We searched the GOS data set for PolB-like sequences using the Pfam hidden Markov profile (PF00136). This resulted in a set of 1,947 sequences (from 23-562 amino acid residues). These sequences are referred to as 'PolB fragments' in this study. We next built a reference MSA of PolB homologs from known organisms (Additional data file 1). The reference MSA (Additional data file 2) corresponds to the polymerase domains of PolB homologs and contains 101 sequences, which were selected to achieve the widest possible taxonomic/paralog coverage (but with a non-exhaustive sampling for closely related species) for the analysis of the GOS metagenomic data. The reference MSA was used to generate a maximum likelihood tree (that is, the reference tree; Figure 1). Although the phylogenetic reconstruction did not provide statistical support for most of the basal branches, many peripheral groupings (supported by bootstrap values ≥ 70%) were coherent with the current taxonomy of viruses and cellular organisms. In this tree, we identified eight viral groups: poxviruses; chloroviruses; phaeoviruses; mimivirus and related algal viruses (Pyramimonas orientalis virus PoV01, Chrysochromulina ericina virus CeV01 and Phaeocystis pouchetii virus PpV01); iridoviruses grouped with ascoviruses; herpesviruses; baculoviruses; and one phage group. The PolB homologs from African swine fever virus (ASFV, Asfarviridae), Emiliania huxleyi virus 86 (EhV-86, Phycodnaviridae), Heterosigma akashiwo virus 1 (HaV, Phycodnaviridae) and the phage RM378 did not show well supported clustering with other PolB sequences. We also identified eleven groups in the reference tree for cellular PolB homologs: seven archaeal groups, one bacterial group and three eukaryotic groups (α, δ and ζ subtypes). Each of the GOS PolB fragments was then examined for its phylogenetic position using the reference MSA and the reference tree. To reduce the computation time and to streamline tprocess of summarizing results, we reduced the size of the reference MSA. Specifically, we selected 51 representatives from the 101 reference sequences and removed the remaining sequences. The reference tree was also reduced so that the resulting tree contains only the selected 51 representatives, while we conserved the original topology of the full reference tree shown in Figure 1. The reduced reference tree has 99 branches (including internal branches). A constraint on this topology defines 99 possible branching positions for each of the GOS PolB fragments. We aligned, one by one, each of the PolB fragments on the reduced reference MSA using the T-Coffee profile method. Based on the resulting profile MSA containing 52 sequences, the likelihoods for all 99 possible branching positions (thus 99 different topologies) were computed by ProtML 35. A statistical significance for the best tree among the 99 topologies was assessed by the RELL (resampling of estimated log likelihoods) bootstrap method 3637. We considered the branching position of a PolB fragment to be supported when the RELL bootstrap value for the best topology was ≥ 75%.

Our phylogenetic mapping method could assign the best branching position for 1,423 PolB fragments, of which 1,224 (86%) were mapped on viral branches. The best branching position was statistically supported by the RELL method for 869 PolB fragments, of which 811 (93%) were mapped on viral branches. Figure 2 and Additional data file 3 show the taxonomic distribution of the GOS PolB fragments. The largest fraction of the PolB fragments was mapped on the phage group. Of 866 cases of mapping within the phage group, 633 were supported. This appears consistent with the current estimate of the large number of phage-like particles and their genetic richness in marine environments 3. The second largest number of supported mappings was found to fall into large eukaryotic viruses commonly found in aquatic environments. Among them, the 'Mimiviridae group' (mimivirus, PoV01 and CeV01 17) represented the largest fraction, with 115 supported cases. The chlorovirus group gathered 51 supported cases of mapping. The iridovirus/ascovirus group and the branch leading to HaV showed five supported mappings each. In contrast, no PolB fragment was mapped for the groups for baculoviruses or herpesviruses commonly found in terrestrial animals. Interestingly, we found two PolB fragments mapped with good support on the ASFV branch (JCVI SCAF 1101668126451, JCVI SCAF 1101668152950). When these two PolB fragments were compared to the NCBI non-redundant amino acid sequence database (NRDB) using BLASTP, they were most similar to the ASFV PolB sequence. ASFV is pathogenic to domestic pigs and is currently the sole representative of the Asfarviridae family 38. Concerning cellular organisms, eukaryotic homologs gathered few mappings, as expected from the sample filtration threshold used in the GOS metagenomic study. Two archaeal groups - the group III containing crenarchaeotes (for example, Pyrobaculum aerophilum, Cenarchaeum symbiosum) and the group IV containing euryarchaeotes (for example, Thermoplasma acidophilum, an uncultured euryarchaeote Alv-FOS1) - had 23 and 17 supported cases of mapping, respectively. The bacterial group presented ten supported mappings.

