Efficient Augmentation of Biomedical Knowledge Graphs through Data Integration and Edge Inference

Knowledge graphs have emerged as powerful frameworks for representing and reasoning over biomedical data, enabling applications from drug repurposing to disease mechanism discovery. However, biomedical knowledge graphs face challenges, including data fragmentation, incomplete coverage of biologically meaningful relationships, and opacity of inference methods. This dissertation investigates augmentation strategies for biomedical knowledge graphs by integrating external data sources and developing an explainable edge-inference model. First, we describe a systematic evaluation framework to quantify the impact of integrating structured pathway databases into knowledge graphs, demonstrated through incorporating Reactome pathway data into ROBOKOP. Reactome contributed 9,332 new semantic metapath patterns and produced architecture-dependent performance improvements: complex embedding models (ComplEx, RotatE) showed gains of 6–7%, while translational models showed minimal or negative effects. Critically, 98.8% of improvements were mechanistically explainable through identifiable Reactome pathways. Second, we present RELATE (Relation Extraction with LLMs and Ontology Constraints), a framework for extracting structured knowledge from unstructured biomedical literature. By embedding Biolink Model semantics into the relation extraction pipeline through LLM-based named entity recognition followed by ontology mapping, RELATE achieved 52% exact match accuracy and 89.7% semantic similarity on ChemProt benchmark data, producing 1,964 validated triples from 947 articles across three disease domains. Third, we introduce EDGAR (Enrichment-Driven GrAph Reasoner), an edge-inference framework combining statistical enrichment analysis with resistance-based scoring to predict missing relationships. EDGAR identifies known answers through one-hop lookup, discovers statistically significant commonalities via hypergeometric enrichment, and infers new candidates sharing those enriched features. Validation on Alzheimer’s disease drug repurposing confirmed biological validity, with all identified gene targets corroborated in the AGORA AD Knowledge Portal. Unlike black-box embedding methods, EDGAR provides fully interpretable predictions traceable to specific enrichment rules. Together, these contributions advance the construction of more comprehensive, semantically coherent, and scientifically actionable biomedical knowledge graphs, supporting researchers in generating testable hypotheses and accelerating translational discovery.