The Drosophila melanogaster strain BNE was derived from field-caught female flies from Brisbane, Australia, and is described in detail elsewhere (Brownlie et al., 2009; Yamada et al., 2011). Two genetically paired fly lines, one infected by wMel the other Wolbachia-free, were maintained at ~25°C on a 12/12 h light/dark schedule throughout the study. Tetracycline treatments were performed as described previously (Hoffmann et al., 1986) to generate a genetically identical fly line that lacked the Wolbachia infection. To reconstitute gut flora standardized methods were used (Chrostek et al., 2013) and all experiments were conducted at a minimum of seven generations post tetracycline treatment. To minimize genetic drift between these fly lines, approximately every 10 generations reciprocal crosses (BNE-wMel Female × BNE Tet male; BNE Tet female × BNE-wMel Male) were performed using 1-week-old adults. Fly lines were reared on three types of diets.

Cornmeal fly diet was made from yellow corn meal medium (Sigma). As described in detail elsewhere (Brownlie et al., 2009), the amount of available dietary iron was reduced by substituting the water that was used to make up the medium with an aqueous extract of black tea (Camellia sinensis)—which reduces iron availability by chelating free iron—or increased by the addition of a FeCl₃ solution to the cornmeal fly diet to a final concentration of 10 mM.

The total content of nine biologically relevant metals (manganese, iron, cobalt, nickel, copper, zinc, cadmium, lead, and arsenic) present in flies reared on each of the food types (listed above), was determined using inductively coupled plasma mass spectrometry (ICP-MS) at the Advanced Center for Isotope Research Excellence at The University of Queensland. The only metal responsive to diet was iron. Pools of ten 5-day old flies were used for each analysis (Wolbachia-infected and uninfected controls) and replicated 15 times. Comparisons of the average total iron levels across diet and infection status were made using two-way ANOVA (GraphPad Prism).

Wolbachia infection density in 5-day old flies reared on each food type were estimated using an established relative qPCR assay that compares the abundance of the single-copy Wolbachia ankyrin repeat gene WD0550 to that of the single-copy D. melanogaster gene Act88F (McMeniman et al., 2008). Briefly, ten adult flies reared on each food type were collected and genomic DNA isolated using a QIAGEN DNeasy Blood and Tissue Kit according to manufacturer instructions (QIAGEN, Doncaster, VIC) and target genes amplified using SYBR-Green pre-mix (Qiagen). Comparisons of the average relative density of Wolbachia were made using one-way ANOVA (GraphPad Prism).

A total of 20 five-day old mated female flies per sample were collected from each combination of strain (Wolbachia-infected, or uninfected controls) and diet (low-, standard-, or high-iron), snap frozen in liquid nitrogen and stored at −80°C. Table 1 lists the number of samples for each of these six groups. The frozen flies from each sample were funneled into pre-labeled microfuge tubes containing 400 μl of ice-cold acetonitrile (50%, v/v) and homogenized using silica beads in a tissue-lyser (QIAGEN) system. Samples were then centrifuged at 14,000 × g for 10 min at 4°C. The supernatant was transferred to another set of pre-labeled and pre-weighed microfuge tubes, and stored at −80°C until lyophilized.

The freeze-dried extracts were thawed and dissolved in 185 μL of 200 mM sodium phosphate buffer, pH 7.4, 10 μL 5 mM sodium 2,2-dimethyl-2-silapentane-5-sulfonate-d₆ (DSS) in D₂O as chemical shift reference, and 5 μL 5 mM difluorotrimethylsilanylphosphonic acid (DFTMP) in D₂O as internal pH indicator, yielding a final sample volume of 200 μL with final concentrations of 250 μM DSS, 125 μM DFTMP, and 7.5% D₂O. Samples were transferred into 3-mm NMR tubes for measurement. For sample numbers and group IDs see Table 1.

¹H NMR spectra were recorded on a Bruker AV900 NMR spectrometer (Bruker Biospin, Rheinstetten, Germany) operating at a ¹H frequency of 900.13 MHz and equipped with a 5 mm self-shielded z-gradient triple resonance probe. For each sample a 1D NOESY spectrum was acquired at 298 K with the noesypr1d pulse sequence [(RD) – 90° – t1 – 90° – τ_m – 90° – acq] (Bruker Biospin pulse program library). The transmitter frequency was set to the frequency of the water signal, and water suppression was achieved by continuous wave irradiation during both the relaxation delay of 3.0 s and the mixing time (τ_m) of 100 ms. After 16 dummy scans, 256 transients were collected into 32,768 data points using a spectral width of 14 ppm, leading to a total experiment time of 18.9 min per spectrum. Samples were manually changed after each 1D spectrum, and spectra were recorded in four different datasets. All spectra were processed using TOPSPIN version 3.5 (Bruker Biospin, Rheinstetten, Germany). The free induction decays (FIDs) were multiplied by a sine bell window function shifted by π/2 along the direct dimension before Fourier transformation, manual phase and baseline correction. The resulting spectra were referenced to the DSS signal at δ = 0 ppm.

