Mosquito microbiome diversity varies along a landscape-scale moisture gradient

Abstract

Microorganisms live in close association with metazoan hosts, and form symbiotic microbiotas that modulate host biology. Although the function of host-associated microbiomes may change with composition, hosts within a population can exhibit high turnover in microbiome composition among individuals. However, environmental drivers of this variation are inadequately described. Here, we test the hypothesis that this diversity among the microbiomes of Aedes albopictus (a mosquito disease vector) is associated with local climate and land-use patterns on the high Pacific island of O’ahu, Hawai’i. Our principal finding demonstrates that the relative abundance of several bacterial symbionts in the Ae. albopictus microbiome varies in response to a landscape-scale moisture gradient, resulting in turnover of the mosquito microbiome composition across the landscape. However, we find no evidence that mosquito microbiome diversity is tied to an index of urbanization. This result has implications toward understanding the assembly of host-associated microbiomes, especially during an era of rampant global climate change.

Introduction

Symbiotic or host-associated microbiomes represent communities of microorganisms that live on or within a host and can modulate the host’s behavior, metabolism, fecundity, developmental schedule, and immune system [1, 2]. Although often crucial for normal host organismal function, the microbiomes of individuals may vary substantially within a species or host population, and the origin of this diversity is inadequately resolved [3, 4]. Environmental variation has been proposed as a hypothesis to explain this diversity, as a substantial proportion of the microbiome diversity is acquired from the environment [5]. Niche effects associated with environmental variation may shape the regional species pool of symbionts that colonize host individuals and directly augment β-diversity in microbiomes among different environments. In addition, microbiome function changes with microbiome taxonomic composition, and hosts may associate with different microorganisms in response to these different environmental pressures [6].

Mosquitoes represent an ideal system to test the hypothesis that β-diversity in symbiotic microbiomes is associated with environmental variation [7–10]. Mosquitoes harbor simple microbiomes, generally including 10–100 common bacteria taxa [11, 12]. While the majority of these taxa are acquired from the environment, some mosquito lineages are infected vertically with the bacterial endosymbiont Wolbachia sp. [13]. Individual mosquitoes also show a high turnover in the composition of their microbiotas between individuals within and among populations [9, 14]. While the functional impact of this turnover is poorly understood, diversity in the microbiome is an important modulator of mosquito biology, including their capacity to sustain disease transmission [15, 16].

Here, we explore the β-diversity of the microbiome of a medically important human disease vector, Aedes albopictus, across two major contributors to environmental heterogeneity across the landscape: local climate and land-use. Climate variables are well known modulators of ecological community structure, including microbiomes [17–19]. In addition, land-use patterns such as urbanization drastically alter habitat structure and increase pollution, which may in turn promote changes in microbiome structure and composition [8, 20–22]. The high Pacific island of O’ahu, Hawai’i represents a premier natural laboratory to test the association between the diversity of the mosquito microbiome and the environment. A principal driver of local climate heterogeneity is the topographic structure of the island and prevailing moisture laden trade winds that arrive from the northeast [23, 24]. The trade winds encounter steep mountain ranges on the island’s windward flanks, rise in elevation, and cool via adiabatic processes. This cool air holds less moisture, and forms clouds and rain on the mountainous and windward sections of the island. Winds that move over these mountainous regions to leeward sides of the islands are far less humid, and less clouds and rain form in these regions. The degree of environment variation along this gradient relative to geographical distance is immense, as the environment of O’ahu may transition from humid rainforest to arid coastal or leeward dessert in a few kilometers. In addition, the pattern and levels of land-use change dramatically across the island of O’ahu, a populous island with approximately a million residents. We hypothesize that both local climate and the level of urbanization shape the mosquito microbiome over this landscape at very fine spatial scales.

