Dissecting the contributions of time and microbe density to variation in immune gene expression

Abstract

Widespread differential expression of immunological genes is a hallmark of the response to infection in almost all surveyed taxa. However, several challenges remain in the attempt to connect differences in gene expression with functional outcomes like parasite killing and host survival. For example, temporal gene expression patterns are not always monotonic (unidirectional slope), yielding results that qualitatively depend on the time point selected for analysis. They may also be correlated to microbe density, confounding the strength of an immune response and resistance to parasites. In this study, we analyse these relationships in an mRNA-seq time series of Tribolium castaneum infected with Bacillus thuringiensis. Our results suggest that many extracellular immunological components with known roles in immunity, like antimicrobial peptides and recognition proteins, are highly correlated to microbe load. On the other hand, intracellular components of immunological signalling pathways overwhelmingly show non-monotonic temporal patterns of gene expression, despite the underlying assumption of monotonicity in most ecological and comparative transcriptomics studies that rely on cross-sectional analyses. Our results raise a host of new questions, including to what extent variation in host resistance, infection tolerance and immunopathology can be explained by variation in the slope or sensitivity of these newly characterized patterns.

1. Introduction

Natural variation in host immunological defences drives heterogeneity in host resistance and infection tolerance, disease dynamics and the evolution of microbial virulence [1,2]. A central focus in the field of disease ecology, therefore, has been to characterize and quantify immunological variation across individual hosts and populations in both model and non-model organisms, and to discern the ecological and evolutionary factors [3] that give rise to immunological heterogeneity.

The decreasing cost of next-generation sequencing, combined with the difficulty of developing protein quantification tools for non-model organisms, has driven the popularity of transciptomics analyses for characterizing immunological variation [4,5] in populations or genotypes that demonstrate different phenotypes in response to microbial infection. In a typical experiment, hosts or their tissues from different treatment groups or genotypes are challenged with live or heat-killed microbes and then sacrificed at one or two time points post-challenge in order to harvest mRNA. These samples undergo deep sequencing, and the differential gene expression between treatments is compared to a control group or to each other (e.g. [6]). The interpretation of these results generally assumes that a group of individuals who have a higher fold change in immune genes have higher immunocompetence. However, this interpretation of cross-sectional data may err in cases where different groups differ in the temporal dynamics of microbial clearance, as individuals who do not clear an infection may have a higher level of immune gene expression than those who have already resisted their parasites and regained homeostasis [7]. Thus, higher immune gene expression, as with other markers of immunocompetence [8], may actually indicate lower functional resistance, depending on when the individuals are sampled relative to the dynamics of infection and on the induction and decay dynamics of immune gene expression.

While several studies have described patterns of temporal dynamics in insect immune gene expression after exposure to live or heat-killed microbes (e.g. [9–11]), even those that use live microbes rarely measure temporal changes in microbe density beyond the initial dose, despite the potential for interaction between differences in host resistance and differences in host gene expression. Moreover, recent evidence from Bombus terrestris [12] and Drosophila melanogaster [13] indicates that antimicrobial peptides (AMPs), key antibacterial effectors and the focus of innumerable ecological immunology studies are correlated to microbe loads. These results underscore the urgency for disentangling the contributions of temporal and microbial dynamics to immunological and resistance variation.

In this study, we develop a set of algorithms to disentangle the relative contributions of time and microbe load to patterns of systemic host gene expression across the acute infection phase. We take advantage of an mRNA-seq dataset [14] that we previously generated from red flour beetles (Tribolium castaneum) septically infected with live Bacillus thuringiensis (Bt). These hosts were periodically sacrificed from the time of inoculation through the phase of peak host mortality, yielding sequential cross-sectional samples from which both individual host gene expression and microbial density were determined. By delineating the relative temporal and microbial contributions to gene expression dynamics, our study highlights new considerations and approaches to quantifying immunological variation across individual hosts, genotypes and populations.

2. Material and methods

The data used in these analyses are derived from experiments designed to query differences in the dynamics of gene expression between trans-generationally primed and unprimed flour beetles [14]. Priming is a phenomenon that provides increased protection against microbes to which a host (or its parent) has been previously exposed, resulting in higher resistance or survival following infection. While our original analysis revealed some differential dynamics in gene expression between primed and unprimed individuals upon Bt infection (e.g. treatment effects upon the expression of genes driving metabolism), no genes associated with canonical immune pathways demonstrated priming-dependent patterns of expression [14]. We wondered whether the expression dynamics of canonical immune genes instead would reflect correlations to microbe density or non-monotonicity with regard to time, regardless of priming status. Therefore, in the present study, we decided to take advantage of our time series and its attendant microbe density data to investigate the prevalence of these patterns across the infection-induced transcriptome.

