A comparative analysis of machine learning algorithms for genomic prediction across fish populations

Genomic prediction can shorten selection cycles and accelerate individual breeding processes. Machine learning methods are widely employed for genomic prediction, but existing studies have not yet clarified which methods are generally applicable to fish genomic prediction. This study aims to conduct a systematic and comprehensive comparison of commonly used machine learning methods and GBLUP, explore their performance in predicting categorical and continuous traits in fish from multiple aspects including prediction accuracy and stability, and ultimately find the machine learning methods suitable for fish genomic prediction, thus providing some guidance for molecular breeding. Using eight fish datasets reported in the literature (covering four categorical traits and fifteen continuous traits), this study identifies the most suitable fish genomic prediction methods by comparing their predictive performance. In addition to conventional evaluation metrics, this study also used two approaches for algorithm performance assessment: one is the Legarra-Reverter (LR) validation method, and the other is the Wilcoxon signed-rank test. The results show that, among the regression features, Kernel Ridge Regression method ranked first in terms of average accuracy and demonstrated statistical significance compared to the other 8 methods in statistical tests, as well as strong robustness. Similarly, the Multilayer Perceptron also ranked first in terms of accuracy among the classification features. Thus, this study provided specific recommendations for predicting fish phenotypic values: use KRR for continuous traits, and the MLP for categorical traits. The study results offer important references and practical guidance for research related to fish genomic prediction.