Bees are a highly diverse and critical group of pollinators, represented by over 20,000 species worldwide1,2, which contribute an estimated value of at least $200 billion in agricultural services per year3. The ongoing and alarmingly pervasive decline in bee populations around the globe thus represents a major ecological and economic issue. Bee health and fitness are strongly affected by the bacteria, fungi, and viruses they are exposed to in their environments;4,5 and there is evidence that both beneficial and harmful microbiota may be regularly transmitted within and between species4,5,6,7,8,9,10. Recent studies exploring hologenomic data consider the genomes of a host species and its microbiome in concert. These works indicate that host species may be continuously coevolving with their microbial communities11. As such, hologenomic data can be used to achieve insights into a host species and its environmental ecology simultaneously12. Accordingly, alongside studies of bee behavior, genetics, ecology, and evolution, research into bee microbiomes has begun to reveal an intimate loop of influence between microbes, bees, and their shared ecosystems12,13,14,15,16.

To date, microbiome studies have primarily focused on the gut microbiomes of economically salient bee groups, such as honey bees (genus Apis) or bumble bees (genus Bombus), applying targeted amplicon sequencing methods to identify bacterial communities and make inferences regarding their influences on their hosts10,17. We now appreciate that the microbiomes of highly social honey bees and other members of the subfamily Apinae (including bumble bees, stingless bees, and orchid bees), collectively termed corbiculate bees, are composed of relatively small suites of bacteria, which nonetheless provide major benefits through facilitating nutrient uptake and immune functions for their hosts18,19,20. Similar research among wild bees—which provide the majority of pollination services21—promises to offer similarly invaluable insights despite remaining in its early stages16. For example, we are gaining an appreciation for the importance of pollen-borne microbes for wild bee development and fitness13,22 and learning that the composition of microbial communities within plant-pollinator networks is likely highly dynamic23.

Ongoing advancements in next generation sequencing methods have allowed for the production of metagenome data—offering massively expanded insights into the environmental elements potentially influencing organismal health and biology12,24,25,26,27. Compared to the more targeted efforts of 16S sequencing, metagenomic datasets capture a wholistic profile of bacteria, fungi, viruses, plants, and other taxa, associated with host organisms at the time of sampling. To date, these methods have been applied primarily to honey bees12,25,26,27,28,29, allowing for fuller characterization of host microbiomes alongside identification of potentially pathogenic elements26,29. Other studies have focused on potential roots of honey bee disease. For example, recent efforts by Galbraith et al.25 identified a suite of bee viruses, including evidence of 127 novel viral contigs, from both managed and wild bee species around the globe; an important finding considering previous metagenomic research revealed that honey bee viruses may be readily circulated by wild bee populations28. These studies also highlight a paucity of metagenomic data from any wild bees beyond the corbiculates, data that promises to help capture a more complete picture of wild bee health and ecology.

One particularly suitable candidate group for this work are the Ceratina small carpenter bees, a well-studied wild bee genus which is globally distributed and features considerable social diversity among its members30,31. All Ceratina form small burrows in the soft pith of woody-stemmed plants (especially of Rubus or Rhus spp.) in which they usually rear a single brood per year31. This form of stem nesting makes the Ceratina highly tractable for empirical study, and three species (C. japonica, C. calcarata, and C. australensis) have already emerged as powerful models for research into behavioral ecology32,33, sociogenomics34,35,36,37, and the combined influences of both nutrition and microbial composition on development and behavior14,38,39. Generation of metagenomic data for these species in particular would provide novel insights into wild bee ecology while greatly advancing the comparative study of bee health overall.

Here, we present the metagenomes of three globally distributed species of Ceratina, C. australensis (Australia;35), C. japonica (Japan;32), and C. calcarata (North America;34), and use these data to address four focal questions: (1) Do metagenomic data characterize host Ceratina species? We hypothesize that metagenomes will be largely unique to hosts. Ceratina likely do not feature sufficient sociality to standardize their microbiome (as seen among corbiculates;20); rather, these three species are likely strongly affected by the microbiota of their local environments7,8,9 and should therefore reflect their highly isolated and distinct environments in their metagenomes. (2) Does local environment drive metagenomic composition among populations of C. australensis across its distribution? We hypothesize that variation among C. australensis metagenomes is directly tied to environmental distinction among populations, as has been seen in previous amplicon sequencing microbiome studies within the species40,41. (3) Does the C. australensis metagenome reveal signals of sociality? Previous research has shown that about 13% of C. australensis females demonstrate a form of cooperative breeding across their distribution42, but that selection on social traits may be highly population-specific and spread across many genetic loci43. We thus hypothesize that any metagenomic influence associated with social phenotypes in the species will be tied to local population rather than consistently distributed across populations. (4) How does the core bacterial microbiome of Ceratina compare to the microbiomes of other bee species drawn from across a globally distributed spectrum of bee families and social forms? We hypothesize that, despite considerable social polymorphism within the group37, our largely solitary Ceratina species are less able to mitigate their exposure to microbial species from their environment. As such, we predict the core Ceratina microbiome will share more elements with those of other solitary and weakly social bees from distantly related families than they will with their sister subfamily of the highly social corbiculate bees20. This study defines the first core carpenter bee microbiome and offers the most comprehensive comparative metagenomic assessment of wild bees to date. As the genomes of the featured taxa have also been recently sequenced34,35,37, the metagenomic data presented here also represents the completion of the first wild bee hologenomes. Overall, this study thus provides an exciting and highly encouraging demonstration of the power of metagenomic methods to yield a wealth of insights into wild bee ecology and health.


