A virtual cell is an AI model that predicts how a living cell will respond to a change, such as a drug, a gene edit or a disease state, before anyone runs the experiment.
If it worked reliably, it would turn much of early drug discovery from trial and error into simulation. That promise has pulled in nonprofits, big tech, pharma and a wave of startups.
However, the field’s own benchmarks show that today’s models are still far from that goal.
The Virtual Cell Tracker follows the field as it moves: the models, the datasets that feed them, the deals that fund them, and the evidence on whether any of it works. It is updated in place, so the tracker is always the current picture. Every figure links to its source. The tracker is free to read with a sign-up on BiopharmaTrend.
The latest big story: a $1.8 billion bet that data, not models, is the bottleneck
In October, Biohub (the Chan Zuckerberg science organization) expanded its Virtual Biology Initiative to $1.8 billion. The US Department of Energy, NIH, Google DeepMind, Isomorphic Labs and Meta joined.
The goal is open, standardized data for training virtual cells. There is a catch. Biohub’s head of science, Alex Rives, told Reuters that commercial funders get an embargo period with the data before it becomes public, and, as reported by The Next Web, he told Axios the embargo lasts one year. Rives expects a first dataset in about a year and accurate predictive models within five.
This is the clearest sign yet that the field sees data as its bottleneck. Academic groups and startups gain in the long run, because they could never produce perturbation data at this scale themselves. In the short run, DeepMind, Isomorphic and Meta get a hand in shaping the data and a year’s head start on training with it.
What the money can’t buy is proof that more data solves generalization. This year’s independent benchmarks and follow-up studies show that models still struggle once they’re tested on cell types and perturbations they haven’t seen.
How we got here
The idea of simulating a whole cell goes back to Masaru Tomita’s 2001 call to treat whole-cell simulation as “a grand challenge of the 21st century.”
The landmark came in 2012, when Markus Covert’s group at Stanford published a model of Mycoplasma genitalium covering all 525 of its known genes. Mechanistic platforms such as VCell (launched in 1999), CompuCell3D, Morpheus, CHASTE and BioDynaMo let researchers simulate reaction kinetics, diffusion and tissue growth from hand-built equations. Those approaches broke down on real complexity: processes spanning atomic to cellular scales, highly nonlinear dynamics, and too many unknown parameters. After 2012 the field went largely quiet for a decade.
The revival is data-driven rather than equation-driven. Cheap single-cell sequencing, pooled CRISPR screens (Perturb-seq) and transformer models made it possible to learn cell behavior directly from hundreds of millions of measured cells.
A December 2024 Cell perspective led by Charlotte Bunne, with co-authors from Stanford, the Chan Zuckerberg Initiative, Genentech, Google and others, named the goal: the AI Virtual Cell.
It is a set of linked foundation models that represent cells across scales, predict their response to perturbations, and run experiments in silico.
Since then, the race has split into three parts: who has the data, whose models generalize, and who can prove either.
Access The Virtual Cell Tracker for free at BiopharmaTrend.