The assignment of peaks to specific metabolites (Supplementary Table 1) was performed with Chenomx NMR Suite, version 8.4 (Chenomx Inc., Edmonton, Canada) as well as by comparing spectra with the online databases Human Metabolome Database and Biological Magnetic Resonance Bank (Wishart et al., 2018, Ulrich et al., 2007). In addition, the assignment of metabolites to peaks in the 1D loadings plots of the multivariate statistics (see below), was aided by calculating the covariance matrix (STOCSY) (Cloarec et al., 2005a) at full spectral resolution—either over all NMR spectra, or only over the groups of spectra that were part of the respective multivariate model. This analysis also uncovered covariation between different individual metabolites.

Assignments were confirmed by 2-dimensional (2D) ¹H–¹³C heteronuclear single quantum coherence (¹³C–HSQC) spectra measured on selected samples. The 2D NMR spectra were acquired on a Bruker Avance 900 spectrometer operating at a ¹H frequency of 900.13 MHz, equipped with a 5 mm self-shielded z-gradient triple resonance cryogenic probe. In all 2D spectra, the ¹H carrier frequency was positioned on the water resonance. The ¹³C–HSQC and experiments were performed with spectral widths of 14.03 ppm in the ¹H dimension. The spectral width in the ¹³C dimension were 110 or 120 ppm, and the ¹³C carrier frequency was set at 45 or 50 ppm, respectively. A total of at least 128 increments with 128 transients were recorded into 4,096 data points in the direct dimension, and a relaxation delay of 1.1 s was used. GARP decoupling of the ¹³C channel was used during the acquisition time. The spectra were multiplied by a squared sine bell window function shifted by π/2 along the direct and indirect dimensions before two-dimensional Fourier transformation.

To correct for pH and ionic strength-dependent shift variations of individual NMR signals, the ¹H NMR spectra were aligned in specific segments with the icoshift algorithm (Savorani et al., 2010) in MATLAB (MathWorks, Massachusetts, U.S.A) and then data-reduced with an in-house MATLAB script to consecutive integral regions of 0.001 ppm width (“buckets”), covering the range of δ = 10.0–0.25 ppm. The chemical shift region at δ = 5.10–4.67 ppm was excluded to eliminate the effects of imperfect water suppression. For each spectrum, the resulting integral regions were normalized to the total intensity of the spectrum to correct for inter-sample differences in weight and dilution. Subsequently, the bucketed data matrices were imported into the SIMCA 16 software package (Sartorius Stedim AB, Umeå, Sweden) for multivariate statistical analysis.

The data matrix with the bucketed 1D spectra was Pareto scaled for the subsequent analysis. An initial principal components analysis (PCA) was performed to determine whether there were any sample outliers and to investigate inherent differences in the samples (Hotelling, 1933). Nine samples were considerably outside the tolerance range of three standard deviations of the average score values in at least two latent components. Three additional samples were outside the 99% Hotelling's T2 limit. All twelve samples were removed as outliers in the subsequent analysis.

To maximize the distinction between classes, a supervised orthogonal partial least-squares (OPLS) analysis was performed on all remaining spectra (M1) (Trygg and Wold, 2002). To characterize global differences as a result of infection or diet, three OPLS models were fitted, one comparing all infected to all uninfected samples, irrespective of diet (M2), and one model each comparing all low- vs. standard-iron and standard- vs. high-iron diet samples, respectively (M3, M4), irrespective of infection status. Finally, to compare systematic differences between pairs of sample groups in more detail, a series of pairwise OPLS models was subsequently fitted (M5-M13). Supervised multivariate methods need additional input data about the class membership of individual samples, which are provided in the form of a Y–table, against which the PLS or OPLS algorithm performs a regression of the X-matrix. In our study the group identity of the six sample groups (0–5) was used as the Y–table for M1, and M2-M13 used the Wolbachia infection status (0 or 1) or the type of diet (0, 1, or 2), as appropriate.

In SIMCA, the number of OPLS components (A) for all models was optimized by cross validation. R²X, R²Y, and Q² were used to evaluate model quality. R²X and R²Y are the fraction of the sum of squares explained by the latent components in the model, representing the variance of the X and Y variables, respectively, and Q² is the predictive ability parameter of the model, which is estimated by cross validation. All OPLS models were validated by CV-ANOVA (Eriksson et al., 2008). The figures of merit of all multivariate models are listed in Supplementary Table 2.

