The Nature of Genetic Variation for Complex Traits Revealed by GWAS and Regional Heritability Mapping Analyses

Abstract

We use computer simulations to investigate the amount of genetic variation for complex traits that can be revealed by single-SNP genome-wide association studies (GWAS) or regional heritability mapping (RHM) analyses based on full genome sequence data or SNP chips. We model a large population subject to mutation, recombination, selection, and drift, assuming a pleiotropic model of mutations sampled from a bivariate distribution of effects of mutations on a quantitative trait and fitness. The pleiotropic model investigated, in contrast to previous models, implies that common mutations of large effect are responsible for most of the genetic variation for quantitative traits, except when the trait is fitness itself. We show that GWAS applied to the full sequence increases the number of QTL detected by as much as 50% compared to the number found with SNP chips but only modestly increases the amount of additive genetic variance explained. Even with full sequence data, the total amount of additive variance explained is generally below 50%. Using RHM on the full sequence data, a slightly larger number of QTL are detected than by GWAS if the same probability threshold is assumed, but these QTL explain a slightly smaller amount of genetic variance. Our results also suggest that most of the missing heritability is due to the inability to detect variants of moderate effect (∼0.03-0.3 phenotypic SDs) segregating at substantial frequencies. Very rare variants, which are more difficult to detect by GWAS, are expected to contribute little genetic variation, so their eventual detection is less relevant for resolving the missing heritability problem.

Keywords: additive genetic variance; complex traits; fitness; missing heritability; quantitative trait variation.

Figure 1

Distribution of number of QTL segregating in the population and their contribution to the additive genetic variance V_A. For simplicity, QTL are classified as rare (gene frequency q ≤ 0.05), common (q > 0.05), small effect (a ≤ 0.02 and 0.04 pSD, for ρ between 0 and 0.75 and a ≤ 0.18 pSD for ρ = 1), and large effect (larger a). (A) Number of QTL for each class as a proportion of all QTL. (B) Proportional contribution to the total additive variance of each class of QTL.

Figure 2

Numbers and effects of significant SNPs detected in the Chip analysis (blue) or the Sequence analysis (red). (A) Average number of significant SNPs per genome for different values of the correlation between mutational effects on fitness and the quantitative trait (ρ). (B) Number of significantly associated tag SNPs for each QTL. (C and D) Distribution of frequencies of significant SNPs found in the Chip analysis (C) or the Sequence analysis (D). B–D combine results for all ρ values. For the Chip analysis, MAF > 0.05 was assumed for all ρ values.

Figure 3

Average number of QTL detected (responsible for the significance of the SNPs shown in Figure 2) per genome in the Chip analysis (blue) or the Sequence analysis (red) and percentage of total additive variance V_A explained by them for different values of the correlation between mutational effects on fitness and the trait (ρ). (A and B) Results for all QTL detected. (C and D) Results for QTL directly detected (i.e., the QTL itself was significant in the analysis). (E and F) Results for QTL detected indirectly through a significant neutral tag SNP.

Figure 4

Distribution of frequencies q of all detected QTL (combining all genomes and values of ρ) by (A) the Chip analysis and (B) the Sequence analysis.

Figure 5

Distribution of the frequency of significant SNPs for which detection is due to a given number of causal QTL for the Chip analysis (blue) and the Sequence analysis (red) (results combining all simulations and values of ρ).

Figure 6

Proportion of additive variance detected (A) and missed (B) by single-SNP GWAS (scenario ρ = 0.5) for different classes of QTL as a function of their effects (a, in pSDs) and frequencies.

Figure 7

Allelic spectrum showing the simulated estimated homozygous effects E(a) of significant SNPs plotted against their MAF. Red dots refer to significant SNPs (263) found in simulations for the scenario ρ = 0.5. Black dots refer to the 39 variants (for different disease traits) with genome-wide significance from Table S5 of UK10K Consortium (2015). Homozygous effects E(a) are obtained by multiplying the β values given in the table (in SDs) by 2 under the columns for whole-genome sequenced samples and genome-wide analysis [WGS + GWA-based (four-way)]. MAF values are obtained from the corresponding column of EAF (Effects allele frequency).