Number of samples needed to identifying significant differences between metagenomes

Fig. ​Fig.11 plots the sample size (S) of proteins that were sampled against the number of significantly different subsystems when the Sargasso data was compared to the SEED data at different confidence levels. At the most stringent level (99% confidence that the difference does not occur by chance) there are about 75 unique phylosubsystems in the Sargasso sample (solid red line) and about 150 unique phylosubsystems in the SEED. In either case at least the sample size needs to approach 300,000 proteins to achieve statistical significance. In contrast, at lower confidence levels, there are more significantly different phylosubsystems, and the samples need to be smaller to detect them. Thus with a 90% confidence that the difference do not occur by error there are approximately 100 unique subsystems in the Sargasso data, and a sample size of about 150,000 proteins is required to identify them all.

Open in a separate window

Figure 1

Effect of sample size on identifying differences between phylosubsystems. The red lines reflect the number of phylosubsystems overrepresented in the Sargasso Sea dataset. The blue lines represented the number of phylosubsystems overrepresented in the SEED dataset. Three different confidence levels (90, 95, and 99%) are plotted.

The AMD dataset has many fewer phylosubsystems than the SEED dataset, and only those subsystems present in the both samples are used in the comparison to identify significantly different phylosubsystems. This limits the analysis to 284 of the 523 different phylosubsystems. Of those that are present in the AMD and the SEED, only 19 are significantly more abundant in the AMD sample (Table ​(Table2)2) and 58 are more abundant in the SEED dataset. Saturation was reached at ~145,000 samples.

Subsystem differences between the SEED and Sargasso Sea metagenomes

Statistically different phylosubsystems between the SEED and Sargasso Sea metagenomes are shown in Fig. ​Fig.2A.2A. The effect of sample size is apparent. For example, statistically differences between RNA metabolism, oxidative phosphorylation, and membrane transporters can be detected at small sample sizes (e.g., 100,000 proteins). However, the differences between carbohydrate and amino acid metabolism, as well as most cofactors, vitamins, and pigments are not statistically different until >150,000 proteins have been sampled. Out of the 77 phylosubsystems that are significantly overrepresented in the Sargasso Sea, 69 are from the Bacteria, 5 from the Archaea, and 3 from the Eukarya.

Open in a separate window

Figure 2

Significantly different subsystems between Sargasso Sea and SEED datasets. (A). Each subsystem was bootstrapped with between 10 and 400 samples per bootstrap, and subsystems that are significantly different with 99% and 2,000 bootstraps are highlighted. Those subsystems that are significantly more prevalent in the SEED database are colored blue, and those subsystems that are significantly more prevalent in the Sargasso Sea dataset are colored red. (B). Magnified view of several different subsystems. Subsystems from amino acid synthesis, carbohydrate utilization, cofactor synthesis, fatty acids, nucleotide synthesis, and photosynthesis are shown in more detail. Colors are as described for (A). E = eukaryotic subsystem, B = bacterial subsystem, A = archaeal subsystem.

Fig. ​Fig.2B2B highlights some of the differences for phylosubsystems. A more detailed description of all of the subsystems with statistically significant differences in occurrence between environmental data sets is given in the supplemental material [see Additional file 3]. Some ecologically important differences between the Sargasso Sea and the SEED database are discussed below with data extracted from the Supplementary Table.

Potential osmoregulation by amino acids in the Sargasso Sea

The phylosubsystems involved in the synthesis of serine (S), threonine (T), and glycine (G) are overrepresented in the Sargasso Sea metagenome (Table ​(Table3).3). For example, there are 503 proteins per million proteins in the Sargasso Sea database that are similar to bacterial glycine synthesis proteins, and only 390 proteins per million proteins in the SEED database that are similar to bacterial glycine synthesis proteins. Exactly these three amino acids are also the most abundant amino acids found in the Sargasso Sea [10,11]. However proteins in the Sargasso Sea metagenome do not contain significantly more S, T, or G (Fig. ​(Fig.3),3), suggesting that the observed bias in amino acid synthesis subsystems and in the water samples is not related to protein synthesis. The primary organic osmolytes used by marine bacteria are small organic, uncharged, solutes that have little effect on the intracellular biochemistry, such as glycine betaine [12-15]. The subsystems for the production of betaine are also overrepresented in the Sargasso Sea compared to the SEED database (Fig. ​(Fig.2B).2B). Similarly, serine and threonine are small, polar, uncharged amino acids that are perfect compatible solutes for balancing osmotic pressure. Serine has also been previously shown to be as effective as glycine betaine at protecting enzymes from the effects of increased osmolality [15]. In contrast, genes encoding the osmoprotectants proline, sucrose, and trehalose [13,14,16,17] were underrepresented in these samples. Therefore, we predict that marine microbes are synthesizing glycine, serine, threonine, and betaine as osmolytes, an hypothesis that can be tested experimentally.

Open in a separate window

Figure 3

Fraction of amino acids in metagenomes. The fraction of each amino acid in all the predicted proteins in the three data samples was counted and compared.

Photosynthesis in the Sargasso Sea

As previously observed [6,7], there was a strong bias towards subsystems involved in photosynthesis in the Sargasso Sea metagenome. This bias includes subsystems for the Calvin-Benson cycle, chlorophyll biosynthesis genes, the cytochrome B6-F complex, Photosystem I, Photosystem II, isoprenoid biosynthesis, and carotenoid biosynthesis.

Some phylosubsystems involved in one-carbon metabolism, including the synthesis and degradation of carbohydrates, cell walls, and capsules are more abundant in the Sargasso Sea. In contrast, the genes for the utilization of complex carbon sources including lactose, arabinose, fructose, mannose, galcitol and inositol are all underrepresented in the marine environment suggesting these are not significant sources of carbon in this environment.

