A network-based meta-analysis of the honeybee gut microbiome: geographic, seasonal, and pesticide-associated shifts

Abstract

The honeybee (Apis mellifera) gut microbiome is essential for pollinator health, yet its functional responses to environmental stressors remain poorly understood. We conducted a global meta-analysis of honeybee gut microbiomes from seven geographic regions, including newly generated data from Armenia, and applied a novel co-abundance network approach, tsantsR, to uncover community-level co-abundance patterns and functional adaptations. We identified a conserved core of six and ten phylotypes in 16 S and WGS datasets, respectively. In Armenia, seasonal and environmental factors, such as urbanization, were linked to compositional shifts, including higher relative abundance of Commensalibacter in autumn and of Bombilactobacillus in urban colonies. Analysis of pesticide and dietary exposure revealed distinct microbial responses. Oxalic acid and neonicotinoid exposure were associated with shifts in opportunistic pathogens, including Klebsiella, Hafnia-Obesumbacterium, and Serratia. In contrast, the herbicide glyphosate was linked to a potential adaptive response characterized by disruption of core taxa such as Snodgrassella alvi and upregulation of pathways hypothesized to be involved in glyphosate degradation, primarily linked to enrichment of Pseudomonas.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1038/s41598-026-62172-4.

Introduction

The Western honey bee (Apis mellifera) is a key pollinator^1,2. In recent years, honeybee populations have declined, with annual losses reaching 30–45%^3–6. Several factors have been linked to this decline, including climate change⁶, heavy use of pesticides⁷, use of antibiotics⁸, and pathogen infection⁹, but no single factor has yet been implicated.

The gut microbiome plays a vital role in bee health by regulating metabolism, pollen digestion, immunity, development, and behavior¹⁰. The core microbiome of adult workers comprises highly specialized bacteria. Five consistently reported phylotypes include Gilliamella, Snodgrassella, Bombilactobacillus (a.k.a Lactobacillus Firm-4), Lactobacillus (a.k.a. Lactobacillus Firm-5), and Bifidobacterium^10–13. Other genera, e.g. Frischella, Bartonella, Commensalibacter, and Bombella, are also common^11,14,15, while Apilactobacillus kunkeei, and Apibacter are reported at low abundance¹⁶.

Environmental factors and health status affect the honeybee gut microbiome, often leading to increased mortality^4,17. In particular, pesticides can disrupt the microbiota, altering behavior and immune function^18–20. Neonicotinoid insecticides, e.g. thiacloprid, clothianidin, and imidacloprid, not only directly impair the nervous system affecting honeybee behavior, but may also disrupt the gut microbiome^21–24. Oxalic acid, an acaricide commonly used against Varroa parasites and considered relatively safe for bees²⁵, can nonetheless lower the gut pH and alter the microbiome composition at high doses¹⁸. Glyphosate, the most widely used herbicide, affects honeybee behavior and physiology²⁶. In plants, it targets the EPSPS enzyme in the shikimate pathway, crucial for the synthesis of aromatic amino acids. Since many bacteria possess glyphosate-sensitive class I EPSPS enzymes, concerns exist about its adverse impact on the bee gut microbiome^13,27–29. Among the glyphosate-affected species, Snodgrassella alvi is most consistently reported to decrease in relative abundance. However, the overall impact of glyphosate on the gut microbiome remains inconclusive^4,28,29.

Overall, conflicting evidence remains regarding how pesticides affect the taxonomic composition and functional potential of the honeybee gut microbiome. For example, studies on neonicotinoids yield contradictory results, some report negligible impacts in laboratory conditions, while others document significant shifts after long-term field exposure^24,30. Similarly, while oxalic acid is linked to an increase in relative abundance of Gilliamella, Bifidobacteria, Lactobacillus, and Snodgrassella in vitro, these effects are not consistently observed in vivo¹⁸. The impact of glyphosate also varies considerably with dose, exposure duration, and the bee life stage^28,31,32. Importantly, these studies do not fully explain the functional consequences or the adaptation mechanisms underlying these shifts.

Here, we performed a meta-analysis by aggregating honeybee gut microbiome datasets from eight published studies across six geographic regions and incorporated newly generated data from Armenia, an understudied region in apicultural research (Fig. 1). We employed a unified bioinformatics workflow, thereby reducing technical variability and enabling robust comparisons of community structure. We introduced a novel co-abundance network analysis to identify exposure-associated bacterial communities and elucidate their putative functional roles. By extending beyond single-taxon analyses, our network-based approach captures coordinated microbial responses and potential functional adaptations to these environmental pressures.

Fig. 1.

Schematic representation of the data collection and analysis workflow.

Results

Dataset collection

We analyzed 516 gut microbiome samples from adult worker honeybees, comprising 16 S and WGS data from eight published studies from Uruguay, Denmark, New Zealand, Southern US, China, and the UK, and one newly generated 16 S dataset from Armenia under different exposure conditions (Fig. 2; Table 1).

Fig. 2.

Geographic distribution and sample composition of studies included in the meta-analysis. The map shows study locations and the number of samples per exposure (red) and control (green) group. Colors indicate the sampling location: (AM: Armenia, DN: Denmark, NZ: New Zealand, US: ST: Southern US, CH: China, UK: United Kingdom). Shapes indicate whether the data are from 16 S or WGS sequencing.

Table 1.

Summary of dataset features included in this study.

Dataset ID	Data description	Sample type A	Exposure	Bee source	Exposure/condition	# Samples
AM	16 S, V3-V4	Entire gut	Untreated	Apiary	Untreated	21
UR4	16 S, V3-V4	Entire gut	Herbicide	Lab	GlyphosateB	19 (10)
DN18	16 S, V4	Entire gut	Insecticide	Apiary	Acetamiprid / Thiacloprid / Oxalic acidC	182 (39/35/36)
US: TX-1619	16 S, V4	Crop-Rectum	Insecticide	Apiary	ImidaclopridB	69 (34)
US: TX-1728	16 S, V4	Entire gut	Herbicide	Apiary/Lab	GlyphosateB	84 (59)
US: ST33	16 S, V4	Entire gut	Untreated	Apiary	Untreated	44
NZ17	16 S, V3-V4	Crop-Rectum	Supplement	Apiary	Sugars (Dihydroxyacetone / Manuka honey / Invert sugar) or antibacterial (Methylglyoxal)C	53 (8/16/9/8)
CH34	WGS	Entire gut	Untreated	Apiary	Untreated	32
UK35	WGS	Entire gut	Untreated	Apiary	Untreated	23

^A We exclusively focused on subjects that included all possible gut sites where bacteria can reside. ^B Using sucrose as a control. ^C Using untreated and sucrose-treated samples as controls.

