Vivodyne, a biotechnology startup, claims there's a significant data deficit within the AI-driven drug discovery sector, and it has developed a solution to address this issue.
The company's innovative modular robotic labs, known as HIVE, can cultivate 20 types of human tissues and autonomously administer doses and conduct monitoring. This process generates causal biological data that is currently lacking, as most existing data primarily relies on animal testing or studies involving single cells and proteins, rather than living tissues.
Andrei Georgescu, the CEO and co-founder of Vivodyne, raises a pertinent question: “Without human testing, what can these AI models achieve?” He emphasizes that they might only succeed in “curing cancer in mice.”
Dario Amodei, CEO of Anthropic, echoed this sentiment over the weekend, noting that claims of AI curing cancer have become more of a cliché than a viable expectation, stating, “what will actually work is curing cancer."
Interestingly, Amodei himself had previously suggested that AI could play a role in cancer cures, while Sam Altman has often referenced this potential as a driving reason for OpenAI’s pursuit of artificial general intelligence and expansive data centers. Similarly, Demis Hassabis of Google DeepMind mentioned last year that AI might eradicate all diseases within a decade.
However, the actual results from AI-driven drug discovery have been modest. Although a few AI-designed drugs have reached human trials—a single candidate even entered Phase III testing—the obstacles many of these initiatives face are not necessarily ones that current AI technologies can overcome.
Despite AlphaFold's significant contributions to understanding the fundamental elements of life, it has yet to lead to the creation of a new drug. Isomorphic Labs, founded with the intention of building upon AlphaFold's technology, has reported delays, pushing back its anticipated trials initially scheduled for 2025 to late this year. The company has stressed that successful drug discovery needs “highly accurate predictive models covering a wide range of biochemical properties and interactions.”
Georgescu advocates for a “sanity check” in the industry, expressing concern that prevailing models lack the necessary data to comprehensively understand human biology. This challenge is a broader issue for the pharmaceutical industry, where 90% of drugs that prove effective in animal models fail to gain regulatory approval for human use.
What sets Vivodyne apart is its foundation; the company emerged from the University of Pennsylvania in 2021, following Georgescu's completion of a PhD in bioengineering there. Vivodyne asserts that its human tissues closely mimic the functionality of genuine human organs—its liver cells boast a 94% predictive accuracy in toxicity tests compared to human trials, its airway tissue demonstrates a 96% match with real human tissue behavior, and its bone marrow shows a perfect 100% concordance in tests involving 20 different chemotherapy agents.
Recently, Vivodyne, which has successfully raised nearly $80 million across two funding rounds led by Khosla Ventures, inaugurated what it describes as the world's largest "human data center" just outside San Francisco. Georgescu states that his team is already achieving double the throughput of all animal trials conducted across the U.S.
Vivodyne aims to streamline the drug development process by gaining a better understanding of which candidates are likely to succeed prior to the costly clinical trial phase, which typically requires tens of millions of dollars. While it is not disclosing its corporate partners, Vivodyne claims to collaborate with several major pharmaceutical companies to tackle a challenge that Georgescu likens to automotive crash testing—automakers have confidence in their vehicles meeting safety standards before actual testing, unlike drug developers who often lack such assurance going into clinical trials, where most candidates do not achieve FDA approval.
Looking beyond immediate concerns, Georgescu envisions that these autonomous biology labs will play a crucial role in generating the causal data needed to train new AI models focused on human biology. He refers to a recent study published in Nature Methods, which suggests that existing cellular data does not follow clear scaling laws when training generative AI models.
Georgescu highlights that most training currently involves static snapshots of cells, meaning models are unaware of the processes that lead to certain cellular states. “They learn merely that ‘this is cell state A’ and ‘this is cell state B,’ but not that ‘cell state B results from inflaming cell state A,’” he explained.
In contrast, Vivodyne’s HIVE systems are conducting hundreds of thousands of live experiments where diseased tissues are subjected to various stimuli. Georgescu anticipates that this approach will yield the reinforcement learning necessary for AI models to accurately interpret human biology, facilitating more significant advancements in healthcare.
He believes this foundational understanding will be essential not only to address current medical challenges but also for a future where complex diseases will require therapies that target multiple biological pathways simultaneously. Georgescu emphasizes, “To explore combination therapies, the search space expands dramatically—it cannot be approached experimentally. We must pinpoint the effects we want and determine the causes required to achieve them. Establishing causality in human biology is fundamental to all of this.”
