利用基因體選拔確認最佳的基因型

Abstract

在植物育種中，基因體選拔 (genomic selection) 可以基於基因型資料去挑選出優良的品系且能夠省去調查外表型的繁重工作。而在基因體選拔中則需要在族群中挑選出最佳的基因型組合去建立訓練族群 (training set)，而使得挑選出來的訓練族群所建立的模型能夠有優良的預測效果。在本篇研究中，利用基因組BLUP (GBLUP) 預測模型估計基因型值，並以正規化累計折損增益(Normalized Discounted Cumulative Gain; NDCG) 作為評估的指標。由於在育種中，大多數時間在意的是優良的品種的表現，因此採用 NDCG 基於排序正確性的方法作為評估標準。我們提出利用廣義決定係數 (generalized coefficient of determination; CD)來找出建立訓練集的最佳的基因型，並以四個資料集分別為兩組水稻資料、一組小麥資料與一組大豆資料使用R語言進行分析模擬試驗。結果顯示 CD 方法在遺傳率低且訓練族群大小較小時所挑選出的訓練集表現對於優秀品系的正確排序能力能夠優於其他方法。

In plant breeding, genomic selection can select superior lines based on their genotypes without laborious phenotyping. For genomic selection, we have to choose the best genotypes to build the training set, which will have an excellent prediction performance. In this study, we predicted the genetic values by the GBLUP model and evaluated the performance by the Normalized Discounted Cumulative Gain (NDCG). As breeding primarily focuses on the performance of outstanding varieties, we utilize the NDCG score, a criterion based on their ranking quality. We proposed a method to find the best genotypes for building the training set by using the generalized coefficient of determination (CD) and illustrating the performance by four datasets, including two rice datasets (tropical rice and 44K rice), a wheat dataset, and a soybean dataset. We implement our simulation and analysis in R language. The simulation results show that the CD method outperforms in selecting great lines with the correct order when the trait heritability is low, or the training set size is small.