Bridging Generative AI and Synthesis-on-Demand Chemical Spaces for Structure-based De Novo Drug Design

Generative AI offers a powerful alternative for de novo drug design but frequently produces molecules that lack viable synthetic routes. While synthesis-on-demand (SoD) chemical spaces provide access to trillions of readily synthesizable compounds, their rapid expansion has created a computational bottleneck for effective virtual screening (VS). This study proposes a bridging strategy to navigate these challenges by identifying structurally similar analogs of AI-generated leads within SoD chemical spaces. We evaluated five generative methods (AutoGrow4, LS-MolGen, MoLeR, MORLD, and REINVENT2.0) targeting the KEAP1-NRF2 protein-protein interaction. Using the Analog Hunter tool, we searched six SoD spaces, including the 7-trillion-molecule eXplore library, to find accessible analogs of the top AI-generated leads. Overall, the workflow identified 1701 synthesizable analogs, while being substantially less resource-intensive than VS and maintaining high chemical diversity, with the REINVENT2.0-eXplore pairing benefiting the most from this bridge strategy. Furthermore, 1037 of these analogs were predicted to be more potent than the reference inhibitor iKEAP1. These results show that bridging generative AI with chemical spaces can overcome the synthetic gap in de novo design, providing a scalable and efficient framework for discovering synthetically accessible and high-potency drug leads.