A Roadmap to Build an AI-Powered Virtual Yeast That Simulates a Eukaryotic Cell

A team of scientists has laid out an ambitious blueprint to construct a virtual yeast cell—an AI-driven simulation that captures how a living eukaryotic cell responds to genetic changes, drugs, and environmental shifts. If successful, the effort could accelerate target discovery, optimize biofuel and pharmaceutical production, and provide a testbed for understanding fundamental cellular logic shared with human cells.

Writing today in Nature, researchers led by Tiannan Guo at Westlake University, with first author Liujia Qian, propose a detailed strategy to build a ‘virtual yeast’ using the budding yeast Saccharomyces cerevisiae. The perspective article does not announce a finished model; rather, it describes a multi-year framework that integrates artificial intelligence, massive multi-omics datasets, and automated experimentation to gradually construct a predictive digital simulation of a eukaryotic cell.

Rather than attempting to model every molecule at once, the team divides the cell into eight interconnected functional modules, each capturing a core cellular process—such as mitochondrial energetics, membrane trafficking, biosynthetic networks, or stress responses. Each module is built as a specialized AI tool trained on relevant data, from imaging to proteomics. A large language model (LLM) acts as an orchestrator, dynamically routing a biological question to the appropriate modules and coordinating their outputs. This agent-like design allows the system to reason across scales, predicting how genetic or environmental perturbations propagate across cellular functions.

The virtual yeast rests on three data pillars: prior biological knowledge (curated from decades of yeast research), spatially resolved cellular architecture, and dynamic multi-omics measurements collected after systematic perturbations. A key component is a closed-loop ‘AI flywheel’ in which the virtual model identifies experiments that would maximally improve its predictions, robotic platforms perform those experiments, and the resulting data are fed back to refine the model—all with progressively less human intervention.

As a first implementation step, the team describes a prototype metabolic and synthetic biology module. They selected 12 genetically diverse yeast strains from a panel of 969, subjected them to over 200 environmental and chemical perturbations, and generated more than 15,000 time-resolved proteomic profiles and 5,000 metabolomic measurements, alongside growth curves. This dataset forms the training foundation for the prototype AI framework WAY-AL v0.1, which will guide subsequent rounds of experimentation. While this module focuses on predicting metabolic fluxes and optimizing compound production, it serves as a template for expanding to other cellular functions.

The authors are careful to emphasize that the virtual yeast is not a digital twin that replicates every detail of a living cell. Instead, it is a pragmatic, function-centred tool that can focus on specific tasks—such as predicting metabolite yields or identifying how stress granules remodel the proteome—while remaining experimentally testable. Major challenges remain, including integrating data across vastly different modalities, capturing single-cell variability, and handling the pleiotropic effects of genetic background.

The consortium behind the project envisions a 5- to 10-year development timeline, with progressive integration of additional modules and validation against real-world experiments. If successful, the yeast platform could provide a transferable blueprint for building virtual cells of more complex eukaryotes.

Reference

Qian L, Zhou Z, Zhou P, et al. Towards the construction of a virtual yeast. Nature. 2026;655:59–70.
DOI: https://doi.org/10.1038/s41586-026-10574-9