A weighted integration method based on graph representation learning for drug repositioning

The time-consuming and expensive nature of traditional drug discovery necessitates a cost-effective approach to facilitate disease treatment. Drug repositioning, discovering innovative applications for existing drugs, has become a viable strategy that is essential for facilitating drug discovery due to its cost-effectiveness and shorter development cycle. While existing methods assume neighbors of the target node are independent, they neglect potential neighbor interaction features. We propose a weighted integration method based on graph representation learning for drug repositioning (called WIGRL) to comprehensively consider neighborhood features and neighbor interaction features, with encoders designed for similarity networks of drugs and diseases, respectively, and a network of associations between the two. Firstly, WIGRL utilizes graph convolutional network modules to obtain the neighborhood properties of nodes in similar networks. Secondly, neighbor interaction properties in similar networks are captured by graph attention network modules. Next, projection encoders are introduced to represent the association features in the association network. Finally, a more representative, unified vector is formed by simultaneously fusing information from diverse networks. After that, the decoder receives this vector to predict associations. The findings of the experiments conducted on the Fdataset, Cdataset, and LRSSL benchmark datasets demonstrate that WIGRL outperforms the existing SOTA approaches in identifying the most real positive associations and obtains the most outstanding average metrics (AUROC of 0.9331 and AUPR of 0.5654). Notably, in the case study, WIGRL discovered new associations not recorded in the dataset, validated by clinical trials and authoritative sources. Additionally, it identified novel therapeutic candidates for two neuro- degenerative diseases.