Marine DNA viral macro- and micro-diversity from pole to pole

Summary:

Microbes drive most ecosystems and are modulated by viruses that impact their lifespan, gene flow and metabolic outputs. However, ecosystem-level impacts of viral community diversity remains difficult to assess due to classification issues and few reference genomes. Here we establish a ~12-fold expanded global ocean DNA virome dataset of 195,728 viral populations, now including the Arctic Ocean, and validate that these populations form discrete genotypic clusters. Meta-community analyses revealed five ecological zones throughout the global ocean, including two distinct Arctic regions. Across the zones, local and global patterns and drivers in viral community diversity were established for both macrodiversity (inter-population diversity) and microdiversity (intra-population genetic variation). These patterns sometimes, but not always, paralleled those from macro-organisms and revealed temperate and tropical surface waters and the Arctic as biodiversity hotspots and mechanistic hypotheses to explain them. Such further understanding of ocean viruses is critical for broader inclusion in ecosystem models.

ETOC summary

A global survey of ocean virus genomes vastly expands our understanding of this understudied community and reveals the Arctic as un expected hotspot for viral biodiversity.

Graphical Abstract

Introduction:

Biodiversity is essential for maintaining ecosystem functions and services (reviewed by Tilman et al., 2014). In the oceans, the vast majority of biodiversity is contained within the microbial fraction containing prokaryotes and eukaryotic microbes, which represents ~60% of its biomass (Bar-On et al., 2018). Meta-analyses looking at changes in marine biodiversity show that biodiversity loss increasingly impairs the ocean’s capacity to produce food, maintain water quality, and recover from perturbations (Worm et al., 2006). To date, marine conservation efforts have focused on specific organismal communities, such as fisheries or coral reefs, rather than conserving whole ecosystem biodiversity. However, emerging studies across diverse environments show that the stability and diversity of higher trophic level organisms rely upon diversity throughout the food web (e.g. Soliveres et al., 2016). Despite being the foundation of the food web, most marine microbial biodiversity numbers are based on a few well-studied locations (e.g., Hawaii Ocean Time Series, Bermuda Atlantic Time Series, and San Pedro Ocean Time Series). For ocean microbes and their viruses, global surveys that parallel century-old global terrestrial and decades-old marine macro-organismal global biodiversity surveys (Reiners et al., 2017) are only now emerging (e.g. de Vargas et al., 2015; Sunagawa et al., 2015; Brum et al., 2015; Roux et al., 2016; Ser-Giacomi et al., 2018; Table S1). Key to assessing biodiversity changes across marine ecosystems is improving our understanding of current microbial biodiversity levels, distribution patterns, and their ecological drivers.

Despite their tiny size, viruses play a large role in marine ecosystems and food webs. For example, mortality due to viruses is credited with lysing approximately 20–40% of bacteria per day and releasing carbon and other nutrients that impact the food web (reviewed by Suttle, 2007). Beyond mortality, viruses can alter evolutionary trajectories of microbial communities by transferring ~10²⁹ genes per day globally (Paul, 1999) and biogeochemical cycling by metabolically reprogramming host photosynthesis, as well as central carbon metabolism and nitrogen and sulfur cycling (reviewed in Hurwitz & U’Ren, 2016). Finally, as the oceans are estimated to capture half of human-caused carbon emissions (Le Quéré et al., 2018), it is notable that genes-to-ecosystems modeling has placed viruses as central players of the ocean ‘biological pump’ (Guidi et al., 2016). Many of these discoveries are very recent as ocean viral genome sequence space is just now being explored at the level of viral macrodiversity, i.e., inter-population diversity, throughout the global oceans -- at least for the most abundant double-stranded DNA viruses sampled (Table S2).

In spite of this progress in studying marine viral macrodiversity, virtually nothing is known about microdiversity, i.e., intra-population genetic variation. This is due to the controversy surrounding the existence of viral species (Gregory et al., 2016; Bobay et al., 2018). In eukaryotic organisms, where species boundaries are more widely accepted, such microdiversity has been studied and is thought to drive adaptation and speciation to promote and maintain stability in ecosystems (Hughes et al., 2008; Larkin & Martiny, 2017). This is likely also true in viruses since even a few mutations can alter host interactions and ecological and evolutionary dynamics for the genotype (e.g. Marston et al., 2012; Petrie et al., 2018). In nature, viral microdiversity measurements have been limited to marker genes (e.g. genes encoding major capsid proteins), which capture neither community-wide variability (Sullivan, 2015) nor genome-wide evidence of selection (e.g. Achtman & Wagner, 2008). Recently, deeper metagenomic sequencing and population genetic theory-grounded species delimitations (Shapiro et al., 2012; Cadillo-Quiroz et al., 2012) have begun to reveal such microdiversity in microbes, and this has elucidated unknown features of speciation, adaptation, pathogenicity and transmission (e.g. Snitkin et al., 2011; Schloissnig et al., 2013; Rosen et al., 2015; Lee et al., 2017; Smillie et al., 2018). Although parallel species delimitations are now available for viruses (Gregory et al., 2016; Bobay et al., 2018), no datasets are yet available to explore genome-wide microdiversity in viruses, particularly at the global scale.

Here we leverage the Tara Oceans global oceanographic research expedition sampling to establish a deeply-sequenced, global-scale ocean virome dataset and use it to assess the validity of the current viral population definition and to establish and explore baseline macro- and micro-diversity patterns with their associated drivers across local to global scales. These data have been collected and analyzed in the context of the larger Tara Oceans Consortium systematically- sampled, global-scale, viruses-to-fish-larvae datasets (de Vargas et al., 2015; Sunagawa et al., 2015; Brum et al., 2015; Lima-Mendez et al., 2015; Pesant et al., 2015; Roux et al., 2016), and help establish foundational ecological hypotheses for the field and a roadmap for the broader life sciences community to better study viruses in complex communities.

Results & Discussion:

The dataset.

The Global Ocean Viromes 2.0 (GOV 2.0) dataset is derived from 3.95 Tb of sequencing across 145 samples distributed throughout the world’s oceans (Fig. 1A and Table S3; see Methods). These data build on the prior GOV dataset (Roux et al., 2016) by increased sequencing for mesopelagic samples (defined in our dataset as waters between 150m to 1,000m) and upgrading assemblies, both of which drastically improved sampling of the ocean viruses in these samples (results below). Additionally, we added 41 new samples derived from the Tara Oceans Polar Circle (TOPC) expedition, which traveled 25,000 km around the Arctic Ocean in 2013. These 41 Arctic Ocean viromes were generated to represent the most significantly climate-impacted region of the ocean, and an extreme environment. No such metagenome-based viral data exist for the Arctic region (Deming & Collins, 2017), and more generally, for many planktonic organisms, systematic sampling is uneven throughout the Arctic Ocean (CAFF State of the Arctic Marine Biodiversity Report) due to geopolitical and physical challenges of sampling these regions.

Fig. 1. The Global Ocean Viromes 2.0.

(A) Arctic projection of the global ocean highlighting the new sampling stations of viromes in the GOV 2.0 dataset. Datasets from non-arctic samples were previously published in (Brum et al., 2015; Roux et al., 2016). (B) Histograms of the average assembled contig lengths for viral populations >10 kb shared between GOV and GOV 2.0. B-inset. More than 92% of the unbinned GOV viral populations were reassembled and identified in GOV 2.0 >10 kb populations. (C) Pie charts showing how many of the 488,130 total viral populations comprising GOV 2.0 can be annotated and, of those, their viral family level taxonomy. (D) Barplot showing the host affiliations for each viral population at the domain level.

The first step to studying viral biodiversity from the assembled GOV 2.0 dataset (see Methods and Fig. S1A) was to identify contigs that likely derive from viruses using tools that collectively utilize homology to viral reference databases, probabilistic models on viral genomic features, and viral k-mer signatures (see Methods). These putative viral contigs were then assigned to ‘populations’, which are currently defined as viral contigs ≥10 kb where ≥70% of the shared genes have ≥95% average nucleotide identity (ANI) across its members (Brum et al., 2015; Roux et al., 2016; Roux et al., 2018; population definition also discussed below). This process identified 195,728 viral populations in the GOV 2.0 dataset, which is a ~12-fold increase over the 15,280 identified in the original GOV dataset and assemblies (Roux et al., 2016) and augments prior marine viromic work (Tables S2). Of these original GOV viral populations, 12,708 were represented by single contigs and, of these, most (92%) were recovered in GOV 2.0 (Fig. 1B-inset), with average lengths increased 2.4-fold from 18 kbp to 44 kbp (Fig. 1B). Outside these GOV-known and now improved viral populations, an additional 180,448 new GOV 2.0 viral populations were identified -- derived mostly (58%) from improved assemblies and deeper sequencing of the original GOV samples, and the rest (42%) from the 41 new Arctic Ocean viromes. Finally, new methods to identify shorter viral contigs (see Methods) were applied and these identified another 292,402 contigs as viral (5–10 kb length and/or circular), which, when added to the earlier data and clustered at ≥95% ANI, resulted in a total of 488,130 viral populations (N50= 15,395; L50=105,286; mean read depth per population = 17x). Ninety percent of the populations could not be taxonomically classified to a known viral family, but the 10% that could were predominantly dsDNA viral families and bacteriophages (Fig. 1C, D).

