Using Genetic Distance to Infer the Accuracy of Genomic Prediction

Abstract

The prediction of phenotypic traits using high-density genomic data has many applications such as the selection of plants and animals of commercial interest; and it is expected to play an increasing role in medical diagnostics. Statistical models used for this task are usually tested using cross-validation, which implicitly assumes that new individuals (whose phenotypes we would like to predict) originate from the same population the genomic prediction model is trained on. In this paper we propose an approach based on clustering and resampling to investigate the effect of increasing genetic distance between training and target populations when predicting quantitative traits. This is important for plant and animal genetics, where genomic selection programs rely on the precision of predictions in future rounds of breeding. Therefore, estimating how quickly predictive accuracy decays is important in deciding which training population to use and how often the model has to be recalibrated. We find that the correlation between true and predicted values decays approximately linearly with respect to either FST or mean kinship between the training and the target populations. We illustrate this relationship using simulations and a collection of data sets from mice, wheat and human genetics.

Multiethnic polygenic risk scores improve risk prediction in diverse populations

“…In addition, this approach attains 30% relative improvement in prediction accuracy for height in an African cohort. These relative improvements are robust to overfitting, consistent with simulations and reduce the documented gap in risk prediction accuracy between European and non-European target populations (Bustamante, De La Vega, & Burchard, 2011; International Schizophrenia Consortium et al, 2009; Popejoy & Fullerton, 2016; Rosenberg et al, 2010; Scutari et al, 2016; Vilhjálmsson et al, 2015); we note that there are at least 35 phenotypes for which there are published GWAS data sets in Europeans and at least one non-European population (with minimum sample size of 8,000) that are listed in the NHGRI-EBI GWAS Catalog (MacArthur et al, 2017), where our approach could potentially be valuable (S21 Table). Intuitively, our approach leverages both large training sample sizes and training data with target-matched LD patterns.…”

“…Existing training data sets have much larger sample sizes in European populations, but the use of European training data for polygenic risk prediction in non-European populations reduces prediction accuracy, due to different patterns of linkage disequilibrium (LD) (or potentially due to different causal effects) (International Schizophrenia Consortium et al, 2009; Rosenberg et al, 2010; Scutari, Mackay, & Balding, 2016; Vilhjálmsson et al, 2015). For example, ref.…”

“…However, all of these methods are inadequate for polygenic risk prediction in non-European populations, because they consider training data from only a single population. Existing training data sets have much larger sample sizes in European populations, but the use of European training data for polygenic risk prediction in non-European populations reduces prediction accuracy, due to different patterns of linkage disequilibrium (LD; or potentially due to different causal effects) (International Schizophrenia Consortium et al, 2009;Rosenberg et al, 2010;Scutari, Mackay, & Balding, 2016;Vilhjálmsson et al, 2015). For example, Vilhjálmsson et al (2015) reported a relative decrease of 53-89% in schizophrenia risk prediction accuracy in Japanese and African-American populations compared to Europeans when applying PRS methods using European training data.…”

Theoretical and empirical quantification of the accuracy of polygenic scores in ancestry divergent populations

“…Although increasingly emphasised in the recent GWAS literature, it is worth noting that the "loss of accuracy" problem is not utterly new. Indeed, a number of studies in the animal breeding literature have previously reported lower accuracy of genomic selection across genetically distant breeds 4,5 , consistent with the observation of limited transferability of GWAS findings across diverse human populations 6,7 . These studies also highlight major factors influencing that loss such as differences between populations in causal variants effect sizes, in alleles frequencies and in linkage disequilibrium (LD) between causal variants and SNPs assayed in GWAS.…”

Human Demographic History Impacts Genetic Risk Prediction across Diverse Populations

“…For example, a previous study showed that the accuracy of breeding values and genomic prediction decays approximately linearly with increasing divergence between the discovery and target population. 8 Additionally, multiple individuals with African ancestry have received false positive misdiagnoses of hypertrophic cardiomyopathy that would have been prevented with the inclusion of even small numbers of African Americans in these studies. 9 Further, a previous study finding that 96% of GWAS participants are of European descent 1 has recently been updated; although the non-European proportion of GWAS participants has increased to nearly 20%, this is primarily driven by Asian individuals, and the proportion of individuals with African and Hispanic/Latino ancestry in GWASs has remained essentially unchanged.…”