On Breed Composition Estimation of Cross-Bred Animals Using Non-Linear Optimisation

Abstract

Genetically admixed animals are common in most quantitative genetic analysis, and usually are a result of intended crosses between two or more pure breed populations to enhance productivity. Disregarding the genetic heterogeneous architecture of admixed individuals may lead to poor or even wrong inference about the quality, quantity and genome location of genetic factors affecting phenotypes, and it could reduce the accuracy of estimates of genetic merit. In this article a nonlinear optimisation approach (constrained genomic regression, CGR) is presented to describe the marker genotype of a focus animal as a linear function of marker allele frequencies of possible populations of origin. The algorithm was tested on a beef cattle data set consisting of 11639 animals from 11 different breeds with marker genotypes of 4022 single nucleotide polymorphisms, which were used to generate 5000 artificially cross-bred animals. For comparison the data set was also analysed with the ADMIXTURE software (ADM). CGR outperformed ADM with a maximum difference between the true and estimated breed proportion of 0.25 and 0.28 for the 5 and 25 cross-over data set respectively. For ADM this parameter was 0.83 and 0.66. The mean squared estimation error was 15 and 5 times larger for ADM compared to CGR for the 5 and 25 cross-over data set respectively. In addition, CGR always outperformed ADM in terms of speed by factor 20.