Untargeted plasma metabolites in Dutch cohorts
In this study, we examined plasma metabolomes in 1,679 fasting plasma samples from 1,368 individuals from two LLD5 sub-cohorts (LLD1 and LLD2) and the GoNL6 cohort (Extended Data Fig. 1 and Supplementary Table 1). The LLD1 cohort was the discovery cohort, with information about genetics, diet and the gut microbiome available for 1,054 participants. Moreover, 311 LLD1 subjects were followed up 4years later (LLD1 follow-up). We also included two independent replication cohorts: 237 LLD2 participants for whom we had genetic and dietary data and 77 GoNL participants for whom only genetic data were available (Extended Data Fig. 1 and Supplementary Table 1). Untargeted metabolomics profiling was done using flow-injection time-of-flight mass spectrometry (FI-MS)10,11, which yielded plasma levels of 1,183 metabolites (Supplementary Table 2). These metabolites covered a wide range of lipids, organic acids, phenylpropanoids, benzenoids and other metabolites (Extended Data Fig. 2a). As we observed weak (absolute rSpearman<0.2) correlations among the 1,183 metabolites (Extended Data Fig. 2b), data reduction was not required and, consequently, all metabolites were subjected to subsequent analyses. We validated the identification and quantification of some metabolites (for example, bile acids, creatinine, lactate, phenylalanine and isoleucine) by comparing their abundance levels from FI-MS with those previously determined by liquid chromatography with tandem mass spectrometry (LC-MS/MS)12 or NMR13 (rSpearman>0.62; Extended Data Fig. 2c,d).
To compare the relative importance of diet, genetics and the gut microbiome in explaining inter-individual plasma metabolome variability, we calculated the proportion of variance explained by these three factors for the whole plasma metabolome profile and for the individual metabolites separately. We have detailed information on 78 dietary habits (Supplementary Table 3), 5.3million human genetic variants and the abundances of 156 species and 343 MetaCyc pathways for each individual of the LLD1 cohort. Diet, genetics and the gut microbiome could explain 9.3, 3.3 and 12.8%, respectively, of inter-individual variations in the whole plasma metabolome, without adjusting for covariates (see the Methods section Distance matrix-based variance estimation; false discovery rate (FDR)<0.05; Fig. 1a and Supplementary Table 4), whereas intrinsic factors (age, sex and body mass index (BMI)) and smoking collectively explained 4.9% of the variance. Together, these factors explain 25.1% of the variance in the plasma metabolome (Fig. 1a).
a, Inter-individual variation in the whole plasma metabolome explained by the indicated factors, estimated using the PERMANOVA method. All, all of the indicated factors combined; smk, smoking status. b, Venn diagram indicating the number of metabolites whose inter-individual variation was significantly explained by diet, genetics or the gut microbiome, as estimated using the linear regression method (FDRF-test<0.05). c, Inter-individual variations in metabolites explained by diet, genetics or the gut microbiome, as estimated using the linear regression method (the lasso regression method was applied for feature selection) with a significant estimated adjusted r2>5% (FDRF-test<0.05). The blue bars represent dietary contributions to metabolite variations, the yellow bars indicate genetic contributions and the orange bars indicate microbial contributions. The other colors indicate the metabolic categories of metabolites (see legend). The yaxis indicates the proportion of variation explained. TMAO, trimethylamine N-oxide.
Next, we tested for pairwise associations between each metabolite and the dietary variables, genetic variants and microbial taxa. We observed 2,854 associations with dietary habits (Supplementary Table 5), 48 associations with 40 unique genetic variants (metabolite quantitative trait loci (mQTLs); Supplementary Table 6), 1,373 associations with gut bacterial species (Supplementary Table 7) and 2,839 associations with bacterial MetaCyc pathways (Supplementary Table 8) (see the Methods sections Associations with dietary habits, QTL mapping and Microbiome-wide associations). In total, 769 metabolites were significantly associated with at least one factor (Fig. 1b and Supplementary Tables 58). We then performed interaction analysis to assess the role of dietmicrobiome, geneticsmicrobiome and dietgenetics interactions in regulating the human metabolome using an interaction term in the linear model (see the Methods section Interaction analysis). Among these, 185 metabolites were associated with multiple factors and seven were affected by either geneticsmicrobiome, geneticsdiet or dietmicrobiome interactions (Supplementary Table 9).
