A pseudo-label supervised graph fusion attention network for drug–target interaction prediction

Abstract

Listen

Translate article

Drug–target interaction (DTI) prediction can reveal new drug targets and assist in drug repositioning. It can also help identify the most potential candidate drugs for specific targets, advancing new drug discovery. Graph Convolutional Networks (GCNs) have been employed to explore potential relationships between drug–target pairs (DTPs) due to their strong learning capabilities. However, existing methods primarily rely on static graphs constructed from topological structures. These graphs may contain missing or meaningless edges, limiting the ability of GCNs to capture node embeddings. Furthermore, the lack of labels in practical DTI prediction presents a significant challenge. To address these issues, this paper introduces a pseudo-label supervised graph fusion attention network for DTI prediction (PSF-DTI). Specifically, we establish a far-neighbor graph to capture robust differential information between DTPs, compensating for the limitations of traditional topology graphs. Additionally, we create an adaptive graph to dynamically update edge information for more accurate graph structures. During training, we assign pseudo-labels to unlabeled data based on feature similarities between DTPs, mitigating the impact of label scarcity. Comparative experiments with seven state-of-the-art algorithms on public datasets demonstrate the superior performance of PSF-DTI. Extensive ablation experiments validate the effectiveness of the proposed approaches. Our findings suggest that PSF-DTI offers significant advantages in DTI prediction, providing innovative methods and perspectives for future drug discovery and repositioning.