Role of Dietary Flavonoid Compounds in Driving Patterns of Microbial Community Assembly

Cohort characteristics.The characteristics of the study population are presented in Table 1. The mean ± standard deviation total-flavonoid intake was 460 ± 303 mg/day, with the highest tertile of high total-flavonoid consumers having more than threefold-greater mean total-flavonoid intake than low total-flavonoid consumers. This biologically substantial difference of 570 mg/day equates to nearly three cups of black tea, nine medium sized oranges, or nearly five glasses of red wine (36, 37). High total-flavonoid consumers were more likely to have their samples collected in winter and be of healthy weight (as defined as having a body mass index of <25 kg/m²). The cohort consisted of only four current smokers, two from the low total-flavonoid consuming group, and two from the moderate consumers. There was no difference in fecal microbial alpha diversity across the different levels of total-flavonoid intake.

Similar to non-U.S. countries (38–40), the flavanol polymer and monomer classes made the greatest contributions to total-flavonoid intake in this cohort, contributing 62% and 12%, respectively (Table 2). Despite the fact that the study population was a U.S. population, the foods contributing most to flavonoid intake are common features of many dietary patterns (41, 42).

Composition of the subclass-specific microbial communities.We first sought to identify subclass-specific communities (SMCs) in order to summarize the species that were differentially more, and less, prevalent in high flavonoid subclass consumers. For all flavonoid subclasses, the greatest degree of discrimination between low and high subclass consumers was observed along eigenvector 1 of the canonical discriminant models, and thus, the eigenvector 1 canonical discriminant coefficients were used to determine which microorganisms comprise SMC scores. This ecosystem-based analysis identified six communities of bacteria, one for each flavonoid subclass.

The Lachnospiraceae (NCBI taxonomy identifier NCBI:txid186803), Prevotellaceae (NCBI:txid171552), Coriobacteriaceae (NCBI:txid84107), and Bacteroidaceae (NCBI:txid815) families were the most bacterial common families to be identified, representing 20%, 10%, 10%, and 9%, respectively, of inclusions into individual SMC scores.

Members of the Bacteriodaceae family made predominantly inverse contributions to subclass-specific microbiome communities. Antimicrobial effects of flavonoids on Bacteriodaceae have been observed in chicks and broilers (43, 44), and tea polyphenols have been reported to reduce the abundance of Bacteriodaceae in the fecal samples of swine (45). In vitro, tea flavonoids and their metabolites have been shown to inhibit the growth of bacteria belonging to the Bacteroides genus (46). Together, these results suggest that this broad antimicrobial effect of flavonoids on Bacteriodaceae is less likely to be linked to specific enzymes, and more likely to occur as a result of broadly conserved structural features in the Bacteriodaceae family.

Of the 73 species identified for inclusion into the SMC scores, 56% (n = 41) represented species that were unique to just one flavonoid subclass, with anthocyanidins having the most nonshared species (Table 3). Anthocyanidins are differentiated from all other flavonoid subclasses by being the only flavonoid subclass to exist naturally in ionic form.

The species identified in each SMC are phylogenetically diverse (Table 3). This is to be expected given that the compounds comprising each flavonoid subclass also have diverse structures, and there are numerous factors dictating how compounds elicit antimicrobial effects. In agreement with our data, a comprehensive review of the BRENDA (47) database (see Table S1 in the supplemental material) reveals that there is great phylogenetic diversity in the microorganisms reported as being capable of flavonoid metabolism. As is the case for flavonoid-microbe literature, the BRENDA database does not provide a comprehensive summary of all bacteria involved in flavonoid metabolism, but rather, it is useful in indicating the diversity of ways in which bacteria metabolize flavonoids. For example, we identified Bifidobacterium adolescentis (NCBI:txid1680) was inversely discriminatory for flavonol intake (Table 3, panel a). The BRENDA database identifies B. adolescentis as a flavanol monomer metabolizer by acting upon sucrose and (+)-catechin to produce d-fructose, (+)-catechin 3′-O-alpha-d-glucoside, and (+)-catechin 3′,5-O-alpha-d-diglucoside. B. adolescentis catalyzes a similar reaction involving (−)-epicatechin to produce (−)-epicatechin 3′-O-alpha-d-glucoside, (−)-epicatechin 5-O-alpha-d-glucoside, and (−)-epicatechin 3′,5-O-alpha-d-diglucoside (47). The precise mechanism by which flavonol compounds interact with B. adolescentis, both enzymatically and structurally, remains unclear.

