Search

Recent years have seen an increasing amount of genomic data, biobank data and electronic health records (EHR). Combination of these data into genetic and phenomic studies provides a broader view of population features as well as disease characteristics, and holds great potential in supporting more powerful inference and prediction to identify and quantify the health risks. In this dissertation, we propose several statistical methods for genome by phenome analysis using large biobank data. In Chapter 2, we propose a new method for genetic association test with family disease history, TAPE (mixed-model-based Test with Adjusted Phenotype and Empirical saddlepoint approximation), which controls for increased phenotype correlation and accounts for case-control imbalance. Family disease history contains important latent information and can be incorporated into GWAS to increase power, especially when diseases have low prevalence. However, in the presence of related samples, existing methods fail to address increased phenotypic concordance among closely related samples due to similar family history, leading to reduced testing power with limited unrelated samples. In addition, existing approaches cannot adjust for the unbalanced phenotypic distribution, causing type I error inflation when rare variants are tested. We show through simulation studies and analysis of UK-Biobank data of white British samples and KoGES data of Korean samples that the proposed method is computationally efficient and gains greater power for detection of variant-phenotype associations than common GWAS with binary traits while yielding better calibration compared to existing methods. In Chapter 3, we present PRSbils, a polygenic risk prediction method with bilevel continuous shrinkage prior to utilize the functional annotation information. Polygenic risk scores (PRS) have been widely used in predicting human complex traits, yet an ineligible gap exists in leveraging the functional roles of genetic variants and capturing non-infinitesimal genetic architectures simultaneously. The bilevel continuous shrinkage prior in PRSbils accommodates the varying genetic architectures both on the variant-specific level and on the functional annotation level. GWAS summary statistics is used as input, and local LD patterns are accounted for through an external LD reference panel. We conducted simulation studies and applied PRSbils to binary and quantitative traits in three real world data sources, and demonstrated that the proposed method holds the potential for improving PRS predictive performance by incorporating functional annotations. In Chapter 4, we propose IMCD, an inductive matrix completion based disease risk prediction method, which incorporates multi-phenotypic information such as disease-disease similarity in addition to individual similarity. We observe two major challenges in utilizing information from other phenotypes for risk prediction of the target phenotype. The first challenge considers the situation where information for the target phenotype is available for a set of training samples. Existing cross-phenotype risk prediction methods are often trained using solely individual-level information such as PRS from other diseases, and cannot incorporate the abundance of disease or individual similarity information that holds the potential to improve prediction. The second challenge considers the situation where target phenotype information is not available for both the training and the test samples. Currently there is limited literature for predictive methods under this situation. The proposed IMCD method can be used under both the above-mentioned situations. We show through simulation studies and real-world data analysis that the IMCD approach can improve predictive performance especially when there was a higher similarity between the neighboring diseases and the target disease.