A Framework for Optimizing Breeding Pairs Using Artificial Intelligence

Abstract

<p>While selective breeding has played an important role in improving the economic performance of animals, traditional selection methods depend on animal-based data such as phenotypic or Estimated Breeding Values. The advent of novel genotyping technologies have led to genomic data, which directly probed into the genotypic configuration of the animals.</p> <p>This allows the exploitation of non-additive genetic components such as the dominance effects, which previously were not exploitable in selective breeding due to their dependence on the genotypic configurations of the parents, an aspect not made available through animalbased data. The use of such components has been relegated to crossbreeding systems, and rarely in within population mating systems. </p> <p>For this reason, the aim of this thesis is to explore the optimization of breeding pairs and mating decisions, with emphasis on the use of genomic data. This thesis will explore the use of such data in the exploitation of additive and dominance genetic components while constraining the inbreeding level increment. To cover the large sample space of possible solutions, this project will be conducted using artificial intelligence for the optimization of breeding pairs. The optimization method proposed in this study was validated using a simulated dataset. </p> <p>It is noted that there could be factors such as genetic architecture and data sizes that would affect the usability of genomic data in the optimization of breeding pairs, which was the reason this project starts by investigating the impact of these factors on the power and false positive rate of detecting quantitative trait loci (QTL) in a Genome-Wide Association Study (GWAS), a tool widely used for the detection of QTL and estimating the effect sizes of genomic regions. This study suggested significant impacts of sample sizes and number of markers, as well as genetic architecture of the traits on the power and false positive rates of the GWAS. This study also explored the performance of GWAS using two commonly used multiple testing correction methods, and also proposed a scoring method that could be used to test the optimality of thresholds between different multiple testing correction methods. </p> <p>From the findings of this foundational work, techniques that could improve the performance of GWAS experiments have been explored. One such techniques was the calculation of optimal threshold that takes into account the effects of genetic architecture and data size. For this calculation, a method based on Receiver Operating Characteristics was developed to calculate the optimal threshold of a GWAS. Simulation studies suggested this method performed better in binary classifications and marker selection for genomic predictions, with the use of this optimal threshold resulting in an increment of accuracy of genomic prediction up to 16.8% compared to that of the Bonferroni method, and 7.0% compared to the Benjamini-Hochberg FDR method. </p> <p>The calculation of optimal threshold requires information on the genetic architecture of the trait, and this has become the basis for the next part of the thesis, where a novel method that estimates the genetic architecture parameters such as number of QTL and shape of the effect size distributions was proposed, while taking into account the impact of various confounding factors such as correlation between markers, heterogeneity in linkage disequilibrium structures, and allele frequency distribution. Using this method, the estimated number of QTL with effect sizes 0.1 σ_e ranged from 69.9% to 167.0% (an average of 109.8%) of the true number of QTL, and for effect size 1.0 σ_e it ranged from 101.6% to 175.8% (an average of 123.6%). The method was developed to be able to estimate the QTL effect size, similar to a GWAS, but taking into account the impact of the confounding factors. This method would also allow the detection of QTL with smaller effect size with more confidence. New statistical tests designed to be powerful at the tail of the QTL distribution were developed, and an observation was made on the preference of utilization of test statistics for optimization of breeding pairs over the estimated effect size of the markers. </p> <p>For the final chapter, a framework for the optimization of breeding pairs was developed that could optimize both the additive and dominance genetic component while constraining the increment in inbreeding coefficient. For this framework, a genetic algorithm was used. Using the EBVs, this method successfully improved the additive genetic component by up to 87.0% compared to a truncation genomic selection method. Using heterozygosity as a mean of optimizing the dominance component, the genetic lift from the dominance component in offspring is approximately twice the additive genetic gain, although the lift only occurs in the first generation.</p> <p>This project is important for livestock producers or species conservationists who wished to improve the additive and non-additive genetic components in their breeding herds by using genomic data. It is anticipated that this framework could be further developed into a fullfledged product that could be utilized in a commercial setting.</p>