Introduction

A long-standing challenge is to understand processes governing the composition of ecological communities (Clements, 1916; Gleason, 1927; Diamond, 1975; Weiher and Keddy, 1995; Chesson, 2000; Hubbell, 2001). It is broadly recognized that community assembly is simultaneously influenced by factors that are relatively deterministic and factors that are more stochastic (for example, Dumbrell et al., 2010; Ofiteru et al., 2010; Langenheder and Szekely, 2011). The deterministic class includes selection imposed by the abiotic environment (‘environmental filtering’) and both antagonistic and synergistic species interactions. In contrast, the stochastic class includes unpredictable disturbance, probabilistic dispersal and random birth-death events (see Chase and Myers, 2011). The most immediate goals are to uncover how and why the relative influences of stochastic and deterministic factors vary through space and time (Chase, 2007, 2010; Chase et al., 2009; Myers and Harms, 2009, 2010; Chase and Myers, 2011).

Although there is a rich literature examining microbial community composition across disparate habitats, few studies examine how the relative influences of stochastic and deterministic processes change with environmental conditions (but see Horner-Devine and Bohannan, 2006; Bryant et al., 2008; Kembel et al., 2011). Approaches commonly applied to microbial communities generally lead to inferences regarding whether or not community composition is significantly related to specific environmental factors (for example, Fuhrman et al., 2006; Sun et al., 2006; Bork et al., 2009; Zinger et al., 2011; Gilbert et al., 2012). These inferences are clearly important, but we suggest that applying the concepts and statistical tools that are more commonly used in macroorganism ecology provides a complementary approach that can significantly extend the understanding of microbial communities. We specifically couple phylogenetic community composition with randomization procedures, referred to as ‘null models’ (Webb, 2000; Webb et al., 2002; Graham and Fine, 2008), and examine deviations from null model expectations across temporal and spatial environmental gradients. Doing so characterizes how the relative influences of stochastic and deterministic processes change with environmental conditions (Chase and Myers, 2011; Chase et al., 2011).

Using phylogenetic information to infer ecological processes requires that phylogenetic distance between taxa approximate their ecological niche difference (for example, the difference in habitat requirements). When phylogenetic distance does approximate niche difference, niches are said to have ‘phylogenetic signal’ and to be ‘phylogenetically structured’ (Losos, 2008). Previous work suggests that closely related bacteria have similar habitat associations (Andersson et al., 2010) and that bacteria may be ecologically coherent even at the phylum level (Fierer et al., 2007, Philippot et al., 2010). On the other hand, horizontal gene transfer can move ecological attributes among bacteria (Doolittle, 1999a, 1999b; Welch et al., 2002). Wiedenbeck-Cohan (2011) draw from Haldane’s (1932) classical work to demonstrate that horizontal gene transfer does not, however, ‘scramble’ ecological niches among taxa (see also Cohan and Koeppel, 2008; Philippot et al., 2010). Horizontal gene transfer should instead improve the performance of recipients (Cohan and Koeppel, 2008; Wiedenbeck and Cohan, 2011). Rates of adaptive evolution may therefore increase because advantageous functions can be acquired from other taxa, instead of relying on mutation (Cohan and Koeppel, 2008). Rapid adaptive evolution suggests that bacterial niches should have phylogenetic signal only among close relatives. There should be no phylogenetic signal across deep evolutionary time because adaptive evolution fueled by horizontal gene transfer should not consistently direct all members of a given clade toward a specific niche. We test this expectation before examining phylogenetic community composition.

If across recent evolutionary time more closely related taxa are more ecologically similar, ecological theory links specific patterns of phylogenetic community composition to specific types of processes (Kraft et al., 2007; Fine and Kembel, 2011; Pei et al., 2011). First, if community structure and dynamics are primarily stochastic, community phylogenetic composition and dynamics should not differ significantly from expectations based on random community assembly. Second, if environmental filtering is the most influential process, coexisting taxa should be more closely related than expected by chance at a given point in time. Third, if environmental change deterministically shifts community composition, phylogenetic turnover between assemblages should be greater than expected after controlling for observed turnover in operational taxonomic units (OTUs) composition (referred to as ‘taxonomic turnover’). Fourth, if competition is strong whereby coexisting taxa must be ecologically differentiated from each other, coexisting taxa should be more distantly related than expected by chance for a given point in time. This mechanism of competition is not, however, related to patterns of phylogenetic turnover.

