Artificial Intelligence Researchers Race to Build Functional ‘Virtual Cells’

Researchers are increasingly turning to artificial intelligence to unlock the remaining mysteries of cellular biology, with a growing effort to build what are known as virtual cells. These computer models aim to replicate the complex internal environment of a cell, potentially allowing scientists to predict how any cell type might respond to specific environmental changes or therapeutic interventions.
Proponents of the technology believe that virtual cells could eventually enable personalized medicine, allowing for simulated drug trials tailored to an individual’s unique biology. Others, such as Rory Stark, a computational biologist at Altos Labs, suggest these models could help researchers learn how to guide cells from unhealthy or aging states toward healthier, younger ones. The field has seen a surge in interest and investment; the nonprofit research organization BioHub recently announced commitments of nearly $2 billion toward data and modeling efforts, while companies like Xaira Therapeutics have secured significant private funding.

Efforts to refine these simulations are accelerating. The Arc Institute, a California-based research nonprofit, hosts the Virtual Cell Challenge, which drew over 1,600 teams this year to test the performance of various models. This follows the 2023 unveiling of GeneFormer, an early model created by Dr. Christina Theodoris and her colleagues, which was trained on data from approximately 30 million cells to identify gene activation patterns. While some researchers, including Simon Anders of the University of Heidelberg, have questioned the immediate advantage of such models over conventional statistical methods, developers maintain that these A.I. systems offer the unique ability to propose novel biological predictions that human researchers might not otherwise identify.
Technological strategies vary among developers. While many current models utilize large language models—similar to those powering chatbots—to predict cellular changes, other researchers, including those at the startup GenBio.AI, are implementing “world models.” These programs, already utilized in technologies like autonomous vehicles, are designed to track and predict how complex systems evolve over time. Dr. David Baker, a co-founder of GenBio.AI and Xaira Therapeutics, who shared the 2024 Nobel Prize for his work on protein structure prediction, views the transition to whole-cell modeling as a vital next step in biological research.
Despite the enthusiasm, significant hurdles remain. Experts note that current datasets, while large, may be insufficient to capture the vast complexity of human biology. Dr. Alex Marson of the University of California, San Francisco, is currently involved in the “Billion Cell Project,” a BioHub-backed initiative designed to generate more robust data by systematically observing how immune cells respond to the disruption of individual genes. Even so, some within the scientific community remain cautious. Alex Rives, head of science at Biohub, described the current state of the technology as “aspirational,” suggesting that even a billion-cell dataset may not be enough to reach the goal of a truly general model. Andrea Califano of Columbia University noted that while A.I. is poised to impact biology, the field currently faces risks of overhype.