As interactions were limited, we further assessed the proportion of variance of each metabolite that was explained by these factors using an additive model with the least absolute shrinkage and selection operator (lasso) method (see the Methods section Estimating the variance of individual metabolites). In general, the inter-individual variations in 733 metabolites could be explained by at least one of the three factors (FDRF-test<0.05; Supplementary Table 10). In detail, dietary habits contributed 0.435% of the variance in 684 metabolites; microbial abundances contributed 0.725% of the variance in 193 metabolites; and genetic variants contributed 328% of the variance in 44 metabolites (adjusted r2; FDRF-test<0.05; Supplementary Table 10). We also estimated the explained variance of metabolites using Elastic Net14, which is designed for highly correlated features, and found that the estimated explained variances were comparable between linear regression and the Elastic Net regression (Supplementary Fig. 1).
We further compared the variance explained by each type of factor (diet, genetics or the microbiome) and assigned the dominant factor for each metabolite if one factor explained more variance than the other two. Inter-individual variations in 610 metabolites were mostly explained by diet, 85 were explained by the gut microbiome and 38 were explained by genetics (Supplementary Table 10). Hereafter, we refer to these as diet-dominant, microbiome-dominant and genetics-dominant metabolites, respectively. The dominant factors of metabolites highlight their origin. For instance, ten out of the 21 diet-dominant metabolites for which diet explained >20% of the variance (FDRF-test<0.05; Supplementary Table 10) were food components based on their annotation in the Human Metabolome Database (HMDB)15. Similarly, of the 85 microbiome-dominant metabolites, 23 were annotated in the HMDB as microbiome-related metabolites (including 15 uremic toxins). Furthermore, out of the 38 genetics-dominant metabolites, ten were lipid species and eight were amino acids. Taken together, our analysis highlights that one factoreither dietary, genetic or microbialcan have a dominant effect over the other two in explaining the variances of plasma metabolites, with diet or the microbiome being particularly dominant. However, we also found that the variances in 185 metabolites were significantly attributable to more than one factor (Supplementary Table 10), including six metabolites associated with both genetics and the microbiome and 153 metabolites associated with both diet and the microbiome. For example, genetics and the microbiome explained 4 and 5%, respectively, of the variance in plasma 5-carboxy--chromanol (Fig. 1c)a dehydrogenated carboxylate product of 5-hydroxy--tocopherol16 that may reduce cancer and cardiovascular risk17. Another example is hippuric acida uremic toxin that can be produced by bacterial conversion of dietary proteins18, with 13% of its variance explained by diet and 13% explained by the microbiome (Fig. 1c).
Temporal changes in plasma metabolites can reflect changes in an individuals diet, gut microbiome and health status. When assessing the plasma metabolome in the 311 LLD1 follow-up samples, we indeed observed a significant shift in the plasma metabolome, with a significant difference in the second principal component (PPC1 paired Wilcoxon=0.1 and PPC2 paired Wilcoxon=1.3105; Fig. 2a). Baseline genetics, diet and microbiome, together with age, sex and BMI, could explain 59.4% of the variance in the follow-up plasma metabolome (PPERMANOVA=0.004) (Supplementary Fig. 2). We also observed that temporal stability can vary substantially between different metabolites (see the Methods section Temporal consistency of individual metabolites; Supplementary Table 11). Previously, we had assessed the changes in the gut microbiome in the LLD1 follow-up cohort and linked these to changes in the plasma metabolome7. Here, we further checked the temporal variability of the plasma metabolome and assessed the stability of diet-, microbiome- and genetics-dominant metabolites over time. Interestingly, the temporal correlation of the microbiome-dominant metabolites was similar to that of the genetics-dominant metabolites (PWilcoxon=0.51; Fig. 2b), whereas the temporal correlation between diet-dominant metabolites was significantly lower than between microbiome- and genetics-dominant metabolites (PWilcoxon<3.4105; Fig. 2b). However, the dominant dietary, microbial and genetic factors identified at baseline also explained similar variance in metabolic levels in the follow-up samples (Extended Data Fig. 3 and Supplementary Table 10). Our data also revealed a positive correlation between stability and the amount of variance that could be explained: the more variance explained, the more stable a metabolite is over time (Fig. 2c). For a few metabolites, we could not replicate the variance explained at baseline at the second time point, and these metabolites also showed weak or no correlation in their abundances between the two time points. For example, N-acetylgalactosamine showed very weak correlation between the two time points (r=0.13; P=0.02), and its genetic association was not replicated at the second time point.