We apply these pattern-to-process linkages to better understand factors governing (i) the composition of subsurface microbial communities; and (ii) changes in the relative influences of stochastic and deterministic processes. Our study further evaluates the utility of including phylogenetic information and associated randomization procedures into analyses of microbial systems. This is achieved through comparison with a companion study (Lin et al., 2012), which examines the same data set from a conceptual and statistical perspective more commonly applied to microbial systems.

Materials and methods

Data collection

A detailed description of sampling methods is provided in Lin et al. (2012). Briefly, nine wells spread across three clusters of three wells each were sampled within the Hanford Integrated Field Research Challenge site (http://ifchanford.pnnl.gov/); North of Richland, WA (Bjornstad et al., 2009). Clusters were ∼30 m apart, forming vertices of an equilateral triangle, and the three wells within each cluster were within 0.6 m of each other. Clusters A, B and C were ∼270 m, 250 m and 280 m, respectively, from the Columbia River. All wells were ∼15 cm in diameter and each well within each cluster was screened at one depth: 10, 13 or 17 m below ground surface. An extensive characterization of all environmental variables potentially relevant to bacterial physiology is not yet available for our system. However, it is known that sediment texture varies vertically (Bjornstad et al., 2009) and that the magnitude of river water intrusion varies with depth and with horizontal distance from the Columbia River (McKinley et al., in review). From December 2009 to September 2010, 354 samples were taken across 41 points in time. Sampling interval varied from biweekly to every two weeks. Water table elevation was also monitored (McKinley et al., in review; Vermeul et al., 2011). Sampled groundwater (4 l) was passed through a 0.2-μm polyethersulfone filter (Supor200, Pall Corporation, Port Washington, NY, USA). DNA was extracted from half of each filter using a PowerSoil-htp 96-well DNA isolation kit (MoBio Laboratories, Carlsbad, CA, USA). PCR was conducted following Costello et al. (2009) and Fierer et al. (2008), and pyrosequencing was carried out on a Roche FLX 454 (Roche Diagnostics, Indianapolis, IN, USA). Sequences were analyzed in QIIME (Caporaso et al., 2010), using PyNast (http://pynast.sourceforge.net) and the Greengenes core set (http://greengenes.lbl.gov) for alignment and FastTree (http://www.microbesonline.org/fasttree/) to estimate a phylogeny containing all OTUs observed across all samples. OTUs were defined using a 97% sequence similarity cutoff. Each sample was rarefied to 500 sequences and, subsequently, the 1000 most abundant OTUs were retained for analyses.

Phylogenetic signal

To ask whether closely related OTUs have similar habitat associations (that is, to test for phylogenetic signal in habitat associations), the abundance-weighted mean subsurface depth and water table elevation were first found for each OTU. For example for depth, we took all records of a given OTU, recorded the depth of each record and the OTU’s abundance in each record, and then found the abundance-weighted mean of depth. This is the OTU’s ‘niche value’ for depth. The analogous procedure was used to estimate OTU niche values for water elevation. The niche values estimate the depth and the water elevation where each OTU is most abundant, thereby characterizing two axes of each OTU’s niche in terms of the habitats where it occurs (for similar approaches in plant and microbial ecology, see Andersson et al., 2010; Pei et al., 2011; Hardy et al., 2012). For each niche axis, between-OTU niche differences were related to between-OTU phylogenetic distances. To summarize major trends in this relationship, between-OTU niche differences were placed in phylogenetic distance bins and median niche difference was found in each bin. Median niche differences were then regressed against phylogenetic distance bin mid-points (similar to Andersson et al., 2010). Maximum phylogenetic distance was 3.64 (arbitrary units) and a bin interval of 0.01 units was used, proving good resolution of the niche difference-to-phylogenetic distance relationship. This allowed identification of the phylogenetic distance threshold beyond which niche differences no longer increased with phylogenetic distance (Figure 3). Using all data (not medians) we then conducted Mantel tests with 999 randomizations above and below the phylogenetic distance threshold (see also Diniz-Filho et al., 2010). We further compared the niche difference-to-phylogenetic distance linear regression slopes above and below the phylogenetic distance threshold to the expectations under Brownian niche evolution. For both niche axes 999 replicates of Brownian niche evolution were run across the empirical phylogeny using ‘sim.char’ in package ‘geiger’. Variance of the evolving niche values was set equal to the empirically observed variance in niche values for each niche axis. In Brownian evolution, moving from the root to the tips of the phylogeny, niche values evolve via Brownian motion: niche values take a random walk across the niche axis, where descendants have niche values similar to their ancestor.