We examined the phylogenetic mapping result and the sequence diversity of the PolB fragments classified in large eukaryotic virus groups (that is, NCLDVs). From those mapped on NCLDV branches, we selected long PolB fragments that generated a profile MSA showing at least 150 non-gapped sites. We computed a single alignment of these long PolB fragments together with the reference PolB sequences from large eukaryotic virus groups. A maximum likelihood tree (Figure 3) based on the alignment was perfectly consistent with our one-by-one mapping result (Figure 2) in terms of taxonomic assignment. The Mimiviridae group contained 16 PolB fragments showing substantial sequence variations. Twelve of them were significantly closer (bootstrap 100%) to CeV01 or PpV01 (both viruses of haptophytes) than to mimivirus or PoV01 (a green algal virus). Three of the rest were grouped with either mimivirus (bootstrap 89%) or PoV01 (bootstrap 96%). The last one (JCVI SCAF 1096627348452) was placed at the basal position of the Mimiviridae group. Although this basal positioning was not statistically supported, it was consistent with our one-by-one phylogenetic mapping result. The mimivirus PolB shared 47% identical amino acid residues with its closest homolog (JCVI SCAF 1101668170038). A large and diverse group containing 27 PolB fragments (bootstrap 92%) was also found beside the chlorella virus group (Paramecium bursaria chlorella viruses 1, K2 and NY2A). The DNA polymerase gene from the recently released Ostreococcus virus OtV5 genome (GenBank: EU304328) 39 was found grouped together with these PolB fragments. The grouping of a PolB fragment with ASFV PolB was also confirmed (bootstrap 100%).

We investigated the abundance of viral PolB genes relative to bacterial PolB genes in the GOS data set. Here, we used read coverage as a proxy to measure the abundance of the cognate DNA molecules in the samples. We computed the read coverage of each contig harboring a PolB fragment mapped on the reference tree with significant support, and then obtained the median of the read coverage values for each set of contigs mapped on the same branch (Additional data file 3). PolB sequences mapped on viral branches exhibited low median coverage values ranging from 1.31 for the ASFV branch to 2.00 for a phage branch. The median coverage value for the contigs mapped on the mimivirus branch (12 contigs) was 1.32. The viral contig with the largest read coverage (6.68) was the one mapped on the cyanophage P-SSM4 branch. In contrast, a higher median coverage value (8.40) was found for bacterial contigs mapped on the branch leading to Shewanella frigidimarina. One of the bacterial contigs exhibited a read coverage of 29.17. Viral branches were thus characterized by a large number of mapped contigs exhibiting a low coverage. This is consistent with numerous and very diverse viral populations 40. On the other hand, the bacterial branches exhibited a lower number of mapped contigs with a larger read coverage. This is consistent with numerous but less diverse populations of bacterial species, although our results concern only bacteria having PolB homologs.

GOS metadata provide physicochemical and biological parameters associated with each sampling site, such as water temperature, salinity, chlorophyll a concentration, and sample's water depth. These data offer additional dimensions to analyze the viral PolB fragments identified by our phylogenetic mapping. Here we compared the relative abundance of the predicted viral PolB fragments and the associated metadata across different GOS sampling sites (Figure 4a).

Predicted viral PolB fragments were detected in all of 44 GOS sampling sites (Figure 4b). The relative abundance of different virus groups showed substantial variation across these samples. This is consistent with the diverse ecosystems covered by the GOS expedition.

