Introduction

Prokaryotes exhibit remarkable genome flexibility, with strains from the same species often containing dramatically different gene content [1–4]. Intraspecific differences in gene content are often characterized by a “core” genome (genes common to all strains) and “accessory” genome (genes found in a fraction of strains) [5]. While the core genome might represent a set of species-specific indispensable genes, we do not yet understand whether the accessory genome of a species is the result of neutral or adaptive evolution. Indeed, this is the subject of an ongoing debate: do the majority of prokaryotic accessory genes have negligible or positive fitness effects, i.e. are they neutral or adaptive?

Recent population genetics arguments support roles for both neutral and adaptive evolution as possible factors driving accessory genome evolution [6–9]. For example, microbial species with more accessory genes also tend to have larger effective population sizes, as expected of genetic variation in a population under neutral evolution [9]. On the other hand, models in which microbial genomes evolve in large, migrating populations, suggest that acquired genes can often be beneficial, as expected under adaptive evolution [7]. However, these studies have only addressed broad aspects of microbial populations such as effective population size, migration, and the fitness effects of gene loss and gain. In response, subsequent criticisms of these studies have strongly expressed the need for more functional, gene-explicit and ecological analyses [10–11]. Here I present the first such systematic analysis of 96 phylogenetically diverse prokaryotic species, which suggests that prokaryotic accessory genomes often provide significant metabolic benefits.

I chose to study metabolism as a possible explanatory factor for three reasons. (1) Metabolic genes dominate the functional content of accessory genomes [12] (S8A Fig). (2) Metabolic interactions between microbes—especially interdependencies—can often be adaptive [13–14]. For instance, microbes that obligately cross-feed, i.e. that critically depend on exchanging metabolites with each other, can grow faster than their wild-type counterparts [13]. Such a fitness benefit can also drive genomes, in many cases, to lose genes and become metabolically dependent [14,15]. If different genes are lost between different conspecific strains, this can lead to both metabolic interdependence, as well as accessory genomes (since different strains will have different metabolic repertoires) [16]. (3) Databases such as KEGG contain already-curated genomes for several fully-sequenced strains. KEGG contains high-quality gene and reaction annotations, allowing us to accurately predict the biosynthetic capabilities of each strain under different conditions [17].

In this study, I ask to what degree accessory genes can metabolically benefit conspecific strains. For this, I have used genome-scale metabolic network reconstructions of 1,339 prokaryotic strains (corresponding to 92 bacterial and 4 archaeal species) from the KEGG database over 59 distinct nutrient environments. In general, my analyses reveal two beneficial roles for the accessory metabolic content of prokaryotes. First, I find that the accessory genome of most species harbours extensive biosynthetic potential, with several accessory genes providing strains with additional nutrient utilization abilities. Second, I find that pairs of strains from the same species often display a remarkable potential for metabolic interdependence, which scales with the amount of accessory genome content. These interdependencies have the ability to alleviate strain-specific auxotrophies in a particular niche through the exchange of secreted metabolites. My results are, from a metabolic standpoint, consistent with a possible adaptive evolution of accessory genomes.

Results and discussion

To obtain a large set of species pangenomes, I first collected a list of all prokaryotic species in the KEGG GENOME database, and filtered those that had complete genomes for 5 or more conspecific strains. This gave me 1,339 genomes (96 species), which I used in all my subsequent analyses (S1 Table). To account for potential biases due to uneven phylogenetic sampling, I verified that a more restrictive choice of one species per genus did not significantly impact my results (55 species; S1 Fig). For each strain, I then extracted all annotated genes and metabolic reactions from the KEGG GENES and REACTION databases, respectively. To quantify accessory genome content for every species, I used the well-studied genome fluidity measure, φ [18]. For this, I calculated, across each pair of conspecific strains, the fraction of all genes in the pair that were unique to each strain. The average of this fraction over each species gave me its genome fluidity φ (see Methods).

