이미지: TechCrunch AI
Summary
- Vivodyne is producing the human-body data needed for AI drug discovery directly, using HIVE, a robotic lab that cultures, doses, and monitors 20 types of human tissue
- The company presented validation results including 94% accuracy in predicting liver toxicity, 96% match in airway tissue response, and 100% match in bone marrow chemotherapy response
- Dario Amodei wrote over the weekend that AI's promises to cure cancer have become more cliché than credible, while Vivodyne's CEO countered that without human data, AI will end up curing cancer only in mice
- 회사
- Vivodyne, 2021년 펜실베이니아대에서 스핀아웃, CEO는 Andrei Georgescu
- 핵심 장비
- HIVE — 20종 인체 조직을 배양하고 자율적으로 투약·모니터링하는 모듈형 로봇 실험실
- 정확도 검증
- 간세포 독성 예측 94%, 기도 조직 일치 96%, 골수 화학요법 20종 100% 일치
- 투자
- 코슬라 벤처스 주도로 두 차례에 걸쳐 약 8천만 달러 미만 유치
- 시설
- 지난주 샌프란시스코 인근에 '세계 최대 인간 데이터센터'라고 자칭하는 시설 개소
- 처리량 주장
- 미국 내 전체 동물실험 대비 2배 처리량 달성
- 업계 통계
- 동물실험을 통과한 신약의 90%가 인체 대상 규제 승인을 받지 못함
AI That Only Cures Cancer in Mice
Anthropic CEO Dario Amodei wrote over the weekend that claims AI will cure cancer have become more cliché than credible. Ironically, he himself made the same argument in a past essay, and Sam Altman has repeatedly cited curing cancer as justification for OpenAI's push toward AGI and its massive compute investments. Google DeepMind's Demis Hassabis also said last year that AI could cure all diseases within a decade.
Actual results have fallen short of such statements. Some AI-designed drugs have reached human clinical trials, and one has advanced to Phase 3, a large-scale human trial, but critics point out that the real barriers in drug development aren't problems today's AI can solve. AlphaFold, which won a Nobel Prize, made major strides in understanding the basic structures of life but has yet to actually deliver a new drug. Isomorphic Labs, built on top of AlphaFold, pushed back its first clinical trial — originally planned for 2025 — to later this year, and in a February announcement said that developing real new drugs requires high-precision predictive models that broadly capture biochemical properties and interactions.
Human Tissue Grown by Robots
Vivodyne frames this gap as a data problem. CEO Andrei Georgescu, who earned his bioengineering PhD at the University of Pennsylvania, spun the company out of the university in 2021. The company's HIVE is a modular robotic lab that cultures 20 types of human tissue, then autonomously administers drugs and monitors the responses. The company's diagnosis is that most of the biological data used to train today's AI models comes from animal testing or single-cell/protein studies, which differ from how living tissue actually responds.
Vivodyne presented the following figures as evidence that its tissue behaves similarly to actual human organs.
| Tissue | Validation Metric | Match/Accuracy |
|---|---|---|
| Liver cells | Toxicity prediction vs. human clinical data | 94% |
| Airway tissue | Match with actual human tissue response | 96% |
| Bone marrow | Match with response to 20 chemotherapy drugs | 100% |
In the pharmaceutical industry, 90% of new drugs that pass animal testing and enter clinical trials fail to win regulatory approval for human use. Georgescu compared this to automotive crash testing, noting that automakers generally enter testing confident they'll pass NHTSA standards, whereas pharmaceutical companies rarely have that same confidence heading into clinical trials.
$80 Million and a "Human Data Center"
Vivodyne has raised just under $80 million across two funding rounds led by Khosla Ventures. Last week, it opened a facility near San Francisco that it calls "the world's largest human data center." Georgescu said the facility's throughput already amounts to twice the total volume of animal testing conducted in the United States. The company hasn't disclosed its partners but says it is working with several major pharmaceutical companies. The goal is to more accurately screen candidate compounds for their likelihood of success before they enter clinical trials, which can cost tens of millions of dollars, thereby accelerating the development path.
Models That Understand Cause and Effect
Georgescu's larger ambition goes beyond a simple validation service. He believes the data generated by HIVE will become the material for training new AI models that understand human biology. He cited a paper published last month in Nature Methods showing that no clear data scaling law emerges when training generative AI models on existing cell data. Georgescu explained that current models only learn static snapshots of cells — distinguishing "state A" from "state B" — without ever learning the causal relationship that "state B is the result of inflammation occurring in state A."
HIVE, by contrast, is simultaneously tracking hundreds of thousands of experiments in which specific stimuli are applied to diseased tissue. Georgescu said he expects this data to accumulate through a reinforcement-learning approach, forming the foundation for models that truly understand human biology. "What can these models do without human data? They'll just cure cancer in mice," he said. His conclusion is that as combination therapies targeting multiple pathways at once become necessary, the space of combinations to explore grows exponentially — making it essential to first understand cause and effect.
Editor's Take
The discourse around AI drug discovery has, for the past few years, rested on the premise that "better models will yield new drugs." Vivodyne's claim shakes that premise itself — the problem isn't the model, it's the data the model is fed. This isn't just a marketing line; it's a wall the industry has actually run into since AlphaFold. Even with a Nobel Prize-winning model, Isomorphic Labs' first clinical trial slipped by over a year, and in the meantime, statements from Amodei, Altman, and Hassabis about "conquering disease within a decade" have piled up without concrete results to back them.
What's notable here is that Vivodyne positions itself as a data company, not an AI company. Over the past two years, most drug-discovery startups have competed on the claim that "our model is more accurate," but Vivodyne treats that competition as essentially meaningless. No matter how good a model trained on animal-testing data gets, if it can't predict human responses, the structure in which 90% of candidates fail at Phase 3 won't change. This isn't a problem solved by throwing more compute at it — it requires rebuilding the experimental infrastructure itself, making it a fundamentally different kind of bet than the "bigger model" strategies pursued by OpenAI and Anthropic.
The practical lesson for domestic pharmaceutical and biotech companies is clear: when investing in AI drug-discovery pipelines, the priority should be checking what data a model was trained on — and whether that data comes from animal testing or actual human tissue — rather than focusing on model performance metrics alone. Building a robotic lab with an initial investment of $80 million isn't something that can be replicated domestically right away, but partnering with organizations that have secured human-tissue-like data ahead of the outsourced clinical-trial stage is an option worth considering now.
Over the coming months, the key thing to watch is whether the outcomes of Vivodyne's partnerships with unnamed "several major pharmaceutical companies" translate into actual clinical drug candidates. Until those results materialize, this announcement will remain in much the same position as Isomorphic Labs' "delayed clinical trial."
