Adding Genetically Distant Individuals to Training Populations Reduces Genomic Prediction Accuracy in Barley

Abstract

One of the most important factors affecting genomic prediction accuracy appears to be training population (TP) composition. The objective of this study was to evaluate the effect of genomic relationship on genomic prediction accuracy and determine if adding increasingly unrelated individuals to a TP can reduce prediction accuracy. To accomplish this, a population of barley (Hordeum vulgare L.) lines from the University of Minnesota (lines denoted as MN) and North Dakota State University (lines denoted as ND) breeding programs were used for model training. Predictions were validated using two independent sets of progenies derived from MN × MN crosses and ND × ND crosses. Predictive ability sharply decreased with decreasing relationship between the TP and validation population (VP). More importantly, it was observed that adding increasingly unrelated individuals to the TP can actually reduce predictive ability compared with smaller TPs consisting of highly related individuals only. Reported results are possibly conditional on the relatively low marker density (342 single nucleotide polymorphisms [SNPs]) used. Nevertheless, these findings suggest plant breeding programs desiring to use genomic selection could benefit from focusing on good phenotyping of smaller TPs closely related to the selection candidates rather than developing large and diverse TPs.

Validating Genomewide Predictions of Genetic Variance in a Contemporary Breeding Program

“…This is not unexpected, since any error associated with the predicted marker effect will more strongly influence ĝ V than m (Zhong and Jannink, 2007;Lado et al, 2017). Although studies in barley have consistently reported a pattern of diminishing returns in prediction accuracy with increasing TP size (Lorenz et al, 2012;Lorenz and Smith, 2015;Sallam et al, 2015), simulations have indicated that TP size may be particularly important when predicting V G (Lehermeier et al, 2017). Interestingly, the predictive abilities of m SP for FHB severity (r MP = 0.69) and plant height (r MP = 0.62) were greater than those for m. We might expect that the predictive ability for m SP would be the intermediate of that for m and V G (i.e., the pattern observed for heading date), given the relative contributions of the mean and variance to the superior progeny mean, along with the unequal impact of prediction error described above.…”

“…These estimates did not change appreciably when removing families with very low estimates of V G (data not shown). Close relatedness is imperative for prediction accuracy (Habier et al, 2007), and genomewide selection research in barley has confirmed this (Lorenz et al, 2012;Lorenz and Smith, 2015). It is worth noting here the small sample size used to estimate predictive ability (n = 14 or 26).…”

Genomic assisted selection for enhancing line breeding: merging genomic and phenotypic selection in winter wheat breeding programs with preliminary yield trials

“…Accordingly, genomic selection is expected to give more accurate predictions if lines included in the training population are closely related to (Asoro et al 2011; Lehermeier et al 2014; Lorenz and Smith 2015) or even come from the same population as the selection candidates (Windhausen et al 2012; Charmet et al 2014). The underlying population structure can be readily deciphered when multiple large bi-parental populations (Heffner et al 2011a; Schulz-Streeck et al 2012; Riedelsheimer et al 2013; Lehermeier et al 2014) or larger heterotic groups (Technow et al 2013; Lehermeier et al 2014; Spindel et al 2015) are directly involved in the development of varietal candidates.…”

“…Furthermore, breeders frequently introgress foreign material in their breeding pools and lines are often derived by crosses between introduced and their own germplasm, resulting in an unclear population structure in such mixed line breeding populations (Sallam et al 2015; He et al 2016; Michel et al 2016). Simulation (Habier et al 2013) and empirical (Lorenz and Smith 2015) studies clearly showed that adding distant relatives to prediction models can have detrimental effects on the accuracy, thus there is serious need for an appropriate training population design to achieve high prediction accuracies with genomic selection in line breeding.…”

“…These efforts are often obstructed by different genetic backgrounds (Schulz-Streeck et al 2011, 2012) and linkage phase change between the training and selection population (Riedelsheimer et al 2013; Lorenz and Smith 2015). The incorporation of preliminary yield trials into the genomic selection framework could promote a more targeted pre-selection of marker sets due to prior knowledge of the genetic variation in different selection populations.…”