Horticulture and Crop Science Exit Seminar
Title: Accuracy of genomic selection in a soft winter wheat (Triticum aestivum L.) breeding program
Genomic selection (GS) is a new marker assisted selection tool that utilizes data from lines in a training population (TP) to predict performance of other related lines by generating their genetic estimated breeding values. The selection process is complicated by genotype by environment interaction (GEI), as the performance of lines in one environment may not predict their performance in other environments. It is critical to evaluate and optimize the prediction accuracy of GS with the existence of GEI. The prediction accuracy of GS can be evaluated by testing the GS model on different types of validation populations (VP). This study utilized subset of lines from the TP as the VP, and also utilized a VP composed of lines that were not included in, yet were genetically related to the TP. Our objectives were: 1) to assess the GEI patterns and generate trait stability indices; 2) to evaluate GS accuracy for traits and trait stability indices through different modeling approaches for within and between-population predictions; 3) to assess the effects of different optimization approaches on GS accuracy for between-environment predictions for agronomic traits within population; 4) to validate the GS accuracy from different optimization approaches for between-environment predictions across populations. An elite population (EP) of 273 lines and a yield population (YP) of 294 lines were phenotyped in independent sets of environments. A total of up to 24 different environments, representing four years across locations in five different states were assessed. The EP and YP were both phenotyped for yield (YLD), test weight (TW), plant height (HGT), and heading date (HD), and were genotyped with a common set of 3,537 single nucleotide polymorphism (SNP) markers. The EP was additionally phenotyped for seven quality traits. We produced useful GS prediction accuracy for within-population predictions for all traits (r ranging from 0.33 to 0.74) and most trait stability indices. We observed that ridge regression Best Linear Unbiased Prediction model was as predictive as other GS models, including the ones incorporating GEI term. The best approach to optimize the TP for between-environment accuracy was to subset markers that had the significant and stable effects coupled with eliminating least predictive lines in the TP. The between-population prediction for TW, HGT and HD were useful (r exceeded 0.29) though the between-population prediction accuracy for YLD was low (r less than 0.18). The EP and YP environments were separated in two distinct clusters based on the marker effects of YLD, and further supported the hypothesis that the low GS accuracy for YLD was mainly due to the marker effects by population interaction. This suggests that in order to obtain maximum GS accuracy for complex traits such as yield, the population to be predicted could consist of the same lines as in TP, but would be grown under different environments, or the new population to be predicted may be directly derived from TP. Our findings are directly applicable for wheat breeders in North-Eastern U.S. to best design GS schemes, and to implement GS in their breeding programs to achieve higher genetic gains with reduced costs and time than conventional breeding methods.