ReactionSeek: LLM-powered literature data mining and knowledge discovery in organic synthesis

Area of Science:

• Computational chemistry
• Cheminformatics
• Artificial intelligence in science

Background:

• Scientific literature contains vast amounts of chemical data, but it is largely unstructured and inaccessible to AI.
• Manual data extraction is time-consuming, limited, and requires custom software, hindering AI-driven chemical discovery.
• Existing methods fail to efficiently leverage AI for extracting complex chemical information from diverse sources.

Purpose of the Study:

• To develop an automated framework for multi-modal data mining from organic synthesis literature.
• To overcome the data curation bottleneck in AI-driven chemical discovery.
• To enable efficient extraction and standardization of chemical information for AI applications.

Main Methods:

• Developed ReactionSeek, a framework combining large language models (LLMs) and cheminformatics tools.
• Utilized sophisticated prompt engineering with minimal custom code for data extraction.
• Validated the framework on the Organic Syntheses collection, extracting textual, graphical, and semantic data.

Main Results:

• Achieved over 95% precision and recall for key reaction parameters from the Organic Syntheses collection.
• Generated a large, AI-ready dataset for chemical discovery.
• Created an interactive Synthetic Chatbot (SynChat) for natural language querying of chemical data.
• Revealed decades-long trends in catalysis through autonomous analysis.

Conclusions:

• ReactionSeek provides a general solution for automated data curation from scientific literature.
• The framework significantly advances AI-driven archive mining and knowledge discovery in chemical sciences.
• Enables efficient and scalable application of AI to chemical discovery challenges.