Genomic adaptations in information processing underpin trophic strategy in a whole-ecosystem nutrient enrichment experiment.

Several universal genomic traits affect trade-offs in the capacity, cost, and efficiency of the biochemical information processing that underpins metabolism and reproduction. We analyzed the role of these traits in mediating the responses of a planktonic microbial community to nutrient enrichment in an oligotrophic, phosphorus-deficient pond in Cuatro Ciénegas, Mexico. This is one of the first whole-ecosystem experiments to involve replicated metagenomic assessment. Mean bacterial genome size, GC content, total number of tRNA genes, total number of rRNA genes, and codon usage bias in ribosomal protein sequences were all higher in the fertilized treatment, as predicted on the basis of the assumption that oligotrophy favors lower information-processing costs whereas copiotrophy favors higher processing rates. Contrasting changes in trait variances also suggested differences between traits in mediating assembly under copiotrophic versus oligotrophic conditions. Trade-offs in information-processing traits are apparently sufficiently pronounced to play a role in community assembly because the major components of metabolism-information, energy, and nutrient requirements-are fine-tuned to an organism's growth and trophic strategy.

Introduction

Traits that influence the informational underpinnings of metabolism may be crucial to performance and community assembly but ecologists have largely focused on the proximal energetic and stoichiometric features of metabolism (Leal et al., 2017; Sibly et al., 2012; Sterner and Elser, 2002). Organisms must be able to store, copy, and translate the information contained in genetic material. And, they must be able to update their transcriptome and proteome adaptively in response to altered environmental conditions. For an organism to grow and reproduce rapidly, rates at every step of the metabolic network must be sufficiently high such that no single step is unduly rate-limiting, including the information processes that underpin biosynthesis and regulate metabolic networks. This necessary integration of functions is a hallmark of all organisms.

The structure and size of the genome affect the rate, efficiency, and robustness of the information processes that support metabolism, growth, and reproduction (Appendix 1). There are necessary tradeoffs in the costs and benefits of these features (Appendix 1; see also Smith, 1976), which should consequently make individual organisms more competitive in and better-suited to only particular ranges of growth and trophic conditions (Roller and Schmidt, 2015). Organisms that are best suited to compete in environments where resources are abundant (copiotrophs) must have the capacity for intracellular rates of information processing that are sufficiently high to support high rates of metabolism and reproduction. However, maintaining the genomic and structural capacity for rapid growth is costly, potentially placing copiotrophic taxa at a disadvantage in stable, nutrient-poor environments where growth rates are chronically slow (Giovannoni et al., 2014). Oligotrophic environments may thus instead favor organisms (oligotrophs), which have information processing machinery that is less costly to build, maintain, and operate, thereby increasing resource use efficiency and growth efficiency (Koch, 2001; Roller and Schmidt, 2015). Genomic traits that affect the rates and costs of biochemical information processing within cells can thus influence the degree to which an organism is optimized for oligotrophy versus copiotrophy.

This oligotrophic-copiotrophic strategy continuum is reminiscent of the classic slow-fast life history continuum (Stearns, 1992), of classical r/K selection theory (MacArthur and Wilson, 2001; Pianka, 1972; Pianka, 1970), and of their subsequent developments dealing with the evolution of interspecific variation in rates of resource use, mortality, growth, and reproduction (e.g., Dobson, 2012; Grime and Pierce, 2012; Krause et al., 2014; Sibly and Brown, 2007). In conjunction with research on the role of functional traits and niches in shaping the assembly of communities (e.g., Fukami et al., 2005; Litchman and Klausmeier, 2008; McGill et al., 2006; Okie and Brown, 2009; Roller and Schmidt, 2015), this work suggests that traits that are associated with the biological rates and efficiencies of resource use, growth, and reproduction play an important role in community assembly. It is unclear, however, whether the traits that are specifically related to rates and costs of biochemical information processing have sufficiently pronounced tradeoffs or physiological effects to play an important role in the evolutionary ecology of organisms and in the assembly of communities, although there are some promising indications (Dethlefsen, 2004; Roller et al., 2016).

Here, by coupling metagenomic analysis with a trait-based framework that synthesizes theory and hypotheses from genomics, ecology, and evolutionary cell biology, we investigate the role that several universal genomic traits play in determining the response of a planktonic microbial community to nutrient enrichment in a whole-ecosystem experiment. To date, relatively few studies have coupled metagenomics with a trait-based framework to clarify the drivers of community assembly (Burke et al., 2011; Chen et al., 2008; Mackelprang et al., 2011; Raes et al., 2011). Even fewer (indeed, none that we are aware of) have deployed such approaches in the context of whole-ecosystem experimentation to test ecologically relevant hypotheses under field conditions. We focus on a set of four information-processing traits that are hypothesized to affect the ability of organisms to obtain the high maximum growth rates necessary for thriving in copiotrophic environments, or their ability to reduce the energetic and resource requirements necessary to persist under nutrient-poor conditions:

Multiplicity of genes essential to protein biosynthesis. Copiotrophs are predicted to have higher copy numbers of rRNA operons and tRNA genes, because higher numbers of these genes increase their maximum overall transcription rates, helping maintain higher abundances of ribosomes (which are constructed from rRNA and proteins) and tRNAs. In turn, the larger pools of ribosomes and tRNAs facilitate the increased translation rates of protein synthesis necessary to achieve high growth rates (Condon et al., 1995; Higgs and Ran, 2008; Rocha, 2004; Roller et al., 2016). However, higher numbers of these genes incur costs, potentially putting organisms that have more copies of rRNA and tRNA genes at a disadvantage under low-resource conditions. In particular, more DNA has to be synthesized, maintained, and regulated, and there may be an increased risk of transcribing overly large phosphorus-rich pools of rRNA and tRNA, increasing phosphorus requirements (Elser et al., 2003; Elser et al., 1996; Godwin et al., 2017; Makino et al., 2003) and reducing growth efficiency (Roller et al., 2016).

Genome size. Organisms that have smaller genomes are predicted to do better in stable and oligotrophic environments, as they require fewer resources (such as phosphorus) to maintain and replicate their genomes, have higher carbon-use efficiency (Saifuddin et al., 2019), and have smaller cells with increased surface-area-to-volume ratios that facilitate resource uptake (Giovannoni et al., 2014). By contrast, organisms that have larger genomes should do better in complex or copiotrophic environments, where they can take advantage of their typically higher intrinsic growth rates (DeLong et al., 2006) and their more diverse and flexible gene and metabolic networks (e.g., Konstantinidis and Tiedje, 2004; Maslov et al., 2009; Saifuddin et al., 2019) to facilitate substrate catabolism and to respond more rapidly to feasts following famines.

