ORIGINAL ARTICLE

Defining seasonal marine microbial communitydynamics

Jack A Gilbert1,2,3, Joshua A Steele4, J Gregory Caporaso5, Lars Steinbruck6, Jens Reeder5,Ben Temperton1, Susan Huse7, Alice C McHardy6,8, Rob Knight5,9, Ian Joint1,Paul Somerfield1, Jed A Fuhrman4 and Dawn Field10

1Plymouth Marine Laboratory, Prospect Place, Plymouth, UK; 2Institute of Genomics and Systems Biology,Argonne National Laboratory, Argonne, IL, USA; 3Department of Ecology and Evolution, Universityof Chicago, Chicago, IL, USA; 4University of Southern California, Department of Biological Sciences,Los Angeles, CA, USA; 5Department of Chemistry and Biochemistry, University of Colorado at Boulder,Boulder, CO, USA; 6Department of Algorithmic Bioinformatics, Heinrich-Heine University, Dusseldorf,Germany; 7Josephine Bay Paul Centre for Comparative Molecular Biology and Evolution, Marine BiologicalLaboratory, Woods Hole, MA, USA; 8Max-Planck-Institut fur Informatik, Max-Planck Research Group forComputational Genomics and Epidemiology, Saarbrucken, Germany; 9Howard Hughes Medical Institute,Boulder, CO, USA and 10NERC Centre for Ecology and Hydrology, Wallingford, UK

Here we describe, the longest microbial time-series analyzed to date using high-resolution 16S rRNAtag pyrosequencing of samples taken monthly over 6 years at a temperate marine coastal site offPlymouth, UK. Data treatment effected the estimation of community richness over a 6-year period,whereby 8794 operational taxonomic units (OTUs) were identified using single-linkage preclusteringand 21 130 OTUs were identified by denoising the data. The Alphaproteobacteria were the mostabundant Class, and the most frequently recorded OTUs were members of the Rickettsiales (SAR 11)and Rhodobacteriales. This near-surface ocean bacterial community showed strong repeatableseasonal patterns, which were defined by winter peaks in diversity across all years. Environmentalvariables explained far more variation in seasonally predictable bacteria than did data on protistsor metazoan biomass. Change in day length alone explains 465% of the variance in communitydiversity. The results suggested that seasonal changes in environmental variables are moreimportant than trophic interactions. Interestingly, microbial association network analysis showedthat correlations in abundance were stronger within bacterial taxa rather than between bacteria andeukaryotes, or between bacteria and environmental variables.The ISME Journal (2012) 6, 298–308; doi:10.1038/ismej.2011.107; published online 18 August 2011Subject Category: microbial population and community ecologyKeywords: 16S rRNA; microbial; bacteria; community; diversity; model

Introduction

Only recently with the introduction of moleculartechniques satisfactory descriptions of naturalmicrobial assemblages have been generated (Fiererand Jackson, 2006; Rusch et al., 2007; Costello et al.,2009; Caporaso et al., 2011). In this paper, wesummarize a 6-year time series of 16S rRNA tagpyrosequencing of samples taken from a long-timeseries station in the English Channel. The aim wasto understand seasonal variability and to try todetermine which environmental factors might havethe greatest influence on the varying diversity.

In contrast to terrestrial environments that areessentially static, the marine environment has theadded complication that the dispersion and move-ment of populations will be driven by hydrography.This adds to difficulties of interpretation of results,particularly if the sampling design is Eulurian(a fixed site) rather than Lagrangian (moving withthe water flow). The Western English Channel hasbeen studied intensively for more than 100 years(Southward et al., 2005), and this wealth of dataprovide a robust context with which to exploretemporal microbiological complexity. Inferences canbe drawn regarding how bacterioplankton assem-blages may potentially interact with the environ-ment as well as with specific groups of organisms.

Previous efforts to determine which factors mightaffect microbial communities have largely focusedon the relative importance of temperature andnutrient concentrations (Cullen, 1991; Kirchmanet al., 1995; Morris et al., 2005; Fuhrman et al.,

Received 14 March 2011; revised 13 July 2011; accepted 14 July2011; published online 18 August 2011

Correspondence: JA Gilbert, Institute of Genomics and SystemsBiology, Argonne National Laboratory, 9700 South Cass Avenue,Argonne, IL 60439, USA.E-mail: gilbertjack@anl.gov

The ISME Journal (2012) 6, 298–308& 2012 International Society for Microbial Ecology All rights reserved 1751-7362/12

www.nature.com/ismej

2006; Fuhrman, 2009; Gilbert et al., 2009). These areobvious candidates because of the strong effect oftemperature on biological processes (Nedwell andRutter, 1994) and the fact that nutrient availabilitycan drive niche structure through resource parti-tioning (Church, 2009). Of greatest relevance to thepresent study is the recent demonstration thatbacterioplankton diversity followed a latitudinalgradient, with maximum potential richness beingprimarily driven by temperature, with many otherfactors modulating an intricate network of richnessat any particular temperature (Fuhrman et al., 2008).

The aim of the current study was to furthercharacterize seasonal patterns of bacterioplanktondiversity in the Western English Channel, beyond aninitial 1-year study by Gilbert et al. (2009). Usingthese data, we tested three competing alternativehypotheses about potential drivers of diversitypatterns, namely whether the observed seasonalpatterns correlate with (1) varying concentrationsof inorganic nutrients, (2) annual water–temperaturecycle or (3) the population structure of the eukar-yotic phytoplankton and zooplankton. The nullhypothesis was that the seasonal patterns in micro-bial community composition in the Western EnglishChannel showed no relationship with any of thephysical or biological factors measured in this study.

