A drug program usually begins with a room full of people arguing. Someone from target discovery thinks a protein looks promising. Someone from safety points out that the protein shows up in healthy tissue too. Someone from clinical operations asks what the trial would even look like, and how many years it would take before anyone knows whether the argument was worth having. That friction is slow, expensive, and — most pharmaceutical scientists would tell you — indispensable.
A team at Stanford Medicine has now built a version of that room where nobody is human. On September 17, the journal Science published a description of what its creators call the Virtual Biotech: a system of as many as 37,000 artificial intelligence agents, arranged into corporate divisions, reporting up to a chief scientific officer agent. It read through the published results of more than 55,000 clinical trials. It found a statistical signal that predicts which drugs make it to market. And it designed a lung cancer therapy that a major pharmaceutical company, working independently, later built for itself.
What Happened
The project comes out of the lab of James Zou, an associate professor of biomedical data science at Stanford University, with graduate student Harrison Zhang as first author. Rather than build one very capable AI scientist, Zou's team built an organizational chart. A CSO agent sits at the top. Below it are divisions that mirror the ones inside a real biotech: target discovery, molecule design, safety and clinical trials. Within each division, agents specialize further — one reads genetics data, another handles genomics and single-cell data, and so on down the line.
To test whether the structure did anything useful, the team pointed it at a problem that has frustrated the industry for decades: figuring out in advance which drug candidates will survive. The CSO assigned 37,075 agents to the task, one agent per late-stage trial, and had them synthesize results scattered across more than 55,000 studies. The agents then cross-referenced those outcomes against data showing which genes are switched on in which cell types.
The pattern that emerged was specific and quantifiable. Drugs aimed at proteins that behave like switches — active in a narrow set of cell types rather than broadly expressed across the body — were roughly 40% more likely to advance from phase 1 to phase 2, about 48% more likely to reach market, and associated with 32% fewer adverse events. That is not a subtle effect. It is close to a coin flip's worth of difference in a business where most candidates fail.
The second demonstration was narrower and, for most readers, more striking. The team asked the CSO whether a protein called CD276 would make a sensible target in lung cancer. CD276 — also known as B7-H3 — suppresses immune responses and shows up heavily in lung tumours while staying relatively scarce in healthy tissue, which is exactly the switch-like profile the agents had just learned to favour. The system confirmed the target using existing data and then designed a treatment strategy: an antibody that recognises CD276, chemically tethered to a cancer-killing payload. An antibody-drug conjugate, in the industry's vocabulary.
The agents worked only from literature published before January 2025. Several months after the design was finished, Merck independently developed and validated the same therapeutic concept, which went on to receive breakthrough therapy designation from the U.S. Food and Drug Administration. Zou has described this as external, third-party validation of what his agents produced — a claim that is unusual in AI research, where benchmarks are usually graded by the people who wrote them.
Why It Matters
The reflexive way to read this result is as a story about speed. Tens of thousands of workers who never sleep, never take meetings, never go home — the pharmaceutical executive's fantasy, as the coverage in Nature put it. But the more interesting claim buried in the work is architectural, and it cuts against where most of the industry has been spending its money.
For two years the dominant assumption has been that capability scales with the model: build a bigger, smarter single system and it will eventually do everything a team of specialists could. Zou's group ran the comparison directly, putting a multi-agent team against a single agent on the same scientific problem. The distributed version won, and the reason it won is almost sociological. Agents in the virtual lab disagree with each other. They have to argue their positions. Zou has said the agents "get into debates and disagreements," and that this friction produced more creative reasoning and fewer compounding errors than one model reasoning alone.
That framing has consequences outside biology. If the returns come from the structure rather than the individual model, then the engineering problem shifts from fine-tuning weights to designing environments — the shared context, the incentives, the guardrails that let thousands of agents work on the same open-ended question without collapsing into noise. Zou's team hit that wall almost immediately. Scientific databases were built for humans and for pre-AI software, and wrapping them in a protocol layer does not fix an interface that was never meant for agents in the first place. Their answer was a tool called Paperclip, which maps scattered databases and unstructured papers into a unified file system that agents navigate the way a programmer navigates a repository. The claimed result was better accuracy at more than an order of magnitude less time and cost.
There is also a quieter precedent here. Before the Virtual Biotech, the same group ran a Virtual Lab of five to eight agents modelled on Zou's actual Stanford team, complete with an AI principal investigator. That smaller system designed nanobody proteins targeting recent COVID variants, and when they were made and tested in a wet lab, the AI-designed versions bound better than the human-designed ones they were compared against. The scale-up from eight agents to 37,000 was not a leap of faith. It was an extrapolation from something that had already survived contact with a laboratory bench.
Reaction
The enthusiasm has been real, and so has the hedging. The Science paper arrived the same week that another Stanford Medicine group published work turning scientific manuscripts into AI agents that can talk to each other and run analyses, and the two together have been read as evidence that the AI-scientist genre has moved past demos. Coverage in the trade press has leaned hard on the Merck coincidence, which is the kind of detail that travels well.
