Summary
A virtual biotech using up to 37,000 AI agents identified CD276 as a lung-cancer target and proposed an antibody linked to an anticancer drug. The work is at the computational discovery stage, with experimental and clinical validation still ahead.
A virtual biotech comprising as many as 37,000 artificial-intelligence agents has identified a potential lung-cancer treatment strategy: an antibody that recognises the protein CD276 and is tethered to an anticancer drug. The result comes from a computational drug-discovery study and is still at the stage of a proposed candidate rather than a treatment tested in patients.
The system also found a signal in clinical-trial data suggesting that drugs aimed at proteins active in particular cell types were nearly 50% likelier to reach the market than other drugs. The findings illustrate how large groups of specialised AI agents can divide up parts of the drug-development process, while leaving the biological testing required for a medicine to advance.
A virtual biotech organised like a company
The system was designed to imitate the structure of a biotechnology company. A “chief scientific officer” agent directed other agents working in areas such as therapeutic-target identification and clinical-trial design. The agents could interact with large language models or with one another and carry out multistep tasks.
For the analysis of clinical-trial outcomes, the team assigned 37,075 agents to examine the published results of individual later-stage trials. The trials covered drugs developed for a wide range of conditions and numbered more than 55,000 in total. Other agents searched datasets showing which genes were active in different cell types.
That analysis produced a possible marker of drug-development success. Drugs targeting proteins active in specific cell types were reported to be nearly 50% likelier to reach the market than other drugs. This is a predictive signal from analysis of existing data, rather than a guarantee that targeting any particular protein will produce a successful medicine.
The team used versions of Anthropic’s Claude as the underlying large language model. The system’s lead researcher, James Zou of Stanford University, said the approach could also use other advanced language models, including open-source models run on researchers’ own computers.
From clinical data to a CD276 treatment strategy
In a separate demonstration, the AI team investigated whether CD276 could be a useful therapeutic target for lung cancer. Earlier work had suggested that the protein dampens immune responses and is produced at high levels in lung tumours.
Using previously collected data, the system supported CD276 as a candidate target and proposed a targeted-drug strategy. The design links a CD276-recognising antibody to an anticancer compound, allowing the antibody to provide the molecular recognition while carrying the drug payload. With assistance from external reviewers, the researchers judged this approach to be promising.
Human oversight was part of the process: people set the system’s objectives, reviewed its output and helped assess the proposed treatment strategy. The study therefore demonstrates an AI-organised research workflow, rather than an autonomous route from an idea to an approved medicine.
The candidate still requires laboratory experiments to determine whether the CD276-linked design behaves as predicted, followed by the testing needed to assess safety and effectiveness in humans. The report describes no experimental validation or clinical trial of the proposed treatment. Whether CD276 can lead to a useful lung-cancer therapy remains an open research question.