a, Principal component analysis of metabolite levels at two time points (Euclidean dissimilarity). The green dots indicate baseline samples and the orange dots indicate follow-up samples (n=311 biologically independent samples). The KruskalWallis test (two sided) was used to check differences between baseline and follow-up. b, Temporal stability of metabolites stratified by the dominantly associated factor for each metabolite. The Wilcoxon test (two sided) was used to check the differences between groups. Each dot represents one metabolite. The yaxis indicates the Spearman correlation coefficient of abundances of each metabolite between two time points (n=311 biologically independent samples). In a and b, the box plots show the median and first and third quartiles (25th and 75th percentiles) of the first and second principal components (a) or correlation coefficients (b); the upper and lower whiskers extend to the largest and smallest value no further than 1.5 the interquartile range (IQR), respectively; and outliers are plotted individually. c, Correlation between metabolite stability and the metabolite variance explained by diet (left), genetics (middle) and the microbiome (right). The xaxis indicates the inter-individual variation explained by each factor and the yaxis indicates the Spearman correlation coefficient (two sided) of abundances of each metabolite between the two time points. The dashed white lines show the best fit and the gray shading represents the 95% confidence interval (CI) (n=311 biologically independent samples).
Having established the variances in metabolites explained by diet, genetics and the gut microbiome and the dominant factors that explained most of this variance, we focused on detailing specific associations and on the potential implications of our findings for assessing diet quality and improving our understanding of the genetic risk of complex diseases and the interaction and causality relationships among diet, the microbiome, genetics and metabolism.
We observed 2,854 significant associations (FDRSpearman<0.05) between 74 dietary factors and 726 metabolites (Fig. 3a and Supplementary Table 5; see the Methods section Lifelines diet quality score prediction). Associations with food-specific metabolites can, in theory, be used to verify food questionnaire data. For instance, the strongest association we observed was between quinic acid levels and coffee intake (rSpearman=0.54; P=1.61080; Fig. 3b). Quinic acid is found in a wide variety of different plants but has a particularly high concentration in coffee. Another example is 2,6-dimethoxy-4-propylphenol, which was strongly associated with fish intake (rSpearman=0.53; P=1.51076; Fig. 3c). This association is expected as this compound is particularly present in smoked fish according to HMDB annotation15. In addition, we also detected associations between dietary factors and metabolic biomarkers of some diseases. For example, 1-methylhistidine is a biomarker for cardiometabolic diseases including heart failure19 that is enriched in meat, and we observed significant associations between 1-methylhistidine and meat (rSpearman=0.12; P=7.2105) and fish intake (rSpearman=0.11; P=3.1104) as well as a lower level of 1-methylhistidine in vegetarians (rSpearman=0.15; P=9.7107; Fig. 3d).
