Enhancing Drug-Drug Interaction Predictions in Biomedical Knowledge Graphs Through Integration of Householder Projections and Capsule Network Techniques

Investigating Drug-Drug Interactions (DDIs) is crucial for optimizing drug treatments, improving therapeutic outcomes, and reducing adverse reactions, aiding in safer drug therapy strategies. Knowledge Graph Embedding (KGE) models, vital in DDI research, map drugs and their interactions in the Biomedical Knowledge Graph (BioKG), revealing potential and unreported drug interactions. However, due to diverse drug relationship complexities and significant Relationship Mapping Properties (RMPs), existing models are limited in capturing complete relationship patterns, struggling with prevalent RMPs. This study introduces HPCap, combining Householder projections with capsule networks to enhance modeling of complex drug interactions in BioKG. Householder projections first address complex RMPs, followed by embedding entities and their relationships into a projection space for dynamic interaction modeling. A capsule network with reverse dot-product dynamic routing captures in-depth information, enhancing entity representation. Experimental results show HPCap's superior performance over contemporary methods on three BioKG datasets.