The T. castaneum beetles used in these experiments were derived from age-matched adult cohorts from a 2-year-old laboratory stock population originating from the Carolina Biological Supply (#144356). The creation of the mRNA-seq dataset analysed in this study is described in detail in [14] and in the electronic supplementary material. Briefly, we performed two separate experiments where we administered an LD50 dose of live B. thuringiensis serovar Berliner (ATCC 10792) vegetative cells (average 15 CFU/injection), grown overnight in Nutrient Broth medium (Sigma–Aldrich #70149), to larval T. castaneum in two treatment groups: immunologically primed (whose mothers had received a heat-killed dose of Bt) or unprimed (unmanipulated mothers). Our original analysis of this dataset [14] revealed that only a very small portion of genes induced after infection had differential expression attributable to differences in priming treatment, mostly relating to metabolic physiology and translational processes. No canonical immune genes were differentially expressed between primed and unprimed individuals.

The entire dataset includes 48 individuals, sampled over five time points from inoculation to the end of the mortality period (with 6 to 12 individual hosts per time point, ranging from 2 to 12 h post infection), as well as an uninfected cohort (n = 12, referred to here as 0 h). All sacrificed larvae were immediately frozen in liquid nitrogen and stored at −80°C. The mRNA-seq libraries were prepared, sequenced, aligned and assembled into count data as described in [14]. Raw sequencing reads and processed count data have been deposited as accession GSE95387 on NCBI's Gene Expression Omnibus [15]. Bt load (data publicly available in [16]) was also estimated from these samples against a host-normalized standard curve using RT-qPCR with Bt-specific primers as described in [14,17].

Studies that interpret differences in gene expression at a single point in time must assume a fairly linear monotonic (slope does not switch signs) or switch-like correlation between gene expression and time. To determine the validity of these assumptions, we considered two broad alternative patterns of gene expression: (i) genes that show non-monotonic induction and decay dynamics, and (ii) genes that correlate with microbe density. Time and microbe load largely covary across the acute infection period, but there was enough naturally occurring variation in microbe load across individuals, time points and treatments to supply leverage for disambiguating the two factors. Overall, microbe loads span over five orders of magnitude.

(a). Patterns of expression related to time and microbe load

The analysis pipeline is presented in figure 1. In short, outputs of a systematic suite of statistical models enabled classification of expressed genes into categories defining qualitatively different dynamics. Analyses were conducted on the full data as well as a subset that included all data except 12 h primed individuals (subset 1). Microbe loads for those individuals were very high relative to all other samples [14]. Because we could not exclude the possibility that these individuals were moribund or otherwise experiencing a fundamentally different physiological state than all of the other samples that fell within the lower microbe density distribution, and because they strongly inflated the favourability of the Bt density term in the linear models, we excluded them from some analyses as indicated through the use of ‘subset 1’. Genes that were likely to be related to Bt density, time or non-monotonic time were identified by running all gene count data from subset 1 through a linear model of the form log(expression) ∼ log(Bt) + time(linear) + polynomial time(second degree) using the R library DESeq2 [18]. Because transgenerational priming status and experimental ID did not have a significant effect on the expression of most genes or any immune genes [14], but did covary to some extent with microbe density and time, respectively, and because having two extra factors in the model reduced the statistical power for detecting the effect of the focal factors on higher-order gene expression dynamics, these two factors were dropped from the model. Neither of these dropped factors qualitatively affect the gene lists for each focal factor, and the results with the full model (log(expression) ∼ priming status + exp.ID + log(Bt) + time(linear) + polynomial time(second degree)) can be viewed in the electronic supplementary material, tables. We chose to log-transform the expression and Bt density data to reduce the skew of the distributions, which spanned several orders of magnitude, and because of previous observations of log-linear relationships between immune gene expression and microbial density in D. melanogaster [13]. To identify genes that could be related primarily to Bt density, we separated genes that had significant FDR-adjusted p-values for Bt density but not for the time terms.

Figure 1.

