The proportion of conserved DNA sequences with no clear function is steadily growing in bioinformatics databases. Studies of sequence and structural homology have indicated that many uncharacterized protein domain sequences are variants of functionally described domains. If these variants promote an organism's ecological fitness, they are likely to be conserved in the genome of its progeny and the population at large. The genetic composition of microbial communities in their native ecosystems is accessible through metagenomics. We hypothesize the co-variation of protein domain sequences across metagenomes from similar ecosystems will provide insights into their potential roles and aid further investigation.
We calculated the correlation of Pfam protein domain sequences across the Global Ocean Sampling metagenome collection, employing conservative detection and correlation thresholds to limit results to well-supported hits and associations. We then examined intercorrelations between domains of unknown function (DUFs) and domains involved in known metabolic pathways using network visualization and cluster-detection tools. We used a cautious “guilty-by-association” approach, referencing knowledge-level resources to identify and discuss associations that offer insight into DUF function. We observed numerous DUFs associated to photobiologically active domains and prevalent in the Cyanobacteria. Other clusters included DUFs associated with DNA maintenance and repair, inorganic nutrient metabolism, and sodium-translocating transport domains. We also observed a number of clusters reflecting known metabolic associations and cases that predicted functional reclassification of DUFs.
Critically examining domain covariation across metagenomic datasets can grant new perspectives on the roles and associations of DUFs in an ecological setting. Targeted attempts at DUF characterization in the laboratory or in silico may draw from these insights and opportunities to discover new associations and corroborate existing ones will arise as more large-scale metagenomic datasets emerge.
Citation: Buttigieg PL, Hankeln W, Kostadinov I, Kottmann R, Yilmaz P, Duhaime MB, et al. (2013) Ecogenomic Perspectives on Domains of Unknown Function: Correlation-Based Exploration of Marine Metagenomes. PLoS ONE 8(3): e50869. https://doi.org/10.1371/journal.pone.0050869
Editor: Iddo Friedberg, Miami University, United States of America
Received: July 22, 2012; Accepted: October 24, 2012; Published: March 14, 2013
Copyright: © 2013 Buttigieg et al. This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Funding: This study was funded by the Max Planck Society (http://www.mpg.de/en). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The authors have declared that no competing interests exist.
In recent years, genomic sequencing projects have revealed a large number of novel genes across a wide range of organisms and environments. Many of these have poor sequence-level similarity to genes that have been characterized in a laboratory setting and, consequently, have not been annotated with functional roles. These ‘hypothetical’ genes are becoming increasingly prevalent in bioinformatics databases. For example, the Pfam 24 database  stored some 11,912 protein domain families derived from conserved sequence data with ∼26% dubbed “domains of unknown function” (DUFs). This proportion is predicted to soon overtake that of functionally characterized domains , and calls for community action  and cross-disciplinary efforts  towards their identification have been made.
Several groups have employed biochemical and molecular techniques in DUF characterization. Among these, Deng et al.  characterized members of a Pfam family, formerly known as DUF62, as S-adenosyl-l-methionine hydroxide adenosyltransferases using a range of enzymatic analyses. Similarly, Weinitschke et al. characterized members of the DUF81 family as the sulfite exporter TauE by examining co-transcription of genes in the metabolism of C2 sulfonates . Computational approaches have also been used to aid DUF characterization. Goonesekere et al.  applied secondary structure analysis and the use of three-dimensional homology models to functionally annotate 8 DUFs. On a larger scale, Jaroszewski et al.  used structural genomics approaches to determine the three-dimensional structures for more than 250 DUF families, granting insights into their potential activities. This latter study reported that the majority of DUFs analyzed were either divergent structural variants of well-characterized families or showed notable substructure similarity to known proteins. The authors inferred that these variations may have been conserved as they extended an organism's functional repertoire in an ecologically adaptive manner. Here, we explored a marine metagenomic dataset to gain an ecological perspective on DUF functionality.
Metagenomics allows insight into an entire ecological community's genetic content , , providing a rich source of in situ sequence data from a range of ecosystems. These metagenomic resources are ripe for mining, with intra-ecosystem variation in microbial function cited as a high-value target for exploration . If genomic conservation of DUFs is adaptive in nature, then it is probable that these DUFs are co-selected with other domains linked to that adaptation. We thus hypothesized that the abundances of co-selected domains will co-vary across metagenomes from similar ecosystems. By identifying co-varying groups of domains, it may be possible to speculate on DUF functionality using a ‘guilty-by-association’ approach . This approach is often used in the identification of novel metabolic modules by relating experimental perturbations to subsequent gene expression patterns . In this model, genes with similar responses are grouped into putative modules to guide more rigorous future investigation. Further, correlative approaches have been employed to explore ecosystem-level interactions between microbial taxa and their environment , .
Here, we propose that correlations of protein domain sequences found in intra-ecosystem metagenomic data may provide insight into the functional roles of DUFs. We critically examined the correlation of protein domains of known and unknown function across the Global Ocean Sampling (GOS) collection , comprising a globally distributed set of epipelagic microbial community metagenomes. We found several sets of associations that show promise in guiding DUF characterization in an experimental setting.
Results and Discussion
Pfam versions in analysis and interpretation
Here, we queried raw sequences of selected GOS metagenomes against version 24 of the Pfam database . Of the 3,069 DUFs in Pfam v24, we detected 2,531 in our analysis. In subsequent Pfam database releases, many models and their functional descriptions have been updated to reflect the growing knowledge surrounding them. The associations presented here are best validated by domain-level knowledge. We thus referred to Pfam v26  when discussing our results, allowing several correlative hypotheses to be evaluated against later functional characterizations.
