HYPON: embedding biomedical ontology with entity sets

Constructing high-quality biomedical ontologies is one of the first steps to study new concepts, such as emerging infectious diseases. Manually curated ontologies are often noisy, especially for new knowledge that requires domain expertise. In this paper, we proposed a novel ontology embedding approach HYPON to automate this process. In contrast to conventional approaches, we propose to embed biomedical ontology in the hyperbolic space to better model the hierarchical structure. Importantly, our method is able to consider both graph structure and the varied-size set of entities, which are largely overlooked by existing methods. We demonstrated substantial improvement in comparison to thirteen comparison approaches on eleven biomedical ontologies, including two recently curated COVID-19 ontologies.