We analyzed 200 relative metabolite levels measured on 2,526 participants of the Framingham Heart Study (FHS) offspring cohort, free of overt HF14. Baseline characteristics of the study samples are provided in Supplementary Materials (Table S1). Three distinct platforms were applied to measure metabolites in the plasma samples: (1) amino acids and amines (AAA) 15, (2) bile acids, organic acids, nucleotides, sugars, and other intermediary metabolites (BONS) 16, and (3) lipids of varying classes 17. The list of metabolite names, including 42 AAAs, 54 BONSs, and 104 lipids, is provided in Supplementary Materials (Table S2). Genotyping was conducted for over 550,000 single nucleotide polymorphisms (SNPs) 18. The participants were followed up and the first HF incident was recorded.
Using systems approaches, we here analyzed these data and presented two key aspects of the analyses: first, characterizing interconnectivity of metabolites and revealing them as networks, and second, incorporating the network structure to identify the effect of metabolites on HF incidence. For the former, we focused on metabolites measured in each platform and identified three data-driven networks of AAA, BONS, and lipids, where a link represents a partial correlation, a correlation that is not attributable to the other metabolites in the network. To identify causal networks, we augmented the networks with polygenic factors, generated by extracting information from genetic variations. Some polygenic factors satisfied MR assumptions to facilitate direction identification in the AAA network. However, there was not such polygenic factors for BONS/lipid networks. We extracted network properties and revealed the role of each metabolite individually and subnetworks.
In the second aspect of the analysis, we linked the networks to HF incidence. We used a model that was controlled for the effect of confounding metabolites in addition to sex, age, and BMI. For each metabolite, we provided 4 properties: effect size, significant level (p-value), hazard ratio, and connectivity with the other metabolites. Fig. 1 illustrates the study design and analysis workflow. The association of HF incidence with each single metabolite without considering the possible confounding by other metabolites are provided in (Tables S3-S8).
First aspect: Metabolomics and insights into regulation in health
Uncovering AAA connectivity and distinguishing broadcasters from receivers. In total, 8 out of 42 AAAs were directly influenced by polygenic factors that satisfied Mendelian randomization assumptions in the AAA causal network. The list of these metabolites is provided in the supplementary, (Table S9). In constructing the AAA causal network, we assumed that the exogenous variables (lifestyle/genetic variation) are the only sources of variation in essential AAs (threonine, lysine, isoleucine, valine, methionine, histidine, phenylalanine, tryptophan, and leucine). The AAA network is depicted in Fig. 2A with a focus on four metabolites (arginine, asparagine, alanine, and creatine) influenced only by essential AAs.
The connectivity/link between each two AAAs were highly significant (p-value < 1´10-15). In addition, some of AAAs had a high number of connectivity (>= 6), i.e., directly connected to several metabolites, (Fig. 2B). Some of the AAAs were “broadcasters”4, i.e., metabolites directly affecting multiple metabolites with effects propagated downstream to other metabolites, (Fig. 2C); threonine, as an example of a broadcaster, affects 5 other metabolites directly (out degree = 5) and through each of them, affects multiple other metabolites indirectly. Unlike broadcasters, some metabolites were “receivers”, i.e., metabolites with no or very low impact on other metabolites but receiving impact from multiple other metabolites, (Fig. 2D); hydroxyproline, as an example of a receiver, is influenced by 6 other metabolites directly (in degree = 6).
Uncovering BONS connectivity and distinguishing broadcasters from receivers. In the data-driven network of 54 BONS, we observed some metabolites with high connectivity (>=7) and some with very low connectivity (<= 2) (Fig. 3A). The hexose monophosphate among the high connected metabolites belongs to the carbohydrate super pathway and the glycolysis sub-pathway and is composed of fructose-1-phosphate, fructose-6-phosphate, glucose-1-phosphate, and glucose-6-phosphate. More information about this group of metabolites is provided in (Supplementary Materials).
