HiGLDP: a hierarchical graph neural network for predicting lncRNA-disease associations through multi-omic integration

Background: Long noncoding RNAs (lncRNAs) have emerged as crucial regulators in the pathogenesis of complex human diseases. Despite significant advances, identifying disease-associated lncRNAs remains challenging due to the vast noncoding transcriptome and the complexity of lncRNA interaction networks. Results: We propose HiGLDP, a computational framework for predicting lncRNA-disease associations through the integration of multi-omic data and advanced graph neural network techniques. HiGLDP constructs comprehensive similarity networks for lncRNAs and diseases using genomic, transcriptomic, and proteomic information, which are refined using random walk with restart (RWR) and denoising autoencoders (DAE). The bipartite lncRNA-disease association network is transformed into an interconnected graph with relationship nodes, while an association feature graph is constructed based on cosine similarity. A hybrid graph neural network architecture combining graph convolutional networks (GCN) and graph attention networks (GAT) is employed to capture both local and global graph structures, followed by a multilayer perceptron (MLP) for association classification. Comprehensive evaluations demonstrate that HiGLDP achieves superior predictive performance, accuracy, and robustness compared with existing methods. Case studies further validate its effectiveness in identifying novel lncRNA-disease associations. Conclusions: HiGLDP provides a robust and interpretable computational framework for lncRNA-disease association prediction. By integrating multi-omic information with hybrid graph learning, it offers valuable insights into lncRNA-disease interactions and represents a meaningful advancement in predictive modeling in this field.