PolB fragments classified in the phage group were found in 42 (95%) of the 44 sample sites; the two samples without phage PolB fragments were GS08 (Newport Harbor, Richmond, USA) and GS32 (mangrove). In most samples (32 sites), putative phage PolBs exhibited a higher abundance relative to putative eukaryotic viral PolBs. On the other hand, the relative abundance of eukaryotic viral PolBs was higher than that of phage PolBs in 12 sampling sites. We found a significant positive correlation between the relative abundance of phage PolBs and water temperature (p = 0.001; Fischer's exact test with no correction for multiple testing): phage-type PolBs showed a higher relative abundance than eukaryotic viral PolBs in tropical waters (T ≥ 20°C), while a reversed tendency was observed in temperate water (T < 20°C). Interestingly, among eukaryotic viral PolBs, putative Mimiviridae PolBs showed the most widespread distribution, being detected in 38 (86%) of the total sites. One of these sampling sites (mangrove located on Isabella, Ecuador) exhibits only viral PolBs classified in the Mimiviridae group. This is the sole mangrove site of all the GOS sampling locations. Mimiviridae PolBs were also relatively abundant in two of the three samples from a hydrostation located in the Sargasso Sea. Three samples correspond to different size fractions: 3.0-20.0 μm for GS01a; 0.8-3.0 μm for GS01b; and 0.1-0.8 μm for GS01c. Putative Mimiviridae PolBs were identified in the GS01a and GS01c samples. The GS01a sample, which was targeted to small eukaryotes, might have contained host species infected by putative viruses of the Mimiviridae group. PolB fragments grouped with chloroviruses were also widely distributed. They were detected in 16 (36%) samples. The relative abundance of this putative eukaryotic virus group showed a significant positive correlation with chlorophyll a concentration, a measure of primary productivity in oceanic regions (p = 0.00002; Fisher's exact test with no correction for multiple testing).

The sample exhibiting the broadest taxonomic richness of viral PolBs was from Chesapeake Bay (GS12, MD, USA), which is an estuary. The GOS metagenomic sequences from this site exhibited PolB fragments classified in phages, chloroviruses, Asfarviridae and Mimiviridae. Notably, this site is a highly eutrophic estuary with an extremely high chlorophyll a concentration. PolBs classified in Asfarviridae were also detected in another estuary site (GS11, Delaware Bay, NY, USA), which is close to Chesapeake Bay.

Contigs harboring putative viral PolB homologs were relatively small, ranging from 0.4-12.5 kb (average 1,874 bp) for contigs mapped on eukaryotic viral branches and 0.5-8.8 kb (average 1,885 bp) for phages. To examine the presence of additional open reading frames (ORFs) in these contigs, these putative viral contigs were searched against NRDB using BLASTX. We detected several genes or gene fragments that are usually specific to viruses. For example, several contigs (for example, JCVI SCAF 1096626858151, JCVI SCAF 1096626920680) containing PolB fragments assigned to the chlorovirus group also harbor an ORF most similar to the OtV5 putative major capsid gene. Several putative phage-type contigs (for example, JCVI SCAF 1096628232224, JCVI SCAF 1096626847406) mapped on the cyanophage P-SSM4 branch exhibited ORFs similar to regA (translation repressor of early genes) or uvsX (recA-like recombination and DNA repair protein genes). The presence of such 'virus-specific' genes next to the 'virus-like' PolB homologs corroborates the validity of our phylogenetic mapping approach.