Although the focus of this study is DNA viruses, a remarkable diversity of RNA viruses has been described in nature, though largely outside of marine systems. For example, transcriptome sequencing from plants (Roossinck et al., 2010), arthropods (Shi et al., 2016), and birds and bats (reviewed in Greninger, 2018) have shown a genomic and phylogenetic diversity of RNA viruses far beyond those in culture (Shi et al., 2018). In the oceans, however, RNA viral diversity and abundance remains largely unknown. The few estimates of marine RNA virus abundance are based on the relative quantification of RNA and DNA from purified viral particles and genome size extrapolations and suggest that up to half of the viral particles in seawater are RNA viruses (Steward et al., 2013, Miranda et al., 2016). Direct RNA virus counts are not yet available for any environment due to the lack of RNA-specific stains. To date, our understanding of marine RNA viral diversity is based on single-gene surveys that target subgroups of viruses (reviewed in Culley, 2018) and a few viromes generated from extracellular viral particles (Culley and Steward, 2007; Culley et al., 2006; Miranda et al., 2016; Steward et al., 2013; Urayama et al., 2018, Zeigler-Allen et al., 2017) or from RNA viral sequences identified in metatranscriptomes (Carradec et al., 2018; Moniruzzaman et al., 2017; Urayama et al., 2018; Zeigler-Allen et al., 2017). Together, these studies suggest that the marine RNA virosphere is composed of a large diversity of positive-polarity ssRNA and dsRNA viruses diverge from established taxa, with an apparent predominance of viruses that infect eukaryotes (Culley, 2018). Due to current methodological limitations, comprehensive, systematic assessments of marine RNA viral diversity on the global scale are not yet available, and are excluded from our analysis.

Validating viral ‘population’ boundaries.

Defining species is controversial for eukaryotes and prokaryotes (Kunz, 2013; Cohan, 2002; Fraser et al., 2009) and even more so for viruses (Bobay et al., 2018), largely because of the paradigm of rampant mosaicism stemming from rapidly evolving ssDNA and RNA viruses, whose evolutionary rates are much higher than dsDNA viruses [reviewed by (Duffy et al., 2008)]. The biological species concept, often referred to as the gold standard for defining species, defines species as interbreeding individuals that remain reproductively isolated from other such groups. To adapt this to prokaryotes and viruses, studies have explored patterns of gene flow to determine whether they might maintain discrete lineages as reproductive isolation does in eukaryotes. Indeed, gene flow and selection define clear boundaries between groups of bacteria, archaea and viruses, though the required scale of data are only available for cyanophages and mycophages among viruses (Shapiro et al., 2012; Cadillo-Quiroz et al., 2012; Gregory et al., 2016; Bobay et al., 2018).

Because measuring gene flow requires extensive datasets not yet available for many groups, the term ‘species’ is rarely used for prokaryotes or viruses, and instead discrete lineages are described as ‘populations’. Separate from these population genetic theory grounded observations, evidence of discrete lineages, or sequence-discrete populations, is to use metagenomic read-mapping to evaluate naturally occurring sequence variation across organisms. Sequence-discrete populations have now been observed for prokaryotes (Konstantinidis & Tiedje, 2005) and more recently for some dsDNA viruses (viral-tagged metagenomes and 142 isolate genomes for marine cyanophages; Deng et al., 2014, Gregory et al., 2016; Table S4). Buoyed by this and signatures of at least some dsDNA viruses obeying the biological species concept (Bobay et al., 2018), viral ecologists have established the definition of viral populations described above (Brum et al., 2015; Roux et al., 2016; Roux et al., 2018). Notably, however, only deeply sequenced groups, cyano- and myco-phages, have been evaluated to date (Gregory et al., 2016; Bobay et al., 2018), and an emergent hypothesis suggests that phages evolve with different modes and tempos driven by differing temperate or obligately lytic lifestyles (Mavrich & Hatfull, 2017). Thus, there is a need to evaluate how generalizable this empirically-derived ≥95% ANI cut-off viral population definition is in nature.

To test this, we permissively mapped metagenomic reads against our 488,130 GOV 2.0 viral populations by allowing ‘local’ matching as low as 18% nucleotide identity, and statistically identifying ‘breaks’ in the resulting read frequency histograms (see Methods). This revealed that, on average, the break occurred such that reads <92% nucleotide identity failed to map (Fig. 2C; full results Table S5), which resulted in a genome-wide signature of ≥95% ANI for nearly all (99.9% or 487,875) of the GOV 2.0 viral populations, including the smaller <10 kb viral populations (Fig. 2D). This implies that the observed viral populations in the dataset are predominantly and detectably sequence-discrete. This result is consistent with data from viral-tagged metagenomes (Deng et al., 2014) and gene-sharing networks of prokaryotic virus genomes (Iranzo et al., 2016, Bolduc et al., 2017), which also showed that sampled viral genome sequence space is clustered at each ‘species’ and ‘genus’ levels, respectively. Thus, while ssDNA and RNA viruses have variable and elevated genome evolutionary rates that can erode species boundaries [reviewed by (Duffy et al., 2008)], it appears that virtually all metagenome-assembled dsDNA viral populations form discrete genotypic clusters and can be appropriately delineated via a ≥95% genome-wide ANI cut-off.

Fig. 2. GOV 2.0 viral population have discrete population boundaries.

(A) Barplots showing the read mapping results for the most abundant viral population >10kb in length for each of the top four viral families. Despite differences in read boundaries across the representative viral populations, there is no difference in the average read boundaries across the different viral families. (B) Histogram showing the read distribution frequency break (i.e. read boundary) between spuriously mapped reads and legitimate reads mapping to the genome. (C) Histograms showing the average percent identity of reads mapped to each genome after removing spuriously mapped reads.

Meta-community analysis reveals 5 ecological zones.

Having organized this global sequence space into discrete and biologically meaningful populations, we next sought to use metagenome-derived abundance estimates to establish patterns and drivers of viral population diversity across the global ocean across multiple levels of ecological organization (Fig. 3). This revealed that the 145 GOV 2.0 viral communities robustly assorted into just five meta-communities, denoted ecological zones, whether assessed using Bray-Curtis dissimilarity distances in principal coordinate analysis (Fig. 4A), non-metric multidimensional scaling (Fig. S2A), or hierarchical clustering (Fig. S2B) and after accounting for variable sample sizes (see Methods). We designated these 5 emergent ecological zones as the Arctic (ARC), Antarctic (ANT), bathypelagic (BATHY), temperate and tropical epipelagic (TT-EPI) and mesopelagic (TT-MES), and used these for further study. Depth ranges overlapped with those previously defined (Reygondeau et al., 2018), with epipelagic, mesopelagic, and bathypelagic being waters of depths 0 to 150 meters, 150 to 1,000 meters, and deeper than 2,000 meters, respectively.

Fig. 3. Ecological levels of organization.

Schematic showing the different ecological levels of organization studied in this paper.

Fig. 4. Viral communities partition into five ecological zones with different macro- and micro- diversity levels.

(A) Principal coordinate analysis (PCoA) of a Bray-Curtis dissimilarity matrix calculated from GOV 2.0. Analyses show that viromes significantly (Permanova p = 0.001) structure into five distinct global ecological zones: ARC, ANT, BATHY, TT-EPI, and TT-MES zones. Ellipses in the PCoA plot are drawn around the centroids of each group at 95% (inner) and 97.5% (outer) confidence intervals. Four outlier viromes that did not cluster with their ecological zones were removed (Fig. S3A) and all the sequencing reads were used (see Fig. S3B and Methods). (B – right) Scatterplots showing correlations between macro- (Shannon’s H’) and micro- (average π for viral populations with ≥ 10x median read depth coverage; see Methods) diversity values for each sample across GOV 2.0. The larger circles represent the average per zone. (B – left) Boxplots showing median and quartiles of average microdiversity per ecological zone. (B – bottom) Boxplots showing median and quartiles of macrodiversity for each ecological zone. Zonal samples were randomly downsampled to n = 5 to account for zone sampling difference. All pairwise comparisons shown were statistically significant (p<0.01) using two-tailed Mann-Whitney U-tests. (C) Positive (blue) and negative (red) Pearson’s correlation results comparing macro- (upper) and micro- (lower) diversity with different biogeographical and biogeochemical parameters at the global scale (see Fig. S3E, Table S3 for all abbreviations, and Methods). The significance of the correlations is indicated by the size of the black circles on top of the bars, and the variables on the x-axis are ordered from the strongest to the weakest correlation with macrodiversity (except for the top four variables correlating with microdiversity for readability).

