MolSAGE: Self-evolving Agent Experience for Efficient State-of-the-Art Drug Discovery

Building state-of-the-art (SOTA) predictive models for drug discovery requires expensive search over tools, architectures, and training strategies. Current LLM-based agents can find SOTA solutions through extensive trial and error, but repeating this search on every new task limits their efficiency. We propose \method (Self-evolving Agent Experience), a framework that accumulates and reuses experience across tasks to build SOTA drug discovery models efficiently. MolSAGE integrates cross-task experience into its search algorithm and agent workflow, using successful solutions, lessons from failures, and execution feedback to make model development more effective and efficient. We construct a diverse molecular-property benchmark in which \method accumulates experience from 25 non-antibacterial tasks and outperforms five competing agents overall on six held-out antibacterial activity prediction tasks. In antibiotic screening, MolSAGE identified 17 hits among 107 tested compounds, an observed hit rate approximately 82\% higher than the previous expert-driven campaign. Follow-up assays characterized C3 and C32 against methicillin-susceptible and methicillin-resistant \textit{Staphylococcus aureus} and across three human cell types. On a shared 271-compound test set, MolSAGE also improved antibacterial predictive discrimination over the published predictor. These results connect cumulative, transferable model-development experience to experimentally active antibiotic candidates and improved screening performance.