A Knowledge Graph Completion Method Applied to Literature-Based Discovery for Predicting Missing Links Targeting Cancer Drug Repurposing

Abstract

Cancer literature contains a rich body of implicit knowledge which can play an important role in drug repurposing. However, classical knowledge retrieval techniques used in Literature Based Discovery (LBD) suffer from the problem of incomplete knowledge extraction resulting in a large number of knowledge entities being missed. Recently, knowledge graphs (KGs) have been used to represent literature-derived knowledge and support knowledge discovery by representing relations between concepts. Knowledge Graph Completion (KGC) has been proposed as a method to augment knowledge represented as a KG by predicting potential missing relations between concepts in a KG. We posit that KGC methods can be applied to LBD with the goal of augmenting KGs and finding implicit knowledge by reasoning over the KG. In this paper, we present KGC methods (such as FocusE-TransE) to predict missing relations between head and tail entities, rather than the standard head or tail prediction task. Our focus is the generation of a cancer-focused drug repurposing KG, via LBD, replicating recent cancer drug repurposing discoveries. We utilized a time-slicing approach to construct incomplete KGs using semantic triples extracted from cancer literature. Next we apply our KGC methods to augment the base KG, and apply discovery patterns on the augmented KG to generate drug-gene-disease semantic paths that replicate recent cancer drug repurposing discoveries. Further, we assessed the LBD output by comparing drug-disease associations reported in the literature. Our work presents a scalable knowledge discovery framework combining KGC, LBD, and associations measures to discover meaningful implicit knowledge from the literature.