Before a World Model of Life, Biology Needs a Better View

1. Guest Background

Alex Rives approaches the virtual-cell question with experience in both protein modeling and scientific institution building. He is Biohub's head of science and was previously co-founder and chief scientist at Evolutionary Scale. During this interview, he explicitly invokes his background studying protein biology and building protein models. That experience supplies a concrete reference for his argument about learning across evolution, although it does not establish that equivalent capabilities already exist for cells or human physiology.

Biohub, backed by the Chan Zuckerberg Initiative, has a mission to cure or prevent all disease. Asked what that ambition means, Rives describes a world without cancer, infectious disease, and chronic or degenerative conditions. His answer grounds the mission in recognizable disease categories rather than a promise of immortality.

The Cognitive Revolution conversation with Nathan Labenz examines the scientific infrastructure that such an ambition would require. Labenz repeatedly asks how cellular knowledge might become treatment or long-term health improvement. Rives concentrates on the preceding steps: observing biology, generating informative data, learning cellular behavior, and checking predictions experimentally. The resulting episode is an analysis of how predictive biology might be built, with protein-modeling experience informing a much broader and still unfinished research agenda.

2. What the Episode Covers

The organizing ambition is programmable medicine. Rives wants progress in understanding disease, detecting it, and intervening appropriately, eventually enabling researchers to read and write biological programs. Labenz initially divides the problem into understanding causal structure and designing something that produces the desired effect. Rives recognizes the familiar drug-development framework, then widens the discussion: cells execute programs, respond to inputs, and participate in processes through which disease unfolds. This information-theoretic perspective supplies a research direction; the episode does not establish that biology can already be edited with the predictability of software.

A nearer modeling objective is more specific. Given a cellular or tissue state and a set of inputs, predict the distribution of states that could follow. Rives emphasizes that this transition function would be probabilistic. Molecular measurements and imaging would provide complementary evidence about the relationship between inputs and responses. He specifically names transcriptomics, gene sequencing, and epigenetic measurements as molecular readouts to combine with observations of cellular responses through imaging. The intended output is therefore more useful than a descriptive label: it concerns how a system changes under specified conditions, without assuming that every cell has one predetermined outcome.

The data effort is correspondingly contextual. Labenz introduces the announcement as a $1.8 billion initiative. Rives describes collaboration involving the Department of Energy, NIH, Google DeepMind, Isomorphic, and Meta to coordinate data generation through the Virtual Biology Initiative. Priorities include spatial biology, cells situated among neighboring cells and tissue, and molecular measurements linked to imaging. He describes an envisioned scale of hundreds of billions or more cells while making clear that the longer-term roadmap is still being developed. These are proposed directions and scale ambitions, not a completed dataset demonstrated in the interview.

Two imaging examples show why scale means more than sample count. Rives explains that the laser phase plate changes the phase of an electron beam to improve electron-microscopy contrast. In cryo-electron tomography, better contrast helps distinguish macromolecules in noisy observations and investigate their interactions. The limitation is equally consequential: this detailed view covers only part of a cell and cannot supply an entire tissue picture.

At the organism level, Rives describes transparent zebrafish imaging that follows cells through embryonic development. Labeling immune cells also allows researchers to observe their movement and relationships with surrounding tissue. These examples reveal different aspects of biology. Rives accordingly separates measurement progress into parameters per cell, field of view and simultaneous cell count, observation over time, and molecular detail. The episode's data strategy is to improve and connect these dimensions so that models can learn cellular behavior in context.

3. Core Views: Reasoning, Examples, and Limits

Rives sets a demanding standard for a virtual cell: it should meaningfully predict biology and help experimental researchers decide what to do. That requires connecting molecular interactions to cellular responses and the tissue environment. He identifies three interdependent needs—modeling, scalable data generation, and measurement tools—and judges today's virtual cells insufficient as experimental oracles. That assessment belongs to the guest; the conversation supplies no comparative benchmark establishing a universal boundary for the field.

His argument about architecture is conditional. With appropriate inputs and outputs, unlimited data, and unlimited compute, he suggests that different architectures might converge as function approximators. Cellular biology is far from that limit, so inductive biases still matter. The reasoning extends beyond a shortage of examples. If an important process cannot be observed, collecting more records of what is already visible does not necessarily expose it. Model design must work with the information available, while measurement development changes what can become available.

An information-theoretic model offers an intermediate ambition. Rives thinks that level of cellular understanding might become attainable in roughly five years. A world model able to connect molecular structures, development, and tissue interactions remains farther away. His point-mutation example explains why the molecular scale cannot simply be discarded: a very small change can cause disease. Yet he also says that measuring an entire organism at molecular resolution is not currently possible. Together, these claims motivate learning informative relationships across scales without claiming a complete physical reconstruction.

