Wednesday, December 15 2021
11:00am - 1:00am
PhD Thesis Presentation
Thesis Proposal presentation for Megan Duff

Abstract:
A central problem in personalized medicine is understanding how variation in oneís genome leads to disease risk. To address this problem, Meuwissen et al. (2001) introduced genomic prediction models to predict an individualís phenotype using all of their genetic data. Genomic prediction models outperformed existing methods, which built predictive models using only known causal single nucleotide polymorphisms. Due to their success, there has been an expansion of methods used for genomic prediction; however, most methods require very large sample sizes to create meaningful models. These datasets might not be available to single investigators, especially ones studying diverse (non-European) populations. In these situations, transfer learning has the capability to increase prediction performance. Transfer learning is a popular machine learning methodology, which aims to increase prediction accuracy in a target population by transferring knowledge from a model trained on a source dataset for the same or a similar phenotype. It has been shown to dramatically increase prediction accuracy for computer image recognition and text classification. Therefore, the central hypothesis of this work is that transfer learning will generalize existing models for European descent populations to improve phenotype prediction accuracy for diverse (non-European) populations. The proposed work to address this hypothesis is split into two aims. The first aim is to compare current genomic prediction methods, e.g., GBLUP and Bayes B, to transfer learning methods where both the target and source datasets are individuals of European descent. This aim will help inform optimal model building and the potential predictive ability of transfer learning for aim two, which addresses transfer learning across different population ancestries. The second aim is to develop a methodology for utilizing transfer learning in genomic prediction, when the source set is of European descent and the target set is non-European. This proposed work has implications for personalized medicine, through helping communities who are underserved by current genetics research. This work is addressing the central problem of identifying the relationship between genetic variation and disease for non-European populations through using research that has already been conducted on European descent populations.
Speaker:Megan Duff
Affiliation:
Location:Zoom - see email


Download as iCalendar

Done