Methods

We use a cross-sectional study design on O’ahu to test for an association between relatively stable landscape variables (local climate, land-use) and the composition of the mosquito microbiome at fine spatial scales. Several studies have used cross-sectional data to gain insights into distinct drivers of variation in host-associated microbiomes, including links between microbiome composition and local climate and land-use [9, 18, 25–27]. Specifically, Aedes albopictus females were captured from 50 sites across the island of O’ahu (Figure 1) from January through September 2019 using hand nets and hand held aspirators, and were transported from the field to the lab at 4°C. Sites ranged from <250 m to 50423 m apart. Upon arrival to the lab, mosquitoes were sterilized with 75% ethanol for 1 min and rinsed twice in sterile 1X PBS. Midguts were immediately dissected from carcasses and pooled per sampling site (i.e. one pool per sampling site). Each pool contained mosquitoes caught within a 50 m radius area during a sampling event of 5 to 60 minutes. The number of midguts per pool ranged from 2–10 with a mean, median, and mode of 9, 10, and 10, respectively. Midgut pools were stored at −80°C until DNA extraction. DNA was extracted and purified on a KingFisher Flex (Thermo Fisher Scientific, Waltham, MA) using a NucleoMag Tissue Kit (Macherey-Nagel GmbH & Co. KG, Düren, Germany) following the manufacturer’s protocol.

Figure 1.

A map of O’ahu demonstrating the moisture gradient (Principal Component 1 score) and the urbanization index (as dot size) for each of the 50 sample sites across the island. See methods for the calculation of the Principal Component 1 score and the estimation of the urbanization index as the proportion of built-up structures in the area. Specific details on the map’s creation is detailed in an R markdown file in the supplementary material.

The V4 fragment of the bacterial 16S rRNA gene was amplified using Earth Microbiome Project primers (http://www.earthmicrobiome.org/emp-standard-protocols/16s/) with a polymerase chain reaction (PCR). Each PCR contained the following components: 10 μL of 2X Platinum II Hot-Start PCR Master Mix, 7.2 μL of nuclease-free water, 2.0 μL of template DNA, 0.4 μL of 10 μM forward primer, and 0.4 μL of 10 μM reverse primer. PCR amplifications were performed in an Applied Biosystems SimpliAmp Thermal Cycler (Thermo FisherScientific, Waltham, MA) under the following conditions: initial denaturation at 94°C for 2 min; 35 cycles of denaturation at 94°C for 15 s, annealing at 50°C for 15 s, extension at 68°C for 7 s; and a final extension at 68°C for 3 min. After visualization on a 2% agarose gel, the PCR products were purified and normalized to approximately 1.25 to 2.50 ng/μL using a Just-a-Plate kit (Charm Biotech, Cape Girardeau, MO). Equal amounts of each sample were pooled into a single library and subsequently purified and concentrated using a 1.2X volume of AMPure XP. The purified 16S rRNA amplicon library was sequenced using a MiSeq Reagent Kit v3 on the Illumina platform at the Advanced Studies in Genomics, Proteomics and Bioinformatics core facility at the University of Hawai’i at Mānoa. The sequencing run also included two no-template amplifications to assess contamination in the molecular analysis pipeline and a ZymoBIOMICS Microbial Community Standard (Zymo Research, Tustin, Ca).

Bioinformatic analysis was carried out with the MetaFlow|mics pipeline for microbiome marker data [28]. Raw paired fastq reads were preprocessed using the dada2 R package. Reads were filtered with the filterAndTrim() function, truncated at position 220 (190 for the reverse read), and discarded if they contained at least one base below quality 2 or a number of expected errors above 3. Denoising was performed using the learnError() and dada() functions with default parameters. Using the mergePairs() function, reads were merged if they overlapped by at least 20 bases, and a maximum of one mismatch was allowed. Mothur was used along with the Silva database (version 132) (downloaded from https://mothur.org/wiki/Silva_reference_files) to align and annotate the sequences. Sequences with a start or stop position outside the 5^th – 95^th percentile range (over all sequences) were discarded. Potential chimeras were removed with chimera.uchime() and clustered at 100% similarity thresholds with chimera.vsearch(). Taxonomies were assigned using classify.seqs() and classify.otus(). The lulu R package was used to refine amplicon sequence variants (ASVs). Two ASVs were merged if each of the three conditions were satisfied: i) they co-occur in every sample, ii) one of the two ASVs has a lower relative abundance than the other in every sample, and iii) if they share a sequence similarity of at least 100%. Finally, all singletons (ASVs with reads in only one sample) and ASVs with no annotation at the kingdom level were removed.

