Factors Affecting Response to Recurrent Genomic Selection in Soybeans

Abstract

1Herein we report the impacts of applying five selection methods across 40 cycles of recurrent 2 selection and identify interactions with other factors on genetic response using simulated families 3 of recombinant inbred lines derived from 21 homozygous soybean lines used for the Soybean 4 Nested Association Mapping study. The other factors we investigated included the number of 5 quantitative trait loci, broad sense heritability on an entry mean basis, selection intensity, and 6 training sets. Both the rates of genetic improvement in the early cycles and limits to genetic 7 improvement in the later cycles are affected by interactions among the factors. All genomic 8 selection methods provided greater rates of genetic improvement (per cycle) than phenotypic 9 selection, but phenotypic selection provided the greatest long term responses. Model updating 10 significantly improved prediction accuracy and genetic response for three parametric genomic 11 prediction models. Ridge Regression, if updated with training sets consisting of data from prior 12 cycles, achieved greater rates of response relative to BayesB and Bayes LASSO GP models. A 13 Support Vector Machine method, with a radial basis kernel, resulted in lowest prediction 14 accuracies and the least long term genetic response. Application of genomic selection in a closed 15 breeding population of a self-pollinated crop such as soybean will need to consider the impact of 16 these factors on trade-offs between short term gains and conserving useful genetic diversity in 17 the context of goals for the breeding program.18 19 4 Background 20 Plant breeding programs consist of 1) recurrent genetic improvement projects, 2) variety 21 development projects 3) trait introgression projects and 4) product placement projects (Fehr, 22 1991). Genetic improvement is assessed using realized genetic gain, which is an estimate of 23 change of the average genotypic value for traits of interest across cycles of selection and inter-24

Strategies to assure optimal trade-offs among competing objectives for genetic improvement of soybean

“…As reported previously (Ramasubramanian and Beavis 2020), there were 3818 polymorphic loci in the combined population of 20 families and an average of 773 polymorphic loci within each of the families for the initial founding sets of lines. The variance among families was ∼ 34 polymorphic loci.…”

“…Initial sets of soybean lines were generated by simulating crosses of 20 contemporary homozygous lines representing diversity of soybean germplasm adapted to MZ’s II and III with IA3023 to generate in silico F1 progeny (Ramasubramanian and Beavis 2020). Individual F1’s from each of the 20 crosses were self-pollinated in silico for four generations to generate 100 lines per family forming populations of 2000 lines organized into 20 families with genotypic information at 4289 genetic loci (Song et al 2017).…”

“…In a companion manuscript we reported an investigation of various factors on response metrics to recurrent selection of soybean lines derived from founders of the SoyNAM population (Ramasubramanian and Beavis 2020). The combinatorial set of factors consisted of phenotypic selection (PS) and four commonly used GS methods, training sets (TS), selection intensity (SI), number of QTL (nQTL), and heritability (H) on an entry mean basis.…”

“…These results are consistent with reports by Goddard (2009), Jannink (2010) and Liu et al (2015) demonstrated that the full genotypic potential of the founders is eliminated more quickly with GS than PS. In terms of factors that a soybean breeder can control, we found that SI’s of 1.75 and use of Ridge Regression Genomic Prediction (RRGP) models updated every cycle with training data from all prior cycles of selection provided rapid response in the early cycles of selection and retention of genetic diversity for continued response to selection in later cycles, but we noted that further improvements might be made if the populations were organized into islands and mating designs other than the hub network were employed (Ramasubramanian and Beavis 2020). Herein we investigate strategies that soybean breeders can employ to find optimal trade-offs between maximizing genetic gain from selection and retaining useful genetic diversity.…”

“…For both non-isolated and island models we applied three selection methods and four mating designs. To evaluate the potential of these combinations of methods to realize genetic gains while retaining useful genetic diversity, we compare outcomes from simulated recurrent selection applied to contemporary soybean germplasm adapted to MZ II and III using a set of metrics (Ramasubramanian and Beavis 2020). The metrics include the standardized genotypic value (R s ), the most positive genotypic value ( M gv ) among RIL’s selected in cycle c, the standardized genotypic variance ( SV g ), the average expected heterozygosity (Hs), the lost genetic potential of populations based on the number of favorable alleles that are lost.…”