The value of semantics in biomedical knowledge graphs

Knowledge graphs use a graph-based data model to represent knowledge of the real world. They consist of nodes, which represent entities of interest such as diseases or proteins, and edges, which represent potentially different relations between these entities. Semantic properties can be attached to these nodes and edges, indicating the classes of entities they represent (e.g. gene, disease), the predicates that indicate the types of relationships between the nodes (e.g. stimulates, treats), and provenance that provides references to the sources of these relationships.

Modelling knowledge as a graph emphasizes the interrelationships between the entities, making knowledge graphs a useful tool for performing computational analyses for domains in which complex interactions and sequences of events exist, such as biomedicine. Semantic properties provide additional information and are assumed to benefit such computational analyses but the added value of these properties has not yet been extensively investigated.

This thesis therefore develops and compares computational methods that use these properties, and applies them to biomedical tasks. These are: biomarker identification, drug repurposing, drug efficacy screening, identifying disease trajectories, and identifying genes targeted by disease-associated SNPs located on the non-coding part of the genome.
In general, we find that methods which use concept classes, predicates, or provenance improves achieve a superior performance over methods that do not use them. We thereby demonstrate the added value of these semantic properties for computational analyses performed on biomedical knowledge graphs.