Constructing Biomedical Knowledge Graph Based on SemMedDB and Linked Open Data

Biomedical knowledge graphs (BMKGs), which may facilitate precision medicine and clinical decision support, have become more and more important in healthcare practice and research. A lot of challenges still remain in their construction and curation due to the complex and high knowledge demanding nature of the task. Most of the current BMKGs are manually compiled, which is particularly time-consuming and labor-intensive. Some are automatically generated but rely heavily on the quality of the source data. Furthermore, most of them may not fully integrate or represent the most recent biomedical advancement. To tackle these problems, we propose a novel approach to building a BMKG leveraging the SemMedDB and Health Science Linked Open Data (LOD). Carefully checking the inconsistent predications in the SemMedDB, we detected 462,188 conflicting pairs of semantic triples. What's more, further mining of semantic relationships among different datasets, we found over 30 new relationships linking disorders, genes and drugs. Our methods explore a new way to improve the quality of SemMedDB and facilitate BMKGs-based knowledge discovery.

I. Introduction

A knowledge graph (KG), which describes knowledge in the form of“Subject-Predicate-Object” triple [1], is one of the most popular approaches for representing knowledge on the current Web [2]. In the health sciences, biomedical knowledge graphs (BMKGs) link biomedical entities (such as diseases, proteins, and drugs) through certain defined relationships. They are important tools to computationally analyze the comprehensive body of biomedical knowledge. BMKGs have been widely applied in multiple tasks, such as extracting disease biomarkers [3], identifying disease mechanisms [4], and predicting health benefits (i.e., efficacy) of a drug over a placebo [5], which could further facilitate precision medicine and clinical decision support.