Specialization-generalization trade-off in a Bradyrhizobium symbiosis with wild legume hosts
© Ehinger et al.; licensee BioMed Central Ltd. 2014
Received: 3 December 2013
Accepted: 10 March 2014
Published: 19 March 2014
Skip to main content
© Ehinger et al.; licensee BioMed Central Ltd. 2014
Received: 3 December 2013
Accepted: 10 March 2014
Published: 19 March 2014
Specialized interactions help structure communities, but persistence of specialized organisms is puzzling because a generalist can occupy more environments and partake in more beneficial interactions. The “Jack-of-all-trades is a master of none” hypothesis asserts that specialists persist because the fitness of a generalist utilizing a particular habitat is lower than that of a specialist adapted to that habitat. Yet, there are many reasons to expect that mutualists will generalize on partners.
Plant-soil feedbacks help to structure plant and microbial communities, but how frequently are soil-based symbiotic mutualistic interactions sufficiently specialized to influence species distributions and community composition? To address this question, we quantified realized partner richness and phylogenetic breadth of four wild-grown native legumes (Lupinus bicolor, L. arboreus, Acmispon strigosus and A. heermannii) and performed inoculation trials to test the ability of two hosts (L. bicolor and A. strigosus) to nodulate (fundamental partner richness), benefit from (response specificity), and provide benefit to (effect specificity) 31 Bradyrhizobium genotypes.
In the wild, each Lupinus species hosted a broader genetic range of Bradyrhizobium than did either Acmispon species, suggesting that Acmispon species are more specialized. In the greenhouse, however, L. bicolor and A. strigosus did not differ in fundamental association specificity: all inoculated genotypes nodulated both hosts. Nevertheless, A. strigosus exhibited more specificity, i.e., greater variation in its response to, and effect on, Bradyrhizobium genotypes. Lupinus bicolor benefited from a broader range of genotypes but averaged less benefit from each. Both hosts obtained more fitness benefit from symbionts isolated from conspecific hosts; those symbionts in turn gained greater fitness benefit from hosts of the same species from which they were isolated.
This study affirmed two important tenets of evolutionary theory. First, as predicted by the Jack-of-all-trades is a master of none hypothesis, specialist A. strigosus obtained greater benefit from its beneficial symbionts than did generalist L. bicolor. Second, as predicted by coevolutionary theory, each test species performed better with partner genotypes isolated from conspecifics. Finally, positive fitness feedback between the tested hosts and symbionts suggests that positive plant-soil feedback could contribute to their patchy distributions in this system.
Specialized biotic interactions contribute to processes that structure communities and adapt populations to local partners and environments[2, 3]. But, why do specialists exist if a generalist organism can occupy more environments and partake in more beneficial interactions[4, 5]? The “Jack-of-all-trades is a master of none” hypothesis asserts that specialists persist in a heterogeneous environment because the fitness of a generalist utilizing a particular habitat is lower than that of a specialist adapted to that habitat[5–7]. This hypothesis is based on the assumptions that adapting to new habitats involves fitness costs and that traits adaptive in one subset of environments are negatively genetically correlated with traits adaptive in other environments[4, 8–11], leading to local adaptation[2, 3]. When the habitat is a host or interaction partner, this trade-off could lead to co-adaptation.
Fitness feedbacks between plants and specialized soil-based microbes significantly affect plant productivity, community composition, and the distribution and abundance of plants[13–21]. Strong plant-soil feedbacks are often negative; due to antagonistic interactions between plants and pathogens[13, 21–27]. However, many plants engage in mutualistic interactions with soil-borne symbionts. Is there sufficient specialization among plant hosts to soil-borne mutualistic symbionts to impact plant community structure and diversity via positive feedbacks?
Multiple arguments predict that mutualists interacting with partners acquired from the environment are likely to be generalists. Since mutualists have higher fitness when partnered than when not, any partner should be better than no partner. If so, then selection should disfavor specialized mutualists (i.e., those with reduced interaction breadth relative to the breadth of available partners) that resist associating with common partners. Similarly, specialized mutualistic lineages should be vulnerable to perturbations in partner populations[30–36] and the geographic distribution and spread of specialized mutualists could be limited by partner availability[37–44]. Generalized mutualists might also access a wider range of environmental conditions by utilizing partners with differing ecological tolerances[45–48]. Finally, evolutionary convergence of traits among mutualists within an interaction group[49–51], but see, could reduce variance in partner benefit and thereby weaken the effectiveness of selection to specialize[53, 54].