a, Summary of the associations between diet and metabolites. The bars represent dietary habits, with the bar order sorted by the number of significant associations. Association directions are colored differently: orange indicates a positive association, whereas blue indicates a negative association. The length of each bar indicates the number of significant associations at FDR<0.05 (Spearman; two sided). b, Association between plasma quinic acid levels and coffee intake. The x and yaxes indicate residuals of coffee intake and the metabolic abundance after correcting for covariates, respectively (n=1,054 biologically independent samples). c, Association between plasma 2,6-dimethoxy-4-propylphenol levels and fish intake frequency (n=1,054 biologically independent samples). The x and yaxes refer to residuals of fish intake and metabolic abundance after correcting for covariates, respectively. d, Differential plasma levels of 1-methylhistidine between vegetarians and non-vegetarians (n=1,054 biologically independent samples). The yaxis indicates normalized residuals of metabolic abundance. The Pvalue from the Wilcoxon test (two sided) is shown. The box plots show the median and first and third quartiles (25th and 75th percentiles) of the metabolite levels. The upper and lower whiskers extend to the largest and smallest value no further than 1.5 the IQR, respectively. Outliers are plotted individually. e, Association between the diet quality score predicted by the plasma metabolome (yaxis) and the diet quality score assessed by the FFQ (xaxis) (n=237 biologically independent samples). In b, c and e, each gray dot represents one sample, the dark gray dashed line shows the linear regression line and the gray shading represents the 95% CI. In b and c, the association strength was assessed using Spearman correlation (two sided; the correlation coefficient and Pvalue are reported) and in e, the prediction performance was assessed with linear regression (F-test; two sided; the adjusted r2 value and Pvalue are reported).
Given the relationship between diet, metabolism and human health, we wondered whether the plasma metabolome could predict diet quality. For each of the Lifelines participants, we constructed a Lifelines Diet Score based on food frequency questionnaire (FFQ) data that reflected the relative diet quality based on dietdisease relationships8. To build a metabolic model to predict an individuals diet quality, we used LLD1 as the training set and LLD2 as the validation set. The resulting metabolic model included 76 metabolites, 51 of which were dominantly associated with diet. The diet score predicted by metabolites showed a significant association with the real diet score assessed by the FFQ in the validation set (r2adjusted=0.27; PF-test=3.5105; Fig. 3e). We also tested four other dietary scores (the Alternate Mediterranean Diet Score20, Healthy Eating Index (HEI)21, Protein Score22 and Modified Mediterranean Diet Score23) and found that the HEI predicted by plasma metabolites was also significantly associated with the FFQ-based HEI (r2adjusted=0.23; PF-test=6.5105; Supplementary Table 12).
Genetic associations of plasma metabolites may provide functional insights into the etiologies of complex diseases. After correcting for the first two genetic principal components, age, sex, BMI, smoking, 78 dietary habits, 40 diseases and 44 medications, QTL mapping in LLD1 identified 48 study-wide, independent genetic associations between 44 metabolites and 40 single-nucleotide polymorphisms (SNPs) (PSpearman<4.21011; clumping r2=0.05; clumping window=500kilobases (kb); Fig. 4a and Supplementary Table 6). All 48 genetic associations were replicated in either LLD1 follow-up or the two independent replication datasets (LLD2 and GoNL; Supplementary Fig. 3 and Supplementary Table 6). We also assessed the impact of physical activity, as assessed by questionnaires24, on the genetics association of metabolism, but found its influence to be negligible (Supplementary Fig. 4). Functional mapping and annotation (FUMA) of genome-wide association studies (GWAS)25 analysis revealed that the identified mQTLs were enriched in genes expressed in the liver and kidney (Extended Data Fig. 4) and related to metabolic phenotypes (Supplementary Table 6).
a, Manhattan plot showing 48 independent mQTLs identified linking 44 metabolites and 40 genetic variants with P<4.21011 (Spearman; two sided). Representative genes for the SNPs with significant mQTLs are labeled. b, Association between a tag SNP (rs1495741) of the NAT2 gene and plasma AFMU levels. c, Association between a SNP (rs13100173) within the HYAL3 gene and plasma levels of N-acetylgalactosamine-4-sulfate. d, Association between a tag SNP (rs17789626) of the SCLT1 gene and plasma mizoribine levels. e, Differences in coffee intake between participants with different genotypes at rs1495741. f, Correlations between coffee intake and AFMU in participants with different genotypes at rs1495741. g, Differences in bacterial fatty acid -oxidation pathway abundance in participants with different genotypes at rs67981690. h, Correlations between bacterial fatty acid -oxidation pathway abundance and 5-carboxy--chromanol in participants with different genotypes at rs67981690. In be and g, the xaxis indicates the genotype of the corresponding SNP and the yaxis indicates normalized residuals of the corresponding metabolic abundance (n=927 biologically independent samples). Each dot represents one sample. The box plots show the median and first and third quartiles (25th and 75th percentiles) of the metabolite levels. The upper and lower whiskers extend to the largest and smallest value no further than 1.5 the IQR, respectively. Outliers are plotted individually. The association strength is shown by the Spearman correlation coefficient and corresponding Pvalue (two sided). In f and h, the xaxis indicates the normalized abundance of coffee intake (f) or the bacterial fatty acid -oxidation pathway (h) and the yaxis indicates the normalized residuals of the corresponding metabolic abundance. Each dot represents one sample (n=927 biologically independent samples). The lines indicate linear regressions for each genotype group separately. Areas with light gray shading indicate the 95% CI of the linear regression lines. The association strength per genotype is shown by the Spearman correlation and the corresponding Pvalue (two sided).