Metagenomic data characterizes host species

A total of 221 families (40% of all 556 families identified) were hosted in each of the Ceratina species we measured. This set of 221 families contained prominent core taxa shared across the Ceratina genus, including the bacterial genera Burkholderia, Bacillus, Paenibacillus and Lactobacillus (Data S3). This set also included notable virus families (e.g., Potyviridae, Secoviridae, and Retroviridae) and mite genera (e.g., Varroa).

Host species accounted for most of the detectable variation in relative community abundance overall (Figs. 1, 2; PERMANOVA, R2 ≥ 0.626, p < 0.001) and within all assessed groups except plants at both the family and genus levels (e.g., bacterial families, R2 = 0.8567, p < 0.001; Figs. 2a–f, S3, S4; Data S3, S4). Accordingly, a random-forest classifier (RFC) quickly achieved 100% accuracy binning samples by host species (Figs. 3, S12; Data S5) and performed with greater than 95% accuracy even when trained on just 10% of available sample data. Bacteria made up 17 of the 20 most important families for overall RFC model structure (85%), and included some which were also identified during diversity analysis (e.g., Pectobacteriaceae, Data S6, S7). Some of this set also included those families that were significantly overrepresented in one species or another (negative binomial distribution analysis (NBDA), padj <0.05; Fig. S5; Data S8–S10), such as the fungi Wallemiaceae in C. calcarata over C. australensis and the bacteria Peptococcaceae in C. australensis over both C. calcarata and C. japonica among others. Overall, these WGS-derived metagenomic data reliably and confidently characterized three known host bee species sampled from three highly isolated distributions.

Fig. 1: Microorganismal community co-occurrence network among host Ceratina species.
figure 1

Results of weighted gene co-expression network analysis (WGCNA) showing the topmost strongly positively and significantly correlated modules for each host species (full outputs in Data S9–S11) overlaid on a world map outlining endemic distributions of each host. A selection of bacterial, fungal, or viral hub families, which co-occur extensively with other members of their micro-communities, are highlighted (see legend). The featured Ceratina australensis module (cor = 0.99, p = 7.2e−139) was found to be predominantly composed of bacterial families (53%), compared to the more diverse core communities of C. calcarata (cor = 0.97, p = 8.4e−27; bacteria = 19%) and C. japonica (cor = 1.0, p < 1.0e−200; bacteria = 13%).

Fig. 2: Principal coordinate analyses (PCoA) based on Bray-Curtis dissimilarities between Ceratina species C. australensis, C. japonica, and C. calcarata.
figure 2

Plots display relative abundance distributions of family level data from (a) Bacteria, (b) Fungi, (c) Viruses, (d) Plants, (e) Arachnids, and (f) Nematodes. Host species had a highly significant effect on community variance among all phyla (p < 0.001) except plants (p = 0.09; full results in Data S4).

Fig. 3: Summary comparison of performance accuracy among three separate random-forest classifiers.
figure 3

Random-forest classifiers were individually trained to assign samples either to species, population of origin, or social phenotype using total metagenomic data and between 10 and 90% of samples for training. Overall average classification accuracy by model is presented. While models trained to assign species or population of origin approached 100% accuracy with increasing training set size, metagenomic data were unable to resolve social phenotype with accuracy better than chance. Full results are in Data S5.

Examining community composition, members of bacteria Burkholderia, Pseudomonas, Lactobacillus, and Bacillus were found to be among the largest contributors to variation among all host species (combined average contribution = 22.3%; Data S7). Other notable contributors included the plant families Fabaceae and Chenopodiaceae and viruses Podoviridae and Potyviridae. NBDA identified a total of 286 significantly differentially represented families (DRFs) and 472 differentially represented genera (DRGs) among host bee species (padj < 0.05; Data S8–S10). Notably, at both the family and genus levels, C. australensis featured the greatest numbers of significantly overabundant phyla (DRF = 110; DRG = 221) compared to both C. japonica (DRF = 30; DRG = 45) and C. calcarata (DRF = 22; DRG = 29). Here we focus on results of NBDA at the genus level. Notable significantly overrepresented taxa included the bacterial genus Lactobacillus in C. japonica and Burkholderia in both C. australensis and C. calcarata; fungal genus Saccharomyces in C. japonica and C. australensis; and bacteria genus Serratia in C. calcarata and C. japonica. Additional taxa of notable abundance included the plant genus Nicotiana in C. australensis.