Specialized mutualists nonetheless exist, e.g.[55–57]. Indeed, meta-analysis of bipartite interactions found that mutualistic webs are more specialized than antagonistic ones. Further, mutualism theory predicts that, when available partners vary in quality, selection favors mechanisms by which mutualists can select partners[58–63], which can lead to specialization. However, if more-beneficial partners are rare or distributed unpredictably, then selection for specialization via partner choice might be weak[45, 64, 65]. How frequently symbiotic mutualistic interactions are specialized is therefore an open question.
Here, we examined specialization in a group of wild legumes that interact with Bradyrhizobium genotypes in coastal sand dunes of Sonoma County, California. Bradyrhizobium is a genus of rhizobia that can infect legume roots and fix nitrogen endosymbiotically within root nodules in an interaction that is generally mutualistic[66, 67]. Dominant genes that either restrict nodulation with or alter the effectiveness of particular rhizobial genotypes have been described in both natural and managed plant populations[57, 68], but the effect of such genes on fitness feedbacks has been little studied. Moreover, little is known about whether legume-rhizobium interactions are sufficiently specialized to contribute to structuring communities via plant-soil feedbacks or coevolution.
We quantified the realized richness and phylogenetic breadth of rhizobial symbionts collected from plants growing in situ, and measured the frequency with which each host species associated with each of the identified Bradyrhizobium genotypes. Association frequencies quantify the interaction or link strength of each pairwise interaction and were used to calculate the Paired Differences Index PDI, for each host. This index summarizes variation in a partner’s link strengths[69, 70] with all available partners without making any assumptions about their statistical distribution.
Realized association specificity[72, 73] is strongly influenced by the local sampling environment. Therefore, we experimentally paired partners under controlled greenhouse conditions to examine fundamental association frequency[4, 75] of two host species. We also used the greenhouse experiment to quantify the fitness effect of each host on the inoculated Bradyrhizobium genotypes, which is a measure of the host’s potential impact on the rhizobium population and therefore its functional role within the community.
Finally, we used these data to test whether this system supports the “Jack-of-all-trades is a master of none” hypothesis, as well as its underlying assumption that, as frequently interacting partners become co-adapted, they become less adapted to other partners and consequently exchange weaker benefits.
We isolated Bradyrhizobium bacteria from nodules of four native California legumes growing on the Bodega Dunes at Bodega Marine Reserve and Sonoma Coast State Park, Sonoma County. Two species, Acmispon strigosus and Lupinus bicolor, are fast-growing annuals that are patchily distributed across the Bodega Dunes. In contrast, A. heermannii is a decumbent suffrutescent perennial and L. arboreus is a large upright perennial shrub, both widely distributed across the dunes.
Two DNA regions (NifD and ITS) were PCR amplified and sequenced from two to eight isolates from each plant, which yielded 84 sequenced isolates, including 81 NifD amplicons and 82 ITS amplicons (Genbank accession numbers in Additional file1: Table S1). Sequences that differed by one nucleotide were considered distinct, resulting in 22 ITS genotypes, of which 13 occurred in multiple nodules, and 22 NifD genotypes, of which 13 occurred in multiple nodules. There were 39 unique genotype combinations.
The observed PDIs of the four hosts (L. arboreus: 0.264, A. strigosus: 0.301, A. heermannii: 0.395, L. bicolor: 0.433) all differed significantly from the expected joint PDI (0.069 ± 95% confidence limit of 0.002) and from each other, which indicates that each of the four hosts specialized on a subset of the available NifD types, but differed in degree of specialization; L. bicolor was most specialized.
Analyses of population structure of Bradyrhizobium sampled from in situ nodules
(a) ITS region
(b) nifD locus
AMOVA of (a) ITS and (b) NifD sequences from Bradyrhizobium from in situ nodules
Sources of variation
Sum of squares
% of variance
Between species w/in genus
(b) NifD locus
Sources of variation
Sum of squares
% of variance
Between species w/in genus
The conservative analysis presented above does not test for between-host differences in response specificity, which is typically detected as a significant interaction between test host species and Bradyrhizobium genotype G × G interaction, in an ANOVA including data from both host species. In a less conservative analysis, host by genotype (G × G) interactions were prominent in the best fitting model (model 16 in Additional file4: Table S4). A significant G × G interaction of test host with Bradyrhizobium groups isolated from different host species (test host by species of origin interaction, F3,25 = 19.03, p > F < 0.0001; Additional file5: Table S5), occurs because variation in response to different Bradyrhizobium genotypes was large in A. strigosus and almost absent in L. bicolor (Additional file6: Table S6). Further, the G × G interaction due to test host species by Bradyrhizobium genotype nested within species of origin (covariance parameter estimate ± 1 S. E. = 0.07235 ± 0.024; Additional file7: Table S7) accounted for 7.24% ± 2.28% of the variance in plant dry weight. This significant interaction indicates that even within a Bradyrhizobium group isolated from the same host species, genotypes differed in the symbiotic benefits they provided to the two different test host species.
