Deep Learning for genomic prediction accounting for heterosis in crossbreeding systems
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Background
Crossbreeding is used in animal breeding to combine desirable traits from different breeds and to exploit hybrid vigor, and many approaches have been proposed to improve the prediction of genetic values of crossbred animals. This study aimed to assess whether deep learning methods can enhance the prediction of genetic values in crossbred populations by effectively capturing heterosis. We tested several prediction models that varied in how they incorporate genetic information, including purebred and crossbred data, breed composition, and breed-of-origin of alleles, on both simulated and real crossbred data. Finally, we also proposed an approach in which additive genetic effects and heterosis are predicted separately and then combined using effect-specific weights.
Results
Our results show that the presence of heterosis reduced the accuracy of traditional statistical methods but improved the performance of deep learning models. When heterosis was present, the highest prediction accuracy for total genetic value was achieved by combining additive genetic values, predicted using statistical method, with the predicted heterosis values, using a weighted sum. The explicit prediction of heterosis using deep learning and a regression method yielded similar results. The integration of breed-of-origin of alleles achieved the highest prediction accuracy for the additive genetic effect, outperforming both methods based on breed composition, and those that train the GBLUP model by combining purebred and crossbred animals. When applied to real data, adding the predicted heterosis did not increase prediction accuracies of crossbred animals, a result that we attributed to the homogeneity of the real crossbred population and its limited sample size.
Conclusions
The usefulness of deep learning for crossbred prediction depends on the degree of heterosis exhibited in the traits. Incorporating both additive genetic effects and heterosis considering the breed-of-origin of the alleles in the genomic information and effect-specific weighting leads to more accurate and robust predictions in our simulated data, particularly when crossbred data is heterogeneous. Even when prediction accuracy is not significantly improved, deep learning can still provide valuable insight into the degree of heterosis across traits, offering a deeper understanding of crossbred genetic architecture.
