Prediction of complex phenotypes using the Drosophila melanogaster metabolome

Abstract

Understanding the genotype-phenotype map and how variation at different levels of biological organization is associated are central topics in modern biology. Fast developments in sequencing technologies and other molecular omic tools enable researchers to obtain detailed information on variation at DNA level and on intermediate endophenotypes, such as RNA, proteins and metabolites. This can facilitate our understanding of the link between genotypes and molecular and functional organismal phenotypes. Here, we use the Drosophila melanogaster Genetic Reference Panel and nuclear magnetic resonance (NMR) metabolomics to investigate the ability of the metabolome to predict organismal phenotypes. We performed NMR metabolomics on four replicate pools of male flies from each of 170 different isogenic lines. Our results show that metabolite profiles are variable among the investigated lines and that this variation is highly heritable. Second, we identify genes associated with metabolome variation. Third, using the metabolome gave better prediction accuracies than genomic information for four of five quantitative traits analyzed. Our comprehensive characterization of population-scale diversity of metabolomes and its genetic basis illustrates that metabolites have large potential as predictors of organismal phenotypes. This finding is of great importance, e.g., in human medicine, evolutionary biology and animal and plant breeding.

Fig. 1. Conceptual illustration of the NMR cluster-guided phenotypic predictions.

All pairwise correlations among the NMR features were computed, which was used in a hierarchical clustering of NMR data points. The dendrogram was then sequentially cut into K clusters (25, 50, 75, 100, 125, 150 and 200 clusters), and each individual cluster was then used in the MFBLUP model. NMR data points within one cluster were used to construct a metabolomic relationship matrix that was used as covariance matrix in the MFBLUP model. The MFBLUP model was fitted for all clusters within the seven levels of K clusters.

Fig. 2. Genetic variation for the D. melanogaster metabolome.

Panel (A) shows in solid blue line the average NMR intensity across all DGRP lines (intensity axis not shown) as function of chemical shift. For each NMR data point, we estimated the heritability (h²); the points in grey represent non-significant estimates of h², and points in green are significant estimates of h². Panel (B) is a histogram of the significant heritability estimates.

Fig. 3. Prediction accuracies using genomic and metabolomic data.

For each panel, the barplot shows the maximum mean prediction accuracies (+standard error) for the different models. GBLUP and MBLUP are based on single component prediction models, whereas the two MFBLUP models are based on two components. The global maximum prediction accuracy obtained across all levels of clusters (K_cl = {25, 50, 75, 100, 125, 200}) is shown in the MFBLUP bar (indicated with white arrow). The prediction accuracy when combining the significant clusters is shown in the MFBLUP2 bar (indicated with white square and circle). Significant improved predictive performance is indicated by asterisks above the bars, see Supplementary Table S3 for all comparisons. The heatmaps on the right side of the panels show all prediction accuracies for the NMR cluster-guided prediction model within K_cl cluster level. The columns correspond to NMR data points (fixed across the K_cl cluster levels) and each cell is one cluster of NMR data points (link between NMR data points and clusters can be found in Supplementary Table S4). The predictive performance of each cluster within K_cl cluster level is indicated with the colour scale. Within cluster level significant prediction accuracies are indicated with white squares, and the cluster with the highest significant prediction accuracy is indicated with asterisk. Across all cluster levels, the highest prediction accuracy is indicated with white arrow (which then corresponds to the orange bars on the left-side panel). The set of significant clusters that give the highest predictive performance is marked with white squares with black circle (corresponds to the light green bars in the barplot).

Fig. 4. Contributions to predictive clusters from the metabolite NMR spectra at the K_cl = 200 level.

Clusters are indicated with coloured dots on the average of all NMR spectra (black line). Panel (A) shows clusters: 1, 2, 5, 21, 32, 45, 65, 74, 83 and 95; panel (B) clusters: 112, 121, 129, 139, 145 and 154; panel (C) cluster: 171; and panel (D) cluster: 187. Selected major metabolites in these regions are identified. The location of the nicotinamide ribotide signals resonating at the highest ppm values is also indicated in panel (A).