Building a World Model of Life: Biohub's Alex Rives on Why Better Measurement Must Come Before Virtual Cells

Alex Rives, head of science at Biohub and formerly of the ESM protein language model project, describes the mission of curing all diseases in practical terms. He is not talking about living forever. He means a world without cancer, infectious disease and chronic or degenerative disease, afflictions that have been fundamental to the human condition for a long time.

Reaching it requires a deep, fundamental understanding of biology, so scientists can identify core mechanisms and know how to detect and intervene in disease. It also requires going beyond the current therapeutic paradigm of pills and molecules toward far more programmable medicines. Rives talks about being able to read and write the programs of biology.

When host Nathan Labenz splits the problem into understanding causal structure and then finding an intervention, Rives calls that the classic drug development framework, in which progress is still possible. He adds an information-theoretic view in which life is a system of rules and code, and disease is part of the program that often unfolds in regular ways. That wider view, he says, opens up new tools for making progress.

Labenz suggests the data may simply not be good enough for models to learn what is going on. Rives agrees that the cell is incredibly complex, built from a dynamic collection of matter, where atomic-level interactions determine events at the cell scale, and where cells are embedded in tissues and organisms and execute programs to sense, respond and communicate.

He names three limits. One is the modeling paradigm, since the transformer lacks the multi-scale structure to represent different levels of complexity. More fundamental is observability: so much about the cell cannot be measured, and many scientific advances come from technologies that change what we can see. The third is generating data at scale that reveals the cell's information architecture and programs.

Progress has to happen on all three at once. Rives says models currently called virtual cells do not meet the definition of a model that could simulate the cell in a meaningful way and serve as an oracle for experimental biology.

Asked whether today's techniques would work with enough good data, Rives says that in the limit of infinite data and compute most architectures converge as function approximators. But cellular biology is very far from that limit, unlike protein biology, where sequencing at scale gets closer. When data is scarce, inductive bias becomes important.

His ultimate aim is a world model of life that could project a cell down to atomically resolved structures, model its programs, and let development and tissue interactions unfold. He says that is far off. What he thinks is possible in roughly five years is an information-theoretic model that treats the cell as an information-processing machine whose genome encodes the programs that run life.

Self-supervised pre-training on vast datasets can capture the latent variables behind input-output relationships, he says. The first thing to learn is a transition function that takes a cell state and a set of inputs to the next state. It would be probabilistic, predicting a distribution of downstream states, and Rives says the technology to create such a model now exists.

Rives explains that in April Biohub announced the Virtual Biology Initiative, and days before the interview it announced flagship partnerships with the Department of Energy, the NIH, Google DeepMind, Isomorphic and Meta as part of a $1.8 billion expansion. The aim is a coordinated effort to generate data for a general model of cellular biology, with an open call to the worldwide scientific community and other funders.

The priority is contextual maps of cellular biology, such as spatial biology, which sees cells in the context of tissue and neighboring cells, and linking molecular measurements to imaging. The scale should reach hundreds of billions of cells or more, creating the first datasets to train contextual, information-theoretic cell models.

On the measurement side, he describes the laser phase plate, a decades-long project first imagined by Holger Müller at Berkeley and invested in by Biohub for years. It is phase contrast for electron microscopy, using a powerful standing-wave laser to shift the electron beam's phase, which makes the noisy images of cryo-electron tomography clear enough to identify individual molecules and how they interact.

Rives says many historic scientific paradigm shifts have come from new measurement technologies, and biology differs from other AI domains mainly in its limits on observation. Biohub wants to raise the number of parameters measured per cell, widen the field of view and number of cells, and open up the temporal and molecular dimensions.

On whether a cell or a few cells tell us enough, Rives says this is a multi-scale problem in which a single point mutation can cause disease, and in the near term it is not possible to measure a whole system to the molecular level. A real cell model has to include context in living systems, not isolated cell cultures, though cells in culture still carry the core machinery and can teach a lot.

Biohub's imaging institute builds technologies at several levels. The laser phase plate gives an atomic-resolution view of part of a cell, other tools observe and perturb fields of cells in culture, and whole-organism imaging in transparent zebrafish lets them track every cell of a developing embryo or follow immune cells in a living fish. The model must then integrate across these hierarchies.

On the old joke to append "in mice" to biology headlines, Rives notes that model organisms, even yeast, have given most of our molecular understanding of disease, but are imperfect proxies that can lead to false cures or missed mechanisms. Organoids are the latest improvement. He thinks AI helps: ESM and AlphaFold came from training across all of evolution, not just human proteins, and he says that broad information is essential, because core molecular machinery is conserved even as programs are rewired.

Rives expects the principles from other AI fields to carry over. Self-supervised pre-training, which realises the dream of Hinton, LeCun, Bengio and other pioneers that you need not hand-encode the variables, will play out in cellular biology as it has with proteins. A phase of learning statistical patterns would be followed by an elicitation phase bringing out what is relevant to a given problem.

He expects reinforcement learning with verifiable rewards to be possible for biology, with an eventual goal of empirical AI that asks a question, performs the experiment, learns and proposes a better one in a recursive cycle. The difficulty is that biology is empirical, whereas math and code allow fast computational verification, so biology will need systems for verification in the laboratory.

On timescales, he says atlases of cells across health and disease are within reach, with first drafts in one to two years and a pretty comprehensive map within five given a focused, coordinated effort. Such a model could identify cell states relevant to disease and answer how to reverse a pathological state into a healthy one.

On human health records, Rives admits he has thought less about them and says his worry is that without the right basis one just learns correlations, with models of patient histories that cannot link back to underlying causal biology. He thinks records could help get existing drugs to the right people, but says a step change needs the fundamental science that Biohub is focused on.

On whether language and biology models will fuse into a joint latent space, he says he is not sure. Language is a powerful substrate for reasoning and for tapping centuries of human knowledge, but the facts biologists most want are empirical and not in text. Fusing the spaces would need bridging data that does not yet exist, though he thinks it is probably possible eventually.

He adds that protein language models develop features for unknown protein biology, concepts we have no words for yet, since the function of most of the billions of sequenced proteins is unknown. Asked about a continuous-monitoring study of many people, he says such data would be valuable and will likely happen, but Biohub sees itself as working at the technology layer. Personally, he says he barely thinks about his own health data and only tracks sleep with a watch.