Comparing machine learning algorithms and linear model for detecting significant SNPs for genomic evaluation of growth traits in F۲ chickens

High-density Single Nucleotide Polymorphisms (SNPs) panels are expensive, especially in developing countries. However, methods have been developed to detect critical SNPs from these panels and design low-density chips for genomic evaluation at lower cost. This study aimed to determine the efficiency of Random Forest (RF) and Gradient Boosting Machine (GBM) algorithms, and Linear Model (LM) in identification of SNPs subsets to predict Genomic Estimated Breeding Values (GEBVs) for Body Weights at ۶ (BW۶) and ۹ (BW۹) weeks in broiler chickens and compare the predicted GEBVs with those obtained by the ۶۰K SNP panel. The data were collected on ۳۱۲ F۲ chickens that genotyped with ۶۰K Illumina SNP BeadChip. After applying quality control, the remaining ۴۵,۵۱۲ SNPs were ranked based on p-values, mean square error percentage, and relative influence, obtained by LM, RF and GBM methods, respectively. Then, subsets of top ۴۰۰, ۱,۰۰۰, ۳,۰۰۰ and ۵,۰۰۰ SNPs, selected by each method, were employed to construct genomic relationship matrices for the prediction of GEBVs with genomic best linear unbiased prediction model. Results indicated that predicted accuracies by RF and GBM were generally higher than LM. A Subset of ۱,۰۰۰ SNPs selected by RF and GBM algorithms compared to the total SNPs increased accuracy from ۰.۳۸ to ۰.۶۴ and ۰.۶۶ for BW۶, and from ۰.۴۲ to ۰.۶۰ and ۰.۶۶ for BW۹, respectively. The findings of the present study provide that machine learning methods, especially GBM, can perform better than LM in selecting important SNPs and increasing the accuracy of genomic prediction in broiler chickens.

کلیدواژه ها:

genomic evaluation ، body weight ، broilers ، machine learning