As expected from the conservative analysis, test host species differed significantly in dry weight (least square means [logits] of shoot dry weight ± 1 S. E. for A. strigosus = -4.80 ± 0.10, L. bicolor = -2.40 ± 0.11; t9 = -18.36, p < 0.0001) and groups of Bradyrhizobium genotypes isolated from different host species differed significantly in average effect on test host shoot dry weight (species of origin effect F3,25 = 24.56, p > F < 0.0001; Additional file8: Table S8). In particular, shoot dry weight produced by genotypes isolated from L. bicolor differed from that produced by genotypes isolated from other species (Additional file6: Table S6 and Additional file8: Table S8). When averaged across the test hosts, there was no significant variance in shoot dry weight due to inoculation with different Bradyrhizobium genotypes isolated from the same host species (strain(sp_orig) covariance parameter estimate ± 1 S. E. = 0.01929 ± 0.020; Additional file7: Table S7).
Response specificity, as indicated by the paired difference indices calculated from within-block growth responses stimulated by the different Bradyrhizobium genotypes, differed significantly between host species and was more than three-fold greater in A. strigosus (mean PDI ± 1 S. E. = 6.6 ± 0.7) than in L. bicolor (mean PDI ± 1 S. E. = 1.7 ± 0.1).
Bradyrhizobium fitness was estimated from nodule number, total nodule area, and area of the average nodule. Both nodule number and total nodule area differed significantly among Bradyrhizobium genotypes on each host (Additional file2: Table S2, Figures 2 and3).
On A. strigosus, over 30% of variance in total nodule area was due to Bradyrhizobium genotype: genotypes isolated from congeners produced significantly greater total nodule area than did those isolated from lupines. Ten percent of variance in nodule number was due to Bradyrhizobium genotype: genotypes isolated from congeners produced slightly fewer nodules than did those isolated from lupines. Seventy percent of variance in average individual nodule area was due to Bradyrhizobium genotype: genotypes isolated from congeners produced significantly larger nodules.
On L. bicolor, only 15% of variance in total nodule area was explained by Bradyrhizobium genotype. Although genotypes isolated from congeners produced significantly greater nodule area, the difference was slight. Bradyrhizobium genotype explained almost 15% of variance in nodule number, but there was no significant difference between genotypes isolated from congeners versus Acmispon species. Less than 20% of variance in the average individual nodule area on L. bicolor was due to Bradyrhizobium genotype. Genotypes isolated from congeners produce marginally significantly larger nodules than did those isolated from Acmispon hosts.
PDI values for two Bradyrhizobium fitness components
Mantel tests of the correlation between genetic distance matrices and phenotypic difference matrices
(a) A. strigosus dry weight
(b) A. strigosus average nodule area
(c) L. bicolor dry weight
(d) L. bicolor average nodule area
For A. strigosus, there was an overall positive relationship between total nodule area and host dry weight (genotype mean correlation between total nodule area and host dry weight = 0.77, prob(H0: r = 0) < 0.0001). Bradyrhizobium genotypes that were more beneficial for A. strigosus also produced on that host fewer nodules (genotype mean correlation between nodule number and host dry weight = -0.35, p < 0.0001) that were larger (genotype mean correlation between average nodule area and host dry weight = 0.71, p < 0.0001).
A weak positive relationship between total nodule area and host dry weight (genotype mean correlation between natural log of nodule area and host dry weight = 0.36, prob(H0: r = 0) < 0.05) suggests a weaker but still positive fitness feedback between L. bicolor and Bradyrhizobium symbionts. In L. bicolor, there was no relationship between host dry weight and nodule number (genotype mean correlation between natural log of nodule number and host dry weight = 0.15, prob(H0: r = 0) > 0.4).
Nodule data were lost from 14 L. bicolor plants. Seven out of 100 control plants had nodules, indicating cross-contamination (perhaps water draining from containers was splashed from the greenhouse bench). These plants (two A. strigosus and five L. bicolor) were removed from analyses. Cross-contamination probably occurred late in the experiment because the nodulated control plants were still very small and had formed only a few small nodules. We are confident that these potential cross contaminations do not undermine our conclusions, since any broader cross contamination would only reduce phenotypic differences observed between Bradyrhizobium genotypes.
