USPTO-LLM: A Large Language Model-Assisted Information-enriched Chemical Reaction Dataset

USPTO-LLM: Supplemental Material

We proposed the USPTO-LLM dataset, designed to revolutionize chemical reaction prediction with high-quality, information-enriched data. Developed by Shen Yuan, Shukai Gong, and Hongteng Xu from Renmin University of China, USPTO-LLM addresses the limitations of current learning-based chemical synthesis and retrosynthesis models caused by a scarcity of high-quality datasets. This innovative dataset incorporates reaction conditions and multi-step reactions, providing 247K meticulously annotated chemical reactions. Through an automatic data annotation pipeline, the dataset utilizes Heterogeneous Directed Graphs (HDGs) for precise representation of complex chemical processes. The USPTO-LLM enables more accurate reactant prediction under varied conditions and captures intricate multi-step reactions, offering a challenging yet essential resource for cutting-edge AI models.