Why this evidence?
It is linked to a specific model version, scientific task, dataset, split, metric and source. That makes the claim inspectable rather than a detached marketing score.
State / Stack: Cell perturbation prediction
Open evaluation of cell-state prediction under perturbation.
Model versionStack
TaskCell perturbation prediction
DatasetPerturbation prediction datasets
SplitPerturbation-prediction datasets
MetricCell-state prediction
Replicationemerging
Reported byModel developers
Review statuscurated
Evidence confidence · limited
Confidence is multidimensional, not a universal model score.
Evidence completenessstrongModel version, task, dataset, split, metric, source and provenance fields.
Independent validationlimitedemerging
Source qualitylimitedPreprint / open evaluation
Reproducibility evidencelimitedReflects documented replication status, not a universal reproducibility score.
Version specificitystrongStack
Context applicabilitystrongDepends on task, split and explicit caveats; users must still validate their own context.
Contradiction reviewclearNo direct contradiction signal is currently queued.
BioAtlas reports evidence dimensions separately so a strong source cannot hide weak applicability, incomplete replication or unresolved contradiction.
Why?
Why should this evidence influence a decision?
Why not a universal score?
Performance can change with dataset, split, preprocessing, metric and context of use. BioAtlas therefore keeps confidence dimensions separate.
What could change the conclusion?
Independent replication, a better matched prospective dataset, a version change, a contradictory result or a more relevant validation protocol can reopen this evidence record.
Evidence boundary
What this claim does not prove.
- Cell line, tissue, dose and timepoint shifts can dominate performance.
- In-silico perturbations require experimental validation.
BioAtlas groups benchmark claims only when task, dataset, split, metric and protocol context align. This record is not a universal model score.
Model context
State is Arc Institute's first virtual-cell model: given a starting transcriptome and a perturbation (drug, gene edit or cytokine), it predicts how gene expression will shift. Trained on ~170M observational and 100M+ perturbational cells across 70 contexts, it pairs a State Embedding module with a State Transition transformer. Its 2026 successor, Stack, learns cell biology in-context to generalize to unseen conditions.
Full evidence passport →Known model 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.