Introduction

Since the discovery of their abundance in marine bacterioplankton communities two decades ago (González and Moran, 1997), members of the marine Roseobacter lineage have emerged as important model organisms for marine microbial ecology. The group spans multiple described genera (at least 45), encompasses a comparatively large sequence variation among 16S rRNA genes (up to 11%) and has a poorly resolved within-taxon phylogeny (Buchan et al., 2005; Wagner-Döbler and Biebl, 2006; Brinkhoff et al., 2008). The recent availability of genome sequences (currently, 5 closed and 27 draft) from cultured members of the Roseobacter lineage provides a detailed inventory of the metabolic and ecological capabilities of each strain (albeit limited by the accuracy of annotation), a basis for comparative analyses among strains, and a means to examine predictive frameworks for the lineage.

Genome sequences of other ocean microbes have been used to explore niches and resource partitioning within taxa. Multiple genome sequences and robust phylogenies for Prochlorococcus have revealed that the distribution of ecologically important gene systems (for example, light harvesting (Garczarek et al., 2000; Hess et al., 2001; Bibby et al., 2003) and DNA repair mechanisms (Scanlan et al., 2009)) strongly correlate with the phylogenetic structure of this genus (Rocap et al., 2003; Coleman and Chisholm, 2007). In addition, genome content differences among both Prochlorococcus and Synechococcus strains have been linked to variations in the environments from which the strains were isolated (West and Scanlan, 1999; Johnson et al., 2006; Martiny et al., 2006; Palenik et al., 2006; Dufresne et al., 2008). For members of the Vibrionaceae, small-scale differences in environmental conditions based on microenvironment and season have been shown to drive lineage adaptation (Hunt et al., 2008) and presumably genome content. The phylogenetic- and environment-based frameworks used to interpret data from studies such as these have facilitated the development of predictive community structure models for marine microbes (Follows et al., 2007; Rabouille et al., 2007).

Members of the Roseobacter lineage have been characterized as ecological generalists (Moran et al., 2004, 2007; Polz et al., 2006). Although the first cultured roseobacters were aerobic anoxygenic phototrophs (AAnPs) (Shiba et al., 1979), numerous heterotrophic strains have since been found (Blankenship et al., 1995; Shimada, 1995; Buchan et al., 2005). Cultured roseobacters have a surprisingly flexible suite of mechanisms for energy and carbon acquisition, including carbon monoxide and hydrogen sulfide oxidation (King, 2003; Moran et al., 2004), and anaplerotic CO2 fixation (Sorokin et al., 2003; Moran et al., 2004; Swingley et al., 2007). Together with the observation that this lineage's genomes are large and variable, a picture of considerable trophic versatility among roseobacters has emerged (Buchan et al., 2005; Wagner-Döbler and Biebl, 2006; Brinkhoff et al., 2008).

The 32 Roseobacter genomes provide an unprecedented opportunity to examine the scope of extant gene systems and to explore various ecological and evolutionary perspectives that might distinguish functionally differentiated clusters within this lineage. In this study, we consider three theoretical ecological/evolutionary frameworks as possible predictors of the gene repertoires of the 32 Roseobacter strains. As a robust Roseobacter phylogeny has yet to emerge from rRNA gene analysis (Buchan et al., 2005; Brinkhoff et al., 2008), we first develop a well-supported phylogeny for the lineage from a concatenation of conserved, single-copy genes and within this phylogenetic structure we examine the evolutionary relationships as possible constraints on genome content and predictors of the genetic capabilities of each strain. Next, we explore lifestyle strategy (heterotroph, photoheterotroph or autotroph) as a possible driver of genome attributes (that is, imposing or releasing bacteria from constraints on genome content). Finally, we ask whether environmental conditions (defined here by the geographical location of isolation) might best explain the observed differences in genetic traits and the retention or acquisition of specific gene systems.

Materials and methods

Genome sequencing, annotation and completeness index

The 32 Roseobacter genomes publicly available as of 15 August 2008 were used in analyses (see Table 1 for genome details). Sequencing and annotation methods for Ruegeria pomeroyi DSS-3 (Moran et al., 2004; formerly Silicibacter), Ruegeria sp. TM1040, Jannaschia sp. strain CCS1 (Moran et al., 2007), Roseobacter denitrificans OCh 114 (Swingley et al., 2007) and Dinoroseobacter shibae DFL12 (Wagner-Döbler et al., 2009) are described elsewhere. The remaining genomes (Table 1) were sequenced and auto-annotated by the J Craig Venter Institute as part of the Moore Foundation Microbial Genome Sequencing Project (see http://moore.jcvi.org/moore/ for details).

