Abstract #T83

# T83
Effect of prediction method and cross-validation approach on accuracy of DGV for feed efficiency traits.
Rafael M. O. Silva*1, Daniela A. L. Lourenco2, Breno O. Fragomeni2, Luciana Takada1, Rafael Espigolan1, Maria E. Z. Mercadante3, Fernando Baldi1, Guilherme C. Venturini1, Joslaine N. S. G. Cyrillo3, Ignacy Misztal2, Roberto Carvalheiro1, Lucia G. Albuquerque1, 1Univ Est Paulista Julio de Mesquita Filho,FCAV-UNESP, Jaboticabal, SP, Brazil, 2The University of Georgia, Athens, GA, 3APTA Center for Beef Cattle, Animal Science Institute, Sertaozinho, SP, Brazil.

Accuracies of direct genomic values (DGV) for feed efficiency traits obtained with different methods and cross-validation approaches were compared. After quality control 437,197 SNP genotypes were available for 761 Nellore cattle provided by the Institute of Animal Science, SP, Brazil. Methods of analysis were traditional BLUP, ssGBLUP, GBLUP, and BayesCPi. The traits were residual feed intake (RFI), feed conversion ratio (FCR), average daily gain (ADG) and dry matter intake (DMI). Model included fixed effects of contemporary groups (sex, year of birth, and pen), month of birth, and the covariable age of dam (linear and quadratic effects); and, as random, the additive animal effects. Three cross-validation approaches were considered: WPRO – validation was done on animals that did not have progeny; UNREL – the data set was split into 3 less related subsets; RAN – the data set was randomly divided into 4 subsets and the validation was done in each subset at a time. The accuracy of DGV was calculated as the Pearson correlation between corrected phenotype and the DGV divided by square root of heritability. Accuracies ranged from 0.01 (with UNREL) to 0.78 (with RAN) for studied traits. The inclusion of genomic information increased more than 10% of the average accuracy of predictions over traditional BLUP; on average, GBLUP showed more accurate predictions of DGV than BayesCPi (0.33 and 0.29, respectively). For the RAN cross-validation approach, accuracies were 50% higher with ssGBLUP than GBLUP, especially for traits with high heritabilities (ADG- h2 = 0.55, and DMI – h2 = 0.58). The most accurate predictions were obtained using RAN, ranging from 0.28 to 0.78. On the other hand, the UNREL cross-validation approach provided the less accurate predictions, ranging from 0.001 to 0.29. With WPRO accuracies of DGVs were from 0.12 to 0.69. These results show that accurate genomic prediction can be obtained for all analyzed traits, especially for those with high heritability. Accuracies of DGV are higher when animals in validation are more related to those in training. São Paulo Research Foundation (FAPESP) grant 2013/01228-5 associated to 2009/16118-5.

Key Words: beef cattle, genomic selection