A heterogeneous graph neural network with automatic discovery of effective metapaths for drug–target interaction prediction

Abstract

Predicting potential Drug–Target Interactions (DTIs) is a critical step in drug discovery. In recent years, graph machine learning methods based on heterogeneous networks have garnered significant attention and have demonstrated advantages in predicting potential DTIs. However, existing methods require the specification of metapaths for different types of nodes to learn node features, which fails to fully and effectively mine the high-order relationships between nodes. Moreover, most methods do not generate meaningful embeddings for nodes in the network, relying solely on the network’s topological structure to learn feature representations of drugs and targets. To address these limitations, we propose a novel DTI prediction framework based on heterogeneous networks, named ADEM, which can Automatically Discover Effective Metapaths for any type of node. In addition, this framework employs similarity constraints as prior knowledge to generate meaningful embeddings for nodes in the network. ADEM outperforms classical and state-of-the-art models on three benchmark DTI datasets. Experimental analysis validates the effectiveness and rationality of the ADEM framework and its individual modules. Furthermore, the metapaths discovered by ADEM can enhance the performance of existing models in DTI prediction tasks.