Prompt Engineering Approaches for Working with Biomedical Knowledge Graphs through LLMs

Biomedical knowledge graphs (BioKGs) are widely applied in the biomedical field to represent biological entities and their relationships. Through their simple data model, they facilitate the integration of heterogeneous information and the development of downstream ML applications (e.g., node classification, link predictions, knowledge extraction). However, their construction by integrating structured data (e.g., relational data, json, csv) or extracting facts from scientific literature (e.g., PubMed articles, digital health records) requires many efforts, and intelligent tools supporting the user in this activity are still lacking. In this paper, we describe the activities that our group is carrying out at the University of Milan to support the construction of BioKGs by applying prompt engineering approaches in combination with general-purpose Large Language Models (LLMs) with the aim of reducing the generation of incorrect facts due to hallucinations that are not acceptable in sensitive areas like precision medicine.