Multimodal graph representation learning for drug repositioning

Abstract

In this study, we introduce a novel multimodal graph representation learning framework (MGRL-DR). By integrating bilinear graph convolution, graph attention networks, and autoencoders, the representational capability of MGRL-DR is significantly enhanced. In data preprocessing, the integration of heterogeneous networks and the introduction of a molecular feature enhancement module strengthen the feature representation of drugs and diseases. Ten-fold crossvalidation on four datasets (Fdataset, Cdataset, LRSSL, and Ldataset) demonstrates that MGRL-DR outperforms comparative methods across AUROC, AUPR, and other performance metrics. In the case study, MGRL-DR discovered new associations not recorded in the dataset, validated by clinical trials and authoritative data sources. This method provides an efficient new avenue for drug discovery.