Introduction

Bacterial activity in the ocean is a key driver of biogeochemical cycles; this activity is mediated by bottom-up controls (for example, resource availability and competition), top-down controls (for example, predation and viral lysis) and also bacteria–bacteria interactions (for example, allelopathy or living in consortia). The microbial loop thus links bacteria, protists and viruses, creating a complex microbial community where the bacteria consume organic carbon produced by other organisms following natural death, grazing by protists or viral infection (Azam et al., 1983; Sherr and Sherr, 1988; Fuhrman and Suttle, 1993; Bratbak et al., 1994; Fuhrman, 1999). The dominant top-down controls, or sources of bacterial mortality, in the open ocean are thought to be viral lysis and protistan grazing. Although widely accepted as important, the relative contribution of each control remains a subject of debate and no doubt differs based on location, time, physiological status or identity of the bacteria.

Many studies have sought to quantify grazing and viral lysis to determine the impact of top-down or bottom-up controls on structuring microbial communities (Fuhrman and Noble, 1995; Strom, 2000; Simek et al., 2001; Sherr and Sherr, 2002; Evans et al., 2003; Weinbauer et al., 2003, 2007; Zhang et al., 2007; Baudoux and Veldhuis, 2008; Longnecker et al., 2010; Staniewski et al., 2012). Most studies enriched or removed grazers and viruses to investigate short-term or episodic impacts from which long-term influences were inferred. For example, reduction in grazer activity affected bacterial diversity of active cells, but removal of viruses only affected rates of activity and not diversity of the active cells (Longnecker et al., 2010). Another study similarly suggested that viruses may help control the abundance of rare organisms through selective mortality (Bouvier and del Giorgio, 2007), although the net influence of viruses on bacterial communities has been reportedly mixed (Schwalbach et al., 2004; Hewson and Fuhrman, 2006). Bottom-up controls also affect bacterial diversity yet the net effects may still depend upon viral or protistan activity (Moebus, 1996; Middelboe, 2000; Gasol et al., 2002; Corno and Jürgens, 2008; Sandaa et al., 2009; Ory et al., 2010; Bouvy et al., 2011). These investigations collectively revealed close couplings between viruses, bacteria and protists, but questions remain in our understanding of how ‘top’ communities of viruses and protists affect bacteria at natural concentrations over long timescales.

Ecological networks of trophic interactions have historically been used to characterize complex food webs by the positive and negative interactions within (Sole and Montoya, 2001; Dunne, 2002; Montoya et al., 2006; Olesen et al., 2011). Network analysis has only recently been applied to microbes, and our ability to interpret these networks is still under development (Fuhrman and Steele, 2008; Chaffron et al., 2010; Steele et al., 2011; Eiler et al., 2012; Gilbert et al., 2012). Positive correlations may suggest co-occurrence due to (1) similar preferred conditions, (2) commensalism or (3) a mutualistic relationship between organisms cooperating within the same niche. Indirect relationships, where a third party benefits from an interaction between two others, are potentially common (Miki and Jacquet, 2008, 2010) and may also be detected as a correlated set of three or more. Indirect relationships may also appear as a single correlated pair if the indirect partner is rare or only weakly correlated. Negative correlations, or time-lagged positive ones, may suggest the presence of predation (protist–protist, protist–bacteria and bacteria–bacteria), viral lysis (virus–bacteria) and competition (any two taxa). Time-shifted correlations between viruses and bacteria could represent a succession of taxa perhaps resulting from a lysis event, whereas correlations of bacteria to environmental, viral or protistan parameters could indicate to what extent the environment, viral pressure or grazing activity drives bacterial abundance and community structure.

