Tipping elements in the human intestinal ecosystem

The microbial communities living in the human intestine can have profound impact on our well-being and health. However, we have limited understanding of the mechanisms that control this complex ecosystem. Here, based on a deep phylogenetic analysis of the intestinal microbiota in a thousand western adults, we identify groups of bacteria that exhibit robust bistable abundance distributions. These bacteria are either abundant or nearly absent in most individuals, and exhibit decreased temporal stability at the intermediate abundance range. The abundances of these bimodally distributed bacteria vary independently, and their abundance distributions are not affected by short-term dietary interventions. However, their contrasting alternative states are associated with host factors such as ageing and overweight. We propose that the bistable groups reflect tipping elements of the intestinal microbiota, whose critical transitions may have profound health implications and diagnostic potential.

Temporal dynamics of a bi-stable abundance distribution The logarithmic abundance distribution for Dialister spp. exhibits two peaks of low and high abundance with the estimated state frequencies of n 1 =72% and n 2 =28% across the 1,006 western adults, respectively. In stationary state, the flow between the two states is balanced, and the ratio of the switching rates r should be inversely related to state frequencies n (r 1 /r 2 = n 2 /n 1 ) assuming a stationary continuous-time Markov process. Hence the more frequent states are relatively more stable in the stationary state, while the absolute switching rates determine the overall mixing between the states during a given time interval. The tipping point (dashed  Table 3). Uncultured bacterium clones Eldhufec308 and Eldhufec312 exhibited clear bimodality ( Supplementary Fig. 4). Similar abundance distributions characterized also the phylotypes within the UCII group, which could not be identified further down than order Clostridiales.

Supplementary Note 2 Cross-hybridization control
Cross-hybridization can reduce the accuracy of observations in microarray analyses. We controlled this based on pre-calculated cross-alignment tables between the taxonomic groups targeted by the HITChip microarray.
Cross-hybridization was negligible (<10%) between the bimodal groups and other taxa. The B. fragilis group, targeted by 40 probes, was an exception with 43% of shared probes with the B. ovatus group. Since this could potentially contribute to the bimodality of the B. fragilis abundance distribution, we investigated the 16 probes that were specific for the B. fragilis group. Bimodal abundance patterns were detected in 25% of the unique probes, suggesting that this group may contain both bimodal and smoothly varying higher-level phylotypes. The highly correlated P. oralis and P. melaninogenica groups had 13-26% shared probes. The correlation between these two groups remained high (0.81) after excluding the shared probes, however, confirming positive association. To avoid potential biases associated with different DNA extraction methods, the correlations between phylogenetic groups were calculated based on the 401 samples analysed with the mechanical lysis (see Methods).

Supplementary Note 3 HITChip phylogeny
The HITChip phylogeny is binned at three taxonomic levels based on 16S sequence similarity, roughly corresponding to species/phylotype (98% 16S sequence similarity; n=1033) and genus (90%; n=130) levels, that are further classified into 23 higher-level groups including 10 phyla. In the present work we primarily focus on the genus level, which is less prone to cross-hybridization between closely related targets than the higher-resolution phylotype-level analysis.

Supplementary Note 4 Community diversity
We quantified the overall community diversity by the Shannon index of the probe-level HITChip data (3631 probes targeting 1033 phylotypes) within the subset of 255 samples that were also analyzed for metagenomic richness (gene count) 2 .

Supplementary Note 5 Bimodality significance
As the standard Potential Analysis approach does not provide p-values for the observed bimodal patterns, we derived empirical pseudo-pvalues for each taxon as the fraction of bootstrap samples that did not support the bimodality to obtain the following estimates of false discovery rates 5 : Prevotella <2%; Dialister spp., UCI and UCII 10%; and B. fragilis 25%.
Supplementary Note 6 DNA extraction bias When all samples with varying DNA extraction methods were included, we observed bimodal patterns also in bifidobacteria and ruminococci that thus appear to be confounded by different extraction methods, highlighting the necessity to take the extraction method into account in meta-analyses that integrate data across independent studies.

Supplementary Note 7 Kaplan-Meier analysis
This is an approximation, however, as the changes in bacterial abundance are continuous and reversible unlike the life/death processes, and natural continuous fluctuations in bacterial abundance may induce some observed state shifts.