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

Microbial enzyme production and catalysis systems are crucial aspect of biotechnological research. However, building them from trustworthy published experimental data presents a major obstacle for both manual and automated techniques. Here, we introduce MEPAM (Microbial Enzyme Production and Catalytic Activity based on LLM), a question-answering system designed to accurately address inquiries related to enzyme production and catalytic reactions. Specifically, by training three machine learning models with >0.98 accuracy, we identified 11,068 high-quality, relevant articles from the Web of Science. Leveraging DeepSeek-V3 with zero-shot learning, we developed an ontology-driven knowledge representation that extracted 12,434 entities and 35,918 relations with 0.78 extraction accuracy and constructed a structured knowledge graph. Compared to few-shot learning and other machine learning methods, our framework achieved significantly higher extraction accuracy. Using this framework, we developed MEPAM based on retrieval-augmented generation and prompt engineering. Finally, using MEPAM, we extracted a comprehensive network involving the expression profiles, precise culture conditions, and substrate preferences for cellulase, demonstrating the strong utility of this tool. Compared with traditional LLMs, particularly GPT-4o, MEPAM exhibited superior performance, achieving significantly higher answer accuracy (0.86 vs. 0.52) and nearly eliminating hallucinations. MEPAM is available at http://180.76.108.212. This framework provides context-rich, verifiable insights, thus bridging predictive modeling with experimental validation to facilitate the exploration of microbial enzymatic systems.