Towards the Construction of a Virtual Yeast: AI-Driven Whole-Cell Modeling Advances
核心洞察
A comprehensive perspective article in Nature outlines the design principles and collaborative strategies needed to build AI-driven virtual cells (AIVCs), using yeast as a model system.
The virtual yeast initiative leverages decades of foundational datasets, including nearly 1 million mapped genetic interactions and a near-complete protein–protein interaction map.
Key AI frameworks such as State (搜索), ProteinTalks (搜索), and AlphaFold 3 (搜索) are highlighted as critical tools for predicting cellular responses, protein dynamics, and biomolecular interactions.
The construction of a fully predictive, multiscale computational model of a living cell—long a holy grail of systems biology—is now within reach, according to a landmark perspective published in Nature. The article, titled "Towards the construction of a virtual yeast," lays out a comprehensive roadmap for building an AI-driven virtual cell (AIVC) using Saccharomyces cerevisiae, the most extensively characterized eukaryotic model organism, as the proving ground.
The perspective defines the AIVC concept and articulates the design principles, data requirements, and collaborative strategies necessary to realize AI-driven simulations that span from atomic-level protein structures to whole-cell physiology. "This perspective article defines the concept of the AIVC and outlines the design principles and collaborative strategies needed to realize AI-driven, multiscale simulations of a living system," the authors note, referencing foundational work by Bunne et al. (2024).
A Foundation Built on Decades of Data
The virtual yeast effort stands on an unparalleled foundation of experimental datasets. Landmark studies have mapped nearly 1 million genetic interactions in yeast, revealing "the global wiring diagram of cellular function and establishing a quantitative framework for decoding genotype-to-phenotype relationships" (Costanzo et al., 2016). Complementing this, high-throughput affinity purification–mass spectrometry has generated a near-complete yeast protein–protein interaction map, "uncovering the dense and modular architecture of the cellular interactome" (Michaelis et al., 2023).
Population-scale genomics further enriches the resource base: genome evolution analysis across 1,011 S. cerevisiae isolates has revealed "the evolutionary trajectories and domestication history of S. cerevisiae, providing a comprehensive resource for genotype–phenotype studies" (Peter et al., 2018). More recently, telomere-to-telomere assemblies of 142 strains have characterized the genome structural landscape of the species (O'Donnell et al., 2023).
AI Frameworks Powering the Virtual Cell
Central to the AIVC vision are cutting-edge AI frameworks capable of predicting cellular behavior across diverse contexts. The perspective highlights State (搜索), "a scalable AI framework trained on perturbation transcriptomic data, from over 100 million cells, that predicts cellular responses across unseen contexts, advancing the dynamic modelling of perturbation effects central to AIVC development" (Adduri et al., 2025). Complementing this transcriptome-centric approach, ProteinTalks (搜索)—"a neural ordinary differential equation-based foundation model trained on 38 million perturbed protein measurements"—learns cellular protein network dynamics to predict drug efficacy, synergy, and resistance, "laying a proteome-centric foundation for virtual cell development" (Sun et al., 2025).
Structural biology has been revolutionized by AlphaFold 3 (搜索), which enables "accurate structure prediction of biomolecular interactions" (Abramson et al., 2024), while Evo 2 facilitates "genome modelling and design across all domains of life" (Brixi et al., 2026). These tools, combined with transfer learning approaches that "enable predictions in network biology" (Theodoris et al., 2023), form the computational backbone of the virtual yeast project.
Integrating Multiscale Data
The perspective emphasizes the "three data pillars" framework proposed by Qian, Dong, and Guo (2025): a priori knowledge, static architecture, and dynamic states. These pillars serve as the foundation for building AIVCs and are integrated through closed-loop active learning systems that "autonomously refine virtual cell models."
Spatial omics technologies are critical to capturing the dynamic architecture of the cell. Recent advances include SUM-PAINT, a high-throughput super-resolution method "achieving virtually unlimited multiplexing at sub-15-nm resolution and revealing a new VGLUT1+ Gephyrin+ synapse subtype" (Unterauer et al., 2024), and PLATO, "a microfluidics-based and AI-based framework enabling high-resolution spatial proteomics across whole tissues and revealing distinct tumour subtypes in human breast cancer" (Hu et al., 2025).
The pioneering whole-cell model of Mycoplasma genitalium—"the first mechanistic whole-cell model that integrates all molecular processes... using diverse mathematical formalisms to unify fundamentally different cellular processes and experimental measurements, enabling genotype-to-phenotype prediction" (Karr et al., 2012)—serves as a proof of principle for the yeast endeavor.
From Yeast to Human Biology
The implications extend far beyond yeast. The perspective notes that systematic humanization of yeast genes has revealed "conserved functions and genetic modularity" (Kachroo et al., 2015), positioning humanized yeast as a platform "to model human biology, disease and evolution" (Kachroo et al., 2022). With yeast already serving as a chassis for the complete biosynthesis of cannabinoids and the semi-synthetic production of artemisinin, a predictive virtual yeast model could accelerate drug discovery, metabolic engineering, and our fundamental understanding of eukaryotic cell biology.
The roadmap also envisions autonomous closed-loop experimentation, building on demonstrations such as "a mobile robotic chemist" that established "a closed-loop system for self-driven experimentation, foreshadowing active-learning frameworks for AIVC evolution" (Burger et al., 2020).