Weighted gene co-expression network analysis (WGCNA) corroborated previous analyses and identified a total of 22 modules capturing co-occurrent community members associated with each host Ceratina species (Fig. 1; Data S11–S13). This analysis further highlighted modules (i.e., communities) which were highly significantly and positively correlated with each species (i.e., C. australensis, cor = 0.99, p = 7.2e−139; C. calcarata, cor = 0.82, p = 1.7e−05; C. japonica, cor = 1.0, p < 1e−200; Data S14). The most strongly C. australensis-associated community was predominantly composed of bacteria (b = 53%) and fungi (f = 22%) with relatively few viruses (v = 8%), and included the Aerococcaceae, Acidobacteriaceae, and Deinococcaceae as hub members. Though C. japonica’s key community featured proportionally far fewer bacteria than C. australensis or C. calcarata overall (b = 9%, f = 27%, v = 13%), hub members included the Cohaesibacteraceae and Helicobacteraceae. Finally, C. calcarata’s relatively balanced mix of bacteria, fungi, and viruses (b = 19%, f = 35%, v = 24%) included the Hafniaceae as strongly co-occurrent hub member. Of 19 hub bacterial taxa in the strongly C. australensis-associated module, 16 were also found to be among the 50 most important taxa for RFC analysis. These families included Acidobacteriaceae, Borreliaceae, Chromatiaceae, Leptospiraceae, and Spiroplasmataceae (Data S5; S14). Of these, Acidobacteriaceae and Borreliaceae were also both found to be significantly overrepresented in C. australensis during NBDA (Data S7).

Despite considerable variation in overall community composition, functional enrichment was highly concordant among host species (Data S15). Bacterial communities in all three bee species were primarily enriched for metabolic (e.g., fructose and mannose metabolism, map00051) and genetic information processing tasks (e.g., protein processing in endoplasmic reticulum, map04141; Nconcordant = 382, 88%; Data S15). Terms uniquely enriched by species included the KEGG pathway response to nicotine (map05033) in C. australensis; circadian rhythm (map04710) in both C. calcarata and C. japonica; and the Toll-like receptor signaling pathway (map04620) in both C. australensis and C. japonica.

Regional environment influences diversity of the C. australensis microbiome

Core bacterial and fungal groups (defined in this study as >50% prevalence and 1% relative abundance, but see also ref. 44) were calculated for each C. australensis population of origin (i.e., Queensland, QLD; Victoria, VIC; and South Australia, SA), revealing 13 bacterial and 22 fungal genera present across sampled populations. Core bacterial taxa included Burkholderia, Bacillus, Acinetobacter, and Paenibacillus; core fungi included Aspergillus, Saccharomyces, and Penicillium (Data S16). These core genera were identified within a total of 383 families (75% of 509 families total) present in all three host C. australensis populations. Other notable taxa identified across populations included several potential plant pathogenic virus families Betaflexiviridae, Luteoviridae, Partitiviridae, Potyviridae, Secoviridae; with families Bromoviridae and Virgaviridae additionally detected in VIC and QLD (Data S17).

Populations had significant differences in overall community composition (PERMANOVA, R2 = 0.148, p = 0.007; Fig. 4c, d; Data S18) and among phyla (PERMANOVA, p ≤ 0.005; Figs. 5, S6, S7). Post hoc testing revealed significant differences in community composition specifically between QLD and SA (Tukey, padj = 0.00041) as well as QLD and VIC (Tukey, padj = 0.0323; Fig. 5e; Data S18). RFC corroborated multivariate analyses, which supported QLD as the most distinct metagenomic population, followed by SA, and then VIC (Figs. 4b, S12; Data S5, S6). The most informative families for RFC accuracy included bacteria (e.g., Clostridiaceae, Cellulomonadaceae), fungi (e.g., Agaricaceae, Hypocreaceae), and viruses (Mimiviridae, Peribunyaviridae), some of which were also identified during diversity analysis (e.g., Clostridiaceae; Mimiviridae; Data S5; S19). Of these highly informative families, the Clostridiaceae and Agaricaceae were found to be overrepresented in QLD and VIC compared to SA; the Mimiviridae and Peribunyaviridae in QLD compared to both VIC and SA; and the Hypocreaceae in SA compared to QLD (Data S20).

Fig. 4: Microorganismal community composition by C. australensis population.
figure 4

a A selection of highly differentially abundant genera (Log2FC > 2.0) detected among sampled populations. Bacterial families accounted for the greatest number of highly abundant taxa cross populations. b Classification accuracy of a random-forest model trained to assign population of origin to samples as a function of the size of the training set. All trials performed with significantly greater accuracy than would an estimated random draw (p < 0.05). While total model accuracy approached 100% with inclusion of 40+ samples, performance was consistently better for Queensland and South Australia compared to Victoria. c and d) PCoA plots are shown for total metagenomic abundance data analyzed by family (c) and genus (d) across C. australensis populations of origin. Ordination was comparable at both levels of resolution; and PERMANOVA revealed that overall community composition was significantly distinguished by population of origin; greatest differences were between Queensland and both South Australia and Victoria.

Fig. 5: Principal coordinate analyses (PCoA) based on Bray-Curtis dissimilarities between three focal Ceratina australensis populations: Queensland (QLD), Victoria (Victoria), and South Australia (SA).
figure 5

Plots display relative abundance distributions of family level data from (a) Bacteria, (b) Fungi, (c) Viruses, (d) Plants, (e) Arachnids, and (f) Nematodes. Population of origin had a highly significant effect on community variance among all phyla (p < 0.001); in cases of heterogeneous dispersion, significantly distinct populations are indicated (a, e, f; full results in Data S18).

