Drug repositioning with metapath guidance and adaptive negative sampling enhancement.

Abstract

Objective: Drug repositioning plays a pivotal role in expediting the drug discovery pipeline. The rapid development of computational methods has opened new avenues for predicting drug-disease associations (DDAs). Despite advancements in existing methodologies, challenges such as insufficient exploration of diverse relationships in heterogeneous biological networks and inadequate quality of negative samples have persisted.

Methods: In this study, we introduce DRMGNE, a novel drug repositioning framework that harnesses metapath-guided learning and adaptive negative enhancement for DDA prediction. DRMGNE initiates with an autoencoder to extract semantic features based on similarity matrices. Subsequently, a comprehensive set of metapaths is designed to generate subgraphs, and graph convolutional networks are utilized to extract enriched node representations reflecting topological structures. Furthermore, the adaptive negative enhancement strategy is employed to improve the quality of negative samples, ensuring balanced learning.

Results: Experimental evaluations demonstrate that DRMGNE outperforms state-of-the-art algorithms across three benchmark datasets. Additionally, case studies and molecular docking validations further underscore its potential in facilitating drug discovery and accelerating drug repurposing efforts.

Conclusion: DRMGNE is a novel framework for DDA prediction that leverages metapath-based guidance and adaptive negative enhancement. Experiments on benchmark datasets show superior performance over existing methods, underscoring its potential impact in drug discovery.