Questioning an ecosystem

One of the main challenges in gut microbiota studies is to grasp how the host responds to modifications of a complex bacterial ecosystem while considering inter-individual environmental variation.¹⁰ This challenge requires integration of multiple-level interactions between the host system variables and characteristics of the bacterial taxa, and one approach to tackle this challenge is addressed by multivariate statistical tools. Indeed, bacterial processes are influenced by the surrounding environment that can impact gene expression and thereafter also impact other microorganisms. If a bacterial taxon impacts the health status of the host under specific environmental conditions, then the impacts of this bacterial unit should also be studied taking into account the status of other members of the ecosystem. An example that justifies this approach is the increasing evidence that pathogen colonization is not necessarily followed by symptom development.¹¹ Pathogenicity is generally observed when colonization is concomitant with dysbiosis,¹² demonstrating that colonization by a pathogen is often not sufficient to trigger disease development. Thus, pathogenicity and infection need to be considered together with potentially co-occurring modifications of the bacterial ecosystem. Furthermore, a literature search will often reveal reports of identical taxa having opposite effects on host health.¹³ For instance, Lactobacillus that have been the object of many studies due to their important role in food production have been reported to have opposite impacts on obesity.¹⁴ These apparently discordant findings could be explained by associated modifications of the surrounding microbial environment.

Here, we therefore explore once again the association between VFM and gut microbiota composition, but now aiming to take into account the impacts of bacterial taxa within their ecosystem as a whole using multivariate analysis in the TwinsUK cohort.^15-17 These analyses were performed in the original discovery sample of 1313 twins from Beaumont et al.,⁹ who had available gut microbiota profiles and VFM phenotype estimates. The distribution of the phenotype in this data set is shown in Figure 1A. Collapsing of 16S rRNA gene sequencing (16S) data to OTUs in the data set of 1313 individuals was pursued here using an alternate approach compared with that of Beaumont et al.⁹ Specifically, our previous study used open reference clustering with Greengenes v13_8 at 97% sequence similarity, while the present analysis was based on SUMACLUST De Novo clustering in QIIME at a similarity threshold of 97%, with later taxonomic assignment against the same reference database.¹⁸ The purpose of this approach was to assess if the results are reproducible at the OTU level when changing the method for quantification, as there is evidence for differential accuracy between reference based and de novo clustering of OTUs. Several studies have compared the use of reference based and de novo clustering approaches, with some evidence that de novo clustering produces OTUs more accurately clustered by sequence identity.¹⁹ Carrying out comparisons between methods within the TwinsUK data we also found that de novo methods, in particularly SUMACLUST, produced more heritable OTUs.¹⁸ Beaumont et al.⁹ utilized open reference clustered OTUs, combining reference and de novo clustering. We chose to use the SUMACLUST de novo OTUs in the present study given the evidence regarding the differences in clustering approaches and to demonstrate robustness of our results in respect to these differences.

Open in a separate window

Figure 1.

VFM is strongly associated with gut microbiota composition. A. VFM distribution within the extended TwinsUK data set (n = 1313). B. Plot of the observed scores against the cross-validated scores generated by the O-PLS DA calculated using VFM as a predictor and 155 OTUs (belonging to the Bacteroides, Clostridium, Blautia, Coprococcus, Lachnospira, Oscillospira and Ruminococcus genera) relative abundance as a matrix of independent responders, each score represents one of the 1313 individuals used to generate the model and is color-coded in accordance to individual VFM level. C. Diversity distribution of the 155 OTUs used to generate the O-PLS DA model and of the 26 OTUs that were considered as significantly associated to VFM from the same model. D. Association loadings between VFM and the 14 significant OTUs; genera highlighted in blue are OTUs that were negatively associated with VFM in the Beaumont et al paper, while those in red are positively associated, and purple represents a mixture of positive and negative associations.