Phylogenetic community composition

To characterize phylogenetic community composition within each sample (unique point in space and time), we quantified mean-nearest-taxon-distance (MNTD) and the nearest-taxon-index (NTI) (Webb et al., 2002) using ‘mntd’ and ‘ses.mntd’ in package ‘picante’. Note that NTI is the negative of the output of ‘ses.mntd.’ MNTD finds the phylogenetic distance between each OTU within a sample and its closest relative also found in the sample. The abundance weighted mean is then taken across these phylogenetic distances.

where is the relative abundance of OTU i in community k, nk is the number of OTUs in k, and is the minimum phylogenetic distance between OTU i and all other OTUs j that are also in k.

To evaluate the degree of non-random phylogenetic community structure, OTUs and their relative abundances were randomized across the tips of phylogeny (null.model=‘taxa.labels’ in ‘ses.mntd’). Good Type I and Type II error rates have been found when using this randomization under conditions similar to our natural system (see Figure 3 in Kembel, 2009). NTI quantifies the number of standard deviations that the observed MNTD is from the mean of the null distribution (999 randomizations) (Figure 1). For a single community, NTI greater than +2 indicates coexisting taxa are more closely related than expected by chance (phylogenetic clustering). NTI less than −2 indicates coexisting taxa are more distantly related than expected by chance (phylogenetic overdispersion). A mean NTI taken across all communities that is significantly different from zero indicates clustering (NTI>0) or overdispersion (NTI<0) on average (Figure 1; Kembel, 2009).

Figure 1
figure 1

Summary of methods characterizing within community phylogenetic composition. After estimating a phylogeny that includes all OTUs across all sampled communities, only the OTUs within a given community are retained (observed OTUs shown as gray circles; numbers are OTU identities). In the example, there are two observed OTUs in the community and the phylogenetic distance between them is two branch length units (sum of gray branches). Abundances are not shown, but are assumed equal so that MNTDobs is 2 units. To generate a null expectation for MNTD the observed OTUs and their relative abundances are randomly placed on the phylogeny (red circles; note that OTU identities are the same, only their positions on the phylogeny have changed). MNTD is then re-calculated, providing one null MNTD value (MNTDnull). Red branches connect the two (randomly placed) OTUs and their sum is 8, so MNTDnull=8 units. The randomization is repeated (blue circles), providing a second MNTDnull value (4 units), and so on until 999 randomizations are completed. This provides a distribution of MNTDnull values to which MNTDobs is compared (Step 3). In the example, MNTDobs (vertical dashed line) is much smaller than the average MNTDnull value. The equation for the NTI is given under Step 3 and measures the difference between MNTDobs and mean MNTDnull (given as ) in units of standard deviations, where standard deviation is measured on the MNTDnull distribution. Repeating Steps 1–3 across all communities provides the NTI distribution in Step 4 (actual distribution shown in Figure 4). An NTI distribution with a mean greater (less) than zero indicates niche-based processes cause OTUs to be, on average, more closely (distantly) related than expected under random community assembly.