Materials and methods

Sampling, DNA extraction, 16S rDNA V6amplification and pyrosequencingSeawater samples were collected on 72 instancesfrom January 2003 to December 2008, from the L4sampling site (501 15.000 N, 41 13.020) of the WesternChannel Observatory (http://www.westernchannelobservatory.org.uk). Sampling, extraction, amplifi-cation, and sequencing protocols and environmentalparameter analysis were performed simultaneouslyon the same samples as described previously byGilbert et al. (2009); extensive information can befound in Supplementary Information (Supplemen-tary Tables S1–S3). Bacterial diversity was exam-ined in the context of the broad range of bioticand abiotic variables that are routinely measured atthe Observatory. These included phytoplanktonand zooplankton species abundance, the concentra-tions of ammonia, nitrateþnitrite, phosphate,silicate, total organic carbon and nitrogen, salinity,chlorophyll, photosynthetically active radiation,North Atlantic Oscillation data, day length, primaryproductivity and temperature. Statistical analysesused the routines of PRIMER (Clarke and Warwick,2001; Clarke and Gorley, 2006).

Sequence data analysisAll sequence data were treated as reported pre-viously (Gilbert et al., 2010), using the same qualitycontrol that included random resampling to stan-dardize the sequencing effort as described below,

Sequence data noise reduction using Single-LinkagePreclustering (SLP; Huse et al., 2010) and analysis(sample similarity derived from Bray–Curtis indicesweighted on taxon abundance matrices) also fol-lowed previous protocols. In addition, several noisereduction strategies such as SLP (Huse et al., 2010)and denoiser (Reeder and Knight, 2010) werecompared to examine the impact of pyrosequencingerrors on community diversity patterns observedin the data (see Supplementary Figure S1a). It isimportant to stress that both known and unknownbiases associated with these techniques meant thatthese data could not be seen as quantitative, andhence all analyses are based on relative changesderived through comparison. As the same sequen-cing and sampling effort was applied to eachsample, the operational taxonomic unit (OTU)richness (S) was used as a diversity metric, whichshowed a 97% correlation to two extrapolativeestimators of diversity (Chao1 and Ace) over the 72samples (Supplementary Figure S1b). Changes incommunity diversity and relationship to environ-mental parameters were examined using variousnonparametric multivariate methods, discriminantfunction analysis (DFA), and association networks(see Supplementary Information).

To determine whether microbial communities inthe Western English Channel demonstrated seasonalpatterns over many years, 747 496 16S rDNA V6sequences were analyzed, including those pre-viously published for the year 2007 (Gilbert et al.,2009). To compensate for potential overestimationin diversity resulting from pyrosequencing andamplification errors, a clustering technique wasused. SLP grouped OTUs at 2% sequence identityand an average-linkage clustering followed, basedon pair-wise alignments (Huse et al., 2010), whichresulted in 8794 OTUs. To remove sequencing effortbias, each sample was randomly resampled to thesmallest individual sample sequencing effort (4505)as described before (Gilbert et al., 2009). Thisresulted in a total of 4204 OTUs (for all 72 samplescombined). Approximately, 53% of the OTUs wererepresented by only a single sequence (singletons).These results, in terms of relative abundance, wereconfirmed using a second denoising technique,Denoiser (Reeder and Knight, 2010), which generatedgreater total richness (21 130 OTUs). However, com-parison between Denoiser, SLP and no-denoising/filtering indicated that overall, the same patterns ofcommunity diversity were evident with each techni-que (Supplementary Figure S1). SLP constituted byfar the most conservative OTU predictions, and wastherefore used for subsequent analysis.

Results

Seasonal variations in diversity and persistenceBacterioplankton were very diverse at this stationand a total of 8794 different OTUs (defined using

Marine microbial community structureJA Gilbert et al

299

The ISME Journal

SLP) over a 6-year period were identified. Figure 1summarizes the taxonomic identify of all the OTUssequenced and also gives an indication of thepersistence of OTUs in microbial communities atL4 over a 6-year time period. Although this studyhas shown high diversity of bacterioplankton in theEnglish Channel, as with other studies of naturalassemblages, the majority of sequences could not beidentified to species. Indeed, only 6 of the 10 mostabundant OTUs could be annotated below the levelof Class and, of the top 100 most abundant OTUs,only 2% could be identified to the species level.The taxonomic level to which the OTUs could beidentified was—Phylum (9%), Class (32%), Order(10%), Family (26%), Genus (21%). This was trueusing a number of different annotation strategies(that is, GAST (Sogin et al., 2006); BLAST againstGreengenes (DeSantis et al., 2006), SILVA (Pruesseet al., 2007) and RDP (Maidak et al., 2001); RDPclassifier (Maidak et al., 2001); data not shown,

references in Supplementary Information). Theseresults suggest that a large fraction of as-of-yetuncharacterized lineages were present, even amongthe most abundant taxa, and highlights the difficul-ties associated with accurate annotation of shortread-length tag sequences from hypervariable 16SrRNA regions (Wang et al., 2007; Liu et al., 2008).

