Agentic Generative Active Learning with Adaptive Acquisition Function Selection for RNA-Targeted Molecular Optimization

RNA is an attractive therapeutic target beyond the druggable proteome, but scarcity of RNA-ligand data limits the discovery of RNA-targeted small molecules. Generative Active Learning (GAL) offers a data-efficient solution, but existing approaches typically rely on a fixed acquisition strategy throughout optimization. We introduce an agentic GAL framework that dynamically adapts acquisition function through multi-agent debate. At each round, exploration and exploitation agents deliberate based on the search state, while an orchestration agent selects the acquisition strategy. We apply our framework to \textit{de-novo} optimization of small molecules targeting pre-miR-21, using LOGICS as the generator, an ECFP-based random forest as the surrogate, and DeepRNA-DTI ensemble as the computational oracle. Under the identical oracle budget, the framework discovered more active compounds and a broader range of active scaffolds than fixed acquisition strategies. These results show that agentic acquisition control enables GAL to adapt the exploration-exploitation trade-off as molecular search progresses.