Climate data (including the average monthly temperature, solar radiation, relative humidity, rainfall, cloud cover frequency, and the enhanced vegetative index) for the corresponding month of sampling were obtained for each of the 50 collection sites from the Hawai’i Climate Atlas. This resource generally collates climate variable data across several years to generate estimates of monthly averages across Hawai’i at a spatial grain of 250 m square plots (see Giambelluca et al. [23] for a specific description of methods). Principal component analysis (PCA) was used to collapse correlated variation between these climate variables along orthogonal axes. We characterized land-use by estimating the proportion of built-up structures vs natural/permeable surfaces using satellite imagery (Google, Maxar Technologies) in Google Earth at eye altitudes of 500–750m. We used the ruler tool to generate three 250-meter transects (a minimum dispersal range of Ae. albopictus [29]) beginning at the geocoordinate of each midgut pool sample. We determined the orientation of each transect as a random number between 0–360 and used the generated number as a directional bearing. We then used the ruler tool to estimate the proportion of built-up or impermeable surfaces (e.g. roads, parking lots, buildings, houses, etc.) visible along the transect in the satellite image. We calculated an index of “urbanization” as the mean proportion of built-up surface area among the three transects for each sampling point. In general, landscape-use patterns (i.e. the scale of human development) did not change drastically between the time at which the satellite images were taken and when mosquitoes were collected among the sampling sites reported here.

Generalized linear mixed models assuming a negative binomial error distribution were implemented in glmmTMB [30] to test individual bacteria ASV responses to climate (estimated by PCA score) and urbanization. All models including the null model included the following:

1. An offset of the log sum of total read counts to control for differences in sequencing depth among samples

2. A random intercept of ASV taxon to account for global variation in abundance reflected in raw read counts

3. A zero-inflation variable for ASV taxon

4. Fixed effects of climate PC scores and urbanization index

We controlled for potential spatial autocorrelation of these ASV read counts by fitting a spatial exponential covariance structure with the geocoordinates of the 50 sampling sites. We used AICc to select among candidate models that included spatially structured or unstructured covariance between collection sites, and a random intercept of ASV identity or a random slope for environmental index for each ASV taxa (see supplementary file for the annotated code associated with this analysis). We estimate a pseudo-R² (${\mathrm{\Omega }}_0^2$) for each model using methods in Xu [31]. Our analysis is detailed in the supplementary material as an R-markdown file.

We limit our mixed model analysis to ASVs that had a prevalence greater than 0.10 among the pooled samples. This cutoff was chosen a priori primarily to avoid model fitting errors associated with the random slope effects, and ensured that all ASVs had a minimum of five samples with non-zero read counts. As such, the results reported here pertain to the most widely distributed mosquito symbionts in this population, and ignore rare or severely geographically restricted taxa.

Results

PCA of climate variables from the 50 sampling sites revealed patterns of correlation consistent with a windward/mountain to leeward moisture gradient (Table 1). Specifically, Principal Component 1 (PC1), which explained 58% of the variance in these data, had positive loadings on temperature and solar radiation, and negative loadings on relative humidity, rainfall, enhanced vegetation index (EVI), and cloud cover frequency. Thus, PC1 scores separated sites along windward/mountain to leeward moisture gradient (Figure 1), with positive values associated with hot, sunny, and dry leeward areas and negative values associated with cool, cloudy, wet, and “green” windward and mountain areas.

Table 1.

Loadings on climate variables in a principal component analysis across the 50 sites where mosquitoes were collected

Climate Variable (Proportion of Variance)	PC1 (0.58)	PC2 (0.20)	PC3 (0.12)	PC4 (0.05)	PC5 (0.03)	PC6 (0.02)
Temperature	0.48	0.17	0.36	-	-	0.79
Solar Radiation	0.44	−0.36	0.27	0.13	−0.70	−0.31
Relative Humidity	−0.31	−0.69	−0.22	0.46	-	0.42
Rainfall	−0.48	-	0.15	−0.69	−0.44	0.28
Enhanced Vegetation Index	−0.34	−0.10	0.86	0.20	0.27	−0.17
Cloud Cover Frequency	−0.36	0.60	-	0.52	−0.49	-

