Biotech startup Vivodyne says AI drug discovery fails because of data, not algorithms

The biotech startup Vivodyne argues that the fundamental bottleneck in AI-driven drug discovery is a lack of appropriate data, not flawed models. Founded in 2021 by CEO and co-founder Andrei Georgescu, who earned a PhD in biomedical engineering at the University of Pennsylvania, the company built a system called HIVE—modular robotic labs that can grow 20 types of human tissue, dose them and monitor them autonomously. The output is causal biological data that current models simply do not see: most existing data come from animal studies, single-cell cultures or isolated proteins, not from living tissue that behaves like a human organ.

The problem is data, not the algorithm

Georgescu calls for a “sanity check” of the field, noting that existing models lack the data needed to capture the complexity of human biology. This shortfall is well known in pharma: 90% of drugs that succeed enough in animal studies to reach clinical trials fail to obtain regulatory approval for humans. Vivodyne’s tissues, the company says, show high fidelity to real human responses: liver cells achieve 94% prediction accuracy for toxicity compared with clinical trials, respiratory tract tissue shows 96% alignment, and bone-marrow tissue reached 100% alignment in a test of 20 different chemotherapeutic drugs.

90% fail to translate from animals to humans

The 90% failure rate underscores the gap between animal models and human outcomes. Vivodyne’s claim is that its engineered tissues better mimic true human physiology, potentially narrowing the attrition gap that forces pharmaceutical firms to spend tens of millions of dollars on costly clinical trials.

World’s largest human data hub

Last week, after raising just under $80 million in two rounds led by Khosla Ventures, the company opened what it calls “the world’s largest human data hub” near San Francisco. Georgescu says his team is already achieving double the throughput of all animal experiments conducted in the United States. The goal is to shorten the path to viable drug candidates by understanding what will work before committing tens of millions of dollars to clinical testing. While Vivodyne does not name its partners, it reports collaborations with several large pharma companies. Georgescu likens the situation to automotive crash-test standards: a car manufacturer is confident the vehicle will meet NHTSA requirements before testing, whereas drug companies rarely enter clinical trials with comparable certainty, and most drugs fail FDA approval.

A bigger vision: causal data to train new models

Beyond accelerating development timelines, Georgescu sees the autonomous labs as a source of the causal data required to train the next generation of models on human biology. He points to a study published this month in *Nature Methods* that found no clear scaling laws when training generative models on biological data, suggesting that simply adding more of the same data will not solve the problem and that a fundamentally different type of information is needed.