I constructed metabolic networks for every individual strain, where each network contained the set of reactions corresponding to the strain’s genome in KEGG. I included gap-filled reactions when curated models were available [19], though I verified that their addition did not impact my results (S2 Fig). I used these reaction networks to infer the biosynthetic capabilities of each strain under several different conditions. To define these conditions, I selected 59 different carbon sources, previously shown and commonly used to sustain the growth of diverse microbial metabolisms in laboratory experiments [20–22] (S2 Table). I associated with each carbon source a different nutrient environment or condition. In each condition, I included exactly one of the 59 carbon sources, say glucose, along with a set of 30 commonly available metabolites, which I assumed were always available (for example, water and ATP; S3 Table).

To assess biosynthetic capability, I curated a list of 137 crucial biomass precursor molecules, often essential for growth (hereafter, “precursors”) from 70 experimentally verified high-quality metabolic models [23] (S4 Table). Finally, to calculate what each strain could synthesize in a particular environment, I used a popular network expansion algorithm: called scope expansion [24–25]. This algorithm determines which metabolites each strain can produce—its “scope”—given an initial seed set of already available metabolites. To start with, only those reactions whose substrates are available in the environment can be performed, and their products constitute the initial set of metabolites that can be produced. These metabolites can then be used as substrates for new reactions that can then be performed, and step by step, more metabolites can be produced. When no new reactions can be performed, the algorithm stops, giving the full set of metabolites that could be synthesized in the given environment. Such a calculation sidesteps the need for arbitrary assumptions of binary (yes/no) growth and optimality typically used in more complex metabolic modeling approaches such as flux balance analysis [26] and is well-known for its ability to infer what metabolic networks can synthesize in diverse conditions [27–28].

I first investigated the capabilities of individual strains. Specifically, I was interested in the extent to which the accessory genes in each strain expanded the set of precursors that could be synthesized. For each species, I calculated, via network expansion, the list of precursors that could be produced per strain per condition. I then counted how many unique precursors each strain could synthesize across all conditions, i.e. by the accessory genes alone. From this, I computed, for every species, its accessory metabolic capacity α, defined as the average number of precursors (per strain per condition) produced exclusively due to the accessory genome.

I found that while for 18 species this quantity was zero, for the majority of prokaryotic species (81%), this number lay between 0.1 and 15.0 (mean 3.1; median 2.0). Further, α scaled positively with genome fluidity φ (Spearman's rho = 0.44; P value = 7 x 10⁻⁶; Fig 1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. The accessory genomes of prokaryotes harbour extensive biosynthetic potential.

Scatter plot of genome fluidity φ versus accessory metabolic capacity α for the 96 prokaryotic species in this study. Each point represents the average number of precursors that could be synthesized by the accessory genome content alone in each strain of a species. The Venn diagrams on the right provide a schematic representation of open pangenomes (high φ, small core) versus closed pangenomes (low φ, large core). The solid black line represents a linear regression and the gray envelope around it, the 95% prediction interval. rho corresponds to Spearman’s nonparametric correlation coefficient and the P value to a one-way asymptotic permutation test for positive correlation.

https://doi.org/10.1371/journal.pgen.1007763.g001

Since I observed that the accessory genome of different strains typically imparted different biosynthetic capabilities to different strains, I wondered if, in the same conditions, metabolic interactions between conspecific strains could further expand these capabilities. This could, for instance, indicate a potential dependence of an auxotrophic strain on another strain, i.e. a strain that cannot produce a crucial precursor in an environment. Such auxotrophies have been previously shown for example, in different strains of Escherichia coli co-inhabiting the human gut [29].