Each population had distinct taxonomic diversity and community composition (Data S19). Members of bacterial genus Burkholderia, accounted for the greatest percentage of dissimilarity among any two populations (~5%), with slightly higher average abundances in SA over VIC and VIC over QLD. NBDA corroborated diversity analyses and identified a total of 153 significantly DRFs and 191 DRGs among C. australensis populations (adjusted p < 0.05; Data S20, S21) with a majority of uniquely overrepresented taxa identified in QLD (DAF = 57; DAG = 62) followed by SA (DAF = 37; DAG = 58) and VIC (DAF = 17; DAG = 14). These results also corroborate RFC performance by reinforcing that QLD may be the most distinct of the three populations. By contrast, SA featured more than twice as many strongly overrepresented genera as QLD or VIC (Nlog2FC>2 = 78, Fig. 4a), including Lactobacillus, Nicotiana, and Polerovirus.

Although there were some notable variations in community composition across the three C. australensis populations, functional enrichment was again largely concordant across locations (Data S22). All three populations shared KEGG enrichment for metabolic processes (e.g., fructose, map00051; tyrosine, map00350), environmental information processing (e.g., quorum sensing, map02024), and protein signaling (e.g., oxidative phosphorylation, map00190; Data S22). Unique KEGG pathway enrichment for each population included B-cell receptor signaling (map04662) in QLD, nicotine addiction (map05033) in SA, and Toll-like receptor signaling (map04620) in VIC.

Metagenomic signals of sociality in C. australensis

Comparisons of solitary to social nests across populations revealed no universal effect of C. australensis sociality on metagenomic diversity (e.g., bacteria; PERMANOVA, R2 = 0.022, p = 0.379; Figs. S8, S9; Data S23) or relative abundance across populations (Data S24, S25). Reassessing the effects of sociality within each population revealed evidence of significant variations in diversity and abundance at both the family and genus levels across all phyla (Figs. S10, S11; Data S26–S28), including overrepresentation of Acetobacter in solitary over social hosts (Log2FC = 37.18, p = 6.35E−06) and both Streptococcus and Fusobacterium in social over solitary in the Victoria population (Data S28). These differences in overall community composition were not uniform, however; the bacterial genera Bombella and Gluconobacter, for example, were overrepresented in the social hosts of SA but also in the solitary nest individuals of VIC (Data S28). Lastly, a RFC was unable to ever perform significantly better than a simulated random draw, and failed to surpass 50% accuracy in assigning samples to their correct social or solitary bins across all populations (Figs. 3, S12; Data S5).

Core wild bee microbiomes in comparison with corbiculates

Comparison of bacterial community composition among 38 bee species (Data S29, S30) revealed greater overall similarity between the Ceratina (Xylocopinae) and either the leafcutter (Megachilidae) or sweat bees (Halictidae) than with corbiculate bees from their sister subfamily Apinae (Fig. 6). Although the Ceratina are collectively host to a richly diverse suite of bacteria, their core microbial community is primarily composed of Burkholderia (average relative abundance = 22.3%), Pseudomonas (4.6%), Bacillus (4.5%), and Acinetobacter (2.1%). Genera from this set, such as the Pseudomonas, are effectively undetected among corbiculates, but occur with high relative abundance among both the Megachilidae (e.g., Osmia bicornis, 5.4%) and Halictidae (e.g., Augochlora pura, 5.1% and Megalopta genalis, 14.5%). Other members of the Ceratina microbiome such as Escherichia and Erwinia, which represent prevalent but low abundance organisms, were also detected in the Megachilidae and Halictidae. Only the highly diverse bacterial genus Lactobacillus was detected with relatively high prevalence and abundance across all major bee lineages and in every host species considered with notable exceptions in the oil digger bees (Centris atripes and Anthophora abrupta) and the stingless bee, Tetragonula fuscobalteata (20 Data S29, S30).

Fig. 6: Highlighted selection of some of the most prevalent and abundant bacterial genera identified among bee lineages.
figure 6

The Ceratina core microbial genera (>50% prevalence and >1% relative abundance) are also found in other solitary and socially polymorphic bees. The previously identified corbiculate core microbiome is largely limited to that group. Other key bacterial genera may have their origins in the environment, and may be commensal in lineages with less well-established intranidal environments. Only the highly diverse Lactobacillus genus has been identified in almost all bee lineages studied to date (full dataset for all species can be found in Data S30).


Here we present a first comparative metagenomic analysis of three wild bee species drawn from three continents—Australia, North America, and Asia. We consider the implications of variations in microbiome among host species, investigate how differences in C. australensis population of origin may reflect on species ecology and sociobiology across its range in Australia, and compare the first-ever core microbiome among the carpenter bees to similar datasets among other bee lineages.

