Classification of cattle breeds based on the random forest approach

Abstract

The determination of breed composition or the population of origin has several practical applications, including the authentication of brand mono-breed products that have recently been developed in several livestock species, including cattle. Therefore, the aim of the present study was to verify the applicability of a random forest (RF) approach to distinguish between different dairy and beef cattle breeds based on a reduced panel of single nucleotide polymorphisms (SNPs). A total of 1,370 animals from 17 cattle breeds were genotyped with a 50 K SNP microarray. After SNP pruning, 5,296 SNPs were retained for further analysis. Six methods were used for determining panels of the most informative SNPs: Wright's fixation index (FST), principal component analysis (PCA), random forest (RF) based on the Gini index (GI; two variants) and RF based on the mean decrease in accuracy (MDA; two variants). There was a similar distribution of selected SNPs per BTA, especially for the FST and PCA methods. Two panels of 96 SNPs obtained with two variants of RF-GI and RF-MDA contained the same SNPs but with a different ranking. The percentage of correct classification on the test set (10% of cases randomly selected from the whole dataset) was 100% for the Angus, Hereford, Holstein, Brown Swiss, and Jersey breeds. For the Simmental, Piedmontese, Romagnola, Shorthorn, Norwegian Red, Charolais, Cika, Tyrol Grey, Limousin, Austrian Pinzgau, Slovak Pinzgau, and Slovak Spotted breeds, it was 75.00 – 87.50%, 0.00 – 66.67%, 66.67 – 100.00%, 80.00 – 100.00%, 0.00 – 50.00%, 75.00 – 100.00%, 0.00 – 66.67%, 90.91 - 100.00%, 91.67%, 81.82 - 100.00%, 93.33 - 100.00%, 55.56 - 100.00%, respectively. The overall correct classification rate (for all breeds together) was 87.14%, 94.29%, 90.00% and 88.57% for RF-GI, RF-MDA, FST and PCA, respectively. Consequently, the most accurate method for distinguishing between cattle breeds was RF based on the MDA, although the differences amongst models were not large.