Researchers built a multi‑agent system that parsed tens of thousands of clinical trials, produced a gene‑target scoring scheme tied to trial success, and independently proposed a lung cancer approach a pharma later pursued.
The system has built-in access to the Open Targets database, a massive public repository of clinical trial data.
Now, researchers at Stanford have built a virtual biotech company with 37,000 AI agents that work together to analyze drug targets and design therapies.
# What the study built Stanford researchers assembled a "virtual biotech" made up of as many as 37,000 specialized AI agents to analyze drug targets and design therapeutic strategies. The system models organizational roles you'd find in a real company: a virtual chief scientific officer (CSO) delegates tasks to divisions focused on target discovery and validation, safety assessment, delivery strategy, and clinical‑trial review.
# How the system works A human user submits a query and the CSO breaks it into jobs for specialist agents. Agents have their own tools and databases and can search registries, papers, press releases, and public repositories. The system includes built‑in access to the Open Targets database of trial data. For one test, the CSO assigned a separate agent to each of 37,075 Phase II and III trials to determine outcomes that weren't clearly recorded in the public record. That sweep took roughly six hours.
# What the agents discovered about target selection
# The B7‑H3 example The team asked the system to evaluate B7‑H3, a protein associated with lung cancer. Agents found B7‑H3 was common in fibroblasts near tumors and assembled evidence those fibroblasts suppress nearby immune cells. Based on available data up to January 2025, the virtual biotech proposed tagging B7‑H3‑expressing cells with an antibody to deliver a toxic payload—an antibody‑drug conjugate approach.
A similar strategy was pursued independently by a major pharmaceutical company: an antibody‑drug conjugate targeting B7‑H3, ifinatamab deruxtecan, was granted FDA breakthrough therapy status in August 2025. The researchers described that parallel as an independent validation of the system's reasoning.
# Strengths and limits Strengths: the approach scales human literature review and data curation quickly, unifies dispersed evidence across trials and tissue databases, and generates testable target hypotheses and scoring rules that can reprioritize candidates.
Limits: proposing targets is only one step. Wet‑lab validation, toxicology, and multi‑phase clinical trials remain necessary and time‑consuming. Better candidate selection can reduce obvious dead ends but cannot accelerate the formal experimental and regulatory steps required to bring a drug to market.
# Practical takeaway A large, organized multi‑agent system can rapidly clean and re‑analyze public clinical data and generate target hypotheses that align with later industry choices. This can help research teams prioritize targets with a higher historical likelihood of success, while acknowledging the remaining experimental and regulatory work needed to translate those priorities into approved treatments.
While many AI companies rely exclusively on public datasets and external academic partnerships to train computational models, running an in-house laboratory allows researchers to perform high-throughput experiments, synthesise proprietary biological data, and validate machine learning outputs in real time.
A.I. is already a powerful tool for scientific discovery. Now, some biomedical researchers are teaching it to do even more.
The discovery adds tangible evidence to the investment case for AI in life sciences, a theme that has leaned heavily on promise rather than results. Sentiment toward AI-enabled drug discovery, genomics tooling and gene-editing platforms could benefit from a concrete example of agents cutting early-stage research timeli
One biotech AI system searches 38 million biomedical papers, while another cut trial-related work from six months to less than eight weeks
AI (Artificial Intelligence) နည်းပညာက အá€á€¯á€¡á€á€»á€á€”်မှာ နေရာá€á€á€¯á€„်းမှာ ရှá€á€”ေပါပြီዠဒါပေမဲ့ “AI ကá€á€¯ ဘယ်ကနေ စလေ့လာရမလဲአအမြန်ဆုံး á€á€á€ºá€™á€¼á€±á€¬á€€á€ºá€¡á€±á€¬á€„်â
Loading more related stories...
Open the app view to save this story, compare related coverage, and continue from the same source.