Although there are significant seasonal variationsin OTU frequency throughout a 6-year period(Figure 2), there are also strong repeating patterns.As other studies of marine microbial diversity havedemonstrated, the Alphaproteobacteria were themost abundant Class. The OTUs most frequentlyrecorded were members of the Rickettsiales andRhodobacteriales. Other OTUs with high frequencywere the Flavobacteriales (Class: Bacteroidetes) andthere were also peaks in the Gammaproteobacteria(Vibrionales and Pseudomonadales).

Alpha diversity of the observed OTUs (S) wasrelatively constant across the time series, but

Figure 1 Persistence of OTUs in microbial communities at L4 over a 6-year time period. Median OTU abundance, calculated for all timepoints, over a 6-year period is set proportional to node size on a logarithmic scale. Only OTUs found in at least 5% of the time-seriessamples (X4) are shown. This includes 22.53% of the OTUs, representing 97.48% of the sampled organisms. Node coloring shows thedifferences in persistence over time, with the color scale from orange (5%), yellow (16%), green (35%), blue (66%), red (100%) reflectingincreasing persistence.

Marine microbial community structureJA Gilbert et al

300

The ISME Journal

showed distinct cyclical patterns with maxima inwinter and minima in summer (Figure 3). The meanS per time point was 286, with an average minimumof 179 in summer and maximum of 437 in winter.This pattern was further confirmed by permutation-based analysis of variance (of S) for all taxa, and for arange of phyla (Supplementary Table S4). S wasmost similar when comparing the same time of year,and differences between seasons and among yearswere both highly significant. Seasonal differencestended to be greater than inter-annual (greaterpseudo-F values although there were fewer d.f.).This lack of significant interaction terms suggestedthat the seasonal cycle was consistent across years.Overall persistence (Figure 1) was linked to abun-dance; OTUs that were present at more than threetime points accounted for 97.48% of the sequences.In total, only 12 OTUs were found at every 1 of the72 time-points, yet these were exceptionally abun-dant, comprising B35% of all the sequence reads.

Seasonal trends in most abundant bacteriaThe two most abundant Orders were Rickettsialesand Rhodobacterales, and they had different seaso-nal abundances. The Rickettsiales sequences weredominated by the SAR11 clade and tended to peakin winter (Figure 4). At this time, light and primaryproduction were low, and inorganic nutrient con-centrations were at their maximum. In contrast, theRhodobacterales, which were dominated by theRoseobacter clade, tended to peak in Spring andAutumn, when nutrient concentrations were loweryet primary productivity was higher. This is con-sistent with what is known from single-strain-levelstudies; SAR11 are considered to be obligateoligotrophs, while the Roseobacter clade containsmany genera whose cultured representatives tend togrow in organic nutrient-rich media, and may belikely to respond at times when rates of primaryproduction are higher.

Rare taxa may dominate the assemblageThe largest bacterial ‘bloom’ occurred duringAugust 2003, and this constituted a single Vibrio sp.,

Figure 2 Plot representing the seasonal dynamics (grouped as anaverage of seasons; Winter: January–March; Spring: April–June;Summer: July–September; Fall: October–December) of taxagrouped at the taxonomic level of Order in the L4 6-year timeseries. Frequency is recorded based on abundances within aresampled abundance of 4101 sequences per sample. Only Orderswhose average frequency peaked above 10% of the resampledcommunity abundance were included.

Figure 3 Alpha diversity (observed OTUs) plotted as the log ofspecies richness (S) by month spanning 6 years of marine watersampling at the L4 site in the Western English Channel. A cyclicpattern is observed in alpha-diversity, with species richnesspeaking in the winter months.

Figure 4 Plot representing the seasonal dynamics of the bacterial Orders, Rickettsiales and Rhodobacterales, and environmentalparameters, chlorophyll a and soluble reactive phosphorus (SRP) in the L4 6-year time series. Frequency is recorded based onabundances (abundance of sequences per taxa) within a resampled abundance of 4505 sequences per sample.

Marine microbial community structureJA Gilbert et al

301

The ISME Journal

which represented 54% of the sequences.Yet, for the rest of the time series, this taxon wasrelatively rare, having an abundance of 0–2%.Interestingly, this peak was correlated with anincrease in the relative abundance of the diatom,Chaetoceros compressus. This diatom was alsotypically present at low abundance, between0.002–0.2% of total phytoplankton biomass(Supplementary Table S2). However, in August2003, C. compressus accounted for 1.2% of totaleukaryotic phytoplankton. Our data do not distin-guish between a causal relationship—a specificdependence of a bacterial species on a specificphytoplankton species—and simple co-occurrence,which might be a response to unusual environmen-tal conditions. Certainly at this time point, thehighest total organic nitrogen and carbon concentra-tions, and second highest chlorophyll a concentra-tion were measured in the whole time seriesbetween 2003 and 2008 (Supplementary Table S1).

Seasonal succession in the community compositionis robustThe dataset of environmental and biologicalvariables was examined to investigate potentialrelationships between bacterioplankton and theenvironmental and eukaryotic abundance data.The community composition (rather than richness)was used, after determining whether seasonalpatterns in community composition were as robustas those for species richness. Three different subsetsof the bacterial OTUs, that is, the most abundant,most common and most variable (see Supplemen-tary Materials) were defined. These definitions wererobust across the different denoising strategies (thatis, the same OTUs (based on sequence identity, withthe same taxonomic inference defined). Using DFA,an eigenvector technique that, in this case, searches

out the taxa which are best able to predict the month(Fuhrman et al., 2006), we found that for eachsubset, the bacterial community could correctlypredict the month with 100% accuracy, showed aclear repeating pattern (Figure 5), and was able toexplain 460% of the variance in the communitystructure (Supplementary Table S5).

