How Are AI Agents Discovering New Drugs? A New AI Scientist Story
What Is the Stanford Virtual Biotech?
Artificial intelligence is becoming a useful tool for scientific research.
Researchers are now using AI to analyze large amounts of information, look for patterns, and explore new ideas for treating diseases.
A recent study from Stanford Medicine shows how this approach could be used on a much larger scale.
Researchers created a system called a Virtual Biotech. It uses many AI agents that can work on different research tasks, similar to how people with different jobs work together in a biotechnology company.
The system was used to analyze clinical trials, study possible drug targets, and propose a possible treatment strategy for lung cancer.
So, what exactly did the AI agents do, and what can this research tell us about the future of drug discovery?
What Happened?
Researchers at Stanford Medicine created a Virtual Biotech, a computer-based research system that uses many AI agents to work on different parts of drug research.
The research was led by James Zou, a Stanford Medicine professor, with Harrison G. Zhang as the first author.
The researchers designed the system to work more like a biotechnology research company than a normal chatbot.
Instead of asking one AI system to handle everything, they created many AI agents with different jobs.
A main AI agent can help organize the work, while other agents can focus on areas such as clinical trials, biology, drug targets, and other research questions.
The research was published in Science under the title “The Virtual Biotech: A Multi-Agent AI Framework for Therapeutic Discovery and Development.”
The study shows how groups of AI agents can work together to handle large amounts of scientific information.
What Does 37,000 AI Agents Mean?
One of the most notable numbers in this research is more than 37,000 AI agents.
But this number needs some explanation.
The researchers created 37,075 clinical-trialist AI agents to collect and organize information from clinical trials.
These agents were given specific tasks rather than all working on the same problem.
For example, an AI agent could analyze information from an individual clinical trial and organize its results.
The overall study examined 55,984 clinical trials.
Stanford Medicine also reported that the system was able to analyze and organize information from about 50,000 clinical trials in less than a week.
In simple terms, the researchers divided a huge research task into many smaller tasks and allowed many AI agents to work on them at the same time.
It is similar to having a very large research team where each person is given a smaller part of the work.
The difference is that these workers are AI agents.
How Does the Virtual Biotech Work?
 |
Figure 1. AI agents working together to analyze biomedical data and explore potential treatment strategies |
The Virtual Biotech was designed to organize AI agents in a way that resembles a real biotechnology research company.
A main AI agent can help coordinate the research and assign tasks to specialized AI agents.
Other agents can then focus on different questions.
For example:
- Some agents can study clinical trials.
- Some can examine biological information.
- Some can investigate possible drug targets.
- Others can compare information from different sources.
The results can then be combined to help answer a larger scientific question.
This is different from using a normal AI chatbot.
A chatbot usually responds to one request at a time.
The Virtual Biotech is designed to let many AI agents work on different parts of a complicated research problem at the same time.
→ Source: Harrison G. Zhang — AI Research for Biomedicine
What Did the AI Agents Study?
One major part of the research involved clinical trials.
A clinical trial is a study that tests a medical treatment in people to learn how well it works and how safe it is.
The AI agents analyzed information from tens of thousands of clinical trials.
The researchers wanted to find patterns that might help explain why some drugs have better development outcomes than others.
The system also examined information about genes and different types of cells.
This allowed the researchers to ask an important question:
Are there biological features that are connected with a drug's chance of success?
The analysis found some interesting patterns.
What Did the AI Find?
The researchers looked at two characteristics related to drug targets.
The first was cell-type specificity.
In simple terms, this describes how closely a drug target is connected to one particular type of cell.
The second was called bimodality.
This term sounds complicated, but the basic idea can be explained with a light switch.
Some biological activity can behave more like an on/off switch, while other activity can behave more like a dimmer with many levels.
The researchers found that certain combinations of these characteristics were associated with better drug-development outcomes.