GC content. Genomic GC content (the percentage of DNA composed of the nucleotide bases guanine (G) and cytosine (C)) varies greatly among taxa. The reasons for this variation are controversial, as multiple different selective and neutral forces may be operating (Bentley and Parkhill, 2004; Hildebrand et al., 2010). However, researchers have proposed that G and C have higher energy costs of production and more limited intracellular availability compared to A and T/U (Rocha and Danchin, 2002). In addition, DNA and RNA that have a higher GC content have more nitrogen (Bragg and Hyder, 2004). Thus, lower genomic GC content may be favored in oligotrophic environments, whereas the metabolic and resource-sparing benefits of low GC content should be less consequential in resource-rich environments, leading to a relaxed role for GC content in the ecology and evolution of copiotrophs.

Codon usage bias. Eighteen of life’s 20 proteinogenic amino acids can be encoded in the genome by more than one of life’s 61 different proteinogenic codons (nucleotide triplets), leading to redundancy in the genetic code. However, these synonymous codons have different kinetic properties, including different translation rates and probabilities of mistranslation (Higgs and Ran, 2008). In highly expressed genes that are essential to growth (such as genes encoding ribosomal proteins), there should be increased selection for biasing the usage of certain synonymous codons over others in order to increase the accuracy and speed of translation, especially in organisms that have fast growth rates (Hershberg and Petrov, 2008; Higgs and Ran, 2008; Plotkin and Kudla, 2011; Vieira-Silva and Rocha, 2010). We thus predict that copiotrophic environments should favor organisms that have higher codon usage bias in their ribosomal protein genes, whereas codon usage bias should play a relaxed role in oligotrophic environments.

Further details and background on these traits are provided in Appendix 1.

Research on the genomic traits described above has revealed correlations of these traits with growth and trophic strategy in a variety of eukaryotic and prokaryotic species (Vieira-Silva and Rocha, 2010) and started to unravel the mechanisms by which these traits influence fitness (Appendix 1), suggesting that they may play an important role in ecology (Freilich et al., 2009; Lauro et al., 2009; Weider et al., 2005). However, the work published to date has tended to look at only one or two of these traits at a time (Zeigler Allen et al., 2012; Raes et al., 2007). More work is also required to help resolve incongruities in the literature, such as different views on the evolutionary ecology of bacteria genome size (e.g., DeLong et al., 2010 versus Vieira-Silva and Rocha, 2010) (Appendix 1). It is also necessary to develop a less fragmented understanding of the evolutionary and physiological ecology of these genomic traits (Bentley and Parkhill, 2004; Freilich et al., 2009; Vieira-Silva and Rocha, 2010) and their role in community assembly across a wide range of organisms and environments.

Importantly, most ecological studies of these traits have been based on studies of microbial isolates, comparative analyses, or sampling across environmental gradients (Zeigler Allen et al., 2012; Foerstner et al., 2005; Freilich et al., 2009; Raes et al., 2007; Roller et al., 2013; Vieira-Silva and Rocha, 2010). Experimental work examining their role in mediating the structure of communities under natural conditions is extremely limited. Given (1) the complexity of the biotic and abiotic interactions that shape communities, (2) that most microbial taxa are uncultivable in isolation (Ho et al., 2017), and (3) that community traits can have widely different responses to temporal/experimental versus geographic variation in abiotic variables (e.g., Sandel et al., 2010), direct experimentation in the field with complex communities is required to better establish the validity of inferences about these purported molecular adaptations for ecological dynamics in nature.

Our study site is Lagunita, an oligotrophic, highly phosphorus-deficient pond in Cuatro Ciénegas, a biological reserve in Mexico (Lee et al., 2015; Lee et al., 2017). Because of its strong nutrient limitation, this ecosystem offers a useful setting for a fertilization experiment to evaluate the role of information-processing traits in community assembly and in the trophic strategies of organisms. Our study is noteworthy as one of the first whole-ecosystem experiments to involve experiment-level replicated metagenomic assessments of community response. If it is true that individual genomic features do indeed affect the ability of organisms to survive and reproduce as a function of nutrient availability, then collective measures of these features at the metagenomic level (which aggregates the genomes of all of the individuals constituting a community) should likewise exhibit these characteristics and should reflect the organisms' responses to experimental fertilization (Krause et al., 2014; Wallenstein and Hall, 2012).

Results

Biomass and chlorophyll a concentrations increased substantially in response to nutrient enrichment (biomass—198% mean increase, p=0.009; chlorophyll a—831% mean increase, p=0.001; Appendix 1—figure 1), as did the ratio of phosphorus to carbon (P:C) in seston biomass (19.5% mean increase, p=0.014, Figure 1A). We observed changes in several components of the predicted genomic signatures of growth and trophic strategy (Figure 1). As the percentages of bacteria, Archaea, Eukarya, and viruses making up the community were not discernibly different between unfertilized and fertilized treatments (p=0.46, 0.37, 0.39, 0.21, respectively), these genomic changes reflected changes in the abundances of taxa at finer phylogenetic scales, in particular of bacterial taxa, which made up 94% of the metagenomes (see Appendix 1 and Appendix 1—figures 1, 2, 3, 4 for details). Furthermore, the genomic changes reflected widespread changes within the community—they were not driven by just a few specific populations, as 188 genera showed changes in abundance (with p<0.05) and no single genus dominated the community (the highest relative abundance of a taxon at the genus level was 4%). Consistent with predictions, the mean estimated genome size of bacteria was 25% higher in the fertilized treatment (p=0.011), with nutrient enrichment explaining 75% of the variation between samples in mean genome size (Figure 1). The GC content of open reading frames of DNA was 9.9% higher in the fertilized treatment: 54% compared to 49% in the unfertilized treatment (p=0.007, R2 = 86%).

Appendix 1—figure 1.

Chlorophyll a concentration increased with nutrient enrichment, whereas it remained relatively invariant in the unfertilized treatment.

Figure 1.

Community-level trait responses to nutrient enrichment.

(A–F) As predicted, the mean value for each information-processing trait was higher in the fertilized treatment (Fert) than in the unfertilized treatment (Con). Because the faster-growing organisms generally have more P-rich ribosomes, seston P:C ratio also increased after fertilizer treatment (as shown in panel [A]). Boxes show 25/75% quantiles, center horizontal lines show medians, and vertical lines show the data range. In panel (E), a community’s codon usage bias is the inverse mean effective number of codons (1/ENC) of a metagenome’s ribosomal protein sequences, with higher 1/ENC values indicating increased codon usage bias.

Appendix 1—figure 2.

