The plant materials and experimental designs
The Diversity and Elite panels consisted of 378 and 252 lines (Table S2), respectively. The Diversity panel originally included 500 lines described by Carlson et al. (2019) that was a core set of worldwide collection of oat germplasm, and we further selected for lines with visible anther extrusion for the convenience of collecting developing seeds for RNA sequencing. The Diversity panel was planted at Ithaca, NY, and the Elite panel was planted at Madison, WI, Crookston, MN, and Brookings, SD, respectively. An augmented incomplete design was used for both panels. The Diversity panel included 18 blocks of 23 plots each, one common check across all blocks and six secondary checks replicated in three blocks each. The Elite panel included 12 blocks of 25 plots each, one common check across all blocks and two secondary checks replicated in six blocks each.
Phenotype evaluation and analysis
Plant height was evaluated for five randomly selected plants in each plot after anthesis. Days to heading was defined by the days from seeding to heading in > 50% of total plants. 100 randomly selected seeds from each plot were dehulled with a hand dehuller for evaluation of hundred kernel weight, hundred hull weight and groat percentage. After dehulling, 50 randomly selected seeds were delivered to the Proteomics and Metabolomics Facility at Colorado State University for metabolite analysis, and the other 50 seeds were used for measuring seed length, width and height with an electronic micrometer. Fatty acids were identified and quantified with targeted GC-MS, then normalized to concentration (mg/g of oats) against the internal standard (C17:0) (details were described in the Supplemental Methods).
Genotype analysis
Genotypic data of the two panels were downloaded from T3/oat (https://triticeaetoolbox.org/oat/). SNPs were filtered using the following criteria (i) minor allele frequency (MAF) > 2%; (ii) site missingness < 60%; and (iii) site heterozygosity < 10%. After initial SNP filtering, lines were selected if (i) call rate > 80% and (ii) heterozygosity < 10%. A total of 73,014 markers and 568 lines (368 for the diversity panel, 232 for the elite panel, 32 in common) met these criteria and were used for further analyses. Subsequently, missing genotypes were imputed using the linear regression method glmnet described by Chan et al. (2016). The imputed genotypic data was used for constructing a neighbor-joining tree based on Rogers’ distance using the ape package (Paradis et al. 2004), and the tree was visualized with the ggtree package (Yu 2020).
Transcript profiling
RNAseq was based on developing seeds at 23 days after anthesis (DAA). The 23 DAA was chosen based on our pilot study (Hu et al. 2020) that showed 23 DAA had slightly higher correlation between transcript and metabolite abundance than other sampled seed developmental time points. Seed sample collection, RNA extraction, library construction procedures were described in details by Hu et al. (2020). Pooled libraries were sequenced using Illumina NextSeq500 with a 150 nt single-end run. The RNAseq reads quality trimming, transcript abundance quantification, and library size normalization followed Hu et al. (2020).
Metabolite profiling and network analysis
Metabolite analysis was based on physiologically mature seeds because they have the highest level of health-promoting compounds and those compounds are stable at room temperature until germination. GC-MS non-targeted analysis and LC-MS phenyl–hexyl analysis were done at the Proteomics and Metabolomics Facility at Colorado State University. Details of chemical analysis, raw mass spectrometry data processing, metabolite annotation, and normalization were described in Supplemental Methods. The normalized metabolomics data were used for network analysis with the WGCNA package (Zhang and Horvath, 2005) following the tutorial at https://horvath.genetics.ucla.edu/html/CoexpressionNetwork/Rpackages/WGCNA/Tutorials/FemaleLiver-02-networkConstr-man.R. Module identification included the following steps: (i) Correlation network adjacency was calculated using the soft thresholding power 4, which was selected based on the scale independence chart as described in the WGCNA tutorial; (ii) To minimize effects of noise and spurious associations, we transformed the adjacency matrix into Topological Overlap Matrix (TOM), and calculated the corresponding dissimilarity (1-TOM); (iii) We then used hierarchical clustering to produce a hierarchical clustering tree of metabolite features based on TOM dissimilarity matrix with method = "average"; (iv) Modules were identified using the cutreeDynamic function with the following parameters: method = "hybrid", distM = dissTOM, deepSplit = 2, pamRespectsDendro = FALSE, minClusterSize = 20.
