Stanford's all-agent biotech put one AI agent on each of 37,075 clinical trials
James Zou's agent company read the results of tens of thousands of finished trials and found that drugs aimed at switch-like, cell-type-specific targets reached market about 48% more often. No experiment has tested that yet.
The Scientist · Science desk
What happened
- A paper describing the virtual biotech company appeared in Science on September 17, with Stanford's James Zou as senior author and graduate student Harrison Zhang as lead author.
- The setup copies the organizational chart of an established biotech, with a chief science officer agent leading specialized divisions that work in parallel on molecular target identification and clinical-trial design.
- In a lung-cancer demonstration the system confirmed the protein CD276 as a target from previously collected data and designed a CD276-recognizing antibody tethered to an anticancer drug.
Why it matters
- What it limits constraint Nature reports the predictions were never validated through experiments, so the CD276 design remains a hypothesis that some wet lab has to fund and run before anyone can price it.
- What doesn't add up contradiction phys.org credits the agents with a cancer therapy a major pharmaceutical company later built independently; Nature describes the same lung-cancer work as prompted by Zou's team and blessed by outside reviewers. The two accounts assign design credit differently.
- What's now possible capability Because Zou says the agents do not depend on one vendor's model, a group with local hardware and an open-source model can attempt the same trial-cataloguing screen without buying frontier API capacity.
- Who has to decide decision Trial sponsors weighing whether to add single-cell gene activity measurement now have a specific reason to spend on it, since the scores that predicted success can only be computed where that data exists.
Next story · Product · Oct 7
Cisco builds Webex agents that keep working after the customer hangs upConfirmed2 publishers
The 37,000 figure tracks the reading list. Nature reports that the chief scientific officer agent handed out 37,075 agents, one to each later-stage trial [4], and 37,075 rounded to the nearest thousand gives the 37,000 "employees" in the Stanford account, none of them human and none on a payroll [2][19]. Staffing followed the size of the corpus. The divisions describe how the work was split.
Zou did not turn a swarm loose on the literature. Each agent took a single trial and retrieved that trial's safety and effectiveness data [7]. Every output stays attached to one named study. Zou said the agents analyzed and catalogued some 50,000 trials in less than a week, work he put at years for human staff [6]. Nature describes the corpus as more than 55,000 published trial results [5].
The predictive signal came from agents working only on trials that had single-cell gene activity data [8]. They built one score for how narrowly a drug hit a specific cell type and a second for bimodality, meaning whether the targeted gene's activity behaves like an on-off switch or a dimmer [8]. Drugs against switch-like targets advanced from phase 1 to phase 2 40% more often, reached market 48% more often, and had 32% fewer adverse events than broad-acting drugs [9]. The pattern held across cancers, brain diseases, heart diseases, kidney and lung conditions [17]. Nature gives the market figure as nearly 50% and ties it to cell-type specificity; the Stanford account ties it to switch-like gene behaviour [10]. Neither account gives the number of catalogued trials that carried the single-cell data the scores needed [22].
The comparison runs across trials that already reported results, so it ranks features that approved drugs happened to have. Zou's explanation is mechanistic: a target that acts like a switch and sits in one cell type may be easier, and therefore safer, to control with a drug [20]. "The science the agents discovered is really exciting, and it shows that these single-cell features can be used to make better drugs. It points to the importance of collecting this kind of data," he said [11].
The lung-cancer demonstration had a human hand at both ends. Zou's team directed the CSO agent to investigate CD276, a protein that earlier work had linked to dampened immune responses and high expression in lung tumours [14]. The system confirmed the candidate using previously collected data and developed a strategy, a CD276-recognizing antibody tethered to an anticancer drug [15]. Zou and his collaborators, helped by external reviewers, concluded this was a promising avenue [16]. Nature reports that other scientists note the Virtual Biotech has not been vetted in real-world drug discovery and that its predictions were not validated through experiments, let alone clinical trials [13]. The Stanford account describes the same output more warmly, as a cancer therapy that a major pharmaceutical company later independently built [21].
The agents ran on versions of Claude, from Anthropic, and Zou says any advanced large language model will do, including open-source models researchers can run on their own computers [18]. "We want to see how far these agent teams of AI scientists can help us to really accelerate drug discovery and development," Zou told Nature [12].
What to watch
- Wet-lab or animal data on the CD276 antibody-drug conjugate, which would move it from reviewer judgement to experiment.