Here, we queried a seasonally variable, semi-oligotrophic, surface ocean bacterial community monthly over 3 years to determine the links between protistan, viral and environmental factors using culture-independent community fingerprinting methods, community similarity metrics, local similarity analysis (LSA) and construction of association networks. Past research at the San Pedro Ocean Time series (SPOT) has investigated correlations among the smallest plankton, specifically bacteria, protists and archaea over time (Fuhrman et al., 2006; Fuhrman and Steele, 2008; Steele et al., 2011). In this study, association networks (built from LS correlations between individual bacterial, protistan, T4-like myoviral and environmental parameters, including possible time lags) revealed clusters of operational taxonomic units (OTUs) that likely reflect ecologically relevant interactions. We focused specifically on the T4-like myovirus family in lieu of the entire viral community; T4-like viruses are diverse, abundant, detectable through cultivation-independent methods and include bacteriophages of marine cyanobacteria and SAR11/Pelagibacter (Filée et al., 2005; Comeau and Krisch, 2008; Clokie et al., 2010; Chow and Fuhrman, 2012; Zhao et al., 2013). Over this time series, our observations on correlated OTUs revealed: (1) many T4-like virus OTUs significantly correlated to individual bacterial OTUs (reflective of bacterial hosts susceptible to multiple viruses), (2) single viral OTUs significantly correlated with multiple bacterial OTUs (suggestive of a virus’ capability to infect multiple hosts), (3) protistan OTUs significantly correlated with multiple bacterial OTUs (as evidence of non-selective grazing in the case of a phagotroph or broad bacterial use of nutrients released from phototrophs) and (4) protistan OTUs significantly correlated with a single bacterial OTU or taxonomic group (due to selective grazing or nutrient transfer). Finally, inter-correlated clusters of parameters emerged that detailed potential ecological niches and microbial guilds worth further investigation.

Materials and methods

Sample collection

Seawater (∼20 l) was collected approximately monthly at 0 or 5 m at the University of Southern California's Microbial Observatory at SPOT (33' 33° N, 118' 24° W) and filtered for free-living protistan (0.7–20 μm), bacterial (0.22–1 μm) and viral (30 kDa–0.22 μm) community DNA from March 2008 to January 2011, as previously described (Countway et al., 2005; Fuhrman et al., 2006; Vigil et al., 2009; Countway et al., 2010; Steele et al., 2011; Chow and Fuhrman, 2012; Kim et al., 2012; Chow et al., 2013). Molecular data was unavailable for October 2008 (virus), January 2009 (all), March 2009 (bacteria), October–November 2009 (bacteria) and January 2011 (protist). Bulk seawater samples were also collected and analyzed for bacterial and viral abundance by SYBR green epifluorescence microscopy, bacterial production by thymidine and leucine incorporation, and nutrient concentrations using colorimetric methods (Chow et al., 2013).

Fingerprinting microbial communities

Bacteria and viruses

Bacterial community composition was determined by Automated Ribosomal Intergenic Spacer Analysis (ARISA) (Fisher and Triplett, 1999; Brown et al., 2005; Chow et al., 2013; Needham et al., 2013). T4-like myovirus communities were analyzed by terminal restriction fragment length polymorphism (TRFLP) of g23, which encodes the major capsid protein (Chow and Fuhrman, 2012). Viral fingerprints were obtained from both terminal fragments (5′ and 3′). ARISA and g23-TRFLP products were run in duplicate on non-adjacent lanes on an ABI377 by slab gel electrophoresis with internal size standards (Bioventures, Murfreesboro, TN, USA) every 25 bp (50–900 bp) or 50 bp (900–1400 bp). Peaks were identified in DAx (van Mierlo, Inc, Eindhoven, The Netherlands). Fragments, 400–1210 bp (ARISA) and 50–500 bp (g23-TRFLP), were rounded to the nearest 0.1 bp and dynamically binned (Ruan, et al., 2006b; Chow and Fuhrman, 2012). The resulting bins were manually curated to merge bins <0.1 bp wide with the nearest neighbor; each assay was binned independently. ARISA OTUs were assigned an identity by matching ARISA lengths with known sequences and their ARISA products (Chow et al., 2013; Needham et al., 2013); terminal fragments from in silico analysis of publicly available T4-like viral genomes were used to assign identities to environmental g23-TRFLP OTUs.

Protists

Dominant taxa within protistan assemblages were characterized by 18S rDNA-based TRFLP using Euk-A and Euk570-R primers for PCR and HaeIII for digestion (Countway et al., 2005; Vigil et al., 2009); fragments were analyzed on a Beckman CEQ 8000 (Brea, CA, USA). Fragments from Ostreococcus sp. and Phaeocystis globosa cultures were used as positive controls for calibrating and verifying fragment sizes. Protistan OTUs were identified from in silico digestion of 1341 18S rRNA gene sequences from October 2001 at SPOT (Kim et al., 2012).

Peak analysis for all communities

Unique fragment lengths were considered as individual OTUs. Relative abundance of each OTU was calculated by dividing a peak’s area by the total area within the monthly fingerprint. Bacterial and viral OTUs <0.1% of the community and protistan OTUs <0.5% of the community were removed from further analysis, and the remaining peaks were normalized by sample to determine relative abundance per month; each community thus totaled to 100%.

