Improvement of Genomic Predictions in Small Breeds by Construction of Genomic Relationship Matrix Through Variable Selection

Abstract

Genomic selection has been increasingly implemented in the animal breeding industry, and it is becoming a routine method in many livestock breeding contexts. However, its use is still limited in several small-population local breeds, which are, nonetheless, an important source of genetic variability of great economic value. A major roadblock for their genomic selection is accuracy when population size is limited: to improve breeding value accuracy, variable selection models that assume heterogenous variance have been proposed over the last few years. However, while these models might outperform traditional and genomic predictions in terms of accuracy, they also carry a proportional increase of breeding value bias and dispersion. These mutual increases are especially striking when genomic selection is performed with a low number of phenotypes and high shrinkage value—which is precisely the situation that happens with small local breeds. In our study, we tested several alternative methods to improve the accuracy of genomic selection in a small population. First, we investigated the impact of using only a subset of informative markers regarding prediction accuracy, bias, and dispersion. We used different algorithms to select them, such as recursive feature eliminations, penalized regression, and XGBoost. We compared our results with the predictions of pedigree-based BLUP, single-step genomic BLUP, and weighted single-step genomic BLUP in different simulated populations obtained by combining various parameters in terms of number of QTLs and effective population size. We also investigated these approaches on a real data set belonging to the small local Rendena breed. Our results show that the accuracy of GBLUP in small-sized populations increased when performed with SNPs selected via variable selection methods both in simulated and real data sets. In addition, the use of variable selection models—especially those using XGBoost—in our real data set did not impact bias and the dispersion of estimated breeding values. We have discussed possible explanations for our results and how our study can help estimate breeding values for future genomic selection in small breeds.

Genomic prediction of blood biomarkers of metabolic disorders in Holstein cattle using parametric and nonparametric models

“…Feature selection and ML learners have the power to deal with heterogeneous and large datasets, providing prediction accuracies and detecting genomic regions that impact the relationships between genotype and phenotype [ 30 , 31 ]. Efficient feature selection from GBM enables the preselection of SNPs with biological relevance to the target trait.…”

“…Several studies have noted that ML, such as random forest (RF), support vector regression (SVR), gradient boosting machine (GBM), artificial neural networks (ANN), and stacking ensemble learning, can be used for GP, and outperform parametric models (i.e., GBLUP and Bayesian regression) in situations where the reference population is small [ 26 – 29 ]. On the one hand, ML techniques have also been explored as a potential strategy for variable selection and then used in genomic BLUP (GBLUP) [ 30 , 31 ]. On the other hand, Azodi et al [ 32 ], with data on plants, and Bellot et al [ 33 ], with human data, found that some ML models (e.g., deep neural networks) had a lower stability (i.e., show variation in prediction accuracy when trained on different validation designs or populations) than linear models.…”

Combining genetic markers, on-farm information and infrared data for the in-line prediction of blood biomarkers of metabolic disorders in Holstein cattle

“…Previous studies adopted different strategies to preselect predictors by directly excluding uninformative markers via machine learning [ 45 – 47 ] or assigning weights to markers according to their contributions to trait variability [ 48 ]. Piles et al [ 47 ] and Li et al [ 49 ] showed that feature selection strategies improved the predictive ability of complex traits.…”

“…S 5 –S 9 ). Fragomeni et al [ 50 ] and Mancin et al [ 45 ] highlighted the advantages of removing non-informative SNP, where better accuracy was achieved by constructing the G matrix by considering the window region where the QTL was identified or by using only QTL information.…”