Title:Simulation Analysis on Genomic Selection of Grouper (Epinephelus coioides) Breeding for Categorical Traits
Volume: 1
Issue: 1
Author(s): Zhiyuan Ma and Xinxin You*
Affiliation:
- Shenzhen Key Lab of Marine Genomics, Guangdong Provincial Key Lab of Molecular Breeding in Marine Economic Animals, Shenzhen BGI Academy of Marine Sciences, BGI Marine, Shenzhen 518083,China
Keywords:
gBLUP, rrBLUP, BayesA, BayesC, categorical traits, grouper, genomic selection.
Abstract:
Background: The basic principle of genome selection (GS) is to establish a model of genome
estimated breeding value (GEBV) by using single-nucleotide polymorphisms (SNPs) covering
the entire genome. Despite the decreasing cost of high-throughput genotyping, the GS strategy remains
expensive due to the need for phenotyping and genotyping for a large number of samples.
Simulation analysis of genome selection is a popular, lower-cost method to determine an optimal
breeding program of GS.
Objective: To evaluate the utility of simulation data to study the influence of different factors on
algorithms. This could be helpful for developing genome selection breeding strategies, especially
for stress and resistance traits of fish.
Methods: Real data of orange-spotted grouper (Epinephelus coioides) were obtained from a previous
genome-wide association study. Ammonia tolerance, different population sizes, SNP density,
QTL number, kinship (base mutation rate), and heritability were considered. All of the phenotypes
and genotypes were generated by AlphaSimR simulation software. Four genome selection algorithms
(gBLUP, rrBLUP, BayesA, and BayesC) were tested to derive GEBV, and their accuracies
(area under the curve, AUC) were compared.
Results: In different scenarios, the AUC ranges from 0.4237 to 0.6895 for BayesA, 0.4282 to
0.6878 for BayesC, 0.4278 to 0.6798 for gBLUP, and 0.4346 to 0.6834 for rrBLUP. The mean AUC
of these four algorithms was not significantly different (0.547–0.548). The accuracies of the four
genome selection algorithms were similar but had different predictive performances in specific scenarios.
The gBLUP was most stable, and the rrBLUP was slightly better at predicting low heritability
traits. When the number of individuals was small, the BayesA and BayesC algorithms were more
robust.
Conclusion: A practical GS scheme should be optimized in accordance with marker density, heritability,
and reference population size. Adequate preliminary research is necessary. The results provide
a framework for the design of genomic selection schemes in E. coioides breeding.