Comparison of our virome-inferred ecological zones to those inferred for the oceans in other ways was telling. Our zones differed from traditional oceanographic biogeographical biomes (e.g. Longhurst), where four biomes and ~50 provinces have been designated across surface ocean waters based on annual cycles of nutrient chlorophyll a (Longhurst et al., 1995, Longhurst, 2007), and from mesopelagic ecoregions and biogeochemical provinces based on biogeography and environmental climatology, respectively (Sutton et al., 2017; Reygondeau et al., 2018). However, they were similar to those observed for marine bacterial communities, which clustered by mid-latitude surface, high-latitude, and deep waters (Ghiglione et al., 2012). This implies that the physicochemical structuring of marine microbial communities is likely the most important factor in structuring marine viral communities, perhaps reflecting a relative stability in host range of viruses in the oceans (de Jonge et al., 2018). To evaluate this physicochemical structuring, we examined the universal predictors and drivers of viral ecological zones, across one (Fig. 5A) and multiple ordination dimensions (Fig. 5B; see Methods). This suggested that temperature was the major driver structuring these ecological zones, as previously shown from global microbial surveys (Sunagawa et al., 2015) and our own smaller ocean virome surveys, where we posited previously that temperature likely directly impacts microbial community structure, and indirectly viral community structure (Brum et al., 2015). Moreover, temperature has been shown to play an important role in virus-host interactions, especially in the Arctic (Maat et al., 2017).

Fig. 5. Ecological drivers of global viral macrodiversity.

(A) Regression analysis between the first coordinate of a PCoA (Fig. 4A) and temperature showed that samples were separated by their local temperatures with an r² of 0.82. (B) Potential ecological drivers & predictors of beta-diversity across GOV 2.0 for the first two dimensions (Goodness of fit r² using a generalized additive model) and across all dimensions (Mantel test based on Spearman’s correlation). Temperature was uniformly reported as the best predictor of viral beta-diversity globally. (C) Regression analysis between viral macrodiversity at the deep chlorophyll maximum (DCM) layer and areal chlorophyll a concentration (after cube transformation) showed that the negative correlation between viral macrodiversity and nutrients (Fig. 4C) is mediated (at least partially) by primary productivity. The Shannon’s H outlier 32_DCM (Fig. S3) and a chlorophyll a concentration outlier (173_DCM; Fig. 5D) have been excluded from the regression analysis. (D) Boxplot analysis of areal chlorophyll a concentrations showing a single outlier concentration that fell above the fourth quantile of the data points (function geom_boxplot of ggplot).

To look for specific viral adaptations in each ecological zone, we identified genes under positive selection by evaluating the ratio of non-synonymous to synonymous mutations observed in gene sequences using the pN/pS equation (Schloissnig et al., 2013). Of 1,139,501 genes tested from populations with enough coverage (≥10x mean read depth; mean number of populations assessed per sample: 14,852 viral populations), 124,882 genes were identified as being under positive selection in at least one sample. Most (82%) of the positively selected genes were functionally unannotatable, with the remaining 18% annotatable as predominantly genes related to structure or DNA metabolism (Table S6). In model systems, such genes are often under strong selective pressures during adaptations to new hosts (Marston et al., 2012; Jian et al., 2012; Enav et al., 2018). Thus, we hypothesize that host availability in each ecological zone is a strong selective pressure on our marine viral populations. Given the lack of functional annotations for most of the genes, we clustered all translated GOV 2.0 viral genes into protein clusters (PCs) based on sequence homology (sensu Holm & Sander, 1998) to identify positively selected zone-specific PCs. This resulted in 823,193 PCs, of which ~10% (79,588 PCs) appeared under positive selection, with a subset of these specific to a single zone (ARC = 80%; ANT = 33%; BATHY = 37%; TT-EPI = 75%; TT-MES = 69% of positively selected PCs per zone; see Table S6). These findings of many zone-specific positively-selected PCs is indicative of niche-differentiation. However, functional stories from these data are challenging as 85% of these zone-specific PCs were of unknown function, with the remaining mostly being the structural and DNA metabolism genes described above. This suggests that we have a lot to learn about the function of genes that most likely drive niche-differentiation across the ecological zones.

Viral macro- and micro- diversity, and potential drivers, within and between ecological zones.

To explore diversity patterns across ecological zones, we calculated per sample diversity using Shannon’s H’ for macrodiversity and a newly established method for community-wide microdiversity. This new method for community-wide microdiversity is limited in that it can only assess well-sampled, abundant populations because it estimates the average nucleotide diversity (or π) from the mean of π from 100 randomly subsampled well-sequenced populations sampled 1,000 times (see Methods). These zone-normalized (see Methods) comparisons revealed that macrodiversity was highest in TT-EPI (p < 0.05), closely followed by the ARC, and lowest in TT-MES and ANT (Fig 4B –bottom), whereas microdiversity was highest in TTMES (p < 0.05) and lowest in ARC (Fig. 4B –left). At the zonal level, a negative trend between macro- and micro- diversity emerges (Fig. 4B-right), although we note that the small number of zonal points limits our statistical inferences, even in this global dataset.

Recent work suggests that higher micro-diversity can impede the maintenance of macro-diversity by promoting competitive exclusion (Hart et al., 2016). Thus we posit that, if the zonal level negative macro/micro diversity trends are real, this may result from increased intrapopulation niche variation that reduces interpopulation niche variation resulting in competitive exclusion by the superior competitors, which may occur slowly and may be why it only appears at this regional scale (Fig. S4). Because estimates of microdiversity in our dataset and even currently available single virus genomics approaches (Martínez-Hernández et al., 2017) remain limited to only the most abundant populations, testing such a hypothesis awaits critically-needed advances and scalability in single-virus genomics technologies.

At the per-sample level, however, macro- and micro- diversity were not correlated, even within each zone (Fig. 4B – right). Although these are the first data available for viruses, for larger organisms, macro- and micro-diversity are often correlated across habitats sharing similar species pools, presumably due to habitat characteristics altering immigration, drift, and selection (Vellend & Gerber, 2005). These ecological correlations are generally positive and significantly stronger in discrete habitats (e.g. islands) in contrast to more connected communities like the ocean [reviewed in (Vellend et al., 2014)]. Thus we posit that the lack of correlation between marine viral macro- and micro- diversity at this per-sample level is driven by differences in local drivers (Fig. 4C). Consistent with this, local potential drivers differed as nutrients strongly (and negatively) correlated with viral macrodiversity, whereas photosynthetically active radiation (PAR; an indicator of productivity) best (and positively) correlated with viral microdiversity in the epipelagic waters (Fig. 4C).

Mechanistically, these results suggest several possible hypotheses. We interpret that, at the viral macrodiversity level, decreased host diversity in algal blooms, which themselves rely on nutrient pulses (Farooq & Malfatti, 2007), could skew viral rank abundance curves towards dominance by increasing abundance of bloom-associated viral populations. Even though algal blooms were not targeted in the Tara Oceans expedition, we did find that viral macrodiversity negatively correlated with chlorophyll a (Fig. 5C), and particulate inorganic carbon concentration (PIC; Fig. 4C), which is commonly used as a proxy for coccolithophore abundance (Groom & Holligan, 1987). Additionally, viral macrodiversity negatively correlated with the relative abundance of coccolithophores based on the V9 region of the 18S rRNA genes in the sequencing reads (Fig. 4C). For viral microdiversity in epipelagic waters, we interpret that PAR is potentially the main driver (Fig. 4C). PAR is known to impact host diversity, particularly in nutrient-poor surface waters, by inhibiting photoautotrophs through overwhelming their photosystems with too many electrons that can back up and even damage the photosystems (Feng et al., 2015). Further PAR can inhibit the growth of the dominant heterotroph, SAR11 (Ruiz-González et al., 2013), and can stimulate other key microbes such as Roseobacter, Gammaproteobacteria and NOR5 (Ruiz-González et al., 2013). We hypothesize that the shorter-term impacts of high PAR in the surface waters on host communities may create new niches for viruses, whereby microdiversity increases to enable differentiation of existing viral populations. As above, advances in single-virus genomics would be invaluable for testing this hypothesis.