The strongest association we found was between the caffeine metabolite 5-acetylamino-6-formylamino-3-methyluracil (AFMU) and SNP rs1495741 near the N-acetyltransferase 2 (NAT2) gene (rSpearman=0.52; P=1.71066; Fig. 4b), which showed strong linkage disequilibrium (r2=0.98) with a SNP, rs35246381, that was recently reported to be associated with urinary AFMU26. AFMU is a direct product of NAT2 activity and has been associated with bladder cancer risk27. Interestingly, the plasma level of AFMU was associated not only with coffee intake (rSpearman=0.29; P=9.21022; Supplementary Table 5) and the genotype of rs1495741, but also with their interactions (Supplementary Table 9). Individuals with a homologous AA genotype had a similar level of coffee intake, but their correlation between coffee intake and plasma AFMU level was significantly lower compared with individuals with GG and GA genotypes (Fig. 4e,f).
Pleotropic mQTL effects were also observed at several loci, including SLCO1B1, FADS2, KLKB1 and PYROXD2 (Supplementary Table 6). For example, three associations (related to three metabolites, two of them lipids) were observed for two SNPs (rs67981690 and rs4149067; linkage disequilibrium r2=0.72 in Northern Europeans from Utah) in SLCO1B1, which encodes the solute carrier organic anion transporter family member 1B1. Expression of the SLCO1B1 protein is specific to the liver, where this transporter is involved in the transport of various endogenous compounds and drugs, including statins28, from blood into the liver. The SLCO1B1 locus has also been linked to plasma levels of fatty acids and to statin-induced myopathy29. Furthermore, we detected a geneticsmicrobiome interaction between rs67981690 and microbial fatty acid oxidation pathways in regulating plasma levels of 5-carboxy--chromanol (P=1.5103), where the association of the bacterial fatty acid oxidation pathway with plasma levels of 5-carboxy--chromanol was dependent on the genotype of rs67981690 (Fig. 4g,h).
To identify novel mQTLs, we performed a systematic search of all published mQTL studies from 2008 onwards (Supplementary Table 13). This approach identified three novel mQTLs in our datasets (Supplementary Table 13) that were either not located close to previously reported mQTLs (distance>1,000kb) or not in linkage disequilibrium (r2<0.05). The first two novel SNPsrs13100173 at HYAL3 and rs11741352 at ARSBwere associated with N-acetylgalactosamine-4-sulfate (Fig. 4c,d), which is associated with mucopolysaccharidosis30. Interestingly, N-acetylgalactosamine-4-sulfate can bind to HYAL proteins (HYAL1, HYAL2, HYAL3 and HYAL4), suggesting that mQTLs can also pinpoint potential metaboliteprotein interactions. The third novel mQTL was rs17789626 at SCLT1, which was associated with mizoribinea compound used to treat nephrotic syndrome31.
We established 4,212 associations between 208 metabolites and 314 microbial factors (114 species and 200 MetaCyc pathways) (FDRLLD1<0.05; PLLD1 follow-up<0.05; Supplementary Tables 7 and 8). Interestingly, many of the metabolites that were associated with microbial species and MetaCyc pathways are also known to be gut microbiome related based on their HMDB annotations15. For instance, we observed 919 associations with 25 uremic toxins, 142 associations with thiamine (vitamin B1) and 117 associations with five phytoestrogens (FDR<0.05; Supplementary Tables 7 and 8). Uremic toxins and thiamine have been shown to be related to various diseases, including chronic kidney disease and cardiovascular diseases32,33. Phytoestrogens are a class of plant-derived polyphenolic compounds that can be transformed by gut microbiota into metabolites that promote the hosts metabolism and immune system33,34.
