Factors influencing phenomic prediction: A case study on a large sorghum back cross nested association mapping population

International audience

anglais. Abstract: Plant breeding is crucial to develop varieties able to cope with climate change and support food and feed value chains. Genomic prediction (GP) has been a major step in increasing their efficiency and recently, phenomic prediction (PP) has gained attention as a promising complementary approach to GP, potentially further increasing this efficiency. Factors impacting PP are not fully clarified. Thus, we studied the impacts of spectra preprocessing, prediction methods, population structure, training set size, near infrared reflectance spectroscopy (NIRS) acquisition environment, and wavelength selection on a large multi-parental sorghum population including 2498 BC1F3:5 families from 29 crosses with a strong population structure. Using 51,545 single nucleotide polymorphisms and 1154 NIRS features, we show that PP can reach predictive abilities (PAs) similar to GP, that it is less affected by population structure, and can reach its maximal PA with smaller training sets than GP, but its performances are trait dependent. We also show that NIRS can be acquired in a reference environment to perform prediction in other environments and that it is possible to randomly select wavelengths to perform predictions. Finally, we show that spectra preprocessing and statistical methods have an inconsistent impact on PA. Our study confirms that PP is a relevant trait prediction method that deserves attention to optimize breeding schemes. The main challenges for the future will be to better understand the information contained in the spectra and disentangle their genetic and proxy components to optimize the use of PP in breeding programs.Plain Language Summary: Plants grown by farmers are the result of a breeding process. Efficient plant breeding requires the estimation of breeding values retrieved by statistical models and genetic data. Phenomic prediction refers to the use of these models with a new type of data: infrared spectrum, which is easier to collect and less expensive than genetic data. We found that using infrared spectrum allows breeding value estimations sometimes as good as with genetic data, and that it may lead to more robust statistical models. However, phenomic prediction performance is variable and depends on many factors that are not fully understood. Our work is the first of this kind on sorghum, which is a staple crop in West Africa. Sorghum could benefit a lot from phenomic prediction as breeding programs in the south often have limited financial means.