TY - THES T1 - Genomic prediction in rye A1 - Bernal-Vasquez,Angela-Maria Y1 - 2018/02/20 N2 - Technical progress in the genomic field is accelerating developments in plant and animal breeding programs. The access to high-dimensional molecular data has facilitated acquisition of knowledge of genome sequences in many economically important species, which can be used routinely to predict genetic merit. Genomic prediction (GP) has emerged as an approach that allows predicting the genomic estimated breeding value (GEBV) of an unphenotyped individual based on its marker profile. The approach can considerably increase the genetic gain per unit time, as not all individuals need to be phenotyped. Accuracy of the predictions are influenced by several factors and require proper statistical models able to overcome the problem of having more predictor variables than observations. Plant breeding programs run for several years and genotypes are evaluated in multi environment trials. Selection decisions are based on the mean performance of genotypes across locations and later on, across years. Under this conditions, linear mixed models offer a suitable and flexible framework to undertake the phenotypic and genomic prediction analyses using a stage-wise approach, allowing refinement of each particular stage. In this work, an evaluation and comparison of outlier detection methods, phenotypic analyses and GP models were considered. In particular, it was studied whether at the plot level, identification and removal of possible outlying observations has an impact on the predictive ability. Further, if an enhancement of phenotypic models by spatial trends leads to improvement of GP accuracy, and finally, whether the use of the kinship matrix can enhance the dissection of GEBVs from genotype-by-year (GY) interaction effects. Here, the methods related to the mentioned objectives are compared using experimental datasets from a rye hybrid breeding program. Outlier detection methods widely used in many German plant breeding companies were assessed in terms of control of the family-wise error rate and their merits evaluated in a GP framework (Chapter 2). The benefit of implementation of the methods based on a robust scale estimate was that in routine analysis, such procedures reliably identified spurious data. This outlier detection approach per trial at the plot level is conservative and ensures that adjusted genotype means are not severely biased due to outlying observations. Whenever it is possible, breeders should manually flag suspicious observations based on subject-matter knowledge. Further, removing the flagged outliers identified by the recommendedmethods did not reduce predictive abilities estimated by cross validation (GP-CV) using data of a complete breeding cycle. A crucial step towards an accurate calibration of the genomic prediction procedure is the identification of phenotypic models capable of producing accurate adjusted genotype mean estimates across locations and years. Using a two-year dataset connected through a single check, a three-stage GP approach was implemented (Chapter 3). In the first stage, spatial and non-spatial models were fitted per locations and years to obtain adjusted genotype-tester means. In the second stage, adjusted genotype means were obtained per year, and in the third stage, GP models were evaluated. Akaike information criterion (AIC) and predictive abilities estimated from GP-CV were used as model selection criteria in the first and in the third stage. These criteria were used in the first stage, because a choice had to be made between the spatial and non-spatial models and in the third stage, because the predictive abilities allow a comparison of the results of the complete analysis obtained by the alternative stage-wise approaches presented in this thesis. The second stage was a transitional stage where no model selection was needed for a given method of stage-wise analysis. The predictive abilities displayed a different ranking pattern for the models than the AIC, but both approaches pointed to the same best models. The highest predictive abilities obtained for the GP-CV at the last stage did not coincide with the models that AIC and predictive ability of GP-CV selected in the first stage. Nonetheless, GP-CV can be used to further support model selection decisions that are usually based only upon AIC. There was a trend of models accounting for row and column variation to have better accuracies than the counterpart model without row and column effects, thus suggesting that row-column designs may be a potential option to set up breeding trials. While bulking multi-year data allows increasing the training set size and covering a wider genetic background, it remains a challenge to separate GEBVs from GY effects, when there are no common genotypes across years, i.e., years are poorly connected or totally disconnected. First, an approach considering the two-year dataset connected through a single check, adjusted genotype means were computed per year and submitted to the GP stage (Chapter 3). The year adjustment was done in the GP model by assuming that the mean across genotypes in a given year is a good estimate of the year effect. This assumption is valid because the genotypes evaluated in a year are a sample of the population. Results indicated that this approach is more realistic than relying on the adjustment of a single check. A further approach entailed the use of kinship to dissect GY effects from GEBVs (Chapter 4). It was not obvious which method best models the GY effect, thus several approaches were compared and evaluated in terms of predictive abilities in forward validation (GP-FV) scenarios. It was found that for training sets formed by several disconnected years’ data, the use of kinship to model GY effects was crucial. In training sets where two or three complete cycles were available (i.e. there were some common genotypes across years within a cycle), using kinship or not yielded similar predictive abilities. It was further shown that predictive abilities are higher for scenarios with high relatedness degree between training and validation sets, and that predicting a selection of top-yielding genotypes was more accurate than predicting the complete validation set when kinship was used to model GY effects. In conclusion, stage-wise analysis is recommended and it is stressed that the careful choice of phenotypic and genomic prediction models should be made case by case based on subject matter knowledge and specificities of the data. The analyses presented in this thesis provide general guidelines for breeders to develop phenotypic models integrated with GP. The methods and models described are flexible and allow extensions that can be easily implemented in routine applications. KW - Roggen KW - Genotypisierung KW - Prognose KW - Statistik CY - Hohenheim PB - Kommunikations-, Informations- und Medienzentrum der Universität Hohenheim AD - Garbenstr. 15, 70593 Stuttgart UR - http://opus.uni-hohenheim.de/volltexte/2018/1436 ER -