What researchers are trying
STATE and successors model transitions between cellular states, while newer conditional-transport and in-context systems target larger perturbation atlases and unseen conditions.
Virtual cells that predict perturbation response
Can models forecast how cell populations respond to unseen drugs, gene edits, cytokines and environmental changes across biological contexts?
Why it matters
Reliable virtual cells could prioritize experiments, reveal mechanisms and reduce the search space before high-dimensional wet-lab screens.
Evidence boundary
Recent strict evaluations show marked performance drops under unseen contexts and metric-dependent rankings; simple baselines remain competitive on some global trends.
Organizations represented
Arc Institute · Virtual Cell research community
Demonstrated evidence
What has actually been shown.
- Large-scale benchmark and challenge ecosystems now compare perturbation models across Tahoe-100M and other datasets.
Unresolved questions
- 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?
Signals to watch next
- Virtual Cell Challenge results
- Cross-dataset OOD tests
- Mechanistic explanations
- Prospective experimental validation
Connected evidence graph
Related BioAtlas model passports.
State / Stack
The first virtual-cell model predicting how cells react to drugs.
4/7 evidence fields documentedCZI Virtual Cells (rBio, TranscriptFormer)
A moonshot to build AI models of every human cell.
1/7 evidence fields documentedscGPT
A GPT for the language of single cells.
4/7 evidence fields documentedGeneformer
A transfer-learning foundation model of gene networks.
2/7 evidence fields documentedPrimary and evaluation sources