When growing wild, all four legume species that we examined specialized on subsets of the nodulating community of Bradyrhizobium genotypes. However, accounting for genetic breadth of symbionts in wild-collected nodules shifted the ranking of host specialization. When Bradyrhizobium genotype was ignored (i.e., by simply counting the total proportion of symbionts that were non-identical at the focal loci), L. bicolor ranked as most specialized and L. arboreus as most generalized. However, Bradyrhizobium isolates from wild-collected nodules were genetically clustered by host genus and each Lupinus species hosted a phylogenetically broader group of symbionts than did either Acmispon species. Clustering was strongest when isolates were categorized by NifD genotypes, which suggests that hosts respond to trait loci located on the Bradyrhizobium symbiosis island. The observation that species in Acmispon exhibit greater realized association specificity than do those in Lupinus agrees with previous surveys of wild-collected nodules from plants in these genera[66, 78–80]. However, neither L. bicolor nor A. strigosus exhibited fundamental association specificity in greenhouse inoculation tests: both hosts could nodulate all tested Bradyrhizobium strains.
A gap between fundamental and realized association frequency is not uncommon and indicates that environmental context strongly influences nodulation. The complex biotic community within natural soil could influence a genotype’s nodulation success[81–83]. For example, soil pathogens could trigger systemic acquired resistance in the host, which might affect nodulation by some rhizobial genotypes but not others.
The availability in soil of multiple rhizobial genotypes also provides scope for plants to actively prefer certain genotypes. Host-symbiont recognition signaling is a well-studied feature of legume-rhizobium interactions. In particular, it has been hypothesized that, during nodule formation and prior to nitrogen fixation, plants can recognize and prefer more-beneficial genotypes via pre-infection partner choice; reviewed in. It is unclear however, whether this host ability would be evolutionarily durable, as it requires genetic linkage between at least three sets of genes in two different organisms (nitrogen fixation genes and signaling genes in the bacterial genome and receptors to the bacterial signals in the host plant genome). Such linkage could be stable under very specific conditions, but could also be broken by any incident of recombination. In several legume-rhizobium systems, including A. strigosus, multiple-genotype inoculation experiments have found that more-beneficial and/or co-evolved genotypes are more likely to occupy nodules[81, 88, 89]. However, none of these studies could definitively distinguish plant choice from interactions among rhizobia, which could be due to varying abilities to compete for soil resources, withstand direct interference by other genotypes, colonize the rhizosphere, and utilize a range of rhizosphere resources[93, 94]. In experiments with Bradyrhizobium strains collected from our site, neither in vitro growth rate nor interstrain interference was correlated with nodulation rate on A. strigosus, but these and other hypotheses should be further tested in more complex conditions where bacterial traits not directly related to symbiont effectiveness might influence nodulation ability.
Realized association frequency might also be restricted by patchily distributed plant and symbiont genotypes[95–97], which could be caused by coevolution and/or soil heterogeneity. In the Bodega Dunes, two episodes of Holocene dune advance[99, 100] left contrasting soils that might structure both plant and Bradyrhizobium populations: L. bicolor is restricted to mid-Holocene dunes whereas A. strigosus and A. heermannii occur only on poorly stabilized late-Holocene dunes (E. L. S. and T. J. M., personal observation). In contrast, L. arboreus occurs across dunes of both ages. We specifically sampled nodules from L. arboreus across its habitat range to control for potential confounding of host and Bradyrhizobium distribution; indeed, L. arboreus nodules harbored genotypes from across each gene network. However, the genotypic composition of isolates from L. bicolor did not differ significantly from those isolated from L. arboreus, which suggests that soil habitat is not the only determinant of genotype in wild-collected nodules of Lupinus.
The “Jack-of-all-trades is a master of none” hypothesis was supported by evidence that specialist A. strigosus obtained more fitness benefit from its genetically narrower group of beneficial symbionts than generalist L. bicolor obtained from its genetically broader group of compatible symbionts.
As predicted by the co-adaptation hypothesis, each host received greater fitness benefit from genotypes isolated from congeners or conspecifics (genotypes isolated from L. bicolor provided particularly poor benefit to A. strigosus). Although the sample sizes used to estimate fitness are not large, this pattern of response specificity has been detected in several studies of unmanaged legumes[102–104] and suggests that symbiont effectiveness may be increased by host-symbiont coevolution[45, 89, 98, 105].
A. strigosus plants were highly specialized in their fitness effect on the tested Bradyrhizobium genotypes, which is consistent with previous evidence that this host imposes absolute sanctions on less effective genotypes. There was much less variance among Bradyrhizobium genotypes in the benefits they obtained from L. bicolor.
