The impact of genotype transformations and aneuploidy on genomic prediction accuracy in polyploid species: a simulation study

The accuracy of genomic prediction for key traits in sugarcane remains low relative to other crops, even with growing reference sets. This is potentially due to high-level polyploidy, frequent aneuploidy and complex genetic architectures. Here we aim to assess the impact of genotype transformations typically used in polyploid crops on the accuracy of genomic prediction. We focus on the effect of using “diploid” genotype calls and aneuploidy. We simulated a sugarcane genome with multi-species ancestry and genome duplications, resulting in 96 chromosomes. We included a range of allele dosage to phenotype maps across ploidies including additive linear, dominant, non-linear sigmoid and log shaped maps. Genomic prediction accuracies were compared from six algorithms including genomic BLUP (GBLUP), extended GBLUP, BayesR, Multilayer perceptron, Convolutional neural network and Attention network. Accuracy of prediction for phenotypes controlled by a higher proportion of additive effects were more sensitive to diploid transformation and aneuploidy events. Conversely, accuracy of prediction for phenotypes involving directional dominance were less affected by diploid transformation but were still substantially affected by aneuploidy. For traits controlled by non-linear allele dosage to phenotype maps, the predictive accuracy was significantly affected by diploid transformation, but the influence of aneuploidy was not obvious. Extended GBLUP performed well and gave higher accuracies than other linear based models across most of scenarios, while the DL models were more competitive when the allele action across ploidies became more complex. Overall, our results demonstrate that genotypes that capture allele dosage should be used in genomic prediction for complex polyploids.