Scores plots and loadings plots were used to interpret the various OPLS models. Spectral features representing elevated metabolite levels were identified from bivariate 1D loadings plots, in which loadings coefficients p were plotted against the chemical shift values of their respective variables, and the correlation-scaled loadings coefficients |p(corr)| were superimposed on the loadings plot as a heatmap color scale (Cloarec et al., 2005b). These plots contain the same information as a traditional S-plot, which indicates which variables influence the model with high reliability and are of relevance in the search for significantly altered metabolites.

To obtain a list of significantly changing metabolites (Supplementary Table 3), we used a combination of the variable importance in projection (VIP) score, |p|, and |p(corr)| (Galindo-Prieto et al., 2014). Generally, variables with VIP scores greater than 1 are considered significant (Eriksson et al., 2013). This criterion was converted into cutoffs for |p| and |p(corr)| by inspecting the S-plots for variables with VIP values of >1. For all diet pairwise comparisons (Models M8-M13), cutoffs of |p| ≥ 0.015 and |p(corr)| ≥ 0.6 were defined. For all Wolbachia v. uninfected pairwise comparisons (Models M5-M7) a cutoff for |p| ≥ 0.015 and |p(corr)| ≥ 0.45 was defined. The respective cutoffs for the global comparisons (Models M2-M5) were |p| ≥ 0.015, and |p(corr)| ≥ 0.5 for the diet comparisons (Models M4-M5) and |p(corr)| ≥ 0.4 for the Wolbachia v. uninfected comparison (Model M2).

The list of metabolites observed to change significantly in the multivariate analysis (Supplementary Table 4) was then further explored using univariate statistical analysis. Only metabolite signals with minimal peak overlap were used in this analysis (Supplementary Table 5). Metabolite signal intensity in each spectrum was calculated by peak integration of the spectral region(s) containing signals of the respective metabolite with an in-house MATLAB script on the non-normalized full resolution spectra. Subsequently, the signal intensities from each spectrum were averaged for each of the six experimental groups. The resulting average signal intensities for each metabolite are listed in Supplementary Tables 6, 7. In both tables, average signal intensities lower than the limit of quantification (LOQ) for the respective metabolite were labeled “ < LOQ” and omitted from further analysis. The LOQ for each metabolite was determined as follows. The noise level for each spectrum was calculated as the standard deviation of the baseline noise in the region between δ = 11−10 ppm. The noise levels for all spectra were then averaged to obtain the mean baseline noise level. This value is the error that is contributed by the baseline noise to the intensity of each data point of a spectrum. To obtain the total error contributed by the baseline noise to the integrated intensity of a metabolite, the average baseline noise level was multiplied by the number of data points in the integration regions for the particular metabolite (Supplementary Table 5). The LOQ was then conservatively defined as 5 times the resulting value, i.e., the threshold at which the baseline noise contributes more than 20% error to the intensity integral of a metabolite.

For each pairwise comparison the relative fold change in signal intensities was calculated as the ratio of the averaged signal intensities between the two groups (Supplementary Table 8). To determine significantly changing metabolites across each of the nine pairwise comparisons, the signal intensities of each metabolite (Supplementary Tables 6, 7) were subjected to Mann–Whitney U-tests. The Benjamini–Hochberg method for multiple testing correction was then applied with a false discovery rate of 0.05 on the obtained p values. Metabolite intensities with a p value less than the calculated critical value, Q, were deemed significantly altered in the respective comparison, and the Benjamini–Hochberg corrected p values are listed in Supplementary Table 9.

Genetically paired lines of Drosophila melanogaster (BNE) that differed by their Wolbachia (wMel) infection status were reared on one of three diets: a standard fly diet that contained black tea extract reducing the available dietary iron, standard fly diet, or a standard fly diet supplemented with additional iron (Table 1). The total content of iron, manganese, cobalt, nickel, copper, zinc, cadmium, lead and arsenic was determined using mass spectrometry in 5-day old flies. Only the total iron content was affected (Supplementary Figure 1), as had been observed previously (Brownlie et al., 2009). Furthermore, there were no significant differences in iron levels when comparing uninfected controls and Wolbachia-infected insects. To determine if altering dietary iron influenced Wolbachia densities within 5-day old flies, a standard quantitative-PCR assay was used (McMeniman et al., 2008). Wolbachia densities were equivalent in hosts reared on all diets (Supplementary Figure 2). Taken together, these two results mean that any metabolic changes we observe across the three different diets are directly related to the iron content in the insect.