To assess whether gut microbiome composition causally contributes to plasma metabolite levels, we carried out bi-directional MR analyses (see the Methods section Bi-directional MR analysis). Here, we focused on the 37 microbial features that were associated with at least three independent genetic variants at P<1105 and with 45 metabolites (Supplementary Table 14). At FDR<0.05 (corresponding to P=2103 obtained from the inverse variance weighted (IVW) test)35, we observed four potential causal relationships at baseline that could also be found in the follow-up in the microbiomes to metabolites direction (Fig. 5ad and Supplementary Tables 15 and 16) but not in the opposite direction (Supplementary Table 17), and these outcomes were maintained following weighted median testing (P<0.03; Supplementary Fig. 5). To ensure that the data followed MR assumptions, we performed several sensitivity analyses, including checking for horizontal pleiotropy (MR-Egger36 intercept P>0.05; Supplementary Table 15) and heterogeneity (Cochrans Q test P>0.05; Supplementary Table 15) and leave-one-out analysis (Extended Data Fig. 5). We did not use causal estimates derived using the MR-Egger method to filter the results, as its power to detect causality is known to be low36. These sensitivity checks further confirmed the reliability of these four MR causal estimates.
a, Analysis of the association between adenosylcobalamin biosynthesis pathway abundance and 5-hydroxytryptophol levels. b, Glycogen biosynthesis pathway abundance versus 5-sulfo-1,3-benzenedicarboxylic acid levels. c, E. rectale abundance versus hydrogen sulfite levels. d, Veillonella parvula abundance versus 2,3-dehydrosilybin levels. In the top panels of ad, the xaxis shows the SNP exposure effect, and the yaxis shows the SNP outcome effect and each dot represents a SNP. Error bars represent the s.e. of each effect size. The bottom panels of ad, show the MR effect size (center dot) and 95% CI for the baseline (blue) and follow-up (green) datasets of the LLD1 cohort, estimated with the IVW MR approach (two sided) (n=927 biologically independent samples at baseline and n=311 biologically independent samples at follow-up).
We further found that increased abundance of microbial adenosylcobalamin biosynthesis (coenzyme B12) was associated with reduced plasma levels of 5-hydroxytryptophol (Fig. 5a)a uremic toxin related to Parkinsons disease37. We also found that plasma hydrogen sulfite levels were related to Eubacterium rectale (Fig. 5c)a core gut commensal species38 that is highly prevalent (presence rate=97%) and abundant (mean abundance=8.5%) in both our cohorts and in other populations39,40,41. As a strict anaerobe, E. rectale promotes the hosts intestinal health by producing butyrate and other short-chain fatty acids from non-digestible fibers42, and a reduced abundance of this species has been observed in subjects with inflammatory bowel disease39,43 and colorectal cancer44 compared with healthy controls. As a toxin, hydrogen sulfite interferes with the nervous system, cardiovascular functions, inflammatory processes and the gastrointestinal and renal system45. Our results thus reveal a potential new beneficial effect of E. rectale.
To further investigate the metabolic potential of individual bacterial species, we applied newly developed pipelines to identify microbial primary metabolic gene clusters (gutSMASH pathways)46 and microbial genomic structural variants (SVs)47. These two tools profile microbial genomic entities that are implicated in metabolic functions. By associating 1,183 metabolites with 3,075 gutSMASH pathways and 6,044 SVs (1,782 variable SVs (vSVs) and 4,262 deletion SVs (dSVs); see Methods), we observed 23,662 associations with gutSMASH pathways and 790 associations with bacterial SVs (FDRLLD1<0.05; PLLD1 follow-up<0.05; Supplementary Tables 1820). These associations connect the genetically encoded functions of microbes with metabolites, thereby providing putative mechanistic information underlying the functional output of the gut microbiome. In one example, we observed that the microbial uremic toxin biosynthesis pathways, including the glycine cleavage pathway (in Olsenella and Clostridium species) and the hydroxybenzoate-to-phenol pathway (in Clostridium species) responsible for hippuric acid and phenol sulfate biosynthesis, were associated with the hippuric acid (Olsenella species: rSpearman=0.15; P=9.3107; Clostridium species: rSpearman=0.18; P=5.9109) and phenol sulfate (rSpearman=0.17; P=4.2108; Extended Data Fig. 6a) levels measured in plasma, respectively (FDRLLD1<0.05 and PLLD1 follow-up<0.05; Extended Data Fig. 6b).
