CPM: Knowledge Graph Tensor Factorisation Model for Discovering Drug Targets

The discovery of novel therapeutic applications for drugs is crucial for drug development. Representation learning methods can uncover unknown relationships between drugs and targets within drug knowledge graphs. However, existing methods often suffer from information loss and insufficient representation of entities and relationships. To address these challenges, we propose a knowledge graph completion model based on CP decomposition, CPM. This model decomposes the drug-relation-target third-order tensor into a sum of rank-one tensors, achieving dimensionality reduction, reducing computational complexity, and minimizing the number of parameters. Simultaneously, it more comprehensively represents the entities and complex relationships in drug knowledge graphs. On the ADKG dataset, CPM shows an improvement of more than 1% across all four evaluation metrics (MRR, Hit@1, Hit@3, Hit@10) compared to ConvE. On the DDG dataset, CPM outperforms Tucker with a 1.4% increase in Hit@3 and a 1% increase in Hit@1. Furthermore, we use CPM to predict the head and tail entities of the triples (?, targets, EGFR) and (DB00149, targets, ?). Among the top 10 candidate entities, we successfully identify 6 and 5 entities, respectively. Source code is released in https://github.com/pole-min/CPM-Knowledge-Graph-Tensor-Factorisation-Model-for-Discovering-Drug-Targets.