关键词: 分子生成, 化学空间, 分子碎片, 深度学习, 智能化工, 物质发现

Abstract:

Molecular generation has emerged as a cost-effective and rapid approach for advancing the design and optimization of solvents for separation, reaction, catalysts, functional materials, pharmaceuticals, and other molecules. Existing molecular generation models, mainly based on deep learning frameworks, suffer from limited transparency and struggle to explore local chemical spaces effectively. In this work, we propose a function-driven molecular intelligent generation framework based on molecular fragment chemical space. Using molecular functional indicators as the direction for generation, and the “scaffolds-decorations” set of generated molecules as the basis, this framework explored the molecular fragment chemical space to facilitate targeted molecule generation. In addition, by using the chemical space deconstruction model proposed in this work, new structures from neighboring chemical spaces of excellent molecular structures are derived, thus enriching the variety of new molecules. By demonstrating the generation of drug-like molecules as an example, this framework starts from a smaller set of excellent molecules (644) and ultimately generates five times more excellent molecules (3158) of the same level, which illustrates the framework's ability to efficiently evolve a multitude of novel and high-quality molecules on the basis of diverse samples. This framework can combine functional objectives and constraints in actual chemical processes to promote new optimal designs such as green solvents at the process scale.

Key words: molecular generation, chemical space, molecular fragments, deep learning, intelligent chemical engineering, materials discovery