Next, we carried out a mediation analysis to investigate the links between diet, the microbiome and metabolites. For 675 microbial features that were associated with both dietary habits and metabolites (FDR<0.05), we applied bi-directional mediation analysis to evaluate the effects of microbiome and metabolites for diet (see the Methods section Bi-directional mediation analysis). This approach established 146 mediation linkages: 133 for the dietary impact on the microbiome through metabolites and 13 for the dietary impact on metabolites through the microbiome (FDRmediation<0.05 and Pinverse-mediation>0.05; Fig. 6a,b and Supplementary Table 21). Most of these linkages were related to the impact of coffee and alcohol on microbial metabolic functionalities (Fig. 6a).
a, Parallel coordinates chart showing the 133 mediation effects of plasma metabolites that were significant at FDR<0.05. Shown are dietary habits (left), plasma metabolites (middle) and microbial factors (right). The curved lines connecting the panels indicate the mediation effects, with colors corresponding to different metabolites. freq., frequency; PFOR, pyruvate:ferredoxin oxidoreductase; OD, oxidative decarboxylation; HGD, 2-hydroxyglutaryl-CoA dehydratase; TPP, thiamine pyrophosphate. b, Parallel coordinates chart showing the 13 mediation effects of the microbiome that were significant at FDR<0.05. Shown are dietary habits (left), microbial factors (middle) and plasma metabolites (right). For the microbial factors column, number ranges represent the genomic location of microbial structure variations (SVs) in kilobyte unit, and colons represent the detailed annotation of certain gutSMASH pathway. c, Analysis of the effect of coffee intake on the abundance of M. smithii as mediated by hippuric acid. d, Analysis of the effect of beer intake on the C. methylpentosum Rnf complex pathway as mediated by hulupinic acid. e, Analysis of the effect of fruit intake on urolithin B in plasma as mediated by a vSV in Ruminococcus species (300305kb). In ce, the gray lines indicate the associations between the two factors, with corresponding Spearman coefficients and Pvalues (two sided). Direct mediation is shown by a red arrow and reverse mediation is shown by a blue arrow. Corresponding Pvalues from mediation analysis (two sided) are shown. inv., inverse; mdei., mediation.
Coffee contains various phenolic compounds that can be converted to hippuric acid by colonic microflora48. Hippuric acid is an acyl glycine that is associated with phenylketonuria, propionic acidemia and tyrosinemia49. We observed that hippuric acid can mediate the impact of drinking coffee on Methanobrevibacter smithii abundance (Pmediation=2.21016; Fig. 6c). We also observed that hulupinic acid, which is commonly detected in alcoholic drinks, can mediate the impact of beer consumption on the Clostridium methylpentosum ferredoxin:NAD+ oxidoreductase (Rnf) complex (Pmediation=2.21016; Fig. 6d)an important membrane protein in driving the ATP synthesis essential for all bacterial metabolic activities50.
Of the dietary impacts on metabolites through the microbiome (Fig. 6b and Supplementary Table 21), one interesting example is a Ruminococcus species vSV (300305kb) that encodes an ATPase responsible for transmembrane transport of various substrates51. This Ruminococcus species vSV mediated the effect of fruit consumption on plasma levels of urolithin B (Pmediation=2.21016; Fig. 6e). Urolithin B is a gut microbiota metabolite that protects against myocardial ischemia/reperfusion injury via the p62/Keap1/Nrf2 signaling pathway52. Taken together, our data provide potential mechanistic underpinnings for dietmetabolite and dietmicrobiome relationships.