Removal of rows (sites) and columns (Pfams) dominated by zero abundances was performed to decrease the sparseness of the dataset and focus on observed variation. The dataset was thus reduced from a matrix of 80 GOS metagenomes and 3,587 Pfam domains to 69 sites and 1,863 Pfam domains. A total of 670 of the domains retained were DUFs. We standardized abundances by their site maxima to mitigate the effect of differing sample size across metagenomes; however, the co-variation of domain abundances due to sample size may still provide insights into their potential associations, assuming proportionality. Thus, we used the unstandardized matrix (UM) and its standardized counterpart (SM) in further analysis. Totals of 975 (UM) and 167 (SM) Pfams were retained following removal of those with no correlations stronger than a rho of 0.80 and Bonferroni-corrected P-values below 1×10−6. These included 225 (UM) and 75 (SM) DUFs. Correlation of domains in the UM resulted in largely Gaussian distribution of rho coefficients (mean≈0.41, s.d.≈0.26); however, a slight second peak around a rho of 0.80 was observed (Figure S1). Correlation of domains in the SM displayed a more Gaussian distribution of rho values (mean≈0.03, s.d.≈0.19); however, with a positive skew (Figure S2). Based on these distributions, we chose a correlation threshold and conservative significance values to focus on robust associations.
In the UM, a total of 94 DUFs showed biased correlations to Pfam domains assigned to a single metabolic category (Table 1). Of these, 56 were biased towards photobiological processes (Tables S1 and S2), many of which were also associated with a septum formation inhibitor domain (MinC_C) involved in the control of cell division. In the analysis of the SM, 48 DUFs showed biased correlations, with 45 in favor of photobiological processes (Table S3). Of these 45, 15 showed some bias (≥40% of the maximum correlated category) towards domains in carbohydrate and/or coenzyme metabolism. The three remaining DUFs were associated to domains in inorganic ion transport and metabolism (DUF1008); carbohydrate transport and metabolism (DUF111); and transcription (DUF37). Such biased connectivity of DUFs to domains in a particular functional category is not a guarantee of related function; however, it provides associative context for hypothesis generation.
General network characteristics.
The networks generated from the correlation of the UM and SM described above comprised 975 nodes and 166,232 edges and 167 nodes and 1,897 edges, respectively. The network derived from the unstandardized abundances contained two regions of highly interconnected vertices, each with a mesh-like topology (Figure 1, Boxes 1 and 2; see Figure S3 for node labels). The larger of these regions (Box 1) contained a core of highly interconnected nodes with domains from a wide variety of metabolic categories, as well as ‘spokes’ of less interconnected nodes. The highly-enmeshed topology of this large connected complex (LCC) prevents speculation on DUF function by visual inspection alone. However, a spoke of the UM LCC (Figure 1, Box 3) containing DUF2805 and DUF37 was defined by domains linked to protein biosynthesis at the level of translation (Table 2).
Nodes represent Pfam domains and edges correlations greater than a Spearman's rho of 0.80. Shorter edges indicate stronger correlations. A large network with two enmeshed regions (Box 1 and 2) bridged by a small number of nodes (Box 4) dominates the graph. Several small networks of functionally related nodes are also present (Box 5, inset). Node colors represent functional categories; refer to Figure S7 for description. See text for detailed descriptions.
Network regions with photobiological relevance.
The smaller region of this network (Figure 1, Box 2) was dominated by DUFs and included ten domains implicated in photobiological processes (Table S4). These included a central photosystem II domain, a subunit of the photosystem I reaction center, a domain that targets precursors of light-harvesting prosthetic groups, and a proton-pumping domain essential for photoheterotrophic cell growth. The taxonomic distribution of the DUFs in this region (after Pfam v26) was heavily weighted toward the Cyanobacteria and other photobiologically active taxa, such as the Viridiplantae. It is premature to connect the DUFs in this region directly to phototrophic mechanisms; however, a photoresponsive mode of life may have some bearing on their function and association in marine metagenomes.
Additionally, we noted the domains bridging this putative module to the bulk of the LCC were also photobiologically relevant (Figure 1, Box 4; Table S5). For example, eight domains were involved in cobalamin metabolism and the large and small subunits of RuBisCO were also present. Cobalamin (vitamin B12) is characterized by a corrin ring, which is chemically similar to the porphyrin ring found in heme, chlorophyll, and cytochrome. Further, cobalamin-dependent methyltransferases are involved in C1 metabolism and CO2 fixation .
We observed an SM-derived network (Figure 2, i; see Figure S4 for node labels) with similar membership to the bridge and photobiological regions discussed above (Table S6). This network featured 112 Pfams, which included DUFs (n = 72), photoreactive domains (n = 10), and coenzyme transport and metabolism domains (n = 8). We also observed a network (Figure 2, ii) with similar membership to the spoke comprising domains related to protein biosynthesis discussed above (Figure 1, Box 4).
Abundances were standardized by site maxima. Nodes represent Pfam domains and edges correlations greater than a Spearman's rho of 0.80. Shorter edges indicate stronger correlations. The largest network (i) is dominated by DUFs and domains linked to photobiological processes. Domains linked to urea metabolism were also present in this network (hollow arrowheads). Smaller networks featured domains linked to translation (ii), phosphonate metabolism (iii), and cyanophage activity (iv). Numerous pairs of functionally related domains were also present. Node colors represent functional categories; refer to Figure S7 for description. See text for detailed descriptions.
Networks with known functional relatedness.
