Alex Rives thinks the AI models now called "virtual cells" are missing something more basic than a better design. "So much about the cell, you know, we just really can't measure," he said. "Sometimes we can't see it at all."
Rives is head of science at Biohub, the Chan Zuckerberg Initiative-backed research organization whose mission is to cure or prevent all disease. He was previously co-founder and chief scientist at EvolutionaryScale, which built the ESM series of protein language models. He took the Biohub role when the EvolutionaryScale team joined the organization, which Biohub announced in November 2025. He spoke with host Nathan Labenz on The Cognitive Revolution in an episode published October 20, 2026. The conversation came shortly after Biohub and its partners announced a $1.8 billion expansion of their Virtual Biology Initiative, a coordinated effort to generate biological data.
A virtual cell, in the sense Rives uses, is a computer model that can predict what a real cell will do: how it responds when a gene is altered, a drug is added or its surroundings change. A good one would let scientists try experiments on a computer first. Rives said none of today's models are there yet, and he laid out what he thinks it will take.
A world without disease, and medicine as programming
Labenz opened by asking whether "curing all disease" meant living forever. Rives called himself "a practical person" and described something narrower: a world without cancer, without infectious disease and without chronic or degenerative disease.
He also wants to change what medicine looks like. Today's therapeutic paradigm, he said, is largely one "of pills, of molecules." Biohub hopes to move toward "far more programmable medicines" that let scientists "read and write the programs of biology."
Labenz proposed splitting the challenge in two: find what causes a problem in the body, then find something that fixes it. Rives said that is "the classic framework of drug development" and that AI can still do a lot within it. He also described a more "information theoretic" view, which treats life as a program and cells as systems running rules and code to build an organism. In that view disease is part of the program, and disease "often unfolds in very regular ways," he said. If scientists could understand those regularities, he argued, they would have new tools for making progress.
Three things holding virtual cells back
Labenz's starting premise was that existing virtual cell models "don't work super well," and that the likeliest reason is low-quality data that may not contain the signals a model would need. Rives answered by widening the diagnosis.
He described the cell as a kind of mesoscale, a middle level for thinking about biology that is shaped from below and from above. Events down to individual atoms and molecular interactions decide what happens in the cell, while the cell sits inside tissue, physiology and a whole organism, sensing its environment and communicating with its neighbors.
He named three limits. The first is modeling. The transformer, the architecture behind most of today's large AI systems, "doesn't have that kind of multi-scale structure" needed to represent those different levels. The second is generating data at scale. The third, which he called "even more fundamental," is observation: having the tools to measure what matters inside a cell. Many advances in biology, he said, came from new technologies that changed what scientists could see.
Rives said progress is needed on all three. Until it comes, the models called virtual cells do not meet his definition of a model that could simulate a cell in a meaningful way and "serve as an oracle for experimental biology." Later he put Biohub's aim plainly: to "drive up the number of parameters that we can measure in every cell." One piece of that measurement work is the laser phase plate, an advance in microscopy that Labenz said is letting scientists begin to see directly how individual proteins interact inside the cell.
What a working model would look like
Labenz pushed on the modeling question. Given enough good data, wouldn't today's techniques probably work? And would a successful model track every protein bumping around, or simply take a perturbed cell and predict its next state?
Rives said that with unlimited data and computing power, model architectures would probably end up the same, since they are all "function approximators." Biology is far from that point. Protein biology comes closest, because gene sequencing produces enormous amounts of data. Cellular biology does not. When data are scarce, he said, inductive bias becomes important. Inductive bias means the assumptions about a problem that are built into a model's design.
His ultimate target is what he called "a world model of life." Such a model could describe a cell down to the atomically resolved structures of its parts and also capture its programs, letting development and the interactions within tissue play out and showing what would happen. "I think we're very far from that," he said.
What he considers achievable is an "information theoretic" model of the cell, which treats the cell as an information-processing machine whose genome encodes its programs. He thinks that level of understanding "could be attainable via AI" in about five years. The method would resemble how large AI models are pretrained today: self-supervised learning, in which a model finds patterns in huge datasets without hand-written labels. Here the data would link information at the genetic, transcriptomic (which genes are switched on) and protein levels to how the cell responds, as seen through sequencing, epigenetic measurements and imaging.