These patterns for most abundant, common andvariable subsets are similar to those reported forsimilar subsets in a Californian near-surface bacter-ioplankton time series (Fuhrman et al., 2006),suggesting that seasonal succession patterns ofmarine surface water bacterial communities intemperate regions may be conserved across differentbiomes. The Californian study was based on auto-mated ribosomal RNA intergenic spacer analysisfingerprint technology, but the sequence-basedannotation provided by this study allowed consid-erably better predictions for the bacterial taxacontributing most strongly to these signals. Inthis instance, these were members of the Alphapro-teobacteria (for example, SAR11 and Rhodobacter-iaciae groups), the Gammaproteobacteria (forexample, Pseudomonas, Pseudoalteromonas, andVibrio groups), the Cyanobacteria, and the Bacter-oidetes (for example, Flavobacteriaceae group;Supplementary Table S6).

Seasonal variance in community compositionThe relative significance of environmental versusbiological factors in describing the seasonal varia-tion in bacterioplankton assemblages was investi-gated using DFA. DFA, via multiple regression usingenvironmental factors and eukaryotic counts, wasused to predict the first discriminant function (DF1)from each subset of the community (that is,most abundant, most common and most variable).Environmental parameters explained 49–91% of the

Figure 5 Annual repeating patterns from the bacterioplankton community sampled monthly from 2003–2008 in the English Channeldetermined by DFA where the model used the bacterioplankton community to predict the month. Upper row of graphs shows the time-series analysis of the first discriminant function (DFA1) over 72 months. The lower row shows the autocorrelation of the discriminantfunction with up to a 50-month lag. The lines in the lower row represent correlations with Po0.05.

Marine microbial community structureJA Gilbert et al

302

The ISME Journal

variance in DF1, while eukaryotic variablesexplained 18–51% of the variance (SupplementaryTable S6). This suggests that that the seasonallyresponsive members of the microbial communitywere responding to changing environmental factors,while interactions between the bacteria and theeukaryotes may have had a less comprehensiveinfluence. Obviously, as shown for the Vibrio bloomin 2003, this trend is not absolutely uniform,and blooms of rare taxa can be influenced by thepresence of eukaryotes. However, as defined by therobust annual cyclicity, the community recoversfrom these ‘rare-bloom’ events, suggesting an overallbottom-up influence on the community compositionand structure. Essentially this suggests that nutrientconcentrations, physical parameters and biology alldemonstrate significant influence in an extraordina-rily complex matrix.

Annual day length cycle explains most of thevariability in the seasonal pattern of species diversityTo test whether changes in nutrients or temperatureprovided the best correlation with changes incommunity diversity, distance-based linear model-ling was used (described in detail in SupplementaryMaterial). This showed that, although a significantfit could be ascribed to a combination of tempera-ture and photosynthetically active radiation and therichness of all OTUs, the most significant fitwas always to the annual change in day length(Supplementary Table S7). This was best modelledby a cosine term (DX1) with the peak centered onDecember 22. When day length (DX1) was combinedwith serial day (D), it described 66.3% of thevariance in OTU richness. However, when examin-ing the phototrophic Cyanobacteria (SupplementaryTable S7), the relationship of richness to daylength was not always evident, for example, diver-sity peaked in spring but not in winter, and hencecoincided with the lowest annual temperatures atL4. To account for the Cyanobacteria and to signi-ficantly improve the fit of our model (dAIC4 �2),a second seasonal artificial term centered onMarch 22 (a sine-derived term—DX2) was addedthat closely tracked temperature. Also, because mostof the taxa show subtle changes in their seasonalcyclicity over these years, it was possible tosignificantly improve the model further by addinga linear time trend term (D). However, this did notimprove the fit for the cyanobacterial communitydiversity, which was remarkably stable over the 6years. Strikingly, the Cyanobacteria were unique inthat a combination of photosynthetically activeradiation, temperature and nitrate/nitrite concentra-tion provided as good a fit as the artificial descrip-tors (DX1, DX2 and D; Supplementary Table S7).Not unexpectedly, this suggests that, unlike othergroups, the species diversity of these primaryproducers can be well defined by a combinationof light availability, nitrogen availability and

temperature, reflecting a different set of nichescompared with the other potentially heterotrophicbacterioplankton.

Discussion

The repeating cycles in bacterioplankton diversityin this Eulerian study raise the question of whetherunique water masses pass through the EnglishChannel, and whether those water masses containcharacteristic bacterioplankton assemblages. This isalmost certainly not the case as the hydrography ofthe Western English Channel has been studiedextensively (Southward et al., 2005). From theearliest studies in the 1930s using drift-bottles, itwas known that there was a strong flow through theEnglish Channel from west to east. Later modellingand observational studies showed the importance ofwind over a very wide shelf region (including theNorth Sea) in determining flow through the WesternEnglish Channel (Pingree and Griffiths, 1980).Southerly winds resulted in the greatest net trans-port of water along the English Channel through theStraits of Dover and into the southern North Sea;westerly winds were less effective.