Full size image

Turnover in phylogenetic composition through time (phylogenetic β-diversity) was quantified using Beta Mean Nearest Taxon Distance (βMNTD) and Beta Nearest Taxon Index (βNTI), which are the between-assemblage analogs of MNTD and NTI, respectively (Figure 2; Fine and Kembel, 2011; Webb et al., 2011). Instead of examining βMNTD and βNTI across all possible pairwise comparisons, we quantified these metrics for pairwise comparisons through time within each well. This was done to specifically relate temporal dynamics of community composition to the temporal dynamics of river elevation. βMNTD is the abundance-weighted-mean phylogenetic distance among closest relatives occurring in two different communities (Figure 2).

Figure 2
figure 2

Summary of methods for characterizing turnover in community phylogenetic composition between a given pair of communities. After estimating a phylogeny that includes all OTUs across all sampled communities, only the OTUs within a given pair of communities (k and m) are retained (observed OTUs shown as gray circles; numbers are OTU identities). In the example there are two OTUs in each community. The mean of (gray) branch lengths connecting each OTU in community k with its closest relative in community m (and vice versa) gives βMNTDobs=8 branch length units (abundances assumed equal). To generate a null expectation for βMNTD the observed OTUs are randomly placed on the phylogeny (red circles) and βMNTD is re-calculated, always using minimum phylogenetic distances (red branches) connecting OTUs between the two communities. In the example OTU1 and OTU4 are closest relatives (4 units apart) and OTU3 and OTU2 are closest relatives (2 units apart). Taking the mean gives βMNTDnull=3 units. The randomization is repeated (blue circles), providing a second MNTDnull value (8 units, using blue branches connecting closest relatives between communities), and so on until 999 randomizations are completed. This provides a distribution of βMNTDnull values to which βMNTDobs is compared. In the example, βMNTDobs (vertical dashed line) is larger than the average βMNTDnull value. βNTI measures the difference between βMNTDobs and mean βMNTDnull (given as ) in units of standard deviations (equation shown in Step 3). Repeating Steps 1–3 across all pairwise comparisons provides the βNTI distribution in Step 4 (actual distribution shown in Figure 4). A βNTI distribution with a mean greater (less) than zero indicates that niche-based processes cause phylogenetic turnover between communities to be, on average, greater (less) than expected under random community assembly.

Full size image

where is the minimum phylogenetic distance between OTU i in community k and all OTUs j in community m (see above for other variables). A null distribution of βMNTD is found by randomizing OTUs across the phylogeny and recalculating βMNTD 999 times. βNTI is the number of standard deviations that the observed βMNTD is from the mean of the null distribution. βNTI<−2 or>+2 indicates less than or greater than expected phylogenetic turnover, respectively, for one pairwise comparison. A mean βNTI taken across all pairwise comparisons that is significantly different from zero indicates greater (βNTI>0) or less (βNTI<0) than expected turnover on average (Figure 2; Fine and Kembel, 2011; Webb et al., 2011).

Importantly, the randomizations for βNTI control for observed taxonomic turnover. High levels of taxonomic turnover (for example, large Bray–Curtis dissimilarity) may be associated with less than expected, greater than expected, or as expected (random) levels of phylogenetic turnover. For a single pairwise comparison, βNTI>+2, therefore indicates observed phylogenetic turnover is above what is expected given observed taxonomic turnover.

Parametric and non-parametric tests were used to ask whether the mean (or median) of NTI and βNTI distributions were significantly different from zero. To ask whether NTI showed systematic trends through time or with river water elevation we examined linear regressions of NTI against Julian day and against water elevation. Significant regressions would imply systematic changes in the influence of niche-based processes, but non-significant regressions do not imply static community composition: there can be significant turnover in phylogenetic composition (large βNTI) with no change in NTI. To evaluate the influence of changes in river water elevation on phylogenetic turnover, we regressed βMNTD and βNTI against temporal differences in river elevation. Two sets of regression analyses were applied to the βMNTD and βNTI data. The first set examined data across all comparisons within a well. The second examined only comparisons between samples taken in mid-May and each subsequent sample until early-July, which is the time period before and during river water intrusion into the sampled well field (Lin et al., 2012). For both sets, three statistical models were fit (1-exponential, logarithmic, and linear) using maximum likelihood, and the model with the smallest Bayesian Information Criterion (Schwarz, 1978) was chosen as the best model. Although our data come from time series, characterizations of community structure and turnover in community structure are quantified as deviations from a null model. The assumption of statistical independence is therefore upheld.

