Chan Zuckerberg Biohub Coordinates $1.8B Effort to Train AI on Living Cells

The Chan Zuckerberg Biohub is organizing a $1.8 billion research program to train artificial intelligence models that predict how cells behave, with a first shared dataset targeted for release in roughly a year. The effort pools money and resources from Meta, Google DeepMind, Isomorphic Labs and the US Department of Energy, according to details of the initiative. The goal is to give biology the kind of large-scale predictive models that language and image research have had for years.

The money is not going to a single lab or a single model. Participants are funding three separate inputs: experimental data generation, laboratory equipment, and compute for training. That split matters because the bottleneck in cellular AI has been measurement, not architecture. Models can only learn what the underlying experiments can observe, and wet-lab work is slow, expensive and hard to standardize across institutions.

Biohub's role is coordination. The organization, backed by Mark Zuckerberg and Priscilla Chan, is positioning itself as the convener that sets data formats, quality standards and access rules so that datasets produced at different sites can be combined. Meta and Google DeepMind bring model-building experience, Isomorphic Labs brings drug-discovery context, and the Department of Energy brings national-lab instrumentation and supercomputing capacity.

Predicting cell behavior is a harder target than predicting protein structure. A folded protein is a relatively static object, while a cell is a moving system: genes switch on and off, signals propagate, and the same cell can respond differently depending on its environment and history. Models that handle that complexity could shorten the search for drug targets, flag toxicity earlier, and let researchers run virtual experiments before committing to bench work.

The one-year timeline for a first dataset is the concrete benchmark to watch. If the consortium delivers usable, well-annotated data on schedule, it gives academic groups and startups a common foundation to build on rather than each lab hoarding its own measurements. If it slips, the program risks repeating the pattern of ambitious biology consortia that produced press releases faster than reusable data.

The open question is access. A $1.8 billion program funded largely by four large institutions could either seed a broad open ecosystem or concentrate the best cellular models inside a handful of companies that already have compute and talent. How Biohub licenses the data, and whether outside researchers can train on it, will determine whether this becomes shared infrastructure or a private advantage.