Genomic Prediction for Rusts and Leaf Spotting Diseases in Wheat – a Comparison of Models

The unceasing plant-pathogen arms race and the ephemeral nature of resistance genes has been a major challenge to breeders. One promising approach that could accelerate the gain from selection for disease resistance is ‘genomic prediction’ which utilizes dense genome-wide markers to estimate the breeding values for quantitative traits. While several genomic prediction models have been proposed, our objective was to compare them with a multiple linear regression(MLR) model that uses only quantitative trait loci(QTL) linked markers and also with a pedigree model. For this, we used CIMMYT’s 45^th International Bread Wheat Screening Nursery (267 entries) that was genotyped using Genotyping By Sequencing markers and phenotyped for different diseases in replicated trials at El Batan(Mexico), Njoro(Kenya) and Toluca(Mexico). Our results show that the mean prediction accuracies were the highest for seedling resistance to tan spot (0.72±0.08) and to leaf rust (0.64±0.11), followed by stem rust (0.58±0.06), Stagonospora nodorum blotch (0.55±0.07), stripe rust (0.48±0.14), tan spot (0.45±0.08), Septoria tritici blotch (0.44±0.1) and leaf rust (0.41±0.08). Among the models, MLR gave the lowest mean prediction accuracies (0.34±0.13), while Reproducing Kernel Hilbert Spaces(RKHS) with markers and pedigree gave the highest (0.56±0.13). The Genomic-BLUP (0.53±0.12); Genomic-BLUP mixed model (0.52±0.14); Bayesian Ridge-Regression (0.53±0.12); BayesA (0.53±0.12); BayesB (0.52±0.12); BayesCπ (0.53±0.12); Bayesian LASSO (0.52±0.12); RKHS markers (0.53±0.12) and RKHS pedigree (0.52±0.13) yielded similar accuracies. Overall, our results indicate that (i)prediction accuracies depend on the genetic architecture of resistance (ii)using genome-wide markers is advantageous than using only QTL-linked markers (iii)using both pedigree and markers results in slightly higher accuracies.