Also supporting the co-adaptation hypothesis, genotypes of Bradyrhizobium were better adapted to hosts related to those from which their ancestors were isolated. This pattern was strongest on A. strigosus: genotypes isolated from congeneric hosts produced greater total nodule area and therefore likely yielded more progeny than did genotypes isolated from lupines.
As a consequence, symbiont benefits to A. strigosus fitness positively fed back to symbiont fitness via increased allocation to nodules: genotypes beneficial to this host realized greater fitness benefits than did those that were not. An overall positive correlation between NifD genotypic mean values of total nodule area and host shoot dry weight on A. strigosus suggests that strong positive fitness feedback between mutualists in this partnership is driven by traits encoded in the symbiosis island. This feedback pattern was much weaker on the generalist, L. bicolor, and was not significantly correlated with ITS genotype.
Positive partner feedback can produce positive frequency dependence and drive the most beneficial partners to local fixation[49, 106], a process that has been predicted for other legumes that exhibit symbiont specificity.
Much stronger positive fitness feedback might allow specialist A. strigosus to outcompete generalist L. bicolor in local patches. However, if L. bicolor were a better colonizer, which is expected of generalists, then it might persist in a spatially patchy environment through a competition-colonization trade-off. In that case, A. strigosus and L. bicolor would not coexist at a local scale, but both might persist at a larger spatial scale via patch dynamics. This process could also spatially structure Bradyrhizobium populations. Indeed, wild A. strigosus plants were found in association with a genetically narrow range of Bradyrhizobium genotypes that were most beneficial to their growth, suggesting that these genotypes dominate Bradyrhizobium populations in areas inhabited by this host. In contrast, as predicted from the weaker correlations between plant and rhizobial fitness components on L. bicolor, wild L. bicolor plants associated with a genetically wider range of Bradyrhizobium genotypes that were not necessarily the most beneficial.
A field survey detected differences in partner specificity among four sampled host species and greenhouse experiments revealed that at this site A. strigosus clearly specializes on a genetically narrower range of symbionts than does L. bicolor. The fitness benefits that these two hosts received from symbiotic partners affirmed two important tenets of evolutionary theory. First, as predicted by the Jack-of-all-trades is a master of none hypothesis, specialist A. strigosus obtained greater benefit from its genetically narrow group of beneficial symbionts than generalist L. bicolor obtained from a genetically more diverse group. Second, as predicted by the co-adaptation hypothesis[2, 3, 50], each test species performed better with partner genotypes isolated from conspecifics. Further, host fitness benefit translated directly into symbiont fitness via increased allocation to nodules: nodules occupied by genotypes isolated from conspecific hosts received the greatest benefit. These patterns were strongest in A. strigosus. Positive fitness feedback between this specialized host and its co-adapted symbionts could drive positive plant-soil feedback and contribute to natural distribution patterns observed in the field.
During the end of March and the beginning of April 2007, we isolated Bradyrhizobium bacteria from nodules of four native California legumes growing on the Bodega Dunes at Bodega Marine Reserve (38°19′01″N, 123°04′18″W) and Sonoma Coast State Park (38°20′34″N, 123°03′32″W). Portions of these dunes were stabilized in the mid-twentieth century by planting European beach grass, Ammophila arenaria[110, 111], but no samples were collected in areas occupied by A. arenaria. The four host legumes belong to two genera, Acmispon (A. heermannii Greene, and A. strigosus Nutt. Ex Torr. & Gray, previously Lotus heermannii and Lo. strigosus, respectively) and Lupinus (L. bicolor and L. arboreus), which are from distantly related clades of papilionoid legumes.
We harvested and cultured isolates from up to 20 nodules from each of four to six seedlings of each host species using the procedures of Sachs and colleagues. Briefly, intact plants were excavated, the root systems washed in tap water; nodules excised, surface sterilized, rinsed, crushed and individually streaked onto two replica plates of solid modified arabinose-gluconate medium (MAG). Successful cultures were archived in 25% glycerol-MAG at -80°C.
Genomic DNA was purified from 25 μl of each frozen isolate using Gentra Puregene Yeast/Bacteria kits, (Qiagen, Valencia, CA). We PCR amplified the intragenic spacer between 16 s and 23 s ribosomal subunits (ITS, 1256 nt) and portions of the nitrogenase α-subunit gene (NifD, 756 nt) as previously described. Amplification products were sequenced in both directions using an Applied Biosystems 96 capillary 3730xl DNA Analyzer (Foster City, CA) at the University of California, Berkeley, Sequencing Facility. Two to eight isolates from each plant were successfully amplified and sequenced, which ultimately yielded a sample of 84 sequenced isolates (Additional file1: Table S1).
