Incorporating information of causal variants in genomic prediction using GBLUP or machine learning models in a...

Genomic prediction has revolutionized animal breeding, with GBLUP being the most widely used prediction model. In theory, the accuracy of genomic prediction could be improved by incorporating information from QTL. This strategy could be especially beneficial for machine learning models that are able to distinguish informative from uninformative features. The objective of this study was to assess the benefit of incorporating QTL genotypes in GBLUP and machine learning models. This study simulated a selected livestock population where QTL and their effects were known. We used four genomic prediction models, GBLUP, (weighted) 2GBLUP, random forest (RF), and support vector regression (SVR) to predict breeding values of young animals, and considered different scenarios that varied in the proportion of genetic variance explained by the included QTL.

2GBLUP resulted in the highest accuracy. Its accuracy increased when the included QTL explained up to 80% of the genetic variance, after which the accuracy dropped. With a weighted 2GBLUP model, the accuracy always increased when more QTL were included. Prediction accuracy of GBLUP was consistently higher than SVR, and the accuracy for both models slightly increased with more QTL information included. The RF model resulted in the lowest prediction accuracy, and did not improve by including QTL information.

Our results show that incorporating QTL information in GBLUP and SVR can improve prediction accuracy, but the extent of improvement varies across models. RF had a much lower prediction accuracy than the other models and did not show improvements when QTL information was added. Two possible reasons for this result are that the data structure in our data does not allow RF to fully realize its potential and that RF is not designed well for this particular prediction problem. Our study highlighted the importance of selecting appropriate models for genomic prediction and underscored the potential limitations of machine learning models when applied to genomic prediction in livestock.