According to Stanford Medicine, drugs with these characteristics were:
- 40% more likely to move from Phase 1 to Phase 2
- 48% more likely to reach the market
- Associated with 32% fewer adverse events
These results describe patterns found in the researchers' analysis.
They do not mean that AI can predict with certainty whether a new drug will succeed.
Instead, the analysis suggests that certain biological features may help researchers identify more promising drug targets.
How Did AI Investigate a Lung Cancer Target?
The researchers then asked the Virtual Biotech to investigate B7-H3, a possible target for lung cancer treatment.
B7-H3 is a protein found on certain cells.
The researchers were interested in understanding which cells had high levels of B7-H3 and what those cells might be doing around a tumor.
The AI agents examined information from different biological datasets and studied how different cells communicate with each other.
They found evidence suggesting that cells with high levels of B7-H3 may communicate with nearby immune cells in a way that can reduce immune activity around a tumor.
The system then used this information to propose a possible treatment strategy.
What Is an Antibody-Drug Conjugate?
The proposed strategy involved an antibody-drug conjugate, or ADC.
The name sounds complicated, but the basic idea is fairly simple.
An ADC combines an antibody with a powerful drug.
The antibody acts like a guide.
It is designed to recognize a particular target on certain cells.
The attached drug can then deliver a powerful treatment to those targeted cells.
In the B7-H3 example, the Virtual Biotech proposed using an antibody to target cells with high levels of B7-H3 and deliver a drug designed to damage those cells.
In simple terms:
Find the target → attach the treatment → deliver it to the target cells.
This was a proposed treatment strategy, not a finished medicine created by the AI.
Did AI Create a New Cancer Drug?
This is one of the most important points to understand.
It would be misleading to say that AI created a new cancer drug that is now approved for patients.
That did not happen.
The Virtual Biotech analyzed existing information and proposed a possible B7-H3 antibody-drug conjugate strategy.
The researchers then compared this result with later developments outside their own project.
Importantly, the AI system was limited to information available before January 2025.
Later, in August 2025, a pharmaceutical company independently arrived at a similar B7-H3 antibody-drug conjugate strategy.
→ Source: GSK — B7-H3-targeted antibody-drug conjugate GSK'227
The Stanford researchers described this as an independent third-party validation that was consistent with the strategy proposed by the Virtual Biotech.
However, this does not mean that the pharmaceutical company used Stanford's AI system to create its treatment.
The two developments were separate.
Was the AI-Designed Treatment Tested in Patients?
No.
The Stanford Virtual Biotech did not take its proposed treatment and test it directly in human patients.
A possible treatment idea still has to go through many steps before it can become a real medicine.
Scientists need to test the idea in laboratories.
Researchers then need to study its safety and effectiveness.
If the treatment continues to show promise, it must go through clinical trials and regulatory review.
This is an important difference between finding a promising idea and developing an approved medicine.
The Virtual Biotech can help researchers explore possible ideas, but real-world testing is still necessary.
What Does FDA Breakthrough Therapy Designation Mean?
You may also see the term Breakthrough Therapy Designation when reading about some B7-H3 treatments.
This is an FDA designation for certain drugs being developed for serious diseases when early clinical evidence suggests that they may provide a meaningful improvement over existing treatments.
However, it is not the same as FDA approval.
→ Source: U.S. Food and Drug Administration — Breakthrough Therapy
A drug receiving Breakthrough Therapy Designation has not automatically become an approved medicine.
This distinction is important when reading news about new treatments.
Why Is This Research Important?
Drug research requires scientists to work with huge amounts of information.
Researchers may need to examine clinical trials, genetic information, cell data, disease research, and many other sources.
A human research team cannot easily examine all of this information at the same speed.
The Virtual Biotech demonstrates one possible way to use AI to handle some of this work.
Instead of asking one AI system to do everything, researchers can create many AI agents and give each one a specific task.
The agents can work at the same time and then bring their results together.