Table 1 Roseobacter genome characteristics

In all, 5 of the 32 Roseobacter genomes have been assigned closed genome status, and we used these genomes as the basis for our genome completeness index. The protein sequences of 143 universal single-copy bacterial genes (Santos and Ochman, 2004; Santos, personal communication) were used in a BLASTp query against the five closed Roseobacter genomes; manual gene-calling based on alignment score, E-value, and contextual analysis was used to determine the presence or absence of genes in the five closed genomes. Of these 143 genes, 111 were determined to be unambiguously present in all five genomes (see Supplementary Table S1). The protein sequences of these 111 genes were then used in BLASTp analysis against the remaining 27 genomes. The percent presence of these 111 genes in a single genome constituted that genome's completeness index (Table 1).

Phylogenetic tree inference

Out of 111 universal single-copy genes identified in the 32 Roseobacter genomes, only genes that were completely sequenced in all genomes and had no ambiguous start/stop sites were used in phylogenetic analyses. These 70 genes (Supplementary Table S1) were concatenated and aligned with ClustalW in Geneious 4.0 (available from http://www.geneious.com) using Escherichia coli K12 substrain MG1655 as the outgroup. The alignment was imported into ARB (Ludwig et al., 2004), where it was heuristically adjusted, and a filter was created to remove all positions containing gaps in the alignment. The resultant alignment of 25 316 positions was used in subsequent phylogenetic reconstruction analyses in ARB (neighbor joining with point accepted mutation substitution matrix and 100 bootstrap runs) and in RAxML (Stamatakis et al., 2008) at CIPRES (http://www.phylo.org; maximum likelihood analysis with 200 bootstrap runs and Jones–Taylor–Thornton (JTT) substitution model). The best-fit maximum likelihood tree is reported along with bootstrap values from each phylogenetic inference method.

Identification of orthologs and ecologically relevant genes

Orthologs among the 32 genomes were identified by sequential two-way reciprocal best-hit (RBH) analysis, beginning with the R. pomeroyi and Rhodobacterales HTCC2255 genome comparison and continuing by adding each of the remaining 30 genomes one at a time. The RBH Basic Local Alignment Search Tool (BLAST) thresholds were set at E-value <10−5 and amino acid identity >30%. The RBH results were subsequently compiled into a single matrix containing the distribution of all shared genes and used for the genome content comparisons described below (See Supplementary Table S2 for matrix). This relaxed ortholog definition was used because an all-way RBH requirement was unworkable for the large number of genomes, each containing gene families represented by multiple members. We tested whether the order of the sequential best-hit analysis resulted in substantial changes in the ortholog matrix or the outcome of the analyses (that is, by using a different order of adding genomes in the pair-wise RBH), and found it did not.

In addition to whole-genome ortholog identification, a select group of ecologically relevant genes/gene pathways was also identified using representative protein sequences of the target genes from a Roseobacter for which the gene functions had been experimentally verified. If no Roseobacter met this criterion, then a protein sequence was obtained from the closest Roseobacter relative containing the desired experimentally verified gene. All query protein sequences were used in BLASTp analysis against the Roseobacter genome database (http://www.roseobase.org). BLAST E-values, gene neighborhoods and clusters of orthologous group assignments (Tatusov et al., 2003) were manually examined and used to determine the presence or absence of these genes and pathways in each of the 32 genomes.

Classification schemes

The 32 isolate genomes were sorted into groups within each of the three frameworks. First, five deeply branching nodes in our phylogenetic inference best-fit tree were chosen to distinguish isolate groups based on shared ancestry and were designated Clades 1–5. Next, we categorized isolates into lifestyles based on their trophic status: heterotrophic, photoheterotrophic (that is, heterotrophic but likely subsidized by aerobic anoxygenic phototrophy or rhodopsin-based phototrophy) or autotrophic. Organisms were considered AAnPs based on the presence of the puf operon and genes for the synthesis of bacteriochlorophyll a; they were considered rhodopsin-supplemented photoheterotrophs based on the presence of gene orthologs for proteorhodopsin or xanthorhodopsin; and they were considered autotrophs based on the presence of RuBisCO and the Calvin–Benson–Bassham pathway. Although these designations were made from draft genome sequences for many strains, the high genome completeness index suggests they are largely correct. Finally, isolates were classified into one of five broad environmental categories based on the source of isolation: Pacific Ocean, Atlantic Ocean, Indian Ocean, polar oceans or eukaryote-associated (Table 1).

Identification of genes in the Global Ocean Sampling

Global Ocean Sampling (GOS) sample sites, sampling procedures and sequencing methods are described elsewhere (Rusch et al., 2007; Yooseph et al., 2007). A subset of Roseobacter protein sequences representing each of the major biogeochemical pathways and processes that we examined (Supplementary Table S3) was used in a BLASTp query against the unassembled GOS data set at the Community Cyberinfrastructure for Advanced Marine Microbial Ecology Research and Analysis (CAMERA) (Seshadri et al., 2007). Gene hits from the GOS samples were retained for further analysis as potential orthologs at E-value cutoffs ranging between 10−80 and 10−20, depending on the gene. Paired reads were then removed before the resultant matches were used in BLAST analysis against the All Prokaryotic Proteins (P) database. Only gene matches that had a best hit to a gene in a Roseobacter genome were retained for further analysis. Finally, protein sequences from the Roseobacter-like GOS matches underwent a BLASTp query at GenBank, and were eliminated if their top alignment scores were to proteins with a different annotated function than the original query protein.

