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

What It Covers

Alex Rives, head of science at Chan Zuckerberg Biohub, explains why current virtual cell models fail to transform medicine, how a $1.8B partnership with DOE, NIH, Google DeepMind, and Meta aims to generate the cellular data needed to build information-theoretic cell models, and what a true "world model of life" requires across measurement, data, and modeling dimensions.

Key Questions Answered

• •Measurement as the bottleneck: Current virtual cell models fail not primarily because of weak architectures but because biology's observational tools cannot capture the fine-grained causal structure inside cells. Transformers lack the multiscale structure biology requires, but even perfect architectures cannot learn from data that doesn't encode what actually matters. Advancing microscopy and molecular readouts is therefore a prerequisite to advancing AI models of cellular biology.
• •Laser Phase Plate breakthrough: Biohub's laser phase plate—decades in development—applies phase contrast to cryo-electron tomography by using the world's most powerful standing-wave laser to shift electron beam phase. This enables scientists to directly visualize individual proteins and macromolecules inside cells in their native context, opening structural biology at atomic resolution within living cellular environments for the first time at meaningful scale.
• •Transition function as the near-term target: Rather than a full cell simulation, the achievable five-year goal is an information-theoretic transition function: a probabilistic model that takes a cell or tissue state plus a set of inputs and outputs a distribution of possible next states. This framing—borrowed from world-model thinking in AI—gives researchers a concrete, tractable modeling objective that can be trained on scalable molecular and imaging datasets.
• •Cell atlases linking health and disease states: The $1.8B Virtual Biology Initiative prioritizes building spatial biology datasets that link molecular measurements to imaging across hundreds of billions of cells in both healthy and diseased tissue contexts. These atlases, targeting first drafts within one to two years and comprehensive coverage within five, are designed to reveal shared causal mechanisms across disease classes rather than disease-specific correlations.
• •RLVR for biology with lab verification: The LLM training paradigm—pretraining on statistical patterns followed by elicitation—applies directly to cellular biology, but verification must happen in physical laboratories rather than computationally. Rives envisions closed-loop empirical AI systems where models propose experiments, labs execute them, results feed back into model improvement, and the cycle repeats recursively, mirroring reinforcement learning from verifiable rewards seen in math and coding domains.
• •Cross-species training is essential, not optional: ESM and AlphaFold achieved accurate human protein predictions only by training across all of evolutionary sequence space, not on human proteins alone. The same principle applies to cellular biology: model organisms, organoids, and zebrafish whole-organism imaging data contribute conserved mechanistic information that human-only datasets cannot provide. The core molecular machinery—transcription factors, enhancer-promoter interactions—is preserved across vertebrates even as programs are rewired.