Honeybee gut microbiota is composed of a stable core with region-dependent abundance patterns

We first examined the most abundant taxa in untreated (field control) honeybee gut microbiomes using 16 S datasets from Armenia, Denmark, New Zealand, and the southern US, along with WGS datasets from China and UK. Consistent with prior research, the ten most abundant genera across all samples (both 16 S and WGS) included: Lactobacillus (formerly Lactobacillus Firm-5) and Bombilactobacillus (formerly Firm-4), Gilliamella, Snodgrassella, Frischella, Bartonella, Bombella, Commensalibacter, Klebsiella, and Bifidobacterium^36,37(Fig. 3A).

Fig. 3.

Comparison of honeybee gut microbiome composition across studies. (A) Relative abundances of the 10 most abundant bacterial genera in field control samples, with median relative abundances indicated for each study (AM: Armenia, DN: Denmark, NZ: New Zealand, US: ST: Southern US, CH: China, UK: United Kingdom). Bar plots reflect composition among the retained ASVs. (B) Prevalence and median relative abundance of core genera (present in > 80% of field control samples) across studies. Asterisks indicate two WGS-only candidate core genera, Moraxella and Vibrio, identified by Kraken 2 but not supported by genome-wide read mapping (Table S14). (C) Prevalence of species-resolved core taxa across all datasets. (D) Alpha diversity of the honeybee gut microbiome in field control samples, measured using Shannon index; asterisks indicate significant pairwise differences (Kruskal-Wallis post-hoc test, p < 0.05). The central line indicates the median, the box represents the interquartile range (25th − 75th percentiles), and whiskers extend to the 10th and 90th. (E) Principal coordinate analysis (PCoA) based on Bray-Curtis distances, with samples colored by study; between-group differences were assessed using PERMANOVA. Arrows denote the genera discriminating community composition (multiple regression test with permutations, p < 0.05).

The microbiome composition showed weak variation within and between colonies (ANOSIM, R² = 0.183, p = 0.024), yet varied significantly between states within studies (ANOSIM, R² = 0.246, p = 0.002) and among studies (PERMANOVA, ω² = 0.295, p = 0.001). In 16 S datasets, Denmark and the southern US exhibited higher relative abundances of Gilliamella (26% and 24% median relative abundance) and Snodgrassella (19% and 16%), while Bombilactobacillus dominated in Armenia and New Zealand (21% and 22%). Armenia also showed a higher relative abundance of Bifidobacterium. Notably, WGS datasets were dominated by Bartonella (17%-31%), whereas 16 S datasets were dominated by Lactobacillus (22%-42%) with low levels of Bartonella (0%–9%) (Fig. 3A).

Next, we defined core taxa as those present in > 80% of field control samples, analyzing 16 S and WGS datasets separately (Fig. 3B, C). In the 16 S datasets, six core genera were identified: Lactobacillus, Bombilactobacillus, Gilliamella, Snodgrassella, Frischella, and Bifidobacterium. WGS datasets revealed four additional core genera: Entomomonas, Bartonella, Apilactobacillus, and Commensalibacter (Table S3). Entomomonas is supported by a Metagenome-Assembled Genome recovered from our co-assembly (Table S2, bin 20), and Apilactobacillus kunkeei by genome-wide read mapping (Table S14). Two further genera flagged as core by Kraken 2 — Moraxella and Vibrio — were not supported by genome-wide read mapping (Table S14), likely reflecting read misassignment from related taxa absent from the database; we therefore mark these with asterisks in Fig. 3 and do not interpret them as core honeybee taxa. The occurrence of WGS-specific core taxa likely reflects methodological differences between approaches. For Bartonella specifically, the available season metadata (Table S1) argue against seasonality as the primary explanation: the Denmark 16 S (early autumn) and China WGS (early autumn) cohorts show very different Bartonella levels despite matched season, and the Armenia 16 S samples spanning three seasons did not show a consistent Bartonella seasonal pattern.

At the species level, in the 16 S datasets, Lactobacillus resolved to L. apis and L. melliventris, whereas WGS datasets revealed a broader range of core species, including L. apis, L. kullabergensis, L. helsingborgensis, and L. sp. wkB8. Notably, L. melliventris was not identified by Kraken 2 in the China and UK WGS datasets despite its presence among reads, suggesting database annotation issues (see Supplementary Note 1, Fig. S4). Core Gilliamella species were resolved only in WGS samples, dominated by G. apicola along with G. sp. ESL0441, ESL0405, and ESL0443. In Bifidobacterium, B. asteroides was core in both 16 S and WGS, with B. coryneforme and B. indicum additionally core in WGS. In Snodgrassella, S. alvi was the only core species consistently identified in 16 S and WGS. Additional WGS-only core species were Bombilactobacillus bombi, Frischella perrara, Bartonella apis, Commensalibacter sp. ESL0284 and AMU001, Apilactobacillus kunkeei, Entomomonas moraniae, Moraxella osloensis, and Vibrio anguillarum (Fig. 3C).

We then compared microbiome diversity across studies. Alpha-diversity was significantly higher in New Zealand compared to Southern US and Denmark (Kruskal-Wallis, p < 0.001). Armenia also exhibited higher diversity than the Southern US (p < 0.001) (Fig. 3D, Table S4). PCoA revealed that New Zealand and Armenia samples clustered distinctly from other studies, indicating greater within-group homogeneity (Fig. 3E, Table S5).

The features influencing discrimination of Armenia and New Zealand samples were increased Lactobacillus and Bombilactobacillus, and decreased Gilliamella (Fig. 3E). Notably, a subgroup of Southern US samples from Texas differed from those in Florida and Tennessee due to higher Pantoea abundance (e.g. P. sp. SO10, P. vagans, P. agglomerans, P. ananatis). These species may act as plant pathogens or contribute to pathogen defense in both plants and bees, with some linked to urban areas^38–42.

Overall, our analysis reveals a conserved core microbiota across diverse geographic regions, although their relative abundances vary significantly by location and sequencing method. We next investigated how seasonal and environmental factors shape the gut microbiota of honeybees in Armenia.

