Midway evaluation in Statistics - Magnús Ingvi Magnússon

Title of thesis: From GWAS Algorithms to Clinical Interpretation: Statistical Methods for Population-Scale Genomic and Proteomic Data

Student: Magnús Ingvi Magnússon

Doctoral committee:
Dr. Sigrún Helga Lund, Professor at the Faculty of Physical Sciences, University of Iceland
Dr. Gunnar Stefánsson, Professor Emeritus at the Faculty of Physical Sciences, University of Iceland
Dr. Daníel Fannar Guðbjartsson, guest professor at the University of Iceland and Chief Scientist of Science at deCODE genetics.
Dr. Þórunn Rafnar former head of cancer at deCODE genetics.

AbstractThe thesis focuses on statistical methods for large-scale biomedical data with applications in genomics, specifically Genome-wide association studies (GWAS), cancer and proteomics. The scope is broad and encompasses a wide range of topics. We develop an algorithm for case-control GWAS, specifically tailored for optimal performance on Graphical processing units (GPUs). Utilizing the massive parallelization available when working with GPUs, we observe a two orders of magnitude speedups when compared to previously implemented methods for high performance clusters. We show that the method yields comparable association statistics to the aforementioned method, thus conserving computational robustness while achieving a considerable speedup. We then focus on statistical modeling with clinical and molecular oncology. Using multivariable survival models the association between disease-specific survival and features of the tumor microenvironment, genetic markers and plasma protein levels are assessed. We show that histopathological features—particularly stromal architecture and immune infiltration—are strong and independent prognostic factors, often outperforming mutational markers. This work illustrates how statistical modeling can clarify biological mechanisms and improve risk stratification in clinical practice. Finally, we develop an algorithm to aid in defining protein quantitative trait locis (pQTLs). Protein quantitative trait loci are regions in the genome that are associated with the magnitude of protein expression, i.e. the levels of protein expressions. Using results from GWAS performed on protein levels, we develop a method to cluster these signals into regions of interest.