AI is far from curing cancer. The startup says it knows what it needs.

A biotech startup called Vivodyne says the AI ​​drug discovery industry has a data problem and it has built a machine to solve it.

The company built HIVE, a modular robotic laboratory that can cultivate, autonomously administer and monitor 20 types of human tissue, generating the kind of causal biological data that is lacking in today’s AI models. Today’s data primarily comes from animal studies or studies of single cells or proteins rather than living tissues.

“What do these (AI) models do without human testing?” asks Andrei Georgescu, CEO and co-founder of Vivodyne. “They will treat cancer in mice.”

Even Anthropic CEO Dario Amodei wrote over the weekend that the claim that AI cures cancer has become more of a cliché than it is credible — in his words, “what works is what actually cures cancer.”

To be fair, the idea of ​​AI curing cancer is something Amodei himself abandoned in an earlier essay. Sam Altman repeatedly cites cancer treatment as a reason for OpenAI to power AGI and increasingly large-scale computing. And Google DeepMind’s Demis Hassabis said last year that AI could be able to treat every disease within 10 years.

Actual results remain lukewarm. While a small number of AI-designed drugs have advanced to human trials, with one reaching Phase 3 and undergoing extensive human testing, the reality is that the obstacles are not necessarily ones that can be overcome by AI today.

Although Nobel Prize-winning alphafold was a major advance in understanding the building blocks of life, it has not yet actually produced a new drug. Isomorphic Labs, which was founded on Alphafold, is on track to begin its first trial by the end of this year, originally planned for 2025. In February, the company wrote that true drug discovery will require “highly accurate predictive models across a wide range of biochemical properties and interactions.”

Georgescu says this space requires a “sanity check,” meaning existing models don’t have the data to capture the complexity of human biology. This is a challenge already facing the pharmaceutical industry: 90% of drugs that are effective in animal studies and enter clinical trials do not receive regulatory approval for human use.

Vivodine’s plan is different. Vivodyne was spun off from the University of Pennsylvania in 2021 after Georgescu earned a doctorate in bioengineering from the university. The company says its tissues closely match the behavior of real human organs, with its liver cells having a 94% predictive accuracy compared to human tests for toxicity, its airway tissue matching the behavior of real human tissues 96% of the time, and its bone marrow achieving 100% concordance in tests of 20 different chemotherapy drugs.

The company, which last week raised just under $80 million in two rounds led by Khosla Ventures, opened what it calls the world’s largest “human data center” outside San Francisco, and Georgescu’s team says it has already achieved twice the throughput of all animal studies conducted in the United States.

The idea is to speed up the development of drug candidates by better understanding what works before paying for clinical trials, which typically cost tens of millions of dollars. Without disclosing the names of its partners, Vivodyne said it is working with several major pharmaceutical companies to solve a problem that Georgescu likened to car crash testing. While automakers are typically confident that their vehicles will pass NHTSA requirements before testing, drug companies rarely have the same confidence going into clinical trials, and the vast majority of drugs fail to gain FDA approval.

But there is also a bigger vision. Georgescu believes autonomous biology labs are key to generating the kind of causal data that can be used to train new models of human biology. He points out that studies like this one published last month in Nature Methods do not find clear data scaling laws when training generative AI models on existing cellular data.

“All training is done on static snapshots of these cells, and the model is not conditioned in any way by how the cells got to that state,” Georgescu told TechCrunch. “In other words, the model will learn that ‘this is cell state A’ and ‘this is cell state B’, but it will not learn that ‘cell state B is an effect of cell state A that is inflamed’.”

But Vivodyne’s HIVE machine tracks hundreds of thousands of ongoing experiments in which diseased tissue is exposed to some type of stimulus, and Georgescu hopes this will provide the type of reinforcement learning that generates AI models that understand human biology enough to make more meaningful advances in medicine.

Georgescu believes this is key not only to today’s medical challenges, but also to the future, when complex diseases will require drugs that target multiple pathways, unlike the majority of currently available drugs.

“If you need a combination therapy, the area you have to explore explodes, and you can’t do it with an experimental approach,” he told TechCrunch. “You have to say, ‘I want this effect to occur, what causes should I invoke?'” Establishing causal relationships in human biology is the basis of all of this. ”