This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
LETTER Habitat structure and the evolution of diffusible siderophores
in bacteria
Rolf K€ummerli,1,3 Konstanze T.
Schiessl,1,2 Tuija Waldvogel,1
Kristopher McNeill2 and Martin
Ackermann1,2
Abstract
Bacteria typically rely on secreted metabolites, potentially shareable at the community level, toscavenge resources from the environment. The evolution of diffusible, shareable metabolites is,however, difficult to explain because molecules can get lost, or be exploited by cheating mutants.A key question is whether natural selection can act on molecule structure to control loss andshareability. We tested this possibility by collating information on diffusivity properties of 189secreted iron-scavenging siderophores and the natural habitats occupied by the siderophore-pro-ducing species. In line with evolutionary theory, we found that highly diffusible siderophores havepreferentially evolved in species living in structured habitats, such as soil and hosts, because struc-turing can keep producers and their shareable goods together. Poorly diffusible siderophores,meanwhile, have preferentially evolved in species living in unstructured habitats, such as seawater,indicating that these metabolites are less shareable and more likely provide direct benefits to theproducers.
Keywords
Comparative analysis, diffusion, dispersal, microbes, public goods, secondary metabolites, spatialstructure.
Ecology Letters (2014)
INTRODUCTION
Bacteria show a remarkable range of social behaviours,including the formation of multicellular structures, between-individual communication, group-coordinated motility andcoordinated resource acquisition and host attack (Crespi2001; Velicer 2003; West et al. 2007; Nadell et al. 2009).Most of these behaviours rely on the secretion of metabo-lites into the environment, where they can potentially beshared and generate group benefits. Because of these bene-fits to others, secretion of diffusible metabolites fits the defi-nition of a cooperative behaviour, and secreted metabolitesare therefore often referred to as public goods (West et al.2007).While the numerous laboratory studies on secreted metabo-
lites have greatly increased our understanding of the complex-ity of microbial lifestyles, little is known on the degree towhich bacteria share metabolites in natural environments, andwhether secreted metabolites can invariably be considered aspublic goods. To address this issue, we sought to assess thesignature of natural selection on bacterial metabolite secretionacross species and bacterial habitat types using a comparativeapproach. Building on an extensive body of theoretical andempirical work (Vetter et al. 1998; V€olker & Wolf-Gladrow1999; Allison 2005; West et al. 2006; K€ummerli et al. 2009;Driscoll & Pepper 2010; Nicolaisen et al. 2012; Allen et al.2013; Dobay et al. 2014), we argue that diffusible metabolites
face two problems: (1) metabolites might get lost due to diffu-sion; and (2) metabolites could be exploited by cheating indi-viduals that no longer contribute but still benefit from themetabolite produced by others. One possibility to controlboth (1) and (2) would be for natural selection to act on diffu-sivity properties of metabolites, thereby influencing their lossrate and shareability in the community. We can see two solu-tions to satisfy (1) and (2), whereby selection for one or theother solution is expected to depend on the level of habitatstructure the metabolite-producing species lives in. Specifi-cally, we predict that natural selection could preferentiallyfavour highly diffusible and therefore shareable metabolites inspecies that live in structured habitats, such as in soil or host.The reasoning here is that habitat structure, imposed by phys-ical barriers (Wang & Or 2013), keeps both metabolites andproducers together (West et al. 2006), thereby solving (1) and(2). Conversely, for species living in unstructured habitats,such as open seawater, natural selection should rather favourpoorly diffusible, and therefore less shareable metabolites,which stay in the vicinity of and mainly generate benefits forthe producers themselves. These adaptations would also sat-isfy (1) and (2), but would render metabolite secretion lesscooperative. Thus, we propose that the diffusivity propertiesof a secreted metabolite determine its position on a continuumranging from public to private goods (K€ummerli & Ross-Gillespie 2014), and that selection for these properties isdriven by habitat structure.
1Environmental Microbiology, Swiss Federal Institute of Aquatic Science and
Technology, €Uberlandstrasse 133, D€ubendorf, 8600, Switzerland2Department of Environmental Systems Science, Institute of Biogeochemistry
and Pollutant Dynamics, Swiss Federal Institute of Technology,
We tested our predictions by analysing the diffusivityproperties of 189 secreted siderophores (Table S1) producedby 124 species (Fig. 1, Table S2). Siderophores are iron-scav-enging molecules released into the environment to makeinsoluble or chelator-bound iron available for bacterialmetabolism (Miethke & Marahiel 2007). Using a Bayesianphylogeny-controlled comparative approach, we related, foreach siderophore, its diffusion length (V€olker & Wolf-Glad-row 1999), affinity to biological membranes (Xu et al. 2002)and water solubility (Martinez et al. 2000) to the level ofhabitat structure the siderophore-producing species lives in.As a negative control, we looked at pKa-values of sidero-phores, which are indicative of the siderophore’s ability tobind iron at a certain pH (Raymond et al. 1984; Miethke &Marahiel 2007), but are not expected to vary in response tohabitat structure. Furthermore, we conducted sensitivityanalyses to examine the robustness of our correlativeapproach.
