AI virtual cells: models, measurements and evidence for use

AI virtual cells seek to predict the consequences of intervention across cellular settings. Their methods span mechanistic simulators, compositional latent models, regulatory graphs, foundation representations and population generators. Human-derived assays and recent model-informed drug-development guidance sharpen a common question: what evidence makes a prediction useful for a particular decision? This critical narrative Review traces the field's historical branches and organizes current methods through the interfaces connecting baseline context, achieved intervention, biological response and measurement. A statistical description distinguishes dynamic path laws, endpoint distributions and feasible sets, while retaining assay noise, feature coverage and selection into the measured population. We compare the transfer assumptions of major model families and the evidence supplied by strong baselines, calibration and target-context holdouts. The fifteen biological challenges for generative cell biology extend this analysis from molecular responses to cellular function and patient-level outcomes. We then examine how cellular prediction can be combined with exposure models, organoids and microphysiological systems to build evidence for experimental and regulatory use.This deposit contains the full V10-restored Review (27 pages, five original figures), Supplementary Information (21 pages), and an independently compiled TeX source archive with bibliographic evidence records and editable vector figures. The supplement includes a 71-entry method catalog, the fifteen-challenge crosswalk, and dated regulatory and industrial source records. There are 152 unique cited records. The thematic search snapshot is October 15, 2026, with targeted checks through October 20.Preprint; not peer reviewed. This Review synthesizes existing literature and introduces no new biological dataset or comparative model benchmark. OpenAI Codex assisted literature organization, drafting and substantial revision, figure code and document preparation; its use is disclosed in the manuscript and source package. The author is responsible for the content.