What it is
scGPT is a generative pretrained transformer over 33M+ single cells that transfers to cell-type annotation, batch integration, perturbation prediction and gene-network inference. It helped popularize the 'foundation model' framing for single-cell genomics.
Evidence trail
BioAtlas keeps the path from source to decision visible. A connection records provenance; it does not imply that evidence is sufficient for every context.
Model passport
How scGPT represents biology
Category is navigation. These fields describe the model-specific computational transformation and deliberately override broad category defaults.
Biological scale
Modalities & tasks
Registry, claims and frontier intelligence
Version history not yet curated
1 version record · release year not yet normalized. Model-family identity remains separate from capability and access changes.
Explore version lineage →1 normalized claim
Single-cell representation learning · Single-cell downstream tasks
Open claim intelligence →1 connected frontier
Virtual cells · Recent preprint
Inspect research horizon →Connected research frontiers
These records describe active research directions, not guaranteed capabilities of this model. Evidence stages and unresolved questions are preserved separately.
Virtual cells that predict perturbation response
Arc Institute · Virtual Cell research community · 2026-04-30Can models forecast how cell populations respond to unseen drugs, gene edits, cytokines and environmental changes across biological contexts?
Evidence boundary and unresolved questions
Recent strict evaluations show marked performance drops under unseen contexts and metric-dependent rankings; simple baselines remain competitive on some global trends.
- Can models recover perturbation-specific mechanisms rather than average expression shifts?
- How should cell distributions, dose and time be represented?
- Which metrics predict prospective experimental usefulness?
virtual cells · perturbation · single cell · OOD generalization · world modelsOpen frontier record →Inputs and outputs
Inputs
Single-cell or perturbation dataOutputs
Cell statesPerturbation predictionsScientific and technical profile
Scientific principles
Technology
Scientific lineage
These are transparent concept matches—not claims that one scientist alone caused this model. Each connection is based on the model’s recorded domain, scientific principles, technical terms or an explicit lineage link.
Gene regulation and the operon model
François Jacob & Jacques MonodTarget biology, perturbation models, transcriptomic response prediction and virtual cells all require an explicit model of regulated gene programs.
Transformer self-attention
Ashish Vaswani and colleaguesProtein, genome, molecule and single-cell foundation models use attention to learn dependencies across biological sequences and multimodal inputs.
Information, entropy and communication
Claude E. ShannonSequence modelling, cross-entropy training, language models, mutual information and representation learning all use Shannon’s framework.
The epigenetic landscape and cell-fate trajectories
Conrad H. WaddingtonSingle-cell embeddings, trajectory inference, reprogramming and virtual-cell models often represent cell identity as movement through a learned state landscape.
Concerted allostery
Jacques Monod, Jeffries Wyman & Jean-Pierre ChangeuxAllosteric drug design exploits remote pockets to modulate function, selectivity and resistance without competing at the active site.
Programmable CRISPR–Cas genome editing
Jennifer A. Doudna & Emmanuelle CharpentierCRISPR enables target validation, disease models, perturbation atlases, functional genomics and gene-editing therapeutics.
Evaluation evidence
Task-specific evidence only; not comparable as a universal leaderboard score.
Single-cell downstream tasks
Version history not yet curated · Single-cell downstream-task benchmarksPeer-reviewed evaluation across representation and downstream single-cell tasks.
Claim caveats
- Performance varies by preprocessing, batch correction and downstream task.
- Representation quality is not equivalent to causal perturbation prediction.
Known limitations
- Performance depends on the evaluation dataset and operating conditions.
- Task-specific benchmark results should not be compared across unlike domains.
- Outputs require task-specific scientific and experimental validation.
Milestones
Pretrained on 33M+ human cells.
One of the most-used open single-cell foundation models.