The wild bee microbiome is diverse and unique to host species

Carpenter bee host species had distinct metagenomic signatures which were clearly detectable by multivariate, clustering, and machine learning algorithms. Most notably, RFC resolved host species with >95% accuracy even when trained on just 10% of available data, suggesting a very clear and strong signal in taxonomic diversity and abundance data. A total of 43 of the top 50 most informative taxa for host bee species classification accuracy were bacterial families including the Acetobacteraceae and Acidobacteriaceae. Acidobacteriaceae was additionally found to be a hub taxon within the strongly C. australensis-associated module of co-occurrent taxa, and significantly overrepresented in C. australensis compared to both C. japonica and C. calcarata. These results collectively suggest that members of Acidobacteriaceae may feature centrally in C. australensis biology. Both Acetobacteraceae and Acidobacteriaceae contain acidophilic genera and have been detected in other bees as possible mutualists taken up from the environment, often through pollen provisions7. Although the composition, nutritional value, and microbial associations of pollen provisions have been assessed among Ceratina38,39,45, including C. australensis41, additional investigations could offer finer scale insights into the combined influences of floral resources and intranidal conditions on microbiome composition and fitness in this group.

Notably, all of C. australensis’s hub taxa were bacterial families and included those found to play important roles in other bee lineages (e.g., Prevotellaceae in Apis florea;46 Caulobacteraceae in Osmia;47). By contrast, only two of C. japonica’s 14 hub taxa were bacterial families (Cohaesibacteraceae and Helicobacteraceae), with the remainder including wood-rotting fungal families (Grifolaceae and Punctulariaceae) and plant viral families (e.g., Tospoviridae); and similarly, of C. calcarata’s five hub taxa, Hafniaceae was the only bacterial family. Although each of these hub taxa were also found to be significantly overrepresented in their host species, future studies with extended sampling of C. japonica or C. calcarata would better clarify the degree to which these families play roles of considerable biological importance to their hosts.

Despite considerable variation by host species, functional enrichment across our three Ceratina species was found to be largely uniform, arguably highlighting a core profile of microbial activity. In general, bacterial families detected in each of our Ceratina hosts likely contribute to some combination of improving nutrient uptake for their hosts (e.g., via carbohydrate metabolism) and performing tasks associated with immune response (e.g., Toll signaling pathway; T- and B-cell receptor signaling). There is evidence, primarily in honey bees, that gut microbes may strongly aid in host immune and metabolic pathways15,16,18. Our results reinforce previous assessments in suggesting that the microbiomes of solitary bees may also contribute to essential biological processes for their hosts, regardless of variations in ecology, climate, or even overarching community composition47,48. Future studies that incorporate RNAseq methods alongside metagenomic analyses should be able to further tease apart variations in microbial diversity and function among host bee species.

Wild bee microbiomes may be strongly influenced by local environment

We found that the C. australensis microbiome varies significantly across its range, most likely attributable to both physical isolation and ecological variation. In our study, Queensland not only featured the most unique and significantly overrepresented taxa but also the most distinct metagenomic profile across all analyses. Our metagenomic data are thus broadly consistent with former amplicon, microsatellite, and genome-wide datasets used in previous C. australensis studies (e.g.,40,41,43.) which found similar dimensions of structure among these populations. Gradually dispersing from an origin population in Queensland49, C. australensis is thought to have spread south and west following the Murray River to reach its current distribution40. This trajectory suggests that any substantial variation in metagenomic profiles between Queensland and either South Australia or Victoria should reveal relatively novel elements associated with those environments. Comparing variation in microbiomes by population, the bacterial genus Pantoea was found uniquely in the Queensland core, Streptococcus uniquely in the Victoria core, and both Flavobacterium and Desulfovibrio uniquely within the South Australia core. Pantoea is a highly diverse genus and well-studied plant pathogen which can form mutualistic and even commensal associations with insect hosts (e.g., leafcutter ants,50,51. Further, members of Pantoea may be readily vectored among plant hosts by honey bees52 and have even been identified in the guts of Australian stingless bees (Apidae: Meliponini) sampled in Queensland53. Detection of Pantoea in the Queensland C. australensis core suggests it may also be vectored by this species. Streptococcus bacteria are a widely pathogenic group, well-studied in honey bees54 and known in Australian honey bee populations55. Widespread detection of Streptococcus and Enterococcus (formerly classified as Streptococcus), especially in Victorian C. australensis, may indicate a pervasive—if ultimately non-lethal—bacterial challenge for that population. Finally, members of Flavobacterium56 and Desulfovibrio57 are both known to favor saline, marine environments. Although their biological role in C. australensis remains unknown, their strong representation in the South Australian population—sampled exclusively from beach dunes along the Great Australian Bight—helps to illustrate the degree to which the C. australensis microbiome may be directly influenced by regional environment.

Turning to insights regarding C. australensis’s diet, although all populations appear to be closely associated with Solanum (nightshades), Vigna (legumes), and Gossypium (mallows, including cotton), we did find evidence in support of previous studies41 that pollen usage varies considerably across C. australensis’s range. For example, though South Australian C. australensis include more Brassica (canola and mustard) and Chenopodium (goosefoots) than Queensland, northeastern bees are associated with more Medicago (legumes), Glycine (soybeans), and Sorghum (cereals). Functional enrichment offers some insights into floral resources among populations, such as our detecting evidence of nicotine among C. australensis in South Australia. Secondary floral compounds such as nicotine or other plant alkaloids can strongly influence foraging preferences in honey bees58, bumble bees59, and among other wild bee species60. Although the degree to which such compounds might similarly affect Ceratina remains an open question, plant genus Nicotiana was accordingly detected among South Australian C. australensis significantly more than in Queensland. Notably, the distribution of several known Australian members of Nicotiana, including wild tobacco, overlaps closely with that of C. australensis in both South Australia and Victoria61.