Principal coordinates analysis (PCoA) of the community phylogenetic structure inferred from the metagenomes shows that the phylogenetic composition of samples, as indicated by each point, is substantially different between unfertilized and fertilized treatments.

Microbial community phylogenetic composition also varied notably within the unfertilized mesocosms, falling into two clusters driven by variation in the relative abundance of Alphaproteobacterum, whereas enriched communities all shared relatively similar phylogenetic composition, indicating a convergence of effects of fertilization on community composition.

Appendix 1—figure 3.

Relative abundances of taxonomic phyla in the samples from the unfertilized (Cont) and fertilized (Fert) treatments.

Appendix 1—figure 4.

The significant relative abundance changes in taxonomic classes suggested by the statistical analysis of the phylogenetic composition determined by APIS.

The statistical analysis is based on the statistical methods developed in the EdgeR package (Robinson et al., 2010). Note that less than 50% of reads could be annotated. All p<0.01.

Genomic features that are indicative of adaptations for maintaining high rates of transcription and translation were also positively associated with fertilization. The per sequence occurrence rate of tRNA genes and the total number of tRNA genes per community were 93% and 64% higher, respectively, in the fertilized treatment than in the unfertilized treatment (p<0.001 and p=0.065 with R2 = 53%, respectively). The residuals after regressing the log number of tRNA genes and the log total number of reads per sample to control for differences in sequencing depth were also higher in the fertilized treatment than in the unfertilized treatment (p=0.087, R2 = 52%). Likewise, in the fertilized treatment, the per sequence occurrence rate of 16S rRNA genes and total number of rRNA operons per community were 119% and 86% higher, respectively (p<0.001 and p=0.096 with R2 = 50%, respectively). The residuals after regressing log number of rRNA operons versus log total number of reads per sample were also higher (p=0.038, R2 = 52%). Fertilization explained 65% of the co-variation in these two traits (number of rRNA operons and tRNA genes) along a single dimension, which was quantified by principal component analysis and provides a measure of protein synthesis capacity (p=0.031).

Consistent with predictions, nutrient enrichment also increased codon usage bias in ribosomal protein genes according to two measures of codon usage bias—the effective number of codons (ENC) and ENC′ (Figure 1E and Figure 2). Values of ENC and ENC′ for genes vary inversely with the level of codon usage bias: from 20, which signifies extreme codon usage bias in which one codon is used exclusively for each of the amino acids encoded by a gene, to 61, which represents the case in which the use of alternative synonymous codons is equally likely (no codon usage bias). The mean ENC and ENC′ of ribosomal protein sequences detected in the metagenomes decreased with fertilization by 6.7% and 4.8%, respectively, indicating increased codon usage bias (ENC: p=0.018, R2 = 66%; ENC′: p=0.031, R2 = 55%). Median ENC and ENC′ were also lower in the fertilized treatment than in the unfertilized treatment (p=0.006, R2 = 75% and p=0.010; R2 = 75%), and Kolmogorov-Smirnoff tests indicated that the distributions of ENC and ENC′ in ribosomal protein sequences significantly differed between treatments (all p<0.01). In particular, as shown in Figure 2A and B, a greater frequency of sequences exhibited little-to-no codon usage bias (ENC and ENC′ values around 60–61) in the unfertilized communities (Figure 2), whereas in the fertilized treatment, there were notably higher numbers of ribosomal protein sequences with extremely high codon usage bias (ENC and ENC′ <~37).

Figure 2.

Histograms and cumulative distributions of the codon usage biases (ENC′ values) of the ribosomal protein sequences in the metagenomes, with lower ENC′ values indicating increased codon usage bias and thus increased speeds and/or accuracies of translation of ribosomes.

(A) Histograms showing that the unfertilized communities have more sequences with nearly no detectable codon usage bias (values of 60–61), higher median ENC′ values (indicated by the tall vertical lines), and typically fewer sequences with high codon usage bias (ENC′ values from 20 to 40). Bar colors and their overlapping shades indicate different samples. (B) Cumulative distributions showing that, overall, the fertilized communities’ distributions have fatter left-tails, so higher frequencies of ENC′ values < 50 and thus more frequent codon usage bias. The Y-axis shows the proportion of ENC′ values ≤ ENC′ value designated by the plotted curves (fertilized = red curves, unfertilized = black curves).

As suggested by Figures 1 and 2, the variance in the P:C ratio and in the mean and median codon usage bias of communities (quantified by ENC′, the more powerful indicator of codon usage bias) substantially decreased with fertilization by around an order of magnitude or more—by factors of 8, 31, and 27, respectively (P:C ratio–p=0.058; mean ENC′–p=0.056; median ENC′–p=0.069; Appendix 1—table 1). Variance in genome size and mean and median ENC also decreased substantially by factors of 10, 10, and 6, respectively, but these responses are less certain (genome size–p=0.174; mean ENC–p=0.169; median ENC–p=0.275; Appendix 1—table 1). By contrast, variance in the log-transformed number of tRNA genes and rRNA operons substantially increased with fertilization, by 846% and 655%, respectively (rRNA–p<0.001; tRNA–p=0.013), whereas GC content variance did not appear to exhibit any change (p=0.56; Appendix 1—table 1).

Appendix 1—table 1.

Differences in the variance of community-level traits between treatments.

Community-level trait	Variance		Ratio of variances	Bonett's test p-value
	Unfertilized	Fertilized
P:C ratio	2.09 × 10−7	2.06 × 10−8	0.13	0.058
Mean genome size	105,315.4	10,984.8	0.1	0.174
Log number of rRNA genes	0.0056	0.0422	7.55	<0.001
Log number of tRNA genes	0.0148	0.1396	9.46	0.013
GC content	0.0001	0.0002	1.9	0.56
Mean ENC	3.0554	0.3034	0.1	0.169
Median ENC	2.5861	0.413	0.16	0.275
Mean ENC'	2.5494	0.0824	0.03	0.056
Median ENC'	1.832	0.0671	0.04	0.069

Finally, we examined how well nutrient enrichment predicted the covariation of these genomic traits along a single principal component analysis (PCA) axis, which, according to our trait predictions, should quantify where a community’s information-processing traits fall along an oligotrophy-copiotrophy strategy continuum. The single PCA axis captured 78% of the variance in the genomic traits for the communities, and nutrient treatment explained 86% of the variation in community genomic trait composition along this single axis of trophic strategy (general linear model [GLM]–p=0.004; Figure 3).

Figure 3.

The principal component scores quantifying the position of metagenomes along a single dimension representing the oligotrophy-copiotrophy strategy continuum for information-processing traits.