MATERIAL AND METHODS
Data collection
We obtained information on bacterial siderophores from twodifferent sources. First, we made use of the online database‘Siderophore Base’ (http://bertrandsamuel.free.fr/sidero-phore_base/index.php) operated by Dr. Bertrand Samuel(University of Nantes, France). This database provides chemi-cal structures and references to the original literature of alarge selection of bacterial siderophores. Second, we searchedthe Web of Science (Thomson Reuters) for studies with thekeywords ‘siderophore’ and ‘bacteria’. This search resulted ina list of 1412 articles, which we scanned for relevant informa-tion on bacterial siderophores. As we were interested in thechemical properties of siderophores, we only considered sid-erophores for which the detailed chemical structure wasresolved. Furthermore, we only considered compounds forwhich their primary role seems to be the chelation and uptakeof iron via specific receptors. Accordingly, we excluded com-pounds that also bind iron but represent precursors of sidero-phores (e.g. salicylic acid) or bind iron for other reasons (e.g.pseudomonas quinolone system, PQS; Diggle et al. 2007).Finally, for mycobacteria we only considered siderophoresthat are actually secreted out of the cell (i.e. exochelin, car-boxymycobactin). Due to their special cell envelope structure,mycobacteria also possess envelope-embedded mycobactinsthat serve as shuttles, taking iron from secreted siderophoresto feed it into the intracellular space (Harrington et al. 2011).Overall, our searches with the above criteria were completedon April 20, 2012, and yielded chemical structures of 189 sid-erophores (Table S1) from 124 bacterial species (Table S2). Intotal, we had 284 data entries as some siderophores are pro-duced by several species, and some species produce multiplesiderophores.
Quantification of chemical properties of siderophores
We first estimated the diffusion length (d) for each sidero-phore, which is given by
d ¼ffiffiffiffiffiffiDt
p; ð1Þ
where t is the time in seconds and D is the diffusion coeffi-cient in cm2 s�1, which, in aqueous solutions, can be esti-mated by
D ¼ 2:7� 10�4
m0:71; ð2Þ
where m is the molecular weight (g/mol) of the diffusing agent(Schwarzenbach et al. 2003).Next, we estimated the fraction (fb) of siderophores that
binds to bacterial cell envelopes upon secretion (equationsbelow are derived from Schwarzenbach et al. 2003). The equi-librium constant (K) of the chemical reaction describing theassociation/disassociation of siderophores with a lipid mem-brane is given by
K ¼ ½sþ l�½s�½l� ; ð3Þ
where [s], [l] and [s+l] are the concentrations of free sidero-phores, free lipids and siderophore-lipid complexes respec-tively. fb is then given by
fb ¼ ½sþ l�½s� þ ½sþ l� ¼
K½s�½l�½s� þ K½s�½l� ¼
K½l�1þ K½l� : ð4Þ
An approximation for K, which can be obtained for eachsiderophore, is given by
logK ¼ 0:91 logKOW þ 0:5; ð5Þwhere Kow is the distribution coefficient, given by the ratio ofconcentrations of the siderophore in an octanol vs. a waterphase (Schwarzenbach et al. 2003). We estimated Kow at pH =7 for each siderophore using the computer program Marvin-Sketch 5.11.4 (2012) from ChemAxon (http://www.chemaxon.com/products/marvin/marvinsketch/). The above approxima-tion works fine for most compounds, except for amphiphilicmolecules (Schwarzenbach et al. 2003), which usually consistof a hydrophobic tail attached to a hydrophilic head (Sandy& Butler 2009). For this category of molecules, the embed-ment in membranes is mostly determined by the length andcomposition of the hydrophobic tail (Xu et al. 2002). Thusfor amphiphilic siderophores, we only considered the chemicalstructure of the hydrophobic tail to calculate Kow. To separateamphiphilic from non-amphiphilic siderophores, we deter-mined the longest hydrophobic side chain (i.e. the number ofcarbon atoms) of each siderophore and plotted side-chainlength against frequency (Fig. S1). This analysis revealed abimodal distribution with most siderophores (n = 136) havingno or a very short (1–2 C-atoms) side chain. There were onlythree siderophores with an intermediate side-chain length (3–6C-atoms), but a considerable number of siderophores (n = 50)with side chain length of 7–17 C-atoms. On the basis of thisbimodal distribution, we considered molecules with sidechains ≥ 7 C-atoms as amphiphilic. All other molecules (withside chains < 7 C-atoms) were considered as non-amphiphilic.Estimating [l] is difficult because the relevant lipid concen-