Data analysis

Community similarity

Bray–Curtis similarity was determined for each microbial community independently for all monthly pairwise comparisons in PRIMER-E v6 (Clarke and Gorley, 2006). Bray–Curtis resemblance matrices were compared using RELATE (PRIMER-E), a Mantel-type test, with a Spearman correlation and 999 permutations. Correlations of Bray–Curtis similarities for adjacent month comparisons only were calculated by Pearson-product-moment (Sigmaplot11, San Jose, CA, USA).

Co-correspondence analysis and canonical correspondence analysis

Determination of covariance of microbial community data by co-correspondence analysis was completed with coccorresp in R (Braak and Schaffers, 2004; Simpson, 2009) on log-transformed relative abundance data for months where all three microbial community data sets were available (n=28). Any OTU present in <5 months was excluded. Significance testing was completed by cross-validation with the ‘leave-one-out’ method and permutation tests (n=99). Covariance of communities with environmental parameters was determined in R using cca with a stepwise model from the vegan package v2.0.2 (Oksanen et al., 2011); models were validated by analysis of variance. Estimates for chlorophyll-a concentrations and primary production were downloaded for the grid area surrounding SPOT from National Oceanographic and Atmospheric Administration (NOAA) Coastwatch: (a) SeaWiFS, 0.04167 degrees, West US Science Quality for Chlorophyll-a and (b) SeaWiFS and Pathfinder, 0.1 degrees, Global, Experimental data sets for primary productivity (Hooker and McClain, 2000). Environmental data were transformed as follows: log(value) for bacterial production by thymidine and leucine incorporation, calculated turnover time, chlorophyll-a (bottle) and satellite-based chlorophyll-a; log(value +0.01) for NO2, NO3 and PO4; square-root for bacterial and viral abundance and the virus:bacteria ratio; no transformation for salinity, temperature, sea surface height differential, primary production (satellite), day length and monthly change in day length. Missing environmental data were filled with the overall mean of the transformed data; all data were then normalized to a common scale (subtracted means and divided by s.d.) to account for differences in units before completing canonical correspondence analysis analyses.

LSA and network analysis

We determined LS correlations (ranked Pearson’s correlations) by LS analysis (eLSA) using a linear interpolation for missing values and a delay up to 1 month (Ruan et al., 2006a; Steele et al., 2011; Xia et al., 2011, 2013). Any OTU or environmental parameter that occurred in <5 months was excluded, resulting in 227 bacterial OTUs, 376 T4-like viral OTUs (3′: 171, 5′: 205), 70 protistan OTUs and 30 environmental parameters. P-values were determined using statistical approximation followed by permutation testing to reduce computing time while ensuring accuracy (eLSA option: pmix (Xia et al., 2013)). First, P-values were determined for all vs all pairwise relationships using Feller’s theoretical approximation based on the approximate tail distribution of the maximum partial sum of independent identically distributed random variables (ptheo); second, for any pairwise relationship with ptheo<0.05, the more robust yet more intensive, permutation-based (n=2000) P-value (pperm) was determined. Only local similarity (LS) correlations with q-value <0.10 and pperm<0.0015 were retained for further analysis. The q-value (or false-discovery rate) was the more stringent criteria and led to the specified P-value cutoff; by employing q<0.10, no more than 10% of all remaining ‘statistically significant’ LS correlations may be due to error. These remaining LS correlations were visualized in Cytoscape v2.8.2 (Shannon, 2003; Cline et al., 2007; Smoot et al., 2011). Example networks were selected by taxonomic OTU identification (for example, cyanobacteria) or edge type (or example, correlations between specific OTUs). Random undirected networks of equal size by number of nodes and edges were calculated by the Erdős–Rényi model using the Random Network plugin in Cytoscape. Network statistics were calculated with Network Analyzer as undirected networks using the defaults (Assenov et al., 2008).

Results

Monthly covariance in microbial communities and the environment

T4-like virus community structure varied less than bacterial or protistan communities, whether compared between all months (Figure 1a) or between adjacent months only (Figure 1b). Month-to-month shifts in viral, bacterial and protistan communities occurred concurrently over the 3-year period (Figure 1b). Protistan similarity patterns were significantly correlated with the bacterial community when comparing across all months, although not for communities 1 month apart (Table 1). Protistan community similarity between adjacent months (1 month lag only) was positively correlated with primary production estimates (r=0.579, P=0.004) and negatively correlated to sea surface temperature (r= −0.434, P<0.05). Shifts in community composition of T4-like viruses and bacteria between adjacent months and across all months (any length lag) were also positively correlated; in addition, the 5′ (5H) and 3′ (3H) TRFLP assays for the T4-like viruses were highly correlated as expected, given that they are two related measures (Table 1). Bray–Curtis similarities for bacterial community composition between adjacent months were also negatively correlated to sea surface temperature (r=−0.491, P=0.024) such that communities were more similar from month-to-month during colder months. Bray–Curtis similarities between adjacent months for T4-like viral communities were significantly correlated to bacterial abundance (Figure 1c) and bacterial Bray–Curtis similarities between adjacent months (Figure 1d).