The first step, Rives said, is close to the input-output picture Labenz had sketched, which he said was "exactly right": learning a transition function. Given a cell's state and a set of inputs, the function predicts the next state. It will be probabilistic. "The cell is going to transition, I think, to a distribution of possible downstream states," he said. The idea already appears in the research literature. A 2024 perspective by Bunne and colleagues on building a virtual cell discusses modeling transitions between distributions of states, so that a model can allow for biological variability instead of forcing one fixed path.
What the $1.8 billion is for
Biohub launched the Virtual Biology Initiative in April 2026 as a five-year, $500 million commitment. It set aside $400 million for its own measurement technology, infrastructure and data generation and $100 million for outside research and coordinated data generation, and it committed to releasing the data openly. The October 7, 2026 expansion brought the coordinated total to $1.8 billion. That figure is not one new fund. It combines Biohub's original $500 million; more than $500 million over five years announced by the Department of Energy; resources from more than $500 million in earlier federal investment that the National Institutes of Health will coordinate; and $300 million committed jointly by Google DeepMind, Isomorphic Labs and Meta.
Rives said the group has invited scientists worldwide, and other funders, to join, and that a core team is still convening to write the longer-term roadmap. Some outputs are already clear, he said, above all a "dramatic expansion" of contextual maps of cellular biology. That includes spatial biology, which shows a cell among the surrounding cells in its tissue, and work linking molecular measurements to imaging, "at the scale of hundreds of billions or more cells." The aim is to create the first datasets that can train contextual, information-level models of the cell.
That would be a large jump. The Tabula Sapiens human reference atlas covers more than 1.1 million cells from 28 organs of 24 people. A hundred billion cells would be roughly 90,000 times as many.
A five-year map of health and disease
Labenz asked about time. Seeing how a cell responds an hour or a day after a change is one thing, he said; what he really cares about is "how am I going to live to be 120." Do scientists already know enough to say which cell states predict good outcomes?
Rives said that remains "a really big, big part of the climb," but called it "a scaling challenge" that today's technology can address. Through the initiative he wants to build atlases of cells across many physiological contexts, showing what separates disease from health at the cellular level, what makes each disease distinct and what causes different diseases share. He expects "first drafts of that within the next one or two years" and, "with a really focused, coordinated effort," a "pretty comprehensive map" within five.
Such a map would be useful for conventional analysis, much of which he expects to be done by AI agents. It could also train a model that pinpoints the cell states involved in disease. With that model, he said, researchers could ask how to intervene to move a cell from a pathological state back to a healthy one.
Why cells come before health records
Labenz asked whether electronic health records, linked to cell data, would be part of the picture. "To be honest, I haven't thought about that question as much," Rives said. His concern is that without the right foundation, models learn "a bunch of correlations": many models of patient histories with no way to connect them back to the causes in the underlying biology.
He did see value in records. They could improve patient care and help match existing drugs to people who might benefit from them, including people nobody had realized a drug's mechanism could help. He called that "a really important problem," but said it would not produce the step change in understanding and treating disease that Biohub is after.
Labenz then returned to his idea of a million Bryan Johnsons: people who would undergo regular microbiopsies, wear every available sensor and track every dimension they could, if money were no object. Johnson's published Blueprint protocol combines biomarkers, blood tests every three to six months, sleep measurements and imaging. Rives said such data "would be immensely valuable" and expected it to appear over time. Continuous-monitoring devices that measure more molecular signals, he said, could link the bottom-up view of cells with the top-down view of a person's health. But he placed Biohub "at that kind of technology layer." Labenz concluded that the area might be "less neglected than it seems to me from my outside position."
Asked what he does to build up his own data for these future technologies, Rives said he is "not the right person to ask" and is so focused on his research that he barely thinks about it. When Labenz pressed on whether he tracks anything at all, even a regular blood test, Rives offered one device. "I have a watch that tracks my sleep," he said. "That's about as far as it gets."