Viral macro- and micro- diversity, and potential drivers, against classical ecological gradients.

Ecologists have long explored the relationship between diversity and geographic range, which in eukaryotes and bacteria are highly (and positively) correlated and thought to be due to the accumulation of niche-specific selective mutations across populations with large heterogeneous geographic ranges (i.e. the niche variation hypothesis; Van Valen, 1965, Hedrick, 2006, Rosen et al., 2015). No parallel studies have looked at viruses. To explore this for viruses, we determined the geographic range of viral populations based on their distribution within and between ecological zones (Fig. 6A) and then calculated their average π (see Methods) to assess patterns in macro- and micro- diversity, respectively. Viral populations were designated as ‘multi-zonal’ if they were observed in >1 ecological zone, ‘zone-specific regional’ if they were observed in only one zone, but ≥2 viral communities, or ‘zone-specific local’ if they were observed in only 1 viral community within a single zone.

Fig. 6. Size of geographic range positively correlates with microdiversity.

(A) Venn diagram showing the number of viral populations found only in one zone (zone-specific) and those that are shared between and among the five ecological zones (multi-zonal). (B) Stacked barplots showing the number of multi-zonal, regional, and local viral populations found within the species pool of each ecological zone. (C) Boxplots showing median and quartiles of microdiversity (average π for viral populations with ≥ 10x median read depth coverage) per populations found within each zone defined as multi-zonal, regional, or local. Statistics were the same as in Fig. 2.

These analyses first revealed differences in the dominant viral geographic ranges across the different ecological zones. For example, multi-zonal viral populations dominated ANT and BATHY (>60% of viral populations found within zone), both across the zone (Fig. 6B) and within each station (Fig. S5), whereas zone-specific regional viral populations dominated TTEPI and ARC and the multi-zonal and zone specific viral populations were approximately equally represented in TT-MES (Fig. 6B). The high levels of zone-specific viral populations in TT-EPI and ARC, as well as the high levels of viral macrodiversity (Fig. 4B-bottom), are indicative of high endemism and suggest these regions may be biodiversity hotspots for marine viruses. In contrast, the ANT and BATHY are composed mostly of multi-zonal viral populations suggesting that they may be sink habitats that are more dependent on migration (sensu Watkinson & Sutherland, 1995). However, across all ecological zones, viral population microdiversity increased with virus geographic range (Fig. 6C; p < 0.05), presumably from varied ecologies providing differing selective niches for the single, widely-distributed population that then drive differentiation through isolation-by-environment processes (sensu Shapiro et al., 2012). Such findings are new for viruses, but parallel the results for eukaryotes (Hedrick, 2006) and bacteria (Rosen et al., 2015) and suggest a universality to isolation-by-environment processes across organismal kingdoms and viruses.

Ecologists have also long observed, across most flora and fauna, that there are latitudinal patterns in diversity across both terrestrial and marine environments. Briefly, the latitudinal diversity gradient suggests that both macro- and micro-diversity are highest at mid-latitudes and decrease poleward (Pianka 1966, Hillebrand 2004, Mannion et al., 2013, Miraldo et al., 2016). We found that both viral macro- and micro-diversity followed the latitudinal diversity gradient except in ARC, where both increased (Fig. 7A). This high equatorial macro- and micro-diversity was consistent across the Indian, Atlantic, and Pacific Oceans as expected (Fig. 7B & C). The Arctic Ocean, however, was not only unexpectedly elevated in diversity, but it also displayed a unique pattern. Specifically, two distinct zones – definable by climatology-derived water mass nutrient stoichiometry (N*; Fig. 7D; see Comparing ARC-H and ARC-L in Methods) – emerged as high (ARC-H) and low (ARC-L) diversity regions that were significantly differentiable at both macro- and micro-diversity levels (Fig. 7E). Further, ARC-H was characterized by low nutrient ratios (N*; >9X lower in ARC-H than ARC-L on average; p < 5E-04) and drove the divergence from the latitude diversity gradient (Fig. S6A).

Fig. 7. Viral macro- and micro- diversity global biodiversity trends.

(A) Loess smooth plots showing the latitudinal distributions of macro- and micro-diversity. (B & C) Equirectangular projections of the globe showing macro- and micro-diversity levels within each sample, respectively, across the global ocean. Samples collected at different depths from the same latitude and longitude are overlaid and the colors representing their macro- and micro- diversity values are merged. (D) Arctic projection of the global ocean showing the geographical division between ARC-H and ARC-L stations. The patterns are largely concordant with the Arctic division by climatology-derived N*. While we did sample across different seasons, the calculated N* values are not dependent on the season (see impact of the coast, depth, and seasons in Methods). (E) Boxplots showing median and quartiles of macro- (left) and micro-(right) diversity of the ARC-H and ARC-L regions. Statistics were the same as in Fig. 2. (F) Loess smooth plots showing the depth distributions of macro- and micro- population diversity. On all the smooth plots, the line represents the Loess best fit, while the lighter band corresponds to the 95% confidence window of the fit. Abbreviations: N*, the departure from dissolved N:P stoichiometry in the Redfield ratio and a geochemical tracer of Pacific and Atlantic water mass (see Methods).

Mechanistically, we interpret these observations as follows. Prior work in this region has shown (i) strong denitrification in the Bering Strait (Devol et al., 1997), which explains the low N* in the west, and (ii) increasing oligotrophy in the Beaufort Gyre due to increasing vertical stratification, which selects against larger algae and for smaller algae and bacteria in the ARC-H (Li et al., 2009). As above, we hypothesize that shorter-term increased host diversity results in increased viral macro- and micro-diversity in ARC-H. Though our GOV 2.0 dataset is confounded by seasonality of sampling, we posit that this elevated summer-time macro- and micro-diversity in ARC may fuel viral ecological differentiation and represent an unrecognized ‘cradle’ of viral biodiversity beyond the tropics. Though this elevated diversity in the Arctic was surprising, together with a similar deviation seen in mollusks (Valdovinos et al., 2003) and recently reported in ray-finned fish (Rabosky et al., 2018), these results call into question whether this decades-old paradigm needs revisiting and suggests that polar regions may be important biodiversity hotspots for viruses, as well as larger organisms.

Finally, as ocean exploration accelerates, patterns in diversity through the vertical layers of the ocean have become a focus. An emergent depth diversity gradient hypothesis suggests that macrodiversity decreases with depth (Costello & Chaudhary, 2017), which has been explored across the World Register of Marine Species that includes some microbes and viruses (http://www.marinespecies.org/), but microdiversity has not yet been explored for any organism. Overall, our virome-inferred diversity patterns were less obviously consistent with the depth diversity gradient, although deep water ocean data were limited (Fig. 7F). Briefly, viral macrodiversity largely followed the depth diversity gradient with high diversity in the surface waters and decreased diversity with depth, whereas viral microdiversity did not as it decreased until 200 m depth, but then sharply increased (Fig. 7F). This deep water increase coincided with an increase in bacterial macrodiversity in the mesopelagic region (Fig. S6B & C), and in TTMES, this bacterial macrodiversity correlated with viral microdiversity (Fig. S6D).

If more extensive deep water sampling confirms these patterns, we see several scenarios that could explain these data. First, we hypothesize that viral microdiversity may, in part, be driven by an increase in macrodiversity of zone-specific bacterial populations in TT-MES, which we interpret as an expansion of host ‘niches’ available for infection that could drive diversification in viruses (Elena et al., 2009). Second, we hypothesize that the decrease in viral macrodiversity may be driven by increased viral microdiversity of some viral populations in the mesopelagic region that can promote competitive exclusion (sensu Hart et al., 2016) as discussed above. Alternatively, lower cell density in the mesopelagic layer (Sunagawa et al., 2015) may result in less encounters between “predator” and “prey”, reducing viral speciation (as a function of reduced number of viral generations), but selecting for viruses with broader host range. Again, testing these hypotheses will require technological advances to measure in situ host ranges and sensitivities of viruses and cells, respectively, at scales relevant to the diversity in nature.

