Toward Fine-Tuning Large Language Models in Ontology of Microbial Phenotypes Construction

Ontologies are crucial for organizing domainspecific knowledge in biomedical fields, but their manual construction is time-consuming. This study explores the automation of ontology learning using large language models (LLMs) like BERT, RoBERTa, and DistilBERT, focusing on the Ontology of Microbial Phenotypes (OMP). We investigate three key tasks: (1) entity extraction, (2) relation extraction between entities, and (3) ontology verification. These tasks align with broader applications in biomedical annotation and named entity recognition (NER) by enabling the identification and structuring of key terms and relationships within microbial phenotypes. We evaluate LLMs in two scenarios: baseline performance using pre-trained models and fine-tuned performance after training on OMP-specific data. Our approach integrates spaCy for entity extraction, Llama 2 for relation identification, and LLMs for ontology verification. Experiments reveal that fine-tuned models significantly improve accuracy, precision, recall, and F1 scores, particularly for ontology verification. This research highlights the potential of LLMs to enhance ontology learning and support related biomedical applications like biofilm analysis, annotation, and NER, while emphasizing the value of expert curation.