Integrating GWAS and Transcriptomic Data Using PrediXcan and Multimodal Deep Learning Reveals Genetic Basis and Drug...

Alzheimer's disease (AD), the leading cause of dementia, imposes a significant societal and economic burden; however, its complex molecular mechanisms remain unclear. This study integrates multi-omics data with advanced artificial intelligence (AI) methods to uncover the molecular basis underlying AD phenotype regulation and explore personalized drug repositioning strategies based on individual genetic backgrounds. First, we applied the PrediXcan method to identify candidate genes closely associated with AD cognitive diagnosis, selecting from 61 brain-related traits. We validated these findings through individual-level analysis using gene expression and genotype data from 553 dorsolateral prefrontal cortex samples in the ROSMAP database. Simultaneously, we constructed a deep, multi-layer information fusion model (AD-MIF) by integrating genotype and gene expression data and employing autoencoders as well as graph autoencoders for multi-modal feature extraction. The results revealed a 10-20% improvement in the Area Under the Curve (AUC) for predicting AD-related phenotypes. Both approaches showed high consistency across cellular structures, brain regions, and neurobiological pathways, demonstrating their complementary advantages. Gene enrichment analysis indicated that APOE and its interacting gene APOC1 play a central role in cholesterol metabolism, lipid transport, and immune regulation, while genes such as SCIMP and KAT8 are involved in immune signaling, epigenetic regulation, and neuroprotection. After incorporating attention mechanisms, AD-MIF highlighted the importance of key genes, such as POLR2C and TRAPPC4, in regulating neuronal function. Based on predictive results and enrichment analysis, we further identified candidate drugs, including sirolimus, dasatinib, and MGCD-265. In vivo experiments confirmed that MGCD-265, also known as Glesatinib, and dasatinib significantly improve cognitive deficits in the SAMP8 AD model mice by inhibiting neuroinflammation, pathological tau phosphorylation, and Aβ deposition. This study demonstrates the complementary advantages of bioinformatics pipelines and AI-based multi-modal fusion methods in elucidating the complex pathological mechanisms of AD and enhancing phenotype prediction accuracy. It also provides new theoretical support for personalized drug interventions based on individual genetic characteristics, laying a solid foundation for optimizing early screening, prediction, and personalized treatment strategies.