Previous studies use other metrics such as mean pairwise distance (Webb et al., 2002; Webb et al., 2011), Unifrac (Knight and Lozupone, 2005), and phylosor (Bryant et al., 2008). These metrics account for deep evolutionary relationships among taxa, and are relevant when there is phylogenetic signal in OTU niches through both recent and deep evolutionary time. In our system phylogenetic signal occurs over relatively short phylogenetic distances (see Results). By quantifying phylogenetic distances only between nearest relatives, the metrics provide the most appropriate ecological inferences for our system.

Results

Regressing between-OTU niche differences against between-OTU phylogenetic distances showed a steep positive relationship across relatively short phylogenetic distances (∼13–15% of the maximum phylogenetic distance across the entire tree), but no systematic relationship thereafter (Figure 3). Mantel tests showed that across short phylogenetic distances (<13% of maximum), habitat differences increased significantly with phylogenetic distances for both depth (r=0.14; P<0.002) and river elevation (r=0.12; P<0.002). The rates (i.e., linear regression slopes) at which habitat differences increased with phylogenetic distances followed the Brownian evolution expectation (P⩾0.18 for both habitats). Across longer phylogenetic distances (>13% of maximum) the rates at which niche differences changed with phylogenetic distances were much slower than the Brownian evolution expectation (P<0.002 for both habitats).

Figure 3
figure 3

Median habitat differences between pairs of OTUs as a function of between OTU phylogenetic distances. Medians are taken within phylogenetic distance bins. Black and gray data points represent differences in OTU mean depth and mean river elevation, respectively. The vertical dotted line and the transition from solid to dashed curves indicates the point at which the curve fit becomes non-monotonic, thereby approximating the phylogenetic distance threshold where phylogenetic signal is lost (∼13% of the maximum phylogenetic distance across the entire tree).

Full size image

NTI across all local assemblages was significantly different from zero (t353=−31.7, P«0.0001), with a mean of +2 (Figure 4). βNTI showed a skewed distribution with a long tail stretching beyond the threshold value of +2, and a median of 0.16 that was significantly different from zero (Wilcoxon signed-rank test, V=13711837, P«0.0001) (Figure 4).

Figure 4
figure 4

Kernal density estimates for distributions of abundance weighted nearest taxon index (NTI) and its between-community analog (βNTI). Each observation is the number of null model standard deviations the observed value is from the mean of its associated null distribution. Positive (negative) NTI values indicate shorter (longer) nearest taxon distances within a community than expected by chance. Positive (negative) βNTI values indicate greater (less) than expected turnover in phylogenetic composition than expected given the observed degree of taxonomic turnover. For both metrics, individual values below −2 or above +2 are statistically significant, as indicated by the dashed black lines. The dashed gray line is centered on zero, the expectation under neutral community assembly.

Full size image

Testing for depth and well cluster effects with analysis of variance (ANOVA) showed that mean NTI varied across well clusters (F2,349=8.2, P=0.0003) but not across depths (F2,349=2.2, P=0.11). Mean NTI values for well clusters A, B and C were, +1.81, +1.84 and +2.36, respectively, and all were significantly greater than zero (P«0.0001 for all). Note that depth and well cluster are used as categorical variables.

Distributions of βNTI within sites were skewed, and a robust non-parametric alternative is not available for two-way ANOVA with multiple samples from each treatment. As such, we used the non-parametric alternative to one-way ANOVA (that is, the Kruskal–Wallis test) to test for differences in βNTI medians across depths and well clusters. βNTI medians differed significantly across well clusters and depths (P«0.0001 for both), with the largest medians consistently observed at the shallowest (10 m) sampling depth (Table 1).