Multivariate statistical analysis (MVSA) was used to investigate whether there are any systematic differences between the metabolite profiles of the six experimental groups of insects. After identifying 12 outliers in a preliminary PCA, an OPLS model of all remaining samples (model M1) showed definitive differences in the metabolite profile between all six experimental groups, and that all groups can be clearly distinguished from each other in the 3D scores plot (Figure 1). Note that the relationship between the six experimental groups to each other is unconstrained in the OPLS model, and thus their spatial arrangement and the distances between the centers of the six sample groups (Table 2) reflect the biological effects that distinguish the groups from each other. Strikingly, the spatial arrangement of the six groups with respect to each other in the 3D scores plot (Figure 1) is such that none of the first three latent components (t[1], t[2], or t[3]) align directly with the two factors of the study—Wolbachia infection, or dietary iron levels. In fact, five latent components are needed to fully describe the spatial arrangement of all groups with respect to each other. This arrangement of the sample groups, and especially the fact that the six groups are arranged in a non-planar manner shows immediately that (a) there are prominent interactions between Wolbachia infection and dietary iron levels, and (b) the effects of dietary iron are non-linear.

The Euclidean distances between the six groups (Table 2) highlight that the effect of dietary iron levels is more pronounced in the Wolbachia-infected insects than in the uninfected controls. This bears witness to Wolbachia's dependency on the host for iron supply. The biological effect of Wolbachia infection is highly variable across the three different dietary conditions. It is the smallest in the context of the standard-iron diet. In the context of the high-iron diet the effect of infection is among the largest biological effects in the study. This highlights that the effects of infection are highly dependent on the diet context and that there are intricate interactions between infection and diet, which will complicate attempts to look at the two effects independently.

In a next step, we nevertheless attempted to characterize any overarching effects of Wolbachia infection or dietary iron levels in the study. We constructed three OPLS models that looked at each effect independently. Two models investigated the effects of diet, irrespective of infection (standard- vs. high-iron and low- vs. standard-iron diet), and the third model characterized the effects of infection, irrespective of diet. Clear group differences were noted in the scores plots of all three OPLS models (Supplementary Figure 3). In the corresponding loadings plots (Supplementary Figure 4), more dramatic global effects were seen for dietary iron levels than for Wolbachia infection.

Looking at the effect of diet first, the differences between the low- and standard-iron diet (Supplementary Figure 4B, model M4) are characterized by increased levels of complex carbohydrates in the standard-iron diet, and decreased levels of proline, glycine, and propionate. The differences between the standard- and the high-iron diet (Supplementary Figure 4A, model M3) are more severe, with the most noticeable changes being an increase in the levels of both saturated and unsaturated lipids, as well as increased levels of trehalose and tyrosine, and a decrease in the levels of maltose and complex carbohydrates, next to decreasing levels of inosine, guanosine, ortho-phosphocholine, choline, and dimethylamine.

Overall, the differences in dietary iron levels are predominantly reflected in alterations in carbohydrate and amino acid metabolism when comparing low- and standard-iron diets, but are dominated by lipids between standard- and high-iron diets. These overarching effects indicate that a metabolic switch occurs between carbohydrate and lipid metabolism when insects are subjected to iron deficiency vs. iron overload.

Interestingly, the levels of trehalose and lipids are strongly correlated with each other in this study. Thus, in a one-dimensional STOCSY plot the covariance between the anomeric signal of trehalose and all lipid signals is about 0.68, nearly as much as the covariance with the rest of the trehalose signals (0.7–0.8), which are heavily overlapped in the NMR spectra, leading to their correlation to the anomeric proton signal being statistically diluted by contributions from other metabolites. Maltose and complex carbohydrates show a biphasic non-linear behavior with respect to diet, as levels of maltose and complex carbohydrate are highest in the standard-iron diet and lower in both the low- and high-iron diet.

The overarching effects of Wolbachia infection were less pronounced. The metabolites that pass the significance filter (Supplementary Figure 4C, model M2) are generally only weakly correlated with the infection status of the insects—a stark contrast to the pronounced effects associated with dietary iron. Decreased levels of several nucleosides, glycerophosphocholine, asparagine and choline were observed in infected insects together with an increase in trehalose. These weak overarching effects of infection arise due to the strong interactions between Wolbachia infection and dietary iron, and thus no highly correlated overarching metabolite changes were observed when comparing all infected insects to all uninfected controls—a picture also reflected in the Euclidean distances of the global OPLS model.

Due to the intrinsic interplay between Wolbachia infection and dietary iron, it was necessary to examine the detailed effects of infection in the context of the three different diets, and at the differences between the three diets in the context of Wolbachia-infected insects and uninfected controls. Nine pairwise OPLS models (M5-M13) were subsequently fitted to investigate the effects of diet on infection status and vice versa (Supplementary Figure 5). Metabolites significantly changing in these nine pairwise models are listed in Supplementary Table 4. To complement the metabolite alterations identified in these multivariate models, univariate statistical analysis (UVSA) was also undertaken to analyse these metabolite changes semi-quantitatively and to calculate fold-changes and p-values for the significance of the fold-changes (Supplementary Tables 5–9).

