rewire.it
model · family

scGPT

scGPT learns representations of genes and cells from single-cell measurements. Its pretrained checkpoints support task-specific adaptation.

0 evaluations · 0 metric rows

Shared profile: scGPT. This page retains the exact record and its evaluation context.

At a glance

Explanatory profile: source reviewed · Automated source review, 2026-09-16. This does not change the review status of its results.

Inputs, outputs and configuration
PropertyDescription and evidence
Whole-human trainingRepository reports 33 million normal human cellsbowang-lab/scGPT official source · README.md: Pretrained scGPT checkpoints and Tutorials; scgpt/model/model.py at cebd6fae655b9c585a4807daa3ac31bb764f06b4, lines 79–188
Configuration in this recordwhole-humanbowang-lab/scGPT official source · README.md: Pretrained scGPT checkpoints and Tutorials; scgpt/model/model.py at cebd6fae655b9c585a4807daa3ac31bb764f06b4, lines 79–188
Model typeNot extracted or verified for this record.
Context limitsNot extracted or verified for this record.
AccessNot extracted or verified for this record.
Code licenceNot extracted or verified for this record.
Weights licenceNot extracted or verified for this record.

How it works

Conceptual procedure

Schematic of the documented input, computation and output; not an executable configuration.

Conceptual procedureGene IDs and expression. Then: Gene / value encoders. Then: Transformer. Then: Cell and gene representations. Then: Adapted task outputGene IDs and expressionGene / value encodersTransformerCell and gene representationsAdapted task output
Read the diagram as text
  1. Gene IDs and expression
  2. Gene / value encoders
  3. Transformer
  4. Cell and gene representations
  5. Adapted task output
bowang-lab/scGPT official source · README.md: Pretrained scGPT checkpoints and Tutorials; scgpt/model/model.py at cebd6fae655b9c585a4807daa3ac31bb764f06b4, lines 79–188

Gene identifiers and expression values are encoded together and processed by a transformer. The resulting representations support cell embeddings or task heads. Vocabulary, preprocessing and the selected checkpoint must accompany any result.

bowang-lab/scGPT official source · README.md: Pretrained scGPT checkpoints and Tutorials; scgpt/model/model.py at cebd6fae655b9c585a4807daa3ac31bb764f06b4, lines 79–188

Benchmarks and results

Release 2026-09-16-d74d282221a9 · 0 evaluations · 0 metric rows. Different protocols are not a single leaderboard.

No evaluations linked in this release.

Strengths and limitations

Strengths supported by sources

  • The repository provides whole-human and specialised checkpoints, plus workflows for annotation, integration and perturbation tasks.bowang-lab/scGPT official source · README.md: Pretrained scGPT checkpoints and Tutorials; scgpt/model/model.py at cebd6fae655b9c585a4807daa3ac31bb764f06b4, lines 79–188

Limitations and conditions

  • Whole-human, organ-specific and continually pretrained checkpoints are different configurations. A pretraining claim does not establish transfer performance in a new cell population.bowang-lab/scGPT official source · README.md: Pretrained scGPT checkpoints and Tutorials; scgpt/model/model.py at cebd6fae655b9c585a4807daa3ac31bb764f06b4, lines 79–188
Profile review details

Primary project documentation or paper inspected for the explanatory claims and cited locations. Reviewed coverage concerns this narrative, not complete metadata, independent reproduction or a performance ranking.

Stable record: catalog-model-scgpt

Sources and history

Release 2026-09-16-d74d282221a9 · Record review: discovered

Download this release
Technical metadata and extraction receipts

Stable ID: catalog-model-scgpt

areas
cells-tissues
method types
foundation model
entity level
family
version
whole-human
reported name
scGPT
access
Public code and downloadable checkpoints; use the unfine-tuned whole-human model for a new task.
method type
foundation model
missing metadata
checkpoint revision: not_yet_extracted; training data: not_yet_extracted; licence: not_yet_extracted
Related records

Suggest a correction