Predicting for disease resistance in aquaculture species using machine learning models

Predicting disease resistance is one of the most prominent applications of aquaculture selective breeding. Reductions in genotyping costs have allowed the implementation of genomic selection in an abundance of aquaculture species and their related diseases showing promising results. Machine learning (ML) models can be of value for prediction purposes, as suggested by several studies in both plants and livestock. The current study aimed to test the efficiency of various ML models in predicting disease resistance using both simulated and real datasets. More specifically, models like decision trees (DT), support vector machines (SVM), random forests (RF), adaptive boosting (Adaboost) and extreme gradient boosting (XGB) were benchmarked against genomic best linear unbiased prediction for threshold traits backend by Markov chain Monte Carlo (GBLUP-MCMC) both in terms of prediction efficiency and required computational time. Moreover, the model ranking was tested in datasets where the ratio between the two observed phenotypes (resistant vs non-resistant) was unbalanced. Across all tested datasets, XGB ranked first with a slight advantage over GBLUP-MCMC, ranging between 1-4 %. SVM and RF delivered predictions in tight proximity with the ones from XGB and GBLUP-MCMC. In addition, predictions 3-4 % lower compared to GBLUP-MCMC were obtained with Adaboost. On the other hand, the predictions from DT were consistently low (-40 % lower compared to GBLUP-MCMC). All tested ML models had significantly reduced computational requirements than GBLUP-MCMC. In the case of XGB, the computational requirements were reduced more than 20-fold as opposed to GBLUP-MCMC under the settings of the current study. RF delivered both competitive predictions and was highly efficient in terms of the required computational time (-3 min). Overall, the results of the current study suggest that ML models can be valuable tools in aquaculture breeding studies for disease resistance.

Automated summary

The study tested the efficiency of various machine learning models in predicting disease resistance and found that XGB performed the best in most datasets, with a slight advantage over GBLUP-MCMC. SVM and RF models also provided predictions close to XGB and GBLUP-MCMC. Adaboost yielded slightly lower predictions, while DT consistently performed poorly in predictions compared to GBLUP-MCMC.