A number of comparatively small UM-derived networks were also observed (Figure 1, Box 5 & inset), including domain groups known to be functionally related. Detection of correlative associations between such domains is encouraging and lends merit to the potential of metagenomic datasets in functional module detection. For example, we observed a group of phosphonate metabolism domains as well as a group of urea metabolism domains in the UM-derived networks (Figure 1, inset, i and ii resp.). Analysis of the standardized data also revealed the phosphonate network (Figure 2, iii), as well as some of the urea metabolic domains noted above, embedded in the LCC (Figure 2, hollow arrowheads). Phosphonate and urea are significant components of the marine dissolved organic matter pool and numerous microbial genera, such as Prochlorococcus, possess corresponding uptake and utilization capacities –. The urea metabolism domains were associated with a bacterial transglutaminase-like domain (Bact_transglu_N), DUF403, DUF404, and DUF407 in the UM-derived networks. These latter domains occurred on the periphery of the SM-derived LCC (Figure 2, solid arrowhead). In Pfam 26, DUF407 has been reclassified as a circularly permuted ATP-grasp family (CP_ATPgrasp_1) and DUF404 has been merged with this family. Further, Pfam architectures place instances of DUF403, now referred to as Alpha-E, in proximity to both Bact_transglu_N and CP_ATPgrasp_1 in microbial genomes. Goonesekere et al.  also observed that DUFs 404 and 407 co-occur. Together, these domains function in a peptide synthesis/modification system .
We observed a number of associations between N- and C-terminal domain abundances in analysis of both the SM and UM. Examples include arginine-tRNA-protein transferase (ATE_N; ATE_C), coenzyme F420 hydrogenase (FrhB_FdhB_N; FrhB_FdhB_C), and domains involved in eukaryotic DNA polymerase processivity (PCNA_N; PCNA_C).
We also observed several small networks with DUFs associated to domains that may shed light on their roles. Notably, a haemin-degrading domain was associated with DUF1008 in analysis of both the SM (Figure 2, v) and UM (Figure 1, inset, v). DUF1008 has been reclassified as a haem utilization ChuX/HutX domain in Pfam v26, which corroborates this correlative association. Further, a photoreaction center domain was associated with DUF1825 and DUF3110 in both the SM (Figure 2, iv) and UM analyses (Figure 1, inset, iii). These DUFs principally occur in the Cyanobacteria; however, their specific correlation with the photoreaction center domain is likely caused by the co-occurrence of these domains in marine cyanophages, which infect the genera Synechococcus and Prochlorococcus. Previous work has identified a number of cyanophage-borne photosynthetic genes which are transferred to their hosts, including psbA and psbD , both of which bear the Photo_RC domain. Further, DUF3708, annotated as a phosphate ATP-binding cassette transporter, was associated with DUF3333 in the UM-derived networks. Both these DUFs are prevalent in the Alphaproteobacteria and Pfam architectures note them in proximity to the inner membrane component of a binding-protein-dependent transport system (BPD_transp_1). DUF137 was associated with two ATP synthase subunits. We observed these three domains to be prevalent in the Archaea, thus, attributing DUF137 with ATP synthase activity is premature, as taxonomic restriction may also explain their correlation. A similar explanation can be applied to the association of DUF655 and DUF54, the former domain reportedly resembling a ribosomal protein . DUF1297 and DUF1246, associated as a pair, share several Pfam architectures, are listed as Pfam interaction partners, and are also prominent in the Archaea. Similarly, a pair comprising DUF126 and DUF521 shared similar taxonomic distributions, occurring primarily in the Bacteria and Archaea. The remaining networks, often featuring few domains with broad taxonomic distribution and poor characterization, were not amenable to our ‘guilty-by-association’ interpretative approach.
Associations between DUFs in highly-enmeshed regions are difficult to evaluate by visual inspection. Transitivity clustering  offers a means to detect interconnected substructures in these regions, isolating them by adding and removing edges against a cost function (in this case derived from correlation strength). Our application of this method produced 49 clique-like transitivity clusters (TCs) with 3 or more members from the UM (Figure 3; see Figure S5 for node labels) and 13 such clusters from the SM (Figure 4 see Figure S6 for node labels).
Edge-weights (correlations) determine the cost of adding or removing edges during clustering. We observed clusters with domains linked to photobiology; oligotrophic adaptations; DNA maintenance and repair; and iron supply. Node colors represent functional categories; refer to Figure S7 for description. See text and Tables 3, 4, and 5 for detailed descriptions.
Edge-weights (correlations) determine the cost of adding or removing edges during clustering. The largest cluster contained DUFs and domains linked to photobiology (Table S5). Node colors represent functional categories; refer to Figure S7 for description. See text for detailed descriptions.
Much of the highly-enmeshed core of the UM-derived LCC remained in a single cluster (Figure 3: TC1, n = 464) which included 30 DUFs. The density and functional diversity of this region prevented speculation on DUF function. However, domains with known and related functions were clustered, including the urea and phosphonate metabolic components observed above (Figure 3: TC40, TC44; Figure 4: TC12, TC23). Further, the DNA mismatch repair domains, MutS I–V (Figure 3: TC14; Figure 4: TC37), which were initially obscured within enmeshed network regions, were clustered.
We observed prominent transitivity clusters with photobiologically relevant domain membership. A UM-derived cluster (Figure 3: TC2; Table S7) contained most domains from the photobiological region of the original network (Figure 1, Box 2). The membership of the largest SM-derived cluster (Figure 4: TC1; Table S8) was similar and included 28 DUFs, five photobiological domains, two cobalamin metabolism domains, and a fructosamine kinase domain.
