Can AI Scientists Discover Better Drugs?

Can "AI Scientists" discover better drugs, or only automate steps a chemist already knows how to take? Drug design—searching a chemical space of more than 10⁶⁰ molecules—usually breaks into three stages: formulating the objective, predicting candidate properties, and optimizing molecules against it. Today each relies on humans to hand-build the underlying models, which is slow and often suboptimal. I present an AI Scientist that automates all three. On methicillin-resistant Staphylococcus aureus (MRSA), a WHO-designated urgent threat, it trained a property-prediction model that beat the Chemprop model I built by hand in 2023 on the same data, then designed an antibiotic four times more potent than the human-designed lead, with a therapeutic index of 32. Two frameworks drive this: SAGA evolves the design objective through a bi-level loop of LLM agents, and DrugSAGE builds and refines state-of-the-art predictors and optimizers by reusing cross-task experience—surpassing the human leaderboard on Therapeutic Data Commons where prior agents cannot.