The repeated experimental and computational breakthroughs in the two decades since the sequencing of the human genome have provided an unprecedented opportunity to understand the etiology of human diseases. The diminishing cost of genomics data means it is now possible for researchers to obtain complete genome sequence information on hundreds of thousands of individuals, with widespread access to those data via large repositories of electronic health records (EHRs) and biobanks. However, interrelated computational, statistical and privacy questions remain for about how to leverage these data to study the contribution of genetic variation to common diseases. Importantly, there is a need for methodological innovation to minimize computational complexity and respect data privacy concerns while maximizing data access and utility. At the forefront of these innovations are computational methods that leverage non-individually identifiable summary statistics pre- computed on biobank/EHR data to maximize downstream functional understanding and clinical utility. One emerging set of summary statistics are point estimates and standard errors from separate regression models of a phenotype (Y) on individual genotypes (Xi), sometimes with limited covariate adjustment (e.g., Age, Sex and principal components (PCs)). A key limitation of any set of pre-computed summary statistics is not being able to anticipate all possible downstream uses of such statistics. For example, researchers may want to use: (a) different sets of covariates than those considered in pre-computed analyses, (b) sets of genetic variants as predictors, instead of single markers and (c) alternative phenotype definitions that are functions of existing variables (e.g., a researcher want to know about a phenotype, 𝑌𝑌𝐶𝐶, of clinical importance, but only has pre- computed summary statistics on 𝑌𝑌1, 𝑌𝑌2, … , 𝑌𝑌𝑘𝑘, where 𝑌𝑌𝐶𝐶 = 𝑓𝑓( 𝑌𝑌1, 𝑌𝑌2, … , 𝑌𝑌𝑘𝑘)). In this project we will (1) develop a computationally efficient framework to evaluate genetic variants with clinically relevant phenotypes using summary statistics and apply these methods to perform harmonized analyses of clinically relevant phenotypes in multi-cohort studies using summary statistics and (2) validate the feasibility of these methodological innovations within two related consortia currently exploring the role genetic variants on cognitive outcomes and the potential moderating and/or mediating role of dietary polyunsaturated fatty acid (PUFA) levels. Preliminary methods will be implemented in open-source tools (R/python packages), and will also involve extensive testing on both simulated and real data across a wide range of clinically relevant phenotypes. Work will set the stage for a future R01 project to provide additional methodological expansion, more widespread testing and comprehensive dissemination.

Researchers now have access to genome sequence information on hundreds of thousands of individuals, with widespread access to those data via large repositories of electronic health records (EHRs) and biobanks. Sharing summary statistics from these databases does not compromise patient confidentiality, however, the current state of the art is limited by how these summary statistics can be used. The proposed research will provide strategies to better utilize summary statistics on genomics data to accelerate our understanding of human disease.