AI for Bio · Model

scGPT / Geneformer

First-generation single-cell foundation-model baselines for annotation, perturbation and integration

Single-cell & perturbationAcademic labCanada, United StatesMaintained only中文

Overview

NamescGPT / Geneformer · aliases: scGPT, Geneformer, Geneformer V2
Full nameNo separate expanded name
Ambiguous names (not auto-linked)None
FamilyTwo first-generation single-cell foundation models grouped as one baseline entry; teams, corpora and versions are separate
  • scGPT ·
  • Geneformer · †
OrganizationUniversity of Toronto, University Health Network, Vector Institute, Gladstone Institutes, UCSF
Organization typeAcademic lab
CountryCanada, United States
Model classSingle-cell foundation model
DomainSingle-cell & perturbation
First release† partly verified · code release
Basis: Earliest public artifact: scGPT GitHub v0.1.2 (2023-05-09); preprint date not re-checked source
Other milestones
  • Code release: · scGPT GitHub v0.1.2 source
  • Weights release: · Geneformer V2 weights added on Hugging Face source
Latest update · new weights
Basis: Geneformer V2 models added to the Hugging Face repository
Geneformer V2 weights released; scGPT's last tag is v0.2.4 (2025-03-31); later commits are maintenance source
The most recent sourced event, not a claim that nothing changed since.
StatusMaintained only · Only maintenance commits after 2025-06
OpennessCode: Open
Weights: Open
Both open source; Geneformer weights on Hugging Face
Benchmarks reported2 counted (self 1, retrospective 1, unconfirmed 1)
Wet-lab validationNone recorded
Unverified fields† first_release, family.members[1].first_release, benchmarks_reported[1]
As of2026-11-16

What it does

Two first-generation single-cell foundation models: Geneformer rank-encodes genes within each cell, scGPT uses generative pretraining; both are fine-tuned for annotation, perturbation prediction and batch integration. This site uses them mainly as baselines.

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. 2 of 3 pairings are counted.

BenchmarkRelationEval set / protocolSource typeNote and evidence
09 scIB (+Open Problems)Developer report, official protocolPBMC 10KSelf-reportedscGPT only (PBMC10K), self-reported; Geneformer not applicable; family entry †partly verified · evidence
24 Virtual Cell Challenge and perturbation benchmarksRetrospective dataset use—Third-partyIndependent third-party evaluations on Norman and Adamson data. · evidence
21 DREAM5 / BEELINE / CausalBenchUnconfirmed (not counted)—Unconfirmed sourceA third-party attention-edge study, read in abstract only. †disputed · evidence

Other evaluations (outside the 44)

None recorded.

Developers report: Source states no developer claim

Independent evidence: Several independent 2025 studies find they do not reliably beat simple baselines on perturbation and zero-shot tasks

Training data

Group in Figure 7: Public data + wet lab for validation only.

ClassValueDetail
a. Public databasesYesGeneformer: ~29.9M human cells (later ~95M); scGPT: ~33M CELLxGENE cells
b. Private or licensed dataNo
c. Own wet-lab data used for trainingNo
d. Wet lab used only for validationYesGeneformer had one cell-level validation; scGPT had none

Wet-lab validation

None. The developers report no wet-lab validation.

Background

  • Organization: Both academic: scGPT from Toronto/UHN/Vector Institute, Geneformer from Gladstone Institutes and UCSF
  • Funding: Geneformer funding included NIH and American Heart Association grants
  • Products and partnerships: Xaira Therapeutics describes scGPT as the starting point of its X-Cell model
  • Related models: X-Cell, STATE, Tahoe-x1, TranscriptFormer, C2S-Scale, CellFM

Reading notes

  1. Appears far more often as a baseline than as a recommended method
  2. Independent evaluations show perturbation prediction no better than deliberately simple baselines

Caveats

  1. Family entry: the two models were released separately; read dates per model
  2. Exact preprint and journal dates not re-checked

Sources

  1. Hugging Face page ctheodoris/Geneformer (repository)
  2. GitHub repository repos/bowang-lab (repository)
  3. Journal article DOI 10.1038/s41586-023-06139-9 (journal)
  4. Journal article DOI 10.1038/s41592-024-02201-0 (journal)
  5. scGPT bioRxiv preprint (preprint)
  6. Nat Methods 2025 independent evaluation (third-party paper)

Corrections

None recorded.

Part of the AI for Bio benchmarks survey · all 112 models · open in the explorer. Record as of 2026-11-16.