Conclusions:

This study provides a systematic and global-scale view of patterns and drivers of marine viral macro- and micro- diversity that reveals three overarching advances. First, five ecological zones emerge for the global ocean, which contrasts known Longhurst biogeographic patterning in other organisms, but is consistent with observations from the largely co-sampled ocean microbiome (Sunagawa et al., 2015). Second, patterns and drivers of viral macro- and micro-diversity differ per-sample and positively correlate to geographic range. These findings offer hints at underlying mechanisms that impact these two levels of diversity that will guide researchers from discovery to hypothesis-testing as technologies, such as scalable single virus genomics and in situ host range assays, advance towards sampling scales relevant to those in nature. Third, epipelagic waters and the Arctic Ocean emerge from our work as biodiversity hotspots for viruses. While this is surprising given the latitudinal diversity gradient paradigm that the tropics rather than the poles are the cradles of diversity, it is in line with other observations in larger organisms (Valdovinos et al., 2003, Rabosky et al., 2018) and emphasizes the importance of these drastically climate-impacted Arctic regions for global biodiversity. Together, these advances, along with the parallel global-scale ecosystem-wide measurements of Tara Oceans (e.g. de Vargas et al., 2015; Sunagawa et al., 2015; Brum et al., 2015; Lima-Mendez et al., 2015; Roux et al., 2016) provide the foundation for incorporating viruses into emerging genes-toecosystems models (e.g. Guidi et al., 2016, Garza et al., 2018) that guide ocean ecosystem management decisions that are likely needed if humans and the Earth System are to survive the current epoch of the planet-altering Anthropocene.

STAR Methods Text

Contact for Reagent and Resource Sharing

Further information and requests for resources and reagents should be directed to and will be fulfilled by the Lead Contact, Matthew Sullivan (mbsulli@gmail.com).

Experimental Model and Subject Details

Tara Oceans Polar Circle (TOPC) expedition sample collection and virome creation

Between June 2013 and December 2013, 41 samples were collected at different depths from 20 different sites near or within the Arctic Ocean (see full list of samples in Table S3). Physicochemical measurements, sample collection, and DNA extractions were performed using the methods described in (Roux et al., 2016). Extracted DNA was prepared for sequencing using library preparation method described in (Alberti et al., 2017) for viral samples collected during the TOPC campaign (section 4.2) and sequenced using the HiSeq 2000 system (101 bp, paired end reads). Importantly, our sample collection and library preparation methods have known bias towards <0.2um dsDNA viruses (Roux et al., 2017). The TOPC samples were combined with the previously published viromes in (Brum et al., 2015; Roux et al., 2016). Of the previously published dataset, the mesopelagic samples at (Tara stations 37, 39, 56, 68, 70, 76, 78, 111, 122, 137, 138) and the Southern Ocean samples (Tara stations 82_DCM, 84, 85) were sequenced deeper. These combined samples comprise the GOV 2.0 dataset. The number of reads found in each sample can be found in Table S3.

Methods Details

Tara Oceans Polar Circle (TOPC) expedition sample processing and sequencing analyses

Due to different library preparation for the TOPC samples than the original Tara Oceans samples, the previously sequenced mesopelagic samples (Tara stations 68, 78, 111, 137) were prepped using the TOPC library preparation to determine if it impacted our ability to assemble viral populations. We found no significant difference between library preparations in terms of the number of viral genomes assembled and the average genome length (Fig. S7A & B). Additionally, to directly assess the impact of experimental variation between Tara Oceans and TOPC on our ecological interpretations, we applied hierarchical clustering on a Bray-Curtis dissimilarity matrix of our viromes and we found that all of the mesopelagic samples prepared using the TOPC protocols clustered with their respective samples prepared using the original Tara Ocean protocols, and the variation between them was far less than the ecological variation across our viromes (see distances in hierarchical clustering in Fig. S7D). For two surface samples (Tara Stations 100 and 102), we also re-prepped the DNA using the DNA SMART ChIP-Seq kit which allows us to catch ssDNA in the library preparation (Takara) and further sequenced these two samples using the HiSeq 2000 system.

While the Tara Oceans and Malaspina expeditions used the same sampling and storage approaches (described in Roux et al., 2016), the sequencing reads were longer for the latter (101 bp for Tara and 151 bp for Malaspina). Given this, we have performed further analyses to evaluate whether the contribution of this experimental method variation surpasses the ecological variation presented in this study or not. These analyses, which are further described below, showed that ecological variation much better explained the data than experimental methods. To evaluate this, we compared the deep ocean samples collected from the Tara Oceans and Malaspina expeditions to assess their power to predict the correct ecological zone (mesopelagic or bathypelagic) based on the depth of collection (ecological variation) and the sequencing read length (experimental variation). Using three different metrics, namely the r² value in a univariate regression analysis, the bayesian information criterion (BIC) of such constructed univariate model, and the p-value associated with different components in a multivariate regression analysis, we found that the depth of collection, rather than the experimental variation, best predicts the ecological zone (higher r²), with a better model fit (lower BIC), and lower p-value (Fig. S7C). Additionally, we have one Malaspina sample from the mesopelagic ecological zone (the rest are Tara samples), and there is no significant difference between the Malaspina sample and Tara samples in the mesopelagic (Fig. S3C and D). Together these findings demonstrate that the differences between the samples collected during the different expeditions are predominantly the result of ecology and community structure rather than experimental artifact.

All the remaining STAR Methods we used are quantifications and statistical analyses. All the details related to these STAR Methods are therefore provided in the following section, Quantification and Statistical Analyses

Quantification and Statistical Analyses

Viral contig assembly, identification, and dereplication

All samples in the GOV 2.0 dataset (Roux et al., 2016) as well as the previously sequenced TOPC library-prepped mesopelagic samples and the DNA SMART ChIP-Seq kit surface samples were individually assembled using metaSPAdes 3.11.1 (Nurk et al., 2017). Prior to assembly, Malaspina samples from GOV 2.0 were further quality controlled. Briefly, adaptors and Phix174 reads were removed and reads were trimmed using bbduk.sh (https://jgi.doe.gov/data-and-tools/bbtools/; minlength=30 qtrim=rl maq=20 maxns=0 trimq=14 qtrim=rl). Following assembly, contigs ≥1.5kb were piped through VirSorter (Roux et al., 2015) and VirFinder (Ren et al., 2017) and those that mapped to the human, cat or dog genomes were removed. Contigs ≥5kb or ≥1.5kb and circular that were sorted as VirSorter categories 1–6 and/or VirFinder score ≥0.7 and p <0.05 were pulled for further investigation. Of these contigs, those sorted as VirSorter categories 1 and 2, VirFinder score ≥0.9 and p <0.05 or were identified as viral by both VirSorter (categories 1–6) and VirFinder (score ≥0.7 and p <0.05) were classified as viral. The remaining contigs were run through CAT (Cambuy et al., 2016) and those with <40% (based on an average gene size of 1000) of the genome classified as bacterial, archaeal, or eukaryotic were considered viral. In total, 848,507 viral contigs were identified. Viral contigs were grouped into populations if they shared ≥95% nucleotide identity across ≥80% of the genome (sensu Brum et al., 2015) using nucmer (Kurtz et al., 2004). This resulted in 488,130 total viral populations found in GOV 2.0 (see Table S5 for VirSorter, VirFinder, and CAT results), of which 195,728 were ≥10kb.

Viral taxonomy

For each viral population, ORFs were called using Prodigal (Hyatt et al., 2010) and the resulting protein sequences were used as input for vConTACT2 (Jang et al., in press 2018) and for blastp. Viral populations represented by contigs >10kb were clustered with Viral RefSeq release 85 viral genomes using vConTACT2. Those that clustered with a virus from RefSeq based on amino acid homology based on diamond (Buchfink et al., 2015) alignments were able to be assigned to a known viral taxonomic genus and family. For GOV 2.0 viral populations that could not be assigned taxonomy or were <10kb, family level taxonomy was assigned using a majority-rules approach, where if >50% of a genome’s proteins were assigned to the same viral family using a blastp bitscore ≥50 with a Viral RefSeq virus, it was considered part of that viral family.

Viral population boundaries

To determine if our viral populations had discrete sequence boundaries, all reads across the GOV 2.0 dataset (excluding the Tara stations 68, 78, 111, 137 prepped using the TOPC library preparation methods and the DNA SMART ChIP-Seq kit prepped libraries) were pooled and mapped non-deterministically to our viral populations using the ‘very-sensitive-local’ setting in bowtie2 (Langmead & Salzberg, 2012). The percent nucleotide identity (% ID) of each mapped read and the positions in the genome where the read mapped were determined. The frequency of reads mapping at a specific % IDs were weighted based on the length of each read mapped across the genomes. Frequencies of reads mapping at specific % IDs were smoothed using Loess smooth functions (span = 1 to be more permissive of lower % ID reads) to create read frequency histograms (% ID vs. frequency). To determine break in the distribution of read frequencies between the different % IDs, Euclidean distances calculated were calculated between % ID frequencies and then hierarchically clustered in R.