During this search, we found an ORF similar to RimK, a protein involved in post-translational modification of the ribosomal protein S6, in a contig (JCVI SCAF 1096626956347) having a PolB fragment mapped on the cyanophage P-SSM4 branch. In this contig, the rimK homolog was flanked by a phage-specific regA homolog (Figure 5). rimK homologs are found in bacteria, archaea and eukaryotes 41. To our knowledge, no rimK homolog has been found in a viral genome. Using this putative viral RimK homolog as a query of TBLASTN, we screened the entire GOS data set. We identified more than a hundred contigs harboring RimK homologs with higher similarities (BLAST score from 137 up to 732; E-value < 10^-30) than those exhibited by cellular homologs (BLAST score < 132; E-value > 10^-29) in NRDB. The sequences of those putative phage RimK homologs were readily aligned with Escherichia coli RimK along its entire length (not shown), and showed amino acid residues highly conserved in the ATP-graps domain of bacterial RimK 41. Several GOS RimK sequences showed an additional domain of unknown function (DUF785, PF05618, E-value < 0.001) at the carboxy-terminal side of the ATP-graps domain. A DUF785 domain is present also in RimK of some bacteria (at the amino-terminal side of the ATP-graps domain) such as Synechococcus sp. (Q7U6F4) and euryarchaeotes (at the carboxy-terminal side of the ATP-graps domain) such as Halobacteria (for example, Q5V351). Furthermore, many of the GOS contigs encoding RimK homologs exhibited additional ORFs usually specific to phages such as T4-like clamp loader subunit genes, contractile tail sheath protein genes or T4-like DNA packaging large subunit terminase genes (Figure 5). Our phylogenetic analysis indicates that those RimK homologs are closely related to each other and distantly related to bacterial RimK (Figure 6). These results suggest the existence of phages carrying rimK homologs in marine environments.

We retrieved the combined assemblies of the GOS metagenomic data through the CAMERA website 60. The data set was composed of 3,081,849 scaffolds. We extracted all the stop-to-stop ORFs (≥ 60 amino acid residues) from the assembled sequences using EMBOSS/GETORF 61. We obtained a set of 21,406,171 ORFs. Those ORFs were translated into corresponding amino acid sequences. To identify PolB-like fragments in this set, we used the Pfam profile (PF00136, both long and fragment search versions: 'ls' and 'fs') 62 and the HMMER software as a search engine 63 using an E-value threshold of 0.001. We then removed redundancy (due to the double use of 'ls' and 'fs' versions of the Pfam profile) and false positive detections (having the best hit against non-PolB sequences in the NRDB) by BLASTP 64 using an E-value threshold of 10^-5). We extracted only the parts of metagenomic amino acid sequences that were aligned on the Pfam profile representing the polymerase domains of PolB. Thus, additional domains (such as endonuclease domains) were not included in our PolB sequence set. No contig was found to contain more than one PolB homolog. As a result of these processes, we obtained 1,947 distinct PolB-like sequences (from 23-562 amino acid residues); these sequences are referred to as PolB fragments in this study. We parsed the GOS PolB fragments to find intein insertions by the TIGRFAM profiles TIGR01445 (intein amino terminus) and TIGR01443 (intein carboxyl terminus) 65, but none of these fragments had a detectable intein domain. In this study, we did not include the protein priming subfamily of the B family DNA polymerase 28, which is represented by the Pfam profile PF03175. The members of this subfamily are found in eukaryotic linear plasmids of mitochondrion, phages and adenoviruses.

We retrieved PolB homologs from the NRDB, RefSeq 66 and KEGG 67 databases using BLAST using multiple query sequences (E-value < 10^-5) and the PolB Pfam profile (E-value < 0.001). We removed species redundancy using BLASTCLUST 64 while keeping the widest possible taxonomic/paralog coverage (but with a non-exhaustive sampling for closely related species). This resulted in a set of 120 PolB homologs (Additional data file 1). We removed intein sequences in the PolBs of mimivirus 68, HaV 69 and CeV01 (GenBank: ABU23716).

