Training optimization in genomic selection: A comparison of algorithms

Abstract

Improved crop breeding will be essential to match the needs of the growing human population in this century. Genomic selection (GS) is a very promising tool to address some aspects of this challenge, because it can improve the response from selection in plant breeding programs mainly by reducing the generation interval (time) to release a cultivar. One of the key aspects in GS is how to select the individuals that will be used to build the training set (TRS). Several TRS optimization algorithms have been described in the literature, but a study that focuses on comparing the different algorithms under the same conditions is lacking. Here, we have used four different datasets from three different species to compare the effect of TRS optimization algorithms (Coefficient of Determination, CDmean; Prediction Error Variance, PEVmean; Upper bound of reliability, Umean; Partitioning Around Medoids, PAM) against the random sampling approach under four different statistical models (additive vs. non-additive). We performed the optimization in two scenarios; whether the breeder has the genotypic information from the test set (targeted) to build the TRS or not (untargeted). Our results indicated that targeted optimization performed always better than untargeted optimization. In general, all algorithms worked well under all circumstances, but PAM worked better for small TRS than for larger TRS specially with low or mild population structure. We did not find great differences in the performance of CDmean, PEVmean and Umean for untargeted optimization, but for targeted Umean was slightly worse than the rest. Our results also showed that non-additive models performed well under all circumstances. Finally, we described some general guidelines to efficiently use the TRS and increase the rate of genetic improvement in plant breeding programs.