A total of 6,984,003 reads passed through quality filtering, and were allocated among 3,968 ASVs (see Supplementary Table I). Sixty-one ASVs had a prevalence of greater than 0.10, and were included in subsequent analyses of the Ae. albopictus microbiome. Among these ASVs, there were 44 Proteobacteria, 8 Firmicutes, 4 Spirochaetes, 3 Bacteroidetes, and 2 Actinobacteria. wAlbB (the typically dominant Wolbachia lineage in Aedes albopictus) was the most abundant and prevalent ASV among the midgut pools. Additionally, we found evidence of Wolbachia superinfection (two or more strains infecting the same host) of wAlbB and wAlbA (Wolbachia sp.) in 70% of the midgut pools, although the rate of this superinfection would likely be higher if additional tissues were analyzed. Among the environmentally acquired symbionts, a Ralstonia sp. was the most abundant member of the microbiome and occurred in 92% of samples. Also common was a Paraburkholderia sp. (78% of samples) and an unclassified Spirochaetaceae lineage (94% of samples). Other members of the microbiome included several Enterobacteriaceae lineages, an Asaia sp., several Pseudomonas lineages, a Stenotrophomonas sp., a Carnobacterium sp., several unclassified Burkholderiaceae lineages, several Acintobacter sp., several Sphingomonas sp., and a Chryseobacterium sp. See Supplemental Table 1 for a comprehensive list of these 61 ASVs, including their corresponding labels (eg. OTUXXXX, where “OTU” denotes an operating taxonomic unit at the level of an ASV).

A model that included spatially structured covariance and a random slope of PC1 scores representing climate for each ASV taxa was best fitted to the data (AICc weight = 0.68), outperforming models that included i) a spatially structured covariance and random slopes for both climate and urbanization (ΔAICc = 2.0, AICc weight = 0.25), ii) a random slope for climate but a spatially unstructured covariance among sites (ΔAICc = 5.3, AICc weight = 0.05), and iii) a random slope for both climate and urbanization with a spatially unstructured covariance among sites (ΔAICc = 7.3, AICc weight = 0.02). Other models that did not include a random slope for climate for each ASV taxa were very poorly fit to the data (ΔAICc > 49, AICc weight ≈ 0), including a null model that excluded both the environmental random slopes and a random intercept for site (ΔAICc = 169.8). See Table 2 for the full set of candidate models and their AICc results.

Table 2.

AIC table and corresponding pseudo-R2 for generalized linear mixed-effects models

Model structure *	ΔAICc	df	Weight	Pseudo-R2 ** (${\mathbf{\Omega }}_0^2$)
Climate random slope, spatially structured covariance	0	10	0.68	0.72
Climate and urbanization random slopes, spatially structured covariance	2.0	11	0.25	0.72
Climate random slope, random intercept for site	5.3	10	0.05	0.67
Climate and urbanization random slope, random intercept for site	7.3	11	0.02	0.66
Random intercept for site	49.2	9	0	0.64
Urbanization random slope, random intercept for site	51.2	10	0	0.64
Spatially structured covariance	54.2	9	0	0.67
Urbanization random slope, spatially structured covariance	56.2	10	0	0.67
Null model	169.8	7	0	0.43

^*All models include an offset of log of the per sample total read count, a fixed effect of PC1, a fixed effect of urbanization, a random intercept for ASV taxon, and a zero-inflation variable for ASV taxon.

^**A pseudo-R² is estimated as ${\mathrm{\Omega }}_0^2$ (Xu 2003), which is computed as 1 minus the proportion of residual variance of a fitted model to the residual variance of the intercept only model, and is interpreted as a measure of the proportion of explained variation by the fitted model.

The best fit model, which included a random effect structure that accounted for spatially structured covariance and a random slope for climate, explained roughly 72% of the variation in the relative abundance of mosquito symbionts. By contrast, the null model, which included a random intercept and zero-inflation variable for ASV taxon and an offset for total read count per sample, explained only 43%. Additionally, the best fit model explained roughly 5% more variation than the model that only included a spatially-structured covariance. See Table 2 for the pseudo-R² of each model.

Conditional means associated with the random slope effect suggest that several taxa (e.g. Ralstonia sp [OTU002], Paraburkholderia sp. [OTU0005], Achromobacter sp. [OTU0030], Neokomagataea sp. [OTU0037]) are relatively more abundant in hot and dry environments while others (e.g. a Flavobacteriaceae lineage [OTU0010], several Enterobacteriaceae lineages [e.g. OTU0011, OTU0042, OTU0020], wAlbB [OTU0001], Carnobacterium sp. [OTU0012]) are more dominant in cool and wet environments (See Figure 2). See Supplementary Table II for a comprehensive list of the conditional means of the random slope for local climate.

Figure 2.

