RT Dissertation/Thesis
T1 Evaluation of alternative statistical methods for genomic selection for quantitative traits in hybrid maize
A1 Schulz-Streeck,Torben
WP 2013/07/23
AB The efficacy of several contending approaches for Genomic selection (GS) were tested using different simulation and empirical maize breeding datasets. Here, GS is viewed as a general approach, incorporating all the different stages from the phenotypic analysis of the raw data to the marker-based prediction of the breeding values. The overall goal of this study was to develop and comparatively evaluate different approaches for accurately predicting genomic breeding values in GS. In particular, the specific objectives were to:
(1) Develop different approaches for using information from analyses preceding the marker-based prediction of breeding values for GS. 
(2) Extend and/or suggest efficient implementations of statistical methods used at the marker-based prediction stage of GS, with a special focus on improving the predictive accuracy of GS in maize breeding. 
(3) Compare different approaches to reliably evaluate and compare methods for GS. 
An important step in the analyses preceding the marker-based prediction is the phenotypic analysis stage. One way of combining phenotypic analysis and marker-based prediction into a single stage analysis is presented. However, a stagewise analysis is typically computationally more efficient than a single stage analysis. Several different weighting schemes for minimizing information loss in stagewise analyses are therefore proposed and explored. It is demonstrated that orthogonalizing the adjusted means before submitting them to the next stage is the most efficient way within the set of weighting schemes considered. Furthermore, when using stagewise approaches, it may suffice to omit the marker information until the very last stage, if the marker-by-environment interaction has only a minor influence, as was found to be the case for the datasets considered in this thesis. It is also important to ensure that genotypic and phenotypic data for GS are of sufficiently high-quality. This can be achieved by using appropriate field trial designs and carrying out adequate quality controls to detect and eliminate observations deemed to be outlying based on various diagnostic tools. Moreover, it is shown that pre-selection of markers is less likely to be of high practical relevance to GS in most cases. Furthermore, the use of semivariograms to select models with the greatest strength of support in the data for GS is proposed and explored. It is shown that several different theoretical semivariogram models were all well supported by an example dataset and no single model was selected as being clearly the best.
Several methods and extensions of GS methods have been proposed for marker-based prediction in GS. Their predictive accuracies were similar to that of the widely used ridge regression best linear unbiased prediction method (RR-BLUP). It is thus concluded that RR-BLUP, spatial methods, machine learning methods, such as componentwise boosting, and regularized regression methods, such as elastic net and ridge regression, have comparable performance and can therefore all be routinely used for GS for quantitative traits in maize breeding. Accounting for environment-specific or population-specific marker effects had only minor influence on predictive accuracy contrary to findings of several other studies. However, accuracy varied markedly among populations, with some populations showing surprisingly very low levels of accuracy. Combining different populations prior to marker-based prediction improved prediction accuracy compared to doing separate population-specific analyses. Moreover, polygenetic effects can be added to the RR-BLUP model to capture genetic variance not captured by the markers. However, doing so yielded minor improvements, especially for high marker densities. To relax the assumption of homogenous variance of markers, the RR-BLUP method was extended to accommodate heterogeneous marker variances but this had negligible influence on the predictive accuracy of GS for a simulated dataset. 
The widely used information-theoretic model selection criterion, namely the Akaike information criterion (AIC), ranked models in terms of their predictive accuracies similar to cross-validation in the majority of cases. But further tests would be required to definitively determine whether the computationally more demanding cross-validation may be substituted with the more efficient model selection criteria, such as AIC, without much loss of accuracy.
Overall, a stagewise analysis, in which the markers are omitted until at the very last stage, is recommended for GS for the tested datasets. The particular method used for marker-based prediction from the set of those currently in use is of minor importance. Hence, the widely used and thoroughly tested RR-BLUP method would seem adequate for GS for most practical purposes, because it is easy to implement using widely available software packages for mixed models and it is computationally efficient.
K1 Biostatistik
K1 Pflanzenzüchtung
K1 Populationsgenetik
PP Hohenheim
PB Kommunikations-, Informations- und Medienzentrum der Universität Hohenheim
UL http://opus.uni-hohenheim.de/volltexte/2013/861