Application of a Bayesian dominance model improves power in quantitative trait genome-wide association analysis

Abstract

Background: Multi-marker methods, which fit all markers simultaneously, were originally tailored for genomic selection purposes, but have proven to be useful also in association analyses, especially the so-called BayesC Bayesian methods. In a recent study, BayesD extended BayesC towards accounting for dominance effects and improved prediction accuracy and persistence in genomic selection. The current study investigated the power and precision of BayesC and BayesD in genome-wide association studies by means of stochastic simulations and applied these methods to a dairy cattle dataset.

Methods: The simulation protocol was designed to mimic the genetic architecture of quantitative traits as realistically as possible. Special emphasis was put on the joint distribution of the additive and dominance effects of causative mutations. Additive marker effects were estimated by BayesC and additive and dominance effects by BayesD. The dependencies between additive and dominance effects were modelled in BayesD by choosing appropriate priors. A sliding-window approach was used. For each window, the R. Fernando window posterior probability of association was calculated and this was used for inference purpose. The power to map segregating causal effects and the mapping precision were assessed for various marker densities up to full sequence information and various window sizes.

Results: Power to map a QTL increased with higher marker densities and larger window sizes. This held true for both methods. Method BayesD had improved power compared to BayesC. The increase in power was between -2 and 8% for causative genes that explained more than 2.5% of the genetic variance. In addition, inspection of the estimates of genomic window dominance variance allowed for inference about the magnitude of dominance at significant associations, which remains hidden in BayesC analysis. Mapping precision was not substantially improved by BayesD.

Conclusions: BayesD improved power, but precision only slightly. Application of BayesD needs large datasets with genotypes and own performance records as phenotypes. Given the current efforts to establish cow reference populations in dairy cattle genomic selection schemes, such datasets are expected to be soon available, which will enable the application of BayesD for association mapping and genomic prediction purposes.

Fig. 1

Scatterplot of the simulated joint distribution of the absolute value of additive effects ($\left|a\right|$) and dominance coefficients ($h=d/\left|a\right|)$

Fig. 2

Plot of window posterior probabilities of association (WPPA) obtained by BayesC (top) and BayesD (bottom), from the real data analyses

Fig. 3

Estimates of window genomic variances based on the real data analyses. The top and middle panels show the within-window estimates of genomic variances obtained by BayesC and BayesD, respectively. The bottom panel shows the within-window dominance variance obtained by BayesD. The window variances were multiplied by 1000. The window size was 0.5 cM

Fig. 4

Simulated gene effects and BayesC and BayesD results for a single simulated trait. The top left panel shows the simulated additive and dominance effects of the 10 causative mutations with a non-negligible effect for a randomly chosen trait for which dominance was important. The top right panel shows the genetic variances of these simulated causative mutations. The two panels in the middle show the within-window genomic variances obtained by BayesC (left) and BayesD (right). The window posterior probability of association (WPPA) obtained from both methods are shown at the bottom. The positions of the 10 causative mutations are indicated by a circle