ABSTRACT: Breed identification is an important prerequisite for the protection, development and utilization of animal genetic resources. This study developed an accurate identification strategy for chicken breeds using whole-genome sequencing data from 492 individuals belonging to 14 chicken breeds. These breeds include eight local Chinese breeds (Tibetan chicken, Chahua chicken, Daweishan chicken, Liyang chicken, Lindian chicken, Silky chicken, Dongxiang blue-shell egg chicken, and WenChang chicken), three standard chicken breeds (Rhode Island Red, Leghorn, and Light Sussex chicken), two commercial breeds (Cobb broiler and Yellow Plumage Dwarf chicken) and the Red Jungle fowl. We compared three ancestry informative marker (AIM) detection methods (Fst, In, and PCA-correlated SNPs) and four machine learning classifiers (K-NearestNeighbor, Support Vector Machine, Random Forest, and XGBoost) to identify the best breed identification model. A total of 30,831 high-information SNPs (Single nucleotide polymorphism) were detected and selected from these breeds using the three AIM detection methods. We found that several AIM methods performed well, but In was the best. Machine learning classifiers were implemented to fit the important SNP loci, and ROC (receiver operating characteristic curve) curves were generated to evaluate the performance of these machine learning classifiers. The ROC curves and 5-fold cross-validation results indicated that XGBoost was the best machine learning classifier, with the largest AUC (Area Under Curve) (macro-AUC=0.9996). In addition, XGBoost achieved 100% accuracy using only 238 SNPs. In this study, it was observed that utilizing only 238 SNPs was effective for breed identification. We found that the combination of XGBoost and In was the optimal strategy for breed identification. This study provides a new method for breed identification, which is highly important for the breeding and preservation of animal genetic resources.
