When the samples from Costello et al. 6 and the current study are compared directly, the samples cluster by body habitat, showing excellent concordance between the studies (Figure 1a in 6; Figure 1b in the current study) despite differences in sequencing technology (454 and Illumina GA-IIx, respectively), mean read length (229 ± 16 (SD) nucleotides and 123 ± 17 (SD) nucleotides, respectively), and region of the 16S sequenced (V2 and V4, respectively). The UniFrac distances between the 331 time series samples that were sequenced on both Illumina and 454 were significantly correlated, as determined by Procrustes analysis of unweighted UniFrac principal coordinate matrices (M2 = 0.161; Monte Carlo P < 0.001; Additional file 1) and Pearson correlation of UniFrac distances for pairs of samples (r = 0.91; P < 0.001). As observed by Costello et al. 6, gut, oral, and skin bacterial communities were found to be compositionally distinct based on principal coordinates analysis of unweighted UniFrac distances between communities (UniFrac measures community similarity based on the degree to which they share branch length on a phylogenetic tree). The long-term time series shows for the first time that this body-site differentiation is highly stable over greater than one year (Figure 1c), but dynamic within sites over time (Additional files 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15).

While overall compositional differences between body sites and individuals were relatively stable, our data also suggest a surprisingly small temporal 'core human microbiota' within an individual's body sites (Figure 2) when we define the 'core' as those species-level phylotypes in a given body habitat that were observed across all sampling events. These data suggest a minimal core microbiome across time, where the size of the core decreases as: mouth > gut > right palm ≈ left palm > across body sites within an individual > across body sites and individuals.

At this depth of sequencing, many more OTUs are either persistent community members, which appear in a given body habitat and remain for an extended period of time but are not present consistently enough to be considered core members, or transient community members, which appear in a body habitat and disappear soon after (Figure 3). The taxa composing these persistent and transient categories are significantly different (for example, M3 gut: G_indep, 84.78; P = 1.80 × 10^-14). In M3 gut, both the persistent and transient communities are dominated by Clostridia, Bacteroidia, and to a lesser extent Erysipelotrichi. The persistent community is, however, also composed of Betaproteobacteria and Deltaproteobacteria, while the transient community is composed of Actinobacteria, Gammaproteobacteria, Epsilonproteobacteria, and Verrucomicrobiae. Taxonomic summaries of the persistent and transient groups for all body sites from both individuals are presented in Additional file 16.

We applied several techniques to control for the possibility of persistent groups being mistaken for transient groups if they occasionally fell below the detection threshold. First, in computing the maximum number of consecutive observations for an OTU, we counted a single zero-count for an OTU as not interrupting a run of consecutive observations, provided that both adjacent timepoints achieved non-zero counts for that OTU. Second, we re-sequenced 331 of the time series samples on the 454 platform, and recomputed the persistent and transient taxa summaries. We found that, for both individuals, the composition of the persistent gut and oral communities were not significantly different between 454 and Illumina despite a nearly 40-fold difference in sequencing depth. However, the persistent palm communities were significantly different in both cases. A possible confounding factor in this comparison is primer bias, as the V2 region was sequenced on 454. We therefore performed 1,000 jackknife iterations of this analysis by subsampling the Illumina data to 5,000 sequences per sample and recomputing the composition of the persistent group for each individual and body habitat. In this analysis, seven of the eight individual/body site pairs were never significantly different from the persistent community composition on the full Illumina data set. The one exception was the F4 gut community, which was significantly different in approximately half of the iterations. When re-sampled at a depth of 10,000 sequences per sample, the persistent F4 gut community never achieved significant difference from that determined on the full Illumina data set. The full results of these analyses are presented in Additional file 17. The designation of groups as 'persistent' was thus highly reproducible across sequencing platforms and amplicons, although it is still possible that sequencing error or very low abundance taxa occasionally falling below the detection threshold could result in underestimated size of the persistent group.

Differences in UniFrac distances in the left and right palms in adjacent timepoints of both individuals were significantly correlated (Pearson correlation for M3, r = 0.69, P = 2.07 × 10^-46; for F4, r = 0.64, P = 3.77 × 10^-16), possibly due to equilibration of microbial communities across palms by physical contact. We did not see correlations between other body sites. While the magnitude and direction of change in phylogenetic dissimilarity between adjacent timepoints were correlated between the palm sites, the species-level microbial taxa present on each hand were not significantly correlated, confirming previous observations that, at a single timepoint, the left and right hand of a single individual may share relatively few OTUs 13.