Flow chart for the identification of genes that belong to time, Btload, rapid-U (induction–decay) or polytime (slow induction–decay) clusters of differential expression following septic infection with Bt. The original data (data) were subsetted (‘subset 1’, no primed t = 12 (high microbe) data) to facilitate conservative identification of key patterns. Subsetted gene expression counts were fitted with the linear model (∼Bt + time + polytime) to identify genes explained by these factors. The interval t = 2 to t = 12 was also fitted with an approximate time function to identify genes that had slopes significantly different from 0 over time, and this dataset was then compared with genes that showed significant differential expression between 0 and 2 h post infection, to identify genes that show rapid induction and decay dynamics (rapid-U). Detailed information on the analysis can be found in the material and methods section. Red indicates the microbial density dynamics in each dataset over time, whereas blue indicates gene expression dynamics over time.

To identify genes that were likely to show monotonic trends with time, we parsed the genes from the linear model that had strong statistical support for the linear time term, but no significant support for the Bt term and fairly weak support for the second-degree ‘polytime’ term. We found that genes that had strong support for both time and polytime tended to show decay dynamics late during acute infection. While the polytime term was capable of capturing non-monotonic trends that peaked around 6–8 h post infection, we wish to emphasize that the change in the slope direction of gene expression over time coincides with the initial phase of host recovery from infection. Thus, it is unknown whether these genes would still show non-monotonic expression dynamics over this time period if the host were faced with a more persistent infection. Therefore, we have grouped all but the most strongly supported polytime genes (which generally peak at 6 h rather than 8 h) with linear time until future experiments can confirm that their non-monotonic expression patterns are not driven by host recovery alone.

However, a gene that is induced and decays while bacteria are still proliferating is highly likely to be truly non-monotonic in expression. This rapid induction and decay pattern (hereafter ‘rapid-U’) was not well captured in the polytime term of the linear model due to the temporal skew of the peak expression. To capture this pattern, we identified genes that were significantly differentially expressed between 0 and 2 h in both directions, and then showed a significant expression pattern over time in the opposite direction subsequently (2–8 h). A visual inspection of the results revealed a high fidelity to the pattern that we intended to capture (figure 2). We subtracted the results with strong statistical support (p < 0.05 for both steps of pattern definition) in this analysis from the Bt density and time gene lists (although, in practice, this minimally affected all lists because there was little overlap) to improve the categorization. Our analysis is skewed towards the over-conservative capture of genes that are highly likely to be Bt or induction–decay driven, while excluding genes that did not have sufficient statistical support to be categorized into these groups, even if they may belong there.

Figure 2.

Examples of genes that correlate to microbe density, time or non-monotonic time. The gene TC007738 (attacin-2) reflects microbial dynamics over time (a) and is highly correlated to microbe density (b). The gene TC000247 (a serine protease) is significantly correlated to time (c) but not to Btload after accounting for the main effect of time (d). The gene TC015051 (fatty acyl Co-A reductase family) shows a pattern of slow induction and decay dynamics over time (e), as fitted with a polynomial (quadratic) time factor, but has no correlation with microbe density (f). The gene TC010851 (imd, g) shows rapid induction-and-decay dynamics over time, with another peak in three insects with very high microbe loads (12 h primed), but is not otherwise highly correlated to microbe load (h). None of these genes, or others in the categories, showed a significant association with maternal priming treatment after accounting for microbe load. Red lines indicate the median count value for each time point. (Online version in colour.)

The gene sets for each of the four patterns were annotated with Gene Ontology (GO) accession terms, Interpro domain hits and wikigene information using the BiomaRt package in R [19]. These sets were subsequently subjected to GO enrichment using g:Profiler [20], with the Fisher's exact test implementation to identify biological processes, molecular functions and cellular components associated with each pattern.

(b). Comparing our data to previous work and revising to account for new models

Previous transcriptomics studies on T. castaneum have identified suites of genes in the Toll and IMD pathways as well as other canonical immune pathways that are differentially expressed 6–8 h [6,21] post-septic infection, including challenge with Bt [6]. To obtain patterns for these annotated immune genes based on our data, we first queried the Btload, time and rapid-U pattern sets for these genes. If the gene did not appear in any list, we resorted to the traditional cross-sectional method of identifying differentially expressed genes by querying the list of genes that were differentially expressed between 0 and 2 or 6 h.

3. Results

(a). Cross-sectional patterns of differential gene expression

A conventional cross-sectional time point analysis of differential gene expression in our data revealed that 1888 genes were differentially expressed between 0 and 2 h post-infection (1061 up, 827 down; electronic supplementary material, table S1), whereas 1429 genes were significantly differentially expressed between 0 and 6 h post-infection (813 up, 616 down; electronic supplementary material, table S2). A total of 672 genes (electronic supplementary material, table S3) were differentially regulated at both 2 and 6 h post-infection.

