Cost-Effective Implementation of Genomic Selection By Subsetting SNP Markers

Molecular markers covering the whole genome have been successfully used for prediction of genetic merit. However, the use of high-density SNP marker panels may not be feasible in most breeding programs due to costs. An alternative is the selection of a subset of informative markers for generating a cost-effective SNP panel. We evaluated the use of two approaches for such a task, which involves two steps: selecting and fitting. The selecting step consisted of grouping the p available SNPs such that linkage disequilibrium levels are higher between SNPs within groups and lower between SNPs from different groups. Within each group one of two approaches, backward elimination or stepwise regression based on least squares theory, was used to select SNPs. Lastly, SNPs selected from each group were combined, forming a final set of SNPs, and a Bayesian Lasso (BLasso) procedure was used to fit the final model. The procedures were applied to datasets from literature. We performed a 10-fold cross-validation to compare the predictive ability of the proposed approaches against other commom methods: selection of SNPs equally spaced and model fit using ridge regression (RR-Blup); and selection of SNPs with larger estimated effects and model fit using BLasso or using RR-Blup. Also, we compared the accuracy of the prediction from both approaches against the accuracy from the complete model (p SNPs) using BLasso. Both approaches perform better than the complete model in many cases and allowed for an optimized reduction on the final number of SNPs. (FAPEMIG, CNPq, CAPES)