A world model of the virtual cell

The prospect of an AI-driven digital organism (AIDO), such as a virtual cell, has recently captured growing excitement and imagination across the AI and biology communities. We envision a virtual cell as a multi-modal, multi-scale, dynamic, and stateful computational system capable of simulating the activity and behavior of a living cell. Such a system could shift cell biology from trial-and-error experimentation with cell-culture models in the wet lab toward systematic simulation of combinatorial interventions in a digital laboratory. In this paper, we propose the world model as an operational framework for realizing this vision-an emerging AI paradigm that supports action-conditioned simulation, counterfactual reasoning and long-horizon planning in complex dynamic environments. In this formulation, a virtual cell world model (VCWM) represents a persistent cellular state and simulates its evolution under genetic, chemical, environmental, and other biological interventions. In contrast to predictive foundation models optimized for specific tasks or endpoints, a VCWM seeks to model the underlying cellular system, allowing interventions to propagate through an evolving state and generate coherent molecular, structural, interactional, and morphological outcomes over time. We outline an architecture, a data framework, training strategy, and evaluation principles for realizing this vision and discuss how existing biological foundation models can serve as its building blocks. We envisage that VCWMs could transform biological discovery from exhaustive experimental search toward structured navigation of learned cellular worlds, enabling counterfactual exploration, rational intervention design, and ultimately more predictive, designable, and programmable cell biology.