Phenomic selection in wheat breeding: prediction of the genotype-by-environment interaction in multi-environment breeding trials.

KEY MESSAGE: Phenomic prediction of wheat grain yield and heading date in different multi-environmental trial scenarios is accurate. Modelling the genotype-by-environment interaction effect using phenomic data is a potentially low-cost complement to genomic prediction. The performance of wheat cultivars in multi-environmental trials (MET) is difficult to predict because of the genotype-by-environment interactions (G × E). Phenomic selection is supposed to be efficient for modelling the G × E effect because it accounts for non-additive effects. Here, phenomic data are near-infrared (NIR) spectra obtained from plant material. While phenomic selection has recently been shown to accurately predict wheat grain yield in single environments, its accuracy needs to be investigated for MET. We used four datasets from two winter wheat breeding programs to test and compare the predictive abilities of phenomic and genomic models for grain yield and heading date in different MET scenarios. We also compared different methods to model the G × E using different covariance matrices based on spectra. On average, phenomic and genomic prediction abilities are similar in all different MET scenarios. Better predictive abilities were obtained when G × E effects were modelled with NIR spectra than without them, and it was better to use all the spectra of all genotypes in all environments for modelling the G × E. To facilitate the implementation of phenomic prediction, we tested MET designs where the NIR spectra were measured only on the genotype-environment combinations phenotyped for the target trait. Missing spectra were predicted with a weighted multivariate ridge regression. Intermediate predictive abilities for grain yield were obtained in a sparse testing scenario and for new genotypes, which shows that phenomic selection is an efficient and practicable prediction method for dealing with G × E.

Phenomic selection in wheat breeding: identification and optimisation of factors influencing prediction accuracy and comparison to genomic selection.

KEY MESSAGE: Phenomic selection is a promising alternative or complement to genomic selection in wheat breeding. Models combining spectra from different environments maximise the predictive ability of grain yield and heading date of wheat breeding lines. Phenomic selection (PS) is a recent breeding approach similar to genomic selection (GS) except that genotyping is replaced by near-infrared (NIR) spectroscopy. PS can potentially account for non-additive effects and has the major advantage of being low cost and high throughput. Factors influencing GS predictive abilities have been intensively studied, but little is known about PS. We tested and compared the abilities of PS and GS to predict grain yield and heading date from several datasets of bread wheat lines corresponding to the first or second years of trial evaluation from two breeding companies and one research institute in France. We evaluated several factors affecting PS predictive abilities including the possibility of combining spectra collected in different environments. A simple H-BLUP model predicted both traits with prediction ability from 0.26 to 0.62 and with an efficient computation time. Our results showed that the environments in which lines are grown had a crucial impact on predictive ability based on the spectra acquired and was specific to the trait considered. Models combining NIR spectra from different environments were the best PS models and were at least as accurate as GS in most of the datasets. Furthermore, a GH-BLUP model combining genotyping and NIR spectra was the best model of all (prediction ability from 0.31 to 0.73). We demonstrated also that as for GS, the size and the composition of the training set have a crucial impact on predictive ability. PS could therefore replace or complement GS for efficient wheat breeding programs.

IPSIM-Web, An Online Resource for Promoting Qualitative Aggregative Hierarchical Network Models to Predict Plant Disease Risk: Application to Brown Rust on Wheat.

A qualitative pest modeling platform, named Injury Profile Simulator (IPSIM), provides a tool to design aggregative hierarchical network models to predict the risk of pest injuries, including diseases, on a given crop based on variables related to cropping practices as well as soil and weather environment at the field level. The IPSIM platform enables modelers to combine data from various sources (literature, survey, experiments, and so on), expert knowledge, and simulation to build a network-based model. The overall structure of the platform is fully described at the IPSIM-Web website ( www6.inra.fr/ipsim ). A new module called IPSIM-Wheat-brown rust is reported in this article as an example of how to use the system to build and test the predictive quality of a prediction model. Model performance was evaluated for a dataset comprising 1,788 disease observations at 13 French cereal-growing regions over 15 years. Accuracy of the predictions was 85% and the agreement with actual values was 0.66 based on Cohen's κ. The new model provides risk information for farmers and agronomists to make scientifically sound tactical (within-season) decisions. In addition, the model may be of use for ex post diagnoses of diseases in commercial fields. The limitations of the model such as low precision and threshold effects as well as the benefits, including the integration of different sources of information, transparency, flexibility, and a user-friendly interface, are discussed.

Genome-wide prediction of three important traits in bread wheat.

Five genomic prediction models were applied to three wheat agronomic traits-grain yield, heading date and grain test weight-in three breeding populations, each comprising about 350 doubled haploid or recombinant inbred lines evaluated in three locations during a 3-year period. The prediction accuracy, measured as the correlation between genomic estimated breeding value and observed trait, was in the range of previously published values for yield (r = 0.2-0.5), a trait with relatively low heritability. Accuracies for heading date and test weight, with relatively high heritabilities, were about 0.70. There was no improvement of prediction accuracy when two or three breeding populations were merged into one for a larger training set (e.g., for yield r ranged between 0.11 and 0.40 in the respective populations and between 0.18 and 0.35 in the merged populations). Cross-population prediction, when one population was used as the training population set and another population was used as the validation set, resulted in no prediction accuracy. This lack of cross-population prediction accuracy cannot be explained by a lower level of relatedness between populations, as measured by a shared SNP similarity, since it was only slightly lower between than within populations. Simulation studies confirm that cross-prediction accuracy decreases as the proportion of shared QTLs decreases, which can be expected from a higher level of QTL × environment interactions.