Genomic selection is defined as using the molecular marker information that covered the whole genome to estimate individual’s breeding values. Using genome information can avoid many problems caused by pedigree errors so as to improve selection accuracy and shorten breeding generation intervals. According to different statistical models, methods of estimated genomic breeding value (GEBV) can be divided into based on BLUP (best linear unbiased prediction) theory, based on Bayesian theory and others. At present, GBLUP and its improved method ssGBLUP have been widely employed. Accuracy is the most used evaluation metric for genomic selection models, which is to evaluate the similarity between the true value and the estimated value. The factors that affect the accuracy can be reflected from the model, which can be divided into controllable factors and uncontrollable factors. Traditional genomic selection methods have promoted the rapid development of animal breeding, but these methods are currently facing many challenges such as multi-population, multi-omics, and computing. What’s more, they cannot capture the nonlinear relationship between high-dimensional genomic data. As a branch of artificial intelligence, machine learning is very close to biological mastery of natural language processing. Machine learning extracts features from data and automatically summarizes the rules and use to make predictions for new data. For genomic information, machine learning does not require distribution assumptions, and all marker information can be considered in the model. Compared with traditional genomic selection methods, machine learning can more easily capture complex relationships between genotypes, phenotypes, and the environment. Therefore, machine learning has certain advantages in animal genomic selection. According to the amount and type of supervision received during training, machine learning can be classified into supervised learning, unsupervised learning, semi-supervised learning, and reinforcement learning. The main difference is whether the input data is labeled. The machine learning methods currently applied in animal genomic selection are all supervised learning. Supervised learning can handle both classification and regression problems, requiring the algorithm to be provided with labeled data and the desired output. In recent years, the application of machine learning in animal genomic selection has been increasing, especially in dairy and beef cattle. In this review, machine learning algorithms are divided into three categories: single algorithm, ensemble algorithm and deep learning, and their research progress in animal genomic selection were summarized. The most used single algorithms are KRR and SVR, both of which use kernel tricks to learn nonlinear functions and map data to higher-dimensional kernel spaces in the original space. Currently commonly used kernel functions are linear kernel, cosine kernel, Gaussian kernel, and polynomial kernel. Deep learning, also known as a deep neural network, consists of multiple layers of connected neurons. An ensemble learning algorithm refers to fusing different learners together to obtain a stronger supervised model. In the past decade, the related literature on machine learning and deep learning has shown exponential growth. And its application in genomic selection is also gradually increasing. Although machine learning has obvious advantages in some aspects, it still faces many challenges in estimating the genetic breeding value of complex traits in animals. The interpretability of some models is low, which is not conducive to the adjustment of data, parameters, and features. Data heterogeneity, sparsity, and outliers can also cause data noise for machine learning. There are also problems such as overfitting, large marks and small samples, and parameter adjustment. Therefore, each step needs to be handled carefully while training the model. This paper introduced the traditional methods of genomic selection and the problems they face, the concept and classification of machine learning. We discussed the research progress and current challenges of machine learning in animal genomic selection. A Case and some application suggestions were given to provide a certain reference for the application of machine learning in animal genomic selection.

Resilience traits in pig populations allow animals to deal better with infectious disease and suboptimal production environments. The data on daily weight, feed intake and feed behaviors in pigs are collected in test period by automated feeding stations, which facilitate to evaluate the resilience traits. In this study, we adopted the root mean square error (RMSE) of ordinary least squares (OLS) and the negative residuals of quantile regression (QR) to generate four different novel resilience traits using daily records of feed intake and feed duration between 90 and 180 days of age in a population of commercial Duroc pigs. The genome-wide association studies (GWAS) based on single- and two-trait mixed models were carried out on 550 pigs using 48,603 single nucleotide polymorphisms (SNPs) to identify genomic regions associated with resilience traits in growing pigs. We further focused on the GWAS signals to conduct gene annotation, colocalization with multi-tissue eQTL summary statistics of PigGTEx project and identification of enhancers and promoters using the publicly available data. The genomic heritabilities of four novel resilience traits ranged from 0.09 to 0.41. The pairwise genetic and phenotypic correlations ranged from 0.16 to 0.95 and from 0.05 to 0.36, respectively. Twenty-seven SNPs were identified to be significantly associated with these resilience traits. They were distributed on nine chromosomes (SSC1, SSC2, SSC6, SSC7, SSC8, SSC12, SSC14, SSC16 and SSC17). After annotation, 39 QTLs and 49 candidate genes were identified. Several of these are functionally relevant candidate genes including OTUD4, TIFA and CARD14, which are involved in the host immune response, disease susceptibility and signal transduction. Eight unique SNPs were found to be causal in both GWAS and eQTL analyses across 15 tissues. Notably, one SNP (rs80794541) was associated with eQTLs identified concurrently across seven tissues/cell types, including the macrophage cell type. Furthermore, four significant SNPs (rs81467127, rs81356029, rs80794541 and rs81305085) were linked to the function of the primed enhancer, active element, and poised promoter in five pig tissues. Using the porcine fibroblast HiC dataset, SNP (rs81356029) on SSC2 regulates the CARNS1 and SSH3, while SNP (rs80794541) on SSC7 regulates the H2AC6. In this study, we generated four novel resilience traits and identified SNPs significantly associated with these resilience traits in a Duroc pig population. GWAS signals were associated with candidate genes involving in the immune traits, and were linked to the crucial regulatory elements as well. Our findings will contribute to elucidating the genetic mechanism that can enhance genome-enabled breeding and inform further research on resilience in domestic pigs.