Virtual Cells Need Better Measurements Before They Can Predict Cell Behavior

For Alex Rives, Biohub’s Head of Science, the obstacle to a useful model of cellular biology is not simply a shortage of observations. It is that much of the biology a model would need to learn is hard to observe, and often harder still to measure at scale. Rives described three requirements that have to advance together: modeling approaches suited to biology’s complexity, data that reveals how cells work, and technologies that can measure the relevant processes. Current virtual-cell models, he said, do not yet meaningfully simulate cells or serve as an oracle for experiments.

The difficulty starts with the object being modeled. A cell is a dynamic collection of matter: interactions among molecules help determine what happens at the scale of the cell. But the cell is also part of tissue, physiology, and an organism. Cells sense and respond to their environment, communicate with one another, and take part in complex, changing processes. Rives described the cell as a useful mesoscale for thinking about biology, but not a self-contained unit whose behavior can be understood independently of the levels beneath and around it.

That makes measurement part of the modeling problem. Rives said some aspects of the cell cannot be seen at all with current methods, while generating data about other aspects at scale is difficult. In his account, advances in biology have often followed new technologies that change what researchers can observe. Better measurement can open up new questions, not just provide more answers to questions that were already easy to ask.

Biohub’s work on the laser phase plate illustrates that principle. Rives said the project had been in development for years, crediting Holger Müller at Berkeley with imagining the approach decades earlier. The device uses a standing-wave laser to shift the phase of an electron beam, producing phase contrast in electron microscopy.

The contrast matters for cryo-electron tomography, which uses electron microscopy to image sections of cells. Rives said those observations are very noisy; without sufficient contrast, it is difficult to make out individual proteins and other macromolecules in their cellular context. The phase plate is intended to make those molecules easier to see, so researchers can begin to understand what they are and how they interact in situ. Rives expects the technique to open a new view of the dynamic structural biology of living cells.

The phase plate does not provide a molecular-resolution image of an entire tissue or organism. Rives described it as opening a high-resolution view of part of a cell. Other technologies are needed to observe fields of cells, perturb those cells and measure their responses across many observations, or track biological processes at the scale of a whole organism.

Biohub’s imaging work spans some of those levels. Rives pointed to zebrafish, whose transparency permits researchers to image cells in a developing embryo. He described microscopes that can track cells as an embryo develops from one cell toward a fully developed organism. Researchers can also label cell types such as immune cells and follow their movement and interactions with surrounding tissue in a living fish.

The aim is not to produce one universally complete picture with a single instrument. It is to gather observations at different levels and build models that can integrate them. Rives framed the measurement challenge in practical terms: increase the number of parameters researchers can measure in each cell, expand the field of view and the number of cells observed at once, follow cells over time, and see more of what is happening at the molecular level. Each measurement method can contribute a different part of that picture.

The Virtual Biology Initiative was announced in April as an effort to build datasets that could enable virtual biology more broadly. Rives described the later $1.8 billion announcement as a flagship partnership within that initiative. He named the Department of Energy, the NIH, Google DeepMind, Isomorphic, and Meta among the organizations involved in the collaboration.

The distinction matters: Rives did not describe the full $1.8 billion as a single budget for one dataset, nor did he specify a line-item allocation or a fixed inventory of experiments funded by that sum. He described a coordinated effort to generate data and a broader initiative intended to build the datasets needed for virtual biology. A core group was convening to develop a longer-term roadmap, and Biohub was inviting other funders and the worldwide scientific community to participate.

Some priorities were already clear in Rives’s account. One is a dramatic expansion of contextual maps of cellular biology. Spatial biology is a central example: observing cells in tissue, in relation to neighboring cells, while linking molecular measurements to imaging. Rives described the intended scale as hundreds of billions of cells or more. The goal is to move quickly enough to produce initial datasets that can support contextual, information-theoretic models of the cell.

The emphasis on context addresses a limitation of measurements made on cells in isolation. A molecular readout can provide information about a cell, but the tissue around it and its neighboring cells may matter to how it behaves. Linking molecular measurements to images and tissue context gives researchers a way to examine those relationships rather than treating each cell as an isolated sample.

Rives also described the data effort as a way to bring together measurements across biological dimensions: molecular readouts, imaging, and cellular responses. Such datasets could be useful before they produce a sophisticated model. They could support conventional analysis of what distinguishes health from disease and help identify causes shared across different disease classes. AI agents, he suggested, could eventually help scale some of that analysis, but the datasets would have value for classical analysis as well.