For each pair of conspecific strains in each condition, I calculated a metabolic dependency potential (MDP), defined as the average number of new precursors each strain has the potential to synthesize when grown as a pair versus alone. Here I assessed, in every pair, which metabolites that could be produced and secreted by one strain could subsequently allow the production of a new precursor in the other strain that it would not otherwise be able to make (i.e. was auxotrophic for). Note that this method does not count those metabolic interactions that can provide extra (functionally redundant) pathways to produce a precursor and supplement growth, and is thus more likely to represent actual or obligate dependencies. I verified that my approach can successfully predict such obligate dependencies by comparing with some well-documented intra- and inter-species pairs [13–14,30–32] (see Methods) (S3 Fig).

I found that while I could not detect any dependency potential for 17 species, surprisingly, the majority of species (82%) showed an MDP per strain per condition between 0.1 (for Bacillus thuringiensis) and 3.3 (for Ralstonia solanacearum), with a mean 1.7 and median 1.4. Interestingly, the 17 species for which I could not detect any MDP matched those with zero α (the leftover Legionella pneumophila showed low MDP = 0.5). Over all tested pairs with detected dependency potential (48%), commensal interactions were more common than mutualisms (29% versus 19%; Fig 2B). This is because, in species with detected dependencies, not all pairs show dependency potential (on average 46% conspecific pairs do). The auxotrophies relieved by these dependencies varied from those for amino acids, vitamins, carbohydrates, and organic acids, among others (Fig 2C).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Conspecific metabolic dependencies scale with accessory genome content.

a, Scatter plot of genome fluidity φ versus conspecific metabolic dependency potential (MDP) for the 96 prokaryotic species in this study. Each point represents the average number of dependencies detected per strain per condition across all conspecific pairs for one species. Colours represent each species’ phylum-level taxonomic identity. The solid black line represents a linear regression and the gray envelope around it, the 95% prediction interval. rho corresponds to Spearman’s nonparametric correlation coefficient and the P value to a one-way asymptotic permutation test for positive correlation. b, Pie chart of the types of interactions detected in each conspecific pair. c, Pie chart of the types of auxotrophies that the detected dependencies relieve due to pairwise growth. Each of the 137 precursors belongs to a unique chosen category (S4 Table).

https://doi.org/10.1371/journal.pgen.1007763.g002

Strikingly, like α, MDP also scaled positively with genome fluidity φ, suggesting that greater amounts of accessory content can potentially sustain more conspecific metabolic dependencies (Spearman's rho = 0.56, P value = 4 x 10⁻⁹; Fig 2A). For both α and MDP, considering medians instead of means did not impact my results (for α: Spearman's rho = 0.38, P value = 10⁻⁴; for MDP: Spearman's rho = 0.51, P value = 10⁻⁷; S4 and S5 Figs). To verify that such potential conspecific dependencies are indeed ecologically realizable, I repeated my analysis, this time restricting it to those genomes, which were known to co-occur in microbial communities (29 strains across 14 species; see Methods). I found that my observed trend was still valid, namely MDP still scaled with φ, suggesting that several auxotrophies may indeed be reduced through within-species metabolic exchanges in nature (Spearman’s rho = 0.56, P value = 0.03; S6 Fig).

Given the extent to which I detected the potential for obligate metabolic interactions between conspecific strains, I wondered whether such interactions are possibly common among prokaryotes. For this, I extended my study to analyze metabolic dependency potential between inter-specific strains (see Methods). I found that, indeed, strains from all species can metabolically depend on strains from at least one other species to alleviate potential auxotrophies across many different environments (with MDP ranging from 1.6 to 3.2, with mean 2.2; S7 Fig). Interestingly, I found that these interspecific metabolic interactions often involve accessory metabolic genes as well.

Taken together, my results suggest that a considerable fraction of prokaryotic accessory genomes contains potentially beneficial metabolic functions (upto 70% of accessory genes per strain across this study, with median 40%; see S8D Fig). Specifically, I found that the accessory “metabolome”: (1) expands a genome’s biosynthetic potential, possibly allowing for niche-specific adaptations [29,33–34]; and (2) reduces potential auxotrophies via obligate metabolic interactions, also explaining how conspecific strains can coexist despite high competitive potential [35–38]. The accessory genes that impart these functions are often different (median overlap 10%), suggesting that these are indeed distinct, non-redundant benefits. Moreover, apart from these additional biosynthetic abilities, the y-intercept for genome fluidity (at φ = 0.03 for both α and MDP) provides an estimate of metabolic redundancies (such as extra pathways).

