於多環境試驗中進行基因體選種之訓練集最佳化

Abstract

基因型與環境交互作用 (Genotype-by-environment Interaction, G×E) 是植物育種中的常見現象，此因子會影響多環境試驗 (Multi-environment Trials, METs) 中的選拔準確性。將基因體最佳線性無偏預測模型 (Genomic Best Linear Unbiased Prediction Model, GBLUP Model) 納入G×E效應並進行基因體選種 (Genomic Selection, GS) ，能在多環境試驗之中提升預測準確性。本研究使用兩種決定係數 (Coefficient of Determination, CD) 指標應用於訓練集最佳化中的方法，並針對識別優良品種的能力，將其與隨機抽取訓練集之方法進行比較。本研究使用基因演算法 (Genetic Algorithm, GA) ，從三個作物資料集: 水稻 (Oryza sativa L.)、大麥 (Hordeum vulgare L.) 和玉米 (Zea mays L.) 中選出最佳訓練集。以三項評估指標: 標準化折扣累積增益 (Normalized Discounted Cumulative Gain, NDCG) 、Spearman等級相關係數 (Spearman’s Rank Correlation, SRC)、以及名次總和比率 (Rank Sum Ratio, RSratio) 評估上述三種方法的表現。結果顯示，基於 CD 指標選出的訓練集在三項評估指標上都表現得較好。將兩個CD指標的表現進行比較，結果顯示 CDmean(v2) 於SRC與RSratio兩項指標皆優於CDmean.MET，尤其是在使用較大的訓練集規模時。因此，本研究建議使用 CDmean(v2) 在多環境試驗中將訓練集最佳化。

Genotype-by-environment interaction (G×E) is a key factor in plant breeding, impacting multi-environment trials (METs) for accurate selection. Genomic selection (GS) can improve prediction accuracy across environments, especially with Genomic best linear unbiased prediction (GBLUP) models that account for G×E effects. This study evaluates training set optimization using two coefficient of determination (CD) criteria and compares them to random selection based on the ability to identify elite varieties. A genetic algorithm identified optimal training sets from three datasets of rice (Oryza sativa L.), barley (Hordeum vulgare L.), and maize (Zea mays L.), and their performance was assessed using normalized discounted cumulative gain (NDCG), Spearman’s rank correlation (SRC), and rank sum ratio (RSratio). CD-based training sets showed better performance among these evaluation metrics. The performance of the two CD criteria were compared. CDmean(v2) outperformed CDmean.MET in SRC and RSratio especially in larger training set sizes. Therefore, CDmean(v2) was highly recommended to select training sets in multi-environment trials.