Robust Mendelian Randomization Framework with Multi-Omics Data for Alzheimer's Disease and Related Dementias

Project Summary / Abstract Alzheimer's disease (AD) is a progressive brain disorder that slowly destroys memory and other brain functions. Modern large-scale high-throughput multi-omics data may provide a unique opportunity to detect omics biomark- ers that are involved in the process of neurodegeneration in the preclinical and early clinical disease stages, which is highly relevant for both early diagnosis and targeted interventions. Hidden confounding bias is a major threat when using observational multi-omics data for detecting causal omics biomarkers for AD. Mendelian ran- domization is a powerful causal inference approach to removing hidden confounding bias by leveraging omics quantitative trait loci (xQTL) as instrumental variables (IV) to obtain unconfounded causal effect estimates. How- ever, current MR methods may not be well suited for analyzing high-dimensional multi-omics data. First, xQTL may have a horizontal pleiotropic effect on the outcome not mediated by the omics variables, violating the IV exclusion restriction assumption. Second, the genetic associations between xQTL and omics variables might be highly nonlinear and thus traditional linear exposure models might introduce bias and lose power; and there is a lack of MR methods to account for the heterogeneity of AD etiology for personalized intervention. Third, current MR methods may not handle high-dimensional omic exposures well. To address those pressing methodological challenges, we will develop a new robust Mendelian randomization framework with multi-omics data using xQTL as IVs to detect omics biomarkers causally associated with AD outcomes. We will construct uniformly valid and informative confidence intervals for the causal effect accounting for weak and IV selection errors in finite-sample settings. We will develop machine learning algorithms to capture nonlinear structures of the omic exposure model for more robust identification and enhanced xQTL IV strength, and to estimate heterogeneous exposure effects on the AD outcomes for personalized medicine. We will also develop a new double deconfounding framework for the joint analysis of high-dimensional omics biomarkers in a common biological pathway. We will develop software suitable for both high-performance computing clusters and cloud computing platforms following the Findability, Ac- cessibility, Interoperability, and Reusability principle. Finally, we will apply the proposed methods to multi-omics data in multi-ethnic cohorts including Washington Heights/Inwood Columbia Aging Project, the Alzheimer's Dis- ease Sequencing Project Functional Genomics xQTL project and UK Biobank. Our proposal will provide the best available analytical methods to date to resolve confounding concerns for multi-omics data, and will pave the way towards innovative research on establishing causal relationships between omics biomarkers and AD outcomes.