Analysis of phenotypic traits, transcriptomic, and metabolic features
Phenotypic traits, transcriptomic and metabolic features were analyzed following a standard linear mixed model of an augmented design accounting for effects of check genotypes and blocks (Campbell et al. 2021a). For metabolites analysis, batch effect was also included in the model to account for batch variation. All statistical models were described in Supplemental Methods and fitted using the sommer package (Covarrubias-Pazaran 2016).
Single-environment prediction
The additive genomic relationship matrix was calculated with the A.mat function implemented in the rrBLUP package (Endelman 2011), and relationship matrices for transcripts (TRM) and metabolites (MRM) were calculated with the following equations:
$${\text{TRM}} = \frac{1}{{N_{{\text{T}}} }} W_{{\text{T}}} W_{{\text{T}}}^{{\text{T}}} ,$$
(1)
$${\text{MRM}} = \frac{1}{{N_{{\text{M}}} }} W_{{\text{M}}} W_{{\text{M}}}^{{\text{T}}} ,$$
(2)
where NT and NM denoted the number of transcript and metabolite features, respectively, WT and WM are the feature matrices of transcripts and metabolites, and WTT and WMT are transpose of feature matrices.
GBLUP, Transcriptomic BLUP (T), metabolomic BLUP (M), G + T, G + M, and G + T + M models were fitted with the BGLR package (Pérez & De Los Campos, 2014). The equations used to implement G + T, G + M and G + T + M models are:
$$y = Xb + G\alpha + T\beta + \varepsilon ,$$
(3)
$$y = Xb + G\alpha + M\gamma + \varepsilon ,$$
(4)
$$y = Xb + G\alpha + T\beta + M\gamma + \varepsilon ,$$
(5)
where y is a vector of phenotypes, X is a design matrix relating the fixed effects to each genotype, b is a vector of fixed effects, α, β and γ are random effects of genome, transcriptome and metabolome, respectively; G, T, and M are design matrices allocating records to those random effects; ε is random residual effect.
In the Diversity panel, transcriptomics and metabolomics data were collected on the same plots as the phenotypic data and therefore non-genetic (i.e., microenvironmental) factors that affected both omics features and phenotypic traits may induce non-genetic correlations among traits. Therefore, we estimated prediction accuracy as \(c\hat{o}r_{g} \left( {\sqrt {\hat{h}_{{\hat{u}}}^{2} } } \right)\) described by Runcie and Cheng (2019), and used a 50:50 training/testing split of the data to ensure that \(c\hat{o}r_{g}\) could be estimated accurately in the testing partition. This cross-validation procedure was repeated for 50 times with different random partitions. To determine whether there was a significant difference in prediction accuracy between each omics model and the GBLUP model, we performed the Wilcoxon signed-rank test based on prediction accuracies across the 50 cross-validation runs for each pair of models. The Wilcoxon signed-rank test was also applied to multi-environment prediction and prediction of distantly related individuals in this study.
Multi-environment prediction
The metabolomics data were also collected on the same plots as the phenotypic data for the Elite panel, which would bias prediction accuracy if directly using metabolites to predict target phenotypes from the same environment. Therefore, when predicting target phenotypes from one environment, we used metabolites from other two environments to make the metabolomic relationship matrix. For each trait, we fitted six multi-trait mixed models on G, M and G + M kernels with different genetic and residual covariance structures. A standard multi-trait linear mixed model was used, and the equation for the case of genomic SNPs is:
$$y = Xb + Zg + \varepsilon ,$$
(6)
where y = (y1’, y2’, y3’)’, g = (g1’, g2’, g3’)’, ε = (ε1’, ε2’, ε3’)’. y1, y2, and y3 are the column vectors of phenotypic data in each environment. g1, g2, and g3 are the column vectors of random genetic effects in each environment. ε1, ε2, and ε3 are the column vectors of random error terms associated with each environment. X and Z are design matrices relating the fixed and random effects to each genotype. Vectors containing the random effects in Eq. (6) are assumed to follow a multivariate normal distribution, centered at zero, and with covariance structure Cov(g, g’) = G0 \(\otimes\) K, Cov(ε, ε’) = I \(\otimes\) R0, and Cov(g, ε’) = 0, where K is the additive genomic relationship matrix, I is an identity matrix, \(\otimes\) is the Kronecker product, G0 is a 3 × 3 genetic covariance matrix, R0 is a 3 × 3 residual covariance for the three locations. There are various covariance structures for R0 or G0 (Burgueño et al. 2012). In this study, six multi-trait models on three different kernels/combinations (G, M, G + M) with various genetic and residual covariance structure were used (codes and covariance structures of the six multi-trait mixed models were described in Table S3).
