Advancing plant metabolic research by using large language models to expand databases and extract labeled data

Area of Science:

• Plant science
• Metabolomics
• Bioinformatics

Background:

• Scalable data collection in plant science has advanced significantly.
• Machine learning applied to large datasets provides valuable insights in plant metabolic research.
• Large language models (LLMs) offer a new frontier for consolidating scientific literature.

Purpose of the Study:

• To evaluate the efficacy of LLMs in extracting structured data from plant science literature.
• To assess LLM performance in identifying enzyme-product pairs, compound-species associations, and transcribing tabular data.
• To explore the potential of LLMs in creating specialized, comprehensive plant science databases.

Main Methods:

• Testing various prompt engineering techniques and language models for enzyme-product pair identification.
• Applying automated prompt engineering and retrieval-augmented generation for compound-species associations.
• Developing and validating a multimodal LLM pipeline for transcribing table images into machine-readable formats.

Main Results:

• High accuracies (80-90%) achieved for enzyme-product pair identification and table image transcription when models are task-tuned.
• Modest accuracy (50%) for table image transcription.
• Reduced false-negative rates (from 55% to 40%) in compound-species pair identification compared to previous methods.

Conclusions:

• LLMs demonstrate significant potential for automating data extraction and knowledge consolidation in plant science.
• Task-specific tuning of LLMs is crucial for achieving high performance.
• Domain expertise remains vital for effective utilization of LLMs in scientific research.