tration is the one that siderophores encounter directly uponsecretion (i.e. the sum of the lipid content of the producer cell
and its surrounding individuals per volume unit). Because ofthis difficulty, we chose an in silico approach, where we askedwhat the fb would be when mixing a certain number of bacte-ria (nc) with its siderophore, characterised by the KOW-value,in 1 mL water. [l] can then be estimated as
½l� ¼ nccwc1; ð6Þwhere cw is the cellular dry weight and cl is the cellular surfacelipid content. From the literature, we took cw = 2.8*10�13 gand cl = 9% (an intermediate value of the rather broad rangeof estimates yielded by different methods) (Neidhardt et al.1990). For nc, we took three values, 106, 108 and
1010 cells mL�1 to cover the range of bacterial densities typi-cally observed in nature (Whitman et al. 1998). This resultedin [l] = 2.5*10�8, [l] = 2.5*10�6 and [l] = 2.5*10�4.We further estimated the aqueous solubility logS (whereby
S is given in mol/L) for each siderophore using the ‘VirtualComputational Chemistry Laboratory’ (http://www.vcclab.org/lab/alogps/) (Tetko et al. 2001). Using data on moleculartopology, the implemented algorithm yields highly confidentestimates of logS (R2 = 95.4% of the variance explained bythe correlation between true and estimated value). Apart fromdescribing the quantity of a substance that can be dissolved inwater, logS can also be understood as a measure of self-aggre-
Micrococcus_luteus
sivob_muiretcabocy
M
Dickeya_chrysanthem
i
Anabaena_sp
senegorea_retcaboretnE
Serratia_marcescens
Enterobacter_cloacae
Bordetella_pertussis
Streptomyces_tendae
Aeromonas_hy
drophila
suec
avil
o_se
cymo
tpert
S
Amycolatopsis_AA4
Pectobacterium
_carotovora
Alcaligenes_denitrificans
Vibrio_B
LI41
Cupriavidus_necatorErwinia_am
ylovora
Halom
onas_campisalis
Thermobifida_fusca
Pseudomonas_fluorescens
Marinobacter_DS40M
6
Rhizobium_leguminosarum
Vibrio_H
C0601C
5
Arthrobacter_terregens
Streptomyces_W
S116 S
treptomyces_scabies
9W
G_se
cymo
tpert
S
Pseudomonas_rhodesiae
Staphylococcus_hyicu
s
Pseudomonas_putida
Klebsiella_oxytoca
Corynebacterium_glutamicum
Staphylococcus_
aureus
Magnetospirillum_magneticum
Brucella_abortus
Salmonella_enterica
Vibrio_parahaemolyticus
Marinobacter_algicola
Cystobacter_disciformis
Saccharopolyspora_erythraea
Pseudomonas_cichoriiStreptom
yces_nitrosporeus
Sinorhizobium_meliloti
43MT_secy
motpertS
Halom
onas_SL28
Bacillus
_megate
rium
Shigella_flexneri
Ochrobactrum_SP18
Streptomyces_YM5799
Mycobacterium_neoaurum
Shigella_dysenteriae
Pantoea_agglomerans
Mycobacterium_tuberculosis
Ralstonia_solanacearum
Bacillus_lic
heniformis
Shigella_boydii
Arthrobacter_pascens
Pseudomonas_tolaasii
Halom
onas_LOB5
Streptom
yces_griseus
Pseudomonas_monteilii
Azotobacter_vinelandii
Cupriavidus_metallidurans
Paenibacillus
_elgii
Stigmatella_aurantiacaActinomadura_madurae
Acinetobacter_haem
oliticus
Salmonella_typhimurium
snecsenimul_sudbahrotoh
P
sned
exe_
sitsy
conn
aN
Ralstonia_pickettii
Mycobacterium_smegmatis
sitse
p_ai
nisr
eY
Pseudomonas_corrugata
acitilocoretne_ainisreY
Bordetella_bronchiseptica
Pseudomonas_aureofaciens
Vibrio_DS
40M5
Bacillus_anthracis
Halom
onas_meridiana
Streptomyces_am
bofaciens
Streptom
yces_coelicolor
Pseudoalteromonas_haloplan
ktis
Mycobacterium_avium
Halom
onas_aquamarina
Anabaena_cylindrica
Pseudomonas_chlororaphis
Klebsiella_pneumoniae
Paracoccus_denitrificans
Azomonas_macrocytogenes
Pseudomonas_aeruginosa
Marinobacter_aquaeolei
Bacillus_
subtilis
Bacillus
_cereus
Vibrio_vulnificus
Vibrio_
R10
Azospirillum_lipoferum
Synechococcus_PCC7002
Aliivibrio_sa
lmonicida
Pseudoalteromonas_KP204
Vibrio_fluvialis
Listonella
_anguillar
um
sucit
oibit
na_s
ecy
motp
ertS
Rhodococcus_erythropolis
Pseudomonas_entomophila
Sorangium_cellulosum
Streptom
yces_pilosus
Pseudomonas_syringae
Marinobacter_DG
870
Marinobacter_hydrocarbonoclasticus
Burkholderia_cepacia_complex
Vibrio_campbelli
Bacillus
_thuring
iensis
Acinetobacter_baum
annii
Nocardia_tenerifensis
Pseudoalteromonas_luteoviolacea
Pseudomonas_stutzeri
Shigella_sonnei
Vibrio_
cholera
e
Shewanella_putrefa
ciens
Myxococcus_xanthus
Escherichia_coli
Agrobacterium_tumefaciens
ievla_ainfaH
Figure 1 Composite phylogenetic tree of the 124 species examined in our comparative analysis. Colours depict major phylogenetic clades: firmicutes =orange; cyanobacteria = purple; actinobacteria = turquois; a-proteobacteria = red; b-proteobacteria = blue, c-proteobacteria = green; d-proteobacteria =yellow (see methods for detailed information on phylogeny reconstruction and Table S2 for species information).
