Part of the TPC Seminar Series

---

Abstract:

How can we measure whether AI agents learn from experiments and develop an understanding of the systems they investigate? Science sandboxes offer controlled environments in which agents can design experiments, receive feedback, and revise hypotheses. Drawing on regulatory genomics and protein fitness prediction, this talk examines how agents can improve predictive performance while failing to discover the underlying rules, particularly when those rules depart from familiar biological assumptions. These failures provide a basis for evaluating scientific reasoning and identifying where current systems need improvement. The talk will also introduce BroadBox, an effort to build science sandboxes across biology.

Biography:

Arya Rao is a Harvard-MIT MD-PhD candidate whose research uses AI to investigate human biology and develop approaches to precision medicine. Conducting her doctoral research at the Broad Institute with Pardis Sabeti and Sangeeta Bhatia, she develops methods to interpret human genetic variation and design regulatory DNA for targeted control of gene expression. Her work on science sandboxes examines how AI agents conduct experiments and acquire scientific understanding. She is a Paul & Daisy Soros Fellow and Forbes 30 Under 30 honoree, and has received the Massachusetts Medical Society Information Technology Award, the Harvard Medical School Dean’s Innovation Award, the Harvard Presidential Public Service Fellowship, and the Ladders to Cures Accelerator Award. She graduated from Columbia University with a BA in biochemistry and computer science.