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

Nonadditive genetic effects pose significant challenges to traditional genomic selection methods for quantitative traits. Machine learning approaches, particularly kernel-based methods, offer promising solutions to overcome these limitations. In this study, we developed a novel machine learning method, KPRR, which integrated a polynomial kernel into ridge regression to effectively capture nonadditive genetic effects. The predictive performance and computational efficiency of KPRR were evaluated using six datasets from various species, encompassing a total of 18 traits. All the traits were known to be influenced by additive, dominance, or epistatic genetic effects. We compared the performance of KPRR against six other genomic prediction methods: SPVR, BayesB, GBLUP, GEBLUP, GDBLUP, and DeepGS. For datasets dominated by additive effects, KPRR achieved superior prediction accuracies in the wheat dataset and comparable performance in the cattle dataset when compared to GBLUP. For datasets influenced by dominance effects, KPRR matched GDBLUP in accuracies in the pig dataset and outperformed GDBLUP in the sheep dataset. For datasets exhibiting epistatic effects, KPRR outperformed other methods in some traits, while BayesB showed superior performance in others. Incorporating nonadditive effects into a GBLUP model led to overall improvements in prediction accuracy. Regarding computational efficiency, KPRR was consistently the fastest, while BayesB was the slowest. Our findings demonstrated that KPRR provided significant advantages over traditional genomic prediction methods in capturing nonadditive effects.