Some of BONS metabolites had the property of broadcasters, affecting multiple other metabolites directly, as illustrated by kynurenine as an example (Fig. 3B). In contrast, there were some receivers, BONS metabolites with no or very low impact on other metabolites but highly influenced by others, as illustrated by hypoxanthine as an example (Fig. 3C).
Uncovering lipid connectivity and distinguishing broadcasters from receivers as well as subnetworks. In the data-driven network of 104 lipids, we observed five subnetworks corresponding to structural classes of metabolites, lysophosphatidylethanolamine and lysophosphatidylcholine (LPE-LPC), cholesteryl ester (CE), sphingomyelin (SM), and phosphatidylcholine (PC), as well as triacylglycerol and diacylglycerol lipids (TG) (Fig. 4A). We extracted lipids with high connectivity (>=5) (Fig. 4B). The number of lipids in each subnetwork is provided in Fig. 4C.
The connections within LPE-LPC, CE, and SM subnetworks were highly significant (p-value < 5´10-35). We observed that PC lipids were mostly mediators to spread the effect of the LPE-LPC lipids to other subnetworks (Fig. 4A and 4D). There was no individual lipid with broadcaster or receiver properties across the network; however, with the exception of TG, some lipids within each subnetwork demonstrated such properties (Fig. 4E-F).
Second aspect: Link between metabolites and the progression to HF
We identified metabolites associated with future HF risk after controlling for the confounding metabolites and provided four properties for each metabolite including hazard ratio, effect size, the significance level (p-value), and connectivity (Table 1, Fig. 5).
Out of 8 AAAs associated with HF incidence, three were inversely associated, including glycine, asparagine, and serotonin, and all three were associated with polygenic factors. Asparagine was also directly affected by the essential amino acid threonine, and in turn, asparagine directly affects glycine.
Isoleucine and ureidopropionic acid are among the five AAAs associated with increased risk of HF incidence. In the AAA causal network, we observed that isoleucine affected ureidopropionic acid directly, and in turn both contributed to an increase in HF risk (Fig. 5B). Note that the effect of ureidopropionic acid is not attributed to isoleucine. While assessing the effect of ureidopropionic acid on HF incidence, isoleucine is considered a confounder and was included in the Cox model. However, the effect of ureidopropionic acid on HF risk remained significant even after adjusting for isoleucine. The same property was observed for dimethylglycine and trimethylamine N-oxide (Fig. 5B).
Five BONSs were associated with HF incidence: glycocholic acid, inosine, 2-hydroxybutyrate, isocitric acid, and hexoses (hexose metabolites, fructose-glucose-galactose). All of these were directly associated with HF incidence (Table 1). We revealed metabolites directly connected to isocitric acid which may facilitate understanding the unknown mechanisms of isocitric acid association with HF (Supplementary, Fig. S1). The relatively higher out degree of both hexose and isocitric acid (out degree = 5) suggests that there are multiple other BONSs that may increase HF risk, but their p-values did not pass the statistical significance threshold. The out/in degree of glycocholic acid was not identified because the direction of links directly connected to it were not identified. The snapshot of the network with a focus on glycocholic acid is provided in Supplementary (Fig. S2).
After adjusting for confounding lipids, age, sex, and BMI, no lipid was identified associated with HF incidence. However, adjusting for confounding lipids, age, sex, three lipids were associated with HF risk, among them LPC 18:2 had a negative effect on HF (Table 1).
Table 1. Metabolites with significant effect on HF risk after controlling for confounding metabolites. A higher out degree represents a higher impact of the indicated metabolite on the other metabolites. A higher in degree represents a higher impact of other metabolites on the indicated metabolite. Glycine and isoleucine had the highest impact on other AAAs, out degree = 4, among metabolites associated with HF risk. The out/in degree of glycocholic acid was not identified because the direction of links directly connected to it were not identified. Three lipids showed association with HF risk after adjusting for sex and age; adjusting for BMI in addition to age and sex, no lipid was associated with HF risk.