A World Model of Life: Biohub's Alex Rives on Virtual Cells, Measurement, & the $1.8B Data Push

Summary

• Biohub just announced a $1.8B expansion of its Virtual Biology Initiative, coordinating the Department of Energy, NIH, Google DeepMind, Isomorphic Labs, and Meta on data generation for a general model of cellular biology. The effort aims to build contextual maps — spatial biology linking molecular readouts to imaging “at the scale of hundreds of billions or more cells” — with an open call for other funders and the worldwide scientific community to join.
• Rives sees observability as a particularly fundamental bottleneck for virtual cell models, but not the only one. “There’s so much about the cell that we just can’t measure”; progress requires models, data generation at scale, and observation tools. He says today’s models “just don’t meet the definition of a model that could actually simulate the cell in a meaningful way and serve as an oracle for experimental biology.”
• The attainable five-year target is an “information-theoretic model of the cell” — a probabilistic transition function from cell state plus inputs to a distribution of next states — building toward “a world model of life.” In “the limit of infinite data and compute,” architectures probably end up the same; cellular biology is far from that limit, so inductive bias still matters.
• Measurement hardware is where Biohub’s thesis gets expressed: the laser phase plate — “the most powerful standing-wave laser in the world” — brings phase contrast to cryo-electron tomography, helping scientists make out individual molecules and how they interact in the living cell. Biohub also uses whole-organism zebrafish imaging to track every cell from embryo to fully developed organism.
• The LLM playbook transfers to biology: pretraining on statistical patterns, elicitation, then “RLVR for biology” with lab verification — “empirical AI” that asks a question, performs the experiment, learns, and proposes a better one in a recursive loop. Unlike math and coding, verification must happen in the laboratory, not through fast computation.
• Timeline and payoff: first drafts of health-and-disease cell atlases within one or two years, a “pretty comprehensive map” within five — then the question becomes “How do I intervene to reverse a state that’s pathologic?” Rives says he has thought less about EHR data; he worries that without a causal basis, models could learn “a bunch of correlations,” while acknowledging that health records could help get existing drugs and care to the right patients.
• The man building toward curing all disease tracks little about his own health: “I have a watch that tracks my sleep. That’s about as far as it gets.” On fusing language and biology latent spaces, he is unsure; the biological facts we most want to know may not be represented in text and can require experiments to discover.

Deep dive

1. “Curing all disease” means a world without disease — not a promise of immortality — and medicines you can program

• Pressed by Nathan on whether the mission implies living forever, Rives calls himself “a practical person” and sets his sights on a world without cancer, infectious disease, or chronic and degenerative conditions, rather than promising immortality.
• The therapeutic ambition goes beyond the “current medical paradigm, the therapeutic paradigm of pills and molecules” toward “far more programmable medicines” grounded in “reading and writing the programs of biology.” Disease is part of the program and “often unfolds in very regular ways,” which could provide new tools for understanding and intervening.
• Nathan’s two-part taxonomy — learn the causal structure, then design the intervention — Rives accepts as “the classic framework of drug development,” but argues the information-theoretic view of cells as systems of rules and code “opens up a wider aperture” than target-then-molecule.

2. Why virtual cells fall short: you can’t model what you can’t measure

• Rives says the cell is “maybe the right mesoscale” for thinking about biology, spanning molecular events and the larger context of tissue, physiology, and the organism — and notes that the transformer does not have the multiscale structure to represent all those levels.
• More fundamental than architecture is observability: “There’s so much about the cell that we just can’t measure. Sometimes we can’t see it at all.” He points to new measurement and observation technologies as a source of many advances in biology.
• Three legs must advance together — modeling, data generation at scale that reveals “the information architecture and programs of the cell in their statistical structure,” and observation tools. Rives says today’s so-called virtual cells do not meet the definition of a model that could meaningfully simulate the cell and serve as an oracle for experimental biology.

3. The five-year object: an information-theoretic cell model, en route to “a world model of life”

• Architecture agnosticism, properly hedged: in the limit of infinite data and compute, models are “just function approximators” and probably end up the same. Cellular biology is far from that limit; protein biology, with sequencing at scale, is an area where researchers can get closer. Away from the limit, inductive bias matters.
• Of biology’s two ways of thinking — bottom-up from chemical and physical principles versus the emergent informational view of “the cell as an information-processing machine with a genome that encodes the programs that run life” — Rives says the latter could be attainable via AI in the next five years.
• The first concrete target, aligned with Nathan’s input/output framing, is a transition function from state and inputs to the next state — probabilistic, since the cell transitions “to a distribution of possible downstream states.” The eventual model would represent molecular complexity and cell programs, allowing development and interactions to unfold: “a world model of life.”

