Development of a Genomic Prediction Pipeline for Maintaining Comparable Sample Sizes in Training and Testing Sets across Prediction Schemes Accounting for the Genotype-by-Environment Interaction

The global growing population is experiencing challenges to satisfy the food chain supply in a world that faces rapid changes in environmental conditions complicating the development of stable cultivars. Emergent methodologies aided by molecular marker information such as marker assisted selection (MAS) and genomic selection (GS) have been widely adopted to assist the development of improved genotypes. In general, the implementation of GS is not straightforward, and it usually requires cross-validation studies to find the optimum set of factors (training set sizes, number of markers, quality control, etc.) to use in real breeding applications. In most cases, these different scenarios (combination of several factors) vary just in the levels of a single factor keeping fixed the levels of the other factors allowing the use of previously developed routines (code reuse). In this study, we present a set of structured modules that are easily to assemble for constructing complex genomic prediction pipelines from scratch. Also, we proposed a novel method for selecting training-testing sets of sizes across different cross-validation schemes (CV2, predicting tested genotypes in observed environments; CV1, predicting untested genotypes in observed environments; CV0, predicting tested genotypes in novel environments; and CV00, predicting untested genotypes in novel environments). To show how our implementation works, we considered two real data sets. These correspond to selected samples of the USDA soybean collection (D1: 324 genotypes observed in 6 environments scored for 9 traits) and of the Soybean Nested Association Mapping (SoyNAM) experiment (D2: 324 genotypes observed in 6 environments scored for 6 traits). In addition, three prediction models which consider the effect of environments and lines (M1: E + L), environments, lines and main effect of markers (M2: E + L + G), and also the inclusion of the interaction between makers and environments (M3: E + L + G + G×E) were considered. The results confirm that under CV2 and CV1 schemes, moderate improvements in predictive ability can be obtained with the inclusion of the interaction component, while for CV0 mixed results were observed, and for CV00 no improvements were shown. However, for this last scenario, the inclusion of weather and soil data potentially could enhance the results of the interaction model