Higher scores on the vertical axis indicate communities whose information-processing traits are better suited for copiotrophy (increased maximum growth rate in resource abundant environments), whereas low scores indicate adaptation to oligotrophy (increased resource use efficiency in low nutrient environments). Nutrient enrichment explains 86% of the variance in metagenomes along this axis of growth and trophic strategy (R2 = 86%, p=0.004). Within each treatment, circular symbols are spaced horizontally for visual clarity.

Discussion

Our fertilization study demonstrated strong nutrient limitation in the Lagunita ecosystem, as manifested by increased biomass, chlorophyll content, and P:C ratios of plankton in the fertilized pond versus the unenriched mesocosms. Using metagenomics, we found strong differences in information-processing traits between the fertilized pond and unfertilized internal mesocosms, which agree with all five of our directional predictions. We are not aware of any obvious reasons or existing theoretical work to indicate that a difference in habitat size between treatments should lead to the observed trait differences. We thus interpret the observed genomic trait differences as resulting primarily from differences in the growth and nutrient conditions of the two treatments. However, a valuable future experiment could be to repeat the experiment using internal mesocosms for both the fertilized and unfertilized treatments in order to help to rule out potential mesocosm effects. Conservationists and ecologists should also have an involved discourse on whether it is worth conducting future research to verify our interpretation by performing much larger experiments across multiple ponds (with wider environmental impacts).

We interpret these differences as primarily reflecting trait-mediated ecological dynamics—the differential success of lineages present within the pond and colonization by other lineages—rather than microevolutionary changes within populations. Our reasoning is that the relatively short time period of our experiment (32 days) encompasses a relatively low maximum number of generations of replication (Appendix 1), and so limited the opportunity for measurable genomic evolution to occur (e.g., Lenski and Travisano, 1994). Overall, our study suggests that ecologically significant phenotypic information about communities can be uncovered using a read-based approach (metagenomics) that leverages the relevant abundance and DNA characteristics of conserved genetic elements. Importantly, this approach avoids the biases associated with the massive gaps that are present in microbial taxonomic databases and cross-taxonomy assembly efficiencies (Kunin et al., 2010; Sogin et al., 2006; Temperton and Giovannoni, 2012; Yooseph et al., 2010).

An oligotrophy-copiotrophy gradient in information-processing traits

Remarkably, nutrient enrichment explained 50% or more of the variation in each trait and 88% of the co-variation of these five traits along a single statistical dimension quantifying the oligotrophy-copiotrophy strategy continuum (Figure 3). The congruence of all trait responses with our predictions suggests that the effects of these genomics traits on the rates and costs of biochemical information-processing are sufficiently pronounced to play a role in community assembly.

We found that the fertilized community had genomic traits that were expected to augment translational capacity, whereas the unfertilized communities had genomic traits that lower the costs of biochemical information processing. For instance, the fertilized community had increased codon usage bias in ribosomal protein genes, which can improve the speed and accuracy of translation of ribosomal proteins by increasing the rate at which tRNAs bind to mRNA codons of ribosomal protein genes, as well as by possibly reducing the likelihood of mistranslation (Higgs and Ran, 2008; Figures 1E and 2). The resulting increase in the rate of production of ribosomes supports the larger ribosomal pools undergirding higher overall rates of protein synthesis. The fertilized communities also had higher numbers of rRNA and tRNA genes. In conjunction with the observed increase in the biomass P:C ratio, this finding supports the ‘Growth Rate Hypothesis’—that fast-growing cells should have higher P-contents that reflect higher concentrations of P-rich ribosomal RNA, which are maintained by increased copies of rRNA and tRNA genes (Elser et al., 2000).

Mean genome size was also higher in the fertilized pond, reflecting the need to encode a larger array of genes that are needed to support the expanded translational and catabolic capacity of faster growing cells, and possibly also reflecting the benefits of maintaining a streamlined genome that requires minimal resources for maintenance and replication in nutrient-poor conditions (e.g., Giovannoni et al., 2005; Figure 1B). Consistent with our prediction that low GC-content in oligotrophic taxa is a valuable resource conservation strategy, GC content was lower in the unfertilized communities (Figure 1F).

These results contrast with those of Vieira-Silva and Rocha (2010), who found no significant interspecific correlation of GC content and genome size with generation time in a comparative study of bacteria (Vieira-Silva et al., 2010). The difference in results may reflect multiple factors, including: (1) the presence of confounding variables and pronounced statistical error (measurement and biological) inherent to interspecific microbial growth and trait databases that are collated from a variety of sources (e.g., ecosystems, conditions such as temperature), as in their study; (2) an underrepresentation in databases of the unculturable prokaryotic species that may comprise the majority of species in microbiomes, such as Lagunitas, and which typically have different evolutionary ecologies than culturable taxa (Pande and Kost, 2017; Swan et al., 2013). Such differences in results highlight the value of conducting in situ experiments.

Overall, however, our findings are largely consistent with observational studies of metagenomes and genomes along productivity gradients, and with comparisons across species varying in growth rate (Zeigler Allen et al., 2012; DeLong et al., 2010; Foerstner et al., 2005; Lauro et al., 2009; Raes et al., 2007; Roller et al., 2013; Swan et al., 2013). Our experiment on communities in situ suggests that much of the variation in these genomic traits along productivity gradients or across species may similarly be attributed to effects of information-processing costs and rates on an oligotrophy-copiotrophy strategy continuum and related adaptive strategies, such as r/K selection and Grime’s ‘C-S-R Triangle’ (e.g., Grime and Pierce, 2012).

Variance changes in the genomic traits

We also observed changes in the variances of several community-level traits (Appendix 1—table 1). The direction of the changes that occur in response to enrichment differed among the traits, so the variance responses do not simply reflect potential for increased dispersal limitation in unfertilized mesocosms as compared to enriched pond samples (see 'Materials and methods'). Instead, all-else-being equal, the responses suggest that the strength of a trait’s role in filtering the community differs between unenriched and enriched treatments, which could reflect differences in the degree to which a trait affects the costs versus the rates of information processing. In general, traits that have much more of an impact on the costs of information processing should play more of a role in oligotrophic environments, whereas traits that have more of an impact on the rates of information processing (e.g., probably codon usage bias) should play more of a role in copiotrophic environments.

The substantial reduction in variance levels in codon usage bias in response to fertilization are in agreement with this viewpoint: the evolution of low codon usage bias is primarily thought to reflect relaxed selection rather than positive selection for neutral (random) codon usage, although many complex issues remain to be resolved (Hershberg and Petrov, 2008). By contrast, the highly increased variance in rRNA operon and tRNA gene numbers in response to fertilization may indicate that the augmented P and N requirements of taxa that have high rRNA and tRNA gene copy numbers, which tend to have larger pools of ribosomes and other translation machinery, are particularly detrimental in oligotrophic conditions (Stevenson and Schmidt, 2004). In nutrient-rich conditions, on the other hand, some organisms may still do well even with fairly low numbers of these genes, for example by having multiple genome copies (e.g., Mendell et al., 2008) or other mechanisms that increase the concentrations and kinetics of RNA polymerase, aminoacyl-tRNA synthetase, rRNA and tRNA (e.g., Yadavalli and Ibba, 2012).