Calculating viral population relative abundances, average read depths, and population ranks

To calculate the relative abundances of the different viral populations in each sample, reads from each GOV 2.0 virome were first non-deterministically mapped to the GOV 2.0 viral population genomes using bowtie2. BamM (https://github.com/ecogenomics/BamM) was used to remove reads that mapped at <95% nucleotide identity to the contigs, bedtools genomecov (Quinlan & Hall, 2010) was used to determine how many positions across each genome were covered by reads, and custom Perl scripts were used to further filter out contigs without enough coverage across the length of the contig. For downstream macrodiversity calculations, contigs ≥5kb in length that had <5kb coverage or less than the total length of the contig covered for contigs <5kb were removed. For downstream microdiversity calculations, all contigs with <70% of the contig covered were removed. BamM was used to calculate the average read depth (‘tpmean’ -minus the top and bottom 10% depths) across each contig. For the macrodiversity calculations, the average read depth was used as a proxy for abundance and normalized by total read number per metagenome to allow for sample-to-sample comparison. The rank abundance of all the viral populations was calculated using the normalized abundances and the ‘rankabundance’ in the BiodiversityR R package.

Subsampling reads

Unequal sequencing depth can have large impacts on diversity measurements, specifically α-diversity measurements (Lemos et al., 2011). Due to 5x more sequencing depth in TOPC samples and the deeply sequenced mesopelagic and Southern Ocean samples (Table S3), all viromes in the GOV 2.0 dataset were randomly subsampled without replacement to 20M reads for Tara or 10M reads for Malaspina (as many Malaspina samples were <20M reads and there was no significant difference between the 10M and 20M reads assemblies; p = 1) using reformat.sh from bbtools suite (https://sourceforge.net/projects/bbmap/). The subsampled read libraries were assembled using metaSPAdes 3.11.1. Contigs ≥1.5kb that shared ≥95% nucleotide identity across ≥80% of the genome with the 488,130 viral populations in GOV 2.0 were pulled out and grouped into populations to be used as the subsampled GOV 2.0 viral populations. In total, there were 46,699 viral populations. Relative abundances were calculated per sample as aforementioned for macrodiversity calculations, but using the subsampled GOV 2.0 viral populations and the subsampled reads.

Macrodiversity calculations

The macrodiversity α- (Shannon’s H) and β- (Bray-Curtis dissimilarity) diversity statistics were performed using vegan in R (Dixon, 2003). The α-diversity calculations were based on the relative abundances produced from the subsampled reads. Loess smooth plots with 95% confidence windows in ggplot2 in R were used to look at changes in Shannon’s H across latitude (Fig. 7A) and depth (Fig. 7F). For the β-diversity, both the subsampled and the total reads abundances were used to look at community structure (Fig. S3). Principal Coordinate analysis (function capscale of vegan package with no constraints applied) and NMDS analysis (function metaMDS; K=2 and trymax=100) were used as the ordination methods on the Bray-Curtis dissimilarity matrices from both the subsampled and total reads calculated from GOV 2.0 (function vegdist; method “bray”) after a cube root transformation (function nthroot; n=3). The ecological zones that emerged were verified using a permanova test (function “adonis”) and the confidence intervals were plotted using function “ordiellipse” at the specified confidence limits (95% and 97.5%) using the standard deviation method. There were no significant differences in clustering between the subsampled and all reads Bray-Curtis dissimilarity PCoA plots (Fig. S3). Hierarchical clustering (function pvclust; method.dist=“cor” and method.hclust=“average”) was conducted on the same Bray-Curtis dissimilarity matrices using 1000 bootstrap iterations and only the approximately unbiased (AU) bootstrap values were reported. The heatmaps were generated using the heatmap3 package with appropriate rotations of the branches in the dendrograms. Samples that did not cluster with their ecological zone (Tara mesopelagic stations 72, 85, and 102 and Tara surface station 155) were considered outliers and removed from further analyses (Fig. S3A & C).

Microdiversity calculations

Viral populations with an average read depth of ≥10x across 70% of their representative contig in at least one sample in the GOV 2.0 dataset were flagged for microdiversity analyses. We used 10x as the minimum coverage because population genetic statistics were found to be relatively consistent down to 10x based on previous downsampling coverage analyses (Schloissnig et al., 2013). BAM files containing reads mapping at ≥95% nucleotide identity were filtered for just the flagged viral populations. Samtools mpileup and bcftools were used to call single nucleotide variants (SNVs) across these populations. SNV calls with a quality call > 30 threshold were kept. Coverage for each allele for each SNV locus was summed across all the metagenomes. For each SNV locus, the consensus allele was re-verified and those with alternative alleles that had a frequency >1% (1000 Genomes Project Consortium, 2012), the classical definition of a polymorphism, and supported by at least 4 reads were considered SNP loci (Schloissnig et al., 2013). Nucleotide diversity (π) per genome were calculated using equation from (Schloissnig et al., 2013). Due to the variable coverage across the genome, coverage was randomly downsampled to 10x coverage per locus in the genome. For the downsampling, if there was not the target 10x coverage for the locus, all of the alleles were sampled. Nucleotide diversity (π) was calculated for each genome with an average read depth ≥10x across 70% of their contig in each sample. For each sample, π values of 100 viral populations were randomly selected and averaged. This was repeated 1000x and the average of the all 1000 subsamplings was used as the final microdiversity value for each sample. Loess smooth plots with 95% confidence windows in ggplot2 in R were used to look at changes in average π across latitude (Fig. 7A) and depth (Fig. 7F).

Annotating Genes & Making Protein Clusters

Genes were annotated by translating the sequences into proteins and running a combination of reciprocal best blast hit analyses against the KEGG database (Kanehisa et al., 2002), and blast against the UniProt Reference Clusters database (Suzek et al., 2007), searching for matches against the InterPro protein signature database using InterProScan (Zdobnov et al., 2001), and running HMM searches against Pfams (Bateman et al., 2004). A diamond ‘blastall’ alignment search (Buchfink et al., 2015) of all the protein sequences was performed against all the protein sequence was performed and the protocol “Clustering similarity graphs encoded in BLAST results” with a granularity of I=2 from the MCL website (https://micans.org/mcl/; Enright et al., 2002) was used to create protein clusters.

Selection Analyses

Natural selection (pN/pS) was calculated using the method from (Schloissnig et al., 2013). The pN/pS method compares the expected ratio of non-synonymous and synonymous substitutions based on a uniform model of occurrence of mutations across the genome with the observed ratio of non-synonymous and synonymous substitutions. The original method treats each SNP locus as independent from each other. Thus, if two SNPs occur in the same codon, the alternate codon produced from each SNP would be considered in the pN/pS calculation. Thus, if two SNPs occur in one codon, the effect of the SNPs could potentially cancel each other out or amplify a non-synonymous signal leading to false positive selection calls. In order to minimize this bias, SNPs found within the same codon in the same gene were tested for linkage in each metagenome. If SNP alleles from loci within the same codon had depth coverage within 15% of each other within each metagenome, they were considered linked in that sample.

For each codon with SNP loci in a gene, the minimum coverage was identified based on the lowest read depth coverage among the three base pair position. The initial number of the consensus codon was determined based on the lowest coverage of the consensus alleles at the SNP locus or loci if linked. The initial numbers of potential alternate codons was based on the coverage of the alternate allele at that position or the lowest coverage between two linked SNPs. The final coverage of the each codon per SNP locus was calculated by taking the rounded down number of the product of the initial number x (initial number/ minimum coverage for the codon). These codons then subsampled down to 10x. The number of observed non-synonymous and synonymous substitutions were counted and pN/pS was calculated. Genes were considered under positive selection if pN/pS was >1.

Drivers of Macro- and Micro-diversity

Regression analysis between the first coordinate of the PCoA (Fig. 5A) and available temperature measurements was conducted using the lm function in R. The environmental variables were fitted to the first two dimensions of the PCoA using a generalized additive model (function envfit; permutations=9999 and na.rm = TRUE). Then, they were correlated with all the PCoA dimensions using a mantel test (function mantel; permutations=9999 and method=“spear”) after scaling (function scale) and calculating their distance matrices (function vegdist; method “euclid” and na.rm = TRUE). Finally, they were correlated with Shannon’s H and π using Pearson’s correlation (function cor; use=“pairwise.complete.obs”) after removing Shannon’s H outliers based on a boxplot analysis (Fig. S4). Both Pearson’s and Spearman’s correlations are provided in (Table S7).

Subsampling macro- and micro- diversity

Due to unequal sampling across each ecological zone, we chose to normalize the number of samples between each ecological zone by subsampling the down to lowest zone sample size (ANT; n = 5). Shannon’s H outliers were not included in the subsampling. Five samples within each zone were randomly subsampled without replacement and their macro- and micro- diversity values averaged, respectively. We subsampled 1000x and plotted the averages and assessed for significant differences using Mann-Whitney U-tests in ggboxplot from the R package ggpubr (Fig. 4B).

