Mota Lucio F M, Arikawa Leonardo M, Valente Júlia P S, Fonseca Larissa F S, Mercadante Maria E Z, Cyrillo Joslaine N S G, Oliveira Henrique N, Albuquerque Lucia G
Department of Animal Science, School of Agricultural and Veterinary Sciences, São Paulo State University (UNESP), Jaboticabal, 14884-900, SP, Brazil.
Institute of Animal Science, Beef Cattle Research Center, Sertãozinho, SP, 14174-000, Brazil.
Sci Rep. 2025 Jul 1;15(1):22266. doi: 10.1038/s41598-025-06949-z.
Genomic selection (GS) has become a widely used tool in breeding programs, enhancing selection accuracy and leading to faster genetic progress. However, in small populations, GS faces challenges due to limited data and a large number of markers potentially leading to biased predictions. Implementing feature selection strategies is essential to improve prediction accuracy and avoid overfitting. Hence, we compared the predictive ability of genomic best linear unbiased prediction (GBLUP), Bayesian B (BayesB), and elastic net (ENet) models, using all markers and feature selection via GWAS and fixation index (FST) to reduce marker numbers, for growth and ultrasound carcass traits in three Nellore cattle populations differentially selected for yearling body weight (YBW). The populations evaluated included: Nellore Control (NeC), selected for YBW; Nellore Selection (NeS), selected for maximum YBW; and Nellore Traditional (NeT), selected for maximum YBW and lower residual feed intake (RFI) since 2013. Comparing the statistical approaches using GBLUP as the reference, ENet improved prediction accuracy by 10% for growth traits and 12% for carcass traits, while BayesB showed no improvement for growth traits but achieved a 3% gain for carcass traits. When comparing models using all markers to those with variable selection, both GWAS and FST improved prediction accuracy across models, with FST outperforming GWAS in stratified populations. A stricter GWAS threshold (> 1.0% explained variance), compared to a less conservative criterion (> 0.5%), reduced BayesB prediction accuracy (6.8%), while slightly increasing accuracy for GBLUP (1.3%) and ENet (2.4%). Similarly, a more restrictive FST threshold (> 0.2) against a less conservative (> 0.1) resulted in smaller gains for GBLUP (4%) and ENet (5%), but reduced BayesB accuracy (- 4%). Overall, selecting markers through GWAS and FST improves prediction accuracy for both growth and carcass traits, particularly in stratified populations. However, stricter thresholds can negatively impact accuracy, highlighting the need for optimized marker selection strategies.
基因组选择（GS）已成为育种计划中广泛使用的工具，提高了选择准确性并带来更快的遗传进展。然而，在小群体中，由于数据有限以及大量标记可能导致预测偏差，GS面临挑战。实施特征选择策略对于提高预测准确性和避免过拟合至关重要。因此，我们比较了基因组最佳线性无偏预测（GBLUP）、贝叶斯B（BayesB）和弹性网络（ENet）模型的预测能力，使用所有标记以及通过全基因组关联研究（GWAS）和固定指数（FST）进行特征选择以减少标记数量，针对三个因一岁体重（YBW）而进行不同选择的内洛尔牛群体的生长和超声胴体性状进行分析。评估的群体包括：内洛尔对照（NeC），按YBW选择；内洛尔选育（NeS），按最大YBW选择；以及内洛尔传统群体（NeT），自2013年以来按最大YBW和较低的剩余采食量（RFI）选择。以GBLUP作为参考比较统计方法，ENet对生长性状的预测准确性提高了10%，对胴体性状提高了12%，而BayesB对生长性状没有提高，但对胴体性状提高了3%。当比较使用所有标记的模型与进行变量选择的模型时，GWAS和FST均提高了各模型的预测准确性，在分层群体中FST的表现优于GWAS。与较宽松的标准（>0.5%）相比，更严格的GWAS阈值（>1.0%的解释方差）降低了BayesB的预测准确性（6.8%），而GBLUP（1.3%）和ENet（2.4%）的准确性略有提高。同样，与较宽松的FST阈值（>0.1）相比，更严格的阈值（>0.2）使GBLUP（4%）和ENet（5%）的增益较小，但降低了BayesB的准确性（-4%）。总体而言，通过GWAS和FST选择标记可提高生长和胴体性状的预测准确性，特别是在分层群体中。然而，更严格的阈值可能对准确性产生负面影响，凸显了优化标记选择策略的必要性。
