Leveraging Multi-Scale Genomic Information to Enhance Prediction and Understanding of Complex Traits

Understanding the mechanisms underlying complex diseases is essential for developing effective treatments. Advances in sequencing technologies and declining costs have led to an exponential growth in genomic data. To effectively analyze these large datasets and identify the genetic basis of diseases, it is crucial to develop new statistical tools that combine information from multiple data sources. Integrating information from within a single omics layer, such as leveraging genetically related traits or combining total and allele-specific gene expression, can enhance prediction and increase power in biological research. Multi-trait polygenic risk score (PRS) methods improve disease risk prediction by utilizing the shared genetic architecture among correlated traits. However, these methods do not account for vertical pleiotropy, where one trait acts as a mediator for another. For my first project, I present endoPRS, a weighted lasso framework that incorporates information from relevant endophenotypes, measurable biological traits with genetic links to disease progression. Simulation analyses demonstrate the robustness of endoPRS across complex genetic frameworks. By leveraging a genome-wide association study (GWAS) of eosinophil count, endoPRS significantly improves the prediction of childhood-onset asthma. Quantitative trait loci (QTL) mapping identifies genomic regions associated with quantitative traits such as gene expression. Approaches that simultaneously leverage total gene expression and allele-specific gene expression can increase QTL detection power but are limited to bi-allelic variants. They cannot capture haplotype effects, which are essential to analyzing the genetics of multiparent populations. For my second project, I extend two QTL mapping methods, TReCASE and mixQTL, to model founder haplotype effects. These methods substantially increase the power to detect QTL compared to conventional methods in both simulated and real gene expression data from Collaborative Cross Recombinant Intercross (CC-RIX) mice. As access to individual-level genotype-phenotype data is often restricted, my third project extends endoPRS to endoPRS-SS, which estimates PRS coefficients using GWAS summary statistics and LD reference panels. I also extend the endoPRS-SS framework to allow for multiple endophenotypes. Simulations and real data analysis show that endoPRS-SS performs comparably to endoPRS. Moreover, by integrating both monocyte count and executive function scores, endoPRS-SS improves Alzheimer’s disease risk prediction compared to using either endophenotype alone.