Figure 1
figure 1

Month-to-month shifts in Bray–Curtis similarity within microbial communities. (a) Average similarity within each community, observed approximately monthly, over 3 years. ‘All’ indicates the average similarity of all communities (that is, bacteria, viral and protistan) combined into one meta-community. Line, average similarity; box, 25th and 75th percentiles; and error bars, 10th and 90th percentiles. (b) Bray–Curtis similarity between adjacent months for each microbial group, plotted according to the earliest month (that is, March 2008 for comparing March 2008–April 2008). (c) Correlation of bacterial abundance (y axis) and (d) shifts in bacterial Bray–Curtis similarity between adjacent months (y axis) to viral Bray–Curtis similarity (x axis).

Full size image
Table 1 Correlation of variability patterns between microbial communities based on Bray–Curtis similarities
Full size table

Co-correspondence analysis (Braak and Schaffers, 2004) uncovered covariance of communities from the relative abundance of each measured OTU (Supplementary Table S1). In our analysis, one microbial data set was considered as an independent data set (the predictor), while the predictability of variance in a second (the response) was determined. 47% of variance of one T4-like virus TRFLP assay (3H) was predictable by the other (5H), and 20% of T4-like (3H) variability was due to changes in bacterial community composition. No T4-like community variance was significantly predictable by protistan community composition. However, protistan community variance was significantly predictable (P<0.05) by T4-like viral community variance (5H, 9.2% and 3H, 8.4%) and not by bacterial community composition. Co-correspondence analyses were unable to estimate bacterial community variance from T4-like or protistan community data, despite overall correlation in Bray–Curtis community similarities.

Microbial communities were predicted by up to five environmental parameters at P⩽0.05 using canonical correspondence analysis when compared with the null model (Supplementary Table S1). Bacterial variation (12%) could be explained by chlorophyll-a concentration (bottle data) and salinity, and protistan community variance (11.6%) by day length and bacterial abundance. Day length, change in day length, salinity and temperature explained 28.3% of T4-like viruses (3H); viral abundance, ENSO index, day length, change in day length and temperature explained up to 33.5% of variability observed within the T4-like viral community (5H).

Correlations between individual bacterial, viral and protistan taxa in association networks

Many significant LS correlations were observed between viral, bacterial and protistan OTUs and environmental parameters (Table 2). After significance testing by permutation tests and screening by P-values and false-discovery estimates (q-values), 4365 of 223 446 (2%) possible pairwise LS values were statistically significant (Supplementary Figure S1) and formed one global network (Supplementary Figure S2). Most non-significant LS values ranged from −0.5 to 0.5 (that is, relatively weak correlations). LS correlations for both viral assays had largely similar distributions (Table 2). For simplicity, only LS correlations with the 3′ TRFLP (3H) viral OTUs were included in the following network figures (Figures 3, 4, 5, Supplementary Figures S3 and S4). We focused on 3H as month-to-month shifts in Bray–Curtis similarity, and similarities across all months between the 3H-viral and bacterial communities were significantly correlated.

Table 2 Distribution of significant LS correlations (edges) between all microbial OTUs and environmental parameters (P⩽0.0015, q<0.10)
Full size table

Positive and negative interactions were observed between OTUs (that is, bacteria–bacteria, protist–bacteria and virus–bacteria) with co-occurring (not delayed) or time-shifted (delayed by 1 month) LS correlations. Figure 2 depicts simple networks that occurred within the whole community and the underlying relative abundance data. In Network A, a bacterial OTU and a protistan OTU were positively correlated—potentially indicative of co-occurrence, mutualism or predator-prey interaction. In Network B, the positive time-shifted LS correlation observed between one bacterial OTU and one viral OTU might reflect a lytic relationship. Negative and delayed correlations (not shown) between two viral OTUs could indicate a competitive relationship for hosts, whereas a negative correlation between OTUs may indicate lysis (virus–bacteria) or grazing (protist–bacteria). Pearson’s correlation coefficients were close to LS values unless the optimal LS correlations were time-shifted (Table 3).