It has recently been calculated that averageresidence time at the sampling site is on the orderof 2 weeks (Lewis and Allen, 2009), althoughdispersion occurs continuously. The repeating an-nual patterns of bacterioplankton demonstrated inthis study cannot be due to the repeated intrusionof water mass with an annual periodicity. We do notknow how representative these robust annualpatterns are of the entire English Channel. It maybe that the observed patterns represent seasonalchanges in bacterioplankton on the Celtic Sea Shelf,which is advected into the Western English Chan-nel. Given that this advection will largely depend onwind conditions, it seems unlikely that such similarpatterns would occur over a 6-year period. Clearly,further sampling on the European Shelf will berequired to answer the question of the representa-tiveness of this station.

The relationship between OTU richness and daylength is interesting. To the best of our knowledge,this is the only example from a marine dataset wherea single variable has such explanatory capacity(66.3% of the variance in OTU richness). There areexamples from terrestrial systems; for example,tRFLP analysis identified an r2 value of 0.7 betweenbacterial community richness and pH (Fierer andJackson, 2006). Temperature would imply a clearmechanism; we can see no such direct mechanismthat result in day length directly controlling bacter-ioplankton assemblages.

Other environmental factors that could suggestdirect mechanisms did have significant relation-ships. They did not, however, apply to the mostcommon and abundant taxa, but the compositionof the most variable taxa could be significantly

Marine microbial community structureJA Gilbert et al

303

The ISME Journal

predicted by nutrient concentrations (NH4þ , total

organic nitrogen (TON), soluble reactive phosphate,primary production and broad shifts in oceancurrents indicated by the North Atlantic Oscillation(Supplementary Table S6). Overall, we concludethat the monthly pattern and response to broadseasonal changes indicate that the most commonand most abundant bacterial OTUs have temporallydefined niches. In contrast, the most variable OTUshave niches that can be defined temporally as wellas by nutrient pulses and changes in currents.Temporal niche structure suggests taxa with aresilient seasonal pattern, for example, SAR11 andRhodobacteriaciae, although tracking nutrientpulses and currents, are potentially less resilient tochanging environmental conditions. However, therelationship is complex, and potentially a functionof abundance, commonality and variability, as bothSAR11 and Rhodobacteriaciae are in the mostabundant, most common and most variable subset.

Interestingly, interactions were strongest withinthe bacterial and eukaryotic domains rather thanbetween them, and relationships were strongerbetween bacterial taxa than with environmentalvariables. Association network analysis was em-ployed in an attempt to deconvolute the complexnetwork of relationships that were driving theobserved DFA results. However, this revealed thatthe strongest correlations exist between bacterialOTUs (whether abundant, common or variable) and,to a slightly lesser extent among eukaryotes, com-pared with correlations between these two domainsor between either bacteria or eukaryotes andenvironmental factors (Figure 6). Also, the integrityof these relationships was maintained across thethree chosen subsets of OTUs (Figure 6). Evenamong the highly variable OTUs, which mightbe expected to respond to changing conditions

enabling growth from rare to abundant, mostsignificant correlations were still between bacteria(Supplementary Figure S2). Also, at a highlycorrelated (r40.7, Po0.001, qo0.0012) level, therewere many eukaryotic taxa in a loosely intercorre-lated group (Supplementary Figure S2a), but thereare still very few specific connections between theeukaryotes and the bacteria. Mostly the bacteriawere correlated to one another and to the environ-mental factors, and the eukaryotes were alsoconnected to one another and the environmentalfactors. The highly intercorrelated group (Supple-mentary Figure S2b) was almost completely devoidof eukaryotes, but was connected to an herbivorous,parasitic copepod (Poescilostomatoida), and to theseasonal factor DX1, NO2þNO3, and an intercon-nected cluster of Gammaproteobacteria, Bacillusand Actinobacteria OTUs.

Interactions between eukaryotes and bacteriabecame more apparent when moderate correlationswere examined between different subsets of theeukaryotic community and the 300 most abundantbacterial OTUs. Mixotrophic eukaryotes (potentialgrazers on bacteria) and autotrophic eukaryotes bothshowed complex interactions with the prokaryoticcommunity (Supplementary Figure S3). Althoughflagellates (when grouped by size) were correlatedto each other (r¼ 0.59, Po0.001, qo0.0012) and,naturally, to the total number of flagellates, only twobacterial OTUs (a single Rhodobactereaceae OTUand a single Cyanobacteria OTU) are correlated toall three groups (Supplementary Figure S3a). Therewere many bacterial and eukaryotic OTUs, whichcorrelate to two of the flagellate subgroups, and asmaller number, which correlate to only one of theflagellate subgroups. The 5 mm flagellates werenegatively correlated to a Betaproteobacterial and aGammaproteobacterial OTU, and the diatom Paralia

Figure 6 Broad view of correlation network for the microbial community and the environment at station L4. The network shows strongcorrelations (r40.8, Po0.001, qo0.002) between microbial and environmental parameters for the 300 most abundant bacterial taxa (a),the 300 most common bacterial taxa (b), and the 300 most variable bacterial taxa (c). Bacteria are shown in blue, eukaryotes are shown inred and environmental variables are shown in yellow.

Marine microbial community structureJA Gilbert et al

304

The ISME Journal

sulcata in samples with a 1-month lag, whichreflects an increase in those abundant members ofthe community following a decrease in 5 mm-sizedflagellates (Supplementary Figure S3a).

