Abstract
Identifying the complex genetic architecture of Alzheimer's disease (AD) is critical for understanding its pathophysiology. While network-based computational methods assist in this task, they primarily model simple pairwise gene interactions and fail to capture the higher-order associations of genes that drive complex diseases. To address this limitation, we introduce HyperAD, a novel hypergraph neural network framework designed to predict AD risk genes by explicitly modeling these higher-order associations of genes. HyperAD constructs a hypergraph in which functional gene sets from databases such as MSigDB form hyperedges, and uses a two-stage hypergraph message passing neural network to extract high-order association information from the hypergraph. Comprehensive evaluations demonstrate that HyperAD significantly outperforms state-of-the-art methods. We validate the prediction results of HyperAD through multiple lines of evidence. HyperAD-predicted genes are enriched in AD-related biological processes and have significant associations with known related genes in terms of sequence similarity, protein interaction, and miRNA. In addition, their protein expression levels are significantly altered in the brains of AD patients, and they contain both known risk sites and new, high-confidence candidate genes. HyperAD provides a more accurate and biologically insightful tool for prioritizing genes and unraveling the complex genetic landscape of AD.