Mining patents with large language models elucidates the chemical function landscape

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Cheminformatics

Background:

• Small molecule discovery traditionally relies on structure-based methods.
• Leveraging the vast chemical literature offers a complementary approach.
• Chemical function is intricately linked to molecular structure and interactions.

Purpose of the Study:

• To investigate the feasibility of using text-derived data for chemical function prediction.
• To develop and evaluate a large-scale dataset of molecule-function relationships from literature.
• To demonstrate the utility of text-based functional landscapes in drug discovery.

Main Methods:

• Construction of a Chemical Function (CheF) dataset using LLM and embedding techniques on patent data.
• Extraction of 1.5K unique functional labels for ~100K molecules from 188K patents.
• Analysis of semantic coherence and congruence with chemical structure relationships.

Main Results:

• The CheF dataset provides a semantically coherent representation of chemical functionality.
• The dataset approximates the actual chemical function landscape.
• A model successfully predicted functional profiles from structure alone, identifying drugs with target functionality.

Conclusions:

• Text-derived functional landscapes can effectively guide molecular discovery.
• This approach offers a viable alternative to traditional structure-based methods.
• Functional label-guided discovery facilitates the design of novel functional molecules.