Strategies to connect RDF Graphs for Link Prediction using Drug-Disease Knowledge Graphs

Traditionally, drug development is a time-consuming and

costly process. Using the vast amount of available data, it is hoped that new information can be mined or inferred automatically, reducing this cost. In this work, we present steps towards completing the ReDrugS KB, which others have used to predict interactions between various drugs and

diseases. Our goal is to further complete this graph, without human intervention in the process, aiming at a high recall. For the link prediction, we used state-of-the-art embedding techniques for RDF graphs. The embeddings are fed into binary classifiers which predict the relation existence

between entities. The ReDrugS knowledge graph is the combination of the results of many studies organised in 8 million named graphs. These graphs are not entirely disjoint and might even contain contradictory assertions. Hence, a signicant challenge is making the named graphs suitable for graph embedding, by combining them into a single graph. In

this work, we report on classification performance on the merged KG.