A Unified Comparative Evaluation of Genomic Prediction Models Across Four Aquaculture Species

Genomic prediction has been increasingly applied in aquaculture selective breeding; however, systematic evaluations of prediction accuracy across multiple aquaculture species and analytical methods under a unified and comparable framework remain limited. In this study, we conducted a comprehensive comparative assessment of genomic prediction performance across four representative aquaculture species, including Atlantic salmon (Salmo salar), gilthead sea bream (Sparus aurata), common carp (Cyprinus carpio), and rainbow trout (Oncorhynchus mykiss), using ten genomic prediction models including GBLUP, Bayesian and machine learning methods. Prediction accuracy varied widely among species and models, ranging from 0.49 to 0.85, and was strongly associated with trait heritability. High-heritability traits consistently achieved higher prediction accuracies, with rainbow trout and common carp exhibiting the best overall performance (0.75–0.83 and 0.73–0.85, respectively), whereas Atlantic salmon and gilthead sea bream showed lower and more variable accuracies (0.49–0.61 and 0.49–0.66). No single model performed optimally across all species. Machine learning-based approaches achieved the highest prediction accuracy in specific cases but exhibited pronounced species-dependent variability, while GBLUP provided stable and well-calibrated predictions with consistently low bias. Incremental SNP feature selection further improved prediction accuracy by 2.8–4.2% in three species using only 0.54–9.64% of the available markers, whereas no improvement was observed for a low-heritability trait. These results show that genomic prediction performance is highly context-dependent and underscores the importance of jointly considering trait genetic architecture, population characteristics, model choice, and marker selection when optimizing genomic selection strategies in aquaculture breeding programs.