Sequences were aligned using MAFFT with default parameters. Genotypes were identified in MacClade 4.05. Some portions of the ITS included indels and these regions were removed from the analysis.
Specialization is most easily quantified as the number of taxa with which a taxon interacts, termed “partner richness”. However, this measure accounts for neither association frequency[72, 73], phylogenetic distance among partners, e.g., phylogenetic breadth, nor fitness effects of interactions. Resurging interest has stimulated new methods for quantifying ecological specialization[1, 71, 75]; we used the Paired Differences Index.
where n i,j = the number of nodules on the ith host species occupied by the jthBradyrhizobium NifD type and n i = the total number of nodules sampled from the ith host (n ACHE = 19, n ACST = 18, n LUAR = 21, n LUBI = 23).
where P j is the link strength averaged across the four hosts of Bradyrhizobium strain j, P1 is the average link strength of the Bradyrhizobium strain with the strongest link averaged over the four host species, and N = 20 nodules, then used these values to estimate the mean and variance of PDI and calculate 95% confidence limits.
We constructed a molecular network for each DNA region with the neighbor-net algorithm implemented in SplitsTree 4.8, using the Akaike Information Criterion (AIC) implemented in Modeltest (version 3.4) to choose a suitable model of sequence evolution. The NifD genotype network was estimated using the GTR + I model of sequence evolution and the ITS network was estimated with the HKY + I + G model. When there were multiple isolates with the same sequence, we randomly picked one for analysis.
A molecular network approach was favored over a phylogenetic tree approach, as previous studies have repeatedly found incongruence between phylogenies of different gene loci in Bradyrhizobium bacterial strains[66, 79, 123–125]. Moreover, in our sample, SplitsTree found significant recombination in the ITS region and marginally significant recombination within the NifD locus. Genetic networks are standard tools for the analysis of populations in which genomes recombine; such events will induce a reticulated representation of genealogies, which violates the bifurcating model of phylogenetic trees, as reviewed in.
For each DNA region, we computed the matrix of average number of differences between pairs of sequences (pairwise differences) within and between host genera to test whether bacterial communities harbored by the two genera differed in genetic diversity. We further computed matrices of within and between host-species pairwise Φst to test whether pairs of host species harbored genetically distinct groups of bacteria. The Φst statistic for nucleotide diversity is analogous to Wright’s FST, but accounts for sequence divergence as well as genotype frequency. Finally, each DNA region was analyzed with a two-level hierarchical AMOVA model with the following levels: within host species, between species within host genus, and between host genera. Significance tests for each level of the AMOVA, the Φst values, and the pairwise differences were obtained by permuting genotypes between levels or populations. These analyses were performed with Arlequin 3.0b.
Seeds of A. strigosus and L. bicolor were obtained from ripe fruits of many plants growing at Bodega Marine Reserve in May-June 2008, pooled, mixed and stored dry at room temperature. In mid-July, seeds were surface-sterilized and scarified[66, 129], then germinated individually in wells of 96-well culture plates containing 200 ml of sterile ddH2O (incubated at 15°C, 5–7 days). Seedlings of A. strigosus were transferred to bleach-sterilized 38-mm × 140-mm Conetainers (Stuewe & Sons, Corvallis, OR, USA) filled with autoclaved quartzite sand. One-week old seedlings of L. bicolor were transferred to bleach-sterilized 64-mm × 250-mm Deepots (Stuewe & Sons, Corvallis, OR, USA) filled with autoclaved calcine-clay (Turface MVP®, Profile Products, Buffalo Grove, IL, USA). Media were chosen to maximize plant survival. Neither medium provides nitrogen; we have no reason to expect that differences in media would affect symbiont specificity or fitness benefits to hosts.
All transplants were placed in a greenhouse under ca. 50% shade for seven days of hardening, with misting twice daily, and subsequently exposed to full sunlight, watered daily, and fertilized weekly with 10-ml Jensens’s nitrogen-free solution until treatments were applied one month after germination (see below).
For the greenhouse inoculation experiment, we chose 29 Bradyrhizobium strains that represented the range of a preliminary concatenated nifD and ITS network of the sequenced isolates. For reference to previously described Bradyrhizobium genotypes, we added two strains (T11 and T12) from among those described by Sachs and colleagues, for a total of 31 strains. A culture of each strain was initiated from ~50 μl of original stock archived in glycerol, plated onto solid MAG medium and incubated at room temperature until bacteria covered the plate. Bacteria were washed from plates into 50-ml polypropylene tubes containing sterile dd H2O and vortexed. Bacterial concentration of each inoculum was estimated via light absorbance at 600 nm with a spectrophotometer and a previously established standard growth curve; then adjusted to 108 cells ml-1 by dilution in sterile dd H2O. Each treatment plant received 10 ml of bacterial suspension and control plants received 10 ml of sterile dd H2O.
