BridgePRS: bridging the gap in polygenic risk scores between ancestries.

The key appeal of polygenic risk scores (PRS) is the provision of individual-level estimates of genetic liability to complex disease. These proxies of genetic liability enable a raft of applications across clinical and basic research settings. However, while PRS are set to play a pivotal role in the future of biomedical research, their present formulation is suboptimal for application across the multi-ancestry US population. To address this, we propose to develop high-resolution modeling to optimize the computation of PRSs across the multi-ancestry US population, which will: (i) use Bayesian hierarchical modeling to account for the population genetic and statistical causes of low PRS portability between populations, (ii) deconstruct genetic risk into shared, ancestry-focused and gene*environment sub-components, (iii) produce pathway-based PRSs that can help expose the functional causes of the portability problem and explain disease heterogeneity. The key deliverable will be the production of a powerful PRS suite of tools tailored to the multi-ancestry US population. The rationale is that failure to model important structural features that are inherent to clinical populations constitutes a vital loss of information. By modeling this high-resolution data in statistically principled and rigorous ways, researchers will be better placed to perform powerful PRS prediction in all individuals. This will offer unprecedented predictive power and insights into disease mechanisms. In Aim 1, we develop a Bayesian hierarchical PRS method, BridgePRS8, that models differences in LD, effect sizes and allele frequencies between and within ancestries. In Aim 2, we build a novel method, admixPRS, for application to genetically heterogeneous individuals that deconstructs the genome into local ancestry tracts, accounting for known demographic history, and decomposes genetic risk into 3 sub-components. In Aim 3, we develop a pathway-based PRS method for genetically heterogenous populations, PRSet+. Finally, we develop a unifying PRS method, globalPRS, that is powered to calculate PRS in any individual of the US population. Our proposal is significant because the burgeoning application of PRS means that increasing PRS portability will have immediate, high impact across the entire US population. By performing high-resolution modeling to boost PRS predictive power by mirroring the structure of human populations, and exposing gene*environment and pathway-level contributions to the PRS portability problem, our suite of PRS tools have the potential to increase the clinical utility of PRS and our understanding of how genetic risk varies across clinical populations. Our proposal is innovative because we develop the first Bayesian hierarchical PRS tools tailored to model the genetically heterogenous US population, in relation to: ancestry, genetic risk (3-component model), the genome (pathway-level PRS) and phenotype (sub-types). In summary, our proposal will deliver a suite of tools to the field to perform powerful PRS analyses across the US population and to better understand heterogeneity of disease and PRS.