Table 1 Medians for Beta Nearest Taxon Index (βNTI) distributions from each depth within each well cluster.
Full size table

Within most sites (one cluster, one depth), NTI did not vary with river water elevation or through time (Supplementary Figures S1 and S2). Notably, however, at the shallow depth within cluster A (see Figure 1 in Lin et al., 2012) the degree of phylogenetic clustering increased with water elevation and through time (Supplementary Figures S1A and S2A).

Relating βMNTD and βNTI to change in river elevation between sample pairs revealed generally weak relationships, although some were statistically significant (Figure 5, Supplementary Figures S3 and S4). Examining comparisons between samples taken in mid-May to subsequent samples up to early-July revealed relatively strong or weak relationships between river elevation and, respectively, βMNTD or βNTI (Figure 5, Supplementary Figures S3 and S4). One important exception for the May-to-July comparisons was the shallow depth within cluster A. In this site, increasing differences in river elevation were associated with very rapid increases in βMNTD (Figure 5a) and an increase followed by saturation of βNTI above +2 (Figure 5d).

Figure 5
figure 5

(a–c) Beta Mean Nearest Taxon Distance (βMNTD) or (d–f) Beta Nearest Taxon Index (βNTI) as a function of difference in Columbian River water elevation. Each panel shows data for the shallow sampling depth (10 m) across one of the three well clusters (see panel headers). Gray squares represent all comparisons from a given well cluster, and solid gray lines show best-fit regression models (linear, exponential or logarithmic). Solid black squares show results from comparing a sample in each well from mid-May to later sampling dates until early-July, and solid black lines are best-fit regression models to those data. Only significant regressions are shown. Horizontal dashed lines in (d–f) indicate βNTI values of −2 and +2, the values beyond which an individual βNTI value is considered statistically significant.

Full size image

Discussion

Overall, we found that environmental filtering strongly determines local community composition of subsurface microbes, and that stochastic factors have an important but secondary role. The influence of stochastic factors over changes in community composition is, in comparison, much greater. However, the influence of stochastic processes relative to deterministic processes varies through space and time (see also Caruso et al., 2011). Our results further suggest that general rules may govern spatial and temporal changes in the relative influences of stochastic and deterministic processes.

Phylogenetic signal in ecological niches

Using phylogenetic information to make ecological inferences required that we first test for phylogenetic signal in OTU niches. Drawing on established evolutionary theory (specifically Haldane, 1932; Wiedenbeck and Cohan, 2011), we predicted that a positive relationship between bacterial ecological differences and phylogenetic distances should only occur among close relatives. Our results and previous work in bacterial systems (Andersson et al., 2010) are consistent with this expectation: beyond ∼13–15% of the maximum phylogenetic distance observed between any two OTUs, there was no systematic relationship between ecological difference and phylogenetic distance (Figure 3). In addition, across close relatives, the rate at which ecological differences increased with phylogenetic distances was consistent with Brownian niche evolution. Across more distant relatives, however, there was effectively no relationship between ecological differences and phylogenetic distances. Ecological niches of bacteria have therefore evolved in a Brownian-like manner across recent evolutionary time, but in a phylogenetically random manner across deeper evolutionary time (see also Andersson et al., 2010; Diniz-Filho et al., 2010; Hardy et al., 2012). Kembel (2009) found that phylogenetic community composition provides robust inferences of ecological processes for systems like ours, where (i) phylogenetic signal is consistent with Brownian evolution; (ii) the environment is heterogeneous; and (iii) more than one aspect of organismal physiology determines OTU relative performance across environmental conditions, which we assume to be true. We therefore suggest that metrics quantifying phylogenetic distances among close relatives, such as NTI, provide information useful for understanding processes that assemble bacterial communities.