The two host species were planted and inoculated at the same time, but grown in different substrates and separated on the greenhouse bench to prevent shading of A. strigosus by the larger L. nanus. Each host species was arranged into ten 36-seedling blocks: 31 test seedlings, each randomly assigned to a Bradyrhizobium strain, and five seedlings that received the sterile water control inoculum. All treatments were spatially randomized within each block and the 10 blocks were randomly located on the greenhouse bench, for a total of 360 plants per host species. Seven to eight weeks after inoculation, plants were harvested to obtain oven-dried aboveground (shoot) dry weight (a component of plant fitness). We counted, excised, and photographed nodules with a stage micrometer to obtain their total projected area using ImageJ64. Because symbiotic rhizobia reproduce clonally and are ultimately released from legume nodules into the soil, nodule number and mass or projected area have been treated as multiplicative components of rhizobial fitness[67, 129, 132–134]. When obtained using a standard-density, single-genotype inoculum, nodule number indexes the probability that an individual rhizobium cell can initiate a nodule[88, 135]. Nodule size, measured as biomass or projected area, is a good proxy for the number of viable rhizobia in a nodule[129, 132, 136]. Since most nodules are initiated by a single rhizobium cell[88, 137], nodule size indexes rhizobium fecundity. Like many fitness components, nodule number and individual nodule size are often negatively correlated. Nonetheless, when all nodules on a host are occupied by a single genotype, the ratio of total nodule area to nodule number estimates average fecundity on that host for cells of that genotype.
No analysis was required because all hosts nodulated all Bradyrhizobium genotypes.
Because the experimental design technically precludes treating test host as an independent variable in the ANOVA, it was most conservative to analyze each host species separately. These models (Additional file2: Table S2) included random effects of spatial block and Bradyrhizobium genotype, with genotype nested within host species from which it was isolated and further nested within host genus of isolation, along with all appropriate interactions. A planned one degree-of-freedom contrast compared the growth of inoculated versus uninoculated test plants and a posterior 1-df contrast compared the growth of plants inoculated with bacterial genotypes that had been isolated from the each of the two plant genera (i.e., Lupinus, L, versus Acmispon, A). Bacterial genotype effects on dry weight (Additional file3: Table S3) were compared using a Tukey-Kramer test of honestly significant pairwise differences.
Although the experimental design technically precludes treating test host as an independent variable in the ANOVA, if the different growth conditions (different growth media and locations in the greenhouse) of the two test host species did not interact with the effect of Bradyrhizobium genotype, then both species might be analyzed with a single model. Lack of replication of Bradyrhizobium genotype within blocks precluded testing for this interaction within each species. Nonetheless, we cautiously tried this less conservative approach, in which a significant interaction between test host species and Bradyrhizobium genotype G × G interaction, would indicate differences in response specificity between L. bicolor and A. strigosus.
We analyzed the entire data set (both host species) by adding the fixed factor of test host and all its interactions to a nested hierarchy of generalized linear mixed models to examine how host species, Bradyrhizobium genotypes, the species or genus of plant from which they were isolated, and two-way interactions affected plant benefit. The two hosts differed so much in response specificity that they exhibited very different distributions of shoot weight. For this reason, all models employed a gamma probability density function and a logit link function. Corrected Akaike Information Criteria (AICc) were used to find the best model (i.e. the one with the lowest AICc). Relative likelihoods of other models were compared using e(AICmin-AICi)/2, where AICmin is the corrected AICc of the best-fitting model and AICi is the AICc of the model being compared. This analysis was performed with SAS Version 9.3 (SAS Institute Inc., 2002–2010, Cary, NC, USA). The best fitting model (model 16 in Additional file4: Table S4) included the fixed effects of host species (host), host species from which the Bradyrhizobium genotype was isolated (sp_orig), and the host species by species of origin interaction, random effects of block (block) and Bradyrhizobium genotype nested within the host species from which the genotype was isolated (strain(ge_orig)), and the interaction of host species with Bradyrhizobium genotype nested within host species of origin (host* strain(sp_orig)). Thus, the G × G interaction was split between a fixed effect interaction (Additional file5: Table S5) and a random effect interaction (Additional file7: Table S7).
