多環境實驗中 GBLUP 模型之比較

Abstract

基因體選拔 (genomic selection, GS) 可以透過全基因體遺傳資訊預測未觀察的基因型表現。在多環境條件下，基因型與環境 (genotype-by-environment, G×E) 交互作用被視為影響模型預測能力的重要因子。已有多名研究者將 G×E 交互作用納入模型結構中，並提出多種基因體最佳線性無偏預測 (genomic best linear unbiased prediction, GBLUP) 之模型。然而，目前對這些模型在多環境下的預測表現仍缺乏全面性的比較。本研究目的為評估不同 GBLUP 模型在多環境下的預測能力，並探討模型結構對預測性能之影響。我們使用水稻、玉米和大麥資料集基於真實基因體數據生成的模擬表型數據與其真實的表型數據，設計多種情境及交叉驗證方法來評估各模型的性能。結果顯示，MGE 模型的預測能力跟環境間遺傳相關性呈正相關。G×E 交互作用較弱時，MGE 模型顯示出更高且穩定的預測能力。G×E 交互作用較強時，雖然 MGE 模型整體性能有所下降，但相較於其他模型，仍具有良好的競爭力。至於 MGBLUP 模型，則因其模型結構複雜且待估計參數較多，所以在多數場景下測試時，性能仍有侷限。此外，跨環境中使用不同的訓練資料，可能有助於提升考量不同環境間遺傳共變異數的模型之預測能力。

Genomic selection (GS) can predict the performance of unobserved genotypes by utilizing whole-genome genetic information. Under multi-environmental conditions, genotype-by-environment (G×E) interaction is considered an important factor influencing prediction ability. Many researchers have incorporated G×E interactions into model structures and proposed various genomic best linear unbiased prediction (GBLUP) models. However, a comprehensive comparison of these models across multiple environments is still insufficient. This study aims to evaluate the prediction ability of different GBLUP models across multiple environments and to investigate the impact of model structure on prediction performance. Based on real genomic data from rice, DST2 maize, and barley, we generated simulated phenotypic data and also utilized their real phenotypic data. We then designed multiple scenarios and cross-validation methods to assess model performance. The results showed that the prediction ability of the MGE model was positively associated with the genetic correlation between environments. When G×E interactions were weak, the MGE model demonstrated higher and more stable prediction performance. When G×E interactions were strong, the overall performance of the MGE model declined, but it remained competitively comparable to other models. In contrast, the MGBLUP model exhibited limited performance in most scenarios, likely due to its complex structure and the large number of parameters to be estimated. Moreover, using different training data across environments may enhance the prediction ability of models that account for genetic covariance between environments.