To compare gene counts between oceans, the Roseobacter-like metagenomic sequences obtained from the GOS data underwent several normalizations. Counts for functional genes retrieved at each sample location were size-normalized to the length of the recA gene from E. coli K12 substrain MG1655 to account for effects of size on the probability of sampling (Howard et al., 2008). The number of Roseobacter genome equivalents for each sample location was then calculated by averaging size-normalized Roseobacter-like gene counts of the universal single-copy genes recA and rpoB. To estimate per-cell frequency for each examined Roseobacter gene (listed in Supplementary Table S3), the sample gene counts were summed by ocean basin (Atlantic, Pacific and Indian) and divided by the number of Roseobacter genome equivalents for that basin. Only coastal and open ocean GOS sample sites were considered, with estuaries, embayments, lagoon reefs, fringing reefs, freshwater, mangroves, coral reefs, hypersaline lagoons, warm seeps and harbors excluded from the analyses (Supplementary Table S4).

The clade distribution among ocean basins was determined using the recA gene sequence. After retrieving Roseobacter-like RecA sequences by BLASTp analysis against the GOS database at CAMERA (as described above), each individual protein sequence was used as a query sequence in a subsequent BLASTp analysis against all 32 genomes at Roseobase (http://www.roseobase.org). The Roseobacter-like RecA sequences from GOS were then assigned to a clade according to their best match among the 32 genomes, and the occurrences of each clade were summed across samples in each ocean basin.

Statistical analyses

Patterns of ortholog distribution among the 32 genomes were evaluated using the Bray–Curtis Index of Similarity (Legendre and Legendre, 1998). The 32 genomes contained a total of 31 874 orthologs. Similarities between genomes include all orthologs in this matrix, so that both the shared presence and shared absence of a gene are taken into account in the similarity calculation. This similarity matrix was used to create a hierarchical clustering dendrogram based on complete linkage grouping (that is, furthest neighbor analysis). An analysis of similarity (ANOSIM) was used to test for significant differences among a priori assigned genome groups based on phylogenetic clade, trophic strategy or geographical isolation location. The multivariate analyses were performed using the statistical package PRIMER 5 for Windows v. 5.2.7.

The average nucleotide identity was obtained for two Roseobacter strain comparisons, Phaeobacter gallaeciensis 2.10 with P. gallaeciensis BS107 and D. shibae DFL12 with P. gallaeciensis 2.10 according to the method described by Goris et al. (2007). These comparisons were chosen to bracket the amount of sequence heterogeneity observed and to provide context for our ortholog similarity comparisons.

Significance of gene distributions between any two assigned groups (for example, between two clades, between two trophic strategies or between two ocean basins) was assessed with a binomial distribution d-score test (Markowitz et al., 2008).

Results and discussion

Because many of the 32 Roseobacter genomes are in draft status, we developed a completeness index based on the presence or absence of 111 universal single-copy genes. The lowest genome completeness index obtained was 96% (for Octadecabacter arcticus 238 and Rhodobacterales bacterium HTCC2255; 107 out of 111 presumed universal genes were represented), and 18 of 32 genomes had a completeness index of 100% (Table 1). We therefore considered all draft genomes to be good representations of these organisms’ gene content.

Phylogenetic inference and clade distribution

Previous phylogenetic reconstructions of the Roseobacter lineage using 16S rRNA gene relationships have led to the identification of subgroups within the lineage (Buchan et al., 2005; Brinkhoff et al., 2008). However, many of the nodes, especially those distinguishing deep branching points in these phylogenies, do not have statistical support, and therefore do not provide clear phylogenetic relationships for the members of this lineage. We took advantage of the genome sequence data to construct an alignment from the concatenation of 70 conserved single-copy genes; this alignment was subsequently used in phylogenetic tree inference (see Supplementary Table S1 for gene list and Supplementary Figure S1 for concatenated gene, 16S rRNA gene and 23S rRNA gene tree comparisons). The resultant tree topology suggested there are five deeply branching clades within the Roseobacter lineage (Figure 1). Three of the presumed roseobacters, Maritimibacter alkaliphilus HTCC2654, Rhodobacterales HTCC2150 and Rhodobacterales HTCC2255, fell outside these clades. Most members within a single genus clustered together on the tree, although the placement of two members of the genus Oceanicola into different clades suggests that a taxonomic reclassification may be needed for some isolates.