gation among molecules. In other words, the lower the solu-bility of a compound the more likely it aggregates with othermolecules of its own type, which in the case of amphiphilicsiderophores can lead to the formation of micelles (Zhanget al. 2009). Whether self-aggregation compromises the func-tionality of siderophores is not known. If this turns out to bethe case then low logS values could indicate that amphiphilicsiderophores remain associated with the bacterial membraneto keep functionality and do not diffuse away from the cell(Xu et al. 2002).To what extent are these three chemical properties indepen-
dent from one another? To address this question, we con-ducted Pearson’s product-moment correlations between thethree variables (note that fb was log-transformed to obtainnormally distributed residuals). We found no significant corre-lations between diffusion length and fb (r = �0.099, P = 0.94,variance explained = 0.6%), but significant correlationsbetween fb and logS (rho = �0.458, P < 0.001, varianceexplained = 20.7%), and diffusion length and logS (r = 0.624,P < 0.001, variance explained = 38.7%), showing that logS isnot entirely independent from the two other chemical proper-ties (Fig. S2). These analyses indicate that, while the threevariables do correlate with each other to some degree, eachvariable still has considerable independent explanatory power.To test whether chemical properties of siderophores corre-
late more generally with habitat structure, and not only thoseproperties associated with diffusivity, we conducted a controlanalysis with a chemical property that is not supposed to cor-relate with habitat structure. Specifically, we looked at pKa-values of siderophores, which are indicative for the sidero-phores’ ability to bind iron at a certain pH (Raymond et al.1984; Miethke & Marahiel 2007), but should not influence dif-fusivity. A pKa-value stands for the pH at which a given func-tional group is deprotonated. Upon deprotonation, a neutralfunctional group becomes negatively charged (and a positivelycharged one neutral), and can thereby better bind positivelycharged ferric iron. The closer the pKa-value is to the pH ofthe medium, the stronger is the affinity of the particular func-tional group for ferric iron. High pKa-values correspond tomore basic atoms, which are better metal-binding ligands.However, if the pKa-value is too high then the site will remainprotonated and blocked from binding. In addition to pKa-values of specific atoms, ligand-binding strength is also afunction of the number of chelating atoms, the size of the che-late rings, and the strain associated with forming the chelatedstructure. In any case, we expect pKa-values to vary consider-ably, because the siderophores may differ in the pH at whichthey bind iron optimally. However, there is no a priori reasonwhy pKa-values should vary in response to habitat structure.We used MarvinSketch 5.11.4 to calculate the pKa-value thatis closest to but lower than pH 8, a threshold value that ischaracteristic for the siderophore’s ability to bind iron aroundphysiological pH (Raymond et al. 1984).