The proposed maps are not, by themselves, a treatment or a complete model of a person. They are a foundation for asking more specific questions about which cellular states are associated with disease, how those states differ from healthy ones, and what changes might move them. Rives estimated that first drafts of the maps could be available within one or two years. With a focused, coordinated effort, he said, a fairly comprehensive map could be possible within five years. Those timelines concern the mapping effort; they do not guarantee that therapies will follow on the same schedule.

Rives distinguishes a nearer-term information-level model from the more ambitious “world model of life.” The latter would represent biology at multiple levels: molecular structures and the programs that unfold in cells, tissues, and developing organisms. He described that as a long-range goal, not a capability that current virtual-cell models already provide.

One way to study life is bottom-up, building from physical and chemical principles. Another is to regard the cell as an information-processing system whose genome encodes programs that help build living organisms. Rives sees the informational description as a potentially productive level for AI: learning how genetic, transcriptomic, and protein-level information relates to observable cellular programs and responses.

A first step, he said, would be to learn a transition function. Given a cell state and a set of inputs, the model would predict a distribution of possible next states. The output is probabilistic because, in Rives’s account, a cell does not necessarily move from one state to one fixed next state under a given input. The task is to learn the range of possible downstream states and how they relate to the conditions the cell encounters.

That transition function would make a specific kind of question expressible: if the cell is in this state and receives this input, what states might follow? It is not yet the same as explaining every mechanism inside the cell, and it does not by itself establish that a proposed intervention will work in a patient. It would provide a model of how states and inputs relate to subsequent states, which could help researchers identify experiments and interventions to examine.

Rives’s therapeutic ambition is to ask how to move a pathological state toward a healthy one. Mapping states associated with health and disease could help identify which cell states matter to a disease, and how those states change under different inputs. The intervention problem then becomes more than identifying a molecular cause and designing a molecule to act on it: researchers would want to predict which changes in a system might shift it away from disease.

That remains a proposed direction, not a demonstrated capability of the models Rives described. The near-term maps could reveal distinctions between disease and health, or common causes across disease classes; a model built on those data might help identify relevant states and explore possible interventions. Whether a particular intervention actually produces the desired effect would still require evidence, including experimental testing.

Rives connects this ambition to the conventional structure of drug development. Researchers identify a cause at the molecular level, because that is where current technologies can intervene, and then find or design an intervention. AI may create new possibilities within that framework. His broader aspiration is to move beyond a therapeutic paradigm centered on pills and molecules toward more programmable medicines—ways to read and write biological programs.

The further goal is a model able to represent molecular complexity and biological programs together: one that might project cell state down to atomically resolved structures of its components while also allowing development and interactions within tissues to unfold. Rives described this as a world model of biology and said it is far off. The information-theoretic model is a more attainable stage on the way, not another name for the full ambition.

He said that, in the limit of infinite data and computation, different model architectures might converge as general function approximators. But cellular biology is far from that limit. When data is scarce, inductive bias—the assumptions and structure built into a model—becomes important. Rives questioned whether the transformer, as currently used, has the multiscale structure needed to represent the different levels of biological complexity.

Protein biology is a partial contrast. Rives pointed to large-scale sequencing and protein models as an area where researchers can learn from much broader data. Cellular biology, he said, is much further from the point at which data alone might make architecture less consequential. Better modeling can help extract more from limited observations, but it cannot supply information that the measurements never captured.

For Rives, “curing all disease” is a practical ambition rather than a promise of immortality. He described the goal as a world without cancer, infectious disease, chronic disease, or degenerative disease. The path he outlined depends on understanding how disease works, identifying the mechanisms involved, and developing appropriate ways to intervene.

Rives expects self-supervised learning and pretraining to matter in biology, as they do in other areas of AI. These methods can learn representations from statistical patterns without requiring researchers to specify every relevant variable in advance. He sees this approach already playing out in protein biology and expects it to extend to cellular biology.

In his description, pretraining learns from patterns linking inputs and outputs to build a representation of the system behind them. Researchers can then elicit the parts of that representation relevant to a particular problem. But biological predictions ultimately have to be tested against experiments. In mathematics or coding, verification can often happen computationally and quickly. Biology is empirical, Rives said, and requires systems that allow predictions to be checked in the laboratory.

Rives described a future in which AI can ask a question, perform an experiment, learn from the result, propose a better experiment, and repeat the cycle. He also raised the possibility of reinforcement learning with laboratory verification. These are expectations for future systems, not a description of a settled or already functioning pipeline. They depend on models that represent biological complexity in useful ways and on laboratory systems capable of providing feedback.