Notably, several photobiologically relevant domains occurred in separate clusters. In the UM-derived clusters, a photosystem I reaction center domain was clustered with five DUFs, while a domain that has been implicated in the assembly and stability of photosystem I complex in chloroplasts was clustered with four DUFs (Figure 3: TC9 and TC15 resp.). Further, a photosystem I reaction center domain (PsaL) was clustered with DUF2839 (UM) and DUF1824 (SM). Lastly, Photo_RC and DUF1825 were associated in an UM-derived cluster (TC73) while an SM-derived cluster also included DUF3110 (TC5). These DUFs have distinct taxonomic distributions, which may partially account for their clustering (see Taxonomic perspectives, below). Cobalamin synthesis domains, which may participate in pigment synthesis, were clustered with the septum formation inhibitor domain, MinC_C. MinC_C shares a Pfam architecture with an amidase domain involved in cobalamin synthesis (CbiA) and was clustered with a domain which catalyzes the conversion of cobalamin into its coenzyme form; DUF3531 which occurs primarily in the Cyanobacteria and Viridiplantae; and DUF3104 which occurs primarily in the Cyanobacteria (Figure 3: TC27).
Several UM-derived clusters, discussed below, appeared to have a common functional theme, which suggested putative roles for the DUFs they contain.
We observed three UM-derived clusters (Table 3) featuring domains with special relevance in nutrient-limited environments, including capacities to flexibly metabolize nutrients, repress costly biosynthetic pathways, and cope with metabolic stress. Indeed, many marine microbes are adapted to low or intermittent nutrient availability and such adaptations are reflected in their genomes , . A cluster comprising 13 members (Figure 3: TC4) included domains involved in molybdenum cofactor biosynthesis; nitrate assimilation and nitrate inducible dehydrogenase activity; phosphatase activity; oxygenation of nitrogen, sulfur, phosphorus and selenium atoms in xenobiotics; carbon-sulfur and carbon-nitrogen bond chemistry; and a chorismate mutase domain, whose activity promotes phenylalanine and tyrosine biosynthesis at the expense of tryptophan synthesis. Minimizing the synthesis of biochemically costly amino acids such as tryptophan  is a viable strategy to promote ecological competitiveness in oligotrophic conditions. Continuing this theme, a cluster including DUF3047 and DUF2155 (Figure 3: TC12) contained domains that catalyze the cleavage and cannibalization of nitrogen-bearing components from polyamines and purines. Lastly, a four-membered cluster (Figure 3: TC19) contained an organic solvent tolerance protein; a regulator of the acetate and glycerol operons and glyoxalate shunt; a domain involved in monitoring cellular nitrogen levels, the nitrogen stress response, and with possible roles in iron metabolism; and a fatty acid degradation regulatory domain, also found in regulators of sugar biosynthesis operons and iclR activators. These associations suggest that DUFs in these clusters may contribute to metabolic adaptations to oligotrophic conditions.
DNA metabolism and repair.
Two UM-derived clusters were characterized by DNA metabolism and maintenance domains (Table 4). The microbial communities of epipelagic waters face sustained solar irradiation, which causes DNA damage . It is likely the following clusters reflect adaptations to cope with this environmental threat. A cluster of 9 nodes (Figure 3: TC5) included amino acid dehydrogenase domains; domains involved in the synthesis of purines and aromatic amino acids; domains involved in DNA synthesis and repair; and a tRNA synthetase domain. Milligan et al.'s in vitro finding on the ability of several amino acids to repair oxidative DNA damage by reducing guanyl radicals  may offer some insight into the presence of amino acid dehydrogenases in this cluster: epipelagic microbes may similarly couple amino acid oxidation and DNA radical deactivation. Another cluster (Figure 3: TC39) comprised DUF836, classified as a glutaredoxin-like domain; an inorganic pyrophosphatase domain; and a domain involved in recombination, repair of double strand DNA breaks, and resistance to irradiative and chemical DNA-damage. As a glutaredoxin-like domain, DUF836 may participate both in DNA metabolism through glutathione-dependent synthesis of deoxyribonucleotides and in antioxidant defense . Glutaredoxins may also serve in the assembly and transfer of iron/sulfur complexes  and thus may have particular importance in safeguarding DNA integrity and metabolism in iron-limited marine waters.
Seawater iron concentrations have previously been linked to iron metabolism genes in GOS , and the importance of iron supply may offer insight into two further clusters (Table 5). A four-membered cluster (Figure 3: TC23) contained DUF255; domains of an iron/manganese superoxide dismutase; and the Gram-negative bacterial TonB domain. TonB works in conjunction with outer membrane transport proteins in the active uptake of siderophore-bound iron (II) and cyanocobalamin. The ability of the TonB system to divert the proton motive force in aid of active iron transport is likely to be an asset in the typically iron-limited marine water column. Close association of this iron (II) supply-line to iron-dependent superoxide dismutases may be due, once again, to the threat of radicals generated in the irradiated water column. Another association of iron transport and iron-dependent enzyme domains was observed in a cluster (Figure 3: TC24) containing DUF58; a pair of pyruvate oxidoreductase domains; and a periplasmic binding protein domain often involved in iron transport. Pyruvate oxidoreductase has multiple iron/sulfur clusters and participates in the metabolism of short carboxylic acids as well as the carbon-fixing, reductive carboxylate cycle.
