Lecture 08: AI Agents for Accelerating Scientific Discoveries
Link to lecture recording on YouTube
Date: 2025-11-03
Speaker: James Zou
Speaker’s Social Profile: Website / Google Scholar / GitHub / LinkedIn / X (Twitter)
Education:
- Ph.D., 2014, Harvard University
Work:
- Associate Professor, Department of Biomedical Data Science, Stanford University
- Associate Professor (by courtesy), CS and EE, Stanford University
Notes
AI has been very successful as a tool across many domains of science
- traditional way: start with well-defined scientific problem, have a specific target, then use a well-curated AI tool to tackle the problem; e.g. use tools like AlphaFold to predict structures of proteins
- start to explore AI as a co-scientist, try to tackle broader range of research endeavors (generate hypothesis, design experiments, analyze data, write papers); a lot of these shifts driven by AI agents
The Virtual Lab
A team of interdisciplinary AI scientists that work with human scientist on challenging, open-ended research
- agent creation: human user -> Principle Investigator (PI) agent -> student agent with various expertise
- group meeting: all agents meet together and come up with the research plan
- one-on-one meeting: review subtask
Example Virtual Lab team meeting:
- task: design binders to variants of virus
- PI agent spoke up first to explain the objective to other agents
- immunologist spoke up second and made a fairly unorthodox recommendation - design binders in the form of nanobodies
- machine learning agent spoke up next: agree with immunologist because smaller nanobodies make jobs easier so agents can computationally model and predict their structures more reliably
- scientific critic agent (skeptical reviewer providing more conservative feedback): not a lot of public data for nanobodies, need to be very careful about overfitting predictive deep learning models
Human researchers (Prof. Zou and students) wrote ~1.3% of the words; their job is to mostly provide high-level guidance to the AI agents
Virtual Lab School of self-improving AI scientists: frontier models behind these agents are good at broad general knowledge, but do not have the most up-to-date information or may not know how to use the latest tools in a specific technical or scientific domain
- first tell the agents the topics that they should learn about in the school
- for each of these topics, the agent would generate their own learning curriculum
- agents do web search for reference materials
- read and summarize papers
- do supervised fine-tuning to update model parameters
- teacher agents will generate quizzes to test how well the student agents learn the topic; if don’t do well, go back to school, do additional training and fine-tuning before they can graduate
[Incomplete, work in progress]
References