In this study we have characterized the metabolic changes as a result of Wolbachia infection in Drosophila melanogaster under different levels of dietary iron. The Euclidean distances between experimental groups in the global analysis already provide some insight into the relative biological effects. For example, the metabolic differences between infected insects and uninfected controls under standard iron conditions was the smallest distance in the whole study. This highlights that Wolbachia places only a minimal burden on Drosophila under standard-iron conditions, and reflects Wolbachia's adaptation as a mutualistic symbiont. Furthermore, the standard diet is nutritionally rich, and as such masks any obvious Wolbachia-host effects. In contrast, the high-iron diet is substantially different to the standard- and low-iron diets in both the uninfected controls and the Wolbachia-infected insects. This reflects the impost that iron levels in excess of the natural environment have on both groups. Of further interest are the distances between dietary conditions, which are always larger in the Wolbachia-infected groups when compared to the respective distance in the uninfected controls. This is especially evident under conditions of iron deficiency, as the distance between low- and standard-iron conditions in Wolbachia-infected insects is almost twice the corresponding distance for the uninfected controls. This bears witness to Wolbachia's dependency on the host for iron and the extra demands it makes on the host for iron supply.

The Euclidean distances are a result of marked metabolite profile differences between the six groups of insects in this study. In both Wolbachia-infected insects and uninfected controls, differences in dietary iron levels lead predominantly to alterations in carbohydrate and amino acid metabolism, between low- and standard-iron diets, but are dominated by lipids between standard- and high-iron diets. This indicates that a metabolic shift occurs between carbohydrate metabolism and fatty acid metabolism when comparing low- and high-iron levels, highlighting fundamental changes in metabolism. Interestingly, this metabolic shift is slightly different in Wolbachia-infected insects and uninfected controls. Across the board, the changes in carbohydrates and lipids are accompanied by slightly different sets of other metabolites. In addition, under iron deficiency the changes in carbohydrate metabolism are dominated by changes in glucose levels in the uninfected controls, whereas in the Wolbachia-infected insects levels of maltose and complex carbohydrates are affected.

When considering the biological and regulatory mechanisms that are likely to underpin the observed metabolic changes, it is worthwhile to keep in mind that metabolic effects that occur in both Wolbachia-infected insects and uninfected controls are likely to result from regulatory mechanisms in Drosophila itself. Examples of this are the strongly increased lipid levels and decreased levels of complex carbohydrates in the high-iron diet which are observed in both Wolbachia-infected insects and uninfected controls. Our observation is confirmed by other studies who have also observed lipid accumulation under high-iron conditions in a variety of organisms, including Drosophila (Navarro et al., 2010; Wang et al., 2016; Rockfield et al., 2018) and humans (Ahmed et al., 2012). Unsurprisingly, lipid and carbohydrate metabolism are intrinsically interlinked with the energy metabolism and mitochondrial function of the cell. A potential mechanism for this link between different parts of the metabolism is the outer mitochondrial membrane protein mitoNEET, which modulates mitochondrial function and iron content (Ferecatu et al., 2014). Upon iron excess, mitoNEET is upregulated leading to lipid uptake and storage within the host as well as to downregulated β-oxidation, and thus to increased adipose tissue mass (Kusminski et al., 2012). In addition, upregulated mitoNEET increases the rate of glycolysis (Kusminski et al., 2012; Wang et al., 2017), which could explain the lower levels of complex carbohydrates observed in our study. Furthermore, upregulated mitoNEET inhibits iron transport into the mitochondrial matrix, thus limiting both oxidative phosphorylation and ROS damage.

Under iron-deficiency conditions one might expect to see metabolic adaptations that lead to conservation of iron, iron scavenging, and a move to or reliance on metabolic pathways that do not require iron, such as glycolysis. Indeed, alterations to carbohydrate metabolism are among the most prominent metabolic changes we observed in both infected and uninfected insects under the low-iron diet. Interestingly, mitoNEET could also be involved in these metabolic alterations. Recently, a glycogen branching enzyme was established to be a key regulator of iron homeostasis under iron-depleted conditions in Drosophila. This enzyme controls cellular iron homeostasis via binding to the Iron-regulatory protein 1A (IRP1A) and mitoNEET (Huynh et al., 2019). This link between cellular iron homeostasis and glycogen metabolism may explain why carbohydrates change significantly upon iron deficiency. Additionally, the interaction between the glycogen-branching enzyme, mitoNEET, and IRP1A leads to the downregulation of iron-intensive processes as well as increasing the levels of bio-available iron for the cell (Huynh et al., 2019).