We observed several UM-derived clusters that we could not place in a larger interpretive framework, but which included DUFs in a noteworthy functional context (Table S9). For example, DUF484 and DUF2066 were clustered with domains involved in sodium translocation linked to amino acid transport and redox as well as a ligase that initiates the glutathione biosynthesis pathway (Figure 3: TC11). The glutathione pathway is implicated in a wide range of cellular functions, including amino acid transport, and has been specifically linked to sodium-dependent transport in eukaryotic systems , . Next, DUF37, annotated as a haemolytic domain in Pfam v26, was clustered with a number of domains implicated in transcription and ribosomal function (Figure 3: TC7). This cluster is similar to a network derived from the SM (Figure 2, ii) and a spoke of the UM (Figure 1, Box 3; Table 2). Another cluster was characterized by domains capable of phosphorylation (Figure 3: TC10) and contained a DUF (DUF894) reclassified as a transmembrane secretion effector in a major facilitator superfamily (MFS_3); several kinase domains including a DUF (DUF227) reclassified as a kinase targeting insect hormones (EcKinase); and a selenium transferase domain. Phosphorylation plays a key role in selenocysteine formation through selenophosphate in Archaea and Eukarya. A pair of proline metabolism domains was clustered with DUF525 and DUF461 (Figure 3: TC29). Lastly, DUF192, putatively involved with extracellular sugar processing, was clustered with an isomerase domain, which functions in both the pentose phosphate pathway and the Calvin cycle, and a Type I restriction modification DNA specificity domain (Figure 3: TC46). This latter domain targets and either degrades DNA foreign to bacterial cells (such as viral DNA) or methylates DNA. This cluster may reflect the coupling of viral defense mechanisms to sugar scavenging – a possible adaptation to capitalize on the ‘spoils of war’ in resource limited environments. Despite their isolation, the above clusters offer interesting perspectives on the involvement of DUFs in epipelagic community metabolism.
Several SM-derived clusters may also grant interesting perspectives on DUF co-occurrence. An Mg2+-dependent acid phosphatase involved in the biosynthesis of several cofactors including cobalamin and heme was clustered with six DUFs (Figure 4: TC2). These DUFs are prevalent in the Cyanobacteria (see below) with very low representation in the Proteobacteria, a restriction that may account for their clustering. However, domains in several smaller clusters showed dissimilar taxonomic distributions. These included a cluster (Figure 4: TC6) comprising a voltage-gated chloride channel, DUF2930, and DUF2214, the latter predicted to be a membrane protein. Pairs of domains with dissimilar taxonomic distributions were also observed: DUF3531 and DUF3641; a cobalamin-5-phosphate synthase domain (CobS) and DUF3727; a septum formation inhibitor (MinC_C) and DUF3119; DUF2010, reclassified as a mitochondrial PGP phosphatase, and DUF1823; and a divalent ion tolerance protein (CutA1) and DUF92. Comments on other transitivity clusters are available in the supplementary Material S1.
The taxonomic distribution of Pfam domains, particularly those present in the highly abundant marine Cyanobacteria, will inevitably influence their association and condition any interpretation. We retrieved the taxonomic distribution of the DUF families analyzed above from the Pfam web-portal to qualitatively contextualize the observed associations. However, quantifying the degree to which the taxonomic make-up of microbial communities confounds functional associations in metagenomic samples is a non-trivial task. Such assertions are contingent on the taxonomy and functional annotation of the current genome collection, which is unlikely to reflect the true in situ diversity. Further, correct assignment of sequencing reads to known taxa is often problematic. Thus, the taxonomic distributions presented below are intended to provide a tentative context to advise hypothesis generation (as demonstrated above) and are not intended as a basis for phylogenomic profiling.
In the UM-derived clusters (Figure 3, Table 6), four phyla contained greater than 5% of DUF instances, namely the Proteobacteria, Firmicutes, Actinobacteria, and Cyanobacteria. The DUFs of the largest UM-derived transitivity cluster (TC1) had similar distributions to the whole collection, while DUFs whose instances were concentrated in cyanobacterial genomes dominated the second cluster (∼55%). The proteobacterial proportion of this latter cluster was dominated by the Gammaproteobacteria (∼56%) and a proportion of Alphaproteobacteria (∼23%). The genera Rhodobacterales and Rhodospirillales, known to possess proteorhodopsin genes encoding light-powered proton pumps, were present in the alphaproteobacterial division. As reviewed by DeLong and Béjà , there is increasing evidence that proteorhodopsin pumps and light-powered heterotrophy are more broadly distributed in epipelagic bacteria than previously thought. In fact, proteorhodopsin pumps have recently been identified in marine eukaryotes  suggesting the limits of their occurrence in marine microbes has yet to be fully established. This suggests that the correlation of these DUFs may not be entirely due to their restriction to photoautotrophs. Indeed, DUFs with distributions more restricted to photoautotrophs were clustered separately. For example, less than a percentage of organisms bearing DUFs from TC9 were Proteobacteria and those from TC15 were almost exclusively Cyanobacteria (∼95%). Other clusters with conspicuous taxonomic restrictions included TC31, with domains prevalent in actinobacterial genomes, and TC43, with domains prevalent in bacterial, fungal, and plant genomes.
The DUFs in SM-derived clusters (Figure 4, Table 7) were predominately found in cyanobacterial genomes (≥50% of all occurrences). Exceptions included TC6 and TC7, which contained DUFs comparably distributed between proteobacterial and cyanobacterial genomes, as well as TC14 which featured DUFs distributed in the Fungi, Firmicutes, and Cyanobacteria.
Assigning the associations discussed above with meaningful measures of confidence is an immediate concern. False positives may lead experimentalists in fruitless directions, while false negatives may limit functional discovery. Similar difficulties are encountered when attempting to benchmark protein interaction networks and attempts to minimize them rely on the reproducibility of interactions across datasets or the use of well-characterized model systems or gold standards .