We next constructed an alignment of PolB homologs from known organisms (that is, the reference MSA) and generated a phylogenetic tree of PolB homologs (that is, the reference tree). There is a tradeoff between the number of distant homologs included in the reference MSA (contributing to a wider taxonomic/paralog coverage) and the quality of the resulting MSA and tree (contributing to a reliable classification of metagenomic sequences). Among the 120 PolB homologs, we identified 19 highly divergent sequences that decrease the quality of the resulting PolB alignment and tree but that show no close homologs in the GOS PolB fragments. This process was performed through multiple trials of building alignments by T-Coffee 70 and phylogenetic trees by PhyML for the PolB homologs. These 19 sequences correspond to six groups of PolB homologs: eukaryotic DNA polymerase ε, a Trichomonas vaginalis DNA polymerase α-like paralog, PolBs of unclassified herpesviruses (Ostreid, Ictalurid and Ranid herpesviruses), Heliothis zea virus, a nimavirus (shrimp white spot syndrome virus), and PolBs of a group of bacteria related to Prosthecochloris vibrioformis and Chlorobium tepidum. There was no PolB-like fragment in the GOS data exhibiting a best BLAST hit against these groups of PolB homologs. Therefore, the removal of the six groups of PolB homologs from our reference data set does not affect the interpretation of the results described in this manuscript. After discarding these 19 sequences, the final PolB set was composed of 101 sequences. We aligned the 101 PolB sequences using M-Coffee accessible from a public server 71 with the use of default options. M-Coffee is a meta-method for assembling multiple sequence alignments 72. We extracted only the polymerase domain sequences from the alignment (that is, the reference MSA; Additional data file 2). The reference alignment showed four conserved regions (numbered from I to IV) previously described as the signatures of the PolB polymerase domains 33. We next built a maximum likelihood tree based on the reference MSA (that is, the reference tree) using PhyML after removing gap-containing sites 73 with JTT substitution model and a gamma low (four rate categories). Bootstrap values were obtained after 100 bootstrap replicates. We used the phylogeny.fr platform 74 to generated scalable vector graphics from newick formatted trees.

Each of the metagenomic PolB fragments was taxonomically assigned by aligning it against the reference MSA and by examining its phylogenetic position in the reference tree. In order to reduce the computation time and to avoid unnecessary complications in summarizing results within too dense a tree, we reduced the size of the reference MSA and the reference tree. Specifically, we selected 51 PolBs from the 101 PolBs contained in the initial set. We kept the selected 51 PolBs in the reduced set, and deleted the remaining PolBs. The selection of the 51 representatives was carried out in the following way. First, we selected all the PolBs (that is, ASFV, EhV86, HaV, Phage RM378) that were not grouped with other PolBs with a statistical support (≥ 70% bootstrap value) in the initial reference tree (Figure 1). Second, we selected two or three representatives from each of the statistically supported monophyletic groups (≥ 70% bootstrap value). The choice of representatives from a monophyletic group was arbitrary. We simply selected two or three relatively distant sequences from the members of the monophyletic group. To obtain a reduced reference MSA composed of the selected 51 sequences, we extracted a part (that is, lines) of the initial reference MSA (containing gaps). The initial reference tree (composed of 199 branches including internal ones) was also reduced by pruning branches leading to the non-selected leaves using BAOBAB 75.

The reduced reference tree has 99 branches (including internal branches); the constraint on the topology of the reduced reference tree thus defined 99 possible branching positions for each PolB-like fragment extracted from the metagenomic data set. The reduced reference MSA and the reduced reference tree are the basis for our phylogenetic mapping in this study. Each of the PolB fragments from the GOS data set was aligned on the reduced reference MSA (containing gaps) using T-Coffee 70 with a profile alignment option. For the T-Coffee profile alignment, we used the option '-profile comparison = full10'. If a GOS PolB fragment generates an alignment with less than 50 sites after removing gap-containing sites, we discarded the GOS PolB fragment from our analysis. Based on the resulting alignment (51 reference sequences and one GOS PolB fragment), the likelihoods of all 99 possible branching positions (thus 99 different topologies) for the PolB fragment were computed by ProtML 35. A statistical significance for the best tree among the 99 topologies was assessed by the RELL method 3637. We considered the branching position of a PolB fragment to be supported when the RELL bootstrap value for the best topology was ≥ 75%.

Read coverage for a contig was defined by dividing the cumulated size of reads contributing to the contig by the size of the contig.

For the analysis of the relative abundance of PolB sequences, we used the same approach used by Williamson et al. 14. Briefly, we first estimated the average number of reads overlapping with a part of a contig where a PolB domain was encoded, by taking into account the length of the PolB domain (as defined by the Pfam hit) and the length of the contig. The abundance of the PolB-sequences for each viral group for a given sample site was then quantified by the total number of reads associated with the relevant set of PolB-sequences (that is, the sum of the estimated read numbers). For a given site, the viral PolB proportion was computed by dividing the total number of viral PolB reads (for all viral groups) by the total number of reads obtained from the site.