We also observed a 10-fold decrease in variance of mean genome size in the unfertilized mesocosms, implying that genome size has a more variable ecological role in oligotrophic ecosystems, although this change was only suggestive (p=0.17). On the basis of genome streamlining theory alone, which suggests that small genomes should be favored in oligotrophic conditions because of the favoring of cellular architectures that minimize resource requirements (Giovannoni et al., 2014; Morris et al., 2012), we would expect reduced variance in genome size in oligotrophic environments, in contradiction with our results. Consideration of metabolic scaling may provide the explanation. There is presumably an interspecific increase in active mass-specific metabolic rate and rmax with genome size in bacteria (DeLong et al., 2010), so under abruptly enriched conditions, community members that have larger genomes may displace the complex and diverse original community by growing faster. By contrast, it appears that the effects of genome size on metabolic costs may be insufficient to play as strong a role in the evolutionary ecology and assembly of these oligotrophic communities. This suggestion agrees with others who have argued that the elemental costs of DNA play a negligible role in the evolutionary ecology of genome size (Lynch, 2006; Mira et al., 2001; Sterner and Elser, 2002; Vieira-Silva et al., 2010)—the smaller genomes found in many parasites, symbionts, and oligotrophs may instead reflect the mutational loss of genes as the result of relaxed positive selection for keeping non-essential genes (Mira et al., 2001).

Unlike the other information-processing traits, we observed no change in variance of GC content between treatments. Although we expected environmental filtering for low GC content in oligotrophic conditions, it is less clear why there may be filtering for high GC content in response to nutrient enrichment, pointing to a need for further investigations into the implications of GC content for microbial physiological and community ecology.

Concluding remarks

The genomic traits studied here affect the costs and rates of biochemical information processing within cells, and all of these traits responded as predicted to nutrient enrichment. Cells proliferating under nutrient-enriched conditions had increased capacities for transforming and storing information, whereas those persisting in oligotrophic conditions had genomic traits associated with reduced costs for the information processes that underpin metabolism and reproduction. Optimizing tradeoffs in the efficiency and capacity of information processing may thus play a vital role in the evolutionary specialization of a microbe’s cellular biology to the particular trophic conditions of an ecosystem. Information processing traits should be included in the development of trait-based theories and frameworks for microbial community ecology, as apparently all three core components of metabolism—information, energy, and material requirements and transformations—must be closely fine-tuned to the growth and trophic strategy of a microorganism.

Materials and methods

Study site description

The whole-ecosystem fertilization experiment took place in Lagunita, a shallow (<0.33 m) pond averaging 35,000 L in volume and roughly ~12 m x 4 m. It is adjacent to a larger lagoon (Laguna Intermedia from the Churince system) in the Cuatro Ciénegas basin (CCB), an enclosed evaporitic valley in the Chihuahuan desert, Mexico. Despite its aridity, the CCB harbors a variety of groundwater-fed springs, streams, and pools. Past research has also shown that these aquatic environments harbor a high diversity of unique microbiota (Souza et al., 2018; Souza et al., 2006), which have evolved under strong stoichiometric imbalance (high nitrogen (N):phosphorus (P) ratios) and prevalent ecosystem P limitation (Corman et al., 2016; Elser et al., 2005). Lagunita water is high in conductivity, dominated by Ca2+, SO42–, and CO32–, and has an average molar TN:TP ratio of 122 indicative of strong P limitation, as previously demonstrated in this system during a mesocosm experiment completed in 2011 (Lee et al., 2015; Lee et al., 2017). During the summer season, the pond shrinks substantially and the surface water temperature increases.

Experimental design

On 25 May 2012, prior to initiation of fertilization, five replicate enclosures were established in different parts of the pond; these were unenriched treatments that served as reference systems for comparison with the pond after enrichment. As in Lee et al. (2015) and Lee et al. (2017), each unenriched mesocosm consisted of a 40 cm diameter clear plastic tube enclosing around 41 L (based on an average depth of 0.33 m at the time at which the mesocosms were installed). This volume fluctuated slightly during the experiment and decreased very slightly towards the end of the experiment because of evaporation (which decreased the pond volume by 1.4%). The mesocosms were fully open to the atmosphere and sediments. Each mesocosm’s water column was gently mixed periodically during our regular sampling (described below). Thus, with exposure to both the air and the bottom sediments, the unenriched mesocosms were essentially cylindrical ‘cross-sections’ of the ecosystem. The 41 L volume is a typical size for an aquatic mesocosm (e.g., see review of 350 mesocosms by Petersen et al., 1999) and appropriate for microbial studies, encompassing in the order of 30 billion prokaryotic cells (estimated from our cell counts). See Appendix 1 for more details.

The fertilization procedure was based on a previous mesocosm experiment in Lagunita (Lee et al., 2015; Lee et al., 2017). Prior to the initiation of the experiment, a morphometric map of the pond was created, allowing us to estimate the pond's water volume and to adjust that volume estimate as water depth changed through the season. Based on the pond’s volume, we fertilized to increase the PO43– concentration in the water by 1 μM (as KH2PO4). We also added NH4NO3 in a 16:1 (molar) N:P ratio with the added P. The soluble reactive phosphorus (SRP) concentration of the pond was then measured every 3–4 days, after which we added sufficient KH2PO4 to bring the pond's in situ concentration back to 1 μM, along with the appropriate amount of NH4NO3 to achieve a 16:1 molar ratio. Fertilizer was added by mixing fertilizer solution with ~2 L pond water and broadcasting the mixture into all regions of the pond.

