Mutual information stacking method for prediction of the growth traits in pigs.

Abstract

Genomic prediction is a crucial technique for phenotype estimation, with the genomic best linear unbiased prediction (GBLUP) being the most widely adopted method. Yet, GBLUP falls short in capturing the intricate nonlinear relationships between genomic data and phenotypes. Given its ability to more effectively capture nonlinear genetic effects, machine learning (ML) has become increasingly appealing in genomic prediction. However, almost GBLUP and ML methods utilize all single nucleotide polymorphisms (SNPs) data for prediction, ignoring the fact that only a subset of SNPs are effective. This not only consumes computation time but also has poor prediction accuracy. So, this paper proposed a mutual information stacking method (MISM). Firstly, mutual information was introduced to select the SNPs with effect and remove the redundant SNPs. Then, we constructed a stacking model that can capture both linear and nonlinear relationships between SNPs and phenotypes to improve the prediction accuracy. To assess the effectiveness of MISM, we compared its performance on pig growth traits with GBLUP and other ML methods. The statistical analysis results indicated that MISM outperformed other ML models and GBLUP.