Figure 2
figure 2

Two mini networks and the relative abundance of each OTU over time. Each mini network (a and b) depicts microbial OTUs as shapes (bacteria, circles; protists, diamonds; viruses, v-shapes). Lines represent statistically significant LS correlations with LS values shown: solid lines are positive correlations and arrows are delayed correlations, pointing toward lagging OTU. Relative abundance of each node is shown as a percent of each microbial community from March 2008–December 2010 for each network (c and d). Bacterial abundance is shown by the bar graph, whereas the protistan (c) and viral (d) OTU relative abundance is indicated by the line graph. * denotes the time-shifted viral OTU, as described in Table 3.

Full size image
Table 3 Description of LS correlations displayed in Figure 2
Full size table

Bacteria–virus interactions and bacteria–protist interactions differed remarkably in their interconnectivity (Figure 3) despite each network having 18 unique components (unconnected subnetworks). The number of significant correlations between bacterial and T4-like virus OTUs (3H: 1.15% of all possible; 5H: 0.8%) outnumbered bacterial–protistan correlations (0.6%), although this may be because of the higher number of viral OTUs observed (Table 2). The protist–bacteria network (Figure 3a, Supplementary Figure S3) differed from the virus–bacteria network (Figure 3b, Supplementary Figure S4) in terms of the internal structure (that is, number of nodes and how the correlations were distributed between nodes); network statistics are summarized in Table 4. Most LS correlations were positively similar to the giant network of all OTUs; negative correlations were 19.5% (n=16) of protist–bacteria and 25.9% (n=111) of virus–bacteria. The global network retains small-world properties with short path lengths (number of nodes needed to link individual nodes) and a high clustering coefficient ratio such that many nodes are connected to other nodes in close-knit groups—more so than expected by chance alone. Although an OTU in the protist–bacteria network had 1.7 connections on average as opposed to 3.5 connections in the virus–bacteria network, these numbers are identical to those observed in random networks of equal size and far fewer than observed in the global network. However, network density (normalized parameter for the average connectivity) was slightly higher in the protist–bacteria network than virus–bacteria network. Each bacterial OTU was typically associated with a single protistan out, whereas a protistan OTU was often correlated to several bacterial OTUs to form small cliques. A viral OTU was often correlated to multiple bacterial OTUs, which were also correlated to two or more viral OTUs, and resulted in one large interconnected cluster. However, more correlated pairs of one viral to one bacterial OTU were observed than pairs of one protist to one bacterial OTU, which could skew the network density calculations when looking at average correlations per node overall. Network heterogeneity (unevenness of the number of connections per node) is lower in protist–bacteria as compared with a random network of equal size, but higher in virus–bacteria, which would confirm a higher skew in the distribution of connections per viral OTU as opposed to protistan OTU. If secondary connections between OTUs of the same type (that is, bacteria–bacteria, protist–protist and virus–virus) were also shown, many small clusters would be connected although subgroups remained apparent.

Figure 3
figure 3

Broad overview of interactions between (a) protists and bacteria only, and (b) T4-like viruses and bacteria only. Microbial nodes are bacteria, circles; protists, diamonds; viruses, v-shapes. Node color indicates its number of edges according to the scale shown in the upper right. Solid lines are positive correlations with no delay; dashed lines, negative correlations with no delay; sine-wave lines, negative-delayed correlations; and forward-slashed lines, positive-delayed correlations. Arrows point toward the lagging OTU. Note that correlations between similar taxa (for example, bacteria–bacteria, protist–protist and virus–virus) were omitted.

Full size image
Table 4 Global and local network statistics for microbial association networks
Full size table

We examined significant correlations for the five most abundant surface ocean bacterial OTUs. These four putative SAR11 and one Actinobacterium (OCS155_435.5) were significantly correlated to several other viral, bacterial and protistan OTUs (Figure 4). This network comprised 66 nodes and 67 edges. OCS155_435.5 was negatively correlated to bacterial OTUs only. The four SAR11 OTUs were correlated to bacterial and viral OTUs, with several delayed correlations (1 month lag); SAR11_S1_666.4 was correlated to the El Niño Southern Oscillation Index (MEI). Four protistan OTUs, identified as Ostreococcus_259, Ichtyosporea_593, Stramenopile_598 and Unknown_280, were correlated with delay to two of the dominant SAR11 OTUs. The fifth protistan OTU, Unknown_127, was negatively correlated with no delay to SAR11_686.9.