The objection from within the field is narrower and harder to wave away: the Virtual Biotech has not been tested in the actual crucible of drug discovery. Its predictions were not validated through new experiments, and certainly not through clinical trials. The CD276 design was checked by external reviewers and matched by Merck's independent programme, but no molecule from the system has been synthesised, dosed in an animal, or given to a patient on the strength of the agents' recommendation. The 48% figure is an association mined from historical trials, not a prospective forecast that has been allowed to be wrong.
The CD276 example also invites a fair question about difficulty. B7-H3 was not an obscure target. Antibody-drug conjugates aimed at it — ifinatamab deruxtecan, HS-20093, and others — have been in clinical development for years, and the protein has been described in the literature as a promising immune checkpoint across solid tumours. The agents were given a tip-off that previous work had flagged CD276, and they confirmed and elaborated it. That is a genuine and useful thing for a system to do. It is not the same as pulling a target out of nowhere, and the distinction matters when the headline says an AI discovered a drug.
What's Next
The honest next step is the one the critics named: a prospective test. Let the system pick targets before anyone knows the answer, commit to those picks in public, and see how the predictions hold up over the years it takes for trials to read out. Everything published so far is retrospective, and retrospective analysis of 55,000 completed trials is an exercise with a known answer key. The switch-like-gene signal is a hypothesis about the future dressed in data from the past.
There is a regulatory dimension too. The FDA has spent the past two years building out its position on AI in drug development, qualifying its first AI-based drug development tool and deploying agentic systems internally. A design that emerged from an autonomous multi-agent system and converged with a programme that earned breakthrough designation is a useful case study for regulators trying to work out what evidence they should ask for when the provenance of a molecule includes tens of thousands of language models. Nobody has written that rulebook yet.
And there is the infrastructure question, which is less glamorous and probably more consequential. Paperclip is open source. If the bottleneck on agentic science really is the interface between agents and the world's accumulated data — rather than the intelligence of any individual agent — then the unglamorous work of making databases legible to machines is where the next round of gains will come from. That is a problem every scientific field has, not just drug discovery.
Closing Thoughts
It is worth sitting with what the Merck detail actually demonstrates. Two independent processes — one a pharmaceutical company with chemists and budgets and a decade of institutional memory, the other a swarm of language models reading papers published before 2025 — arrived at the same therapeutic design. The optimistic reading is that the agents reproduced expert judgment at a fraction of the cost. The deflationary reading is that both were converging on a conclusion the published literature was already pointing toward, and that the agents were very good at reading the room.
Both readings can be true, and the second one is not an insult. A great deal of scientific work consists of noticing what the accumulated evidence already implies but nobody has yet assembled. If a system can do that reliably, across 55,000 trials and millions of papers, it is doing something researchers cannot do at that scale. The question is whether it can also do the other thing — the leap that contradicts the literature and turns out to be right. Nothing in this paper tests that, and it is not obvious that a system trained on consensus would be the place to look for it.
What lingers is the shape of the finding rather than the finding itself. The team did not make a better scientist. They made a worse one, many times over, and then built an institution around it — divisions, supervisors, meetings, disagreements, a place for arguments to happen. The productive unit turned out to be the organisation, not the individual. Anyone who has worked in a research group already suspected that. It is strange and a little clarifying to see it hold when none of the participants are people.
The drug, for what it is worth, still has to work.
한글 요약
스탠퍼드 의대 제임스 저우(James Zou) 부교수 연구팀이 최대 3만 7천 개의 AI 에이전트로 구성된 가상 바이오텍 회사를 만들어 사이언스 9월 17일자에 발표했습니다. 최고과학책임자(CSO) 역할을 하는 에이전트 아래에 신약 표적 발굴, 분자 설계, 안전성·임상시험 부서를 두고 실제 제약사의 조직도를 그대로 흉내 낸 구조입니다. 대학원생 해리슨 장이 제1저자로 참여했습니다.
이 시스템은 5만 5천 건이 넘는 임상시험 결과를 분석해, 특정 세포 유형에서만 켜지는 스위치형 단백질을 표적으로 삼은 약물이 시판 승인까지 도달할 확률이 약 48% 높고 부작용은 32% 적다는 신호를 찾아냈습니다. 이어 폐암 표적으로 CD276(B7-H3) 단백질을 지목하고, 이 단백질을 인식하는 항체에 항암 약물을 결합한 항체-약물 접합체(ADC) 설계안을 내놓았습니다. 에이전트들은 2025년 1월 이전에 공개된 자료만 사용했는데, 몇 달 뒤 머크(Merck)가 독자적으로 같은 설계를 개발해 FDA 혁신치료제 지정을 받았습니다.
다만 한계도 분명합니다. 이 시스템의 예측은 아직 새로운 실험이나 임상시험으로 검증되지 않았고, 48%라는 수치도 이미 끝난 시험들을 되돌아본 연관성 분석에 가깝습니다. CD276 역시 이전부터 여러 제약사가 ADC 후보로 개발해 온 표적이어서 완전히 새로운 발견이라고 보기는 어렵습니다. 그럼에도 저우 교수팀이 보여준 것은 모델 하나를 키우는 대신 에이전트들이 서로 논쟁하고 반박하는 조직을 설계했을 때 더 나은 결과가 나온다는 점입니다. 성과의 핵심은 개별 지능이 아니라 그 지능들이 놓인 환경에 있었습니다.
참고: Nature · Stanford Medicine News · VentureBeat · bioRxiv