水稻基因組選種之模擬研究─訓練族群預測模型之建立與最低投入試驗規模之確立

Abstract

Genomic selection is a new strategy of marker-assisted selection. The statistical model built by genotypic and phenotypic data of a training population is used to estimate breeding values of targeted traits from marker genotypes of individual plants in a genetic recombinant population derived from the training population, and then genomic selection is conducted based on individual genomic estimated breeding values. The prediction accuracy of genomic estimated breeding values can be affected by several factors, including statistical methods of the prediction model, numbers of genotyped markers, size of training population, and genetic nature of targeted traits. The present study simulated 192 sets of genotypic and phenotypic data similar to rice recombinant inbred populations as in silico training populations, among which effective QTL numbers, population size, marker number, and narrow-sense heritability were assigned at different levels. The prediction accuracy of each 192 sets of data was then calculated and compared, using prediction models built by the RR-BLUP (ridge regression best linear unbiased prediction), BL (Bayesian LASSO), and RKHS (reproducing kernel Hilbert space) methods respectively. Finally, the most effective inputs (population size and marker number) for model-training trails were determined at different levels of heritability, given that the expected prediction accuracy of genomic selection is greater than 0.5 and its prediction accuracy is greater than that of phenotypic selection.