A similar situation was applied to correlationsbetween autotrophic eukaryotes and abundant bac-terial OTUs. The diatom, P. sulcata, correlatednegatively to the total diatom counts with a 1-monthtime lag (Supplementary Figure S3b). This mayindicate a situation where P. sulcata dominatedthe diatom community, while the total number ofdiatoms decreased. These two eukaryotic nodesshared 26 bacterial OTUs that correlated positivelyto P. sulcata and negatively with a 1-month time lagto the total diatom count (Supplementary FigureS3b). These bacterial OTUs may reflect a communityshift indicated by the increase of P. sulcata and the26 Proteobacteria, Bacteroidetes and Verrucomicrobiawhen the total number of diatoms decreased.The winter peak seasonal cycle, DX1, also positivelycorrelated to P. sulcata and negatively correlated,with a 1-month lag, to total diatoms in the same way,possibly implying seasonal community succession.There were positive contemporaneous correlationsbetween P. sulcata and NO3þNO2, between silicateand mixed layer depth, and a negative 1-monthlagged correlation between the North Atlantic Oscil-lation and total diatom counts; these results indicatethat nutrient concentrations may be drivers of thissuccession (Supplementary Figure S3b). Interestingly,there were only positive correlations between bacter-ial OTUs and 2mm flagellates (SupplementaryFigure S3a), even though 2mm flagellates might beexpected to be the major grazers of bacterioplankton.Bacterial OTUs were also positively correlated to totalflagellates, total phytoplankton, coccolithophoresand Emiliania huxleyi (Supplementary Figure S3b).

Many environmental factors were highly corre-lated (r40.7, Po0.001, qo0.0012) with both eukar-yotic OTUs and bacterial OTUs, when both the 300most variable bacteria (Supplementary Figure S4a)and the 300 most common bacteria (SupplementaryFigure S4b) were considered. Strikingly, the seaso-nal index peaking in winter (DX1) was correlatedalmost exclusively to bacterial OTUs, includingProteobacteria (for example, Alphaproteobacteria,Gammaproteobacteria, Nitrospira), unidentifiedbacteria, Deferribacteres and Owenweeksia in boththe common and variable sub-networks (Supple-mentary Figure S4). Cladocera and Echinodermatawere the only eukaryotes that connected to DX1 andthey were negatively correlated with no lag and a1-month lag, respectively. This suggests that seaso-nal factors (for example, day length, which is aproxy for DX1) may be more important for thebacterioplankton than for the eukaryotic commu-nity. The spring seasonal factor, DX2 was correlatedwith a 3-month lag to Cladocera (indicating asummer increase in abundance), and was negativelycorrelated to a Bacteroidetes OTU in the mostvariable subset (Supplementary Figure S4a).

Positive correlations were widespread in themicrobe–environment network. Primary production(monthly average) was correlated to total diatoms,total ciliates, total microzooplankton and a Rhodo-bacteriaceae OTU (which also correlated to dailyprimary production and temperature). Daily primaryproduction (ML primary production, calculatedfrom observed chlorophyll values and integratedover the observed mixed layer depth) was alsopositively correlated to total diatoms, totalphytoplankton, total ciliates and echinodermata(Supplementary Figure S4). This suggests that, asproductivity and nutrients increased, these bacteriaand eukaryotes also increased in abundance, that is,these taxa appear to perform best in a productivesystem. There was little correlation-based evidencefor top-down effects in this system, althoughthis may be a function of a lack of resolution ofbacterivores among the eukaryotes or perhaps alimitation of this kind of analysis.

Local similarity analysis, with its ability to seetime-lagged correlations, also provided insight intothe relationships between environmental factorsthemselves. Although day length was not correlatedto temperature at the 0.7 level, the Winter seasonalcycle (DX1) was negatively correlated to day lengthwith no time delay, and to temperature and primaryproduction with a 1-month time delay (Supplemen-tary Figure S4); that is, day length changed season-ally, followed by a change in temperature. DX1 andday length (which was positively correlated tophotosynthetically active radiation and primaryproduction) may be serving as combinatory signalsof seasonal environmental change, involving factorssuch as changes in input of energy into the system.These combinatory variables may more closely mapthe changes in the whole community of bacterio-plankton as well as the individual bacterial OTUsconnected to them. NO2þNO3 were highly corre-lated with soluble reactive phosphate and silicate(Supplementary Figure S4). However, soluble reac-tive phosphate was correlated only to a Gammapro-teobacteria OTU and a Rhodobacteriaciae OTU,while silicate was not highly correlated to anybacteria or eukaryotes. NO2þNO3 was positivelycorrelated to 12 bacterial OTUs, which werealso positively correlated to DX1, and there were10 bacterial taxa that were positively correlatedsolely to NO2þNO3. The close coupling betweenthese taxa and NO2þNO3 (that is, these taxa wereonly abundant when there was an increased avail-ability of nitrogen) suggests that these taxa may beseasonally nitrogen limited in this ecosystem.