Seasonal and environmental factors are linked to changes in gut microbiota in Armenia, with increased Commensalibacter abundance in autumn

Gut microbiomes of worker honeybees in Armenia were collected from nine regions during 2017, 2020, and 2023, in spring, summer, and autumn. All Armenian samples in this study are pools of 5–6 worker bees per pool (n = 21 pools; see Methods); the analyses below are therefore at pool resolution rather than at individual-bee resolution. The microbiome was largely dominated by Lactobacillus, Bombilactobacillus, and Bifidobacterium (Fig. 4A).

Fig. 4.

Impacts of seasonality and urbanization on honeybee gut microbiome composition in Armenia. (A) Relative abundance of major taxa in Armenian honeybee gut samples, ordered first by season and then by anthropization level. Bar plots reflect composition among the retained ASVs in pooled samples (5–6 worker bees per pool, n = 21). Sample location abbreviations: AR: Ararat, GE: Gegharkunik, KO: Kotayk, SY: Syunik, LO: Lori, VD: Vanadzor, JE: Jermuk, AG: Aghveran, HA: Hanqavan. (B) Alpha diversity of the gut microbiome (Shannon index) analyzed according to seasonal and anthropization factors. The central line indicates the median, the box represents the interquartile range (25th − 75th percentiles), and whiskers extend to the 10th and 90th. (C) Principal coordinate analysis (PCoA) based on Bray–Curtis distances; arrows indicate the most discriminant genera (multiple regression with permutations, p < 0.05). Samples are colored by collection season. The adjacent scatterplot highlights seasonal differences in Commensalibacter abundance, with group differences determined using ANCOM-BC2. (D) PCoA is colored by anthropization level, with an accompanying scatterplot illustrating differences in Bombilactobacillus abundance between urban and natural habitats.

Alpha diversity (Shannon index) was comparable across the 21 samples and did not differ significantly between seasons or anthropization levels (Fig. 4B). Seasonal comparisons among spring, summer, and autumn samples revealed significant differences between autumn and summer (PERMANOVA, ω² = 0.072, p = 0.012) (Fig. 4C, Table S6). These differences were primarily associated with higher relative abundances of Commensalibacter and Snodgrassella in autumn, and of Bombilactobacillus in summer. Differential abundance analysis confirmed a statistically significant rise in Commensalibacter during autumn (summer-autumn, q < 0.001, logFC = 3.94; spring-autumn, q < 0.016, logFC = 4.56) (Fig. 4C, Table S7).

We further compared bees from urban/near-urban apiaries (Ararat and Kotayk, near Yerevan) with those from rural or natural regions. PCoA revealed distinct microbiota profiles between urban and natural sites (PERMANOVA, ω² = 0.077, p = 0.04) (Fig. 4D), primarily due to a higher relative abundance of Bombilactobacillus in the urban samples. However, differential abundance analysis did not yield statistically significant differences (q < 1, logFC = 0.32), likely due to the small sample size (Fig. 4D, Table S7)⁴³. Although some studies have reported increased microbial diversity in urban environments^40,44,45, we did not observe such an effect (Fig. 4B).

Differential Impact of Pesticides and Diet on Honeybee Gut Microbiome

We investigated the impact of pesticides and diet on the composition of the honeybee gut microbiome. All datasets were processed using a uniform preprocessing and analysis pipeline. In the studies considered, exposures were administered via feeding, with controls provided either as sugar water (sugar controls¹⁸ or as field controls). Comparisons were made between treated samples and their corresponding controls.

Oxalic acid exposure was associated with a pronounced shift in beta diversity (PERMANOVA, ω² = 0.115, p = 0.001), while alpha diversity did not differ significantly from sugar controls (Fig. 5A). Differential abundance analysis revealed that oxalic acid was associated with significant decreases in Bombella intestini (logFC < -3.8, adjusted p < 0.001), Bartonella (logFC < -1.4, adjusted p = 0.003), Apilactobacillus (logFC < -1.4, adjusted p = 0.003), alongside an increase in Klebsiella (logFC > 1.2, adjusted p = 0.013) (Fig. 5E, Table S8). These findings are consistent with the original study¹⁸, which employed ALDEx2 for differential analysis⁴⁶ and detected similar taxa, although additional species were reported.

Fig. 5.

Effects of exposures on alpha diversity, community composition, and differential abundance of specific taxa. (A)–(D) Changes in the honeybee gut microbiome following various exposures. Alpha diversity (Shannon index) is presented as boxplots, where the central line indicates the median, the box represents the interquartile range, and whiskers extend to the 10th and 90th percentiles. Group differences in alpha diversity were assessed using the Kruskal-Wallis test (p-values indicated). Beta diversity was evaluated using Bray-Curtis distances and visualized via PCoA, with ellipses representing 95% confidence intervals; PERMANOVA (ω² and p-values) was used to assess differences between groups. The panels depict: (A) Oxalic acid (vs. sugar controls); (B) Glyphosate treatment in adult honeybees from Uruguay and Texas, USA, as well as newly emerged bees in Texas (vs. sugar controls); (C) Manuka honey (vs. field controls); and (D) Inverted sugar (vs. field controls). (E) Differentially abundant taxa identified by ANCOM-BC2 following exposure. Log₂ fold changes (logFC) are plotted on the x-axis, and only taxa with statistically significant differences (adjusted p-value < 0.05) are shown. Taxa names in bold indicate those that passed the sensitivity analysis for pseudo-count addition. Squared taxa belong to the core, round taxa are non-core.

In contrast, exposures to other insecticides, including acetamiprid, thiacloprid, and imidacloprid were not associated with significant shifts in overall microbiome composition (Fig. S5), consistent with previous studies^4,18,24. Nonetheless, differential abundance analysis identified specific taxa that were altered by these insecticides, including increased abundances of Bombella apis, Hafnia-Obesumbacterium, Klebsiella, Lonsdalea, and Serratia, and a decrease in Fructobacillus and Apilactobacillus (Fig. 5E, Table S8).

For glyphosate, we analyzed data from studies conducted in Uruguay⁴ and Texas, USA²⁸. In Texas, two experimental setups were performed, one treating adult bees and another treating newly emerged bees (NEW), while in the Uruguay study, only NEWs were considered. Beta diversity analysis revealed a significant shift in Texas NEW samples (PERMANOVA, ω² = 0.191, p = 0.001). The Texas adult (hive) bees showed a more modest but significant shift (PERMANOVA, ω² = 0.017, p = 0.003); no significant changes were detected in the Uruguay NEWs, in agreement with the original Uruguay report^4,28. Although the Uruguay study reported compositional changes in NEW bees, these findings were based on comparisons involving multiple exposures rather than glyphosate alone (Table S9). Notably, differential abundance analysis showed consistent reductions in S. alvi in Texas and Uruguay NEW samples (Fig. 5E, Table S8)^4,28. In adult bees from Texas, glyphosate treatment was also associated with a decrease in Fructobacillus fructosus and an increase in Lonsdalea, whereas the Uruguay NEWs revealed reductions in Commensalibacter and S. alvi coupled with an increase in Staphylococcus.

