Genomic prediction by modelling genotype by environment interaction for yield in groundnut

Multi-environment trials (METs) are routinely conducted in plant breeding to capture the genotype-by-environment (G×E) interaction effects. We assessed the potential of using genomic prediction (GP) in four groundnut traits (pod yield[PY], shelling percentage[SP], seed weight[SW] and seed weight of 100 seeds[SW100]) observed in four environments. Three prediction models (M1: Environment + Line, M2: Environment + Line + Genomic, and M3: Environment + Line + Genomic + Genomic × Environment) were evaluated under four cross-validation (CV) schemes that simulate realistic problems that breeders face at different stages of their programs. CV2 predicts lines tested in sparse multi-location trials; CV1 predicts newly developed lines; CV0 predicts tested lines in unobserved environments by leaving one environment out; CV00 predicts untested lines in unobserved environments. Under CV2, the average predictive ability (PA) range was 0.6-0.68 (PY), -0.26-0.05 (SP), 0.5-0.59 (SW) and 0.79-0.86 (SW100). For CV1, the PA range was -0.12-0.51 (PY), -0.08-0.35 (SP), -0.10-0.47 (SW) and -0.05-0.47 (SW100). In CV0, the PA range was 0.27-0.35 (PY), -0.18--0.17 (SP), 0.24-0.32 (SW) and 0.59-0.66 (SW100). In CV00, the PA range was 0.01-0.23 (PY), -0.03-0.07 (SP), 0.01-0.19 (SW) and -0.02-0.38 (SW100). In all the four CV schemes, including marker data improved the PA. Incorporating G×E in M3 model improved PA in CV2 and CV1 schemes and also reduced the residual and environment variances. The high PA in CV2 imply that sparse testing could be implemented. The low PA recorded in CV00 shows that it is difficult to predict untested lines in new environments.