BEAD: a Human-in-the-Loop Agentic Framework for Explainable and Prospective Drug Discovery

A critical bottleneck in AI-driven drug discovery is the inability to integrate traditional medicinal chemistry intuition, which frequently results in uninterpretable or chemically implausible molecular outputs. Here, we present BEAD (Bridging the gap between Experiment- and Artificial intelligence-based Drug design), a human-in-the-loop agentic framework for explainable drug design. BEAD places medicinal chemists in the decision-making loop to explore and optimize new scaffolds or molecules. Retrospective studies on MAT2A, PKMYT1, and IL-17A demonstrate that BEAD enables successful cross-target scaffold hopping and scaffold-constrained molecular generation via conversational interactions. Prospectively, we deployed BEAD in a real-world drug discovery targeting PRMT5. A two-stage pipeline was executed from cross-target scaffold hopping-based hit discovery to scaffold-constrained lead optimization. By synthesizing only nine molecules, we successfully identified CRP-276. Crucially, CRP-276 achieved in vivo anti-tumor efficacy equivalent to the clinical reference MRTX1719, alongside superior target selectivity and improved ADMET profiles. Ultimately, this work represents a successful implementation of a powerful human-AI collaborative system, enabling efficient, interpretable and empirically grounded drug design.