In contrast to pesticides, dietary supplements such as manuka honey and inverted sugar significantly decreased alpha diversity compared to field controls, consistent with previous findings from New Zealand¹⁷. Both supplements induced shifts in microbiome composition as determined by beta diversity analysis, though the effect sizes were smaller than those observed with oxalic acid and glyphosate in Texas NEWs (PERMANOVA ω² = 0.056, p = 0.005 for manuka honey; ω² = 0.046, p = 0.039 for inverted sugar) (Fig. 5C, D). At the taxonomic level, manuka honey supplementation was associated with a decrease in Apilactobacillus, a finding not reported in the original study (Fig. 5E, Table S8). Additionally, we did not identify any differentially abundant taxa with inverted sugar, in contrast to the original study, which utilized ANOVA for statistical analysis¹⁷(Table S9).

Co-Abundance Networks Reveal Adaptive Pesticide Responses and Potential Glyphosate Degradation in Honeybee Gut Microbiome

To identify bacterial communities that respond jointly to pesticide and dietary exposures, we applied the tsantsR framework to all the datasets. Co-abundance networks were constructed as described in Methods, with nodes representing taxa and edges representing significant co-abundance relationships (Fig. 6).

Fig. 6.

Workflow for constructing co-abundance networks and identifying maximum-weight connected subgraphs (MWCS). The co-abundance network is computed using Spearman correlations within each exposure group across all studies. Nodes represent species- or genus-level taxa, while edges represent statistically significant co-abundance relationships and should be interpreted as coordinated abundance shifts rather than direct ecological interactions. Log fold changes (logFC) between exposure and control groups within each study were computed using ANCOM-BC2. The absolute logFC values were discretized into node weights to favor MWCS-based selection of subnetworks with a maximum number of positively weighted (differentially abundant) nodes and a minimum number of negatively weighted (unchanged) nodes (see Methods for details).

The glyphosate-associated subnetwork identified in the Uruguay and Texas NEW samples included interconnected taxa, with decreases in S. alvi in both datasets, and a decrease in Serratia only in Uruguay. In the Uruguay NEWs, this subnetwork expanded to encompass additional taxa: decreased Frischella, Commensalibacter, Lactobacillus, Bifidobacterium asteroides, and Bombilactobacillus, together with increased Gilliamella, Fructobacillus (other), Lactobacillus apis and Pseudomonas.

Pathway enrichment within this glyphosate-associated subnetwork in Uruguay NEW samples revealed multiple amino acid metabolic pathways, including tryptophan and tyrosine metabolism, which were enriched following glyphosate exposure (Fig. 7A, B, Table S11, S12).

Fig. 7.

Co-abundance networks of the honeybee gut microbiome under exposures and corresponding KEGG pathway overrepresentation analysis. Panels depict exposure-associated subnetworks for: (A) glyphosate in Uruguay NEW samples, (B) glyphosate in US Texas NEW samples, (C) oxalic acid, (D) acetamiprid, and (E) thiacloprid. For each exposure, the upper panel shows the co-abundance network and the lower panel displays the corresponding functional enrichment results. In the networks, nodes represent taxa and edges denote significant Spearman correlations (ρ > 0.6, adjusted p < 0.2); solid lines indicate positive correlations, and dashed lines negative correlations. Core taxa are represented by rectangular nodes, and non-core taxa by circles. Node size reflects the absolute log fold change (LFC) between exposure and control groups, with a green-to-blue color scale indicating up- and down-regulation, respectively. Nodes for which LFC values were not calculated by ANCOM-BC2 due to sample size limitations are colored gray. Nodes and edges highlighted with red borders indicate the exposure-specific subnetwork identified using the maximum weight connected subgraph algorithm (see Methods). Only KEGG pathways significantly enriched in the exposure groups (FDR < 0.2) are shown.

Furthermore, the enrichment of pathways for xenobiotic and aromatic compound degradation, nitrogen cycling, and two-component system, suggests potential capacity for glyphosate degradation. Pseudomonas is known to degrade glyphosate either by converting it to sarcosine via the C-P lyase pathway or by transforming it into aminomethylphosphonic acid (AMPA) and methylamine^47–50. In our samples, we observed enrichment of the phosphonate and phosphinate metabolism pathway, including the phnJ gene, which encodes C-P lyase; a key enzyme in both degradation routes. This gene is part of the phn operon, which also encodes putative glyphosate transporters (phnC, phnD, and phnE)⁵¹, indicating glyphosate uptake capacity. Additionally, enriched pathways for glycine metabolism, nitrogen cycling, and aromatic compound degradation may further facilitate the breakdown of glyphosate degradation intermediates.

We also observed an enrichment of bacterial chemotaxis and biofilm formation pathways, likely mediated by both Pseudomonas and Gilliamella. This response may enhance overall bacterial stress adaptation⁵². Furthermore, a disconnected subnetwork in these samples showed upregulation of Staphylococcus and Streptococcus, aligning with previous reports of impaired immune function following glyphosate treatment⁵³.

Notably, the glyphosate subnetwork in Texas NEWs only included nodes with negative or unidentified logFC values, while the Texas hive samples (adult bees) did not reveal an exposure-associated subnetwork. This cohort was exposed to a higher glyphosate dose, but over a shorter duration than the Uruguayan bees.

Exposures with oxalic acid and neonicotinoids (acetamiprid and thiacloprid) affected a different subnetwork that included Klebsiella, Hafnia-Obesumbacterium, and Serratia. Klebsiella and Serratia increased in both cases, while Hafnia-Obesumbacterium increased under neonicotinoid exposure, and decreased under oxalic acid treatment. (Fig. 7C). For neonicotinoids, this subnetwork further expanded to include Fructobacillus, which decreased in abundance (Fig. 7D, E). Oxalic acid’s impact also extended to include decreased abundances of Snodgrassella alvi, and Commensalibacter, highlighting the documented sensitivity of these taxa to pesticides¹⁸.