We also compared test host growth response, relative to uninoculated control, to Bradyrhizobium genotypes isolated from conspecifics versus response to those isolated from allospecifics. To calculate growth response (fold increase over uninoculated control), we used shoot weight of the control plants in each block to calculate the mean growth effect on each test host of the ith genotype in each block: (w i – w0)/w0, where w i = shoot dry weight of the plant inoculated with genotype i and w0 = shoot dry weight of the uninoculated conspecific control plant. This variable was analyzed with a generalized linear model including test host, host species of origin, and their interaction, with an exponential distribution and canonical (reciprocal) link function. A planned contrast was used to compare the fitness benefit obtained from Bradyrhizobium genotypes isolated from conspecific versus heterospecific hosts. This analysis was performed with JMP Version 10.0.0 (SAS Institute Inc, Cary, NC, 1989–2007).
Variance among blocks was used to calculate 95% confidence intervals for the PDI of each test host species.
The effect of the two test hosts on the fitness of each bacterial genotype was examined by univariate ANOVAs on nodule number, total nodule area per plant, and average nodule size. All dependent variables were log transformed to meet the ANOVA normality assumptions, pp. 185, 202–204. Standard deviations were not equal among factor levels; so we performed a Welch ANOVA, which allows for heteroscedasticity among factors, pg. 183.
Specificities of effect of the two test hosts on Bradyrhizobium genotype fitness under conditions of individual inoculation were calculated from equation (4), above, by defining p ijk in terms of each of two rhizobial fitness components, nodule number and total projected area of nodules. To meet the assumptions of normality and homoscedasticity between populations, nodule number was natural log transformed and nodule area was square root transformed. Variance among blocks was used to calculate 95% confidence intervals.
The groups of Bradyrhizobium genotypes isolated from congeneric versus heterogeneric hosts were identified by the population genetic analysis. The A group (genotypes T1 - T12) were originally isolated from Acmispon; the L group (genotypes T13 - T31) were originally isolated from Lupinus. Having been classified by host association, the bacteria within one group may not be more related to each other than to bacteria in the other group. Thus, although this contrast compares the average effects of the two bacterial groups, it does not test for a specific association between bacterial genotype and phenotype. To explicitly test for an association between bacterial genotype and symbiotic phenotypes, we used Mantel tests of non-random association between matrices. We computed six matrices: two matrices of pairwise genetic distances between all pairs of genotypes, one for NifD and one for ITS, and four matrices of pairwise mean phenotypic differences between all pairs of bacterial genotypes, two for A. strigosus and two for L. bicolor. Each genetic distance matrix was then compared with the four different phenotype matrices. Under the null hypothesis of these tests, a random association between the elements of the phenotype matrix and those of the genotype matrix would indicate that across Bradyrhizobium genotype pairs, the two genotypes within a pair shared phenotypes independently of the genetic distance between them. In contrast, a positive, non-random association would indicate that the phenotypic difference between the genotypes in a pair increases in concert with an increase in their genetic distance. The two reference genotypes, T11 and T12, were omitted during computation of the matrices, as was genotype T30, which we were unable to sequence at the NifD locus. The ANOVAs were performed with JMP®, Version 7. (SAS Institute Inc., Cary NC, 1989 – 2007); the Mantel tests were performed with the R statistical software.
To estimate fitness feedback between host and symbiont, the JMP statistical package was used to calculate the Pearson product mean correlation between host fitness, measured as shoot dry mass, and symbiont fitness, measured as nodule number, total nodule area per plant, and average nodule size.
The sequence datasets supporting the results of this article are available from Genbank (http://www.ncbi.nlm.nih.gov/genbank/); accession numbers JQ230720 – JQ230882.
Chi-square test statistic
Akaike information criterion
Analysis of molecular variance
Analysis of variance
degrees of freedom
Genotype by genotype interaction
Generalized time-reversible model of DNA evolution with invariant sites (I)
Hasegawa, Kishino and Yano model of DNA evolution with invariant sites (I) and Gamma-distributed among site rate variation (G)
16S-23S rRNA internal transcribed spacer
Modified arabinose gluconate
Gene coding for the alpha subunit of dinitrogenase
probability of obtaining a test statistic at least as extreme as the one observed, assuming a true null hypothesis
Paired differences index
The correlation of haplotypic diversity at different levels of hierarchical subdivision.
We thank Barbara Rotz and the Oxford Tract greenhouse staff for their valuable help; Michiel Vos for insights on genetic analysis; and the Bodega Marine Reserve of the U. C. Natural Reserve System and its outstanding staff. Seed and tissue collections were approved by the Bodega Marine Reserve Director and by permit from State of California Department of Parks and Recreation. This research was funded by grant DEB-0645791 to E. L. S, from the USA NSF. During some of this work, M. E. was supported by a grant from the Swiss NSF.
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly credited. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.