Several aspects of metagenomic data currently hinder the construction of such benchmarks. Firstly, metagenomes are likely to contain genome fragments from organisms with no metabolically well-characterized counterparts. This greatly weakens the credibility of model-systems or –organisms as gold standards. Secondly, studies on the scale of the GOS expedition are, presently, difficult to replicate. The increasing collection of marine metagenomes, fed by initiatives such as TARA Oceans , as well as the growing genome collection driven by programs such as the Genomic Encyclopedia of Bacteria and Archaea , may soon offer the opportunity to construct such standards. However, experimental confirmation or falsification of our assertions by bench scientists is perhaps the most conclusive basis for evaluation. In the spirit of initiatives such as the Computational Bridge to Experiments (ComBrEx) , groups with the infrastructure and expertise to test in silico predictions in vitro may, en masse, provide a degree of confidence estimation for studies similar to ours. DUFs which have been described as at least partially characterized in subsequent releases of Pfam may anticipate such estimation. Relative to Pfam v24, we noted 28 DUF families present in our UM and SM datasets which have been renamed or merged into existing families in Pfam v26 (Table S10). In the UM-derived data, eight of these DUFs were not clustered while a further eight occurred in the highly-enmeshed TC1 (Figure 3). We were thus unable to compare these updates to our results. However, the updated descriptions of six of the remaining DUFs – DUF989, DUF403, DUF404, DUF407, DUF227, and DUF1008 – reflect their clustering in this analysis. For example, a heme-binding domain towards the C-terminus of DUF989, now merged with the Cytochrome-C family, is clustered with families such as CobT, involved in cobalamin synthesis. The chemical similarity of the corrin and porphyrin rings found in cobalamin and heme, respectively, may provide a basis for the co-occurrence of these domains. The remaining DUFs have been compared to their updated descriptions in the sections above.
While the novelty and scale of metagenomic datasets present obstacles in constructing gold standards, these factors are assets in inferring functionality through intra-ecosystemic domain covariation. We attempted to limit false assertions by using conservative detection and correlation thresholds coupled with knowledge-guided curation of our results. These approaches were used to gauge if domain associations offered plausible insights into DUF functionality in an epipelagic setting, and noteworthy associations were discussed above.
We observed that intercorrelation of protein domain sequences across intra-ecosystem metagenomic datasets can provide perspectives on the potential roles of domains of unknown function. Naturally, even strong correlation across metagenomic datasets cannot provide direct functional annotations, as numerous factors may account for domain covariance in natural systems. However, critically evaluating strongly correlated domains with knowledge-level resources can provide an interpretive context to complement more targeted efforts in DUF characterization. The observed association of domains involved in microbial phosphonate metabolism, urea metabolism, and other ecologically relevant capacities encourage this approach. Of the 225 and 75 DUFs retained following correlation analysis in the UM and SM datasets respectively, we detected 94 (UM) and 48 (SM) DUFs with connectivity biased towards a single metabolic category. Further, the results above list 73 DUFs from the UM-derived and 41 DUFs from SM-derived networks and transitivity clusters whose associations may reflect ecological features of the marine epipelagic zone. While these results represent only a fraction of the DUFs detected in the GOS metagenomes, this analysis is an initial step in using ecogenomic variation to assist functional discovery. The opportunity to routinely perform such exploratory analyses and establish quantitative benchmarks is emerging as data from large-scale metagenomic, metatranscriptomic, and metaproteomic sampling campaigns becomes publically available. The perspectives that can be derived from this data will almost certainly forward efforts to characterize DUFs where homology-based approaches cannot.
Materials and Methods
Detecting domains of unknown function
A collection of 10,133,846 unassembled reads from the Global Ocean Sampling expedition metagenomes GS000a-GS023, GS025-GS051, GS108a-GS117b, GS119-GS123, GS148-GS149, and MOVE858  were downloaded from the CAMERA web-portal  and queried against all hidden Markov models (HMMs) present in the Pfam 24 database using the HMMER3 software (version 3.0b3). Hits were deemed significant if their domain independent E-value was less than or equal to 1e-3, their bias composition correction was at least an order of magnitude less than their full score, the length of the query alignment was at least 20% of the query length, and the model alignment was at least 20% of the HMM length. Results were stored in a relational database and cross-tabulated into a “site×Pfam” matrix, wherein the abundance of each Pfam at a given site was enumerated. Pfams assigned to COG functional metabolic categories available from the integrated microbial genomes (IMG) system , as well as an additional category for photobiologically active domains (Table S1), were used in further analyses.
Correlation analysis and network visualization
The site×Pfam matrix described above was imported into the R statistical computing environment (http://cran.r-project.org/). Pfam categories listed in more than one metabolic category were removed. Distributions of total, non-zero abundances across Pfam categories and sites were used to advise data preparation. Categories with less than 20 non-zero abundances across the 80 sites analyzed and sites with less than 1,000 non-zero abundances across the 3,587 Pfam categories analyzed were removed. A copy of the resulting matrix was subject to row standardization, whereby the Pfam abundances across a given site (row) were divided by the maximum Pfam abundance of that site. The Spearman's rank correlation of Pfam categories in both matrices was determined using the rcorr() function from the R package Hmisc. Pfam categories with no correlations greater than a rho of 0.80 and with a P-value less than a Bonferroni corrected cut-off of ∼1×10−6 were removed. DUFs that showed correlations biased towards Pfams in a single metabolic category were noted. Bias was declared when the number of correlations of a DUF to Pfams in a given category was at least double that of correlations to any other category.
The igraph R package  was used to create an adjacency matrix from these correlation results. This was imported into Cytoscape  for visualization. Network vertices, each corresponding to a Pfam category, were connected by an edge if their correlation satisfied the thresholds stated above. The Spearman's rho statistic provided weights in an edge-weighted, spring-embedded visualization. Pfam categories were color-coded according to their assigned metabolic category.
Networks were manually inspected for distinct topological features, particularly those where DUFs were associated with characterized Pfams from a narrow range of COG categories. The domains comprising these features were investigated further by retrieving descriptions from Pfam 26 via its webportal (www.pfam.sanger.ac.uk) and further literature where appropriate.
