Specialization-generalization trade-off in a Bradyrhizobium symbiosis with wild legume hosts

Background 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. Results 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. Conclusions 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.


Background
Specialized biotic interactions contribute to processes that structure communities [1] 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][6][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][9][10][11], leading to local adaptation [2,3]. When the habitat is a host or interaction partner, this trade-off could lead to co-adaptation [12].
Multiple arguments predict that mutualists interacting with partners acquired from the environment are likely to be generalists [28]. 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 [29]. Similarly, specialized mutualistic lineages should be vulnerable to perturbations in partner populations [30][31][32][33][34][35][36] and the geographic distribution and spread of specialized mutualists could be limited by partner availability [37][38][39][40][41][42][43][44]. Generalized mutualists might also access a wider range of environmental conditions by utilizing partners with differing ecological tolerances [45][46][47][48]. Finally, evolutionary convergence of traits among mutualists within an interaction group [49][50][51], but see [52], could reduce variance in partner benefit and thereby weaken the effectiveness of selection to specialize [53,54].
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 [69] or link [70] strength of each pairwise interaction and were used to calculate the Paired Differences Index PDI [71], 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 [74]. 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 [76].
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.

Isolate collection and sequence datasets
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 file 1: 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.
Realized in situ association specificity (realized niche breadth) 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.

Genetic breadth of symbiotic partners
In both gene networks, Lupinus hosts associated with a genetically broader range of bacteria than did Acmispon hosts. Isolates from within a host genus tended to cluster together ( Figure 1) and exhibited little overlap between host genera. Host genera shared no genotypes in the NifD network ( Figure 1A). A single common ITS genotype (S1) was shared between host genera and one genotype from Lupinus (S8) clustered closely to most genotypes from Acmispon ( Figure 1B).

Genetic distance matrices among bacterial communities
Bacterial populations differed significantly between the two host genera but not between species within a genus. For both DNA regions, significantly large Φ st values and corrected between-population mismatch rates occurred between Bradyrhizobium populations isolated from different host genera, but not between populations isolated from different host species within a genus (Table 1). However, bacteria isolated from Lupinus were more genetically variable than those isolated from Acmispon. For both DNA regions, the average pairwise sequence divergences between Bradyrhizobium genotypes from Lupinus were two-to five-fold greater than they were between genotypes from Acmispon ( Table 1).

Distribution of bacterial genetic variance
In a two-level hierarchical AMOVA (Table 2), neither ITS nor NifD sequence explained a significant component of variance between host species within genus. Most genetic variance occurred among plants within species (50.29% for ITS and 40.84% for NifD) and between host genera (49.14% for ITS and 57.92% for NifD). Permutations suggest that the variance component due to host genus was not significant ( Table 2); even though genera harbored distinct genotypes in both networks ( Figure 1) and all inter-genus values of pairwise genetic differences and Φ st were statistically significant for both DNA regions (Table 1). This discrepancy probably arises from a lack of power in the AMOVA, which treats genus as a random effect, even though it is associated with only one degree of freedom.

Fundamental association specificity (fundamental niche breadth)
In the greenhouse inoculation experiment, neither test host exhibited fundamental association specificity; each could nodulate all tested Bradyrhizobium genotypes (Figures 2 and 3).