Eggerthella lenta (NCBI:txid84112) was positively associated with higher intakes of both flavonols and flavanones. Red wine is a major contributor to flavanone intake in this cohort and is also a good source of dietary flavonols. A randomized controlled trial, also in healthy adult men, found that supplementation with red wine increased the fecal content of E. lenta compared to placebo, independent of its alcohol content, and was positively associated with the concentration of red wine metabolites in the urine (48).

Eggerthella lenta is a flavonoid-metabolizing species (49, 50), strains of which are capable of reductively cleaving the heterocyclic C-ring that is characteristic of flavonoid compounds (51). For example, E. lenta rK3 cleaved the heterocyclic C-ring of both (−)-epicatechin and (+)-catechin, giving rise to 1-(3,4-dihydroxy-phenyl)-3-(2,4,6-trihydroxyphenyl)propan-2-ol. This effect does not appear to be strain specific, but rather a feature of E. lenta in general (52). Furthermore, E. lenta JCM 9979 has been shown to convert (−)-epigallocatechin into 1-(3,4,5-trihydroxyphenyl)-3-(2,4,6-trihydroxyphenyl) propan-2-ol (53). Even though the literature shows mechanisms for E. lenta in the context of flavanol monomers, it is reasonable to draw the conclusion that this species is capable of metabolizing flavonoid compounds more broadly.

Conversely, we have direct evidence for other identified species where the literature has investigated the exact subclass species pairs we observed. For example, we found that Adlercreutzia equolifaciens (NCBI:txid 446660) was a positive discriminator for both flavonol and flavanol monomer intake, meaning that A. equolifaciens is present in greater abundance in high, compared to low, consumers of flavonols and flavanol monomers. A. equolifaciens has been shown to metabolize flavanol monomers (49) by catalyzing cleavage of the central C-ring, and then dehydroxylated the 4’ carbon of the resultant metabolite’s B-ring (50, 53–55). Although traditionally thought of as an isoflavone-metabolizing bacterium, more-recent data have demonstrated the capacity of A. equolifaciens to metabolize nonisoflavone compounds (50), further reinforcing the concept that activities of flavonoid-metabolizing bacteria are often not constrained to one single flavonoid subclass.

Inverse discriminators are considered to be species that, when considered a part of a community, are present in smaller relative abundance in high, compared to low, flavonoid subclass consumers. Flavonifractor plautii (NCBI:txid292800) was an inverse discriminator in the microbiome communities for five of the six flavonoid subclasses: flavanol monomers, flavanol polymers, flavanones, flavones, and anthocyanidins. F. plautii has been shown to metabolize many flavonoid compounds from a variety of subclasses (56, 57), providing a biological basis for why one species can contribute to more than one SMC score. However, the relation of flavonoid intake with its relative abundance in fecal samples of humans is unclear, with randomized controlled trials finding no effect of polyphenol-rich diets on F. plautii abundance in healthy adults (58). The role of flavonoids, as distinct from other polyphenolic compounds, in determining the relative abundance of F. plautii requires further elucidation.