(b). Patterns of expression related to time and microbe load

Overall, there were 3100 genes (electronic supplementary material, tables S4 and S5) that fell into one of the investigated patterns of differential expression. There were 600 genes that showed a significant relationship with Bt density (‘Btload’; electronic supplementary material, table S4, and example in figure 2a,b) after accounting for the contributions of time and non-monotonic time. Six genes showed an unambiguous relationship with linear time (‘time’; figure 2c,d; electronic supplementary material, table S4), whereas 71 genes showed a strong relationship with slow induction and decay (‘polytime’; figure 2e,f; electronic supplementary material, table S4). In total, 1217 genes had multiple highly significant terms in the linear model (‘time-or-polytime’; electronic supplementary material, table S4), and thus, we were reluctant to assign them to one category. In addition, there were 737 genes in the ‘rapid-U’ set of induction and decay (figure 2g,h; electronic supplementary material, table S5) that showed a rapid and significant increase followed by decrease over the early time points, whereas there were 469 genes that were quickly down-regulated followed by a return to baseline (‘down-up rapid-U’; electronic supplementary material, table S5). For most rapid-U genes, the induction and decay pattern was independent of microbe load (figure 2h), except at very high microbe loads (10⁵ CFU; t = 12 primed individuals), where gene expression was re-induced.

GO enrichment of each of these pattern gene sets [20] revealed different overarching physiological roles (electronic supplementary material, table S6). The Btload set was enriched for proteolysis, extracellular activity, defence to bacterium, and hydrolase and serine peptidase activity. The set of genes with strong support for slow induction and decay (polytime) was enriched for cell–cell adhesion and membrane components, while the more ambiguous time-or-polytime set was enriched for metabolic and biosynthetic processes. The rapid-U set was enriched for regulation of cellular processes and signalling, intracellular signal transduction, enzyme inhibitor activity and various nucleotide-binding roles.

Distinctive trends therefore emerge when mapping these patterns to canonical immunological pathways (figure 3a). The majority of genes that code for extracellular microbe recognition and signalling, including recognition proteins like GNBP3 (TC003991) and PGRP-SA (TC010611), as well as a host of serine proteases and serpins, are significantly related to microbe load (figure 3a, red). By the same token, all antimicrobial peptides that are significantly differentially expressed are correlated to microbe load, most with correlation coefficients of 0.85 or greater (electronic supplementary material, table S4). All differentially regulated genes that code for intracellular signalling members of both Toll and IMD pathways, on the other hand, exclusively belong to the rapid-U set (figure 3a, purple). None of the genes associated with these major immunological pathways belonged to polytime or linear time patterns, with the exception of serine protease H2s TC000247 and TC000249. Only one gene in the pathways delineated in figure 3, Caspar (09985), was significantly differentially expressed at 2 or 6 h, but did not fall into one of the four patterns; it showed switch-like behaviour, wherein it was slightly but consistently up-regulated after inoculation.

Figure 3.

A revised representation of gene expression patterns for the canonical T. castaneum immunological pathways. Mapping our analyses onto these pathways (a), red represents genes that are correlated to microbe load, purple indicates a rapid-U pattern of expression over time, pink represents a statistically ambiguous placement that could either be Bt load or rapid-U, blue represents a time or slow induction and decay (polytime) pattern, grey represents genes that were up-regulated or down-regulated at 2 or 6 h post-infection without falling statistically into one of the other categories and white represents genes that were not differentially expressed in response to infection. Most genes were up-regulated, but in a few cases they are negatively regulated. The statistical support for categorizations can be found in electronic supplementary material, tables S4 and S5. We provide the cross-sectional 6 h post-infection results (b) from a previous study [6] for comparison, which compiles the results from two host strains. Text is coloured only in cases where not every gene in that category is differentially regulated (figure adapted from Behrens et al. [6], fig. 6) (Online version in colour.)

(c). Comparing our data to previous work and revising to account for new models

Previous studies have explored differential gene expression following wounding and septic injury in T. castaneum. However, they have not accounted for non-monotonic dynamics or the contribution of microbe load to gene expression profiles. We wished to first confirm that our data paralleled data from other studies when analysed in a similar fashion, and then explore how the patterns revealed by our analysis could alter the interpretation of gene expression patterns.

