Decoding enzymatic landscapes: a knowledge graph-enhanced large language model framework for microbial enzyme...

Area of Science:

• Biotechnology
• Bioinformatics
• Enzyme Engineering

Background:

• Extracting reliable experimental data for microbial enzyme production and catalysis is challenging.
• Existing manual and automated methods face significant obstacles in data acquisition.

Purpose of the Study:

• To develop MEPAM, a question-answering system for microbial enzyme production and catalytic activity.
• To create a structured knowledge graph from scientific literature for enhanced data retrieval.

Main Methods:

• Trained machine learning models (>0.98 accuracy) to identify relevant scientific articles.
• Utilized DeepSeek-V3 with zero-shot learning for ontology-driven knowledge representation.
• Constructed a knowledge graph by extracting 12,434 entities and 35,918 relations (0.78 accuracy).
• Developed MEPAM using retrieval-augmented generation and prompt engineering.

Main Results:

• Identified 11,068 high-quality articles from Web of Science.
• Achieved higher extraction accuracy compared to few-shot and other machine learning methods.
• Demonstrated MEPAM's utility by extracting a comprehensive cellulase network.
• MEPAM showed superior performance over GPT-4o, with higher accuracy (0.86 vs. 0.52) and reduced hallucinations.

Conclusions:

• MEPAM provides context-rich, verifiable insights into microbial enzymatic systems.
• The framework bridges predictive modeling with experimental validation.
• Facilitates exploration and understanding of microbial enzymatic systems for biotechnological applications.