Ecosystem-based approach to analyzing the relation of diet (flavonoid subclass intake) with the microbiome.By adopting a bootstrap resampling approach to community identification, weighting all included microorganisms equally, and incorporating only directionality into the model, the SMC score was able to avoid carrying forward artifacts, or unintended patterns, from the original data set. We have previously shown that our profile identification method identifies a reproducible profile that outperforms studies of individual associations in its validity (35). In agreement, this current study also found more consistent associations with the profile identification method compared to individual subclass-microbe regression models. For all flavonoid subclasses, members of the SMC were consistently ranked as positive or inverse discriminators in the bootstrap resamples (Table 3). Sensitivity analyses in which the number of resamples and the random seed were varied did not have a substantial impact on results, adding further evidence to the robustness of the results.

Conversely, partial Spearman correlation models in the sample data set identified 16 associations that remained significantly associated with the respective subclass after controlling for multiple potential confounders. The consistency of the bootstrapped community results, in conjunction with the limited number of microbes that are individually associated with the subclasses, is further support that our community-based analytic ethos is appropriate. The results suggest that flavonoid subclass consumption impacts the microbiome at a community level, rather than at the level of individual species.

When conducting community-based analyses, there is often concern that community membership may be driven by an individual species, or group of species, that is both strongly associated with subclass intake and that contributes substantially to the total species abundance. To test whether this phenomenon was occurring in our analysis, we examined the correlation between all species comprising an individual SMC. As can be seen in Fig. S1 in the supplemental material, there were isolated cases of two species being cocorrelated (P < 0.05), but there was no observable pattern whereby many microbes were highly correlated with an individual species, nor was there any evidence of groups of microbes being highly cocorrelated with each other. As such, we concluded that community membership was not driven by a single organism, or group of organisms, but rather that each member of the community contributed to discrimination between high and low subclass consumers. This reinforces our conclusion that the association of flavonoid subclasses with the microbiome is at the community level, and not at the level of individual species.

Structure of the relationships between flavonoid subclass intake and microbiome composition.The microbiome is an ecosystem comprised of living microorganisms interacting with the human gastrointestinal tract environment in which they reside. As such, we sought to implement a structural equation model aimed at modeling this ecosystem in the context of the nonindependence of human exposure and microbiome variables and the numerous potential, as-yet-unknown, confounders. As outlined in Table 4, despite substantial correlation between the intakes of the various flavonoid subclasses, there were less pronounced correlations between the different values of the SMC scores, which was expected given the large proportion of unique species contributing to each of the microbiome communities of each flavonoid subclass and the method utilized to identify the communities.

The effect of flavonoids on the microbiome is not unidirectional. In fact, compounds from many flavonoid subclasses have been shown to elicit antimicrobial effects on particular microorganisms (31). As such, in addition to incorporating a set of “positive” species (that were in higher relative abundances in high subclass consumers, and lower relative abundance in low flavonoid consumers), the SMC scores also incorporated a set of “inverse” species (that were in lower relative abundances in high flavonoid consumers, and higher relative abundance in low flavonoid consumers). Given the way in which the SMC scores were computed (equation 1), it was our prespecified hypothesis that, compared to low consumers, high subclass consumers would have more of the “positive” microorganisms and fewer of the “inverse” microorganisms. As such, for each subclass, we hypothesized that subclass intake will be positively associated with the value of the corresponding microbiome community score.

The relative importance of diet and lifestyle factors in influencing diet-microbiome associations is poorly understood. When exploring the contribution of individual variables in the multivariate-adjusted model, none of the variables significantly contributed to the model (Table S2). Therefore, in addition to concluding that cocorrelation with other flavonoid subclasses was not a driver of the subclass-specific associations, we were also able to conclude that none of the included covariates significantly altered the diet-microbiome relationship in this analysis. Furthermore, the inclusion of acid-lowering medications and antibiotic medications to the multivariate-adjusted model did not substantially alter results (data not shown).

Even after adjusting for potential confounders and removing the variability that is accounted for by significant covariance with other variables in the flavonoid intake and SMC score data matrices, the intakes of all subclasses remained significantly and positively associated with the values of their respective SMC scores. These results indicate that the SMC scores we identified were indeed reflective of the intake of their respective subclass, and were not simply a marker for flavonoid intake in general, or a diet/lifestyle that is characteristic of high flavonoid consumers.