G × G interaction (Response specificity)
On average, inoculation increased host shoot dry weight over that of uninoculated control plants. The average increase was 2.8-fold for A. strigosus (F 1, 318 = 64.261, p value < 0.0001) and 2.9-fold for L. bicolor (F 1, 314 = 387.191, p value < 0.0001). However, bacterial genotypes differed significantly in their effects on shoot dry weight ( Figure 4, Additional file 2: Table S2 and Additional file 3:  Table S3). Notably, several genotypes failed to improve host growth beyond that of uninoculated controls (17 for A. strigosus and three for L. bicolor, using the Tukey-Kramer honestly significant difference test; 13 for A. strigosus and two for L. bicolor, using the less conservative Student's t-test).
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 [77], 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 file 4: 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, F 3,25 = 19.03, p > F < 0.0001; Additional file 5: Table S5), occurs because variation in response to different Bradyrhizobium genotypes was large in A. strigosus and almost absent in L. bicolor (Additional file 6: 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 file 7: 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; t 9 = −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 F 3,25 = 24.56, p > F < 0.0001; Additional file 8: 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 file 6: Table S6 and Additional file 8: 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 file 7: Table S7).
The fold increase in growth of inoculated plants over uninoculated controls was strongly affected by the interaction of test host with source host ( Figure 5; χ 2 d.f. = 1 = 25.61, p < 0.0001). Further, the planned contrast comparing the fold increase in fitness obtained from Bradyrhizobium genotypes isolated from conspecifics versus heterospecifics was highly significant (likelihood ratio χ 2 d.f. = 2 = 28.76, p < 5 × 10 −7 ). Both main effects were statistically significant (test host species χ 2 d.f. = 1 = 5.059, p < 0.03; host species of origin χ 2 d.f. = 1 = 5.059, p <   Figure 1 Genotype networks generated by the neighbor-net algorithm. (A) NifD locus, based on the distance calculated by MODELTEST, using a GTR + I model of sequence evolution [121]. (B) ITS region, based on the distance calculated by MODELTEST using a HKY + I + G model of sequence evolution [122]. Dotted lines represent the position of a genotype on the network. Each box represents one genotype, each symbol represents a host species (filled circles represent bacteria harvested from nodules of A. strigosus; open circles, A. heermannii; filled triangles, L. bicolor; open triangles, L. arboreus), and numbers indicate the number of strains representing a particular genotype; labels T1 to T31 in italics represent the Bradyrhizobium genotypes used in the greenhouse experiment (as listed in Additional file 1: Table S1). One genotype, T30, appears only in the ITS network, because we were unable to sequence its NifD locus. Genotypes T11 and T12 do not figure in either network because they were addenda from another study, used for reference, but not part of the original sample. 0.03), but the mean growth effect on each host species changed rank between conspecific and heterospecific pair- . 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).

G × G interaction (Effect specificity)
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 file 2: Table S2, Figures 2 and 3).
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.
Paired difference indices (PDI) calculated from host effect on Bradyrhizobium fitness under non-competitive conditions of single inoculations is presented in Table 3. When measured in terms of nodule number, A. strigosus was significantly more specialized than L. bicolor, whereas PDI of the two hosts did not differ when calculated in terms of total nodule area.

Association between bacterial genotypes and phenotypes
The matrices of pairwise phenotypic differences in A. strigosus shoot dry weight and average nodule area were each significantly associated with pairwise genetic differences at the NifD locus; association with the ITS region was not  significant for shoot dry weight and marginal for average nodule area. Traits measured on L. bicolor exhibited no significant associations (Table 4).
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(H 0 : 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(H 0 : r = 0) > 0.4).
(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.

Association richness and phylogenetic niche breadth
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 nonidentical 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][79][80]. However, neither L. bicolor nor A.  T1  T2  T3  T4  T5  T6  T7  T8  T9  T10  T11  T12  T13  T14  T15  T16  T17  T18  T19  T20  T21  T22  T23  T24  T25  T26  T29  T28  T27  T31  T30  C   T1  T2  T3  T4  T5  T6  T7  T8  T9  T10  T11  T12  T13  T14  T15  T16  T17  T18  T19  T20  T21  T22  T23  T24  T25  T26  T29  T28  T27  T31  T30  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 [6] and indicates that environmental context strongly influences nodulation. The complex biotic community within natural soil could influence a genotype's nodulation success [81][82][83]. For example, soil pathogens could trigger systemic acquired resistance in the host, which might affect nodulation by some rhizobial genotypes but not others [84].
The availability in soil of multiple rhizobial genotypes also provides scope for plants to actively prefer certain genotypes. Host-symbiont recognition signaling is a wellstudied feature of legume-rhizobium interactions [85]. In particular, it has been hypothesized that, during nodule formation and prior to nitrogen fixation, plants can recognize and prefer more-beneficial genotypes via preinfection partner choice; reviewed in [62]. 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 [86]. In several legume-rhizobium systems, including A. strigosus [87], multiple-genotype inoculation experiments have found that more-beneficial and/or coevolved 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 [90], withstand direct interference by other genotypes [91], colonize the rhizosphere [92], 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 [87], 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][96][97], which could be caused by coevolution [98] 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 [101]. 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.
Response and effect specificity, fitness trade-offs, and fitness feedback 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][103][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 [64] on less effective genotypes [87]. 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 [67] 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 [83].
Much stronger positive fitness feedback might allow specialist A. strigosus to outcompete generalist L. bicolor in local patches [107]. However, if L. bicolor were a better colonizer, which is expected of generalists [108], then it might persist in a spatially patchy environment through a competition-colonization trade-off [109]. 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 [12]. 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.