Nucleic acid and phosphate metabolism in the Sargasso Sea

Phylosubsystems involved in purine and pyrimidine de novo synthesis and scavenging pathways, as well as ribonucleotide reduction (scavenging ribonucleotides for DNA synthesis) are more abundant in the Sargasso Sea. Similarly, the subsystems involved in the capture of phosphate via the conversion of ADP to ATP coupled to oxidative phosphorylation are also overrepresented in the Sargasso sample. In contrast, nitrogen metabolism phylosubsystems are less abundant in the Sargasso than the SEED, with the sole exception of ammonia assimilation that is marginally overrepresented in the Sargasso sample at larger sample sizes. The Sargasso Sea has previously been reported to be phosphate limited. The concentration of dissolved inorganic phosphate is approximately 4 nM in the Sargasso Sea. By comparison, the North Pacific and typical bacterial minimal media have phosphate concentrations of approximately 100 nM [18,19]. Together, these results suggest that phosphate acquisition is critical for microbial growth the Sargasso Sea environment.

Mobility of Sargasso Sea microbes

Estimates of the percentage of bacteria in the ocean that are motile vary from less than 5% to more than 80% [20,21], and there were far fewer genes encoding flagella in the marine environment compared to the SEED database. However many marine microbes are thought to use alternative, less well characterized, motility systems, such as the motility mechanism characterized in cyanobacteria [22,23] or twitching motility previously shown in marine microbes [24]. This data leads to the hypothesis that marine microbes are generally not using flagella based motility for movement, and future studies on the genomics of twitching and gliding motility may reveal hints about these mechanisms of movement.

Subsystem differences between the SEED and AMD metagenomes

When the AMD and SEED databases were compared, only phylosubsystems that were in both the AMD and the SEED samples were included. This limited the total number of subsystems that were compared for statistically significant differences. There are far fewer phylosubsystems with significantly different distributions between the AMD and SEED datasets, and phylosubsystems that are significantly more common in the AMD dataset are shown in Table ​Table2.2. The different occurrences of subsystems reflect the limited complexity of the AMD environment that contains Bacteria and Archaea [8]. The majority of subsystems that are significantly more common in the AMD data set are from archaeal proteins. In the AMD environment, the production of amino acids does not appear to be critical, and only archaeal arginine and histidine degradation and leucine and chorismate synthesis are overrepresented in these samples. Our limited selection of overrepresented subsystems in the AMD sample presumably reflects the current bias in annotated subsystems in the SEED. As the subsystems continue to evolve and expand, and the NIH Project to Annotate 1,000 Genomes [9] matures the impact of these annotations on the AMD sample and other metagenomes will highlight those areas of metabolism and physiology that are critical to survival in different environments.

Assignment of proteins to subsystems and phylosubsystems

Each protein from the AMD, Sargasso, or the SEED database was compared to proteins in the SEED database previously assigned to particular subsystems by the SEED annotators [9]. A protein was considered a member of a subsystem if the protein had significant similarity (designated as an E value less than 1 × 10^-20) to another protein previously assigned to a subsystem. Each protein was also classified as Bacteria, Eukarya, and Archaea, based on the Domain assignment of the most similar protein. There were a total of 276 annotated subsystems in the SEED. Bacteria had proteins in 257 of the subsystems. Archaea and Eukarya had proteins in 132 and 134 of the subsystems, respectively. This means that each there were a total of 523 potential data points. The term phylosubsystem is used to reflect that fact that the assignments are based both on the subsystem and Domain.

Comparisons of metagenomic databases

A flowchart of the methods used to identify subsystems with statistically significant differences between databases is shown in Fig. ​Fig.44 and described in the supplemental material [see Additional File 2]. In addition, source code to software to calculate these differences is provided as supplemental material [see Additional File 1]. In order to decide whether a subsystem was over-represented between metagenomes, a comparison between the median number of proteins assigned to specific subsystems was performed for a given confidence level. The following steps were carried out: 1) A number N of proteins were drawn at random, with replacement, from each of two metagenomes. Each protein was classified into a subsystem. Then, the difference between the two sets of samples was calculated for each subsystem, resulting on a list of differences between metagenomes by subsystems. These differences were expected to be more pronounced where the differences between metagenomes were more remarkable, but with just one set of differences this was not guaranteed. Thus, this procedure was repeated M times. Then, for each subsystem the median of all the M differences was calculated. 2) To build a confidence interval, a number N of "mixed" samples were drawn at random, with replacement, from a mixture of both metagenomes. For each sample, two random numbers were drawn, the first one to decide which metagenome would be selected, and the second one to choose a particular element of the metagenome, which was the accounted into its corresponding subsystem. This process was repeated for another set of N samples and with these two mixed sets a difference was calculated. This whole process was repeated M times again, and for each subsystem the resulting differences were ordered and used to build a confidence interval. The limits of the confidence interval for each subsystem were estimated by using the corresponding quantile elements on the ordered set of differences of each subsystem as limits to the confidence interval. For example, if M = 1,000 and the confidence level was 90%, the limits for the confidence interval were at the 5% and 96% percentiles on the ordered values of the differences for that subsystem, that is, the 50^th and 951^st elements would be the upper and lower limits on the confidence interval. 3) For any given subsystem, if the median from step 1 was inside the confidence interval determined in step 2 it was considered that the subsystems were not significantly different at the given confidence level. If the median was outside the confidence interval, it was assumed that the difference was significant at the given confidence level.

Open in a separate window

Figure 4

Flow chart of methods used to identify statistical differences between phylosubsystems.