Biohub's Alex Rives says better measurement is the key to virtual cell models as the Virtual Biology Initiative grows to $1.8B

Rives says the models now called virtual cells can't yet stand in for experiments. He sees measurement and observability as a bigger limit than model architecture, though he notes that Transformers have no built-in multiscale structure. This fits a Nature Methods benchmark in which deep models, including scGPT and Geneformer, did not beat simple linear baselines at predicting perturbation effects. [1] [11]

The Virtual Biology Initiative launched in April with $500M over five years and has now grown to $1.8B. The new money includes more than $500M from DOE over five years and $300M from Google DeepMind, Isomorphic Labs and Meta. NIH's share of more than $500M counts datasets it already has, not new funding. These totals come from Biohub's own announcement and coverage that repeats it. The data will be open, but commercial funders get one year of exclusive access first. [2] [3] [5] [7]

Rives's target model is a probabilistic transition function: given a cell or tissue state and a set of inputs, it predicts a distribution of next states, linking genetic, transcriptomic, epigenetic and imaging readouts. The proposed training recipe follows LLMs: self-supervised pretraining, then elicitation, then reinforcement learning with verifiable rewards, where lab experiments do the verifying. He expects training across species to help, as it did for ESM and AlphaFold. [1]

Funding also goes to new measurement tools. One example is a laser phase plate that boosts cryo-EM contrast. Berkeley's results appeared in Science, while Biohub's dual-laser version is described only in a preprint. Rives says the device shows only part of a cell at atomic resolution. He expects first drafts of cell atlases covering health and disease within one to two years. [1] [6] [10]