The assignment success for 22 horse breeds registered in the Czech Republic: The machine learning perspective

The paper demonstrates the dependability of assignment testing in the identification of an appropriate breed to monitor comprehensive genetic information from molecular markers to analyse the collection of real population data covering 22 horse breeds registered in the Czech Republic, including native breeds and genetic resources. If 17 microsatellites are used, the mean number of alleles per locus corresponds to 10.4. The count of alleles at the individual loci ranges between five (HTG07) and 17 (ASB17). The loci ASB02, ASB23, HMS03, HTG10, and VHL20 exhibit the highest gene diversity and observed heterozygosity (both above 80%), with the mean value of 0.77 and 0.73, respectively. The moderate total inbreeding coefficient (5.2%) is estimated across all the loci and breeds. The levels of apparent breed differentiation span from zero between the Czech Warmblood and Slovak Warmblood to 0.15 between the Shetland Pony and Standardbred. The phylogenetic breed relationships are revealed via the NeighbourNet dendrogram constructed from Reynolds' genetic distances, which clearly separate the Coldblood draught, Hot/Warmblood, and Pony horses. Our results reveal that the Bayesian approach (the Rannala and Mountain technique) provides the most intensive prediction power (83.6%) out of the GeneClass tools and that the Bayes Net algorithm exhibits the best efficiency (78.4%) from the WEKA machine learning workbench options, considering the use of the five-fold cross-validation technique. The algorithms could be trained on large real reference data sets, and thus there appears another viable perspective for machine learning in horse ancestry testing. In this context, it is also important to stress the fact that innovated computational tools will potentially lead towards structuring a novel webserver to allow the identification of horse breeds.