The growing integration of artificial intelligence into aquaculture breeding has fuelled expectations that machine learning models will consistently outperform traditional approaches. A new comparative study published in Fishes suggests otherwise: the most complex algorithms do not necessarily deliver the highest predictive accuracy, and performance depends strongly on biological and production context.
The study provides a unified benchmarking of ten genomic prediction models – including GBLUP, Bayesian approaches and machine learning methods such as Random Forest, Support Vector Regression (SVR) and XGBoost – across four economically important aquaculture species: Atlantic salmon, gilthead seabream (Sparus aurata), common carp and rainbow trout.
The results show no universally superior model. Prediction accuracy varied substantially depending on species, trait and heritability. In rainbow trout, where heritability was high, predictive accuracy ranged from 0.75 to 0.83. In gilthead seabream, where disease resistance exhibited low heritability, accuracy dropped to between 0.49 and 0.66.
These findings reaffirm a fundamental principle of quantitative genetics: heritability remains one of the primary drivers of genomic selection performance. Algorithmic complexity alone cannot compensate for limited genetic signal.
Machine learning approaches achieved the highest accuracies in specific cases – for instance, Machine learning approaches achieved the highest accuracies in specific cases — for instance, SVR reached 0.853 in common carp — but their performance was highly species-dependent.
By contrast, GBLUP consistently delivered stable and well-calibrated predictions across all species, reinforcing its role as a robust baseline model in breeding programmes.
From a practical perspective, one of the most significant findings relates to marker optimisation. Through GWAS-based incremental feature selection, the authors improved prediction accuracy while using only a fraction of available SNPs: 9.64% in salmon, 4.58% in carp and just 0.54% in trout, achieving relative improvements of up to 4.2% compared with the full marker panel.
This has clear industry implications: greater predictive accuracy with fewer markers may reduce genotyping costs without compromising selection efficiency.
The study sends a strong message to European aquaculture breeding programmes. Model choice should be guided by trait genetic architecture, population structure and phenotypic data quality. Artificial intelligence can enhance prediction, but it does not replace biological understanding or strategic evaluation.
Aquaculture genetics is entering a phase of optimisation — where context, not complexity, determines success.
Related
EVENTS
-
21AprTue, 21 Apr - Fri, 24 AprGran Via - Fira Barcelona - Barcelona