Can a virtual biotech find new lung cancer treatments?
A research team has built a ‘virtual biotech’ in which thousands of AI agents work together on different parts of drug discovery.
In a study published in Science, the system analysed evidence from 55,984 clinical trials. The researchers also asked it to investigate B7-H3, a protein already being studied as a possible treatment target in lung cancer.
The results are interesting. But the AI did not discover B7-H3, produce a new drug or prove that a treatment works. Its proposal has not yet been tested experimentally.
What did 37,000 AI agents actually do?
The Virtual Biotech was designed to work like a biotechnology company.
A ‘chief scientist’ AI assigned work to teams covering areas such as treatment targets, drug design and clinical trials. More than 37,000 agents each reviewed a later-stage clinical trial.
The researchers then gave the system a more specific task: assess whether CD276, more commonly known as B7-H3, could be a useful treatment target in lung cancer.
B7-H3 was not a new discovery. Previous research had already found high levels of the protein in some lung tumours and suggested that it may help cancer escape the immune system.
The system searched existing genetic, single-cell, spatial and clinical data. It supported B7-H3 as a possible target and proposed using an antibody-drug conjugate, or ADC, to attack it.
An ADC combines an antibody that recognises a particular target with a cancer-killing drug. The antibody is intended to carry the drug towards cells bearing that target.
Our article on TROP2 and lung cancer looks more closely at how ADCs work and the questions surrounding their use.
B7-H3 is already being tested in small cell lung cancer
At the 2026 World Conference on Lung Cancer, interim Phase 3 results were presented for two separate B7-H3-targeted ADCs in relapsed small cell lung cancer.
In the TAISHAN-302 trial, tambotatug pelitecan improved median overall survival to 13.3 months, compared with 9.4 months for topotecan.
In the ARTEMIS-008 trial, median overall survival was 18.5 months with risvutatug rezetecan and 10.3 months with topotecan.
Both trials were conducted in China. Further studies are needed to establish whether the results apply as strongly to wider populations, including people in Europe.
These drugs were developed independently of the Virtual Biotech study. The results do not validate the AI system or prove that its proposal would work. They do, however, show why B7-H3 is receiving attention in lung cancer research.
This comes at a time when treatment for small cell lung cancer is beginning to change after decades of limited progress. Our article on new small cell lung cancer treatments in Europe explains some of the recent developments and why European approval does not automatically mean access.
Our recent small cell lung cancer webinar also covered access to trials, new treatments and the support people need when decisions must be made quickly.
The Virtual Biotech has not produced a treatment
The study shows what a large team of AI agents can do with existing information. It does not show that their proposed treatment is safe or effective.
No new B7-H3 drug from the Virtual Biotech has been made. Its proposal has not been tested in laboratory experiments, animal studies or clinical trials.
Human researchers also remained involved. They chose the lung cancer target for the system to investigate, reviewed its work and brought in external specialists to assess the proposal.
Drug development is not simply a search for a plausible target. A potential treatment must still be made, tested and compared with existing care. Many apparently promising ideas fail during that process.
Could AI shorten lung cancer drug development?
The possible value lies in speed and scale.
A human research team cannot divide tens of thousands of clinical trials between tens of thousands of researchers. An AI system can organise work at that scale and search for connections across large amounts of evidence.
That could help researchers decide which targets deserve closer attention, identify reasons why previous trials succeeded or failed and narrow down the ideas taken into laboratory research.
It could also produce convincing mistakes at scale. The quality of its conclusions depends on the evidence it receives, the questions it is asked and the checks carried out by people with the relevant expertise.
Lung Cancer Europe’s position statement on AI in lung cancer care calls for transparency and evidence, keeping human judgement at the centre and involving people affected by lung cancer in decisions about how AI is used.
Those principles apply to drug research as much as they do to clinical care.
Who benefits from faster lung cancer research?
Finding a possible treatment target more quickly is only one part of the job.
The two B7-H3 Phase 3 trials were conducted in China. Further evidence, regulatory decisions and national reimbursement processes will determine whether these treatments eventually reach people in Europe.
Access to new lung cancer treatments already differs widely between European countries. The Lung Cancer Europe Access to Treatment Atlas allows people to compare reimbursement, biomarker testing, screening and clinical trial information across Europe.
AI may help researchers move through early drug development more quickly. It cannot by itself fix slow approvals, unequal access to testing, limited trial availability or differences in national funding.
The same questions arise elsewhere in medical AI. Our article on AI imaging and Cancer Image Europe looks at how better data could support research and diagnosis, and why representation, governance and access need attention from the start.
What happens next?
The Virtual Biotech now needs to prove itself outside a computer system.
Its B7-H3 proposal would need experimental testing before anyone could know whether it could become a treatment. The system must also be tested on other research questions to see how reliably it performs.
For people affected by lung cancer, the test is straightforward: does this approach help produce safe and effective treatments, and do those treatments reach the people who need them?
That answer will come from laboratory work, clinical trials and access in practice, not from the number of AI agents involved.