Weighted Kernel Ridge Regression to Improve Genomic Prediction
Chenguang Diao,
Yue Zhuo,
Ruihan Mao,
Weining Li,
Heng Du,
Lei Zhou,
Jianfeng Liu
Affiliations
Chenguang Diao
State Key Laboratory of Animal Biotech Breeding, Frontiers Science Center for Molecular Design Breeding (MOE), College of Animal Science and Technology, China Agricultural University, Beijing 100193, China
Yue Zhuo
State Key Laboratory of Animal Biotech Breeding, Frontiers Science Center for Molecular Design Breeding (MOE), College of Animal Science and Technology, China Agricultural University, Beijing 100193, China
Ruihan Mao
State Key Laboratory of Animal Biotech Breeding, Frontiers Science Center for Molecular Design Breeding (MOE), College of Animal Science and Technology, China Agricultural University, Beijing 100193, China
Weining Li
State Key Laboratory of Animal Biotech Breeding, Frontiers Science Center for Molecular Design Breeding (MOE), College of Animal Science and Technology, China Agricultural University, Beijing 100193, China
Heng Du
State Key Laboratory of Animal Biotech Breeding, Frontiers Science Center for Molecular Design Breeding (MOE), College of Animal Science and Technology, China Agricultural University, Beijing 100193, China
Lei Zhou
State Key Laboratory of Animal Biotech Breeding, Frontiers Science Center for Molecular Design Breeding (MOE), College of Animal Science and Technology, China Agricultural University, Beijing 100193, China
Jianfeng Liu
State Key Laboratory of Animal Biotech Breeding, Frontiers Science Center for Molecular Design Breeding (MOE), College of Animal Science and Technology, China Agricultural University, Beijing 100193, China
Nonparametric models have recently been receiving increased attention due to their effectiveness in genomic prediction for complex traits. However, regular nonparametric models cannot effectively differentiate the relative importance of various SNPs, which significantly impedes the further application of these methods for genomic prediction. To enhance the fitting ability of nonparametric models and improve genomic prediction accuracy, a weighted kernel ridge regression model (WKRR) was proposed in this study. For this new method, different weights were assigned to different SNPs according to the p-values from GWAS, and then a KRR model based on these weighted SNPs was constructed for genomic prediction. Cross-validation was further adopted to choose appropriate hyper-parameters during the weighting and prediction process for generalization. We compared the predictive accuracy of WKRR with the genomic best linear unbiased prediction (GBLUP), BayesR, and unweighted KRR using both simulated and real datasets. The results showed that WKRR outperformed unweighted KRR in all simulated scenarios. Additionally, WKRR achieved an average improvement of 1.70% in accuracies across all traits in a mice dataset and 2.17% for three lactation-related traits in a cattle dataset compared to GBLUP, and yielded competitive results compared to BayesR. These findings demonstrated the great potential of weighted nonparametric models for genomic prediction.