The loop would make experiments part of model development rather than only a final test. An AI system could propose a question or prediction; an experiment could produce an observation; and that result could inform the next question. If the result contradicts the model, the system would need to incorporate that evidence and revise what it predicts. Rives called the prospect “empirical AI” and described repeated cycles of question, experiment, and learning as a route to recursive improvement.

That vision also clarifies why the proposed data effort is not only about accumulating more records of cells. A dataset can help a model learn patterns, but experimental feedback is needed to test whether the model’s predictions hold under a perturbation or in a new setting. The laboratory is the verification system; building it into a repeatable loop is part of the technical challenge. Rives did not claim that current models can autonomously carry out this cycle, or that a general laboratory verification system is already in place.

How biological models will connect to language models remains unresolved for him. Rives sees language as a powerful substrate for reasoning and a way to draw on centuries of empirical investigation represented in human knowledge. But the biological facts researchers most want to discover may not be present in text or derivable from existing knowledge. They may be empirical facts that require experiments, which is one reason he sees predictive biological models as critical.

He was not committed to whether language and biological representations will eventually be fused into one latent space or interact through some other arrangement. A fused space could be useful, he said, but it would initially require bridging data that is not abundant now. He also pointed to biological features for which researchers have no words. Most proteins do not have known functions; a protein model can still learn patterns across proteins and represent aspects of biology that have not yet been named. Some of the representations researchers may need, in other words, cannot simply be supplied by existing language.

Rives emphasized that experiments in model systems, including yeast, have yielded important insights into molecular mechanisms relevant to human biology and disease. But model organisms are imperfect proxies. Overreliance can lead researchers to believe a treatment is a cure for cancer when it works only in mice. It can also cause researchers to miss a mechanism relevant to humans because there is no good proxy for it in mice.

Organoids, which can reproduce more of the complexity of tissues in a culture system, are one effort to improve experimental proxies. Rives did not describe them as a solution to the problem. Nor did he suggest that model organisms are interchangeable with people. His broader distinction is between trying to create a perfect substitute for human biology and building models that can learn from information across species.

His example comes from protein biology. Rives said ESM and AlphaFold were not produced by training only on human proteins; they drew on proteins across evolution. In his account, evolutionary information is not merely helpful to predictions about human biology but essential to them. A model trained only on human proteins would not be comparably predictive.

The relevant conservation is not a claim that different organisms are the same. Rives noted that much of the core molecular machinery across vertebrates is preserved, including transcription factors and the ways they interact with enhancers and promoters. The programs built from that machinery have been rewired, even when some of the building blocks remain shared. A model can learn from those shared structures and from variation across organisms without treating a mouse, fish, or other system as a perfect human stand-in.

This distinction shapes the role he expects model organisms to play. They can provide experiments and observations that illuminate shared mechanisms, and they can contribute data to models trained across evolutionary diversity. But an experimental result in one organism does not automatically establish what will happen in humans. The aim is not to eliminate the differences between species; it is to build predictive models that can learn from both shared machinery and those differences.

Rives’s account of cellular modeling makes the same point about experimental systems within a species. Cells in culture can differ from cells in tissue, and transformed cells may differ from states found in situ. Yet they still retain core machinery and programs, so they can be informative. Tissue context adds another kind of evidence. No single setting supplies the complete picture, and no perfect proxy is likely to replace the work of integrating information across settings.

Human health records could contribute another kind of evidence, but Rives said he had not thought as much about how they might connect to cellular observations. He sees potential for health histories to improve care—for example, by helping get existing drugs to the right patients or revealing that a known mechanism may be relevant in a setting where its importance was not recognized.

His concern is that health histories alone could yield correlations without linking them to deeper biological causes. Rives’s emphasis is on understanding the underlying biology that determines health and disease, which he sees as necessary for a larger change in the ability to treat disease. He described Biohub’s focus as the fundamental science and technology that can help make those connections, rather than making linked patient histories the basis of the effort.

He also sees value in continuous monitoring technologies that can measure more parameters over time, including parameters closer to the molecular level. Such measurements might help connect bottom-up cellular observations with top-down views of health. He described that work as valuable and anticipated convergence over time, while saying Biohub sees its own role primarily at the technology layer.

Predicting long-term health from cell states remains a substantial challenge. Rives called it a major part of the climb, while also describing it as a scaling challenge that technology available today could begin to address. His nearer expectation is that cellular maps across health and disease will help distinguish disease states and shared causes. Those maps could support models that identify relevant cell states and explore interventions, but they are not the same as predicting a person’s lifetime health trajectory.

Asked what he does personally to prepare for future health technologies, Rives said he is so focused on research that he barely thinks about his own health data. He wears a watch that tracks his sleep and said he does not regularly track much else.