What it is
State is a transformer model that predicts how cells respond to genetic, chemical and signaling perturbations while accounting for cell-to-cell heterogeneity. It was trained on gene-expression data from over 100 million perturbed cells, with a cell-embedding component trained on observational data from 167 million cells. On large datasets it improved discrimination of perturbation effects by more than 30% and more accurately identified differentially expressed genes across perturbation types. Using its observational embeddings, it flagged strong perturbation effects in cellular contexts where no perturbations were seen during training.
Why it matters
Predicting perturbation responses in unobserved cell contexts is the central obstacle for a usable virtual cell, and models have generally failed to generalize there. A 30%+ gain in effect discrimination, plus the ability to call effects in novel contexts, is a meaningful move toward screening interventions computationally before running them at the bench. The accompanying Cell-Eval framework also gives the field a harder yardstick for these claims.
Underlined numbers link to their source. Every metric and quoted figure is listed under Sources and data below.
Filed undervirtual cell, single-cell, perturbation, transformer, Arc Institute