Beaumont et al.⁹ identified several OTUs that were significantly associated with adiposity phenotypes. All of these OTUs were classified under the 7 following bacterial genera Ruminococcus, Blautia, Oscillospira, Clostridium, Lachnospira, Coprococcus and Bacteroides. To evaluate if these results were consistent in the updated analyses, we extracted all OTUs classified under these 7 genera in the set of 1313 individuals. Out of the 582 OTUs considered in total that were observed in at least 25% of the extended TwinsUK sample (2730 individuals), 155 belonged to one of the 7 genera of interest. Scores for which relative abundance was equal to zero were adjusted to 0.000001 before log transformation. In the current study, we used the log10 transformed relative abundance of 155 OTUs (corrected for sequencing run, sequencing depth, who extracted the DNA, who loaded the DNA and sample collection method) as independent variables in an orthogonal projection to latent structure discriminant analysis (O-PLS DA) using VFM phenotype as a predictor in the 1313 individuals (Fig. 1B). O-PLS DA is a supervised analysis that allows us to determine if a phenotype (for example, VFM) can be used as a predictor of a multivariable system (for example, the gut microbiota).²⁰ Model validity can be assessed by evaluating the goodness of fit and of prediction. Further, it is possible to extract from this model the variables (for example, OTUs) that are significantly contributing to the model. Here, we used an in-house algorithm (provided by Korrigan Sciences Ltd) in MatLab (version R2016b, The MathsWorks inc.), with 7 cross-validations for prediction and zero orthogonal component. To assess the fit of the model we report the observed goodness of fit (R²Y = 0.065) and goodness of prediction (Q²Y = 0.0338) values. We also assessed the fit of the model by permutation, where based on 10000 permutations we report significant estimates for goodness of fit (P-value = 0.0001). Association between specific OTUs and O-PLS DA model scores were further explored by fitting a linear mixed effects regression (LMER) model using the R package lme4. In the analyses described in this section we included several additional covariates for the 16S profiles, including sex, age and long-term summary dietary profiles (see Beaumont et al.⁹) and BMI (fixed effects), and family and zygosity (random effects). A Benjamini–Hochberg false discovery rate (FDR) correction was applied and LMER results are reported at FDR of 1%.

The central role of the Firmicutes

Altogether, 14 of the 155 OTUs used to generate the model appear to significantly contribute to VFM prediction at FDR 1% in the extended data set of 1313 individuals from TwinsUK. When considering OTU diversity of the 155 OTUs used to generate the O-PLS DA model, it appears that over a third of OTUs belonged to the Bacteroidetes phylum and the rest were classified as Firmicutes (Fig. 1C). However, OTU taxonomic classification proportion was considerably modified when considering just the 14 OTUs that were significantly associated with VFM in Figure 1C. Within these 14 OTUs 88% belonged to the Firmicutes phylum and only 12% were Bacteroidetes. The reduction in VFM-associated proportion of Bacteroidetes in the O-PLS DA model compared with the linear regression results is somewhat surprising and suggests a potential redundancy in VFM-associated effects across some Bacteroidetes members. It has been consistently observed that an increase in Bacteroidetes results in or is associated with lean phenotypes.^5,6,21 Our results are in line with previous findings as we also observed a negative association between VFM and Bacteroidetes OTUs, similar to previously published analysis by Beaumont et al.⁹

The analyses may also in part reflect incomplete coverage of bacterial composition and diversity by 16S data, while technologies such as shot gun metagenomics may allow us to detect more subtle and accurate changes of the ecosystem.²² However, the modification in diversity balance indicates the central role occupied by Firmicutes in energy balance and expenditure reflected by adiposity phenotypes that was also observed in Beaumont et al.⁹

In terms of considering GM deleterious or protective effects on cardio-metabolic disease risk, the overall the direction of association between OTUs and VFM were comparable to results previously published in Beaumont et al. (Fig. 1D). As previously observed, OTUs belonging to the Bacteroides,Oscillospira, Lachnospira and Coprococcus genera were negatively associated with VFM. However, none of the OTUs belonging to the Clostridium genus appeared to be significantly associated with VFM. Similarly, all OTUs belonging to the Blautia genus were positively associated with VFM both in the original and current analysis. Lastly, only negative associations between VFM and OTUs within the Ruminococcus genus were observed here, while both positive and negative associations were reported in the previous study.

The most associated components with VFM in the new extended analyses was a class of Oscillospira genus members. Of the 155 OTUs used to generate the initial linear regression model, 25 belonged to this genus (16%), demonstrating that this genus if replicated could be a key player of the host-GM mutualism in relation to fat deposition.

The potential implication of these 4 potentially “beneficial” genera (Bacteroides, Oscillospira, Lachnospira and Coprococcus) needs to be explored in further studies using animal models and human intervention. It is difficult to assess if the reported adiposity phenotype associations are a cause or a consequence of the gut microbiota modifications, and follow up work using animal and in vitro models may shed a light on specific mechanisms involved in the host-gut microbiota cross talk.²³ For example, germ-free mouse models have shown that a GM community associated with health status could induce specific phenotypes within recipient animals.²⁴ However, animal studies may also have some drawbacks in terms of translating the findings to human physiology and pathology²⁵ as they cannot fully recapitulate the symbiotic human host-bacterial co-evolution.²⁶ Nevertheless, results from the extended data set confirm the close links between the gut microbiota and fat metabolism and storage. The Beaumont et al.⁹ study also explored the potential impact of host genetics on these interactions. This topic is addressed in the following section where the mechanisms by which the host might be able to modulate the microbiota will be discussed, and linked to obesity.