For single body sites, within-subject distances were lower than between-subject distances, suggesting a stable pattern, consistent across body sites, in between-subject dissimilarities across time. For example, the between-subject fecal sample distances were significantly higher than the within M3 fecal sample distances (t = 15.52; P < 0.001; one-tailed, two sample t-test) and the within F4 fecal sample distances (t = 33.45; P ≤ 0.001; one-tailed, two sample t-test).

The microbial community dynamics are especially apparent in principal coordinates analysis animations (Additional files 2, 3, 4, 5, 6, 7) where the sample types (subject, body site combinations) are represented as moving traces against a background of the Costello et al. data. The traces give an immediate picture of the variability within each body site, the relative distinctiveness of the sites and the subjects, and the relative speed of change in each site. Blooms of certain taxa contribute to the within site dissimilarities over time, as with the waning and waxing of the relative Proteobacteria abundance in the gut of both M3 and F4 (Additional files 8 and 9, Phylum panel (teal)). Similar patterns are apparent across all taxonomic levels and body sites (Additional files 8, 9, 10, 11, 12, 13, 14, 15). Such visualizations of microbial community dynamics will add another dimension to long-term studies of variable clinical states, such as inflammatory bowel disease or drug treatments, and changes in diet or lifestyle.

Sample collection and DNA isolation were performed as described in Costello et al. 6; and PCR, sequencing, and quality filtering of reads were performed as described in Caporaso et al. 14. Samples were not collected on days 422 through 437.

To facilitate massively parallel sequencing (1,967 samples), barcodes were reused across six lanes in a single Illumina GAIIx, with 374, 372, 364, 271, 265, and 323 samples in lanes 1 through 6, respectively (differing from Caporaso et al. 14, where samples were pooled and run over seven lanes). Sixteen samples were ultimately excluded from the analysis as fourteen samples were identified as potentially mislabeled (discussed below), and the barcodes for two samples were not found in the sequencing output, likely indicating a problem with amplification for those two samples.

To directly compare these M3/F4 time series samples with the samples presented in Costello et al. 6, which sequenced a different variable region (V2) using a different technology (454 FLX), a reference-based OTU picking protocol 18 was applied. After demultiplexing and quality filtering sequences, 97% OTUs were picked against the Greengenes database 19 (pre-filtered at 97% identity) using uclust 20. Reads were assigned to OTUs based on their best match to a Greengenes sequence, and reads that did not match a Greengenes sequence at 97% or greater sequence identity were discarded. The Greengenes taxonomy associated with the best match in Greengenes was assigned to each OTU, and the Greengenes tree was used for phylogenetic diversity calculations. These steps and subsequent data analysis were performed using Quantitative Insights Into Microbial Ecology (QIIME) on AWS.

To identify potentially mislabeled samples, we used the random forests classifier 21. A 2,000-tree forest was trained on the OTU × Sample Abundance matrix after evenly sampling to 500 sequences per sample and removing OTUs present in less than 1% of samples. The posterior probability that a given sample came from each of the body habitats (gut, oral cavity, skin) was estimated using only those trees in the forest that did not contain that sample in their training sets, to avoid overfitting. The classifier considers samples to be mislabeled when their alleged environment labels have a low posterior probability (<60%). Fourteen such samples were identified, and these samples were removed from all analyses.

The temporal core microbiome across body sites and individuals (Figure 2) was computed by varying the minimum number of samples in which an OTU must be observed to be considered part of the core microbiome, and then determining the number and fraction of total OTUs observed in each site (or combination of sites) that are part of the core. To facilitate direct comparison across sample types that contained different numbers of observations (for example, M3 (all) versus M3 gut), we randomly subsampled to exactly 130 observations per sample type, corresponding to the sample type for which we had the fewest observations.

The number of consecutive timepoints containing an OTU (Figure 3) was calculated as the maximum number of consecutive timepoints where an OTU was observed, allowing a zero count at a single timepoint to be considered part of a continuous stretch of non-zero counts if both adjacent timepoints had a non-zero count. This controls for sampling error as, for example, a long contiguous stretch of non-zero counts for an OTU interrupted by a single zero count for that OTU would likely indicate a bad sample, rather than a biologically relevant fact about that OTU in relation to the community. Persistent taxa were defined as those observed in 20% or more of the timepoints, but with at least 90% of those observations being consecutive (that is, they appear and remain present). Transient taxa were defined as those observed in at least 60% of the samples, but with at most 75% of those observations being consecutive (that is, they appear and disappear from the community frequently).

Animations were created in inVUE 22 based on the principal coordinate data presented in Figure 1a, b. inVUE files can be created in QIIME from the principal coordinate matrix and associated metadata file. After installing and opening inVUE, the user can run, pause, and stop the animations associated with different metadata categories.

All sequence data and sample metadata are publicly available under the 'Moving Pictures of the Human Microbiome' project [MG-RAST:4457768.3-4459735.3].