322. Finetuning hyper-parameters increases the prediction accuracy in single-step genetic evaluation

Abstract

Single step genomic best linear unbiased prediction (HBLUP) has been widely used in livestock breeding. The HBLUP method (e.g. BLUPf90) requires hyper-parameters to combine genomic and pedigree relationships and these should be adequately initialised to maximise the accuracy of genomic prediction. In this study, we assess the performance of HBLUP, using various values of hyper-parameters in simulated genomic data. We show that the tuning parameter (tuning GRM relative to the pedigree-based numerator relationship matrix) considerably increases prediction accuracy, confirming previous studies. The scale factor, α, which scales the allele effect size by its frequency, also affects accuracy and the optimal scale factor can vary for each trait. In conclusion, fine-tuning the hyper-parameters of HBLUP is necessary to maximize prediction accuracy and the scale factor should be considered.