Multi-omics prediction of oat agronomic and seed nutritional traits across environments and in distantly related populations

Abstract

Key message Integration of multi-omics data improved prediction accuracies of oat agronomic and seed nutritional traits in multi-environment trials and distantly related populations in addition to the single-environment prediction. Abstract Multi-omics prediction has been shown to be superior to genomic prediction with genome-wide DNA-based genetic markers (G) for predicting phenotypes. However, most of the existing studies were based on historical datasets from one environment; therefore, they were unable to evaluate the efficiency of multi-omics prediction in multi-environment trials and distantly related populations. To fill those gaps, we designed a systematic experiment to collect omics data and evaluate 17 traits in two oat breeding populations planted in single and multiple environments. In the single-environment trial, transcriptomic BLUP (T), metabolomic BLUP (M), G + T, G + M, and G + T + M models showed greater prediction accuracy than GBLUP for 5, 10, 11, 17, and 17 traits, respectively, and metabolites generally performed better than transcripts when combined with SNPs. In the multi-environment trial, multi-trait models with omics data outperformed both counterpart multi-trait GBLUP models and single-environment omics models, and the highest prediction accuracy was achieved when modeling genetic covariance as an unstructured covariance model. We also demonstrated that omics data can be used to prioritize loci from one population with omics data to improve genomic prediction in a distantly related population using a two-kernel linear model that accommodated both likely casual loci with large-effect and loci that explain little or no phenotypic variance. We propose that the two-kernel linear model is superior to most genomic prediction models that assume each variant is equally likely to affect the trait and can be used to improve prediction accuracy for any trait with prior knowledge of genetic architecture.

Generalizable approaches for genomic prediction of metabolites in plants

“…instance, computing relationship matrices from metabolomics data (Campbell et al, 2021a) or metabolomics and transcriptomics data (Hu et al, 2021) led to high mean prediction accuracies (r > 0.4) for fatty acid traits in oat (Avena sativa L.) seed. These studies have demonstrated that genomic prediction is effective for a few to tens of biochemically similar metabolites traits.…”

“…Whole-metabolome phenotypes were measured from mature seeds using untargeted LC-MS and GC-MS in a diverse oat germplasm panel of 375 inbred lines (discovery panel). These phenotypes have been previously described (Brzozowski et al, 2021;Campbell et al, 2021aCampbell et al, , 2021bHu et al, 2021). For each metabolite phenotype, measured as relative signal intensity, deregressed best linear unbiased predictors (drBLUPs) could be calculated for 1,067 of the LC-MS and 601 of the GC-MS signals as in Campbell et al (2021b).…”

“…The metabolites were annotated by comparison to an in-house spectral library RAMSearch (Broeckling et al, 2016) and MSFinder (Tsugawa et al, 2016), and details of measurement and annotation of this dataset are provided in Brzozowski et al (2021). Campbell et al (2021aCampbell et al ( , 2021b, and Hu et al (2021). To further characterize the metabolites, we examined the continuous variables of retention time (a measure of polarity, where, using a reverse phase column, a lower retention time indicates greater polarity), molecular mass, and genomic heritability (de los Campos et al, 2015).…”

“…We tested if metabolite kernels developed in the discovery panel improved prediction accuracy for metabolites in a validation panel evaluated in three environments (MN, SD, and WI) that had 397 LC-MS and 243 GC-MS metabolites. Although the measurements do not allow for direct comparison of all individual metabolites to those in the discovery panel (because of currently no robust method to map all untargeted metabolites from one panel to another and quantify them accurately, Hu et al, 2021), the metabolite Note. Descriptions of the kernels is given in Table 2.…”

“…For this analysis, we removed lines that overlapped with our discovery (diverse) panel, leaving 212 lines in MN and SD and 208 lines in WI. The relationship between the discovery and validation panels are described in Hu et al (2021) named as "discovery" and "elite" panels, respectively. Deregressed BLUPs were calculated as in Campbell et al (2021b), where data was cube-root transformed, and there were 397 LC-MS metabolites and 243 GC-MS metabolites (640 total) for which drBLUPs could be calculated.…”