- A prospective test of whether candidates chosen on the switch-like, cell-type-specific score clear phase 2 more often.
- An ablation running the same trial corpus through a flat pool of agents, with no CSO and no divisions, to show whether the hierarchy mattered.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence58
- Adoption22
- Hype gap+32
- Incentives60
- Confidence62
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
A paper describing the virtual biotech company was published Sept. 17 in Science. James Zou, associate professor of biomedical data science at Stanford Medicine, is senior author and graduate student Harrison Zhang is lead author.
- [2]
The virtual biotech company has 37,000 employees and none of them are human; there is no lab space, no lunch breaks and no payroll.
- [3]
The virtual company mirrors the organizational chart of an established brick-and-mortar biotech: a chief science officer agent leads research teams broken into multiple specialized divisions that work in parallel on core elements of drug design such as identifying molecular targets and designing clinical trials.
- [4]
In the Science study, the chief scientific officer agent assigned 37,075 agents to each tackle a single later-stage trial.
- [5]
Zou's team tasked the Virtual Biotech with analysing the published results of more than 55,000 clinical trials run for drugs for a wide range of conditions.
- [6]
The agents analyzed and cataloged some 50,000 trials in less than a week; Zou said that task would have taken human agents years.
- [7]
Instead of unleashing a cadre of agents into repositories of scientific literature, Zou assigned a single agent to analyze a specific clinical trial and retrieve relevant data about safety and effectiveness.
- [8]
For trials with single-cell gene activity data available, the agents built two scoring systems: one evaluating how specifically a drug targeted a certain cell type, and one measuring bimodality, indicating whether a targeted gene's activity is more like an on-off switch (high bimodality) or a dimmer.
- [9]
Drugs that targeted switch-like genes were 40% more likely to advance from phase 1 to phase 2 trials, were 48% more likely to reach market and had 32% fewer adverse events compared with those that had a broad spectrum of activity.
- [10]
Nature reports that the analysis found drugs targeting proteins active in specific cell types were nearly 50% likelier to reach market, compared to other drugs.
- [11]
"The science the agents discovered is really exciting, and it shows that these single-cell features can be used to make better drugs. It points to the importance of collecting this kind of data," Zou said.
- [12]
"We want to see how far these agent teams of AI scientists can help us to really accelerate drug discovery and development," says James Zou, a computer scientist at Stanford University in California who led the effort.
- [13]
Other scientists note that the Virtual Biotech has not been vetted in the crucible of real-world drug discovery, and its predictions were not validated through experiments, let alone clinical trials.
- [14]
Zou's team directed the CSO agent to investigate whether a protein called CD276 would make a good therapeutic target for lung cancers; previous work had suggested CD276 dampens immune responses and is highly expressed in lung tumours.
- [15]
The system confirmed CD276 as a candidate using previously collected data and developed a strategy to target it: a CD276-recognizing antibody tethered to an anticancer drug.
- [16]
With the help of external reviewers, Zou and his collaborators concluded that the CD276 approach was a promising avenue.
- [17]
The patterns persisted for a variety of conditions, including cancers, brain diseases, heart diseases, kidney and lung conditions.
- [18]
The agents were powered by versions of Claude, developed by Anthropic in San Francisco, and Zou says any advanced LLM will do, including open-source models that researchers can run on their own computers.
- [19]
The 37,000 headcount reported for the virtual company is the per-trial agent count, 37,075, rounded to the nearest thousand.
- [20]
Zou's theory is that a target that behaves like an on-off switch and homes in on a specific cell type may be easier, and therefore safer, to control with a drug than one with a spectrum of activity.
- [21]
The virtual biotech company has uncovered a biological signal that predicts which drug candidates are more likely to succeed and has designed a cancer therapy that a major pharmaceutical company later independently built.
- [22]
Neither published account states how many of the catalogued trials had single-cell gene activity data available for the two scoring systems.
Sources
2 independent publishers whose own reporting we read for this story.
- nature.comHow a team of AIs discovered a promising lung-cancer drug
1 article · September 16, 2026
- phys.orgVirtual biotech company puts thousands of AI scientist agents to work on drug discovery
1 article · September 17, 2026
Topics and entities
Topics
- Agentic AI and tool use
- AI Drug Discovery
- Clinical Trial Evidence
- Single-Cell Genomics
- Antibody-drug conjugates