Augmented Knowledge Graphs for Literature-Based Discovery (AKG-LBD): A Novel Framework to Enhance Semantics-Based LBD For Biomedical Knowledge Discovery

Abstract

The biomedical literature is expanding exponentially, generating a vast amount of knowledge that frequently goes unnoticed. Consequently, there is an urgent need to develop methods to mine knowledge from published literature to facilitate the automated discovery of hidden biomedical knowledge. Literature-Based Discovery (LBD) is a novel paradigm that aims to uncover new knowledge from the literature via transitive inference. Advances in text mining and knowledge extraction methods have enabled semantics-based LBD, which extracts knowledge in the form of subject-predicate-object semantic triples represented in a Knowledge Graph (KG). The subject and object are normalized biomedical concepts, and the predicate denotes the semantic relation between them. Semantics-based LBD has not seen large scale adoption due to several challenges. Firstly, knowledge extraction methods result in incomplete knowledge extraction due to missing semantic relations. Secondly, extracted biomedical entities are represented by granular and ambiguous representations, leading to a large discovery search space. Thirdly, the over-generation of spurious discoveries as output obscures meaningful discoveries. This dissertation investigates semantics-based methods and KG representation learning to develop novel solutions addressing the fundamental challenges in semantic-based LBD. Specifically, we address the challenges by: (i) incorporating state-of-the-art knowledge extraction to acquire semantic-based knowledge from the literature; (ii) utilizing concept disambiguation and semantic alignment techniques to resolve ambiguity and granularity of concept representations; (iii) leveraging a multi-step Knowledge Graph Completion (KGC) methodology to augment the literature-based KG by predicting missing relations using KG embeddings; and (iv) presenting a knowledge filtering and ranking approach based on the principles of information theory to prioritize interesting discoveries. The outcome of this dissertation is the novel Augmented Knowledge Graphs for LBD (AKG-LBD) framework that enhances traditional semantics-based LBD frameworks. The AKG-LBD framework is assessed by replicating biomedical discoveries published in peer-reviewed journals. The results indicate that AKG-LBD can discover meaningful knowledge with high precision relative to baseline approaches. The main implication of this dissertation is that KGC methods, combined with semantic alignment, enhances the performance of semantics-based LBD by generating augmented literature-based KGs. Additionally, the knowledge filtering and ranking methods are capable of prioritizing interesting knowledge which facilitates the exploration of meaningful biomedical discoveries.