Figure 1
figure 1

A consensus maximum likelihood tree of the 32 sequenced Roseobacter genomes. The alignment for tree inference was created from a concatenation of 70 universal single-copy genes contained in each of the Roseobacter genomes and in E. coli K12, which was used as an outgroup. Bootstrap values of >50% for the maximum likelihood best-fit tree (200 iterations) and neighbor-joining tree (100 iterations) are listed at each node. The neighbor-joining bootstrap values are listed in parentheses. (*) demarcates nodes where the neighbor-joining tree did not agree with the maximum likelihood tree. Designated Clades 1–5 are listed to the right of the tree. The scale bar represents 10% sequence divergence.

Buchan et al. (2005) identified 13 major sequence clusters based on 16S rRNA gene sequences within the Roseobacter lineage. Twelve of the 16S rRNA-based clusters can be mapped onto our 70-gene phylogeny (data not shown). Clade 1 contains the 16S rRNA gene sequence clusters RGALL, RATL and TM1040. Clade 2 contains sequence clusters ANT9093, OBULB, SPON and AS-21. Clade 3 contains sequence cluster CHAB-I-5. Clade 4 contains sequence clusters AS-26, DG1128, DC5-80-3 (RCA cluster) and OCT. Clade 5 contains no previously identified sequence clusters, and cluster NAC11-7 is not covered by any of the clades in our study.

Two of the most abundant Roseobacter 16S rRNA gene sequence clusters recovered from marine habitats do not have closely associated sequenced genomes (Buchan et al., 2005), and thus are not included in the 70-gene phylogenetic tree. The first, the DC5-80-3 or RCA cluster, has often been observed as the most abundant Roseobacter group in polar and temperate oceans (Brinkhoff et al., 2008). 16S rRNA genes from RCA distantly group with those from genomes in Clade 4 (Figure 1), a clade that harbors all the sequenced polar Roseobacter isolates thus far. A second abundant marine sequence cluster, NAC11-7, is frequently the dominant Roseobacter taxon found during phytoplankton blooms (Buchan et al., 2005; West et al., 2008). 16S rRNA genes from the NAC11-7 group did not cluster with those of any clade, and were most related to that of Rhodobacterales HTCC2255 (data not shown), which also fell outside the five clades established by the 70-gene phylogeny (Figure 1).

Roseobacter-like recA genes, a robust marker for bacterial phylogeny (Eisen, 1995), were obtained from the GOS data set by BLASTp analysis (see Materials and methods) to ascertain which isolate genomes are most representative of wild roseobacters in surface ocean water. When the set of Roseobacter-like RecA GOS sequences was used in a best-match BLASTp query against the 32 genomes, hits to all five clades were found throughout the major ocean habitats surveyed (Figure 2). In general, the distribution of clades is not remarkably different between the Atlantic and Pacific, or Indian oceans (Figure 2). A large percentage of the Roseobacter RecA sequences from the GOS appear most closely related to one of the three singleton genomes (that is, not belonging to one of the five defined clades). This finding, along with the lack of genomic data for the RCA and NAC11-7 sequence clusters, suggests that representation of oceanic Roseobacter genomes could be improved with additional genome sequences.

Figure 2
figure 2

Clade distribution among ocean basins calculated from BLASTp best match of Roseobacter recA genes from the Global Ocean Sampling (GOS). Atlantic n=71, Pacific n=48 and Indian n=51.

Genome content related to phylogeny

We examined genome characteristics (for example, G+C content, rRNA copy number, genome size; Table 1) and gene content of the 32 roseobacters within the context of the five-clade phylogenetic framework. Whole genomic content comparisons (based on distribution patterns of 31 874 orthologs; see Materials and methods; Supplementary Table S2) indicate weak but significant genome clustering by clade (ANOSIM R=0.410, P0.001, three roseobacters not assigned to clades were excluded from the statistical test), with the within-clade similarity in gene repertoire for Clades 1, 2 and 3 driving this pattern (Figure 3). Generally, neither the examined genome characteristics nor the examined gene distributions segregate strongly based on phylogenetic relatedness (Table 1 and Figure 4). Some exceptions include a greater mean rRNA operon copy number for Clade 1 than for other clades (t-test, P0.01); a strictly heterotrophic composition of Clade 1; a genetic potential for biotin synthesis in Clade 1 (vitamin synthesis in bacteria has been identified as important in bacterial–phytoplankton relationships; Croft et al., 2005; Wagner-Döbler et al., 2009); a lack of Lux-type quorum sensing genes in Clades 4 and 5; a genetic potential for H2 oxidation unique to Clade 3; the absence of sulfur oxidation genes in Clade 4 genomes; and absence of the ppk1 gene for polyphosphate biosynthesis in Clade 1 (whereas all isolates outside Clade 1 have this gene).