Unique variations aside, we found that all three C. australensis populations share most of their core bacterial and fungal genera, a result which echoes previous microbial metabarcoding studies38. Of special note, bacterial genus Burkholderia featured prominently as the most abundant group of each population. Burkholderia are distinguished as one of the most common environmental bacterial genera62, and its members are known to have an expanded range across Australia63. As discussed further below, Burkholderia are known to form mutualistic relationships with some insect hosts64, but it remains unclear what their role may be in C. australensis and among the Ceratina globally. Among the best represented core fungal genera were Aspergillus, Fusarium and Saccharomyces. Depending on species, members of each of these fungal genera may be helpful or harmful as reported from honey bee hosts in which they have primarily been studied to date (e.g.,65,66,67.). This is the first detection of these genera in C. australensis41, and an investigation into whether they may play a commensal or harmful role across its populations, particularly during a time of dynamic climatic change68, remains an important target for future studies.

These data also provide preliminary assessments into the degree to which C. australensis may be vectoring plant diseases in its environment, including those which may be harmful to economically valuable crop species (e.g., plum pox and Prunus69,). We identified seven major plant viral families among our sampled C. australensis populations that had been previously found only in Australian honey bee populations70. Our dataset of viral families—which includes the Secoviridae, Potyviridae, and Luteviridae among others—also suggests these groups may be more widespread than detected by Roberts70. Intriguingly, abundances of these viral families follow similar trends over the landscape in both studies where direct comparisons between sampling areas could be made, with overall highest loads identified among South Australian bees, and lowest among those from Queensland. Notably, both studies also found evidence of Virgiviridae: Tobamovirus, a genus of the very recently established cucumber green mottle mosaic virus, primarily in Queensland.

Metabarcoding studies suggest that the diets of both introduced honey bees and Australia’s wild bee species may overlap considerably71, potentially branching into crop variety. The use of even managed bees as a means of plant pathogen surveillance is of great value, but still very much in its infancy70,72. To the best of our knowledge, our study represents a critical first demonstration that wild bees can also offer invaluable insights into the spread of both established and newly introduced crop diseases. Metagenomic screening of C. australensis and other wild bee populations, which benefit from robust and ever improving reference datasets on NCBI, thus offers ecologically comprehensive insights into both managed and unmanaged landscapes in Australia and elsewhere.

There are detectable metagenomic signals of sociality within C. australensis populations

Previous research has shown that around 13% of C. australensis females consistently demonstrate cooperative breeding across populations in the species, a strategy which is thought to be advantageous under heavy parasite pressure42,73. Metagenomic data indicate that C. australensis populations do experience significantly variable environmental factors across their distribution and, intriguingly, the metagenomes of solitary and social individuals appear to differ significantly within populations. For example, among South Australian C. australensis, the bacterial family Thermomonosporaceae is very well represented in solitary bees, while both Brucellaceae and Prolixibacteraceae are found in social bees. These data suggest there is significant co-variation between bacterial community and host bee sociality. A recent genome-wide association study in C. australensis found that many genetic loci may be tied to social phenotypes, but each locus may have only a relatively small influence43. Signatures of selection by sociality within each C. australensis population detected in that study are concordant with our metagenomic results overall, suggesting wild bees more strongly experience population-level rather than species-level differentiation among social forms43,74.

Social complexity is inversely proportional to microbiome diversity across bees

The carpenter bee core bacterial microbiome presented here represents a first metagenomic profile for non-corbiculate wild bees, a valuable point of comparison outside of the more socially complex bees (i.e., corbiculates) in which most metagenomic and microbiological work has been done to date20,75,76. The carpenter bee core is comprised of 11 bacterial genera, predominantly Burkholderia (22.3%) followed by Pseudomonas (4.6%) and Bacillus (4.5%); all three genera play important roles across insects64,77. In particular, members of the Burkholderia have been found to act as symbionts for a variety of insect hosts in which they have been studied (e.g., the bean bug, Riptortus pedestris78), performing a wide suite of beneficial functions (e.g., nutritional supplementation;64). Among bees, Burkholderia and Pseudomonas are thought to be acquired from the environment (e.g., soils and plants) and have been detected primarily among solitary species;7,9,79 though Pseudomonas may occasionally be present in some corbiculates80,81. Accordingly, we found that the largely solitary carpenter bee core shares almost no microbial members with their highly social sister corbiculate bees from the subfamily Apinae, which features a relatively small core set (Fig. 6). Unlike the corbiculate bees, the nest structure, colony size and social environments of the Xylocopinae, Megachilidae, and Halictidae are relatively small, solitary and exposed to the environment (e.g., soil or decaying branches). Observed consistencies in the composition of small carpenter, leafcutter, and sweat bee microbiomes therefore appears to highlight the importance of the physical rather than social environments in establishing the microbial communities of those host species48,82. Lactobacillus was the only bacterial genus from the carpenter bee core to be detected in all bee species measured, often at comparatively high relative abundances. The Lactobacillus genus is highly diverse, and its members have established as symbionts across a similarly diverse range of both vertebrate83,84 and invertebrate hosts85. Lactobacillus are known to be highly beneficial in honey bees, in which they have been extensively studied, often offsetting the deleterious effects of honey bee diseases like chalkbrood86,87. Species and strains of Lactobacillus very likely play similarly commensal roles in other bee hosts (e.g., Megachilidae88; Nomia melanderi82) and, alongside many other microbes, are critical to the diet, fitness and health of many bees13,89.