There are marked differences in the affected metabolic pathways between Wolbachia-infected insects and controls, which are also dependent on the iron levels in the diet. Indeed, when comparing the metabolite profiles of infected insects and uninfected controls, the major metabolites affected are carboxylic acids, nuclosides, cholines, and amino acids in the low- and high-iron diets. In contrast, in the standard-iron diet the major metabolic effect of infection is a decrease in proline levels. Taken together, these observations mean that the detected metabolic differences between the six study groups are driven by an interaction of both the Wolbachia symbiont and the iron levels inside the host cells. This highlights that signaling pathways and/or mechanisms that regulate iron homeostasis in the cell play a pivotal role in the interaction between Wolbachia and the host.

Earlier studies have shown that Wolbachia affects iron homeostasis of the host by regulating the expression of ferritin (Kremer et al., 2009; Gill et al., 2014). Iron homeostasis and ferritin expression are regulated in multiple organisms including D. melanogaster via the insulin/insulin-like-growth factor signaling (IIS) pathway and the hypoxia signaling pathway (Ackerman and Gems, 2012; Nässel et al., 2015; Altintas et al., 2016). Based on these studies and our results, we postulate that Wolbachia affects both of these pathways in the host. Previous studies have predicted that Wolbachia competes for iron as well as oxygen with the mitochondria of the host cell to run its own oxidative phosphorylation, leading to host cells becoming hypoxic in the presence of Wolbachia (Kremer et al., 2009; Gill et al., 2014; Dutra et al., 2017). Hypoxia depresses the IIS pathway (Wong et al., 2014; Texada et al., 2019; Barretto et al., 2020) and upregulates the hypoxia signaling pathway, as shown in Figure 5. The hypoxia signaling pathway, mediated by hypoxia-inducible factor 1 (HIF-1) (Mylonis et al., 2019), represses ferritin expression, leading to increased bioavaliability of iron in the cell (Kremer et al., 2009). A previous study has indeed shown reduced ferritin expression in Wolbachia-infected host cells (Kremer et al., 2009). Wolbachia competes for the intracellular iron to maintain oxidative phosphorylation, which will lead to the production of reactive-oxygen species (ROS) by the symbiont (Gill et al., 2014). In addition, the increased iron bioavailability in the host cells leads to secondary ROS production (Gill et al., 2014). In support of this interpretation, we have previously shown that ROS are being induced as a result of Wolbachia infection even in co-evolved hosts such as Drosophila (Wong et al., 2015). These metabolic effects observed here are likely to be driven by Wolbachia, modified by the context of the dietary iron levels, rather than being a primarily iron-driven response.

Interestingly, the ATP generated by Wolbachia for its own purposes can potentially be provided to the host (Darby et al., 2012; Gill et al., 2014). In accordance with that hypothesis we observed during Wolbachia infection at low-iron conditions a significant increase of adenosine levels, which are likely to be a proxy for the whole pool of adenosine, AMP, ADP, and ATP, due to our extraction procedure. The increase in adenosine levels we observed is arguably from the symbiont, because recent studies noted that Wolbachia infection in Drosophila yields significantly reduced ATP levels in the host (Carneiro Dutra et al., 2020, Carneiro Dutra et al., 2020).

On the other hand, depressing the IIS pathway leads to the induction of mechanisms that will modulate the increased ROS production induced by Wolbachia infection (Van Heemst, 2010). This aligns with an earlier hypothesis (Zug and Hammerstein, 2015) that in native hosts Wolbachia infection is associated with both increased ROS production and increased mechanisms of redox homeostasis and ROS protection. Firstly, the hypoxia-induced reduction of the IIS pathway leads via phosphoinositide 3-kinase (PI3K)/Akt signaling to an activation of the daf-16/FoxO transcription factor (Barretto et al., 2020), which mitigates oxidative stress resistance through regulating antioxidants (Mattila and Hietakangas, 2017) and increasing ferritin expression (Ackerman and Gems, 2012). The latter effect is likely to provide some counterbalance to the ferritin repression triggered by hypoxia and HIF-1. Secondly, proline catabolism is promoted, which regulates both ROS production and homeostasis (Zarse et al., 2012; Tang and Pang, 2016). In agreement with this interpretation, we observed significantly reduced proline levels in Wolbachia infected insects on the standard iron diet. However, this Wolbachia-related reduction in proline levels is diet dependent, as significant proline accumulation is promoted upon iron deficiency (Kitajima et al., 2003). Likewise, we observe in our data less pronounced changes in proline levels in low-iron conditions.