Figure 3
figure 3

Complete linkage (that is, furthest neighbor) cluster analysis illustrating the gene content similarities among the genomes. Bray–Curtis similarities between all genome pairs were calculated from a matrix containing all 31 874 genes identified in the 32 genomes. In this manner, the similarity calculation was based on both the shared presence and shared absence of genes. For context, the P. gallaeciensis 2.10 to P. gallaeciensis BS107 comparison is 87.4% similar in this analysis compared with an average nucleotide identity (ANI) (Goris et al., 2007) of 97.0%. The D. shibae DFL12 to P. gallaeciensis 2.10 comparison is 46.7% similar here compared to an ANI of 70.4%. The three framework groups of each isolate are illustrated by the shape and color pattern depicted at the tips of the cluster diagram. The phylogenetic clade framework is represented by squares; the lifestyle framework is represented by triangles; and the environment framework is represented by circles. Unique combinations of these three frameworks are illustrated with colored boxes around the names of isolates. Breaks in the three framework groupings are noted by an asterisk next to the strain name. Nodes below the dashed red line indicate groups with 58% similarity.

Figure 4
figure 4

A matrix depicting the presence of select genes or gene pathways in the 32 Roseobacter genomes arranged and color-coded by clade. A colored box containing a dot indicates the presence of the gene/pathway. An ultrametric tree has been placed above the gene matrix for reference. Isolation source indicates the region where the Roseobacter strain was isolated and is coded as: A=Atlantic Ocean, E=Eukaryote Associated, H=polar oceans (high latitude), I=Indian Ocean and P=Pacific Ocean. Gene/pathway abbreviations are as follows: NRPS/PKS, non-ribosomal peptide synthetase/polyketide synthase; Glut/leu/phenyl/val dehydrogenase, glutamate/luecine/phenylalanine/valine dehydrogenase.

The lack of a strong segregation by phylogenetic assignment for genome content (Figure 3) or ecologically relevant gene systems (Figure 4) suggests the importance of gene acquisition by horizontal transfer originating either within or outside the lineage. Other evolutionary processes known to shape genome content (selective gene loss, gene duplication, gene genesis; Snel et al., 2002) are no doubt important in this lineage, but are mechanisms less likely to produce the observed patchy distribution of ecologically relevant genes in the Roseobacter isolates relative to their phylogenetic reconstruction. Although the rates of gene transfer within the Roseobacter lineage is not known, the occurrence in 30 of 32 genomes of gene transfer agent operons (Figure 4), an unusual system for moving chromosomal fragments to close relatives (Biers et al., 2008; Zhao et al., 2009), suggests a mechanism for shaping Roseobacter gene content through frequent within-lineage gene transfers.

Between-genome similarities were generally higher for Roseobacter isolates in the same genus (for example, P. gallaeciensis BS107 and P. gallaeciensis 2.10; R. denitrificans Och114 and Roseobacter litoralis Och149; Figure 3) than for roseobacters belonging to different genera. Nonetheless, blurred gene content boundaries among deeply branching clades would impose a requirement of dozens of taxonomically shallow groups (for example, species level) to accurately represent Roseobacter contributions to ecosystem functions, thus making the phylogenetic framework a cumbersome approach for defining ecological subgroups.

Genome content related to trophic strategy

Owing to the significant versatility in mechanisms for obtaining carbon and energy previously observed for this group (Buchan et al., 2005; Moran et al., 2007), we hypothesized that an organism's trophic strategy could impose or remove constraints on genome content. For example, the ability to use sunlight for energy generation, which is widely distributed within the Roseobacter lineage, might mitigate an organism's energy limitations in the oligotrophic marine environment, while imposing requirements for metals and cofactors specific to phototrophy. Similarly, the ability to fix inorganic carbon might reduce an organism's requirements for substrate transporters. If such interplay between trophic strategy and functional gene repertoire exists, then significant and predictable differences in genome content should be evident between lifestyle categories.

Thirteen of the 32 roseobacters have genes for photoheterotrophy (10 AAnPs, 3 rhodopsin-containing), whereas one has RuBisCO. The remaining 18 are considered heterotrophs here (although some may obtain energy from inorganic compounds such as CO and H2S; Moran et al., 2007) (Table 1). Genome ortholog comparisons suggest moderate and significant differences in genome content among these groups (ANOSIM R=0.545, P0.001). The strength of these differences does not stem solely from the very unique rhodopsin-containing genomes (Figure 3). The differences also are not solely due to the presence of light-harvesting-related genes shared by the AAnP genomes or rhodopsin-containing genomes, as removal of the rhodopsin genes and 29 genes specific for AAnP light harvesting resulted in a similar level of clustering by trophic strategy (ANOSIM R=0.522, P0.001). The lifestyle framework accurately predicts the gene repertoire groupings at similarity levels 58% (Figure 3; red dashed line), which represents the gene content relationships for 19 of the 32 genomes and is the best predictor of the three frameworks analyzed.

