In the latest episode of The Genetics Podcast, host Patrick Short speaks with Dr. James Zou, Associate Professor of Biomedical Data Science at Stanford University. James directs the Stanford AI for Science Lab, where his group builds teams of AI agents that run scientific work with limited human direction.
The conversation centers on the virtual biotech, a simulated drug discovery organization staffed by tens of thousands of AI agents. James explains what those agents learned about past trials and where their data falls short. He also covers turning research papers into agents and his clinical AI work.
The virtual biotech, published in Science, mirrors the structure of a drug developer. A chief scientist officer (CSO) agent manages divisions of specialized agents. Separate divisions handle target discovery, molecule design, safety, and clinical study design.
The team first asked the agents to analyze previous clinical trials. The system created 37,000 clinical trial agents, each assigned to one drug or trial. Each agent read the papers, company reports, press releases, and news coverage to establish what happened and why a trial failed.
The agents then proposed single cell features of drug targets, including cell type specificity and bimodality of gene expression. They matched each trial to a relevant single cell dataset and tested those features against outcomes. James summarizes the result:
"If you have drug targets that are supported by single cell evidence, they're about 48% more likely to make it to the market compared to similar drugs from similar companies looking at similar diseases that are not supported by single cell evidence."
James describes the process as relatively hands off. His team set broad directions, and the agents chose which features to build. Human experts then independently reproduced the results, a step James considers necessary to verify the key assumptions and findings.
James links the virtual biotech to OpenAI's recent work on the Navier–Stokes equations. OpenAI ran about 10,000 agents on that problem over a few days. He sees a shared pattern across both projects:
"We're getting more capabilities now from scaling up the number of agents that work together. It's orthogonal to the previous scaling paradigm, where we're mostly scaling up the size of the individual model or individual agent."
He argues drug discovery suits this approach because it combines many dependent decisions. A good target needs a good molecule, which needs an acceptable toxicity profile and a viable commercial landscape. James says this is often too much for one person to track, while thousands of specialized agents can debate and share full context.
The agents also develop their own working methods. In recent experiments, James' team let agents evolve their teamwork structures. One effective structure was "forced role reversal," where each agent argues for a colleague's proposal before the team ranks all options.
The virtual biotech is open source and draws only on publicly available information. That includes single cell data, pathology data, FDA submissions, and chemistry data. Patient-level data from failed trials is often never published.
James notes that this leaves the agents with a biased view, because public records tend to favor successful cases. His group's next step is to work with drug development companies so the agents can learn from internal data.
For sponsors, this points to a structural gap. Lessons from failed programs often stay inside the organization that ran them. Any model trained on public data inherits that skew.
Another project, Paper2Agent, published in Nature, starts with a problem James describes as knowledge being a static artifact. Papers and datasets sit unused until someone queries them. Paper2Agent converts a paper, its code, and its data into an agent that acts as a virtual author.
Each paper agent first tries to reproduce the original results. It can then answer questions and help new users apply the methods to new problems. James argues that science comes from doing, and that reproducing the work gives agents know-how a summary would miss.
The team has applied the platform to more than 100 papers. Some papers could not be converted because of incomplete methods, missing code, or code errors. James treats that outcome as a quality signal:
"The ability to actually turn a paper into an agent itself, I think is a useful certificate that the paper has enough information and it's like a bar of reproducibility in some sense."
Paper agents can also work together. In one case study, an AlphaGenome agent and an agent built from a large GWAS study identified a splicing variant associated with increased ADHD risk. That finding had not been published before.
On the clinical side, James discusses EchoNet, a deep learning model that reads cardiac ultrasound videos. The FDA has cleared it for several indications. One is cardiac amyloidosis, a condition with available treatments that often goes undiagnosed, so patients miss care they could receive. The team now works with health systems, manufacturers, and vendors on deployment.
James' group also built SleepFM, a foundation model trained on about 2 million hours of sleep lab recordings. From one night of sleep, it predicts risk for more than 100 future diseases across cardiometabolic and neurological conditions. The model transfers well to consumer wearables such as rings and watches.
Combining measurements adds predictive power. Brain waves and heart rhythms each predict many conditions, and the coupling between them predicts better than either alone. James names longitudinal blood proteomics as the largest data gap, since biomarker data remains sparse next to dense wearable signals.
Looking ahead, James expects new agent-first companies built around thousands of AI research agents, with human experts providing supervision. He also sees academia as a neutral, open counterweight to large AI labs. His group releases its tools and data as open source so others can evaluate and apply them.
For drug developers, the episode raises a practical question about evidence. If single cell support changes the odds of reaching market, target selection and the data behind it deserve earlier scrutiny.
Listen to the full episode below.