Estimation & Prediction

Integrative SNP weighting and machine learning approach enhance GEBV prediction accuracy for flavor compounds in chickens

---

Abstract

Improving the accuracy of genomic estimated breeding values (GEBVs) is essential for accelerating genetic progress in poultry breeding, particularly for complex traits influenced by numerous minor-effect loci such as flavor compounds. This study developed an integrative genomic prediction framework that combines statistical models, machine learning algorithms, and biologically informed SNP-weighting strategies to enhance prediction accuracy for nine flavor compounds in Korean native chickens (KNCs). The dataset consisted of 892 individuals genotyped with the Illumina 60K SNP BeadChip and phenotyped for five free amino acids (alanine, aspartic acid, glutamic acid, glycine, and valine), one nucleotide (IMP), and three fatty acids (oleic acid, linoleic acid, and arachidonic acid). All phenotypes were pre-adjusted for sex, data generation year, and carcass weight. To incorporate biological relevance into GEBV prediction, three SNP-weighting scenarios were designed. Scenario 1 served as an unweighted baseline. Scenario 2 utilized SNP effects derived from a conventional genome-wide association study (GWAS), where squared SNP effects were aggregated in 20-SNP windows to emphasize regions with stronger statistical signals. Scenario 3 extended this approach by integrating feature importance scores obtained from a machine learning-based GWAS (ML-GWAS), generating hybrid weights that captured both statistically detectable effects and ML-identified nonlinear signals. Each scenario was implemented in two statistical models (GBLUP, RKHS) using weighted GRMs or kernels, and four machine learning models (Random Forest, Support Vector Regression, XGBoost, ElasticNet) using weighted PCA-transformed genotypes (PC1-PC50). Model performance was assessed via repeated 5-fold cross-validation using accuracy (r/h), RMSE, MAE, and regression calibration parameters (α, β). Among statistical models, RKHS showed the strongest and most consistent performance, reflecting its ability to capture nonlinear genomic relationships. Machine learning models, particularly RF and SVR, demonstrated competitive or superior performance for several traits, often achieving β values close to 1, indicating minimal prediction bias. Scenario 2 produced modest improvements for traits with clear GWAS signals, whereas Scenario 3 consistently delivered the highest accuracies across most traits. Notable gains under Scenario 3 were observed in glycine, IMP, and linoleic acid, where ML-GWAS successfully identified subtle but meaningful SNP contributions that conventional GWAS did not identify. This study demonstrates that integrating weighted genomic information with machine learning substantially enhances the precision and robustness of GEBV prediction for flavor compounds in KNCs. The hybrid weighting framework provides a biologically informed approach to improving genomic prediction and supports the broader application of precision breeding strategies in poultry. Future research should incorporate larger populations and functional validation of key SNPs and genes to further strengthen the biological interpretability of ML-based genomic prediction.

Keywords: 2026

How to Cite:

Cho, E., Kim, M., Choo, H. & Lee, J., (2026) “Integrative SNP weighting and machine learning approach enhance GEBV prediction accuracy for flavor compounds in chickens”, World Congress on Genetics Applied to Livestock Production Digital Archive 2026(1): 2286346. doi: https://doi.org/10.31274/wcgalp.23943

Rights: 1