We thus performed a sustained whole-ecosystem fertilization treatment, with replicate internal unfertilized mesocosms serving as reference systems. Whole-ecosystem manipulation assures that any experimental responses are ecologically relevant, because the manipulated system encompasses the full scale and scope of ecosystem processes that might modulate that response (Carpenter, 1998). Such a whole-ecosystem approach can be especially powerful when coupled to appropriate reference systems. Although our internal unfertilized mesocosms were smaller than the surrounding fertilized pond, we consider them to be pertinent references for investigating the role of genomic traits in community assembly under differing nutrient conditions for several theoretical, empirical, practical, and ethical reasons. Although replicate whole ponds for comparison would be preferred, the availability of multiple ponds at Cuatro Ciénegas for such experimentation is extremely limited given the basin’s arid nature. Indeed, true replication of whole-ecosystem manipulations is very rarely achieved, at least in aquatic ecosystems. We thus followed the recommendations of Carpenter (1989) and Carpenter (1998), relying on the application of a strong experimental treatment at the ecosystem scale and informed by previous experimentation (Lee et al., 2017) together with replication of internal reference mesocosms to assess the impacts of nutrient fertilization. In this way, we maximized the ecological realism of our perturbation by applying it at the ecosystem scale, while retaining the ability to compare manipulated dynamics against a benchmark. We preferred a whole-ecosystem fertilization of the pond over fertilizing internal mesocosms within the pond because the smaller enclosures might have provided an artificial view of how the microbial community responds to a nutrient perturbation at the ecosystem scale. In addition, mesocosms cut off many sources of colonizing species (such as shore sediments/soils in ponds) that can contribute to community reorganization following a perturbation (e.g., Thibault and Brown, 2008). Since this study aimed to understand the role of genomic traits in the assembly of communities (not just the disassembly caused by extinctions), it was thus important to avoid inhibiting community responses driven by the colonization of species.

Given the enormous heterogeneity between communities and water chemistry from different sites within the area, as well as the ability of microbes to disperse between ponds, using internal reference systems rather than other whole ponds is arguably more informative as it avoids introducing confounding factors related to variation between ponds (such as contrasting microbial communities) and instead introduces just one potentially confounding factor, the difference in size between the unfertilized mesocosms and the fertilized pond. We thus consider our approach to be the scientifically appropriate one for this conservation area. The design is a natural first step from experiments in small, homogenous bottles or bags. Scaling such experiments across multiple ponds/lakes may be a future step for experimental metagenomic research but not a responsible current step for research in the ecologically sensitive Cuatro Ciénegas basin area.

Field monitoring, sampling, and routine water chemistry

Following initiation of fertilization, the pond and internal unfertilized mesocosms were sampled every four days to monitor basic biogeochemical and ecological responses (see Appendix 1 for water chemistry sampling details). At the end of the experiment (32 days), we sampled for metagenomics: five water samples from the pond itself (fertilized treatment) and one water sample from inside each of the five unfertilized internal mesocosms. It is worth noting that, given the substantial seasonal changes in temperatures and water chemistry, we think that comparing metagenomic data from the pond pre-fertilization to those from the fertilized pond 33 days later would be an inappropriate approach for gaining insight into the effects of fertilization. Thus, we focus on comparing post-fertilization metagenome data with temporally matched data from the unfertilized mesocosms.

The water inside the mesocosm was gently stirred with a dip net prior to sampling. Sampling involved submerging a 1 L polycarbonate beaker just under the surface of the water. Microbes in the water samples were filtered onto sterile GF/F filters (0.7 µm nominal pore size, Whatman, Piscataway, NJ, USA), frozen immediately in liquid nitrogen, and held at <80° C until laboratory DNA extraction, purification, and sequencing. Given the 0.70 µm pore size, extremely small prokaryotes were not part of our metagenomes and so our results do not apply to these picobacterioplankton. If anything, their inclusion would augment predicted community-level trait responses to fertilization, as picoplankton are slow-growers, tend to do poorly in nutrient-rich waters, have small genomes, and so are likely to decrease in abundance in the fertilized treatment. Routine water chemistry methods were used, as in Lee et al. (2015) and Lee et al. (2017).

DNA extraction, sequencing, annotation, and phylogenetics

DNA was extracted using the MO BIO PowerWater DNA Isolation kit with a slight modification (increasing volume of PW1 solution to 1.5 mL). DNA yield and quality were assessed by PicoGreen assay (Appendix 1) and prepared for sequencing on Illumina MiSeq with 12 samples per v2 2 × 250 bp sequencing run. Raw reads were trimmed of barcodes, quality filtered, and rarefied to 100,000 sequences per sample (Appendix 1). For the quality filtering, we used the standard Qscore of 25. Two samples from the fertilized treatment and one sample from the unfertilized treatment were left out of subsequent analyses because they had sequencing depths less than 20% of the rest of the samples (whose sequencing depth averaged 2.5 × 106 reads) and low-quality scores. Sequences were phylogenetically annotated using the Automated Phylogenetic Inference System (APIS) with default parameters, which is designed to optimize annotation accuracy (Zeigler Allen et al., 2012). PCoA was used to visualize a Bray-Curtis distance matrix of the APIS annotations using the R (R Development Core Team, 2011) package vegan (Oksanen et al., 2016). We also used the statistical package edgeR (Robinson et al., 2010) in R to identify the taxonomic groups at the genus, class, phylum, and domain levels that exhibited discernable changes in relative abundance and to calculate the p-value of these changes. edgeR is specifically designed for dealing with sequencing count data in which there are minimal levels of replication (Robinson et al., 2010).

Trait bioinformatics

We used Phylosift (Darling et al., 2014) to annotate non-sub-sampled libraries, allowing us to count the number of bacterial and archaeal 16S rRNA genes in each sample, which averaged 2607. Counting 16S rRNA genes provides a means of evaluating the copy number of rRNA operons, because 16S RNA genes in prokaryotes are typically transcribed as part of a rRNA operon. Furthermore, the typical situation, at least in the genomes of cultivated organisms, is that each bacterial rRNA operon has a single 16S rRNA gene (Grigoriev et al., 2012). The program tRNAscan-SE v1.4 (Lowe and Eddy, 1997), which is specifically designed for recognizing tRNA genes, was used with the provided general tRNA model in order to count the number of tRNA genes in non-sub-sampled libraries. Variation in the total number of tRNA genes indicates variation in gene copy numbers, as the number of tRNA genes per genome is driven by variation in tRNA gene copy number rather than by tRNA diversity (Higgs and Ran, 2008). In 'Statistics', we describe our method for ensuring that results were not sensitive to variation in DNA sequencing depth. Although the tRNAscan-SE approach that we employed did not distinguish between Bacteria, Archaea and Eukaryotes, the observed rarity of Eukarya and insignificant changes in the groups' relative abundances between treatments (see 'Results') indicate that the tRNAscan-SE results, and data from our other bioinformatic analyses, reflect variation in prokaryotes, namely Bacteria (Archaea are very rare in CCB samples; Lee et al., 2017), rather than in Eukaryotes.

Bacterial genome sizes were estimated according to methods in Zeigler Allen et al. (2012) (Appendix 1). Briefly, length normalized core marker gene counts that were identified as bacterial by APIS were used to determine the number of genome equivalents in a sample. The total number of predicted proteins annotated as bacterial by APIS was then divided by the number of genome equivalents.