The majority of non-light-harvesting gene or pathway-related differences among strains can be traced to hypothetical proteins unique to the AAnPs or heterotrophs, as well as to a number of genes encoding transcriptional regulators and amino acid uptake and synthesis systems (Figure 5). Although trophic strategy was a good predictor of an isolate's gene repertoire, only a few of the ecologically relevant gene systems we examined were strictly differentiated according to this framework. The five C1 utilization pathways identified in Roseobacter genomes (serine cycle, methanol oxidation, trimethylamine oxidation, formaldehyde oxidation and formate oxidation) had a 70% occurrence rate in AAnP genomes (that is, 35 out of the 50 possible occurrences if all 10 AAnP genomes had all five pathways), but only a 42% occurrence rate in the 19 heterotrophs (only 40 out of 95 possible occurrences; d-score, P0.01). The heterotrophs tended to have more genes for six identified aromatic degradation pathways (β-ketoadipate, gentisate, benzoate, phenylacetic acid, homoprotocatechuate and homogenisate) with a 60% occurrence rate (68 out of 114 possible occurrences) compared with 33% in AAnPs (20 of 60 possible occurrences; d-score, P0.01). Compared with the genomes of the other groups, the rhodopsin-containing genomes shared few orthologs that distinguished them as a coherent group (data not shown).

Figure 5
figure 5

Relative abundance of ortholog groups that are overrepresented in the isolate genomes from a particular lifestyle strategy. An ortholog was considered overrepresented when it was 50% more prevalent in the genomes from one lifestyle strategy than the other. The overrepresented orthologs were grouped into functional categories whose relative percent abundance is depicted. The rhodopsin-containing and RuBisCO-containing lifestyle groupings were not considered because of the low number of genomes in these categories.

As noted previously (Buchan et al., 2005; Moran et al., 2007) and strongly reinforced in this analysis, Roseobacter genomes exhibit a remarkably versatile suite of mechanisms for energy and carbon acquisition. Along with the presence of genes for oxidizing carbon monoxide and hydrogen sulfide (King, 2003; Moran et al., 2004), we found evidence for energy generation by H2 oxidation in Roseovarius sp. TM1035, Pelagibaca bermudensis HTCC2601, and in Sagittula stellata E-37, proteorhodopsin- (Rhodobacterales sp. HTCC2255) and xanthorhodopsin-based (O. antarcticus 307 and O. arcticus 238) phototrophy, and CO2 fixation based on the presence of a Form IC RuBisCO and homologs to Calvin–Benson–Bassham cycle genes in P. bermudensis HTCC2601 (Figure 4). This emerging picture of high trophic versatility among cultured roseobacters (Buchan et al., 2005; Wagner-Döbler and Biebl, 2006; Brinkhoff et al., 2008) is in accord with recent shifts away from a perception of marine bacterioplankton communities consisting largely of canonical photosynthetic and heterotrophic cells (Karl, 2002).

The rhodopsin-containing genomes

The three rhodopsin-containing genomes harbored the most unique genome content of any of the isolates (Figure 3). The proteorhodopsin-containing Rhodobacterales bacterium HTCC2255 gene content was unique because it consisted of only 2197 genes, far fewer than any the other genome. The two xanthorhodpsin-containing isolates, O. arcticus and O. antarcticus (which are also the only two polar ocean isolates), are clearly part of the Roseobacter lineage (Figure 1) but possess the greatest number of unique genes among all isolates (2230 genes for O. arcticus and 1822 genes for O. antarcticus; 32 isolate mean=617 and s.d.=437). The majority of these unique genes were annotated as phage or transposase genes with 65 and 52 phage gene annotations and 935 and 574 transposase gene annotations for O. arcticus 238 and O. antarcticus 307, respectively; these numbers are extremely high compared with those in the other Roseobacter genomes (phage mean=21 and s.d.=10; transposase mean=52 and s.d.=38).

Genome content related to ocean environment

Environmental properties are potential drivers of marine bacterial genome evolution by selecting for niche-specific genetic capabilities. For example, the marine cyanobacteria Procholorococcus and Synechococcus exhibit gene content patterns that correlate well with the geographical locations from which they were isolated (Rocap et al., 2003; Palenik et al., 2006; Zwirglmaier et al., 2008; Martiny et al., 2009). For the third framework, we divided the Roseobacter strains into five broad environmental categories that resulted in the following distribution: 15 isolates from the Atlantic Ocean, 1 isolate from the Indian Ocean, 6 isolates from the Pacific Ocean, 2 isolates from the polar oceans and 8 isolates that were cultured in association with eukaryotic organisms (Table 1). We hypothesized that characteristics distinguishing major environments (for example, access to nutrients, differences in temperature) would exert a detectable influence on gene patterns in the Roseobacter genomes.