Classifying multi-zonal, regional, and local viral populations

To determine geographic range, viral populations were evaluated for their distributions across the five ecological zones and plotted using the VennDiagram package in R (Fig. 6A). If present in ≥1 sample in more than one ecological zone, it was considered multi-zonal (58% GOV 2.0 viral populations). If present only in samples found within a single zone, it was considered zone-specific (48% GOV 2.0 viral populations). Zone-specific viral populations were further divided into regional (≥2 samples within a zone) and local (only 1 sample within a zone). The proportion of multi-zonal, regional, and local viral populations found across each zone (Fig. 6B) and across each station (Fig. S6) were calculated by dividing the number of each type by the total number of viral populations found across a zone or station, respectively. To assess the impact of geographic range on microdiversity per zone, stations were randomly subsampled without replacement as described above. Within each sample, π values of 50, 100, and 20 viral populations of each geographic distribution (multi-zonal, regional, and local, respectively) were randomly selected and averaged. All the viral populations with a geographic range were sampled and averaged in samples that lacked enough deeply-sequenced viral populations with particular geographic range. This was repeated 1000x and the averages plotted and assessed for significant differences using Mann-Whitney U-tests in ggboxplot from the R package ggpubr (Fig. 6C).

Comparing ARC-H and ARC-L

The ARC-H and ARC-L regions were defined based on their biogeography; the ARC-H stations were located in the Pacific Arctic region, the Arctic Archipelago, and the Davis-Baffin Bay, in addition to one station (Station 189) in the Kara-Laptev sea, which was separated by a land mass from the rest of the stations in the same area (Fig. 7D). The ARC-L stations were located in the Kara-Laptev Sea (except Station 189), the Barents Sea, and subpolar areas (stations 155 and 210). The departure from the dissolved N:P stoichiometry in the Redfield ratio (N*) was calculated as in (Tremblay et al., 2015) to represent the deficit in dissolved inorganic nitrogen (DIN) in the ratio and as a geochemical tracer of pacific and atlantic water masses. Macro- and micro- diversity values for each station in ARC-H and ARC-L were plotted and assessed for significant differences using Mann-Whitney U-tests in ggboxplot from the R package ggpubr (Fig. 7E).

Comparing GOV to GOV 2.0

Viral populations assembled in the GOV (Roux et al., 2016) were compared to the GOV 2.0 viral populations (Fig. 1B) using blastn. Unbinned GOV viral populations with a nucleotide alignment to a GOV 2.0 viral populations with ≥95% nucleotide identity and an alignment length ≥50% the length were considered present in the GOV 2.0. These results were plotted in a venn diagram using the VennDiagram package in R. The frequency of contig lengths of viral populations that were shared across both samples were plotted using ggplot2 (function “geom_histogram”; binwidth =5000).

Calculating 16S OTU Macrodiversity

Previously published 16S OTU data were taken from (Logares et al., 2014). The macrodiversity α- (Shannon’s H) statistics were performed using vegan in R (Dixon, 2003). Loess smooth plots with 95% confidence windows in ggplot2 in R were used to look at changes in bacterial Shannon’s H down the depth gradient. Differences between surface, deep chlorophyll maximum, and mesopelagic bacterial samples were compared using Mann-Whitney U-tests and plotted in ggboxplot from the R package ggpubr. Finally, viral microdiversity was correlated with bacterial Shannon’s H using Pearson’s correlation (function cor; use=“pairwise.complete.obs”) and a linear regression (Fig. S6D).

Impact of the coast, depth, and seasons

GOV 2.0 samples are largely open ocean samples. Even though the arctic samples were more coastal, we didn’t observe any significant coastal impact on the global macrodiversity (Pearson’s r = −0.25; Bonferroni-corrected p-value = 0.15) and microdiversity (Pearson’s r = 0.11; p-value = 0.23) levels (Fig. 4C). Although nitrate and phosphate levels generally increase with depth, we observed higher negative correlations and significantly lower p-values for these nutrients with macrodiversity levels than between depth and macrodiversity (Fig. 4C) which suggests an impact of nutrients on viral diversity via primary production (Fig. 5C). Additionally, since the sampling was largely at discrete depth layers with different densities in the TT region (epipelagic, mesopelagic, and bathypelagic), rather than sampling gradients, we discerned a clearer signal for the separation between these ecological zones (Fig. 4A). On the other hand, all the arctic epipelagic and mesopelagic samples fell within the same ecological zone due to the absence of a pycnocline in this area (Fig. 4A). Finally, the circumnavigation of the Arctic Ocean spanned multiple seasons (spring, summer, and fall). Based on our previous observation from a time-series data in a sub-arctic system (Hurwitz & Sullivan, 2013), our viral macrodiversity is expected to be lowest during the spring and summer and increase towards the winter season. However, our calculated N* values are not dependant on the season and represent the largest magnitude of change among all of the environmental variables that correlated with macrodiversity between the ARC-H and ARC-L regions.

Assessment of microbial contamination

To quantifying microbial contamination across our samples, we screened our metagenomic reads using singleM (github.com/wwood/singlem) for 16S sequences using the dedicated 16S SingleM package. We found that our viromes are exceptionally clean. Specifically, the number of 16S sequences in our samples ranged from 0–40 per million reads (Table S3), and hence the samples are considered to have “likely negligible bacterial contamination” according to the metric proposed by authors evaluating such signals in published viromes (threshold was 200 16S sequences per million; Roux et al., 2013). In spite of our viromes being exceptionally clean, we sought to evaluate the impact of any variation in 16S, and hence bacterial contamination, however small, on our findings. We found that even though microbial contamination increases with depth (most probably due to the decrease in cell size; linear regression r² = 0.89), this increase was driven mainly by the bathypelagic samples. Briefly, the average contamination in BATHY was 28.7 per million reads (standard deviation = 6.8) as compared to the rest of the samples (average contamination = 1.7 per million reads and standard deviation = 2). These bathypelagic samples were not included in any of the ecological driver analyses due to the unavailability of the environmental data to us. Further, it is clear that our estimates of diversity were not influenced by the minor variations in the negligible contamination in our viroomes as a linear regression between Shannon’s H and the number of 16S reads from deep ocean samples resulted in a negligible r² value (0.06). These data (used for conducting the regression analysis) represent a large range of diversity (3.3–7.8) and the full range of contamination (0–40), but avoid the convolution from the ecological difference between the surface and deep ocean layers. Thus, we conclude that the diversity observations we make in this study are driven by ecological variation far greater than microbial contamination.

Data and Software Availability

Code availability

Scripts used in this manuscript are available on the Sullivan laboratory bitbucket under GOV 2.0.

Data availability

All raw reads are available through ENA (Tara Oceans and TOPC) or IMG (Malapsina) using the identifiers listed in Table S3. Processed data are available through iVirus, including all assembled contigs, viral populations and genes.

Supplementary Material

1

Fig. S1. Related to Figures 1 & 4. Bioinformatic workflow. Flow diagrams showing the bioinformatic workflow for (A) the assembly and identification of viral populations, (B) the population coverages and abundances and how they were used to calculate macro- and micro-diversity calculations, (C) prediction of population boundaries, and (D) how average macro- and micro-diversity calculations per ecological zone were calculated.

5

Fig. S5. Related to Figure 6. Stacked barplots showing the number of multi-zonal, regional, and local viral populations found within the species pool of each station. Ecological zone outliers (see Fig. S3) are excluded.

6

Fig. S6. Related to Figure 7. ARC-H drives the divergence from the latitudinal diversity gradient and microbial 16S OTUs biodiversity deviate from the depth diversity gradient and positively correlates with viral microdiversity in the mesopelagic. (A) Loess smooth plots showing the latitudinal distributions of macro- and micro- population diversity with ARCH and ARC-L regions. The line represents the loess best fit, while the lighter band corresponds to the 95% confidence window of the fit. (B) Loess smooth plots showing 16S OTUs (Logares et al., 2014) macrodiversity distributions down the depth gradient. The line represents the loess best fit, while the lighter band corresponds to the 95% confidence window of the fit. (C) Boxplots showing median and quartiles of surface, deep chlorophyll maximum (DCM), and mesopelagic 16S OTU data taken from (Logares et al., 2014). All pairwise comparisons shown were statistically significant (p<0.05) using two-tailed Mann-Whitney U-tests. (D) Scatterplot showing the positive correlation (Pearson’s correlation r = 0.51; p-value = 0.036) and linear regression (r² = 0.26) between Tara Oceans mesopelagic samples shared between the 16S OTU samples in (Logares et al., 2014) and our viral samples in GOV 2.0.