4. The $1.8B push: DOE, NIH, DeepMind, Isomorphic Labs, and Meta coordinate on data

• The Virtual Biology Initiative was announced in April; the flagship partnership — Department of Energy, NIH, Google DeepMind, Isomorphic Labs, and Meta — was announced “the day before yesterday,” with an open call for other funders and the worldwide scientific community to join.
• The effort aims to dramatically expand contextual maps of cellular biology: spatial biology showing cells in tissue, linking molecular-level measurements to imaging “at the scale of hundreds of billions or more cells,” and producing the first datasets for training contextual information-theoretic models.

5. Measurement is the moat: a laser phase plate and transparent fish

• The laser phase plate, imagined by Holger Müller at Berkeley decades ago and supported by Biohub’s investment for years, uses “the most powerful standing-wave laser in the world” to shift the electron beam’s phase and provide phase contrast in cryo-electron tomography. That contrast helps scientists make out individual molecules in their living context and understand what they are and how they interact.
• Biohub’s program includes increasing the parameters measured per cell, expanding the field of view and number of cells observed simultaneously, and opening up the temporal and molecular dimensions.
• On Nathan’s scale-bridging concern — “a single point mutation in a gene can cause disease” — Rives says identifying an entire system down to the molecular level is not possible in the near future. The answer is to learn at the mesoscale and build models that integrate across biological hierarchies. Cells in culture and transformed cells are imperfect proxies for in-situ tissue, but retain core machinery and can still teach us about cellular biology. Biohub’s imaging institute is developing tools across levels of complexity: zebrafish are transparent, so its microscopes can track every cell from a one-cell embryo to a fully developed organism, or follow labeled immune cells moving through living tissue.

6. “In mice” isn’t the bug — evolution-wide transfer is the feature

• Nathan asks why treatments that work in mice seem not to work in humans more often, given how much biological machinery is conserved. Rives says over-relying on model organisms can fail both ways: you may think a cancer cure works when “it only works in mice,” or miss a human-relevant mechanism because mice do not provide a good proxy.
• The precedent from Rives’ ESM background: ESM and AlphaFold were not trained just on human proteins. Information across evolution is essential; a model trained only on human proteins would not be accurate.
• Across life, and especially vertebrates, much core molecular machinery is conserved, while “the programs are completely rewired, but they’re built with the same building blocks.” Organoids are a newer experimental proxy for human biology, but Rives argues that rather than seeking a perfect proxy, models can learn from information across different systems.

7. Pretraining, elicitation, then RLVR with lab verification — “empirical AI”

• Rives expects biology to follow shared AI principles: pretraining learns the underlying system from statistical patterns, then elicitation brings out “the part of that model which is relevant to the problem.” He connects this to the Hinton, LeCun, and Bengio vision of learning variables and parameters rather than encoding them by hand, with scaling laws as a clear expression of that progress.
• The hoped-for next step is “RLVR for biology” and “empirical AI”: systems that ask a question, perform an experiment, learn from it, propose a better experiment, and repeat the cycle. Unlike math and coding, biology requires systems that can verify predictions in the laboratory.
• Timelines, as hedged: atlases of cellular biology across health and disease are “a scaling challenge” addressable with today’s technology — first drafts within one or two years, and a “pretty comprehensive map” within five with a focused, coordinated effort. Such maps could support classical analysis as well as models that identify disease-relevant cell states and help answer: “How do I intervene to reverse a state that’s pathologic and bring it into a healthy state?”

8. Open questions: EHRs, latent fusion — and Rives’ own biomarkers

• Rives says he has not thought as much about health records. His concern is that patient-history models could produce “a bunch of correlations” without linking them to deeper causal factors in biology. He also sees potential for records to improve care by getting existing drugs and mechanisms to the right people, while arguing that this alone would not reach the biological fundamentals required for a step change.
• On co-training language and biology into one latent space, he is genuinely uncertain. Language can draw on centuries of empirical knowledge, but the biological facts we most want to know are not represented in text, cannot be derived from first principles or existing knowledge, and may only be discovered through experiments. Fusion may be hard at first because bridging data is scarce, though he says it could be interesting eventually. Protein language models already develop features for unknown biology — “novel concepts that we don’t have words or names for yet” — amid billions of sequenced proteins whose functions are mostly unknown.
• Nathan’s million-Bryan-Johnsons proposal — longitudinal human phenotyping with micro-biopsies and sensors — gets a qualified yes: Rives says the data would be “immensely valuable” and likely to happen over time. Biohub focuses on its technology layer, including continuous monitoring that measures more parameters, including molecular ones, to link bottom-up and top-down views.
• Asked for personal health advice, Rives says, “I’m not the right person to ask for that kind of advice” and barely thinks about it: “I have a watch that tracks my sleep. That’s about as far as it gets.”