My findings may additionally help explain the following observations: (1) metabolic functions are enriched in accessory genomes (median 50% in accessory versus 38% in core; S8C Fig); and (2) the variation in accessory metabolic genes exceeds the variation in genes of many other functions (metabolic variation being dominant in 81% of examined cases; S8B Fig).

Previous studies have suggested that the evolution of metabolic dependencies likely occurs via adaptive gene loss [15,39] (e.g. the Black Queen hypothesis). Such a mechanism suggests that metabolic dependency evolution can often lead to reduced genome sizes, but makes no comment on genome flexibility (i.e. gene content variability). My results also indicate that metabolic dependency evolution can impact genome flexibility as well. Specifically, more flexible genomes (with more variable gene content) are more likely to display a potential for metabolic interactions.

Can stochastic accessory gene turnover explain these results? To test this, I repeated my study with randomly assembled pangenomes. Within each species, I retained the core genes in every strain and shuffled the accessory genes between strains (see Methods). During this randomization, I preserved the observed within-species genome size distribution, strain number distribution, and ensured that any change in species’ genome fluidity was insignificant. I found that not only did this significantly diminish the metabolic benefits observed in each species, both measurements of α and MDP yielded non-significant correlations, suggesting that the mere presence of additional accessory genes is unlikely to explain my observed trends (for α: Spearman's rho = 0.01, P value = 0.9; for MDP: Spearman's rho = 0.17, P value = 0.1; S9A and S10A Figs). The measured benefits remained lower than observed, even when I shuffled known operons of genes together instead of shuffling genes one by one (S9B and S10B Figs; see Methods). I believe this is because often, prokaryotic operons do not contain complete metabolic pathways, but instead parts of them (in my data set, each metabolic operon encoded 1.5 reactions on average, while pathways typically had 4 to 5 steps). Collectively, this suggests that accessory gene acquisition is consistent with the coordinated gain of functional and beneficial pathways, which I believe provides further support for the accessory genes being maintained for adaptive reasons.

To summarize, here I addressed the debate on whether the accessory genomes of prokaryotes are beneficial. I found that, indeed, large fractions (about 40%) of the prokaryotic accessory gene pool can contribute to metabolic benefits. Specifically, such genes can allow microbes to produce a larger repertoire of crucial molecules, and facilitate the exchange of others. Since these functions can improve growth in many habitats, my results suggest that adaptation may explain accessory gene maintenance. Note, however that my analyses are only capable of detecting obligate metabolic dependencies and biosynthetic potential, and do not consider signaling, regulation, metabolic redundancy, etc. that could also play important functional roles and might indicate potential benefits due to additional accessory genes. Further work might also explain accessory genomes in those species, where I could not detect additional metabolic functions, if such roles are indeed there. Moreover, even in the context of metabolism, more detailed metabolic models, when available, may be used to probe even more precise fitness effects of intraspecies metabolic variability, including the effect of higher-order interactions. However, these studies would require knowledge of a large number of parameters such as reaction kinetics, thermodynamics, and exact strain biomass compositions before they are feasible. Finally, systematic measurements of the fitness effects of all accessory genes, metabolic and otherwise, are needed for more complete estimates of the fraction of accessory genomes consistent with adaptive versus neutral evolution.

Methods

Supporting information

Acknowledgments

I thank Sergei Maslov, Saurabh Mahajan, Rohini Subrahmanyam, Vishaka Datta, Sandeep Krishna, Deepa Agashe, and Sankarshan Talluri for discussions and comments on the manuscript. I am especially grateful to Christian Kost and two anonymous reviewers for their suggestions and comments on earlier drafts of this manuscript.