Conclusions
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-alltrades 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.

Collection of wild Bradyrhizobium isolates
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  (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 [112], and A. strigosus Nutt. Ex Torr. & Gray [112], previously Lotus heermannii and Lo. strigosus, respectively) and Lupinus (L. bicolor and L. arboreus), which are from distantly related clades of papilionoid legumes [113]. 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 [66]. 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.

Molecular methods and analysis
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) [114] and portions of the nitrogenase α-subunit gene (NifD, 756 nt) [115] as previously described [66]. 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 file 1: Table S1).
Sequences were aligned using MAFFT [116] with default parameters. Genotypes were identified in MacClade 4.05 [117]. Some portions of the ITS included indels and these regions were removed from the analysis.

Quantifying realized in situ association specificity
Specialization is most easily quantified as the number of taxa with which a taxon interacts [118], termed "partner richness". However, this measure accounts for neither association frequency [72,73], phylogenetic distance among partners, e.g., phylogenetic breadth [28], nor fitness effects of interactions. Resurging interest has stimulated new methods for quantifying ecological specialization [1,71,75]; we used the Paired Differences Index.
From field collected samples, we estimated the in situ link strength of each pairwise interaction, P ij , as the proportion of nodules sampled from the i th host species occupied by the j th Bradyrhizobium NifD type: where n i,j = the number of nodules on the i th host species occupied by the j th Bradyrhizobium NifD type and n i = the total number of nodules sampled from the i th host (n ACHE = 19, n ACST = 18, n LUAR = 21, n LUBI = 23).
Observed specificity of each of the four hosts was then measured as the differential frequency of interaction with the 22 different available NifD types, using the Paired Differences Index (PDI) measure [71]: where P i,1 is the strength of the strongest link with the i th host. The PDI values were not normally distributed; therefore, to compare PDIs of the four hosts, we estimated confidence limits from 1000 simulated 20-nodule populations generated as follows. Using the RAND ("TABLE",) function of SAS® 9.3, we specified the frequency with which each genotype appeared in each simulated population, assuming that hosts randomly sampled rhizobia and therefore the observed frequencies of the 22 NifD genotypes in the total sample of 81 nodules represented their availability in the Bradyrhizobium population. From each simulated population, we calculated the joint PDI for all four hosts together: where P j is the link strength averaged across the four hosts of Bradyrhizobium strain j, P 1 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.

Genetic breadth of partners isolated from wild-grown hosts
We constructed a molecular network for each DNA region with the neighbor-net algorithm implemented in SplitsTree 4.8 [119], using the Akaike Information Criterion (AIC) implemented in Modeltest (version 3.4) to choose a suitable model of sequence evolution [120]. The NifD genotype network was estimated using the GTR + I model [121] of sequence evolution and the ITS network was estimated with the HKY + I + G model [122]. 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][124][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 [126].
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 hostspecies pairwise Φ st [127] to test whether pairs of host species harbored genetically distinct groups of bacteria. The Φ st statistic for nucleotide diversity is analogous to Wright's F ST , 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 Φ st values, and the pairwise differences were obtained by permuting genotypes between levels or populations. These analyses were performed with Arlequin 3.0b [128].

Greenhouse experiment Seed source
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 surfacesterilized and scarified [66,129], then germinated individually in wells of 96-well culture plates containing 200 ml of sterile ddH 2 O (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 [66]. One-week old seedlings of L. bicolor were transferred to bleachsterilized 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 nitrogenfree solution [130] until treatments were applied one month after germination (see below).

