Drug repositioning by multi-aspect heterogeneous graph contrastive learning and positive-fusion negative sampling strategy

Abstract

Drug repositioning (DR) is a promising approach for identifying novel indications of existing drugs. Computational methods for drug repositioning have been recognised as effective ways to discover the associations between drugs and diseases. However, most computational DR methods ignore the significance of heterogeneous graph augmentation when conducting contrastive learning, which plays a critical role in improving the generalisation and robustness. The high-order similarity information from multiple data sources is still under-explored. Furthermore, only a limited number of computational DR methods can effectively screen for the most informative negative samples for model training. To address these limitations, we propose a novel DR method called DRMAHGC that employs multi-aspect graph contrastive learning to predict drug-disease associations (DDAs). First, high-order features were generated from the similarity network using a graph-masked autoencoder. Then, heterogeneous graph contrastive learning with structure- and metapath-level augmentation was employed to enhance semantic comprehension and learn expressive representations. Subsequently, the positive-fusion negative sampling strategy was exploited to synthesise informative negative sample embeddings to train the classifier for predicting novel DDAs. Extensive results on three benchmark datasets indicate that DRMAHGC significantly and consistently outperformed the state-of-the-art methods in the DR task. Moreover, the case study of two common diseases further demonstrates its effectiveness and provides novel insights into DRMAHGC in identifying novel DDAs.