Classifying bacterial habitats
We consulted the Bergey’s Manual of Systematic Bacteriology(volumes 1–3, 5: Bergey’s Manual Trust’s 2001–2012) toextract all available information on the habitats the 124 spe-
cies in our data set live in (Table S2). While all species werefeatured in the manual, data on some specific strains weremissing. For these strains, we consulted the original literatureto obtain relevant habitat information (see Table S2). As wewere interested in how habitat structure (i.e. the degree towhich a habitat is compartmentalised) relates to the chemicalproperties of siderophores produced in those environments,we distinguished five broad categories of habitat types. Inincreasing order of habitat structure, we distinguished: (1)marine habitats; (2) freshwater habitats including lakes, riversand ponds; (3) soil and sediment habitats; (4) habitats on sub-strates such as plant and animal tissues, plankton, food, rock,faeces, etc. and (5) habitats within plant and animal hosts.However, most species live in multiple environments, belong-ing to different habitat types, we calculated for each speciesthe habitat structure score (h)
h ¼P
hinh
; ð7Þ
where nh is the number of habitats and Σhi is the sum of thestructure values (1) to (5) as outlined above (Table S2). Forexample, h for Pseudomonas aeruginosa, a species that lives infreshwater, soil and host habitats is 3.3 = (2 + 3 + 5)/3. Weconsidered marine environments as very poorly structuredhabitats (V€olker & Wolf-Gladrow 1999; Hopkinson & Morel2009) because, even when growing on solid substrates, theenormous water volume promotes nearly unlimited bacterialdispersal and metabolite diffusion (Logares et al. 2009). Weregarded freshwater habitats as poorly structured environ-ments because these habitats are also dominated by the diffu-sive nature of water, albeit presumably to a reduced extentcompared to marine environments because of the increasedsubstrate to water ratio (Cole 1982). We classified soil andsediment environments as habitats with intermediate structurebecause of the high density of particles available that allowbacteria to adhere to and to grow in structured communities(Or et al. 2007). On the other hand, soils can also becomeflooded and/or water saturated (at least temporarily), whichpossibly disrupts communities. We regarded substrate-basedenvironments as rather highly structured because interactionsamong bacteria are confined to two dimensions, with themobility of bacteria being greatly restricted, and the possibil-ity to build up structured communities (Beattie & Lindow1999). Finally, we considered habitats within hosts as highlystructured because bacteria are confined to hosts and theircells (Ray et al. 2009), where there are restricted possibilitiesfor dispersal, metabolite diffusion and limited forces to dis-rupt structured communities.We are aware of the fact that classifying bacterial habitat
structure is a difficult endeavour. Some habitats might bemore or less structured than our considerations suggest, andwhile extreme habitats (e.g. marine vs. host) seem clearly dif-ferent from one another, ranking adjacent categories is lessobvious. Hence, to examine the robustness of our analyses,we collapsed adjacent habitat categories in all possible combi-nations, and tested whether we recover consistent associationsbetween habitat structure score and the chemical properties ofsiderophores. Moreover, associations between chemicalproperties of siderophores and habitat structure score could
be driven by a single habitat category, and not reflect a gen-eral trend across categories. To test this possibility, werepeated all analyses by dropping one habitat structure scorelevel at the time.
Phylogeny
We constructed a composite phylogenetic tree of all 124 spe-cies in our data set (Fig. 1). We based the deep branching intofirmicutes, cyanobacteria, actinobacteria and a-, b-, c-, d-pro-teobacteria on the analysis published by Wu et al. (2009).Deep branching of the c-proteobacteria (the most diversegroup in our data set) was based on the phylogeny publishedby Williams et al. (2010). We obtained fine-scale molecularphylogenetic trees for specific groups of bacteria from varioussources as follows: firmicutes (Bergey’s Manual of SystematicBacteriology, BMSB), cyanobacteria (BMSB), actinobacteriaexcept mycobacteria (BMSB), mycobacteria (Mignard &Flandrois 2008), a-proteobacteria (Gupta 2005), b-proteobac-teria (BMSB), c-proteobacteria (BMSB) and d-proteobacteria(Garcia et al. 2011). However, we assembled phylogenies fromvarious sources using different methodologies, we had to stan-dardise branch length. We did this by allocating an arbitrarylength of 1 for each branch connecting two major levels oftaxonomic classifications (kingdom-phylum; phylum-class;class-order; order-family; family-genus, Table S2). Evolution-ary diversifications occurring within these intervals were allo-cated a branch length of 1/b, where b is the number ofbranching events in the interval. However, we assumed thatmost of the evolutionary diversification from the genus to thespecies level occurred relatively recently and is an ongoingprocess, we allocated a reduced branch length (0.2) to thatinterval. Whenever relationships among species/strains in agenus remain unresolved, we put the unresolved group at thebasis of the genus level.
