Saturn-RXN: Generative molecular design with steerable and granular synthesizability control for greener chemical synthesis

Abstract: The design and validation of new molecules with desired properties is crucial for advancing human health or agriculture, yet it remains a costly and time-consuming endeavor. While generative Artificial Intelligence (AI) has accelerated the in silico exploration of novel chemical space, current models struggle to design molecules whose syntheses adhere to the 12 Principles of Green Chemistry. Conversely, heuristic-based computational tools can propose sustainable synthesis routes for known molecules but lack the capability to discover novel scaffolds. In this work, we introduce Saturn-RXN, a generative molecular design framework that explicitly enforces reaction constraints during generation. Leveraging a sample-efficient generative language model and reinforcement learning, Saturn-RXN proposes property-optimized molecules while satisfying specific synthesis constraints, such as avoiding hazardous reaction types or mandating specific building blocks. This approach directly addresses key Green Chemistry principles: Atom Economy (minimizing reaction steps), Waste Prevention/Use of Renewable Feedstocks (utilizing bio-based building blocks or waste-derived feedstocks), and Less Hazardous Synthesis (avoiding deprotections or enforcing milder conditions). We demonstrate the framework's utility in a computational case study, where Saturn-RXN successfully designs novel molecules derived from industrial waste building blocks, promoting greener processes without sacrificing molecular utility. This "sustainable-by-design" approach establishes a new paradigm in molecular discovery. By synergizing human expertise with AI, Saturn-RXN offers a pathway to transform the pharmaceutical and agrochemical industries through the creation of property-optimized molecules via sustainable synthesis.