The TransClust ,  algorithm was run from the Cytoscape plugin, clusterMaker , to detect clique-like clusters. TransClust was run with a maximum sub-cluster size set to 50, a maximum time allowance of 2 seconds to execute each loop in the algorithm, and using edge weight (correlation) as an array source. The resulting clusters were visualized and evaluated as described above.
The DUFs in several transitivity clusters were annotated with their collective taxonomic distribution. Distributions were retrieved from the Pfam website (www.pfam.sanger.ac.uk; 2012-02-10) and the results stored in a relational database. Only clusters with ≥4 members and ≥2 DUFs were examined and those with notable distributions were discussed.
Comments on selected transitivity clusters not discussed in text.
Pfam 24 domains categorized as photobiologically relevant in this analysis.
DUFs with correlative bias towards photobiologically relevant domains (unstandardized data).
DUFs with correlative bias towards metabolic categories (standardized data). Refer to Table 1, footnote for list of abbreviations.
Histogram of correlation strength between unstandardized abundances of Pfam domains across GOS metagenomes. The distribution's mean and standard deviation were ∼0.41 and ∼0.26 respectively. A normal distribution with equal mean and standard deviation is indicated by a blue contour.
Histogram of correlation strength between abundances of Pfam domains across GOS metagenomes, standardized by site maxima. The distribution's mean and standard deviation were ∼0.03 and ∼0.19 respectively. A normal distribution with equal mean and standard deviation is indicated by a blue contour.
High resolution, unannotated version of Figure 1 with individual node labels.
High resolution, unannotated version of Figure 2 with individual node labels.
High resolution, unannotated version of Figure 3 with individual node labels.
High resolution, unannotated version of Figure 4 with individual node labels.
Integrated the Pfam domain taxonomic data into the study: WH. Supervised the project: FOG. Reviewed the manuscript: WH IK RK PY MBD FOG PLB. Conceived and designed the experiments: RK PLB. Performed the experiments: IK RK PLB WH. Analyzed the data: PLB WH MBD PY. Wrote the paper: PLB.
- 1. Finn RD, Mistry J, Tate J, Coggill P, Heger A, et al. (2010) The Pfam protein families database. Nucleic Acids Res 38: D211–222.
- 2. Bateman A, Coggill P, Finn RD (2010) DUFs: families in search of function. Acta Crystallogr Sect F Struct Biol Cryst Commun 66: 1148–1152.
- 3. Roberts RJ (2004) Identifying protein function–a call for community action. PLoS Biol 2: E42.
- 4. Galperin MY, Koonin EV (2010) From complete genome sequence to “complete” understanding? Trends Biotechnol 28: 398–406.
- 5. Deng H, McMahon SA, Eustáquio AS, Moore BS, Naismith JH, et al. (2009) Mechanistic insights into water activation in SAM hydroxide adenosyltransferase (duf-62). Chembiochem 10: 2455–2459.
- 6. Weinitschke S, Denger K, Cook AM, Smits THM (2007) The DUF81 protein TauE in Cupriavidus necator H16, a sulfite exporter in the metabolism of C2 sulfonates. Microbiology 153: 3055–3060.
- 7. Goonesekere NCW, Shipely K, O'Connor K (2010) The challenge of annotating protein sequences: The tale of eight domains of unknown function in Pfam. Comput Biol Chem 34: 210–214.
- 8. Jaroszewski L, Li Z, Krishna SS, Bakolitsa C, Wooley J, et al. (2009) Exploration of uncharted regions of the protein universe. PLoS Biol 7: e1000205.
- 9. Wooley JC, Godzik A, Friedberg I (2010) A primer on metagenomics. PLoS Comput Biol 6 (2) e1000667.
- 10. Gilbert JA, Dupont CL (2011) Microbial metagenomics: Beyond the genome. Ann Rev Mar Sci 3: 347–371.
- 11. Delmont TO, Malandain C, Prestat E, Larose C, Monier J-M, et al. (2011) Metagenomic mining for microbiologists. ISME J 5: 1837–1843.
- 12. Merico D, Gfeller D, Bader GD (2009) How to visually interpret biological data using networks. Nat Biotechnol 27: 921–924.
- 13. Breitling R, Vitkup D, Barrett MP (2008) New surveyor tools for charting microbial metabolic maps. Nat Rev Microbiol 6: 156–161.
- 14. Fuhrman J, Steele J (2008) Community structure of marine bacterioplankton: patterns, networks, and relationships to function. Aquat Microb Ecol 53: 69–81.
- 15. Fuhrman JA (2009) Microbial community structure and its functional implications. Nature 459: 193–199.
- 16. Rusch DB, Halpern AL, Sutton G, Heidelberg KB, Williamson S, et al. (2007) The Sorcerer II Global Ocean Sampling expedition: northwest Atlantic through eastern tropical Pacific. PLoS Biol 5: 398–431.
- 17. Punta M, Coggill PC, Eberhardt RY, Mistry J, Tate J, et al. (2012) The Pfam protein families database. Nucleic Acids Res 40: D290–301.
- 18. Fuchs G (2011) Alternative pathways of carbon dioxide fixation: insights into the early evolution of life? Ann Rev Microbiol 65: 631–658.
- 19. Martinez A, Tyson GW, Delong EF (2009) Widespread known and novel phosphonate utilization pathways in marine bacteria revealed by functional screening and metagenomic analyses. Environ Microbiol 12: 222–238.
- 20. Villarreal-Chiu JF, Quinn JP, McGrath JW (2012) The genes and enzymes of phosphonate metabolism by bacteria, and their distribution in the marine environment. Front Microbiol 3: 19.