We applied a single-environment cross-validation method originally designed for genomic prediction described by Mathew et al. (2018) and extended it to multi-kernel omic prediction (illustrated in Fig. S1). To predict a phenotype in the first environment, we masked 20% of lines for cross-validation and used metabolites from the other two environments to construct the metabolomic relationship matrix. We then used multi-trait models treating phenotypes from all three environments as separate traits for model training but using only the phenotypic data of the masked lines from the first environment as the testing data. We further estimated prediction accuracy of the first environment as \(r\left( {\hat{y},y} \right)/\sqrt {h^{2} }\) (Riedelsheimer et al. 2012), where r(\(\hat{y}\),y) is the Pearson correlation between the observed (y) and predicted (\(\hat{y}\)) phenotypic values and h2 is the heritability of the target trait. To predict the phenotype in the second and third environments, we masked 20% of lines (the same set of lines as those in the first environment) from the second and third environments, respectively, and calculated their prediction accuracies following the same procedure as that applied to the first environment. Finally, we averaged the three prediction accuracies across environments to represent the prediction accuracy of a single run. This procedure was repeated for 50 times with different random partitions.
Prediction of distantly related individuals
Seed fatty acid concentrations were used as target traits for predicting distantly related individuals, which included two steps: likely causal loci prioritization in the Diversity panel (training population) and multiple-kernel prediction in the Elite panel (test population).
We first performed the WGCNA on all metabolite features in the Diversity panel (training population), and identified twenty-six network modules. Based on the metabolites annotation, we performed Fisher's exact test to identify a subset of network modules enriched with lipids and lipid-like molecules. We then performed hierarchical clustering (using correlation based dissimilarity matrix with method = "average") and GWAS on eigenvectors of the twenty-six network modules and PC1 of fatty acids. GWAS was performed based on the linear mixed model (Yu et al. 2006) implemented in the GWAS function of the rrBLUP package (Endelman 2011) with the following parameters: K = GRM (additive genomic relationship matrix), n.PC = 2, min.MAF = 0.02, n.core = 4 (Campbell et al. 2021a). Based on these analyses, we found that a 'darkred' module enriched with lipids and lipid-like molecules, clustered together with PC1 of fatty acids, and its eigenvector had a QTL co-located with the major-effect QTL of fatty acids on chromosome 6A. We finally prioritized 140 markers including significant markers and the markers in LD with them based on the GWAS peak on chromosome 6A identified from the 'darkred' module. A LD threshold of r2 = 0.1 was used as it is frequently recommended for SNP pruning (Kawakami et al. 2014).
The prioritized markers and all rest markers were used to construct two genomic relationship kernels in the Elite panel (test population) and perform a multiple-kernel prediction. The two genomic relationship matrices were calculated with the A.mat function implemented in the rrBLUP package (Endelman 2011). Genomic predictions with GBLUP and BayesB models were used as references to compare with the two-kernel linear model. The fivefold cross-validation was used to estimate prediction accuracies for all models and the prediction accuracy was estimated as \(r\left( {\hat{y},y} \right)/\sqrt {h^{2} }\) (Riedelsheimer et al. 2012). This cross-validation procedure was repeated for 50 times with different random partitions.