The Human Blood Metabolome-Transcriptome Interface

Abstract

Biological systems consist of multiple organizational levels all densely interacting with each other to ensure function and flexibility of the system. Simultaneous analysis of cross-sectional multi-omics data from large population studies is a powerful tool to comprehensively characterize the underlying molecular mechanisms on a physiological scale. In this study, we systematically analyzed the relationship between fasting serum metabolomics and whole blood transcriptomics data from 712 individuals of the German KORA F4 cohort. Correlation-based analysis identified 1,109 significant associations between 522 transcripts and 114 metabolites summarized in an integrated network, the 'human blood metabolome-transcriptome interface' (BMTI). Bidirectional causality analysis using Mendelian randomization did not yield any statistically significant causal associations between transcripts and metabolites. A knowledge-based interpretation and integration with a genome-scale human metabolic reconstruction revealed systematic signatures of signaling, transport and metabolic processes, i.e. metabolic reactions mainly belonging to lipid, energy and amino acid metabolism. Moreover, the construction of a network based on functional categories illustrated the cross-talk between the biological layers at a pathway level. Using a transcription factor binding site enrichment analysis, this pathway cross-talk was further confirmed at a regulatory level. Finally, we demonstrated how the constructed networks can be used to gain novel insights into molecular mechanisms associated to intermediate clinical traits. Overall, our results demonstrate the utility of a multi-omics integrative approach to understand the molecular mechanisms underlying both normal physiology and disease.

Fig 1. Data integration and network analysis workflow for the blood metabolome-transcriptome interface (BMTI).

A: We analyzed fasting serum metabolomics and whole blood transcriptomics data from 712 samples of the KORA F4 cohort. After preprocessing and filtering, a cross-correlation matrix between 440 distinct metabolites and 16,780 unique, gene-mapped probes was calculated. The correlation matrix was transformed into a bipartite network by applying a statistical significance threshold. B: Scientific literature was screened for biological evidence for the strongest metabolite-mRNA associations. All correlating metabolite-mRNA pairs contained in the human metabolic model Recon2 were systematically evaluated with respect to their distance in the metabolic pathway network. C: Aggregated z-scores for each functional annotation were calculated. A pathway interaction network (PIN) was then constructed via cross-correlation of scores between pairs of functional annotations. D: For each metabolite contained in the BMTI, we investigated the promoter regions of associated transcripts for shared regulatory signatures. Similarly, shared regulatory signatures within and between metabolic pathways were examined. As a final step, we identified specific regulatory motifs in the BMTI. E: Both BMTI and the PIN were integrated with the results from an association analysis to the three intermediate physiological traits (HDL-C, LDL-C and TG).

Fig 2. The blood metabolome-transcriptome interface.

A: Distributions of correlation coefficients for metabolite-mRNA, mRNA-mRNA and metabolite-metabolite associations. B: Number of nodes and edges as a function of the absolute correlation coefficient. Red dotted line represents the correlation cutoff used in this study (0.01 FDR). C: Percentages of blood cell type specific markers contained in the BMTI. D: Visualization of the blood metabolome-transcriptome interface. The correlation network consists of 114 metabolites and 522 transcripts connected by 1109 edges. Edge widths represent correlation strengths.

Fig 3. Model-based evaluation of metabolite-mRNA correlations.

A: Schematic representation of the mitochondrial carnitine-shuttle with an explanation of network distance calculation. Note that co-factors are only illustrated for completeness, but are not considered for the calculation of the shortest path between two compounds. Dashed circles indicate unmeasured metabolites. B: Spearman correlation coefficient plotted against the number of pathway steps in human Recon 2. Significant correlations, i.e. those present in the BMTI are displayed as red crosses, whereas all non-significant correlations are plotted as a distribution. NM: no mapping. A distance of infinity (Inf) was assigned if there was no connection in Recon 2. C: Enrichment of significant correlations as determined by Fisher’s exact test. Black bars indicate log10 p-values assessing whether we observe more significant correlations for that particular distance than expected by chance. White bars represent the same test, only for a cumulative distance (i.e. “up to a distance of x”). D: Functional annotations of significantly associated transcripts at distances 0 and 1. At both distances, mainly transcripts belonging to the transport, energy, lipid and amino acid subsystems were observable.

Fig 4. Pathway interaction network (PIN).

A: Bipartite correlation network, where each node represents either a metabolic pathway or a gene set summarized in a GO term, while edges between them represent the correlation of the respective aggZ-scores. B+C: Expanded view on exemplary pathway interactions. Note that zoomed parts correspond to subnetworks from Fig 2D that are explained by one single link of the PIN. Edge widths represent correlation strengths.

Fig 5. Transcription factor binding site analysis.

A: Heatmap of shared significantly overrepresented TFBS between metabolic subpathways. Upper right triangle matrix was left out. Darker colors indicate a higher number of shared TF binding sites. B+C: Identified network motifs of TFs and their respective target genes associated to the same metabolite. Green arrows indicate regulation. Black lines indicate positive correlation; red dotted lines indicate negative correlation.

Fig 6. Intermediate clinical phenotype associations.

Linear regression results of TG, HDL and LDL for each metabolite and transcript (A, C, E), or pathway and GO term (B, D, F) were projected onto the respective networks. Blue colors indicate negative associations, while red colors represent positive associations to the respective phenotype. Color strength of the nodes encodes the-log10 p-value of the respective association. β denotes the regression coefficient and its sign represents the direction of the associations (positive or negative correlation).