KPRR: a novel machine learning approach for effectively capturing nonadditive effects in genomic prediction.

Area of Science:

• Quantitative genetics
• Genomic selection
• Machine learning

Background:

• Nonadditive genetic effects complicate traditional genomic selection for quantitative traits.
• Kernel-based machine learning methods offer potential solutions.
• A novel method is needed to effectively capture complex genetic interactions.

Purpose of the Study:

• To develop and evaluate KPRR, a novel machine learning method integrating a polynomial kernel into ridge regression.
• To assess KPRR's ability to capture nonadditive genetic effects (dominance and epistasis).
• To compare KPRR's predictive performance and computational efficiency against existing genomic prediction methods.

Main Methods:

• Developed KPRR by integrating a polynomial kernel with ridge regression.
• Evaluated KPRR on six diverse datasets across multiple species and 18 traits.
• Compared KPRR against six established methods: SPVR, BayesB, GBLUP, GEBLUP, GDBLUP, and DeepGS.

Main Results:

• KPRR showed superior or comparable prediction accuracies for additive, dominance, and epistatic effects across datasets.
• KPRR demonstrated superior predictive performance in wheat (additive) and sheep (dominance) datasets.
• KPRR was the most computationally efficient method, significantly outperforming others, especially BayesB.

Conclusions:

• KPRR effectively captures nonadditive genetic effects, offering significant advantages over traditional genomic prediction.
• The integration of polynomial kernels in ridge regression provides a powerful approach for genomic selection.
• KPRR represents a computationally efficient and accurate tool for predicting complex quantitative traits.