Application of large language models for building a matrix of interactions of plant-derived bioactive compounds

The number of scientific publications on plant-derived bioactive compounds is constantly growing, and manually tracking them is becoming increasingly difficult.

Problem statement. Information about plants, their bioactive compounds, and biological effects is scattered across a large number of publications, which makes manual analysis and systematization difficult.

Goal. To develop a system that automatically extracts data on plants, bioactive compounds, and biological effects from scientific publications and forms a structured interaction database for further analysis.

Results. An approach to automatic extraction of data from scientific article abstracts using a large language model is proposed. A prototype system has been developed that searches publications in PubMed, processes them with GPT-4o, and accumulates the results in a database. To obtain structured data, the model operates in structured output mode with a fixed JSON schema and zero temperature, which makes the results stable. The dialog module automatically generates PubMed queries based on the user's question. At the current stage, the database contains 247 interaction records.

Practical relevance. The developed prototype combines literature search, structured data extraction, storage, filtering, export, and dialog-based search over the accumulated knowledge base within a single application.