Previous research suggests that even distantly related bee lineages experience similar evolutionary dynamics as a product of their social complexity37,74. Although microbial composition may be closely tied to host species, lineage sociality and individual behavioral caste also appear to play critical roles in shaping community structure20,67. The degree to which there may be any consistent bidirectional influence between microbial community composition and species sociality across major bee lineages remains an open question for future hologenomic research. For example, do the microbiomes of obligately eusocial bees outside of Apinae (e.g., Exoneurella tridentata90) more closely resemble those of other members of their subfamily (Xylcopinae) or those of other eusocial bees (e.g., Apinae) despite phylogenetic distance? Future studies which sample microbiomes from consistent life stages and tissue types across a comparably wide range of bee families may yield refined insights. We are on the cusp of tackling these and other questions as metagenomic methods continue to improve91 and comparable metagenomic studies are performed in bees across independent origins of sociality92,93. These works promise to provide additional datasets that will be invaluable for further illuminating comparisons and that should contribute to a progressively well-defined spectrum of environmental through intranidal origins of bee metagenomes.


Sampling and sequencing

We collected a total of 60 bee samples over 3 years from three species of Ceratina. The sample set of 51 adult female Ceratina australensis were collected in January 2016 from three distinct populations in Australia: Queensland (N = 18; 28.24°S, 152.09°E), Victoria (N = 13; 34.15°S, 142.16°E), and South Australia (N = 20; 34.94°S, 138.50°E). Six females of C. japonica were collected in Sapporo, Japan (N = 6; 43.06°N, 141.35°E) in July 2015, and three females of C. calcarata were collected in Durham, New Hampshire in July 2017 (N = 3; 43.14°N, 70.94°W). All bees were flash frozen in liquid nitrogen during collection to preserve DNA integrity prior to whole body genomic DNA extraction via phenol-chloroform protocol94. Samples were then submitted to Genome Quebec for PCR-free library construction (NEB Ultra II kit) and Illumina HiSeq 2500 (125 PE) sequencing at an average depth of 30M reads per sample.

Sequence data processing

Raw Illumina reads of whole-genome data from C. australensis, C. japonica, and C. calcarata were quality checked with FastQC ( Adapters were removed and reads were cleaned with Trimmomatic95, with the following settings: TRAILING:20 SLIDINGWINDOW:4:20 MINLEN:36. The cleaned read pairs were mapped to the appropriate reference genome for each species with bwa mem using default settings96, and unmapped reads were extracted from the resulting bam files with samtools (97 Data S1). An average of 93% of C. australensis reads mapped to the C. australensis reference genome (NCBI: ASM430768v1;35); 96% of C. calcarata reads mapped to the C. calcarata reference genome (NCBI: ASM165200v1;34); and 87% of C. japonica reads mapped to the C. japonica reference genome (NCBI PRJNA413373;37). Reads were considered unmapped if at least one read from the pair did not map to the reference genome; and paired unmapped reads, in fastq format, were used for further analysis. All sequencing data generated for this study can be accessed via NCBI PRJNA407923 and were handled as compositional data during analysis98.

Taxonomic classification

Taxonomic classification of reads unmapped to the reference genome was carried out with the use of Kraken2 software99. Paired reads were classified with default settings against the latest available nt—NCBI non-redundant nucleotide database. Read counts were used to calculate relative abundance at the family level separately for each of six major biological groups, including: bacteria, fungi, plants, arachnids, nematodes, and viruses. Relative abundance of each family was calculated by dividing number of reads classified to family by the total number of reads classified to a particular group among samples. We then repeated this analysis at the level of genus.

We tested several filtering strategies to ensure low rate of false positive classifications without substantial loss of information. The performance of the following thresholds was assessed: (i) 1% minimum relative abundance within a given sample, (ii) 0.1% minimum relative abundance within a sample, (iii) 0.1% minimum average relative abundance across all samples per species, (iv) minimum number of 10 classified reads (Fig. S1). To explore the effect of the library size on classified taxa and their relative abundance we performed the above-described taxonomic classification for several levels of input reads subsampling, including 0.05–1.25 M paired reads, to construct the rarefaction curve (Fig. S2). All samples were found to be within the rarefied range, and a filter threshold of 0.1% was selected for further analyses.

We also performed taxonomic classification using an assembly-based approach. First, we merged the reads from all C. australensis individuals. Then, we ran the assembly using metaSPAdes100 with paired-end library type and default settings, including k: 21, 33, 55. We then excluded contigs shorter than 300 bp, and classified those remaining based on best BLAST hits101 to the nucleotide (nt) database. We then used bwa mem with default settings to map the reads from each sample against the assembly contigs. We recovered the mapped reads with samtools and assigned each to the taxonomic classification of the contig to which they mapped. Both metaSPAdes and Kraken2 taxonomic assignments were largely concordant with results of BLASTn runs (using default parameters, with max target seqs set to 1 and minimum shared ID set to 70%; Table S2), so we used Kraken2 assignments for further analyses.