Figure 4
figure 4

Top five bacterial OTUs differentially correlate to bacterial, viral and protistan OTUs. Top five bacterial OTUs are highlighted as white circles. All other nodes are bacteria, circles; protists, diamonds; viruses, v-shapes; abiotic, hexagon. Node labels indicate an abbreviated identity (where available) and fragment length. Note that SAR11_S1 indicates SAR11 Surface Clade 1. Solid lines are positive correlations with no delay; dashed lines, negative correlations with no delay; sine-wave lines, negative-delayed correlations; and forward-slashed lines, positive-delayed correlations. Arrows point toward the lagging OTU.

Full size image

Case study: cyanobacteria, possible grazers and viruses

Potential top-down relationships were determined from cyanobacterial, protistan and viral OTUs; this network of cyanobacterial OTUs and their correlated partners has 65 nodes with 66 correlations (Figure 5). Some cyanobacterial OTUs were connected to multiple viral OTUs and others were connected primarily to protistan or other bacterial OTUs; many virus–bacteria correlations were delayed. A non-simple path (a.k.a. a ring structure or cycle) that also included internal rings was identified for a series of correlations between cyanobacterial OTUs, other bacterial OTUs and one viral OTU. Three viral OTUs had TRFLP patterns consistent with cultured isolates by in silico analysis of g23 genes: (1) 3H_408.9: S-SM2 isolated from Synechococcus WH8017; (2) 3H_413.5: Syn9 or Syn19 isolated from Synechococcus WH8012 and WH8109, respectively; and (3) 3H_415.5: S-SSM7 isolated from Synechococcus WH8109 (Sullivan et al., 2003, 2010; Weigele et al., 2007). Protistan OTUs included two potential cyanobacterial grazers: (1) dinoflagellate or Lingulodinium-relative (though possibly a phototroph or mixotroph) and (2) a ciliate. Salinity and temperature were correlated to a potentially low-light Prochloroccus (OTU 907.8).

Figure 5
figure 5

Cyanobacterial OTU correlations to other microbial OTUs reveal potential lytic virus–host relationships, grazing and temporal trends. Cyanobacteria OTUs are noted as white circles and labeled as Prochlorococcus (Pro) or Synechococcus (Syn), followed by ecotype designation (HL: high light; LL: low light; A/B: Synechococcus group). All other nodes are bacteria, circles; protists, diamonds; viruses, v-shapes; abiotic, hexagons. Node labels indicate an abbreviated identity (where available) and fragment length. Solid lines are positive correlations with no delay; dashed lines, negative correlations with no delay; sine-wave lines, negative-delayed correlations; and forward-slashed lines, positive-delayed correlations. Arrows point toward the lagging OTU.

Full size image

Discussion

Potential bottom-up controls on bacterial, protistan and viral communities

We undertook an integrated assessment of bacterial, protistan and T4-like viral communities over 3 years to address microbe–microbe interactions in the surface ocean and their relationship to environmental conditions. ARISA and 18S TRFLP both surveyed the entire domain of bacteria and protists, respectively, whereas g23-TRFLP focused on a specific viral family, the T4-like viruses. ARISA resolved organisms near the species-level (Brown and Fuhrman, 2005; Brown et al., 2005) and detected OTUs >0.1%; g23-TRFLP detected >100 T4-like virus OTUs each month on average, each representing >0.1% of the total T4-like community (Chow and Fuhrman, 2012) and 18S rDNA TRFLP revealed dominant members (>0.5%) of the protistan community (Kim et al., 2012).

Community variability was predictable by environmental factors, but only to a limited extent; each microbial community was discerned by a unique set of environmental factors. The proportion of predictable variability within the T4-like viral community (28.1%) was higher than within both protists (11.6%) and bacteria (11.9%). Bacterial variance was related to chlorophyll-a and salinity, whereas protistan variance was related to day length and bacterial abundance. Day length may have largely influenced the composition of the autotrophic (and mixotrophic) community, whereas prey availability (as indicated by bacterial abundance) may have been a major factor in shaping the heterotrophic (and mixotrophic) community. T4-like virus community shifts were predicted by day length, change in day length (distinguishes seasons), salinity and temperature—all of which relate to abiotic controls on abundance and physiological status of potential hosts. We were unable to significantly predict overall bacterial community structure from T4-like myoviral or protistan community structure. This likely reflects the complex ecology of virus–bacteria–protist interactions, and that bacterial communities can be acted upon by a variety of viral types, not just the T4-likes targeted in this study. However, bacterial, protistan and T4-like viral communities were significantly correlated with one another at the community level, suggesting similar timing of responses to one another or the environment (Table 1; Figure 1).