A heatmap demonstrating the proportion of reads of each symbiont taxa for each sampling site ordered on the x-axis from wettest to driest along the moisture gradient. ASVs along the y-axis are ranked by their response to the moisture gradient, indicated by their corresponding conditional means of the random effect to the Principal Component 1 score (right plot). Larger conditional means suggest a greater abundance in dry and hot regions, while smaller conditional means suggest greater abundance in wet and cool regions.

Discussion

The high islands of Hawai’i offer a unique opportunity to understand how ecological processes change along environmental gradients. This opportunity manifests with very stark gradients, including transitions from rainforest to desert and temperate to tropical thermal regimes over very short geographical distances, and a landscape patchwork of highly urbanized to natural areas. Here, we use this superb natural experiment to demonstrate that spatial turnover in the composition of the mosquito microbiome is strongly associated with local climate over spatial scales that are relevant to an individual host’s life cycle (i.e. less than a few kilometers).

The relative abundance of several common bacterial symbionts in the mosquito microbiome responded differently to a landscape-scale gradient in moisture. The ecological mechanisms that augment β-diversity of mosquito-associated microbiomes across this environmental gradient remain poorly resolved. Our study cannot distinguish if the local species pool of colonizing symbionts changes directly in response to environment and manifests as turnover in the mosquito microbiome, or, if certain microorganisms possess functional traits that increase host fitness in certain environments but not others. Future research should focus on identifying the components of the environmental microbiome that source the mosquito microbiome, and monitor how these microbiomes turnover across space in response to local environmental conditions. Moreover, we have a poor understanding of how the assembly of the mosquito microbiome is mediated by mosquito behavior, physiology, or differential fitness. Future studies should aim to uncouple direct and indirect effects (i.e. host-mediated) of environment on host-associated microbiome assembly, as it will have important implications toward understanding how drivers of β-diversity relate to microbiome function.

The proportion of built-up structures (i.e. urbanization) was not associated with the composition of the local mosquito microbiome. Several studies have demonstrated that common features of urban landscapes directly impact mosquito physiology [32, 33] and the mosquito microbiome [21]. However, we find no evidence that these apparent effects scale to impact the diversity of the mosquito microbiota across the landscape considered in this study. Importantly, our index of land-use focuses on the amount of built-up structures in the area. This metric effectively separates natural areas with very little human use from areas under intense anthropogenic influence but is not able to distinguish environments with moderate patterns of human use. For instance, semi-natural parklands and agricultural fields exhibited moderate surface proportions of built-up structures; however, these forms of land-use can differ greatly in the extent of human chemical and physical pollution, and in modified irrigation patterns that are expected to impact mosquito populations and their microbiome. It is possible that a different, and perhaps, finer assessment of human land-use may detect an effect on mosquito microbiome diversity in a real-world landscape.

We find that a structured covariance based on spatial distance between collection sites fit the relative abundance data for symbionts in the Ae. albopictus microbiome better than an unstructured covariance that modeled a simple random intercept for each site (Table 2). This indicates spatial autocorrelation exists in these data, even after controlling for a spatial moisture gradient. Unmeasured environmental variables may account for this residual spatial autocorrelation. In addition, patterns of dispersal impact the abundance and distribution of species across landscapes. The extent of dispersal limitation in the mosquito microbiome or of microbial symbionts in general is unresolved, but spatial structure in the relative abundance of bacterial symbionts is consistent with dispersal limitation [18]. In the mosquito microbiome, dispersal limitation would have implications for how a microbe might be expected to spread in a population on a real-world landscape, informing disease mitigation strategies associated with paratransgenesis [34]. Future studies must isolate the contribution of dispersal limitation to patterns of β-diversity among microbiotas.

Our results suggest a role for climate change in reshaping mosquito microbiomes in real-world landscapes [35]. Several studies have associated environmental heterogeneity with changes in mosquito biology. Furthermore, a few studies have demonstrated changes in mosquito traits along environmental gradients that are expected to influence the capacity of mosquito populations to sustain vector-borne disease transmission. Our result that a mosquito vector’s microbiome can change in composition over short geographical scales in response to stark variation in local environmental conditions suggests that shifts in climate may be associated with changes in the mosquito microbiome. As the microbiome is an important modulator of mosquito biology and disease-vector capability, this could provide a new dimension for climate change to transform the global ecologies of infectious diseases, including those that pose potent public health burdens and threats to wildlife conservation.

Availability of data and material.

Data used in this manuscript are available as supplementary files.