The oxalic acid–associated subnetwork showed enrichment for stress response pathways, including ABC transporters and⁵⁴ siderophore biosynthesis, while⁵⁵ neonicotinoid-associated subnetworks were enriched for pathways related to stress response, such as the two-component system^56,57 and cationic antimicrobial peptide (CAMP) resistance⁵⁸(Fig. 7D-E, Table S11, S12).

When comparing field controls to sugar supplements, we did not detect any subnetworks associated with manuka honey, while dihydroxyacetone and inverted sugar only revealed one or two node-subnetworks, both including an increased abundance of Apibacter (Fig.S6).

Overall, across pesticide exposures, two consistent signatures emerge: relative reductions of core taxa including Snodgrassella alvi under glyphosate in both Texas NEW and Uruguay NEW cohorts (Fig. 5E; glyphosate-associated subnetworks in Fig. 7A, B), and a Klebsiella / Hafnia-Obesumbacterium / Serratia opportunistic-pathogen signature shared between oxalic acid and neonicotinoid exposures (Fig. 5E; subnetwork in Fig. 7C).

Discussion

This study provides a global view of the honeybee gut microbiome and its responses to environmental and anthropogenic factors. By integrating newly generated Armenian data with publicly available 16 S and metagenomic datasets and re-analyzing them under a unified pipeline, we identified consistent core taxa and context-dependent shifts associated with geography, season, and pesticide exposure (Figs. 1 and 2).

Our results support the established view that honeybees maintain a defined core gut microbiome, composed mainly of six phylotypes: Lactobacillus, Bombilactobacillus, Gilliamella, Snodgrassella, Frischella, and Bifidobacterium. Including metagenomic WGS datasets revealed additional taxa expanding the core to ten genera. Five of the core genera, Lactobacillus, Bombilactobacillus, Gilliamella, Snodgrassella, and Bifidobacterium, are well documented^11,59,60. Others, including Bartonella, Commensalibacter, Frischella, and Apilactobacillus appear context-dependent^2,18,61. In WGS datasets from the UK and China, we also found Entomomonas, a novel genus from Asian honeybee gut⁶². The broader range of core taxa identified in WGS most likely reflects the greater taxonomic resolution of the sequencing approach rather than true biological difference between the datasets studied. While it is known that honeybee microbiota are shaped by diet, climate, and local flora^63–67, the limited metadata on season and local conditions prevented us from confidently linking environmental factors to the observed differences seen across geographies (Fig. 3).

In the Armenian samples, seasonal and environmental effects were evident: Commensalibacter and Snodgrassella showed higher relative abundance in autumn and Bombilactobacillus was more abundant in urban colonies. The rise of Commensalibacter aligns with Swiss data⁶⁸ but contrasts with reports from other regions where Bartonella or Snodgrassella predominate in late seasons^65,69. Although the urban signal was modest and differed from previous scarce reports⁶⁶, it may reflect differences in nectar sources, floral diversity, or pollution levels. Together, these observations underscore the interplay between season, foraging environment, and microbiome composition, and explain why regional studies often report conflicting results (Fig. 4).

Previous studies have linked pesticides to altered gut composition but with inconsistent taxa and unclear mechanisms^4,18,22,24. Our meta-analysis resolves some of this ambiguity by integrating multiple datasets under a unified framework. Glyphosate exposure was associated with reductions in core taxa such as Snodgrassella alvi, Frischella perrara, Commensalibacter, Lactobacillus, Bifidobacterium asteroides, and Bombilactobacillus, and with enrichment of Gilliamella, Fructobacillus, Lactobacillus apis, and Pseudomonas. Newly emerged bees (NEW) from Texas showed marked beta diversity shifts, and those from Uruguay exhibited drastic network changes suggesting higher vulnerability in younger bees, whereas adult foragers in Texas showed more resilient microbiome profiles^4,28(Figs. 5 and 7).

Because focusing on individual taxa may not fully capture the collective responses of microbial communities, we employed co-abundance network analysis to identify coordinated bacterial shifts and their associated functions (Fig. 6). Our co-abundance network analysis further revealed that in Uruguayan NEWs, glyphosate exposure was linked to a distinct subnetwork involving increased Pseudomonas, and Gilliamella, enriched for phosphonate and phosphinate metabolism pathways, including phnJ, consistent with C-P lyase or AMPA-mediated glyphosate degradation. While Pseudomonas species have demonstrated glyphosate utilization in other ecosystems^50,51, this, to the best of our knowledge, is the first potential indication of such a capacity in honeybees. Enrichment of tryptophan and tyrosine metabolism pathways further suggests compensation for glyphosate-induced shikimate pathway inhibition (Fig. 7). Whether this microbial adaptation mitigates glyphosate’s physiological effects on the host remains unknown. The Texas NEWs cohort, exposed to a higher glyphosate dose (169 mg/L versus 10 mg/L in Uruguay) for a shorter duration (2 days versus 7 days in Uruguay), showed a smaller subnetwork, suggesting exposure duration rather than dose is linked to those community responses.

Oxalic acid exposure was associated with the most pronounced beta-diversity shifts, including increases in relative abundance of opportunistic pathogens, such as Serratia, Hafnia-Obesumbacterium, and Klebsiella, and reduced Bartonella, Bombella, Apilactobacillus, and Snodgrassella^18,54 (Figs. 5E and 7C). Although considered “bee-safe,” oxalic acid lowers gut pH, favoring stress-tolerant taxa and pathways of response to low pH and iron limitation (e.g., ABC transporters, siderophore biosynthesis)^54,55. Neonicotinoids (imidacloprid, thiacloprid, acetamiprid) were linked to subtler changes, enriching opportunistic pathogens with xenobiotic detoxification, stress response functions (e.g., two-component system, ABC transporters)^18,56,57, and cationic antimicrobial peptide resistance. These adaptations could confer a selective advantage if the host produces more antimicrobial peptides under insecticide-induced stress⁵⁸ (Fig. 7). Overall, pesticides favored stress-resistant or metabolically adaptable bacteria, while herbicides were linked to compositional and functional restructuring.

Strengths and limitations

A major strength of our study is the standardized re-analysis of diverse datasets using a unified bioinformatics pipeline, which minimizes methodological discrepancies that often hinder cross-study comparisons. Crucially, to address the inherent heterogeneity of public datasets (e.g., varying exposure doses, durations, and bee genotypes), we employed a stratified, within-study comparison design. By calculating effect sizes relative to matched controls within each independent dataset before aggregating results, we controlled for baseline variances, ensuring that the identified microbial shifts reflect consistent responses to exposure rather than confounding environmental factors. Additionally, the new co-abundance network framework, implemented in the tsantsR package, allows us to link these taxonomic shifts with their functional potential.

