Simultaneous fitting of genomic-BLUP and Bayes-C components in a genomic prediction model

Abstract

BackgroundThe rapid adoption of genomic selection is due to two key factors: availability of both high-throughput dense genotyping and statistical methods to estimate and predict breeding values. The development of such methods is still ongoing and, so far, there is no consensus on the best approach. Currently, the linear and non-linear methods for genomic prediction (GP) are treated as distinct approaches. The aim of this study was to evaluate the implementation of an iterative method (called GBC) that incorporates aspects of both linear [genomic-best linear unbiased prediction (G-BLUP)] and non-linear (Bayes-C) methods for GP. The iterative nature of GBC makes it less computationally demanding similar to other non-Markov chain Monte Carlo (MCMC) approaches. However, as a Bayesian method, GBC differs from both MCMC- and non-MCMC-based methods by combining some aspects of G-BLUP and Bayes-C methods for GP. Its relative performance was compared to those of G-BLUP and Bayes-C.MethodsWe used an imputed 50 K single-nucleotide polymorphism (SNP) dataset based on the Illumina Bovine50K BeadChip, which included 48,249 SNPs and 3244 records. Daughter yield deviations for somatic cell count, fat yield, milk yield, and protein yield were used as response variables.ResultsGBC was frequently (marginally) superior to G-BLUP and Bayes-C in terms of prediction accuracy and was significantly better than G-BLUP only for fat yield. On average across the four traits, GBC yielded a 0.009 and 0.006 increase in prediction accuracy over G-BLUP and Bayes-C, respectively. Computationally, GBC was very much faster than Bayes-C and similar to G-BLUP.ConclusionsOur results show that incorporating some aspects of G-BLUP and Bayes-C in a single model can improve accuracy of GP over the commonly used method: G-BLUP. Generally, GBC did not statistically perform better than G-BLUP and Bayes-C, probably due to the close relationships between reference and validation individuals. Nevertheless, it is a flexible tool, in the sense, that it simultaneously incorporates some aspects of linear and non-linear models for GP, thereby exploiting family relationships while also accounting for linkage disequilibrium between SNPs and genes with large effects. The application of GBC in GP merits further exploration.

Multi‐trait genomic predictions using GBLUP and Bayesian mixture prior model in beef cattle

“…Our study demonstrated that ST-BayesCπ had better predictive ability than ST-GBLUP, with higher accuracy across all traits. These results are consistent with previous studies, suggesting Bayesian methods outperform GBLUP methods due to the assumption of different variances of SNP effects [2,16,27,40]. However, some studies based on real data suggested that the accuracies of GBLUP were similar to Bayesian methods [15,27,[40][41][42][43].…”

“…These results are consistent with previous studies, suggesting Bayesian methods outperform GBLUP methods due to the assumption of different variances of SNP effects [2,16,27,40]. However, some studies based on real data suggested that the accuracies of GBLUP were similar to Bayesian methods [15,27,[40][41][42][43]. Additionally, research has suggested that the superiority of GBLUP over Bayesian methods was due to the genetic architecture of the studied trait [44].…”

A review of traditional and machine learning methods applied to animal breeding

“…Bayesian methods that are implemented using MCMC algorithms are time-consuming and computationally demanding when they handle large number of SNPs. Therefore, several iterative (non-MCMC-based Bayesian) methods such as VanRaden's non-linear A/B (VanRaden, 2008), fastBayesB (Meuwissen et al ., 2009), MixP (Yu and Meuwissen, 2011) or emBayesR (Habier et al ., 2007) were developed to overcome the computational demands (Iheshiulor et al ., 2017). The aforementioned methods are computationally fast, and they result in prediction accuracies similar to those of the MCMC-based methods.…”

On the use of whole-genome sequence data for across-breed genomic prediction and fine-scale mapping of QTL

“…Models that simultaneously fit a GBLUP and a BayesC term have been used before in the literature, e.g. [9,32,33], and have been shown to yield high prediction accuracy. Our current implementation of this model is specifically directed at the use of sequence data.…”