Inference of ecological interactions from LS correlation networks

Monthly variation of community structure was apparent and the detailed inquiry that follows revealed underlying relationships between OTUs. Numerous statistically significant correlations between microbial OTUs were observed that potentially represent common, stable relationships in the surface ocean (Table 2). The structure of correlations between protistan and bacterial as compared with viral and bacterial OTUs (Figure 3, Supplementary Figures S3 and S4; Table 4) suggests increased specificity in virus–bacteria interactions. A relationship that switches partners (at a fine taxonomic resolution) often or at random is less likely to result in a significant correlation over a 3-year time series. Network statistics were likely heavily influenced by the presence of one large interconnected subnetwork between viral and bacterial OTUs with at least three apparent hubs in contrast to several smaller hubs of protistan to bacterial OTUs. This is not unexpected. As such, grazing and viral lysis can be selective processes that result in different outcomes (Miki and Jacquet, 2008). For example, grazing can be more influenced by size rather than taxonomy of the prey (González et al., 1990; Monger and Landry, 1992; Simek and Chrzanowski, 1992; Hahn and Höfle, 1999). Taxonomically selective grazing is still thought to occur and may be represented by a protistan OTU linked to one or only a few bacterial OTUs (Figure 3a, Supplementary Figure S3). Protist–bacteria interactions could also represent associations where bacteria obtain nutrients indirectly from particular protists, for example, phytoplankton, and such potential positive (one-way) interactions could be inferred from known protistan phytoplankton (for example, Ostreococcus). Similarly, a virus’ host range may be indicated by the number of bacteria ‘host’ OTUs correlated to a virus OTU such that more connections would suggest a broader host range (Figure 3b, Supplementary Figure S4). Note that the characterized T4-like viruses include several that are known to have a relatively broad host range, compared with other virus families (Sullivan et al., 2003). The large number of correlations between viruses and bacteria supports our observation on the correlated community-level shifts between these two groups.

Gross changes in viral abundance and community structure may have a partitioned, rather than universal impact on the bacterial community. Virus–bacteria interactions were not evenly dispersed throughout the community, as specific bacterial OTUs or clusters were more highly connected than others. Investigation into the five most abundant bacterial OTUs in the surface ocean at SPOT over a 10-year period suggests that three SAR11 OTUs may be equally influenced by interactions with viruses and other bacteria. Recently, a T4-like virus infecting SAR11 was discovered (Zhao et al., 2013); the primers used in this study would result in an uncut fragment of 332 bp based on in silico analysis. Of the potential viral OTUs observed near this length, none were correlated with any individual SAR11 OTUs in this analysis. These viral OTUs were correlated to Flavobacteria OTUs, other Alphaproteobacteria, Ichtyosporea (protist) and other viral OTUs; they were present in 16 months of our study ranging from 6% to 65.7% (3H) and in 28 months ranging from 22.4% to 60.9% (5H). The viral OTUs that did correlate with SAR11 may then represent other unknown pelagiphages. The remaining two most abundant bacteria (Actinobacteria 435 and SAR11_S1_666.4) were correlated to several bacteria, suggesting that bacteria–bacteria interactions may be more crucial to their success (Figure 4). Thus, virus–bacterial interactions defined some niches of our most dominant bacterial OTUs, whereas others were bound more by protistan or bacterial interactions.

Case study: connections between cyanobacteria, co-occurring microbes and the environment

Our network analysis revealed specific virus–cyanobacteria–protist interactions (Figure 5), which suggested that cyanobacteria were differentially responsive to top-down controls similar to their response to environmental pressures. Prochlorococcus and Synechococcus, two marine genera of cyanobacteria, are integral to marine ecosystems as key autotrophs in the microbial loop (Chisholm et al., 1988, 1992; Li, 1994; Campbell et al., 1997; Liu et al., 1997; Partensky et al., 1999; DuRand et al., 2001; Giovannoni and Vergin, 2012). It has been suggested that spatial, temporal and vertical differences in the distribution of specific ecotypes reflect physiological capabilities, adaptation to nutrient utilization and differential mortality (Moore et al., 1998; Martiny et al., 2009; Malmstrom et al., 2010; Partensky and Garczarek, 2010), which was seen in the observed correlations between cyanobacteria OTUs to salinity and temperature.