Reprogramming of lipid metabolism occurs under HIF-1 and insulin signaling. Upregulation of HIF-1 enhances lipogenesis through modulation of proteins involved in fatty acid uptake, synthesis, storage and usage, and diminishes fatty acid catabolism (Mylonis et al., 2019). Similarly, the disruption of the IIS pathway can also induce lipogenesis via the daf-16/FoxO transcription factors (Perez and Van Gilst, 2008). Subsequently, the fatty acids are converted into triacylglycerols and stored in lipid droplets (Mylonis et al., 2019). An increase in localized lipid droplets has indeed been noted as a result of Wolbachia infection (Geoghegan et al., 2017). Similarly, we observed increased lipid levels in Wolbachia-infected insects compared to the uninfected controls under standard and high-iron diet conditions.

When oxygen is sparse, HIF-1 upregulation leads to inhibition of the pyruvate dehydrogenase complex decreasing the flow of pyruvate into the TCA cycle. In uninfected Drosophila, this causes an accumulation of lactate, acetate and alanine as anaerobic end-products (Feala et al., 2007). Wolbachia extracts alanine from the host as precursor of several other metabolites, as it cannot synthesize alanine itself (Jiménez et al., 2019), thus redirecting metabolic flux away from lactate and acetate. Consistent with this interpretation, we observed reduced levels of alanine, acetate and lactate in Wolbachia-infected individuals under standard-iron conditions. Keeping host cells in a hypoxic state that favors the production of an essential metabolite for Wolbachia is a strategy that is advantageous for the bacterium.

Wolbachia is able to synthesize riboflavin and provision it to the host (Jiménez et al., 2019). Riboflavin supplementation significantly contributes to prolonged lifespan, growth and fecundity in both Cimex lectularius and Drosophila melanogaster (Moriyama et al., 2015; Zou et al., 2017)—effects that are similar to Wolbachia infection. In several bacterial species there is evidence for increased riboflavin biosynthesis overcoming iron-restrictive conditions (as reviewed in Sepúlveda Cisternas et al., 2018). The riboflavin biosynthesis pathway starts with GTP, which in turn is formed through the purine biosynthesis pathway from IMP. Thus, increased riboflavin production after Wolbachia infection would be associated with a decrease of IMP and GTP levels compared to the uninfected controls. Indeed, during Wolbachia infection in iron-deficient conditions we observed a significant reduction of both inosine/IMP and guanosine—supporting the hypothesis of riboflavin provisioning by Wolbachia.

Wolbachia infection under low-iron conditions leads to a significant increase of propionate levels with a smaller and correlated increase in acetate levels. This could arise from two potential metabolic pathways. The first possibility is malate dismutation, an ATP generating pathway known in helminths, and nematodes, including C. elegans in which the second half of the TCA cycle runs backwards under low-oxygen conditions, leading to the production of both propionate and acetate (Müller et al., 2012; Stairs et al., 2015). Although the malate dismutation pathway has not been actively studied in either Drosophila or Wolbachia yet, both organisms possess the majority of enzymes required for this pathway, including some enzymes whose presence only seems to make sense if the whole malate dismutation pathway is present. In agreement with this hypothesis, Drosophila are strongly hypoxia and anoxia tolerant (Krishnan et al., 1997; O'Farrell, 2001; Haddad, 2006). Moreover, depressing the IIS pathway in C. elegans leads to high propionate levels (Fuchs et al., 2010). Thus, malate dismutation is a possible contender for the observed propionate production. Secondly, Drosophila possess a direct metabolic pathway for production of propionate from pyruvate (Kanehisa and Goto, 2000), and thus the production of acetate and propionate could be a simple follow-on from anaerobic glycolysis. In both of these scenarios, the production of propionate and acetate is more likely to be an outcome of Drosophila metabolism relying on anaerobic pathways to compensate and accommodate the reduced availability of oxygen induced by Wolbachia infection.

It is interesting to note that the metabolic alterations as a consequence of Wolbachia infection described here are characteristic of metabolic reprogramming, rather than an immune response. A study by Wong et al. confirms this observation as they noted no measurably increased expression of immune genes in wMel-infected Drosophila (Wong et al., 2011). In other words, there is no priming of the immune system by wMel. In addition, we see no evidence for a loss of lipogenesis and glycogenesis in our study, which are metabolic markers for an immune response (Clark et al., 2013; Davoodi et al., 2019). However, we do observe an effect of diet on these pathways irrespective of infection status. Thus, the observed metabolic changes are a result of host-symbiont-environment interactions, rather than an immune response by the host. The above findings also fit with the general hypothesis by Zug and Hammerstein who postulated that strains of Wolbachia that have coevolved with their respective hosts do not stimulate an immune response (Zug and Hammerstein, 2015).

