ConvE-Bio: Knowledge Graph Embedding for Biomedical Relation Prediction

Abstract

Biomedical relation classification aims to automate the detection and classification of biomedical relationships, which has great advantages for various biomedical research and applications. With the development of machine learning, computational model-based approaches have been applied to biomedical relation classification and achieved state-of-the-art performance on some public datasets and shared tasks. Nevertheless, the existing models have some limitations in expressing features of large knowledge graphs. For example, the multilayer Knowledge Graph Embedding (KGE) network structure has fully connected layers and is prone to overfitting. Inspired by the multi-layer convolutional network model ConvE, this paper proposes a novel KGE model named ConvE-Bio for biomedical relation classification. The novel model performs well on the DDI (Drug-Drug Interaction), DTI (Drug-Target Interaction), and PPI (Protein-Protein Interaction) datasets, outperforming the classical baseline algorithms. Results show that ConvE-Bio can be used as a powerful tool in the field of biomedical relation classification for drug development, polypharmacy side-effect prediction and other research.