This could help scientists explore more ideas and analyze large amounts of information more quickly.
However, the research does not show that AI has replaced scientists.
It shows how AI could become a powerful research tool.
What Can AI Agents Do in Drug Research?
This study shows several ways AI agents could support drug research.
They can help:
- Analyze large numbers of clinical trials
- Organize information from scientific studies
- Look for patterns in biological data
- Investigate possible drug targets
- Compare evidence from different sources
- Suggest possible treatment strategies
The important word is help.
AI can process information and suggest ideas, but researchers still need to check whether those ideas are correct.
What Are the Limitations?
AI research systems also have important limitations.
First, AI depends on the information available to it.
If the data are incomplete or poor quality, the results can also be limited.
Second, a computer-based prediction does not automatically mean that something will work in a real human body.
Biology is extremely complicated.
A treatment that looks promising in computer analysis may fail during laboratory testing or clinical trials.
Third, AI-generated ideas still need human review.
Scientists need to decide whether a result makes biological sense and whether it is worth testing.
The researchers also plan to take other findings from the Virtual Biotech into real laboratory research.
This shows why computer analysis and real-world experiments are both important.
Can AI Agents Replace Scientists?
This study does not show that AI agents can replace scientists.
Instead, it shows how AI agents can take on large amounts of research work.
They can analyze information, compare data, find patterns, and suggest ideas.
Human researchers are still needed to evaluate those ideas, design experiments, perform laboratory work, interpret results, and conduct clinical research.
A simple way to think about it is:
AI can help researchers search for promising ideas. Scientists still have to test those ideas.
What Should Beginners Know?
The biggest lesson from this research is not simply that “AI discovered a new drug.”
A more accurate description is:
Researchers created a large team of AI agents that could analyze scientific information, find useful patterns, and propose a possible treatment strategy.
The system did not independently create an approved medicine.
It also did not remove the need for laboratory experiments or human clinical trials.
What makes the research interesting is the scale and organization of the AI system.
Thousands of AI agents can divide a large research problem into smaller tasks and work on many of them at the same time.
This could become an important way for scientists to use AI in future drug research.
Key Takeaways
- Stanford researchers created a Virtual Biotech where many AI agents work on different research tasks.
- More than 37,000 clinical-trialist AI agents were used to analyze clinical-trial information.
- The research examined 55,984 clinical trials.
- The system found biological patterns associated with better drug-development outcomes.
- The Virtual Biotech proposed a possible B7-H3 treatment strategy, but it did not create an approved drug.
- The study shows how AI agents could help scientists analyze large amounts of information and explore new treatment ideas.
FAQ
Q. What is the Stanford Virtual Biotech?
A. The Virtual Biotech is a research system that uses many AI agents to work on different parts of drug research.
Q. What does 37,000 AI agents mean?
A. More than 37,000 specialized AI agents were used to collect and organize information from clinical trials. They were given specific tasks rather than all working on the same problem.
Q. Did AI create a new cancer drug?
A. No. The Virtual Biotech proposed a possible B7-H3 treatment strategy. It did not create an approved cancer drug.
Q. Was the AI's B7-H3 strategy independently supported?
A. Stanford researchers reported that a pharmaceutical company later independently arrived at a similar B7-H3 antibody-drug conjugate strategy. They described this as an independent result consistent with the AI system's proposed strategy.
Q. Can AI agents replace scientists?
A. The study does not show that. AI agents can help with large-scale research and data analysis, but human scientists are still needed for experiments, clinical research, and medical decisions.
Conclusion
The Stanford Virtual Biotech shows how AI agents can be organized into a large research system instead of being used only as individual assistants.
The system was able to analyze huge amounts of clinical and biological information and propose a possible treatment strategy.
But the research also shows an important limitation: a promising AI result is only the beginning of the drug-development process.
AI can help scientists find and explore promising ideas, while laboratory research and clinical trials are still needed to determine whether those ideas can become real treatments.