Building virtual cell dynamics that evolve with biological measurements

CellCompass, an agentic AI for virtual-cell dynamics, achieved the highest aggregate performance across six dynamical tasks on DynBench100, a benchmark of 100 simulated single-cell time-course scenarios.

• Why it matters: Virtual cell models from single-cell measurements must predict time evolution, but new experiments can impose dynamical relationships existing models cannot represent. Translating these measurements into dynamics requires deciding what relations to model, where to include them, and how to validate recovery.
• What they did: CellCompass couples a scientific agent with an extensible dynamical engine for state transitions, population change, fate, and molecular readouts, reusing existing formulations or designing new ones when constraints arise. It was evaluated on DynBench100 and four real single-cell studies using metabolic RNA labeling, clonal barcodes, and time-resolved CITE-seq.
• The result: In drug-response data, metabolic RNA labeling reduced pulse-control identifiability error by 31.1%, and in hematopoietic lineage tracing clonal barcodes improved clone-fate agreement by 54%; in time-resolved CITE-seq, it reduced future surface-protein distribution error by 13.2% for RNA-initialized trajectories. These results enable staged response-program insights, early lineage priming detection, and lineage-specific temporal offsets.