Regardless of the subset of OTUs (for example,abundant, variable, or common) analysed, eachsubset was able to predict the month. In addition,each of the networks appeared to identify many ofthe same connections when we examined 300 taxafrom any of the subsets (Figure 6, SupplementaryFigures S2–S4). OTUs were ranked differently with-in each subset, but they produced similar patterns,

Marine microbial community structureJA Gilbert et al

305

The ISME Journal

which were nearly identical at the r40.8 correlationlevel (Figure 6). This was partly due to the stabilityof the bacterioplankton community at L4 and thedepth of sampling into this community. It was alsoan effect of using statistical analyses that require acertain number of occurrences in order to detect apattern; by design, these analyses would ignore theonce-a-decade occurrence, for example, the spikein Vibrio spp. abundance in the summer of 2003.However, comparing these subsets allowed for abetter sense of the ecology behind these bacterialOTUs. This is demonstrated most clearly whenrestricting the correlations to the 50 most commonand most variable bacterial taxa, and their

relationship to environmental factors (Figure 7).For instance, a SAR11 (Alphaproteobacteria_03_2),although common, changed abundance seasonally(it was the 6th most variable bacterial OTU) andincreased in abundance when inorganic nutrientconcentrations increased. A Rhizobiales member(Alphaproteobacteria_03_121) that correlatedwith NO2þNO3 (Figure 7a) was not as variable(Figure 7b), whereas the Deferribacteres member(Deferribacteres_03_12) that correlated with NO2þNO3 (Figure 7b) was not common (Figure 7a), butincreased in abundance along with increasedNO2þNO3 concentration. Among these observa-tions of common influence, there were also hints

Figure 7 Sub-networks of highly correlated (r40.7, Po0.001) variables built around environmental factors from the 50 most common(a) and 50 most variable (b) bacterial OTUs. Interactions between environmental variables and eukaryotic interactions withenvironmental variables have been removed for clarity. OTU identifications are from http://vampsarchive.mbl.edu/diversity/diversity_old.php. Identifications more specific than the taxonomic order are shown in parentheses. Solid lines represent positivecorrelations, dashed lines represent negative correlations. Black lines show no time delay while red arrows are delayed by 1 month.

Marine microbial community structureJA Gilbert et al

306

The ISME Journal

at ecological differences between these OTUs.Although some taxa seemed to follow inorganicnutrient concentrations (for example, SAR11 andDeferribacteres), others followed system producti-vity (for example, Rhodobacteriales) or temperature(Gammaproteobacteria OTUs; Figure 7). Theseobservations, made possible by extended studiesof microbial assemblages, will lead to deeperunderstanding of microbial niches in the oceanand elsewhere.

This study has confirmed that strong seasonalpatterns occur in this surface water microbial com-munity and that potential drivers of this structurecould be identified from the observatory data. Strik-ingly, the variable with most explanatory power foroverall bacterial richness was day length, whichappears to be as important for describing temporalcommunity structure in coastal temperate seas as pHis for describing spatial microbial structure interrestrial ecosystems. This study has highlightedthe added value of much longer temporal observa-tions of natural communities. Although the overallcommunity succession was robust, subtle changes inthe patterns of individual taxa were observed andwere only detectable because of the long (6 years)time series. Examples of different taxa showingdifferent seasonal cycles were SAR11 and Roseo-bacter, which had nearly exactly opposite peaks inrichness. Additionally, blooms of rare OTUs maybe linked to changes in eukaryotic species andenvironmental variables. Seasonal succession in thecommunity composition was robust and the mostvariable OTUs were best at predicting the time of year.Environmental factors, rather than interactions witheukaryotes, were better at explaining seasonalvariance in bacterial community composition.Meanwhile, interactions were strongest within do-mains rather than between them, and correlativerelationships were stronger between taxa than withenvironmental variables. This may indicate thatbiological rather than physical factors can be moreimportant in defining the fine-grain communitystructure. Finally, in making comparisons of thebacterial OTU subsets, a fundamental stability inthe community has been shown, which suggests thatthe robust seasonal cyclicity noted for the alpha- andbeta-diversity is also self-evident in the interactionsbetween members of the community.

Acknowledgements

We would like to thank Dr KR Clarke for providingextensive expertize in statistical modelling, and MargaretHughes for providing the pyrosequencing technicalsupport. All sequencing data and environmental metadatacan be found in the INSDC SRA under ERP000118 (http://www.ebi.ac.uk/ena/data/view/ERP000118). This workwas supported in part by the US Deptartment of Energyunder Contract DE-AC02-06CH11357. JAF and JAS weresupported by NSF Grant 0703159 and JAS and SH by theSloan Foundation (ICoMM).

DisclaimerThe submitted manuscript has been created in partby UChicago Argonne, LLC, Operator of ArgonneNational Laboratory (‘Argonne’). Argonne, a USDepartment of Energy Office of Science labora-tory, is operated under Contract No DE-AC02-06CH11357. The US Government retains for itself,and others acting on its behalf, a paid-up nonexclu-sive, irrevocable worldwide license in said article toreproduce, prepare derivative works, distributecopies to the public, and perform publicly anddisplay publicly, by or on behalf of the Government.

References

Caporaso JG, Lauber CL, Walters WA, Berg-Lyons D,Lozupone CA, Turnbaugh PJ et al. (2011). Globalpatterns of 16S rRNA diversity at a depth of millionsof sequences per sample. Proc Natl Acad Sci USA108(Suppl 1): 4516–4522.

Church MJ. (2009). Resource control of bacterial dynamicsin the Sea. In: Kirchman DL (ed). Microbial Ecology ofthe Oceans. Wiley & Sons Inc.: NJ, USA.

Clarke KR, Gorley RN. (2006). PRIMER v6: User Manual/Tutorial. Primer-E Ltd.: Plymouth, UK.

Clarke KR, Warwick RM. (2001). Change in MarineCommunities: An Approach to Statistical Analysisand Interpretation, 2nd edn. Primer-E Ltd: Plymouth,UK.

Costello EK, Lauber CL, Hamady M, Fierer N, Gordon JI,Knight R. (2009). Bacterial community variation inhuman body habitats across space and time. Science326: 1694–1697.

