A new initiative involving some of the world's largest technology companies aims to take artificial intelligence to a whole new scale. Google, Meta, and the U.S. government are collaborating on an effort to create virtual models capable of predicting the behavior of real cells.

Led by the Chan Zuckerberg Biohub , this initiative brings together organizations such as Google DeepMind, Isomorphic Labs, and Meta. With combined funding for infrastructure, computing resources, and other assets, the effort represents an investment of nearly $1.8 billion—yet the goal is not simply to build a larger database. The aim is to generate the massive amount of biological data needed to develop what researchers call "virtual cells."

Currently, determining how a cell will respond to a specific change often requires conducting a physical experiment: modifying a gene, applying a substance, or introducing a disturbance, and then observing the outcome. A virtual cell would aim to replicate this process within a computer; a researcher could input a cell's initial state and ask the model to predict the likelihood of a specific outcome—such as what might happen if a gene were deactivated, a protein altered, or another type of intervention performed.

Artificial intelligence would then attempt to predict a cell's response before the actual experiment is conducted; this could allow scientists to virtually test a vast number of hypotheses and take only the most promising ones to the laboratory. However, there is a massive challenge: cells are incredibly complex. Even today's large databases—containing information on hundreds of thousands of cells—may not suffice to train an AI to predict their behavior with true reliability.

It is estimated that models of this kind may need to learn from billions, or perhaps even trillions, of cells; consequently, a fundamental part of the project will be generating new biological datasets on a much larger scale. The plan is to make a major initial dataset available within roughly a year, with the goal—within about five years—of achieving models far more capable of predicting how different cells respond to specific interventions.

This does not mean that a perfect digital copy of a human cell already exists; we are still talking about constructing the necessary data and models to reach that point, and actual experiments would remain crucial for confirming the artificial intelligence's predictions. However, if this technology reaches the expected level, it could transform one of the most time-consuming stages of biological research: in the future, before testing a hypothesis on a real cell, scientists might first be able to determine the likely outcome by testing that same idea on its virtual counterpart.