To quantify the degree of synonymous codon usage bias for each observed bacterial and archaeal ribosomal protein gene sequence, we first used Phylosfit (Darling et al., 2014) to identify the sequences and then used the program ENCprime to calculate two commonly used metrics, the effective number of codons (ENC) (Wright, 1990) and a related measure, ENC′ (Novembre, 2002), for each sequence. ENC and ENC are relatively statistically well-behaved and insensitive to short gene lengths compared to other measures of codon usage bias (Novembre, 2002; Wright, 1990). ENC′ also accounts for departures in background nucleotide composition from a uniform distribution and is therefore considered to provide a more powerful and reliable measure of codon usage bias (Novembre, 2002). Background nucleotide composition for each sample was considered to be the average nucleotide frequencies of all of the samples’ reads.

Statistics

We used t-tests assuming unequal variances to evaluate mean differences between treatments. For community-level genomic traits, these tests were one-tailed with the alternative hypothesis based on the predicted differences described in the 'Introduction'. General Linear Models (GLM) with treatment (unfertilized vs fertilized) as a fixed factor were used to determine the amount of variation (R2) in traits that was explained by treatment. Numbers of tRNA genes and rRNA operons were log-transformed before these analyses to achieve normality. In order to account for potential effects of sequencing depth on the number of rRNA operons and tRNA genes, we also regressed the numbers against a sample’s log total number of reads and then performed GLM analyses on the residuals (ascertaining whether or not, for a given sequencing depth, samples from the fertilized treatment have higher numbers of these genes, that is, higher residuals).

We also employed single-tailed Poisson rate tests to examine differences in the numbers of rRNA operons and tRNA genes. The Poisson rate test is more powerful than the t-test for examining differences in the per sequence rate of occurrences of tRNA genes and rRNA operons between treatments. In the Poisson rate test, the sample size for each treatment was the number of DNA samples and the treatment’s observation length was the mean number of reads per metagenome. Two-sided Bonett’s tests were used to test for treatment effects on variance between samples of community-level traits. Kolmogorov-Smirnoff tests were conducted to evaluate whether or not within-sample distributions of the ENC and ENC′ values of ribosomal protein gene sequences differ between treatments.

Finally, we examined how well nutrient enrichment predicted the covariation (correlation) of these genomic traits along a single axis quantified by principal component analysis (PCA). Should nutrient enrichment explain substantial variation in the communities along this single axis, then the PCA values would provide a measure of molecular adaptiveness to oligotrophic versus copiotrophic conditions. We used median ENC′ as the measure of codon usage bias in the PCA. We calculated the principal component score of each metagenomic sample along the first dimension and used GLM analysis to determine how well fertilization explained variation in the communities’ scores along this dimension. Before performing PCA, in order to give equal weight to each trait, variables were first standardized (z-scored) by subtracting means and dividing by standard deviations. Overall, we aimed to avoid overreliance on significance levels and p-values in judging scientific results (Carver, 1993; Halsey et al., 2015; McGill et al., 2006; Rothman, 2016), so we report p-values and effect sizes and let readers judge the significance of the results for themselves.

We focused on the community response of genomic traits to varying nutrient conditions, rather than on a detailed natural history of the phylogenetic composition of the community, not only because we are interested in interrogating metagenomic changes within a trait-based framework, but also for pragmatic reasons. There are multiple bioinformatic challenges to resolving precisely the phylogenetic composition of entire microbial prokaryotic communities and ensuring that the phylogenetic biases of various molecular methods do not differ between environments or growth conditions (Kunin et al., 2010; Sunagawa et al., 2013). For instance, there are massive gaps in prokaryotic taxonomic databases (Rinke et al., 2013; Temperton and Giovannoni, 2012), and metagenomes generated from low and high growth communities in oceans have different levels of taxonomy blindness (Kalenitchenko et al., 2018; Sogin et al., 2006; Yooseph et al., 2010). Also, DNA sequence assembly introduces a bias, as clonal populations with even low coverage assemble very well whereas high abundance populations with strain diversity will not assemble well.

In the interests of transparency, eLife publishes the most substantive revision requests and the accompanying author responses.

Acceptance summary:

All living things store the rules that shape their lives in a genetic code that makes up their genome. These genomes differ from species to species in ways such as genome size, GC content, or ability to multiply P-rich rRNAs, among others. The full expression of these genome 'traits' requires energy and block-building materials, and therefore they can determine the presence of a species in a given site. The authors describe the results of an experiment in which nutrient-poor springs (oligotrophic conditions) were experimentally NP fertilized (creating a copiotrophic environment). The response of microbial communities in these lakes, assayed using state-of-the-art metagenomic methods, conformed to predictions. This work is unique in that it merges ideas gathered from trait-based macroorganism ecology to microbial communities in whole-lake ecosystems. By doing so, the authors make use of, rather for the first time, fundamental genome-wide traits as a determinant of species composition.

Decision letter after peer review:

Thank you for submitting your article "Genomic adaptations in information processing underpin trophic strategy in a whole-ecosystem nutrient experiment" for consideration by eLife. Your article has been reviewed by two peer reviewers, and the evaluation has been overseen by a Reviewing Editor and Detlef Weigel as the Senior Editor. The reviewers have opted to remain anonymous.

The reviewers have discussed the reviews with one another and the Reviewing Editor has drafted this decision to help you prepare a revised submission.

Summary:

Okie et al. describe the results of an experiment in which nutrient poor springs were experimentally supplemented with a limiting nutrient. The response of microbial communities was assayed using metagenomic methods. In particular, microbial genomic traits were inferred from the metagenomic data, and related to organismal life history. The study has an overarching goal of linking 'trait-based' approaches that have been highly useful in non-microbial ecology to a meta-genomic investigation, which is an important approach to the study of microbial communities. The sampling and field experiment appear to reviewers to be technically sound, allowing for practical constraints that are described and justified. The data processing workflows follow standard practices up to a point – some analyses address novel questions, for which there are no established 'best practice' workflows, but the analytical choices seem reasonable. In sum, the manuscript by Okie et al. could be an interesting and worthwhile contribution to the literature, as it explores novel and interesting ideas, with a relatively strong dataset.

Essential revisions:

1) Is it possible the results are substantially affected by a 'bloom' in one or a few lineages in the P supplemented environments, and if so, does this matter? A hypothetical example might be a substantial increase in the abundance of a particular type of α-proteobacterium, with genome size and GC content (etc) greater than the community average. This kind of possibility (increases in prevalence of one or few particular 'species') would seem to accord quite well with the observed decrease in trait variance. Note, if the observed results were driven partially by growth responses of a few organisms, it would not affect the interpretation of the results too much. It might slightly increase the possibility that the outcomes are attributable to particular organisms, rather than a general response of organisms with certain traits responding in a way that holds generally. This would also affect the notion that decreased variance in a trait is an indication of how much that trait structures the community (e.g. subsection “Variance changes in the genomic traits”, first paragraph).

