Development and application of machine learning-based approaches for estimating disease risk in diverse and admixed populations.

ABSTRACT Polygenic Risk Scores (PRS) quantify the genetic component of an individual’s risk of eventually developing a particular phenotype (generally a complex disease). They, along with more traditional clinical risk prediction metrics, are a crucial component of future personalized medicine protocols, by influencing the optimal timing of monitoring, preventative testing and

potential lifestyle modifications. While many PRS models have been proposed, they suffer from two main drawbacks. Specifically, PRS models for the same disease have not been systematically tested and evaluated for effectiveness using uniform, repeatable protocols; and existing models are generally optimized for risk prediction in European ancestry individuals, and

are much less accurate when applied to people from other populations. This proposal will address these shortcomings by 1) Evaluating the performance of existing PRS for European ancestry individuals in a fair and consistent manner using both public (UK Biobank) and proprietary data 2) Generating an improved ancestry reference panel that includes deeper representation of

indigenous American groups 3) Developing novel machine learning PRS models applicable to individuals with diverse and admixed ancestry 4) Implementing a combined ancestry estimation + PRS application that is commercially available through AWS or DNAnexus Together, this proposed work will ensure that the benefits of improved disease risk prediction

are available to all individuals, regardless of ancestral background.