Taxonomic diversity and dissimilarity analyses

Bray-Curtis dissimilarity matrices were calculated for each taxonomic group and assessed via principal coordinates analysis (PCoA) and non-metric multi-dimensional scaling (NMDS) analysis in the R package Vegan102. We applied PCoA and associated analyses in this instance as these tests are better reveal qualitative differences among sets, rather than a principal components analysis which better assesses quantitative similarities. To test whether species or population of origin had a significant effect on variation in community composition we then performed permutational multivariate analyses (PERMANOVA) via the ADONIS function in Vegan. To ensure that the assumptions of homogeneity of group dispersions was met, we ran BETADISPER and assessed the resulting distances to the group centroids via ANOVA followed by post hoc Tukey’s range test to determine whether there were significant pairwise differences among groups. Next, we performed similarity percentage (SIMPER) analyses within PAST v 4.06103 to identify which taxa were primarily responsible for observed differences between groups.

Functional analysis

We performed functional analysis of reads classified as of bacterial or fungal origin by the taxid information in the Kraken2 classification output. These were then aligned to the nr—NCBI non-redundant protein database using DIAMOND104 with default fast mode, filtering for e-value of 1e−10 and returning maximum five alignments per read. As DIAMOND does not support paired-end mode we performed functional analysis based on forward reads only. We checked taxonomy of the resulting alignments and excluded from further analyses reads that did not have bacterial or fungal hits among the five alignments. Then we used FragGeneScan105 to find genes in the reads, and analyzed the resulting amino acid fasta files with GhostKOALA106 and eggNOG-mapper107 to characterize gene functions based on KEGG identifiers and the reconstructed pathways.

Negative binomial distribution analysis  + weighted gene co-expression network analysis

As usefully applied in other recent metagenomic works (e.g., Rothman et al.17; Kapheim et al.82), significantly differentially abundant phyla were characterized using NBDA to determine DRFs and differentially represented genera (DRGs) (DESeq2108), comparing reads by Ceratina species (i.e., C. japonica vs. C. calcarata vs. C. australensis), C. australensis populations (i.e., Queensland vs. Victoria vs. South Australia), and C. australensis social phenotype across (i.e., solitary vs. social) and within populations (e.g., Queensland social vs. Queensland solitary). We then used weighted gene co-expression network analysis (WGCNA109) to further assess microbiome communality (i.e., co-occurrence) among Ceratina species. WGCNA is a powerful network analysis option, and one which lends itself to illuminating visualization for figures downstream. Following standard protocol to prepare for this analysis110, normalized read data were filtered to remove any taxa that featured too few or no read values in at least one sample set, and any samples that appeared as clear outliers following hierarchical clustering. This preparatory step indicated relatively weak distinctions among both population and sociality data, so only analysis by host species was completed. Sample SJ19 was removed as a group outlier, and samples were then assigned biological trait data (i.e., host species). We selected a soft power of 6, which indicated explanation of well over 90% of the data, and performed the remainder of network analysis following standard protocols. Taxa which were assigned both trait significance and module membership values >0.90 were considered “Hub” taxa, of especially high association with host species and high co-occurrence among other members of their module (i.e., community).

Random-forest classifiers

We first trained three individual RFCs in R (package randomForest111) to assess the degree to which metagenomic communities effectively predicted Ceratina host species, and population of origin and sociality across populations within C. australensis. We then trained a fourth RFC to assess sociality as partitioned by population (6 bins) and then established three additional RFCs to test sociality within each population (2 bins per RFC). We trained each RFC using total metagenomic read data (i.e., all phyla) and eight training set sizes (between 10 and 90% of available samples). We then evaluated classification accuracy on the withheld 90 to 10% remaining samples five times a piece, for a total of forty trials per RFC. Prior to each run, we applied the tuneRF function to determine an mtry value (used to facilitate forest creation during the RFC run) which would minimize error. R package caret112 was then used to produce a confusion matrix to assess overall RFC performance in accuracy, sensitivity, and specificity, as well as statistical support. As models of significant and robust accuracy (e.g., >80%) featured high average specificity and sensitivity scores (e.g., >80%) we focused on reporting overall accuracy and significance in the main text. Post hoc analyses were performed using the R package RandomForestExplainer113 to evaluate the overall degree to which each taxon influenced RFC model accuracy (i.e., individual importance).

Comparative analysis

To determine the variation in microbial community composition across bee lineages, we first characterized the core bacterial community of our three Ceratina species as those genera which were present in more than 50% of samples and featured an average relative abundance >1%14. We then compared the composition and relative abundances of all bacterial genera detected among our Ceratina species to similar datasets from 35 additional bee species (Data S29, S30). Prior to comparison, we consolidated bacterial composition and count data from these additional datasets to the level of genus before recalculating relative abundance and prevalence.

Reporting summary

Further information on research design is available in the Nature Research Reporting Summary linked to this article.