There was a significant but very weak relationship between genome content and environmental origin of the isolate (ANOSIM R=0.296, P-value 0.002), although if only the Atlantic Ocean and Pacific Ocean isolates were compared, this relationship was stronger (ANOSIM R=0.398, P-value 0.002). Examination of ortholog patterns suggests that the relationship is based on the genomic distribution of motility genes, with 33% occurrence in Pacific Ocean isolates compared with 73% in Atlantic Ocean isolates; chemotaxis genes, with 17% occurrence in Pacific Ocean isolates compared with 47% in Atlantic Ocean isolates; denitrification systems, with 5% in Pacific Ocean isolates compared with 29% in Atlantic Ocean isolates; phosphorus uptake systems known to function at low phosphate concentrations (alkaline phosphatases, high-affinity phosphate uptake and phosphonate uptake), with 50% occurrence (12 out of 24 possible) in Pacific Ocean isolates compared with 85% (51 out of 60 possible) for Atlantic Ocean isolates; and aromatic carbon degradation pathways (mixed ocean basin patterns depending on the specific pathway) (Figure 4, all comparisons d-score P-value 0.01). There were also a number of unique carbon and ion transporters, amino-acid metabolism genes, and transcription regulators restricted to each ocean basin, and a large suite of unique genes shared by the two polar isolates (data not shown). We tested whether the higher frequency of coastal strains among the Pacific isolates compared with the Atlantic (Table 1) was the basis for the apparent ocean basin pattern, but found it not to be the case whether comparing whole-genome ortholog patterns (ANOSIM R=0.031, P=0.23 for coastal vs open ocean isolate comparison) or individual gene systems (Supplementary Figure S2).

Geographical patterns in the GOS data set

Despite the many factors that might obscure large-scale environmental imprints (including varied isolation methods, isolation dates spanning several decades and sparse spatial coverage), the ocean basin of isolation seemingly had predictive power for the distribution of select genes/pathways among the Roseobacter genomes. To determine whether this apparent grouping of genome content by geographical origin applies broadly to populations of roseobacters in the world oceans, we probed the GOS data set for similar environmental patterns. As the other two frameworks (phylogeny and lifestyle strategy) require assembled genomes, it is not possible to test for these among the GOS Roseobacter populations.

Homologs to genes listed in Supplementary Table S3 were identified in the GOS peptide sequence database (which currently does not include polar ocean metagenomic data). They were designated as Roseobacter homologs if they had greatest similarity to a gene in a Roseobacter genome in subsequent BLASTp query analysis against all available bacterial genome sequences (the CAMERA ‘All Prokaryotic Proteins (P) database’). Many of the same patterns in gene distribution found for cultured roseobacters were evident in the metagenomic analysis (Figure 6a). Most notable was that all phosphorus acquisition systems known to function at low phosphate concentrations (alkaline phosphatases, high-affinity phosphate uptake and phosphonate uptake) were much more abundant in wild roseobacters from the Atlantic Ocean, where the mean phosphate concentration is lower, than for either the Indian or Pacific Ocean (mean phosphate concentration is 0.06 μM for the Atlantic vs 0.15 μM for the Indian vs 0.53 μM for the Pacific; see Martiny et al., 2009 for details). The phosphate uptake system (pitA), which operates at high phosphate concentrations, had the opposite pattern, being more abundant in the Indian and Pacific Oceans (Figure 6a). Recently, other studies have noted similar trends for phosphorus gene distribution in the Prochlorococcus and SAR11 lineages (Rusch et al., 2007; Martiny et al., 2009), indicating that phosphorus concentration may impart a strong selective force on marine bacterial genomes. Of particular note were Roseobacter genes encoding for phosphonate uptake and assimilation, which exhibited a very large bias in distribution toward the Atlantic Ocean (Figure 6a).

Figure 6
figure 6

(a) Ocean basin (Atlantic, Indian and Pacific) three-way comparison of Roseobacter genes/gene pathways (see Figure 4). The position of each dot indicates the relative abundance of a gene in the Global Ocean Sampling (GOS) data set, based on 61, 52 and 45 Roseobacter genome equivalents in the Atlantic, Indian and Pacific Ocean data sets, respectively. Only genes present at a frequency 10% of genome equivalents in any one ocean basin are depicted. Triangle vertices indicate 100% relative abundance of that particular gene in the representative ocean. Three lines creating an inverted triangle have been drawn to aid in visual interpretation and indicate a relative abundance=50% in a single ocean basin. (b) Gene occurrence percentage of Roseobacter genome equivalents in the GOS data set (n=158) vs isolate genomes (n=32). Gene occurrence percentage >100% (that is, more than one copy per genome equivalent) in the GOS samples are represented as 100%. Gene descriptions are listed in Supplementary Table S2. GOS samples included in the comparison are listed in Supplementary Table S3.

Most representative genes we examined were more prevalent in the isolate genomes than in our per-genome-equivalent calculations for the GOS samples (Figure 6b), an observation that cannot be attributed to sampling disparities as 158 Roseobacter genome equivalents were sampled in the GOS (see Materials and methods). Compared with the roseobacters represented in culture, natural Roseobacter populations in the ocean are more likely to have genes for processing DMSP and utilization of C1 carbon compounds, but less likely to have genes involved in motility, adhesion, quorum sensing, gene transfer and iron uptake (Figure 6b). The higher prevalence of selected genes in the isolate genomes compared with GOS samples may indicate that there are fewer genes per genome in wild cells, could be indicative of the differences in sampling locations between the GOS samples and the isolates, or might reflect a bias during our analysis in selecting genes previously noted in cultured Roseobacter genomes. As the GOS samples were passed through a 0.8-μM filter before sequencing (Rusch et al., 2007), there is also poor representation of sequences from particle-associated cells.

