Inside the Push to Build "Virtual Cells" That Could Predict Disease Before It Strikes

Imagine if a doctor could test a cancer drug on your tumor before ever giving you a dose, not in your body, but inside a computer that has learned to behave like your cells do. No side effects, no wasted time, just a prediction of what might work. That is the promise behind "virtual cell models," an idea that is quickly moving from science fiction toward something closer to a lab tool.

The Chan Zuckerberg Initiative (CZI) is trying to build exactly that: an AI-powered platform for biology that could let researchers run experiments on simulated cells instead of real ones, at least as a first step. Patricia Brennan, CZI's vice president of science technology, and Theofanis Karaletsos, the organization's head of AI for science, discussed the effort at a fireside chat during the 2024 STAT Breakthrough West Summit in San Francisco.

The comparison they reached for was familiar technology most of us already carry. An Apple Watch tracks your heart rate and sleep. ChatGPT can draft a meal plan tailored to your health goals. CZI's vision applies that same logic to the inner workings of a cell, giving scientists a conversational tool that can test a hypothesis, suggest tweaks to an experiment, or help a clinician weigh treatment options based on a tumor's specific mutations.

It is an ambitious idea with a correspondingly ambitious goal attached. CZI's stated mission is to help cure, prevent, or manage all diseases by the end of this century. Virtual cells are being framed as one of the tools that could get researchers there faster.

What a Virtual Cell Actually Does

A virtual cell is not a single image or a static model. It is meant to be a dynamic simulation trained on enormous amounts of biological data, covering genomes, the molecules cells use to build proteins, and the proteins themselves. The goal is to predict how both healthy and diseased cells behave, alone and in groups, and how they respond to outside forces like a new drug or a change in environment.

Think of it less like a photograph and more like a weather model. Meteorologists do not have a crystal ball, but they can run simulations based on huge amounts of atmospheric data to forecast what a storm will likely do. CZI wants something similar for biology: a tool that forecasts cellular behavior with enough accuracy to guide real experiments, cutting down on the cost, time, and risk of testing everything in a lab dish or a living organism.

That lower-risk, lower-cost testing ground is the real selling point here. Lab experiments take months. Clinical trials take years and carry real stakes for patients. A reliable simulation could let scientists try out far more scenarios before committing resources to the ones most likely to work.

Building something this complex requires three things working in concert, according to Karaletsos: data, modeling, and computing power. None of the three is optional, and CZI has been laying groundwork on each front since 2016.

On the data side, CZI has already pulled together nearly 100 million cells' worth of information through Chan Zuckerberg CELL by GENE, an open-source platform that lets scientists access and annotate detailed single-cell data. The organization also draws on resources built by the Chan Zuckerberg Biohub Network, including the Tabula Sapiens cell atlas and the OpenCell protein atlas. "When we want to build large-scale machine learning models of complex things, the gas for this will be data," Karaletsos said at the summit. "No individual entity will be able to collect the right or sufficient data to do this."

The modeling piece means training systems to make sense of different types of biological information at once, sequencing data, imaging data, and findings pulled from scientific literature, so the resulting model can describe and predict cell behavior across multiple angles rather than just one.

Then there is the computing. In September 2023, CZI committed to building a high-performance computing cluster with more than 1,000 GPUs, the specialized chips that power modern AI systems. Once finished, it is expected to be among the largest computing clusters dedicated to nonprofit life science research anywhere. "In many cases, researchers today are limited because of a lack of access to necessary compute," Brennan said. "We want to reduce those barriers to access."

Even once all three pieces are in place, CZI does not consider the job finished. The plan calls for ongoing validation with outside experts and partner organizations, treating the virtual cell platform as something that will keep evolving rather than a product that ships and stays static.

Openness is central to how CZI wants this to work. Grantees and partners across the Chan Zuckerberg Biohub Network are encouraged to publish early results on preprint servers like bioRxiv and medRxiv, a practice CZI has funded for years, well before formal peer review. "We wouldn't have had the breakthroughs we've had in machine learning in recent years without open science," Karaletsos said in a separate interview during the summit. "For us, it's a very natural way to work that typically leads to faster duration times and faster breakthroughs."

That commitment extends to the virtual cell project specifically. CZI plans to share the underlying data, make the predictive models themselves openly accessible, and gather feedback from the researchers who actually use the platform. Brennan was direct about why that matters to the organization. "It's really important to us that whatever we fund, whatever we build, that it's open and accessible to the broader community," she said, "because that is the way that new discoveries will be built on science."

Why It Matters

The technical ambition here is real, but so is the human stakes question underneath it. Drug development is slow and expensive partly because so much of it depends on trial and error in living systems, whether that is cells in a dish or patients in a trial. A working virtual cell platform would not replace that process, but it could narrow it, helping researchers rule out dead ends faster and focus limited resources on the most promising leads.

Karaletsos wants the eventual tool to feel less like specialized software and more like a research assistant, something a scientist can simply talk to rather than something that requires months of training to operate. If CZI's roadmap holds, the people who benefit most will not just be computational biologists. Brennan framed the end goal as giving "not just computational biologists, but all researchers and clinicians" the ability to drive insights faster. For patients waiting on better diagnoses and treatments, that kind of acceleration, if it materializes, could matter more than any single new drug.