Nevertheless, we acknowledge several limitations. Foremost, this study is a computational meta-analysis, and we did not perform new in vivo experimental validation for the identified associations. Consequently, our findings regarding pesticide-associated shifts and metabolic pathways—such as the predicted potential for glyphosate degradation—remain hypothesis-generating. Experimental verification is required to confirm the causal nature of these microbial responses and their physiological impact on the host.

Furthermore, metadata incompleteness limited our ability to fully correct for seasonality and caste effects in some datasets. 16 S-based data lack strain-level resolution, and WGS data were available only for untreated samples, constraining our ability to infer strain-specific functional changes under exposure. Finally, regarding the network analysis, it is important to note that correlation-based methods on compositional data have inherent limitations; such associations should not be interpreted as direct ecological or causal interactions. Rather, they reflect correlated abundance shifts that likely arise from shared environmental pressures or metabolic dependencies. The goal of our network analysis was therefore to identify these co-shifting microbial communities to guide future mechanistic inquiry.

Concluding remarks

This meta-analysis reaffirms a stable honeybee core microbiota while demonstrating that environmental factors and anthropogenic influences induce taxon-specific shifts. Herbicides such as glyphosate disrupt core members while enriching potential degraders, whereas insecticides and acaricides select for stress-tolerant opportunists. These findings suggest that the honeybee gut microbiome actively responds to chemical pressure. Longitudinal and experimental studies are needed to determine the implications of these microbiome shifts for bee health and resilience.

Methods

Dataset search and inclusion criteria

We conducted a systematic search for datasets in the NCBI SRA database using the terms: (microbiome AND “gut” AND (“Apis mellifera” OR “honeybee”)). We included both 16 S rRNA amplicon and whole-genome sequencing (WGS) datasets.

The inclusion criteria for studies were as follows: (i) paired-end sequencing; (ii) study subjects being worker bees (no queens or drones included); (iii) inclusion of the entire gut (including crop, midgut and hindgut); (iv) availability of raw data and corresponding metadata; (v) for 16 S amplicon sequencing, libraries targeting the V4 or V3-V4 regions to minimize batch effects and enable cross-study comparison^43,70 (see Supplementary Note 2 for the PRISMA flow diagram).

In total, we analyzed 55 WGS and 472 16 S samples originating from colonies maintained under natural or laboratory conditions and representing diverse geographic regions spanning Armenia, Uruguay, Denmark, New Zealand, Southern US, China, and the UK (Table 1; Fig. 2). Exposures included herbicides (glyphosate vs. sugar controls), insecticides (oxalic acid and neonicotinoids vs. sugar controls), and sugar supplements (vs. field controls).

The data were downloaded using the fastq-dump tool from the NCBI SRA Toolkit, or directly through personal communication with authors. Detailed information about the datasets is available in Table S1.

DNA extraction and sequencing of the Armenian samples

Honeybees from different colonies were collected from ten provinces of Armenia. Guts were dissected using flame-sterilized forceps under aseptic conditions. The head was separated from the thorax to detach the esophagus, and the entire intestinal tract was carefully extracted through the anterior end by grasping the last abdominal segment with forceps and gently pulling. Additional segments of the abdominal exoskeleton were carefully detached to ensure the rectum was removed intact. Guts from 5 to 6 bees were pooled into a single sample, resulting in a total of 21 pooled samples analyzed in this study.

DNA was extracted using the cetyltrimethylammonium bromide (CTAB) bead-beating method, followed by phenol-chloroform purification and ethanol precipitation, as previously described⁷¹. The extracted DNA was sequenced at Macrogen Inc. (Seoul, Korea) using the Illumina MiSeq platform. The V3-V4 hypervariable regions of the 16 S rRNA gene were amplified with primers 341 F (5′-CCTACGGGNGGCWGCAG-3′) and 805R (5′-GACTACHVGGGTATCTAATCC-3′)⁷².

16 S data processing

Raw paired-end reads were filtered out if they contained any of the following: (i) ambiguous bases, (ii) homopolymers longer than 8 bp (using fastq_qual_trimmer v1.0), or (iii) quality scores less than 28 (using fastq-mcf v.1.0).

The trimmed sequence reads were processed using the R dada2 package (v.1.22.0) to infer exact amplicon sequence variants (ASVs)⁷³. Briefly, reads were truncated with the filterAndTrim function, specifying the truncLen parameter to trim bases with a quality score below 30; reads containing more than two errors (forward) and three errors (reverse) were removed with the maxEE parameter; sequence reads were dereplicated, denoised, and merged using dada2, with pooled sample inference applied for each dataset.

The datasets were merged into a single ASV table using the mergeSequenceTables function of the phyloseq package (v.1.48)⁷⁴. To reduce batch effects, ASVs were truncated to the common V4 region through the following steps: (i) sequences were aligned using the msa R package v.1.30.1⁷⁵; (ii) aligned ASVs were truncated from the 3′ end, based on the C3 conserved sequence (GTGCCAGCAGCCGCGGTAAA), allowing up to five mismatches; (iii) reads were further truncated at the 5′ end, to the most abundant length of 60 bp; (iv) alignment gaps were removed.

Taxonomy was assigned to the ASVs using the SILVA database⁷⁶ via dada2 assignTaxonomy function. The taxonomy-assigned data was converted into a phyloseq object. Chimeric ASVs and those observed in only one sample were filtered out. ASVs unclassified at the genus level, assigned to mitochondria, chloroplasts, or eukaryotes, or not assigned to bacterial or viral lineages were excluded. To improve taxonomic accuracy, we retained only taxa meeting the following criteria: (i) at least 10 ASVs in one or more samples, and (ii) present in more than 70% of samples in each exposure group within a study (e.g. stratified by control or exposure type, and, if applicable, experimental setup (exposure on newly emerged or adult bees)). In this study, an exposure refers broadly to chemical exposures (e.g., pesticide application) and dietary supplementation. An overview of study-specific exposure conditions is provided in Table 1.

Rarefaction curves were constructed using the rarefy function from the vegan package (v.2.6.8) to normalize for library size. Each sample was rarefied to 12,000 reads using the rarefy_even_depth function in phyloseq (Fig. S1A). This rarefaction threshold was selected to ensure consistent normalization across datasets while minimizing sample loss, allowing for comparative analyses using a uniform depth. Following all filtering steps, the initial set of 1,174 ASVs across all datasets was reduced to 67 ASVs retained for downstream analysis. As an exception, the Texas 17 glyphosate dataset (Fig. 5B) was analysed on the non-rarefied data, because rarefying to 12,000 reads would have dropped most of its low-depth samples.

