Abhinav Adduri and colleagues develop State model for generalizable perturbation prediction

The ST model was trained on perturbation data from over 100 million cells, while the SE model learned from observational data covering 167 million cells, giving the system broad knowledge of cellular states. To benchmark State, the authors also built Cell‑Eval, an evaluation framework that goes beyond standard accuracy metrics to assess perturbation distinction, differential‑gene identification, expression‑change prediction and overall perturbation estimation. The work appears in Nature Machine Learning as subscription‑only content.