Relation type-aware knowledge graph embeddings for biomedical data: a semantically adaptive framework

Abstract

Knowledge graphs (KGs) are increasingly used in biomedicine to integrate and reason over heterogeneous datasuch as genes, proteins, drugs, and diseases. However, existing knowledge graph embedding (KGE) methods typically rely on a single, fixed scoring function for all relation types, which limits their ability to model the semantic diversity of biomedical interactions. In this work, we propose a relation type-aware KGE framework that dynamically adapts scoring functions to the structural nature of relations hierarchical, symmetric, or asymmetric thereby improving the semantic fidelity of embeddings. We improve training with a constraint-aware negative sampling approach that creates realistic false examples. This forces the model to focus on learning true biomedical relationships instead of dismissing easy, nonsensical ones. We built our knowledge graph by grabbing biomedical papers from PubMed. Then, we ran them through some natural language processing tools. After that, we used the SemRep server to make subject-relation-object triples, which gave us the graph structure for studies on colorectal cancer. Our KG includes entities and relations at the molecular, clinical, and pharmacology levels. We set up the learning objective as a minimization problem with regularization. This combines margin-based ranking loss with relation-specific changes. We checked our work using MR, MRR, Hits@1, and Hits@10, which let us do a good comparison with baselines like TransE, ComplEx, RESCAL, and ConvE. The results showed our system improved MRR by 9-14% and Hits@10 by as much as 12% compared to other good systems. It was especially good at understanding things like subtype of and interacts with. In short, this research gives a way that can be scaled, is based on math and biology for filling in the blanks in biomedical knowledge graphs. By putting together relation-aware modeling, constraint-guided learning, and several ways of measuring, our method moves forward knowledge graph-based in cancer research and gives a method that can be copied for use in other bio fields.