Cultured cells illustrate the resulting boundary. They retain core machinery and can teach researchers a great deal, according to Rives, but cultured or transformed cells may occupy states substantially different from those in tissue. A successful prediction in one experimental environment therefore leaves a transfer question unresolved. Context can be part of the behavior being modeled, rather than incidental metadata attached afterward.

Model organisms present a related distinction. Rives credits systems including yeast with foundational mechanistic discoveries, but describes two failures of overreliance on mice: a treatment can appear promising because it works only in mice, and an important human mechanism can be missed because mice provide no adequate proxy. Organoids improve experimental representations of tissue complexity without resolving the problem completely. Conversely, he cites ESM and AlphaFold to argue that evolutionary information is essential for accurate human-protein prediction. Conserved components can support learning even where biological programs have been rewired. Integrating evidence across species is therefore a different undertaking from finding one species that perfectly represents humans.

The training discussion preserves another important separation. Rives expects self-supervised pretraining to learn representations from statistical patterns, followed by elicitation of capabilities relevant to particular tasks. He then proposes reinforcement learning with verifiable rewards and empirical AI that asks questions, performs experiments, learns, and proposes better experiments. These are connected stages, but pretraining does not itself produce the full loop. Biology requires laboratory verification, unlike settings where mathematical or coding outputs can be checked quickly through computation. The episode presents this as a future systems-building challenge, not an autonomous laboratory already shown to work.

The largest inferential gap concerns health outcomes. Labenz uses living to 120 to make the time-horizon problem concrete, and Rives acknowledges that linking short-term cellular states to long-term health remains difficult. His forecast of initial health-and-disease atlases in one or two years, and a more comprehensive map within five years under coordinated effort, concerns research infrastructure. Investigating reversal of pathological states would be a further step. Identifying a disease-associated state, establishing its causal role, reversing it, and demonstrating durable benefit remain distinct claims.

Finally, Rives treats language as a valuable route into accumulated scientific knowledge, while insisting that unknown biological facts require experiments. He says billions of proteins have been sequenced and most functions remain unknown; model features may consequently represent concepts without human names. Such features could become discovery leads, but unfamiliarity alone does not establish a mechanism. He remains uncertain about a shared biological-language latent space and points to insufficient bridging data. The episode supports an open research question, not a settled architectural prescription.

4. Learning and Application

For a research team, the first practical consequence is to define a virtual-cell claim precisely. Specify the starting state, the inputs, the predicted output, the time horizon, and the experimental environment. Because Rives proposes probabilistic transitions, ask whether the model represents multiple possible outcomes and whether subsequent experiments can test those predictions. This is an editorial application of his framework, rather than a benchmark supplied by the episode. It turns a broad label into a task with conditions under which it can succeed or fail.

The same discipline can guide data investment. If the research question concerns responses to neighboring tissue, more isolated-cell records may leave the relevant information missing. If it concerns change over time, a detailed snapshot does not directly provide a trajectory. Review parameters, cell count, spatial coverage, temporal coverage, and molecular detail separately, then identify which measurements must be associated. The tradeoff is not reducible to expensive instruments versus cheap samples: each approach must address a particular information gap.

Transfer should also become an explicit part of the study design. Cultured cells can support learning about shared machinery, and evolutionary data can broaden representations, while predictions intended for human biology still need testing in the relevant context. For animal evidence, ask both whether a positive result transfers and whether an absent result reflects a missing proxy. That preserves the value Rives assigns to model systems without treating their observations as interchangeable with human outcomes.

Health-and-disease atlases are best approached as infrastructure for generating and testing intervention hypotheses. A team can first distinguish states, then investigate what happens when those states are changed, and separately examine whether the change improves the intended outcome. Rives's forecasts can inform discussion of research dependencies, but cannot serve as treatment delivery dates. A short-term cellular readout does not, by itself, resolve the long-horizon problem he acknowledges.

EHR work offers a complementary use case. Rives has not deeply considered integration and does not settle regulatory feasibility, so this conversation provides neither an implementation plan nor permission to access records. He nevertheless sees value in matching existing drugs and mechanisms to patients. A project can therefore distinguish better treatment selection from mechanistic discovery, evaluating each on its own terms rather than letting predictive correlations stand in for a causal explanation.

Labenz's proposal for intensive longitudinal tracking, including micro-biopsies and sensors, should likewise be read as a research scenario. Rives values the prospect of connecting health trajectories to richer molecular monitoring and places Biohub's contribution at the technology layer. The useful condition is that measurements actually connect those levels, rather than merely accumulating separate streams. His refusal to give personal health advice—and his mention of only a sleep-tracking watch—provides no basis for recommending that listeners adopt invasive monitoring.

For researchers exploring unnamed model features, the appropriate next step is to formulate competing interpretations and seek experiments that distinguish them. Language systems can organize existing knowledge and help frame questions. Unknown mechanisms still require measurement. Whether the eventual system uses shared representations or cooperating models, its practical value must return to testable biological predictions and evidence that changes the next research decision.

Source

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