Statistical analysis
We used Bayesian phylogeny-controlled statistical analyses totest for associations between the chemical properties of a sid-erophore (d, fb, logS and pKa) and the habitat-structure score(h) of the species that produces the siderophore. For all analy-ses, we used the R package MCMCglmm developed by Had-field (2010), which performs generalised linear mixed models(glmm) based on a Markov chain Monte Carlo (MCMC)algorithm. This powerful procedure allows analysing nonpara-metric data, such as our bounded fb estimate (varying between0 and 1), and our rank-based habitat structure score h. Inaddition, MCMCglmm allows carrying out phylogeny-con-trolled analysis by implementing the phylogenetic tree as arandom factor into the model (Hadfield 2010). This is essen-tial for comparative approaches as ours, to filter out the vari-ance in siderophore chemical properties that is explained byphylogeny, and not by the variable of interest (i.e. habitatstructure). Consequently, we used our composite tree (Fig. 1)as a random factor in all analysis, whereby the relationshipbetween any pair of species is weighted by the shared branchlength. The MCMCglmm method is based on iterativeprocesses and therefore requires the specification of appropri-
ate prior distributions. We chose priors according to the pack-age guidelines (MCMCglmm package course notes; http://cran.r-project.org/web/packages/MCMCglmm). Specifically,for the residual variance structure of the covariates as well asthe random factor, we chose an inverse-Wishart distribution(variance at the limit set to V = 1 and the belief parameter nu= 0.002). This set of parameters results in flat priors, meetingthe requirement of not making any a priori assumptions onthe variance distribution. All models were run for 270,000iterations with a burn in phase of 20,000 iterations and a thin-ning interval of 100 (i.e. only every 100th iteration was consid-ered for analysis). These settings guaranteed that there was noauto-correlation between consecutive data points (d vs. h: cor-relation coefficient r = 0.014 � 0.07; fb vs. h:r = 0.004 � 0.010; logS vs. h: r = 0.014 � 0.009; pKa vs. h:r = 0.031 � 0.006), such that data points could be consideredas independent.From the posterior distributions generated by the Bayes-
ian iteration, we report the model estimates of the parame-ter value (i.e. the slope of the relationship between twovariables), the 95% HPD (highest posterior density inter-val), and PMCMC, with the latter two representing theBayesian equivalent of confidence interval and P-value.Associations between two variables were considered signifi-cant when the 95% HPD interval excluded zero (i.e. theexpected value of the null hypothesis when there is no sig-nificant association between two variables), and PMCMC ≤0.05, whereby PMCMC equals twice the smaller value for theproportion of posterior samples that are above or belowzero. However, Bayesian statistic are based on iterative pro-cesses, with outcomes slightly differing across runs, werepeated every analysis five times and report mean values ofstatistical estimates. All analyses were performed in R 3.0.2(http://www.r-project.org).
RESULTS
We found a significant positive correlation between the diffu-sion length of a siderophore and the habitat structure score ofthe species that produces this siderophore (Fig. 2a, Table 1).Our analysis further revealed a highly significant negative rela-tionship between the fraction of a specific siderophore bindingto lipid membranes and the habitat structure score of the spe-cies that produces this siderophore (Fig. 2b, Table 1). Thisresult was robust across a wide range of lipid concentrations(Table 1). Finally, we also found a highly significant positiveassociation between the water solubility of a siderophore andthe habitat structure score of the species that produces thissiderophore (Fig. 2c, Table 1). Conversely but as expected,there was no significant correlation between a siderophore’spKa-value closest to pH 8 (our negative control trait) and thehabitat structure score of the species that produces this sidero-phore (Fig. 2d, Table 1).How much are these correlations dependent on the specific
habitat structure classification used? To address this question,we conducted sensitivity analyses using broader habitat cate-gories by collapsing adjacent habitat classes in four differentways (Table S3). We found that all 12 sensitivity tests recov-ered the same linear correlations as in the main analysis, with
11 tests being highly significant (PMCMC < 0.002), and one testbeing marginally significant (PMCMC = 0.075, Table S3).To what extent do these correlations reflect general trends
as opposed to being driven by adaptation to one specific habi-tat type? To address this question, we conducted a second setof sensitivity tests, where we omitted species with a specifichabitat structure score from the data set (Table S4). In 12 ofthe 15 sensitivity tests, we found consistent significant linearcorrelations as in the main analysis. In the remaining threecases, the association between habitat structure score and sid-erophore chemical properties was best explained by a qua-dratic fit (significant in two cases, and marginally significant
in one case, Table S4), indicating that, in these cases, the cor-relation between habitat structure score and siderophorechemical property levels off with higher habitat structurescores.
DISCUSSION
Socio-microbiology, and in particular the study of secretedshareable metabolites, is a rapidly growing interdisciplinaryfield. Although of correlational nature, our comparative workon secreted siderophores provides three novel insights for thisdiscipline. First, secretion per se seems to be an insufficient
Table 1 Bayesian statistical analysis investigating the association between siderophore chemical properties and the habitat structure score
d – diffusion length Positive 0.611 [0.127, 1.105] 0.0133
fb – fraction of siderophores binding to membranes Negative �0.007* [�0.011, �0.002] 0.0037
�0.074† [�0.114, �0.035] 0.0005
�0.111‡ [�0.165, �0.056] 0.0003
logS – water solubility Positive 0.321 [0.165, 0.482] 0.0002
pKa closest to pH 8 None 0.074 [-0.254, 0.463] 0.5632
*lipid concentration [l] = 2.5*10�8 g mL�1.
†lipid concentration [l] = 2.5*10�6 g mL�1.
‡lipid concentration [l] = 2.5*10�4 g mL�1.