Ciliates and nanoflagellates are thought to be the predominant grazers of cyanobacteria, and newly identified groups such as lineages of marine stramenopiles are now considered as important bacterivores too (Christaki et al., 1999; Worden and Binder, 2003; Christaki et al., 2005; Massana et al., 2006; Frias-Lopez et al., 2009; Lin et al., 2012). A ciliate (OTU 590) and a dinoflagellate/Lingulodinium sp. OTU were correlated to a high-light Prochlorococcus OTU (828.8) and a Synechococcus (group A-V) OTU, potentially indicative of common grazing controls in the ocean. However, we cannot rule out that this dinoflagellate OTU may be a phytoplankton whose preferred conditions parallel to those of the Synechococcus OTU.

Cyanophage-host systems are some of the best-characterized host–virus models from the ocean. Stable, and seasonally variant, co-existing populations of viruses and their hosts have been observed in the field (Waterbury and Valois, 1993; Suttle and Chan, 1994; Marston and Sallee, 2003; Mühling et al., 2005; Sandaa and Larsen, 2006; Wang et al., 2011); isolate-based laboratory experiments have provided information on the genetic regulation of these interactions (Sullivan et al., 2003; Lindell et al., 2005; Zinser et al., 2009; Weinbauer et al., 2011). The correlations of 3H_415.5 (S-SSM7) and 3H_413.5 (Syn9 or Syn19) to Syn_A.I_1056.1 likely represent known host–virus interactions. The correlated bacterial OTU was a putative Synechococcus; specifically, Synechococcus WH 8109 (original host for Syn19 and S-SSM7) has an empirical ARISA length of 1055 bp, which falls within the OTU-labeled Syn_A.1_1056.1. Other Synechococcus spp. yield ARISA fragments of similar lengths. Both Syn9 and Syn19 have wide host ranges and are capable of infecting Prochlorococcus and Synechococcus (Sullivan et al., 2003). Host range data on S-SSM7 (3H_415.5) and S-SM2 (3H_408.9), to our knowledge, is unavailable, although both were isolated from Synechococcus sp. isolates (Sullivan et al., 2010). With that knowledge, we posit that these specific correlations between viral and putative Synechococcus WH8109-like OTUs 1053 and 1056 represent detection of an ongoing lytic relationship in the surface ocean.

Each cyanobacterial OTU in this network was correlated to at least one other bacterial (non-cyanobacterial) OTU, many of which were known heterotrophs. Prior studies reported that growth rates of Synechococcus are positively affected by lysis of co-occurring heterotrophic bacteria (Weinbauer et al., 2011), and the presence of heterotrophic bacteria in culture with Prochlorococcus either significantly improved or inhibited growth depending on the co-cultured taxa (Sher et al., 2011). The network’s ring structure between these OTUs highlights potential redundancy within the cyanobacterial niches. The main ring includes a series of positively correlated cyanobacterial OTUs with and without delays, potentially illustrating a succession of cyanobacteria. All correlations point toward a positively correlated inner ring of three Prochlorococcus, one Synechococcus and three other heterotrophic bacterial OTUs (highlighted in Figure 5), suggesting that this group forms a guild or clique that follows Synechococcus (upper right) or other Prochlorococcus OTUs (from the left). Thus, these bacteria–bacteria links may identify which bacteria help form unique cyanobacterial cliques, alongside the established environmental features.

Conclusions

Monthly microbial community analysis at SPOT provided a mechanism for exploring relationships between individual bacterial, protistan and viral taxa that influence the seasonal variability in the surface ocean. Ecological relationships in the ocean are complex and the association networks presented here likely represent stable relationships between microbes observed in situ. eLSA has been applied to determine potential interactions that may represent boom-bust relationships over shorter daily timescales (Needham et al., 2013), and others that may occur consistently and have particular ecological significance over longer monthly timescales were shown here. Connectivity of OTUs and observation of independent interconnected clusters indicated that microbial communities are full of potential niches that warrant further investigation. Protistan–bacterial associations were far fewer than virus–bacteria associations, and their connectivity may reflect relative non-selective or size-selective interactions as compared with the web of virus–bacteria interactions that may reflect a virus’ host range. The dominance of positive over negative correlations suggests that on this monthly and seasonal timescale, viral OTUs may be primarily controlled by host availability (that is, viruses following the hosts). The microbial association networks identified factors that were highly correlated to specific OTUs, such as cyanobacteria, and the persistence of those relationships over time. Our association networks support the paradigm that microbes are regulated by both bottom-up and top-down controls, and our findings add another layer of complexity to the bacterial response to changing microbial counterparts and environmental conditions.