Preparation of bacterial inocula
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 [66], 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 [78] and incubated at room temperature until bacteria covered the plate. Bacteria were washed from plates into 50-ml polypropylene tubes containing sterile dd H 2 O 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 10 8 cells ml −1 by dilution in sterile dd H 2 O. Each treatment plant received 10 ml of bacterial suspension and control plants received 10 ml of sterile dd H 2 O.

Experimental design
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 [131]. 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][133][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 [138]. 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.

Fundamental association specificity
No analysis was required because all hosts nodulated all Bradyrhizobium genotypes.

Host response specificity
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 file 2: 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 file 3: Table S3) were compared using a Tukey-Kramer test of honestly significant pairwise differences [139].
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 [77], 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 AIC min is the corrected AICc of the best-fitting model and AIC i is the AICc of the model being compared [140]. 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 file 4: 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 file 5: Table S5) and a random effect interaction (Additional file 7: 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 i th genotype in each block: (w iw 0 )/w 0 , where w i = shoot dry weight of the plant inoculated with genotype i and w 0 = 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, 1989NC, -2007.
To quantify specialization using the host fitness responses, we calculated the fitness effect of each Bradyrhizobium genotype on each test host species separately for each of the k = 10 blocks, where w ijk = shoot dry weight of the ith host species inoculated with the jth strain in the kth block, and w iCk = shoot dry weight of the uninoculated conspecific control plant in the same block. The Paired Differences Index (PDI) [71] of the ith test host can then be calculated for each block by ranking within-block values of p ijk from highest (j = 1) to lowest (j = N), where N = 30 tested strains, to obtain: Variance among blocks was used to calculate 95% confidence intervals for the PDI of each test host species.

Specificity of host effect on symbionts
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 [141], pp. 185, 202-204. Standard deviations were not equal among factor levels; so we performed a Welch ANOVA, which allows for heteroscedasticity among factors [141], 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.

Test of association between bacterial genotypes and phenotypes
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 [142].

Fitness feedback between host and symbiont
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.

Availability of supporting data
The sequence datasets supporting the results of this article are available from Genbank (http://www.ncbi.nlm.nih.gov/ genbank/); accession numbers JQ230720 -JQ230882.

Additional files
Additional file 1: Table S1. Title of data: Field-sampled nodules, their source host, identifying symbols, and accession numbers for NifD and ITS sequences submitted to GenBank. Description of data: * host = host species, pln = plant within species, nod = nodule on plant, strain = strain identifier, T-type = inoculation type in greenhouse experiment (blank cells indicate strains not used in that experiment), g'type NifD and g'type ITS identify the NifD and ITS genotypes, respectively, included on Figure 1. Respective NifD and ITS GenBank sequence IDs and accession number are indicated as seq nifD, seq ITS, accession # NifD, and accession # ITS. Blank cells in these columns indicate isolates that did not amplify.
Additional file 3: Table S3. Title of data: Tukey-Kramer honestly significant differences in trait values between Bradyrhizobium genotypes inoculated on two host species, (a) A. strigosus, (b) L. bicolor. Description of data: Genotypes sharing the same capital letter are not statistically significantly different at p < 0.05. * shoot dry weight.
Additional file 4: Table S4. Title of data: Generalized linear mixed models fitted to variance in host shoot dry weight. Description of data: + estimated G-matrix not positive definite. Host = test host species, block = spatial block, strain = Bradyrhizobium genotype, sp_orig = host species from which Bradyrhizobium genotypes were isolated, ge_orig = host genus from which Bradyrhizobium genotypes were isolated. All models assumed a gamma distribution with a logit link function. Model 16 (bold face) provided the best fit (i.e., minimized AICc). Adding the block by host species interaction improved the fit (e (AICmin-AICi)/2 = 5.21 x 10-13) but the proportion of variance explained by this component (4.91% + 2.42%, Additional file 7: Table S7) was only marginally significant. The next best fitting model (model 1) included the fixed effect of host and the random effects of block and Bradyrhizobium genotype, with interactions of host with strain and with block. Residuals plots suggested that this model did not perform much worse than the best model, but nesting Bradyrhizobium genotype into host species from which it was isolated greatly improved the relative likelihood that the model fit the data (model 16 versus model 1; e (AICmin-AICi)/2 = 1.28 × 10-12)). Examination of Q-Q plots revealed that all other models provided much poorer fits to the data.