1 2 3 4 5
10
15
20
25
30
Habitat structure score
Diff
usio
n le
ngth
in
m
(a) (b)
(c) (d)
1 2 3 4 5
0.0
0.2
0.4
0.6
0.8
1.0
Habitat structure score
Frac
tion
of s
ider
opho
res
bin
ding
to li
pid
mem
bran
es
1 2 3 4 5
–6
–5
–4
–3
–2
–1
0
Habitat structure score
Wat
er s
olub
ility
(log
S)
1 2 3 4 5
0
2
4
6
8
Habitat structure score
pKa
clos
est t
o b
ut lo
wer
than
pH
8
Figure 2 Significant associations between the habitat structure score (h) of bacterial species and three diffusivity properties of their siderophores. (a)
diffusion length d; (b) the fraction of siderophores binding to lipid membranes fb; (c) the water solubility logS. (d) Negative control showing that there is
no significant association between the habitat structure score and a chemical property of siderophores that is important for binding iron (pKa-value closest
to but lower than pH 8), but not for diffusivity. Open squares depict medians for each class of habitat structure score, and dashed lines indicate the trend
criterion for a metabolite to be considered as a highly share-able public good. Instead, our findings suggest that the diffu-sivity properties of the secreted molecule determine the degreeto which a secreted metabolite can be shared in the commu-nity. Thus, secreted metabolites seem to be neither fully publi-cally nor privately available, but rather lay on a continuumbetween these two extremes, whereby the position a certainmetabolite occupies is determined by its diffusivity. Second,habitat structure seems to drive the evolution of the metabo-lites’ chemical properties. Highly diffusible siderophores areassociated with species living in relatively structured habitats,whereas species living in less-structured habitats typically havesiderophores with reduced diffusivity. Third, our findings arein line with evolutionary theory predicting that high environ-mental structuring can favour cooperation because structuringphysically limits individual dispersal, thereby keeping coopera-tors and their metabolites together, whilst keeping cheats atbay (Hamilton 1964; West et al. 2006). In less-structured envi-ronments, meanwhile, where mixing of individuals and metab-olites is less limited by the environment itself, the lowerdiffusivity of siderophores can be understood as a direct adap-tation to limit metabolite loss and exploitation by cheats.However, our analysis is of correlational nature, a key ques-
tion is whether habitat structure is indeed the selective forcedriving the evolution of siderophore diffusivity properties, orwhether other factors are causing the observed associations.For instance, habitat structure score is associated with a shiftfrom aquatic to terrestrial ecosystems, and it might be thatecosystem type and not habitat structure selects for differencesin the siderophore chemical properties. A statistical test of thispossibility is complicated by the fact that many species (48 of124) live both in aquatic and terrestrial ecosystems. Neverthe-less, when contrasting species that live in aquatic ecosystems(N = 60) with species uniquely living in terrestrial ecosystems(N = 64), we find that neither diffusion length (differ-ence = 0.318, 95% HPD interval [�0.739, 1.247],PMCMC = 0.535) nor the fraction of siderophores binding tolipid membranes (difference = �0.049, 95% HPD interval[�0.129, 0.033], PMCMC = 0.247) significantly differ betweenthe two ecosystem types, whereas siderophore solubility is sig-nificantly higher in species uniquely living in terrestrial ecosys-tems (difference = 0.371, 95% HPD interval [0.033, 0.704],PMCMC = 0.034). These analyses indicate that ecosystem typealone can only partially account for variation in siderophorediffusivity properties. Habitat structure score is also associ-ated with a shift from saline to freshwater habitats. Our sensi-tivity analyses, however, show that when excluding speciesuniquely living in marine habitats, we still recover significantassociations between habitat structure score and two of thethree siderophore-diffusivity properties (Table S4). Thus,salinity alone does also not appear to fully explain theobserved variation in siderophore diffusivity properties.Finally, habitat structure score further correlates with nutrientavailability, which in turn supports increasing bacterial densi-ties when moving from aquatic to soil to host habitats (Whit-man et al. 1998). Although higher population densitiescertainly decrease loss rate and improve siderophore sharing,it is insufficient to explain the evolution of diffusible metabo-lites, because non-producing cheats also benefit from
increased metabolite access at higher densities (Ross-Gillespieet al. 2009; Dobay et al. 2014). Although caution is alwaysrequired with correlational data, the above considerations sug-gest that habitat structure indeed represents a major selectiveforce shaping siderophore diffusivity properties.The dilute nature of marine environments has previously