7

Fig. S7. Related to Figures 1 & 4. Library preparation and experimental conditions comparisons. (A & B) Boxplots showing median and quartiles of the number of assembled viral genomes per total reads sequenced and the average genome lengths in TO and TOPC preparations of Tara mesopelagic stations 68, 78, 111, and 137, respectively. All pairwise comparisons shown were not statistically significant using two-tailed Mann-Whitney U-tests. (C) Depth (as an ecological variable) predicts the ecological zone of the deep ocean (mesopelagic or bathypelagic) better than experimental variation between Tara and Malaspina expeditions, with a higher r² (left), lower BIC (middle), and lower p-value (right). The first two metrics were calculated from a univariate regression analysis (using depth alone or experimental variation alone as a predictor of the ecological zone), while the third metric was calculated from a multivariate multiple regression analysis that uses both depth and experimental variation as predictors. (D) Hierarchical clustering of a Bray-Curtis dissimilarity matrix calculated from GOV 2.0 viromes to which four additional viromes (black bars) have been added to control for the impact of experimental variation between the Tara Oceans and Tara Oceans Polar Circle expeditions. The four viromes prepared using the Tara Oceans Polar Circle protocols clustered with their respective original samples, which were prepared using the Tara Oceans protocols indicating that experimental variation was far less than ecological variation.

8

Supplementary Table 1. Related to Figure 1. Examples of marine microbial diversity surveys

9

Supplementary Table 2. Related to Figure 1. List of marine virome datasets used in viral macrodiversity studies

10

Supplementary Table 3. Related to Figure 1. List of viromes in GOV 2.0

11

Supplementary Table 4. Related to Figure 2. List of viral speciation studies

12

Supplementary Table 5. Related to Figure 2. Viral Population Stats and Read Mapping Results

13

Supplementary Table 6. Related to Figure 4. Positive Selection Results using pN/pS

14

Supplementary Table 7. Related to Figure 4. Correlations between environmental variables and macro- and micro-diversity.

2

Fig. S2. Related to Figure 4. Non-metric multidimensional scaling (NMDS) and hierarchical clustering of GOV 2.0. As observed with the Principal Coordinate analysis (Fig. 4A), NMDS analysis (A) and correlation-based hierarchical clustering (B) of a Bray-Curtis dissimilarity matrix calculated from GOV 2.0 structured the viromes into five distinct global ecological zones with an approximately unbiased (AU) bootstrap value ≥ 77 in the hierarchical clustering. Four outlier viromes were removed and all the sequencing reads were used, with justification provided in (Fig. S3, C and D), respectively. Abbreviations: ARC, Arctic; ANT, Antarctic; BATHY, bathypelagic; TT-EPI, temperate and tropical epipelagic; TT-MES, temperate and tropical mesopelagic.

3

Fig. S3. Related to Figure 4. Beta-diversity of the total reads and subsampled reads GOV 2.0 dataset and outlier analyses. PCoA of a Bray-Curtis dissimilarity matrix calculated from GOV 2.0 using all the sequencing reads (A) and after randomly subsampling the reads to the same sequencing depth (B). The dissimilarity matrices from (A) and (B) were used to conduct hierarchical clustering on the samples as shown in (C) and (D), respectively. The four viromes which were removed from (Fig. 4) and (Fig. S2) are highlighted with asterisks; sample 1 (station 155_SUR) is the only surface sample in the North Atlantic Drift Province and could have been influenced by the warm surface currents going northward due to the Atlantic Meridional Overturning Circulation; sample 2 (station 85_MES) is the only mesopelagic sample from the Southern Ocean and could have been influenced by the upwelling of ancient deep ocean water (which is also congruent with the similarity observed between deep water bacterial communities of polar and lower latitude (Ghiglione et al., 2012)); sample 3 (station72_MES) fell outside the 97.5% confidence intervals of all the ecological zones; sample 4 (station102_MES) was located in El Niño-Southern Oscillation region and could have been influenced by the upwellings and downwellings in this area. Additionally, samples 1, 3, and 4 were among the Shannon’s H outliers (Fig. S3E). Viral communities still partitioned into five ecological zones after subsampling the reads as shown by the PCoA (B) and hierarchical clustering (D) plots. (E) Boxplot analysis of viral macrodiversity across GOV 2.0 ecological zones. Outliers that fell below the first quantile or above the fourth quantile (function geom_boxplot of ggplot) of each ecological zone were removed before examining the predictors of viral macrodiversity (Fig. 4C). Outliers: 32_SUR, 155_SUR, 56_MES, 70_MES, 72_MES, 102_MES, MSP131, and MSP144.

4

Fig. S4. Related to Figure 4. Schematic showing the interplay of increased microdiversity and competitive exclusion. Viral populations with more microdiversity usually have larger niche sizes and therefore can outcompete viral populations with smaller overlapping niche sizes. This process of competitive exclusion may not be visible in each community as seen across the three communities. Thus, the average of communities such as across ecological zones can better show this relationship.

Key Resources Table

Reagent or Resource	Source	Identifier(s)
Sequencing Reagents and Kits
NEBNext DNA Sample Prep Master Mix	New England Biolabs, Ipswich, MA	Cat n° E6040S
NEXTflex PCR free barcodes	Bioo Scientific, Austin, TX	Cat n° NOVA-514110
Kapa Hifi Hot Start Library Amplification kit	KAPA Biosystems, Wilmington, MA	Cat n° KK2611
DNA SMART ChIPSeq Kit	Takara Bio USA, Mountain View, CA	Cat N° 634865
Deposited Data
Tara Oceans Viromes Raw Reads	Brum et al., 2015; Roux et al., 2016	European Nucleotide Archive (ENA) - see Table S3 for details
Tara Oceans Polar Circle Raw Reads	This paper	European Nucleotide Archive (ENA) - see Table S3 for details
Malaspania Viromes Raw Reads	Roux et al., 2016	Integrated Microbial Genomes (IMG) with Joint Genome Institute - see Table S3 for details
16S rRNA gene Tara Oceans data	Logares et al., 2014	Supplementary materials in Logares et al., 2014
Biogeographical and Physicochemical data	Pesant et al., 2015	PANGAEA (Data Publisher for Earth & Environmental Science) - see Table S3 for details
N* Arctic Data	This paper	Table S3
Software and Algorithms
nucmer (MUMmer3.23)	Kurtz et al., 2004	https://sourceforge.net/projects/mummer/
bbmap 37.57	https://jgi.doe.gov/data-and-tools/bbtools/	https://jgi.doe.gov/data-and-tools/bbtools/
metaSPAdes 3.11	Nurk et al.,2017	https://github.com/ablab/spades/releases
prodigal 2.6.1	Hyatt et al., 2010	https://github.com/hyattpd/Prodigal
diamond	Buchfink et al.,2014	https://github.com/bbuchfink/diamond
VirSorter v2	Roux et al.,2015	https://github.com/simroux/VirSorter
VirFinder	Ren et al., 2017	https://github.com/jessieren/VirFinder
CAT	Cambuy et al.,2016	https://github.com/dutilh/CAT
blast 2.4.0+	ftp://ftp.ncbi.nlm.nih.gov/blast/executables/blast+/	ftp://ftp.ncbi.nlm.nih.gov/blast/executables/blast+/
vConTACT2	Jang et al., in press 2018	https://bitbucket.org/MAVERICLab/vcontact2
bowtie2	Langmead & Salzberg, 2012	https://github.com/BenLangmead/bowtie2
BamM	https://github.com/Ecogenomics/BamM	https://github.com/Ecogenomics/BamM
Bedtools	Quinlan & Hall, 2010	https://github.com/arq5x/bedtools2/blob/master/docs/content/overview.rst
Vegan (R package)	Dixon, 2003	https://cran.r-project.org/web/packages/vegan/index.html
BiodiversityR (R package)	https://cran.r-project.org/web/packages/BiodiversityR/index.html	https://cran.r-project.org/web/packages/BiodiversityR/index.html
heatmap3 (R package)	https://cran.r-project.org/web/packages/heatmap3/index.html	https://cran.r-project.org/web/packages/heatmap3/index.html
ggplot2 (R package)	https://cran.r-project.org/web/packages/ggplot2/index.html	https://cran.r-project.org/web/packages/ggplot2/index.html
ggpubr (R package)	https://cran.r-project.org/web/packages/ggpubr/index.html	https://cran.r-project.org/web/packages/ggpubr/index.html

Highlights:

• Metagenomic assembly of 145 marine viromes uncovered 195,728 viral populations

• Read mapping revealed discrete sequence boundaries among >99% viral populations

• Viral communities separated into 5 distinct ecological zones in the global ocean

• Viral macro- and micro-diversity did not follow the latitudinal diversity gradient