Caveats
- All evaluations are the authors' own task sets with no third-party reproduction.
- Site figures come from the bioRxiv text; the Science version is paywalled and takes precedence.
- Licence information partly comes from a second-hand source.
Sources
- Science paper (journal)
- bioRxiv preprint (preprint)
- Release list (repository)
- Package registry page (package registry)
- Public data registry page (official site)
Corrections
None recorded.
Part of the AI for Bio benchmarks survey · all 112 models · open in the explorer. Record as of 2026-11-17.