Conclusions

Comparative genomic analysis of a bacterial lineage is a powerful approach for revealing ecological and evolutionary forces that influence genome content, and might form the basis for delineating ecologically differentiated clusters in nature. The substantial 16S rRNA sequence divergence within the roseobacters (11%; Buchan et al., 2005), currently spanning a minimum of 45 described genera, makes this the broadest marine bacterial lineage for which a comparative genomic analysis has yet been undertaken. This taxonomic level is consistent, however, with current methodological resolution in microbial ecology, including target groups for 16S rRNA probes and primers (Alonso-Sáez et al., 2007; Lami et al., 2009), and efforts to assign taxon-specific biogeochemical roles (Alonso and Pernthaler, 2006; Mou et al., 2008; Poretsky et al., 2010).

The three predictive frameworks examined here for the Roseobacter genomes have previously been shown to correlate with the genome content in bacterial taxa, including phylogenetic relatedness in Prochlorococcus (Garczarek et al., 2000; Bibby et al., 2003; Rocap et al., 2003), environmental resource partitioning in Vibrionaceae (a lineage with similar 16S rRNA divergence as the roseobacters; Hunt et al., 2008), and trophic strategies in bacterial endosymbionts and aquatic bacterioplankton (Moran and Baumann, 2000; Lauro et al., 2009). For the Roseobacter lineage, whole-genome content analysis of the 32 genomes produced 23 genome clusters (Figure 3) representing 20 unique combinations of clade, trophic strategy and environmental source. New sequences of Roseobacter strains may well increase the number of known genome clusters, particularly because two environmentally abundant 16S rRNA clades do not yet have reference genome sequences. While all three frameworks had statistically significant predictive power, none emerged as the potential overriding force imprinting Roseobacter genome content. Although other possible explanatory frameworks might have been considered here, all but two of the 23 genome clusters have unique clade–trophy–environment assignments (Figure 3), suggesting that these three frameworks together acceptably classify most of the variability in genome content.

The finding that trophic strategy correlates better than phylogeny or environment with Roseobacter gene inventories (ANOSIM, R=0.545 vs 0.410 vs 0.296) was not anticipated at the outset of our analysis, at least in part because it is not a correlate that has been widely examined for marine bacterial genomes. Nevertheless, the past decade has uncovered remarkable flexibility in the trophic strategies of marine bacterioplankton, suggesting that acquisition of alternate mechanisms for obtaining carbon and energy may be a strong evolutionary force in the ocean. The occurrence of several distinct trophic schemes within the taxonomically broad Roseobacter lineage provided an ideal opportunity to explore whether a bacterium's strategy for obtaining carbon and energy predicts other aspects of genome content. Differences in gene content among trophic groups were unfortunately dominated by hypothetical proteins, which provide little biological insight, although C1 and aromatic carbon oxidation genes and amino-acid transport and metabolism genes contributed to the signal. This concept of lifestyle imprinting of genome content, which has been explored in great detail for bacterial endosymbionts (for example, Moran and Baumann, 2000), may therefore also be important for understanding gene inventories of ocean microbes.

Roseobacter-like genes in the GOS data set showed significant variation in frequency across ocean basins, although only a fraction of all possible genes and gene systems appear to be shaped at this grand scale (Figure 6a). The GOS data set was also valuable for determining how well the genomes from the cultured roseobacters represent the repertoire and stoichiometry of genes in ocean-dwelling ‘wild’ roseobacters, an important perspective for assessing the relevance of this isolate-based genome analysis. Although the mismatch in frequency of some examined genes between isolates and the GOS data set suggests that the currently cultured strains may not yet provide a faithful representation of the prevalent natural Roseobacter populations, many genes and gene systems were indeed present at comparable frequencies (Figure 6b).

Overall, our analysis has firmly established roseobacters as ecological generalists, harboring large gene inventories and a remarkable suite of mechanisms by which to obtain carbon and energy. Further, this comparative analysis has illustrated that members of the lineage cannot be easily condensed into a few ecologically differentiated clusters; rather, each genome is largely unique in its assortment of genes for acquisition and transformation of carbon and nutrients. The fact that the best framework for predicting genome content is lifestyle strategy, not phylogeny, indicates that horizontal gene transfer and homologous recombination may be particularly dominant evolutionary forces in this marine bacterial lineage (possibly facilitated by an unusual gene transfer agent system that is prevalent; Biers et al., 2008; Zhao et al., 2009). Further insights into correlates of genome content, coupled with continued efforts to identify Roseobacter genes that are common in the world oceans, will better elucidate the functional roles of roseobacters in marine ecosystems.