Ask a cell a question — knock out this gene, add that drug — and it answers by changing which genes it expresses. The dream of the virtual cell is to predict that answer before running the experiment. Today’s digest follows the models learning to do it. Dixit et al.’s Perturb-seq built the data engine, measuring how thousands of pooled perturbations reshape single-cell transcriptomes. Norman et al. mapped genetic-interaction ‘manifolds’ of cell states. Lotfollahi et al.’s scGen was first to generalize ‘out-of-sample,’ predicting responses ‘across cell types, studies and species.’ Their CPA predicts response ‘for unseen dosages, cell types, time points, and species.’ Roohani et al.’s GEARS predicts outcomes of gene combinations ’that were never experimentally perturbed.’ And Piran et al.’s biolord generates ’experimentally inaccessible samples.’ State plus perturbation, in silico.
