Vivodyne: teaching AI models cause and effect by growing human tissue in a lab
A University of Pennsylvania spinout runs modular robotic labs that grow twenty kinds of human tissue, generating the causal biological data that today's AI drug-discovery models are missing.
I ran into this story while reading the latest round of pushback against the "AI will cure cancer" line that keeps circulating from AI lab CEOs, and one detail in the rebuttal stuck with me: a startup arguing that today's models don't fail at biology because they're too small, they fail because nobody ever showed them why a cell got sick in the first place.
The startup is Vivodyne, spun out of the University of Pennsylvania in 2021 by bioengineer Andrei Georgescu. It builds HIVE, modular robotic labs that grow twenty kinds of human tissue, dose them with disease triggers and drugs, and monitor the results around the clock, no human running the experiment. The tissue is the pitch: Vivodyne says its liver cells predict toxicity with 94% accuracy against real human trials, its airway tissue matches real behavior 96% of the time, and its bone marrow hit 100% concordance across twenty chemotherapy drugs.
That's the clever part. Most biological AI trains on static snapshots, "this is cell state A," "this is cell state B," with no record of what pushed the cell from one to the other. Without that causal link, Georgescu argues, a model can describe biology but can't predict what an intervention will do to it, which is exactly the gap between curing cancer in mice and curing it in people. HIVE exists to generate that missing layer: hundreds of thousands of tracked experiments where a tissue gets a stimulus and the response is logged as it happens, at a throughput Georgescu says already beats every animal trial running in the US combined.
The company just opened what it calls the world's largest "human data center" outside San Francisco, working quietly with pharma partners it won't name. Georgescu compares it to crash testing: a carmaker trusts a design will pass safety checks before the crash happens, while a drugmaker rarely has that same confidence going into a trial that costs tens of millions and still usually fails.
I don't know if lab-grown tissue data solves drug discovery on its own, but the framing is refreshing after two years of "AGI will cure everything" claims with little to show for it: the fix isn't a bigger model, it's a data source nobody had before.