Su Ruilin, Lv Jingbo, Xue Yahui, Jiang Sheng, Zhou Lei, Jiang Li, Tan Junyan, Shen Zhencai, Zhong Ping, Liu Jianfeng
College of Science, China Agricultural University, Beijing, 100083, China.
College of Animal Science and Technology, China Agricultural University, Beijing, 100193, China.
Genet Sel Evol. 2025 Mar 10;57(1):13. doi: 10.1186/s12711-025-00957-3.
The effectiveness of genomic prediction (GP) significantly influences breeding progress, and employing SNP markers to predict phenotypic values is a pivotal aspect of pig breeding. Machine learning (ML) methods are usually used to predict phenotypic values since their advantages in processing high dimensional data. While, the existing researches have not indicated which ML methods are suitable for most pig genomic prediction. Therefore, it is necessary to select appropriate methods from a large number of ML methods as long as genomic prediction is performed. This paper compared the performance of popular ML methods in predicting pig phenotypes and then found out suitable methods for most traits.
In this paper, five commonly used datasets from other literatures were utilized to compare the performance of different ML methods. The experimental results demonstrate that Stacking performs best on the PIC dataset where the trait information is hidden, and the performs of kernel ridge regression with rbf kernel (KRR-rbf) closely follows. Support vector regression (SVR) performs best in predicting reproductive traits, followed by genomic best linear unbiased prediction (GBLUP). GBLUP achieves the best performance on growth traits, with SVR as the second best.
GBLUP achieves good performance for GP problems. Similarly, the Stacking, SVR, and KRR-RBF methods also achieve high prediction accuracy. Moreover, LR statistical analysis shows that Stacking, SVR and KRR are stable. When applying ML methods for phenotypic values prediction in pigs, we recommend these three approaches.
基因组预测（GP）的有效性显著影响育种进展，利用单核苷酸多态性（SNP）标记预测表型值是猪育种的一个关键方面。机器学习（ML）方法因其在处理高维数据方面的优势，通常用于预测表型值。然而，现有研究尚未表明哪种ML方法最适合大多数猪的基因组预测。因此，只要进行基因组预测，就有必要从大量ML方法中选择合适的方法。本文比较了常用ML方法在预测猪表型方面的性能，进而找出了适用于大多数性状的方法。
本文利用其他文献中的五个常用数据集来比较不同ML方法的性能。实验结果表明，在性状信息隐藏的PIC数据集上，Stacking表现最佳，带径向基核（rbf）的核岭回归（KRR-rbf）紧随其后。支持向量回归（SVR）在预测繁殖性状方面表现最佳，其次是基因组最佳线性无偏预测（GBLUP）。GBLUP在生长性状上表现最佳，SVR次之。
GBLUP在GP问题上表现良好。同样，Stacking、SVR和KRR-RBF方法也具有较高的预测准确率。此外，LR统计分析表明Stacking、SVR和KRR是稳定的。在将ML方法应用于猪的表型值预测时，我们推荐这三种方法。