Processes governing local community composition

To infer the relative influences of stochastic and deterministic processes over local community composition we coupled the observation of phylogenetic signal in habitat associations with patterns of NTI. The term ‘local’ refers to a sample from one point in time and space, in contrast to turnover in community composition, which is discussed in the next section. Mean NTI across all communities was +2 and was significantly greater than zero. This provides good evidence for a dominant role of deterministic environmental filtering (Kembel, 2009). The key inference is that one or more environmental variables limit community membership whereby closely related and ecologically similar taxa coexist to a greater degree than expected if stochastic factors such as random birth-death events and ecological drift governed community composition (Kraft et al., 2007; Fine and Kembel, 2011; Pei et al., 2011). This inference is consistent with previous work on microbial communities in general and subsurface systems in particular (Horner-Devine and Bohannan, 2006; Bryant et al., 2008; Vamosi et al., 2009; Dumbrell et al., 2010; Herrera et al., 2010; Jones and Hallin, 2010; Amaral-Zettler et al., 2011).

Although environmental filtering is a dominant process on average, phylogenetic composition of some communities was statistically random and the relative influence of environmental filtering varied spatially. More specifically, 32% of communities had phylogenetic composition that did not differ significantly from the null expectation (−2<NTI<+2). Stochastic or ecologically neutral factors therefore have an important, but secondary role in the communities studied here. NTI distributions also differed significantly across well groups: well groups A, B and C had mean NTI values of +1.81, +1.84 and +2.36, respectively. In addition, 63%, 67% and 73% of individual communities were significantly phylogenetically clustered (NTI>+2) in groups A, B and C, respectively. These results suggest that the relative influence of environmental filtering over local community composition varies through space and is strongest in group C. In addition, there was generally no relationship between mean NTI and depth, sampling date or river elevation. These patterns collectively suggest it is environmental factors varying horizontally, as opposed to vertically or temporally, that lead to changes in the relative influence of environmental filtering. An important exception was the shallow depth within group A, where phylogenetic composition transitioned from random to significantly clustered with increasing river elevation and through time.

Why is environmental filtering the strongest in group C, and why does its influence increase with river elevation and through time only in the shallow depth of group A? One explanation is that these two locations represent opposite extremes along a gradient of environmental variation through time and space. River water intrusion is a potentially important environmental driver, and the degree of intrusion varies across wells and through time due to changes in river elevation (McKinley et al., in review). Of the three well groups examined here, group C is the furthest from the river and receives the least amount of river water during intrusion events (Lin et al., 2012; McKinley et al., in review). Relative to group C, the shallow depth within group A is environmentally variable through time and space: The shallow-A well receives more river water during intrusion events (Lin et al., 2012) and is more geologically heterogeneous. In terms of geology, the shallow-A well is partially screened within a ‘rip-up clast’ (Bjornstad et al., 2009), which is a meters-thick piece of fine-grained, relatively impermeable material related to the Ringold formation that otherwise underlies the more permeable Hanford formation (in which our sampling locations lie). Rip-up clasts are physically and chemically distinct ‘islands’ within the Hanford formation (Bjornstad et al., 2009). Their presence creates a local environment that is more structurally and potentially more chemically heterogeneous than the surrounding Hanford formation.

Owing to its distance from the river, group C may receive relatively small and infrequent river-water-supplied inputs of limiting resources such as organic carbon and nutrients. It may therefore represent a relatively extreme, low resource environment. Group C was also characterized by the strongest signal of deterministic processes governing local community composition, and previous work also shows greater determinism in extreme or low resource environments (Chase, 2007, 2010; Gerisch et al., 2011; Horner-Devine and Bohannan, 2006; Kembel et al., 2011; van der Plas et al., 2012). Stronger deterministic processes in more extreme environments may therefore be a general phenomenon across both micro- and macro-organism systems (see also Weiher et al., 2011).

