Cost-effective genomic prediction for fertility traits: a comparison of machine learning and conventional models...

Fertility is one of the major factors affecting the efficiency of dairy herds, and genomic selection on milk yield, while ignoring fertility, has resulted in a decline in heifer fertility. Considering the cost of breeding, it is important to enhance the accuracy of genomic selection with a cost-effective way for fertility traits. This study investigates the genomic prediction (GP) of fertility traits in Holstein heifers using both machine learning (ML) and conventional methods, based on data from SNP arrays and low-coverage sequencing. In this study, phenotype and pedigree records from 45 320 Holstein heifers were collected, of which 3 683 cattle were genotyped using low-coverage sequencing. We first estimated the heritability for age at first service (AFS), gestation length of heifers (GLh) and age at first calving (AFC). We then compared the prediction performance of ML methods, kernel ridge regression (KRR), support vector regression, and random forest regression, with genomic best linear unbiased prediction (GBLUP), ssGBLUP (Single-step GBLUP) and BayesR3 regarding GP accuracy and unbiasedness. Inputs for ML include genomic relationship matrices (GRM), principal components, and single-nucleotide polymorphism (SNP). The results revealed that the heritability for the three fertility traits ranged from 0.09 to 0.48. GP accuracy from imputed low-coverage sequencing data was comparable to that of standard SNP arrays. When both pedigree and genomic data were used, ssGBLUP achieved GP accuracy comparable to or higher than those of ML and Bayesian methods, while significantly outperforming GBLUP (P < 0.05) for AFS and GLh. Crucially, using only genomic data, KRR_GRM in AFC improved GP accuracy by up to 28.57% compared to GBLUP (P < 0.05) and showed comparable to or slightly higher performance than BayesR3. Our results highlight the effectiveness of low-coverage sequencing data in breeding applications and the ML's potential to enhance GP accuracy for fertility traits, offering practical insights for dairy breeding programmes.