AI leaders have offered optimistic timelines for using their technology to cure major diseases, but the early record of AI-designed medicines remains limited. Vivodyne believes the bottleneck is not simply a lack of larger models. It is a lack of experimental data showing how human tissue actually responds to potential treatments.
The startup, spun out of the University of Pennsylvania in 2021, has raised just under $80 million across two rounds led by Khosla Ventures. Its approach combines living human tissue with modular robotic laboratories called HIVEs, which can grow, dose and monitor experiments with relatively little manual intervention.
Drug models need experiments, not only predictions
Current biological AI often learns from published observations and existing cellular datasets. Vivodyne chief executive Andrei Georgescu argues that such information is rich in correlations but weak on causality. A system may learn that two biological features appear together without knowing whether changing one produces the other.
Vivodyne’s platform is designed to generate that missing evidence. Its robotic modules cultivate 20 kinds of human tissue and run many controlled interventions, recording what changes after a compound is introduced. The company says it has hundreds of thousands of experiments underway and intends to use the resulting causal data for reinforcement learning and searches across combinations of therapies.
The distinction matters because AI drug discovery has not yet matched its boldest forecasts. TechCrunch notes that only a handful of AI-designed drugs have entered clinical trials and one has reached Phase III. AlphaFold transformed protein-structure prediction, but has not itself produced a new medicine. Isomorphic Labs expects trials by the end of this year after initially targeting 2025.
A study in Nature Methods, cited in the source, also found no clear scaling laws in existing cellular datasets. That result does not show that AI cannot help drug development. It suggests that simply feeding more of the same observations into larger systems may not deliver a predictable jump in performance.
Human tissue could narrow a costly translation gap
Animal testing remains a poor filter for human outcomes. Around 90% of drugs that work well enough in animals to enter clinical trials ultimately fail to secure regulatory approval for people. Clinical studies can cost tens of millions of dollars, making each late-stage failure exceptionally expensive.
Vivodyne wants its tissue models to identify weak candidates earlier. The company claims its liver cells predict toxicity with 94% accuracy relative to human trials, while its airway tissue reaches 96%. It also reports complete concordance from bone-marrow tests across 20 chemotherapy drugs.
Those figures are company claims presented through TechCrunch, not a complete independent validation of every tissue system or disease area. High performance on selected comparisons does not guarantee equal accuracy for new compounds, complex illnesses or long-term effects. Living tissue models also cannot reproduce every interaction occurring across an entire human body.
Still, a reliable intermediate test could be valuable even if it does not replace clinical trials. Reducing the number of unpromising compounds reaching patients would save time, capital and exposure to ineffective treatments.
An automated “human data center” tests the model
Vivodyne has opened what Georgescu calls the world’s largest human data center outside San Francisco. He claims it can produce twice the experimental throughput of all animal trials conducted in the United States. The source does not independently verify that comparison, and the company has not named the pharmaceutical partners using its system.
The facility turns drug development into an infrastructure challenge: tissue cultivation, robotic handling and data generation operate as a unified production line. If the platform works as advertised, its most important output may be neither a single drug nor a single model, but a repeatable supply of intervention data that other AI systems can learn from.
That is a less spectacular promise than an imminent cure for cancer, but it targets a concrete weakness. Algorithms can explore chemical possibilities quickly; they cannot infer dependable human biology from evidence that was never collected. Vivodyne’s bet is that better experiments must scale alongside better computation.
