Predicting chicken body weight from genotypes: Performance comparison of deep convolutional neural networks, artificial neural networks, ensemble models and RR-BLUP

Abstract

Genomic prediction is a cornerstone of poultry breeding, enabling accelerated improvement of complex traits. This study evaluated the predictive performance of machine learning models including a deep convolutional neural network (DeepGS), artificial neural network (ANN) and ensemble model combining DeepGS with ridge regression best linear unbiased prediction (RR-BLUP) as well as conventional statistical RR-BLUP model for genotype-to-phenotype prediction of body weight in chickens using genome-wide marker data. Predictive accuracy assessed via Pearson’s correlation coefficient (PCC) between observed and predicted phenotypes, was similarly high for DeepGS, RR-BLUP and ensemble model (PCC = 0.89) whereas ANN showed slightly lower correlation (PCC = 0.84). Analysis of absolute prediction errors revealed median deviations were comparable among DeepGS, RR-BLUP and ensemble model, with ANN exhibiting greater variability and occasional errors. Examination of top-ranked individuals revealed ANN model better preserved phenotypic ranking among elite birds (PCC = 0.52) compared with RR-BLUP (0.47), ensemble model (0.34) and DeepGS (0.32). Strong concordance among DeepGS, RR-BLUP and ensemble model predictions (PCCs 0.94–1.00) indicates that additive genetic effects dominate body weight variation. Results demonstrate that no single model uniformly outperforms others across all metrics, emphasizing the need to balance prediction accuracy, stability, and ranking performance based on breeding objectives.