To verify that the filtering and rarefaction steps do not drive our conclusions, we recomputed PERMANOVAs for the main contrasts under a 2 × 2 grid of filter and rarefaction settings; bias-corrected ω² effect sizes (MicEco::adonis_OmegaSq) and p-values are reported in Supplementary Table S13.

Batch effect correction

To address potential batch effects resulting from differences in sequencing methodology^43,70, we performed principal coordinate analysis (PCoA) based on Bray–Curtis dissimilarity among samples. Clustering patterns were primarily driven by amplicon region (PERMANOVA: ω² = 0.133, p < 0.001) and library size (ω² = 0.145, p < 0.001) (Fig. S2A). These effects were substantially reduced after trimming all datasets to the common V4 region and standardizing library sizes through rarefaction and abundance-based filtering as described above (Fig. S2B). Although some variation between studies and experimental types (adult vs. newly emerged bees) persisted, residual technical effects were modest following normalization, confirming that cross-study comparisons of gut microbiome composition were not biased by sequencing batch effects.

Core microbiome analysis

We conducted a sensitivity analysis to determine the optimal prevalence cutoff for defining core taxa at the genus level. We computed the variance of the number of shared taxa between dataset pairs at each prevalence threshold varying from 10% to 90%. The variance stabilized at 80%, supporting this value as a robust cutoff for defining the core microbiome in subsequent analyses (Fig. S3).

WGS data processing

Taxonomic classification of the WGS reads was performed with Kraken 2 (v.2.1.2)⁷⁷ and the NCBI Standard database (https://benlangmead.github.io/aws-indexes/k2, accessed on 17 May 2022). 1–9% of reads were assigned to prokaryotes.

Species-level counts were re-estimated using Bracken (v.2.9). The species abundance table was rarefied to 800,000 reads (Fig. S1B) and the taxonomic table was used to construct a phyloseq object.

Diversity analysis and statistics

Shannon alpha diversity was calculated using the phyloseq estimate_richness function. For community composition (multivariate) analyses, pairwise Bray-Curtis dissimilarity between samples was estimated, and principal coordinates were computed using the dist_calc and ord_calc functions in the microViz package v.0.12.4⁷⁸. PERMANOVA analysis was performed using the vegan adonis2 function (v.2.6.8). The vegan envfit function was used to fit environmental variables onto the ordination axes using multiple regression and statistical significance determined with permutation tests. The fitted environmental variables were visualized as vectors, and projected onto the ordination diagram using the ggplot2 geom_segment function (v.2 3.5.1). Both Shannon alpha diversity and Bray–Curtis-based community composition analyses were computed on the filtered, rarefied phyloseq object, with the exception of the Texas 17 dataset. Effect sizes are reported as bias-corrected ω², computed with MicEco::adonis_OmegaSq⁷⁹; ω² is preferred for cross-analysis comparison because R² is upward-biased when sample sizes differ. Alpha-diversity differences were tested using the Kruskal-Wallis rank-sum test; results are tabulated in Tables S4 and S6.

Differential abundance analysis

We performed differential abundance analysis with ANCOM-BC2 (v.2.6.0)⁸⁰, which applies a centered log-ratio (CLR) transformation and bias correction to mitigate the compositional biases inherent to relative-abundance data. ANCOM-BC2 was run on the unfiltered, non-rarefied count table, since the package’s bias-correction model is designed for raw counts. While running ANCOM-BC2, we encountered an error related to object length mismatches that we corrected in our pipeline using the steps suggested previously (https://github.com/FrederickHuangLin/ANCOM-BC/issues/231). ANCOM-BC2 outputs are reported in Tables S7 and S8, including log-fold changes (logFC) with standard errors (SE) and Benjamini-Hochberg-adjusted q-values.

Co-abundance network analysis

We developed an R package, tsantsR, for the construction and functional analysis of taxon-taxon co-abundance networks. Network construction proceeds in four steps:

1. Taxon filtering. Within each (study, exposure-or-control group), we retained taxa with > 20% prevalence. ASVs in a genus with no species-level resolution were aggregated under the genus name; ASVs in a genus with mixed species/genus-level resolution were aggregated and marked with the genus name and an “other” suffix (Table S10).

2. Per-group correlations. Within each (study, exposure group), we computed pairwise Spearman correlations between the relative abundances of filtered taxa. Spearman correlations were used rather than covariance-based inference methods such as SPIEC-EASI because the relatively small group sizes and heterogeneous sampling conditions are not optimal for covariance-based approaches; this choice captures co-abundance patterns reflecting coordinated taxon responses to interventions rather than direct ecological interactions.

3. Bootstrap stability filter. For each (study, exposure group), we performed 200 bootstrap iterations using the phylosmith co_occurrence function: in each iteration, samples were resampled and Spearman correlations recomputed, retaining only taxon pairs with |ρ| ≥ 0.6 and FDR-adjusted p < 0.1, set permissively given the small sample sizes, with false positives controlled through the bootstrap step itself. Pairs surviving in > 80% of iterations with a consistent sign were kept. This bootstrap-with-consistency step is aimed at conservative network reconstruction: only edges that survive resampling at high consistency are retained, minimizing spurious correlations driven by small-sample noise.

4. Cross-group merge. Consistent taxon pairs from all (study, exposure group) cells were merged into a single co-abundance network. Nodes represent species or genera; edge weight is the median Spearman ρ across groups.

Next, we aimed to identify the subnetwork in which the sum of logFC differential abundance weights (compared to controls) was greatest, thereby pinpointing the bacterial communities that collectively respond to the exposure (Fig. 6). Subnetwork extraction proceeds in three further steps:

• 5.logFC computation. We computed log-fold changes (logFC) of taxon relative abundances between exposure and control groups within each study with ANCOM-BC2.
• 6.Node weight discretization. Absolute logFC values were discretized into node weights to emphasize differentially abundant taxa while down-weighting stable nodes, maximizing the signal of coordinated responses. Weights ranged from + 1 to + 3 for differentially abundant taxa: |logFC| ≥ 1.5, weight 3; 1 ≤ |logFC| < 1.5, weight 2; 0.5 ≤ |logFC| < 1, weight 1. Taxa with unchanged abundance (|logFC| < 0.5) were assigned weight − 2, corresponding to the average of the positive weight range; missing values were assigned weight 0.
• 7.Subnetwork extraction. We extracted exposure-associated subnetworks using the maximum weight connected subgraph (MWCS) algorithm of the mwcsr R package (v.0.1.9). MWCS identifies a connected group of nodes that maximizes the sum of their weights, representing the most relevant set of connected taxa associated with an exposure.

