Comparative analysis of genomic selection models for trait improvement in Arkansas rice varieties

Authors: GUO, MINGHAO - University Of Arkansas; SHA, XUEYAN - University Of Arkansas; DEGUZMAN, CHRISTIAN - University Of Arkansas; Jia, Yulin; Edwards, Jeremy

Submitted to: Arkansas Experiment Station Research Series
Publication Type: Experiment Station
Publication Acceptance Date: 8/19/2025
Publication Date: 8/19/2025
Citation: Guo, M., Sha, X., Deguzman, C.T., Jia, Y., Edwards, J. 2025. Comparative analysis of genomic selection models for trait improvement in Arkansas rice varieties. Arkansas Experiment Station Research Series.

Interpretive Summary:

Technical Abstract: Genomic selection (GS) is a powerful tool for accelerating the incorporation of minor effect genes into cultivars. This study evaluated the predictive performance of 13 GS models for seven key agronomic traits in Arkansas rice. Utilizing a panel of 554 rice lines and the LSU550 marker set, we compared 13 models, including linear (RR-BLUP, GBLUP), machine learning (tree-based, kernel-based, neural networks), and spline-based methods. Model performance was assessed using five-fold cross-validation. Results showed that model performance varied depending on the trait. Tree-based models such as RandomForest and GBM achieved the highest predictive accuracies for milling yield, head rice yield, and plant height. Gaussian Process regression performed best for grain chalkiness and thickness. Linear models were consistently accurate across traits and served as reliable baselines. Neural networks and MARS had the weakest predictive performance across most traits. These findings suggest that genomic selection models should be chosen based on the specific trait being analyzed. Future improvements could come from increasing marker density and refining machine learning techniques, particularly for traits with low predictive accuracy such as thickness and whiteness.