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

Summary

This paper presents KPRR, a kernel-based machine learning method designed to improve genomic prediction by effectively capturing nonadditive genetic effects (dominance and epistatic interactions) alongside additive effects. Evaluated across 18 traits in wheat, cattle, pig, and sheep, KPRR demonstrated superior or comparable prediction accuracies to established methods (GBLUP, GDBLUP, BayesB) whilst maintaining the fastest computational performance. The findings suggest that explicitly modelling nonadditive effects enhances genomic prediction accuracy for quantitative traits influenced by dominance or epistatic interactions.

Regional applicability

The methodology is applicable to United Kingdom breeding programmes in livestock (cattle, sheep, pigs) and arable crops (wheat), where genomic selection is widely adopted. The approach could improve breeding value predictions in UK-based genomic selection schemes, particularly for traits with significant nonadditive genetic architecture.

Key measures

Prediction accuracy; computational efficiency; performance across additive, dominance, and epistatic genetic effect types

Outcomes reported

The study developed and evaluated KPRR, a novel machine learning method integrating polynomial kernels into ridge regression for genomic prediction. Performance was assessed across 18 traits from six datasets in wheat, cattle, pig, sheep and other species, comparing prediction accuracy and computational efficiency against six established genomic prediction methods.

Topic tags