All seven steps are implemented in the tsantsR package (https://github.com/abi-am/tsantsR): steps 1–4 in the generate_network function and steps 5–7 in the subgraph_analysis function.

Finally, the networks were visualized in Cytoscape v.3.9.0 using an edge-weighted spring-embedded layout, in which nodes attract or repel each other based on their correlation coefficients⁸¹.

Genome assembly and taxonomic characterization

Metagenome-assembled genomes (MAGs) were reconstructed from the WGS datasets to enhance the accuracy of functional calls in 16 S datasets (see below).

Raw library sizes ranged from 2.1 × 10⁶ to 1.0 × 10⁷ reads. Reads were trimmed and quality-controlled with fastp (v.0.23.2); those < 100 bp, with > 10% low-quality bases (phred quality < 20), > 5 ambiguous bases, or with low complexity (with repetitive or homogeneous patterns) were discarded; Poly-G/X tails were trimmed. Host DNA was eliminated by aligning reads to the Apis mellifera genome (GCF_003254395.2), using Bowtie2 (v.1.3.1). Post-processing yielded 7.8 × 10⁴ to 2.2 × 10⁶ paired-end reads per sample.

Both WGS datasets were co-assembled with MEGAHIT (v.1.2.9), and contigs were binned using MetaBAT, CONCOCT, and MaxBin. The three bin sets were merged via MetaWRAP’s bin_refinement module (v.1.3)⁸². Bin quality was assessed with CheckM (v.1.0.18); only bins with > 80% completeness and < 5% contamination were retained, yielding 15 bins. Finally, MAGs were taxonomically annotated using GTDB-Tk (v.2.4.0)⁸³ (Table S2). To validate the targeted-species assay, representative reference genomes for five species were downloaded from NCBI RefSeq: Frischella perrara (NZ_CP009056.1) and Bifidobacterium indicum (NZ_BTXM01000001.1–NZ_BTXM01000014.1) as positive controls, and Apilactobacillus kunkeei (NZ_CP080568.1), Vibrio anguillarum (NZ_CP031478.1, NZ_CP031479.1), and Moraxella osloensis (deposited as Faucicola osloensis; NZ_CP014234.1–NZ_CP014238.1) as targets. Genomes were concatenated into a single reference, and a Bowtie2 index was built with bowtie2-build. Post-host-removal reads from each sample (China, n = 32; UK, n = 23) were mapped against this combined index with Bowtie2 (v.2.4.4) using default best-alignment-per-read settings and --no-unal. Per-contig coverage was summed to the species level and expressed as percent genome coverage (Table S14).

Gene function annotation

We predicted microbial gene functions for the 16 S datasets using PICRUSt2 (v.2.4.1)^84,85, which infers KEGG Orthology (KO) profiles from ASV taxonomy by phylogenetic placement against reference genomes. As such, the resulting KO assignments are predicted gene contents, not directly observed, and may diverge from the actual genome content of strains in the samples. To replace these inferences with directly observed gene content where possible, we substituted PICRUSt2 KOs with KOs derived from MAGs of matching species. MAGs (Table S2) were annotated with eggNOG-mapper (v.2.1.12)⁸⁵; for species recovered as a MAG, the MAG-derived KO set was used in place of the PICRUSt2-predicted KO set for ASVs of the same species. The resulting merged table therefore combines MAG-derived KOs for the species with MAG support and PICRUSt2-predicted KOs for the rest.

Pathway enrichment analysis

We performed functional enrichment analysis on each exposure-associated subnetwork to determine which pathways may enable these bacteria to adapt and increase in abundance. For this we analyzed the refined KO table to identify enriched KOs in exposure-associated microbial communities. For each subnetwork, we separated nodes into up- or down-regulated groups based on positive or negative logFC values. KO-related gene counts in these groups were compared against the entire network using a hypergeometric test using the stats package in R (v.4.0.2). We computed the total probability of observing the given count (with dhyper) or any higher count (with phyper, lower.tail = FALSE). KOs with p < 0.1 were deemed significantly enriched. This threshold was chosen to increase sensitivity in the context of the modest sample sizes in our datasets, as more conservative cutoffs risk overlooking biologically relevant signals^86–88. Overall false-discovery for the functional-enrichment pipeline is controlled at the subsequent pathway-level FDR step (see below).

Next, we performed KEGG pathway overrepresentation analysis with enrichKEGG in the ClusterProfiler package (v.4.12.6)^89–92, using the underlying network’s KO distribution as a reference. Pathways with FDR-adjusted p ≤ 0.2 were considered enriched. Again, we chose a higher threshold considering the exploratory nature of our analysis and the modest sample sizes^93,94. Only pathways enriched among upregulated (and not downregulated) nodes were retained.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

LN and NV conceived the study. NV performed the analyses, produced the figures, and wrote the manuscript. IB generated the Armenian data. LA developed the *tsantsR* package and conducted the co-abundance network analysis. MZ and HL carried out additional experiments. LN supervised the study and prepared the manuscript. CM supervised computational analysis and revised the manuscript. All authors read and approved the final version of the manuscript.

Funding

The authors acknowledge funding from the Higher Education and Science Committee (HESC) MESCS RA Prospective Directions grant awarded to LN (24FP-2I061), and Yerevan State University Inner Grant 2022 to IB. CM was supported by the MiCRop Consortium (Grant NWO/OCW 024.004.014).

Data availability

Raw 16 S rRNA gene sequencing and WGS data sets evaluated in this study are available from the NCBI Sequence Read Archive (PRJNA719169, PRJNA732842, PRJNA775827, PRJNA432211, PRJNA432210, PRJNA483763, PRJNA531038, PRJNA685398, PRJNA787435 and PRJNA494922). The 16 S amplicon datasets from the Armenian samples generated in this study are deposited in the NCBI SRA database (PRJNA1220608).The *tsantsR* package for co-abundance network analysis is available for download and use at Github: [https://github.com/abi-am/tsantsR](https:/github.com/abi-am/tsantsR) . The rest of the scripts used in this study are available at [https://github.com/abi-am/bee-microbiome-metaanalysis](https:/github.com/abi-am/bee-microbiome-metaanalysis) .

Declarations

Competing interests

The authors declare no competing interests.