Evaluation of Models for Utilization in Genomic Prediction of Agronomic Traits in the Louisiana Sugarcane Breeding Program

Sugarcane (Saccharum spp.) is an important perennial grass crop for both sugar and biofuel industries. The Louisiana sugarcane breeding program is focused on improving sugar yield by incrementally increasing genetic gain. With the advancement in genotyping and (highthroughput) phenotyping techniques, genomic selection is a promising marker-assisted breeding tool. In this study, we assessed ridge regression best linear unbiased prediction (rrBLUP) and various Bayesian models to evaluate genomic prediction accuracy using a 10-fold cross validation on 95 commercial and elite parental clones from the Louisiana sugarcane breeding program. Datasets (individual and pooled in various combinations) were constructed based on soil type (light-Commerce silty loam, heavy-Sharkey clay) and crop (plant cane, ratoon). A total of 3906 SNPs were used to predict the genomic estimated breeding values (GEBVs) of the clones for sucrose content and cane and sugar yield. Prediction accuracy was estimated by both Spearman's rank correlation and Pearson's correlation between phenotypic breeding values and GEBVs. All traits showed significant variation with moderate (42% for sucrose content) to high (85% for cane and sugar yield) heritability. Prediction accuracy based on rank correlation was high (0.47-0.80 for sucrose content; 0.61-0.69 for cane yield, and 0.56-0.72 for sugar yield) in all cross-effect prediction models where soil and crop types were considered as fixed effects. In general, Bayesian models demonstrated a higher correlation than rrBLUP. The Pearson's correlation without soil and crop type as fixed effects was lower with no clear pattern among the models. The results demonstrate the potential implementation of genomic prediction in the Louisiana sugarcane variety development program.