been suggested to disfavour diffusible metabolites (V€olker &Wolf-Gladrow 1999; Martinez et al. 2003; Hopkinson &Morel 2009). One problem that has been recognised early onis the fact that, due to the diffusive environment, the fractionof molecules that are lost or degrade before reaching a recipi-ent might simply be too big (i.e. the net benefit of metabolitesecretion becomes very small). One possible solution to thisproblem would be to evolve more durable metabolites(K€ummerli & Brown 2010), which would guarantee thatsecreted compounds stay in the system for longer. Althoughextended durability could reduce the loss-due-to-diffusionproblem, it does not necessarily solve the cheater problem.This is because in a diffusive well-mixed environment metabo-lite producers and cheats can equally benefit from more dura-ble metabolites. Thus, it seems that the loss-due-to-diffusionand the cheater problem go hand in hand, jointly preventingthe evolution of highly diffusible siderophores in unstructuredenvironments such as the open sea.Our comparative analysis reveals a high level of variation
that is not explained by the significant associations betweendiffusivity properties of siderophores and the level of habitatstructure (Fig. 2). There are at least three factors potentiallycontributing to this unexplained variation. First, we usedcrude categories to classify habitats, which might ignore fine-scale differences in environmental structuring. For instance,while marine habitats can generally be regarded as unstruc-tured, a recent study showed that marine Vibrio species pref-erentially aggregate on organic particles – a considerablystructured habitat (Cordero et al. 2012). Second, we wereusing approximations to estimate diffusivity properties of sid-erophores. The level of precision for these approximationsmay vary across siderophores, thereby introducing additionalnoise to the data. Finally, many species live in multiple habi-tats, which differ in their habitat structure. Consequently, evo-lution might face a trade-off by either favouring a generalistsiderophore that works reasonably well in all habitats, or mul-tiple specialised siderophores that each work optimally in onespecific habitat (Dumas et al. 2013). This complexity is notcaptured in our analysis.Although we focused on siderophores, we propose that our
approach, assessing the position a secreted metabolite occu-pies on the continuum from public to private goods, is a pow-erful tool that can be applied to any class of secretedmetabolites. For instance, when examining the chemical struc-ture of the quorum sensing signalling molecules from the N-homoserine lactone (AHL) family of Gram-negative bacteria(Williams 2007), it becomes evident that two different classesof AHL molecules exist. While all AHL molecules consist of ahomoserine lactone ring linked to a fatty acid chain, thelength of this acyl chain varies considerable: one class ofAHLs possesses a relatively short acyl chain (4–6 C atoms),whereas the other class entails long acyl chains (up to 18 Catoms). In the light of our approach, we would predict that
the various AHL-molecules, too, occupy different positionson a continuum ranging from private to public goods, whichshould significantly impact how fast and across which rangesignalling molecules can be shared in the collective.Important to note is also that some secreted metabolites
may not fall in the category of highly shareable public goodsbased on their poor diffusivity properties, but they might cre-ate a public good through their action in the extra-cellularspace. For instance, P. aeruginosa secretes rhamnolipids forswarming motility on soft agar (Xavier et al. 2011). Rhamn-olipids are amphiphilic molecules that easily associate withbiological membranes, and do therefore, according to ourapproach, rather represent private than public goods. How-ever, the joint secretion of rhamnolipids by many bacterialeads to the formation of a surface-attached film, which itselfcan represent a public good, accessible to all bacteria in thevicinity to swarm upon (Xavier et al. 2011). Similarly, manypathogenic Gram-negative bacteria posses type-three secretionsystems (TTSS), which are complex machineries embedded inthe bacterial membranes, and used to inject toxins into hostcells (Kosarewicz et al. 2012). Although the TTSS itself clearlydoes not represent a public good, it can potentially create oneby causing damage and/or inflammation in the host tissue,which open opportunities for other cells to invade (Barrettet al. 2011).
ACKNOWLEDGEMENTS
We thank Luke McNally for statistical advice, Marta Pintofor assistance with Fig. 1, three anonymous referees for con-structive comments, and the Swiss National Science Founda-tion (grants to RK and MA) and the European Commission(Marie-Curie Reintegration grant No. 256435) for funding.
AUTHORSHIP
RK, KS and MA designed the research. RK, KS and TW col-lected the data, RK, KS, TW and KM performed chemicaland statistical analysis. RK, KS, KM and MA wrote the man-uscript.
REFERENCES
Allen, B., Gore, J. & Nowak, M.A. (2013). Spatial dilemmas of diffusible
public goods. eLife, 2, e01169.
Allison, S.D. (2005). Cheaters, diffusion and nutrients constrain
decomposition by microbial enzymes in spatially structured
environments. Ecol. Lett., 8, 626–635.Barrett, L.G., Bell, T., Dwyer, G. & Bergelson, J. (2011). Cheating, trade-
offs and the evolution of aggressiveness in a natural pathogen
population. Ecol. Lett., 14, 1149–1157.Beattie, G.A. & Lindow, S.E. (1999). Bacterial colonization of leaves: a
spectrum of strategies. Phytopathology, 89, 353–359.Bergey’s Manual Trust (2001–2012). Bergey’s Manual of Systematic
Bacteriology. 2nd edn. Springer, New York.
Cole, J.J. (1982). Interactions between bacteria and algae in aquatic