Interpretable Deep Learning and MultiOme data repurposing in Alzheimer’s target and drug discovery: NIA Artificial Intelligence and Machine Learning Consortium

Background High‐throughput DNA/RNA sequencing technologies have rapidly led to a robust body of genetic and genomic data in the Alzheimer’s Disease Sequencing Project (ADSP); however, approaching AD with a simplistic single‐target approach has been demonstrated effective for developing symptomatic therapies but ineffective when attempted for disease modification. Therapeutic approaches by specifically modulating genetic risk genes are essential for development of disease‐modifying treatments in AD. However, existing data, including genetics, genomics, transcriptomics, proteomics, and interactomics, have not yet been fully utilized and integrated to explore the roles of targeted therapeutic development for AD. Method We will introduce how our team combines tools from artificial intelligence (AI), network medicine, endophenotype models, multi‐omics (genomics, transcriptomics, and proteomics), and electronic health records (EHRs), to identify potential drug targets and repurposable drugs for AD. The fundamental premise of our method is that AD risk genes exhibit distinct functional characteristics compared to non‐risk genes and, therefore, can be distinguished by their aggregated human brain‐specific functional genomic features from various quantitative trait loci (X‐QTL), including expression QTL (eQLT), protein QTL (pQTL), splicing QTL (sQTL), methylation QTL (meQTL), and histone acetylation QTL (haQTL). Result Via applying our deep learning model to the latest AD GWAS data, we identified 156 putative AD‐risk genes that were differentially expressed in AD‐associated microglia and astrocytes from large single‐nuclei RNA‐sequencing data of human postmortem brains with AD. Combining network‐based prediction and retrospective case‐control observations with 10 million individuals, we identified that usage of four drugs (ibuprofen, gemfibrozil, cholecalciferol, and ceftriaxone) is significantly associated with reduced likelihood of AD incidence, after adjusting for various confounding factors. Importantly, gemfibrozil (an approved lipid regulator) is significantly associated with reduced risk of AD compared to simvastatin (another approved anti‐lipid medicine under Phase II AD trials), using an active user design (odds ratio 0.57, 95% CI 0.51‐0.63, P < 0.0001). Conclusion From a translational perspective, the deep learning and network‐based approaches presented here, if broadly applied would significantly catalyze effective treatment development in AD and other n