2) It's often difficult to replicate whole ecosystem experiments or to find appropriate controls. In this case, the authors used enclosures in the lake that were not fertilized. They compared samples from enclosures to the rest of the lake, which were not enriched. The authors pre-emptively try to address issues that might be raised by readers, and conclude that the only really difference is size (i.e., volume) between the non-fertilized and fertilized treatments. I wonder why the authors didn't use enclosures for both the control (non-fertilized) and enriched (fertilized) treatments. The authors argue that whole ecosystem experiments are important because mesocosms do not necessarily reflect observations that are relevant to the whole ecosystem. While this is a defensible argument it is somewhat diminished by the fact that their experiment inherently relies on mesocosms. In fact, one has to believe that the mesocosms are an accurate representation of the whole lake if the results are to be taken at face value. Also, reviewers were not sure the paper makes a compelling case that a whole-ecosystem approach was required to test hypotheses about how genomic attributes change in response to nutrient enrichment. In the end, reviewers were sympathetic to the constraints and applaud the ambition to tackle such questions at the ecosystem scale. Nevertheless, it is hard to completely overlook some of the limitations of the design. Please make sure these and other caveats are properly addressed in the Discussion section.

3) Reviewers were motivated by your questions and therefore were excited to read the manuscript. They were a bit less enthusiastic about how the paper came together in terms of organization and narrative. One of them said “the paper was at least a few drafts away from being "polished", It felt as though the manuscript would have benefited from some additional editing and revision”. So, please make sure the theory and arguments are be tightened up a bit, which will likely improve the flow of the paper.

Essential revisions:

1) Is it possible the results are substantially affected by a 'bloom' in one or a few lineages in the P supplemented environments, and if so, does this matter? A hypothetical example might be a substantial increase in the abundance of a particular type of α-proteobacterium, with genome size and GC content (etc) greater than the community average. This kind of possibility (increases in prevalence of one or few particular 'species') would seem to accord quite well with the observed decrease in trait variance. Note, if the observed results were driven partially by growth responses of a few organisms, it would not affect the interpretation of the results too much. It might slightly increase the possibility that the outcomes are attributable to particular organisms, rather than a general response of organisms with certain traits responding in a way that holds generally. This would also affect the notion that decreased variance in a trait is an indication of how much that trait structures the community (e.g. subsection “Variance changes in the genomic traits”, first paragraph).

We found that a variety of taxa changed in abundance in response to fertilization (see Appendix 1—Figures 3 and 4 and newly added text in Appendix 1, subsection “Results – Changes in Taxonomic Composition”), so our results were clearly not driven by a bloom of just a few particular species. We added statistics and text towards the beginning of the Results to strengthen our answer to these points.

Note that some traits showed an increase in variance with fertilization, so we believe it is very unlikely that our variance results could be explained by just a few species responses.

It is also remarkable that some genomic traits, such as GC content, are not strongly phylogenetically conserved. For example, in α-proteobacterium, GC content varies from less than 30% to greater than 60% (Hildebrand et al., 2010). Thus, even if it were the case that just one phyla responded to nutrient enrichment (which it is important to stress was not the case), the question of which species within the phyla responded and the role of their traits in governing these responses would still be a relevant question.

2) It's often difficult to replicate whole ecosystem experiments or to find appropriate controls. In this case, the authors used enclosures in the lake that were not fertilized. They compared samples from enclosures to the rest of the lake, which were not enriched. The authors pre-emptively try to address issues that might be raised by readers, and conclude that the only really difference is size (i.e., volume) between the non-fertilized and fertilized treatments. I wonder why the authors didn't use enclosures for both the control (non-fertilized) and enriched (fertilized) treatments. The authors argue that whole ecosystem experiments are important because mesocosms do not necessarily reflect observations that are relevant to the whole ecosystem. While this is a defensible argument it is somewhat diminished by the fact that their experiment inherently relies on mesocosms. In fact, one has to believe that the mesocosms are an accurate representation of the whole lake if the results are to be taken at face value. Also, reviewers were not sure the paper makes a compelling case that a whole-ecosystem approach was required to test hypotheses about how genomic attributes change in response to nutrient enrichment. In the end, reviewers were sympathetic to the constraints and applaud the ambition to tackle such questions at the ecosystem scale. Nevertheless, it is hard to completely overlook some of the limitations of the design. Please make sure these and other caveats are properly addressed in the Discussion section.

Thank you for thoughtfully drawing attention to these complex issues. In this new version of the manuscript. we made additions towards the beginning of the Discussion to further address these questions and added further explanation in the Materials and methods that also address your concerns.

Our main idea and logic is that there is an asymmetry in the usefulness of internal fertilized and unfertilized mesocosms. Perturbation of a community (here the planktonic microbial community) usually causes local extinctions, opening up resources and niches for other species to colonize the local communities. From other studies in the same system we know that the pool of species that may potentially be an important source of colonizers in Lagunita pond water is the community of microbes living on the shores of the pond and in the sediments. Unlike a whole-ecosystem nutrient perturbation experiment, fertilizing a mesocosm would not sufficiently replicate many of these ecological and colonization processes that follow a nutrient perturbation – a mesocosm has no natural shore, and in the mesocosm there is decreased mixing between water column and sediments, reducing the maximum rate at which microbes from the sediment can colonize the water column. Since this study aims to understand the role of traits in the assembly of communities (not just the disassembly), it was important that potential colonizers not be artificially cut-off from the fertilized treatment and that is one of the main reasons for our design. We provided a summary of this logic in the Materials and methods, in addition to the other experimental design considerations explained in the Materials and methods.

3) Reviewers were motivated by your questions and therefore were excited to read the manuscript. They were a bit less enthusiastic about how the paper came together in terms of organization and narrative. One of them said “the paper was at least a few drafts away from being "polished", It felt as though the manuscript would have benefited from some additional editing and revision”. So, please make sure the theory and arguments are be tightened up a bit, which will likely improve the flow of the paper.

Thank you for the concerns and efforts to improve the readability and clarity of our study. We carefully revised the entire manuscript to improve the flow and tighten the concepts, arguments and description of the theory, including removing some unnecessary text in the Introduction and Conclusion, add clarifications to the predictions in the Introduction, strengthening the main message in the Discussion, and improving the organization of the Conclusion.

Thank you again, as we believe the polished paper is stronger and more cohesive, with a clearer logical flow and structure.