At the other end of the ‘variation continuum’, unique microbial and geological aspects of the shallow-A well point to important interactions between river water intrusion and geological heterogeneity. In particular, intruding river water may significantly change the resource environment experienced by microbial communities, and this change may be especially large in conjunction with unique rip-up clast geochemistry. The lack of change in mean NTI with river elevation within group B, which received more river water than group A (Lin et al., 2012) but is not within a rip-up clast (Bjornstad et al., 2009), is consistent with this hypothesis. Additional characterization of the physical and chemical environment presented by rip-up clasts will be necessary to directly evaluate their influence over microbial communities. More generally, it appears that the relative influence of environmental filtering can increase toward environmental extremes (group C) and in response to shifting environmental conditions (shallow-A well).

Processes governing community dynamics

An important influence of environmental filtering over temporal community dynamics was indicated by a skewed βNTI distribution with a long tail extending beyond the +2 significance threshold. In addition, the median of this distribution was significantly greater than zero, suggesting that on average environmental filtering deterministically governs community dynamics (for a similar approach see Fine and Kembel, 2011; Kembel et al., 2011). It is interesting to note that across all comparisons 84% were consistent with random (−2<βNTI<+2) phylogenetic turnover. This is a much larger fraction than for NTI (32% consistent with random assembly). Together these results suggest that although, on average, deterministic factors govern composition within and turnover between communities, stochastic ecological drift has a stronger influence on turnover, relative to its influence on within community composition.

Uncovering the environmental factors causing non-random community dynamics is a key challenge, and our analyses provide some insight. Our analyses of phylogenetic community dynamics suggest that the relative influence of environmental filtering over community dynamics (i) is generally stronger in the shallow subsurface and (ii) can increase through time in response to environmental change in geologically heterogeneous regions. Specifically, the βNTI median was largest in the shallow sites across all three well groups (Table 1), where environmental variation was greatest (Lin et al., 2012). Higher levels of temporal environmental variation may therefore deterministically drive community dynamics, consistent with recent studies linking cyclical environmental conditions to cyclical patterns in microbial community composition (Fuhrman et al., 2006; Gilbert et al., 2012). In the shallow-A well βNTI increased to above the significance threshold of +2 with increasing differences in river elevation as the river rose from mid-May to early-July. In all other sites βNTI either did not vary with river elevation or did not increase to above +2, suggesting a unique environmental feature of the shallow-A well. As discussed above, the shallow-A well is indeed geologically unique in that it is partially within a rip-up clast (Bjornstad et al., 2009). The unique geology and βNTI patterns found within the shallow-A well point to an important interaction between intruding river water and geological heterogeneity. That is, river water intrusion into geologically heterogeneous regions of the subsurface may alter the abiotic environment in ways that impose strong directional selection (sensu Vellend, 2010) on microbial community composition.

The βNTI patterns found here complement and extend patterns found in Lin et al. (2012). Most importantly, Lin et al. uncovered large temporal changes in community composition, and found that changes in the relative abundance of specific taxa were especially marked in the shallow-A well (see their Supplementary Figure S6). These dynamics could be driven by deterministic processes associated with river water intrusion, but may also arise from random fluctuations in population sizes due to stochastic factors. The phylogenetic analyses carried out here provide an additional layer of support for the hypothesis that the river intrusion deterministically shifted community composition. However, this was true only for the shallow-A well, suggesting that in all other sites the deterministic influence of river water intrusion did not overwhelm the influence of stochastic factors. This result highlights the fact that deterministic and stochastic processes simultaneously influence community assembly, that their relative influences vary through space and time, and that using phylogenetic tools in microbial community ecology can provide insights not accessible via standard methods.

In summary, understanding gradients in the identity and strength of ecological processes is at the forefront of current ecological research (Chase and Myers, 2011), yet this paradigm has not been broadly embraced in microbial ecology. We argue that simultaneously addressing similar questions and concepts in both microbial and ‘macrobial’ systems will lead to a more general and unified theory of ecological communities. Here we work toward such unification by asking questions and applying statistical tools more common in studies of macro-organisms. Our results collectively suggest that the relative roles of deterministic and stochastic processes follow patterns that may generalize across systems. Further work with both micro- and macro-organisms aimed at understanding what drives gradients in the relative influences of stochastic and deterministic processes will be critical for ultimately predicting the response of communities to environmental change.