The hallmarks of dietary intervention-resilient gut microbiome

Abstract

Maintaining equilibrium of the gut microbiome is crucial for human health. Diet represents an important and generally accessible natural channel of controlling the nutrients supply to the intestinal microorganisms. Although many studies showed that dietary interventions can specifically modulate gut microbiome composition, further progress of the approach is complicated by interindividual variability of the microbial community response. The reported causes of this variability include the baseline microbiome composition features, but it is unclear whether any of them are intervention-specific. Here, we applied a unified computational framework to investigate the variability of microbiome response measured as beta diversity in eight various dietary interventions using previously published 16S rRNA sequencing datasets. We revealed a number of baseline microbiome features which determine the microbiome response in an intervention-independent manner. One of the most stable associations, reproducible for different interventions and enterotypes, was a negative dependence of the response on the average number of genes per microorganism in the community-an indicator of the community functional redundancy. Meanwhile, many revealed microbiome response determinants were enterotype-specific. In Bact1 and Rum enterotypes, the response was negatively correlated with the baseline abundance of their main drivers. Additionally, we proposed a method for preliminary assessment of the microbiome response. Our study delineats the universal features determining microbiome response to diverse interventions. The proposed approach is promising for understanding the mechanisms of gut microbiome stability and improving the efficacy of personalised microbiome-tailored interventions.

Fig. 1. Investigation of computational components in the relationship between alpha and beta diversity.

a Scheme of the two randomisation strategies (see Methods), b–d Relationship between average alpha diversity and beta diversity calculated for the 500 random pairs of samples from the FGFP cohort (p values obtained using linear regression). The colours denote initial data and data randomised by the two strategies (see legend for panel a). Alpha diversity was assessed via the Shannon metric, while beta diversity - via Bray–Curtis (b), Aitchison (c) and RC_bray (d) metrics.

Fig. 2. Variation of microbiome composition in the analysed data.

a Distribution of baseline samples in the species abundance space visualised using the UMAP algorithm (n = 641 samples). Colours denote interventions (abbreviations are disclosed in the text). b Distribution of baseline samples by enterotypes. c Microbiome composition shifts occurring during the interventions visualised in the species abundance space using the UMAP algorithm. Grey lines connect paired samples from the same subject collected before and after the intervention (n = 1242 samples). d Intra- and interindividual variation calculated using RC_bray metric for each of the analysed interventions.

Fig. 3. Associations between microbiome response to interventions and baseline microbiome composition.

a Distribution of baseline samples in the species abundance space coloured in the value of response - RC_bray between samples before and after the intervention (UMAP algorithm). b Seven alternative beta diversity metrics were used for the response calculation: Bray–Curtis (BC), Aitchison (Ait), generalised UniFrac (GUni), weighted UniFrac (WUni), unweighted UniFrac (UUni), Jaccard and inverse Pearson correlation (InvCor). Colours denote partial correlation coefficients between the response and each of the analysed features. Asterisks denote significant associations (partial correlations, FDR < 0.05). Inconsistency for associations with alpha diversity between response metrics had been expected due to the computational component. c Enterotype-wise partial correlation networks between the response, baseline alpha diversity, AGN and B:F. The edges width is proportional to the absolute correlation coefficient. Blue colour denotes significant negative associations, red - significant positive and white – insignificant (significance estimated using partial correlations, FDR < 0.05).

Fig. 4. Associations of microbiome response potential with the baseline microbiome composition features and response to the interventions.

a Partial correlation networks between the response potential, response, baseline alpha diversity, AGN and B:F calculated in each enterotype. The edge's width is proportional to absolute correlation coefficient. Blue colour denotes significant negative associations, red - significant positive and white – insignificant (significance estimated using partial correlations, FDR < 0.05). b Relation between predicted and true response (or response potential) for different interventions. Predicted response was calculated based on baseline species abundances using XGBoost machine learning algorithm. c Partial correlation networks between the response potential, baseline alpha diversity, AGN and B:F calculated in each enterotype on the FGFP cohort.

Fig. 5. Relation between the response, baseline alpha diversity and baseline AGN.

Scatterplot showing negative relationship between alpha diversity and AGN, AGN and response and positive relationship between alpha diversity and response for baseline samples from all analysed interventions (n = 641). Variation of data explained by each axis is given in brackets.