Science has always been a collaboration between human intuition and systematic search. We tell (and romanticize about) stories of how discoveries happen: a researcher, deep in the literature, notices a pattern no one else saw, makes a creative leap, tests a bold hypothesis and runs a study or experiment or comes up with a proof. This story is partly true. It is also partly a bit of an illusion. Much of what we call intuition is pattern matching across a literature we have only partially read, filtered through the biases of our training/expertise, our field’s biases and what is currently considered hot and happens to be fundable.
The bottleneck in early drug discovery (and many other problems) is rarely data. It is the cognitive investment required to traverse it. A researcher (or their graduate student/post-doc) cannot simultaneously absorb the literature on hundreds of compounds, what happens to a signaling network, and the probabilistic weight they should assign to each causal inference. This talk describes a rudimentary system I’ve 90% vibe coded that treats hypothesis generation as a graph traversal (or online discrete reinforcement learning) problem: candidate compounds, a well defined biological target, and a language model like Gemini autonomously tracing mechanistic paths through published PubMed evidence. There is some orchestration involved (which I’ll describe) which leads to assigning confidence to each causal hop, flagging pro-inflammatory liabilities, and surfacing candidates that systematic reasoning would prioritize. Human attention or intuition might eventually reach this (given sufficient time and resources). Aging and biology of aging is an ideal testbed. The target space is broad, the literature is dense and cross-disciplinary, and the stakes of missing a good idea are moderately high. I will describe results, as they stand, on the day of this seminar talk.
The real subject of this talk is a question the results force us to confront: where, precisely, do we add value? The bottleneck was never reading speed or memory. It was knowing which question was worth asking in the first place. If a system can read faster, hold more context, and reason without fatigue and coffee/tea, what remains distinctly human in the scientific process? I’ll try if I can offer some tentative answers.
T32 faculty mentors are strongly encouraged to attend in person.
Please direct any questions/requests to Eric Schafer: eeschafer@medicine.wisc.edu