There was a considerable degree of overlap between our data and previously published data for T. castaneum septic injury gene expression profiles [6,21,22]. In particular, previous investigation into the T. castaneum response to septic Bt infection [6] found 1515 (SB host genotype) or 2036 (Cro genotype) differentially expressed genes at 6 h post-infection relative to uninfected individuals, slightly higher than our total of 1429. About one-quarter to one-third of genes differentially expressed at 6 h in our experiment were also differentially expressed (electronic supplementary material, figure S1) in Bt-septically inoculated SB (360/1429 shared genes; electronic supplementary material, table S7) and Cro (453/1429 shared genes; electronic supplementary material, table S8) genotypes at 6 h post-infection in that study. Within the groups of shared genes, about one-third of genes are specific to septic infection and do not appear in the wounding control data from that study. Only 2–3% of genes differentially expressed at 6 h in our septically infected hosts overlapped with previously published wounding controls but not their septically infected cohorts (electronic supplementary material, figure S1).

Within the canonical immune pathways (figure 3), almost all genes previously identified as differentially regulated at 6 h post-infection [6] (figure 3b) were also captured within our data as being differentially regulated at 6 h post-infection. However, our analysis revealed that most of these genes actually show patterns of microbe or rapid-U regulation (figure 3a). Our differential expression data contained a diminished proportion of Spatzle and Toll receptor genes, but identified more intracellular signalling genes as being differentially expressed. In particular, we found that both dorsal (TC007697) and dorsal2 (TC008096) transcription factors are subject to rapid-U expression dynamics, but dorsal expression decays more quickly over time than dorsal2. We also found that the transcription factor Jun (TC006814) and several of its regulatory genes—IAP (TC001189 and 1192) and Basket (TC006810)—were differentially expressed even though they were not identified as such in the most recent study [6], probably due to their rapid expression decay from 2 to 6 h post-infection. Most of the rapid-U intracellular signalling and regulatory genes, like imd (TC010851) and dorsal2, did not show differential expression in the [6] 6 h wounding control data, but were differentially expressed in septically infected individuals. This indicates that rapid-U dynamics, particularly the rate of post-peak decay, are not exclusively attributable to the response to wounding.

4. Discussion

Unless those factors are explicitly accounted for, the interpretation of gene expression data to suggest that higher differential expression equates to a more robust resistance response (e.g. [23–25]) implicitly assumes that genes are expressed monotonically with regard to time and with no correlation to microbe load. In our analysis of genes associated with canonical immunological pathways (figure 3), 36 of 40 differentially expressed genes violated these assumptions because they were either clearly non-monotonic with regard to time or correlated to microbe density. Moreover, there was an emergent spatial distribution of expression patterns, where extracellular recognition, signalling and effector components largely correlated to microbe load, while intracellular components exhibited early up-regulation followed by a rapid return to baseline. One implication of these dynamics is that studies that measure later time points to compare treatments may actually be comparing differences in the decay phase of gene expression for these genes.

The correlation between microbe loads and AMP expression presents an obstacle for establishing cause-and-effect relationships between AMP mRNA levels and resistance to parasites. For example, a recent study demonstrated that protein-starved bumblebees exhibited lower AMP gene expression in the presence of Crithidia bombii parasite infection [25], and interpreted this as evidence that poor nutrition dampens immune responses. However, a recent meta-analysis [26] suggests that poor host nutrition retards parasite growth rates in invertebrates, and indeed, bees deprived of protein did exhibit a lower parasite load in addition to lower AMP levels [25]. While it is possible that nutrition does affect AMP expression independent of microbe density, our results present an alternative hypothesis that these bees expressed lower levels of AMP transcripts because they had more modest parasite loads, in addition or even in exception to any feedback of nutrition on immune system levels. Understanding the impact of nutrition on investment in immunity is an important frontier in the field of eco-immunology [27], but our results suggest that studies on this topic need to first account for the impact of nutrient limitation on microbial growth. The relationship between microbe-correlated gene expression and microbe density is also likely to affect interpretation of studies investigating the relationship between climate change and host or vector immunity [28], as temperature is likely to affect microbe growth rates and host physiology.