Although we could not account for unmeasured confounders of the subclass-community association, our structural equation modeling approach did enable us to draw conclusions regarding how well our model reflects the observed data. With the Bentler comparative fit and Bentler-Bonett normed fit (NF) all greater than 0.9, we conclude that the model specifying that the subclasses are the major determinants of their respective communities fits the observed data well and is an appropriate biological model for this cohort.

Relation of whole-food intake with microbiome community scores.To test the robustness of our identified communities and ensure translatability of results, we then sought to explore how the SMC scores were related to the wholefoods contributing most to the intakes of the corresponding subclasses outlined in Table 2. Echoing the highly significant associations observed in the subclass analyses, we observed a strong association of blueberry intake with the value of the anthocyanidin microbiome community score (Table 5). In fact, even after accounting for diet and lifestyle factors that may explain the relationship, blueberry consumption accounted for 14.2% of the variance in the community of 20 species that comprised the anthocyanidin microbiome community score (P < 0.0001). To ensure that this association was due to blueberries themselves, and not because blueberries may serve as a proxy for total anthocyanidin intake, we further adjusted the multivariate-adjusted model for intakes of the other whole foods contributing most to anthocyanidin intake, and the significant association remained: R² = 14.4% and P < 0.0001. The anthocyanidin microbiome community score was characterized by having a lower relative abundance of Clostridium bolteae (NCBI:txid208479), which has been previously been shown to be associated with clinical infection (59).

A similar strong association was observed for tea intake, where even after adjusting for multiple potential confounders and the consumption of other flavonol-rich foods, tea consumption explained 5.9% of the unweighted variance in the 20 species comprising the flavonol microbiome community score (P = 0.001). The flavonol microbiome community was characterized by having lower relative abundance of Faecalibacterium prausnitzii (NCBI:txid853) which has been implicated as having beneficial effects on human health (60), and a greater relative abundance of the flavonoid-metabolizing bacterium E. lenta, which has been associated with frailty in the elderly (61) and inactivates the cardiac drug digoxin (62). It is unclear how the health benefits of flavonol intake interact, as shown by an abundant literature of the benefit of tea consumption (63), with the nonbeneficial features of the community of bacteria associated with high flavonol intake, as well as the other community members where their importance to human health is not yet clearly elucidated.

The percent contribution whole foods make to total intake of a given subclass did not appear to drive the strength of associations. In fact, for the flavanol polymer subclass, although all three major whole-food contributors were significantly associated with the value of the flavanol polymer microbiome community score, the top contributor, tea, accounting for 28.1% of total flavanol polymer intake, explained only 2.6% of the flavone microbiome community score variance. Conversely, 6.0% of total variance in the species comprising the flavanol polymer microbiome community was explained by the consumption of blueberries, which contributed only 11.4% of total dietary flavanol polymers. The association of flavanol polymer-rich foods independent from other flavanol polymer-rich foods highlights the potential effect that flavanol polymer-rich foods, generally, have on the community of 20 microorganisms that collectively make up the flavonol polymer microbiome community score.

Due to the observational nature of this study, as well as the measurement error associated with assessing flavonoid intake and microbiome composition, there is an increased likelihood for type two statistical errors to occur, tending results toward the null hypothesis of no association. Measurement error can also contribute to an underestimation of true effects, resulting in effect estimates that are smaller than what truly occurs in the population. On the other hand, there is the chance that the lack of an external validation cohort may contribute to an overestimation of the true magnitude of effect. However, indications of specificity for certain foods, percent contribution of whole foods to subclass intake not driving results, and significant results remaining even after adjusting for other major whole-food subclass contributors, all point toward a lack of bias and an absence of overestimation. Having said this, future studies validating this model in a variety of populations will be useful in establishing population-specific magnitude of effects.