- 21. Kujawinski EB (2011) The impact of microbial metabolism on marine dissolved organic matter. Ann Rev Mar Sci 3: 567–599.
- 22. Iyer LM, Abhiman S, Maxwell Burroughs A, Aravind L (2009) Amidoligases with ATP-grasp, glutamine synthetase-like and acetyltransferase-like domains: synthesis of novel metabolites and peptide modifications of proteins. Mol Biosyst 5: 1636–1660.
- 23. Sullivan MB, Lindell D, Lee JA, Thompson LR, Bielawski JP, et al. (2006) Prevalence and evolution of core photosystem II genes in marine cyanobacterial viruses and their hosts. PLoS Biol 4: e234.
- 24. K-ichi Miyazono, Shirokane M, Sawano Y, Tanokura M (2009) Crystal structure of the DUF54 family protein PH1010 from hyperthermophilic archaea Pyrococcus horikoshii OT3. Proteins 74: 256–260.
- 25. Wittkop T, Emig D, Truss A, Albrecht M, Böcker S, et al. (2011) Comprehensive cluster analysis with Transitivity Clustering. Nat Protoc 6: 285–295.
- 26. Polz MF, Hunt DE, Preheim SP, Weinreich DM (2006) Patterns and mechanisms of genetic and phenotypic differentiation in marine microbes. Philos Trans R Soc B 361: 2009–2021.
- 27. Lauro FM, McDougald D, Thomas T, Williams TJ, Egan S, et al. (2009) The genomic basis of trophic strategy in marine bacteria. Proc Natl Acad Sci USA 106: 15527–15533.
- 28. Seligmann H (2003) Cost-minimization of amino acid usage. J Mol Evol 56: 151–161.
- 29. Meador JA, Baldwin AJ, Catala P, Jeffrey WH, Joux F, et al. (2009) Sunlight-induced DNA damage in marine micro-organisms collected along a latitudinal gradient from 70°N to 68°S. Photochem Photobiol 85: 412–420.
- 30. Milligan JR, Aguilera JA, Ly A, Tran NQ, Hoang O, et al. (2003) Repair of oxidative DNA damage by amino acids. Nucleic Acids Res 31: 6258–6263.
- 31. Fernandes AP, Holmgren A (2004) Glutaredoxins: glutathione-dependent redox enzymes with functions far beyond a simple thioredoxin backup system. Antioxid Redox Signal 6: 63–74.
- 32. Rouhier N, Lemaire SD, Jacquot J-P (2008) The role of glutathione in photosynthetic organisms: emerging functions for glutaredoxins and glutathionylation. Annu Rev Plant Biol 59: 143–166.
- 33. Toulza E, Tagliabue A, Blain S, Piganeau G (2012) Analysis of the Global Ocean Sampling (GOS) project for trends in iron uptake by surface ocean microbes. PLoS ONE 7: e30931.
- 34. Lash LH, Jones P (1984) Renal glutathione transport: Characteristics of the sodium-dependent system in the basal-lateral membrane. J Biol Chem 259: 14508–14514.
- 35. Kannan R, Yi JR, Tang D, Li Y, Ziokovic BV, et al. (1996) Evidence for the existence of a sodium-dependent glutathione (GSH) transporter. J Biol Chem 271: 9754–9758.
- 36. DeLong EF, Béjà O (2010) The light-driven proton pump proteorhodopsin enhances bacterial survival during tough times. PLoS Biol 8: e1000359.
- 37. Slamovits CH, Okamoto N, Burri L, James ER, Keeling PJ (2011) A bacterial proteorhodopsin proton pump in marine eukaryotes. Nat Commun 2: 183.
- 38. Börnke F (2008) Protein Interaction Networks. In: Junker BH, Schreiber F, editors. Analysis of Biological Networks. Hoboken, NJ, USA: John Wiley & Sons, Inc. pp. 207–232.
- 39. Karsenti E, Acinas SG, Bork P, Bowler C, De Vargas C, et al. (2011) A holistic approach to marine eco-systems biology. PLoS Biol 9: e1001177.
- 40. Wu D, Hugenholtz P, Mavromatis K, Pukall R, Dalin E, et al. (2009) A phylogeny-driven genomic encyclopaedia of Bacteria and Archaea. Nature 462: 1056–1060.
- 41. Roberts RJ, Chang Y-C, Hu Z, Rachlin JN, Anton BP, et al. (2011) COMBREX: a project to accelerate the functional annotation of prokaryotic genomes. Nucleic Acids Res 39: D11–14.
- 42. Seshadri R, Kravitz SA, Smarr L, Gilna P, Frazier M (2007) CAMERA: a community resource for metagenomics. PLoS Biol 5: e75.
- 43. Markowitz VM, Chen I-M, Palaniappan K, Chu K, Szeto E, et al. (2010) The integrated microbial genomes system: an expanding comparative analysis resource. Nucleic Acids Res 38: D382–390.
- 44. Csardi G, Nepusz T (2006) The igraph software package for complex network research. InterJournal Complex Sy 1695.
- 45. Smoot ME, Ono K, Ruscheinski J, Wang P-L, Ideker T (2011) Cytoscape 2.8: new features for data integration and network visualization. Bioinformatics 27: 431–432.
- 46. Wittkop T, Emig D, Lange S, Rahmann S, Albrecht M, et al. (2010) Partitioning biological data with transitivity clustering. Nat Methods 7: 419–420.
- 47. Morris JH, Apeltsin L, Newman AM, Baumbach J, Wittkop T, et al. (2011) clusterMaker: a multi-algorithm clustering plugin for Cytoscape. BMC Bioinformatics 12: 436.