Dependence of host gene expression upon microbe load is also likely to complicate interpretation of genotype- or treatment-specific differences in the production of an immune response, for example, if one genotype is more resistant than the other to a focal microbe (as in [6]). At the very least, genes that are correlated to microbe density could be misinterpreted, but there is also emerging evidence that a large set of ‘recovery’ genes do not get activated until insects begin to clear their microbes [13], which could further obscure true variation in immunological dynamics among host genotypes. We suggest that comparison of immune gene expression across genotypes must account for the dynamics of both microbe density and host gene expression. Looking forward to future studies, such analyses might open new opportunities for analysing differences between genotypes or other grouping variables like environmental treatments. For example, genotypes might differ in the slope of the expression-by-microbe density relationship, which would suggest that some genotypes are more immunologically sensitive to microbe burden than others. In addition, a study on Listeria-infected flies [13] suggests that the relationship between AMP expression and microbes might be sigmoidal over the full range of possible microbe loads, rather than linear; it would therefore also be worth investigating shifts in the microbe density that instigates half-maximal expression of immune genes. Finally, our study used live microbes, but it would be interesting to compare these patterns with a dose–response series of heat-killed microbes to investigate whether microbial virulence factors or other feedbacks influence these dynamics.

The variation in microbe and time correlations across the immune system raises some interesting questions for the processes through which investment in immunity and the control of immunopathology are optimized [29]. The inducible production of an immune response with respect to time or wounding is a sunk and invariant cost, inevitably paid whether the invasion involves a few microbes or many orders of magnitude. On the other hand, the production of immunological components in proportion to microbe load represents a marginal cost, paid only as needed. It is worth noting that the IMD pathway-mediated AMPs [30] attacin (figure 2a,b) and coleoptericin showed higher correlations to microbe load than Toll pathway-mediated AMPs did in our study (electronic supplementary material, table S4), suggesting some fine-tuning of signalling in response to the type of peptidoglycan that Bt expresses. However, the gene PGRP-SC2, which digests similar peptidoglycan molecules in D. melanogaster [31] and other insects [32] to prevent them from inducing immunopathology, is also among the set of genes most highly correlated to microbe load. These processes might represent an evolutionary compromise between balancing resistance and tolerance of infection [7,27] under pressure from a community of microbial and parasitic threats that all differ in growth rate, virulence and ability to subvert host defences. Future experiments could begin to quantify the contribution of each of these factors to quantitative variation in relevant parameters like the slope of the relationship between microbial density and AMP expression, and the rates of imd induction and decay (figure 2g), given the intercept and dynamics of a wounding control.

The relevance of different gene expression patterns for functional processes remains an avenue for future investigation. For example, it is not clear to what extent mRNA levels correlate with protein levels, as rates of mRNA degradation, protein translation and protein decay will all influence this relationship [33]. If both mRNA and proteins are dynamically stable, then one could think about mRNA levels as the first derivative of protein levels, where a U-shaped temporal gene expression profile would give rise to a sigmoidal protein profile. However, if both are dynamically unstable, then the proteins should also have a U-shaped pattern. Clearly, intracellular proteins do not decay too rapidly, or the mRNA levels of downstream effectors like AMPs [34] would reflect their U-shaped expression patterns instead of microbe-dependent patterns. Gross spatial considerations may also play a role in patterns of gene expression. For example, the re-induction of expression of otherwise decayed rapid-U genes at very high microbe loads may reflect a shift towards systemic rather than local distribution of microbes in our whole-organism study. This disparity invites future consideration of the relative effects of microbial growth dynamics and protein translation rates on subsequent immunological dynamics and the costs of mounting an immune response.

5. Conclusion

When we infer variation in host immunological competence and the capacity to resist infection from gene expression data at a single time point or without otherwise accounting for within-host dynamics, we are assuming that gene expression is monotonic with regard to time and independent of microbe load. There is emerging awareness that these assumptions may not hold [9–11], but there are also only a small handful of papers (e.g. [12,13]) that consider a correlation between gene expression and microbe load in insect systems. Our results strongly suggest that future studies measure microbe load when sampling for host gene expression, and perform pilot studies to rationally choose the most informative time point for the particular immunological function under investigation, if a full time series is not feasible. Moreover, the spatial, temporal and microbe dependence of gene expression patterns raises interesting questions regarding the control systems that give rise to differences in resistance and tolerance across host genotypes. These considerations are likely to be relevant and generalizable to a wide variety of host–microbe systems in nature.

Data accessibility

This article has no additional data.

Authors' contributions

A.T.T. designed the experiment, collected the data, designed and performed the analyses, and drafted the manuscript. A.L.G. helped to design the experiment and analyses, contributed to the drafting of the manuscript, and provided reagents and resources.

Competing interests

We declare we have no competing interests.

Funding

A.T.T. was supported by Agriculture and Food Research Initiative Competitive (grant no. 2014-67012-22278), from the USDA National Institute of Food and Agriculture.