Previous studies that have compared host and Wolbachia genomes have predicted biochemical and physiological interactions unique to host and symbiont (Min and Benzer, 1997; Foster et al., 2005; Gill et al., 2014; Kosmidis et al., 2014; Newton and Rice, 2020). Indeed, in other host-Wolbachia pairs interactions range from facultative parasitism to obligate mutualism (Gill et al., 2014). Nevertheless, iron has been repeatedly shown to be of crucial importance for the interaction between Wolbachia and host (Gill et al., 2014), and thus the regulatory pathways implicated here are likely to have broader ramifications for Wolbachia-host pairings in general, or potentially even other endosymbiont-host interactions.

In summary, the presence of Wolbachia depresses the IIS cascade in D. melanogaster whilst inducing the hypoxia signaling pathway. This in turn causes both ROS production and ROS adaptations, as well as other metabolic changes, that steer metabolism away from oxygen-intensive pathways and enable metabolite extraction by the symbiont and metabolite provisioning to the host (Figure 5). Finally, this metabolic reprogramming is likely to be driven by Wolbachia, and modified by the dietary iron levels.

Earlier work has suggested that Wolbachia increases insulin signaling (Ikeya et al., 2009) rather than the downregulation inferred by our own work. However, these results were omitting important wild-type or Wolbachia-infected controls when assessing direct activation levels of Akt and FoxO, two of the key proteins in the IIS pathway. Furthermore, the authors' observation of an accumulation of triacylglycerides in the absence of Wolbachia is not supported by the experimental observations made by us—that lipid levels increase in the presence of Wolbachia. Increased IIS signaling in Wolbachia-infected Drosophila was also inferred by a recent study (Carneiro Dutra et al., 2020). However, that study used different strains of Wolbachia and Drosophila than those used by us, and reared flies on a carbohydrate-rich diet. Given the specificity outlined above of the metabolic interactions between strain, host and diet, any differences in results and interpretation are not surprising.

Our study has three limitations. Firstly, Wolbachia are known to establish high infection densities in the fat-body and adult brain of Drosophila (Albertson et al., 2013)—two tissues that are also known to be interlinked by the IIS pathway, supporting the likely involvement of this pathway in Wolbachia infection (Carneiro Dutra et al., 2020). However, because our experimental setup used the entire insect for NMR analysis, we cannot resolve any tissue-specific metabolic changes. Secondly, for the same reason, we can not directly distinguish metabolite changes associated with Wolbachia vs. the host. Lastly, we can only infer the regulatory pathways in our model from the metabolic changes observed here. Other proteins or signaling pathways also regulate iron, e.g., transferrin or iron-regulating proteins (IRPs) (Tang and Zhou, 2013; Cronin et al., 2019), but the pathways inferred here are supported by previous experimental observations (Kremer et al., 2009; Gill et al., 2014; Wong et al., 2014; Dutra et al., 2017; Texada et al., 2019; Barretto et al., 2020) and are thus the most probable explanation for our results. Future experiments aimed at direct manipulation of the IIS and hypoxia pathways will be able to confirm their involvement.

Apart from their direct manipulation as suggested above, the involvement of both the IIS and hypoxia signaling pathways as a consequence of Wolbachia infection opens up several exciting avenues for further discovery. Given the strain-host specificity discussed above, it would be interesting to apply the same NMR techniques to explore the metabolomics of different Wolbachia strain/host pairings, such as facultative Wolbachia strains that infect filarial nematodes, wMel that have been established in novel insect hosts, or Wolbachia strains (e.g., wAu) that impose no obvious effect on host reproduction (Hoffmann et al., 1996; Cao et al., 2019). Moreover, extending such investigations to pairings of Wolbachia and Aedes aegypti would be interesting, given the potential of the former in limiting the spread of tropical diseases carried by the latter (McGraw and O'Neill, 2013; O'Neill et al., 2018; Ryan et al., 2019). Furthermore, due to the strong involvement of lipid metabolism, a complementary lipidomics study would allow for greater metabolome coverage in that area of metabolism. If some of the metabolic effects are indeed localized to certain tissues, then performing tissue-specific metabolomics experiments is an obvious further step. Such experiments could also shed light on which of the metabolic changes are derived from Wolbachia metabolism and which are due to insect metabolism, given the preferential localization of Wolbachia in specific tissues. In addition, the present analyses are a static snapshot of Wolbachia infection, and as such changes in metabolite levels do not directly provide information about changes in metabolic fluxes, as the relationship between metabolic fluxes and metabolite levels is complex. Future fluxomics experiments are well-placed to resolve this point. In this context, it would also be intriguing to further explore the hypothesis that Wolbachia produces riboflavin for the Drosophila host cells, by designing fluxomics experiments that target specifically the fate of riboflavin. The metabolic interactions between Wolbachia and Drosophila also lend themselves to exploration via genome-scale modeling using available genome-scale metabolic models for both organisms (Coquin et al., 2008; Jiménez et al., 2019; Schönborn et al., 2019).