The accuracy of different strategies for building training sets for genomic predictions in segregating soybean populations

Abstract

The design of the training set is a key factor in the success of the genomic selection approach. The nature of line inclusion in soybean [Sorghum bicolor (L.) Moench.] breeding programs is highly dynamic, so generating a training set that endures across the years and regions is challenging. Therefore, we aimed to define the best strategies for building training sets to apply genomic selection in segregating soybean populations for traits with different genetic architectures. We used two datasets for grain yield (GY) and maturity group (MG) from two different soybean breeding regions in Brazil. Five training set schemes were tested. In addition, we included a training set formed by an optimization algorithm based on the predicted error variance. The predictions achieved good values for both traits, reaching 0.5 in some scenarios. The best scenario changed according to the trait. Although the best performance was achieved with the use of full‐sibs in the MG dataset, for GY, full‐sibs and a set of advanced lines were equivalent. For both traits, no improvement in predictive ability resulted from training set optimization. Furthermore, the use of advanced lines from the same breeding program is recommended as a training set for GY, so the training set is continually renewed and closely related to the breeding populations, and no additional phenotyping is needed. On the other hand, to improve prediction accuracies for MG, it is necessary to use training sets with less genetic variability but with more segregation resolution.

Enviromic-based Kernels Optimize Resource Allocation with Multi-trait Multi-environment Genomic Prediction for Tropical Maize

“…As well, adding the interaction effect G × W (M5) increases PA when compared to the main effects models (M1 and M3), but not as much as models containing the G × E, similar to what was already presented by Robert et al (2020) and Jarquin et al (2021); however, G × W has the advantage of predicting new environments (Costa-Neto et al, 2021a;Jarquin et al, 2021;Robert et al, 2020), which has not been tested here. Also, MTMET models gave a small increase in PA compared with STMET, especially for models containing the G × E term (Lyra et al, 2017;Mendonça & Fritsche-Neto, 2020). Furthermore, as Costa-Neto et al (2021a) reported, including the W matrix helps increase PA, especially in the case of untested hybrids and/or untested environments, by better explaining sources of variation, capturing the environment potential per se and its interaction with the genotypes.…”

“…Thus, LA-GA-T optimizes the selection, on a genetic basis, of the n genotypes informed by APY to compose our optimized training set (OTS) (Fristche-Neto et al, 2018). In this context, Mendonça & Fritsche-Neto (2020) used the algorithm designed by Akdemir (2017) to select the most representative genotypes to build the training population. Similar to our findings, they did not notice an increase in PA while using OTS, but reduced the budget.…”

Enviromic-based kernels may optimize resource allocation with multi-trait multi-environment genomic prediction for tropical Maize

“…Nevertheless, the W matrix can optimize complex information, as we saw in our OTS’ scenarios, and optimize trials [ 25 ]. Also, MTMET models showed a small PA increase compared with STMET, especially for models containing the G × E term [ 33 , 35 ].…”

“…Thus, LA-GA-T optimizes the selection, on a genetic basis, of the n genotypes informed by APY to compose our optimized training set (OTS) [ 18 ]. In this context, Mendonça and Fritsche-Neto [ 35 ] used the algorithm designed by Akdemir [ 2 ] to select the most representative genotypes to build a training population. Similar to our findings, they did not notice an increase in PA while using OTS but were able to reduce the budget.…”

Enviromic-based kernels optimize resource allocation with multi-trait multi-environment genomic prediction for tropical maize

“…As well, add the interaction effect G × W (M5) increases PA when compared to the main effects models (M1 and M3), but not as much as models containing the G × E, similar to what was already presented by Robert et al (2020) and (Jarquin et al, 2021), however, G × W has the advantage of predicting new environments Jarquin et al, 2021;Robert et al, 2020), that has not been tested here. Also, MTMET models gave a small increase in PA compared with STMET, especially for models containing the G × E term (Lyra et al, 2017;Mendonça & Fritsche-Neto, 2020). Furthermore, as Costa-Neto et al (2021a) reported, including the W matrix helps increase PA, especially in the case of untested hybrids and/or untested environments, by better explaining sources of variation, capturing the environment potential per se and its interaction with the genotypes.…”