Dimensionality of genomic information and its impact on GWA and variant selection: a simulation study

Abstract

Background: Identifying true-positive variants in genome-wide associations (GWA) depends on several factors, including the number of genotyped individuals. The limited dimensionality of the genomic information may give insights into the optimal number of individuals to use in GWA. This study investigated different discovery set sizes in GWA based on the number of largest eigenvalues explaining a certain proportion of variance in the genomic relationship matrix (G). An additional investigation included the change in accuracy by adding variants, selected based on different set sizes, to the regular SNP chips used for genomic prediction. Methods: Sequence data were simulated containing 500k SNP with 200 or 2000 quantitative trait nucleotides (QTN). A regular 50k panel included one every ten simulated SNP. Effective population size (Ne) was 20 and 200. The GWA was performed with the number of genotyped animals equivalent to the number of largest eigenvalues of G (EIG) explaining 50, 60, 70, 80, 90, 95, 98, and 99% of the variance. In addition, the largest discovery set consisted of 30k genotyped animals. Limited or extensive phenotypic information was mimicked by changing the trait heritability. Significant and high effect size SNP were added to the 50k panel and used for single-step GBLUP with and without weights. Results: Using the number of genotyped animals corresponding to at least EIG98 enabled the identification of QTN with the largest effect sizes when Ne was large. Smaller populations required more than EIG98. Furthermore, using genotyped animals with higher reliability (i.e., higher trait heritability) helped better identify the most informative QTN. The greatest prediction accuracy was obtained when the significant or the high effect SNP representing twice the number of simulated QTN were added to the 50k panel. Weighting SNP differently did not increase prediction accuracy, mainly because of the size of the genotyped population. Conclusions: Accurately identifying causative variants from sequence data depends on the effective population size and, therefore, the dimensionality of genomic information. This dimensionality can help identify the suitable sample size for GWA and could be considered for variant selection. Even when variants are accurately identified, their inclusion in prediction models has limited implications.

Using large-scale whole-genome sequence data for single-step genomic predictions in maternal and terminal pig lines

“…ChipPlusSign also showed greater robustness than Top40k when the genomic prediction was performed using BayesR [8]. Several studies have been conducted to investigate genomic prediction by adding selected variants to regular chip data through either real or simulated datasets [2, 3, 16, 36]. In US Holstein, VanRaden et al [3] investigated the reliability of GEBV for 33 traits when preselected SNP (N = 16k) from WGS were added to a 60k SNP chip.…”

“…Our results agree with the ones from Fragomeni et al [2], especially with ChipPlusSign. In a simulated study, Jang et al [16] investigated the dimensionality of genomic information for variant selection and genomic prediction with sequence data. Their results showed that populations with small Ne obtained a maximum accuracy gain of 0.86% to 1.98% when either significant variants or hundreds of variants with high effect sizes preselected from GWAS were added to a 50k SNP chip.…”

“…Those marginal gains for most traits with ChipPlusSign and Top40k raised a question about the amount of information that has been used for preselecting the variants and performing genomic predictions. Examining the dimensionality of the genomic information can help assess the efficient number of genotyped animals needed to maximize the percentage of discoveries in GWAS and prediction accuracy gains [16]. According to Jang et al [16], in populations with a larger effective size ( Ne = 200), using the number of genotyped animals equal to the number of largest eigenvalues explaining 98% of the variance of G sufficed to capture the most informative variants, although only a tiny proportion of the causative variants was discovered for highly polygenic traits.…”

“…In pigs, the Ne varies from 30 to 50, and the dimensionality of the genomic information or the number of independent chromosome segments segregating in the population ranges from 4k to 6k [18]. Based on Jang et al [16], using a sample size for GWAS of 7k in a population with Ne of 20 allowed identifying causative variants explaining 20% of the additive genetic variance. Moreover, larger sample sizes resulted in larger prediction accuracies with selected variants.…”

“…Using simulated sequence data, Jang et al [16] looked at the dimensionality of the genomic information [17] to assess the number of genotyped animals needed to maximize the percentage of discoveries in GWAS. The authors showed that populations with smaller effective size ( Ne = 20) require more genotyped animals to capture causative variants, whereas for large populations ( Ne = 200), using the number of genotyped animals equal to that of the largest eigenvalues explaining 98% of the variance of the genomic relationship matrix suffices.…”