Predictive assessment of single-step BLUP with linear and non-linear similarity RKHS kernels: A case study in...

Area of Science:

• Animal Breeding and Genetics
• Quantitative Genetics
• Genomic Selection

Background:

• Single-step genomic best linear unbiased prediction (ssGBLUP) is a standard method for genomic prediction, primarily using additive genetic effects.
• The genomic relationship matrix (G) is a linear kernel that limits the capture of complex genetic interactions.
• Non-additive genetic effects can significantly influence complex traits, necessitating more sophisticated modeling approaches.

Purpose of the Study:

• To generalize ssGBLUP by incorporating non-linear kernels, specifically the averaged Gaussian kernel (AK) and the arc-cosine deep kernel (DK).
• To evaluate the performance of these generalized ssGBLUP models against the standard G kernel for body weight (BW) and hen-housing production (HHP) traits in commercial broiler chickens.
• To investigate the impact of varying genotyping rates and selective genotyping strategies on prediction accuracy.

Main Methods:

• Developed and applied ssGBLUP models using three kernels: the standard genomic relationship matrix (G), averaged Gaussian kernel (AK), and arc-cosine deep kernel (DK).
• Evaluated models on a dataset of phenotyped and genotyped broiler chickens for body weight and hen-housing production traits.
• Utilized random replication of training and testing sets across different genotyping rates (20-80%) and selective genotyping scenarios (youngest individuals, random, parent average).

Main Results:

• Prediction accuracy was influenced by kernel type, with non-linear kernels (AK and DK) showing advantages at higher genotyping rates (60-80%).
• The arc-cosine deep kernel (DK) achieved the highest rank correlations for body weight (0.320 ± 0.016) under parent average selection with 80% genotyping.
• The averaged Gaussian kernel (AK) yielded the highest rank correlations for hen-housing production (0.23 ± 0.016) under parent average selection with 80% genotyping.

Conclusions:

• Generalized ssGBLUP models with non-linear kernels (AK and DK) are more effective than the standard G kernel when a substantial proportion of the population is genotyped.
• These advanced kernels can potentially improve genomic prediction accuracy for complex traits influenced by non-additive genetic effects.
• The choice of kernel and genotyping strategy significantly impacts the performance of genomic selection in broiler chickens.