Pea genomic selection for Italian environments

Abstract

Background: A thorough verification of the ability of genomic selection (GS) to predict estimated breeding values for pea (Pisum sativum L.) grain yield is pending. Prediction for different environments (inter-environment prediction) has key importance when breeding for target environments featuring high genotype × environment interaction (GEI). The interest of GS would increase if it could display acceptable prediction accuracies in different environments also for germplasm that was not used in model training (inter-population prediction).

Results: Some 306 genotypes belonging to three connected RIL populations derived from paired crosses between elite cultivars were genotyped through genotyping-by-sequencing and phenotyped for grain yield, onset of flowering, lodging susceptibility, seed weight and winter plant survival in three autumn-sown environments of northern or central Italy. The large GEI for grain yield and its pattern (implying larger variation across years than sites mainly due to year-to-year variability for low winter temperatures) encouraged the breeding for wide adaptation. Wider within-population than between-population variation was observed for nearly all traits, supporting GS application to many lines of relatively few elite RIL populations. Bayesian Lasso without structure imputation and 1% maximum genotype missing rate (including 6058 polymorphic SNP markers) was selected for GS modelling after assessing different GS models and data configurations. On average, inter-environment predictive ability using intra-population predictions reached 0.30 for yield, 0.65 for onset of flowering, 0.64 for seed weight, and 0.28 for lodging susceptibility. Using inter-population instead of intra-population predictions reduced the inter-environment predictive ability to 0.19 for grain yield, 0.40 for onset of flowering, 0.28 for seed weight, and 0.22 for lodging susceptibility. A comparison of GS vs phenotypic selection (PS) based on predicted genetic gains per unit time for same selection costs suggested greater efficiency of GS for all traits under various selection scenarios. For yield, the advantage in predicted efficiency of GS over PS was at least 80% using intra-population predictions and 20% using inter-population predictions. A genome-wide association study confirmed the highly polygenic control of most traits.

Conclusions: Genome-enabled predictions can increase the efficiency of pea line selection for wide adaptation to Italian environments relative to phenotypic selection.

Keywords: Breeding value; Cross-population prediction; Genotype × environment interaction; Genotyping-by-sequencing; Pisum sativum; Predictive ability; Yield.

Fig. 1

Nominal grain yield of six top-performing pea inbred lines out of 306 derived from three connected crosses, three parent cultivars (Attika, Isard, Kaspa) and one commercial cultivar (Spacial) as a function of the environment score on the first genotype × environment interaction principal component axis (PC 1) [environments are Lodi 2013–14 (Lo14), Lodi 2014–15 (Lo15), and Perugia 2013–14 (Pg14); the graph includes the two top-yielding lines in each environment or across environments]

Fig. 2

Intra-population predictive ability for pea grain yield in three environments, for all combinations of three regression models (BL, Bayesian Lasso; rrBLUP, Ridge regression BLUP; G-BLUP, genomic BLUP) and five genotype missing data thresholds. Data averaged across three pea RIL populations and 50 repetitions of 10-fold stratified cross-validation per individual analysis

Fig. 3

Intra-population predictive ability for pea mean grain yield, onset of flowering, lodging susceptibility and individual seed weight across three environments and winter plant survival in one environment, for all combinations of three regression models (BL, Bayesian Lasso; rrBLUP, Ridge regression BLUP; G-BLUP, genomic BLUP) and five genotype missing data thresholds. Data averaged across two (lodging susceptibility) or three (other traits) RIL populations and 50 repetitions of 10-fold stratified cross-validation per individual analysis

Fig. 4

Top 100 genome-wide association scores, ranked in descending order, for single-nucleotide polymorphism (SNP) markers associated with five phenotypic traits of pea. GWAS was performed on stratified data, with each of two (lodging susceptibility) or three (other traits) RIL populations acting as a stratum