Genomic Selection at Preliminary Yield Trial Stage: Training Population Design to Predict Untested Lines

Genomic selection (GS) is being applied routinely in wheat breeding programs. For the evaluation of preliminary lines, this tool is becoming important because preliminary lines are generally evaluated in few environments with no replications due to the minimal amount of seed available to the breeder. A total of 816 breeding lines belonging to advanced or preliminary yield trials were included in the study. We designed different training populations (TP) to predict lines in preliminary yield trials (PYT) consisting of: (i) advanced lines of the breeding program; (ii) 50% of the preliminary lines set belonging to many families; (iii) only full sibs, consisting of 50% of lines of each family. Results showed that the strategy of splitting the preliminary set in half, phenotyping only half of the lines to serve as the TP showed the most consistent results for the different traits. For a subset of the population of lines, we observed accuracies ranging from 0.49-0.65 for yield, 0.59-0.61 for test weight, 0.70-0.72 for heading date, and 0.49-0.50 for height. Accuracies decreased with the other training population designs, and were inconsistent across preliminary line sets and traits. From a breeder's perspective, a prediction accuracy of 0.65 meant, at 0.2 selection intensity, 75% of the best yielding lines based on phenotypic information were correctly selected by the GS model. Our results demonstrate that, despite the small family size, an approach that includes lines from the same family in both the TP and VP, together with half sibs and more distant lines, and only phenotyping the lines included in the TP, could be a useful, efficient design for establishing a GS scheme to predict lines entering first year yield trials.