Cullen JJ. (1991). Hypothesis to explain high-nutrientconditions in the open sea. Limnology Oceanography36: 1578–1590.

DeSantis TZ, Hugenholtz P, Larsen N, Rojas M, Brodie EL,Keller K et al. (2006). Greengenes, a chimera-checked16S rRNA gene database and workbench compatiblewith ARB. Appl Environ Microbiol 72: 5069–5072.

Fierer N, Jackson RB. (2006). The diversity andbiogeography of soil bacterial communities. Proc NatlAcad Sci USA 103: 626–631.

Fuhrman JA. (2009). Microbial community structure andits functional implications. Nature 459: 193–199.

Fuhrman JA, Hewson I, Schwalbach MS, Steele JA, BrownMV, Naeem S. (2006). Annually reoccurring bacterialcommunities are predictable from ocean conditions.Proc Natl Acad Sci USA 103: 13104–13109.

Fuhrman JA, Steele JA, Hewson I, Schwalbach MS,Brown MV, Green JL et al. (2008). A latitudinaldiversity gradient in planktonic marine bacteria. ProcNatl Acad Sci USA 105: 7774–7778.

Gilbert JA, Field D, Swift P, Newbold L, Oliver A, Smyth Tet al. (2009). The seasonal structure of microbialcommunities in the Western English Channel. EnvironMicrobiol 11: 3132–3139.

Gilbert JA, Field D, Swift P, Thomas S, Cummings D,Temperton B et al. (2010). The taxonomic andfunctional diversity of microbes at a temperate coastalsite: a ‘multi-omic’ study of seasonal and diel temporalvariation. PLoS One 5: e15545.

Huse SM, Welch DM, Morrison HG, Sogin ML. (2010).Ironing out the wrinkles in the rare biosphere through

Marine microbial community structureJA Gilbert et al

307

The ISME Journal

improved OTU clustering. Environ Microbiol 12:1889–1898.

Kirchman DL, Rich JH, Barber RT. (1995). Biomassand biomass production of heterotrophic bacteriaalong 140-degrees-W in the equatorial pacific - effectof temperature on the microbial loop. Deep-Sea ResPart Ii-Topical Stud Oceanography 42: 603–619.

Lewis K, Allen JI. (2009). Validation of a hydrodynamic-ecosystem model simulation with time-series datacollected in the Western English Channel. Mar Syst 77:296–311.

Liu Z, DeSantis TZ, Andersen GL, Knight R. (2008).Accurate taxonomy assignments from 16S rRNAsequences produced by highly parallel pyrosequen-cers. Nucleic Acids Res 36: e120.

Maidak BL, Cole JR, Lilburn TG, Parker CT, Saxman PR,Farris RJ et al. (2001). The RDP-II (Ribosomal DatabaseProject). Nucleic Acids Res 29: 173–174.

Morris RM, Vergin KL, Cho JC, Rappe MS, Carlson CA,Giovannoni SJ. (2005). Temporal and spatial responseof bacterioplankton lineages to annual convectiveoverturn at the Bermuda Atlantic Time-series Studysite. Limnology Oceanography 50: 1687–1696.

Nedwell DB, Rutter M. (1994). Influence of temperature ongrowth-rate and competition between 2 psychrotoler-ant antarctic bacteria - low-temperature diminishesaffinity for substrate uptake. Appl Environ Microbiol60: 1984–1992.

Pingree RD, Griffiths KD. (1980). Currents driven by asteady uniform wind stress on the shelf seas aroundthe British Isles. Oceanol Acta 3: 227–235.

Pruesse E, Quast C, Knittel K, Fuchs BM, Ludwig WG,Peplies J et al. (2007). SILVA: a comprehensive onlineresource for quality checked and aligned ribosomalRNA sequence data compatible with ARB. NucleicAcids Res 35: 7188–7196.

Reeder J, Knight R. (2010). Rapidly denoising pyrosequen-cing amplicon reads by exploiting rank-abundancedistributions. Nat Methods 7: 668–669.

Rusch DB, Halpern AL, Sutton G, Heidelberg KB, WilliamsonS, Yooseph S et al. (2007). The Sorcerer II GlobalOcean Sampling expedition: northwest Atlanticthrough eastern tropical Pacific. PLoS Biol 5: e77.

Sogin ML, Morrison HG, Huber JA, Mark Welch D, HuseSM, Neal PR et al. (2006). Microbial diversity in thedeep sea and the underexplored ‘rare biosphere’. ProcNatl Acad Sci USA 103: 12115–12120.

Southward AJ, Langmead O, Hardman-Mountford NJ,Aiken J, Boalch GT, Dando PR et al. (2005). Long-termoceanographic and ecological research in the WesternEnglish Channel. Adv Mar Biol 47: 1–105.

Wang Q, Garrity GM, Tiedje JM, Cole JR. (2007).Naive Bayesian classifier for rapid assignment ofrRNA sequences into the new bacterial taxonomy.Appl Environ Microbiol 73: 5261–5267.

This work is licensed under the CreativeCommons Attribution-NonCommercial-No

Derivative Works 3.0 Unported License. To view acopy of this license, visit http://creativecommons.org/licenses/by-nc-nd/3.0/

Supplementary Information accompanies the paper on The ISME Journal website (http://www.nature.com/ismej)

Marine microbial community structureJA Gilbert et al

308

The ISME Journal