Introduction

Genetic variation at the sequence level has been widely studied within and between species, to identify the forces that drive evolution (e.g., natural selection, genetic drift). The interest has been now turning to changes in gene regulation leading to genetic isolation, and ultimately speciation. Indeed, gene expression can have a strong influence on downstream phenotypes, and therefore its variation is likely to be a target of natural selection (Pavey et al. 2010). Large-scale technologies such as microarrays provide genome wide information that can be used to assess expression evolution (Gilad and Borevitz 2006). Several comparisons are possible: among populations of the same species, among species, as well as between interspecific hybrids and their parents. Natural populations have been previously studied for regulatory variations in several organisms from yeast (Townsend et al. 2003) to hominids (Storey et al. 2007). Two studies have reported differences in gene expression between African and European populations of Drosophila melanogaster with respect to sex biased genes (Meiklejohn et al. 2003) and genes involved in toxicity resistance or flight musculature which are potentially involved in adaptation (Hutter et al. 2008). No evolutionary studies so far have compared expression between closely related species and their hybrids using a set of distinct parental populations (contrasting with studies based on a single strain for each population or species). This is the approach we have adopted in the present work, studying closely related Drosophila species belonging to the melanogaster subgroup. It will allow us to assess at the expression level, not only the extent of population structure but also the importance of misregulation and additivity in hybrids.

We used here different populations of Drosophila simulans and a population of the sibling species Drosophila sechellia. Drosophila simulans and D. sechellia have long been thought to have separated about 400.000 years ago (Hey and Kliman 1993; Kliman et al. 2000; Lachaise and Silvain 2004), but more recent data pointed to a split occurring around 250.000 years ago (McDermott and Kliman 2008). On average, the DNA sequence divergence between these two sister species is around 1.5% (McDermott and Kliman 2008).

While these two species are phylogenetically very close, they are ecologically strongly different. D. simulans, which originates from eastern Africa or Madagascar (Lachaise et al. 1988; Lachaise and Silvain 2004; Dean and Ballard 2004; Kopp et al. 2006), is now a cosmopolitan generalist. D. sechellia is an endemic specialist in the Seychelles islands, and breeds exclusively on Morinda citrifolia, a plant highly toxic for other drosophilids (including D. simulans) (R’Kha et al. 1991). According to Oleksiak et al. (2002), gene expression differences between populations arise mainly from genetic drift (they did not show more differences within population than between). Our design already takes into account the variation within populations, thus differences shown there may equally represent drift or adaptation. Notably we expect a higher expression variation in D. simulans because of the higher diversity of environments occupied by the species compared to D. sechellia. The geographic structure of D. simulans has been studied using nuclear gene sequences or microsatellites (Hamblin and Veuille 1999; Schöfl and Schlötterer 2006; Baudry et al. 2006). These studies have revealed a differentiation between east African populations, other African populations (notably Zimbabwean), and European populations. In contrast, D. sechellia harbours little nucleotide sequence variation, which ranks this species as the least genetically diverse drosophilid (Legrand et al. 2009).

Previous studies which have examined divergence in gene expression between D. simulans and closely related species, have used samples isolated from a single isofemale line (i.e. consisting of the offspring of a single wild-caught female) (Michalak and Noor 2003; Haerty and Singh 2006; Moehring et al. 2007). Thus, they have not taken into account the intraspecific variation. An exception from this is the recent parallel studies of sequence polymorphism and expression of six D. simulans lines originating from five different locations within the species range (Holloway et al. 2007; Lawniczak et al. 2008). The authors revealed interesting insights into the evolution of regulatory sequences between populations. However, each of their population was represented by only one (or two in the case of Madagascar) line, thus they could not take into account intra-population variation.

The present study includes four populations of D. simulans each represented by four isofemale lines, four isofemale lines of D. sechellia and four hybrid “populations” (crosses between D. sechellia males and D. simulans females). This design allowed us to take into account two different types of variances: intraspecific variance (between the different populations of D. simulans), and intra-population variance (between the different lines which represent biological replicates within each population). We considered the four isofemale lines of D. sechellia as a single population as this species does not show geographic structure (Legrand et al. 2009). We used males because they are the most affected sex in Drosophila hybrids (since they are heterogametic they show more hybrid breakdown), therefore suitable to highlight differences linked with species divergence. Regulation breakdown in hybrids can occur due to incompatibilities between alleles at a given locus It can also result from negative epistasis between loci according to Dobzhansky-Muller’s model of hybrid incompatibility (Dobzhansky 1936; Muller 1942). Differences in hybrid expression can also simply result from regulation divergence between genomes. Hybrids have been previously shown to harbour strong regulation breakdown compared to parents in different organisms (Gibson et al. 2004; Haerty and Singh 2006; Moehring et al. 2007). However, other studies have shown large scale additivity in the patterns of expression of the hybrids (Hughes et al. 2006; Rottscheidt and Harr 2007). These contrasting results may be caused by methodological differences (organ-specific vs. whole body/differences in microarray platform) or by different degree of divergences and inbreeding between species (Rottscheidt and Harr 2007). These two non-mutually exclusive hypotheses will be considered here. We extracted RNA from whole body males to consider only general differences, and not tissue specific differences.

Our populations of D. simulans were from Zimbabwe, Kenya, the Seychelles (the last two likely represent the ancestral range) (Lachaise et al. 1988; Lachaise and Silvain 2004; Dean and Ballard 2004; Kopp et al. 2006), and from France (a derived population). This approach allowed us to determine to what extent gene expression shows geographic structure in D. simulans, as well as differentiate between population effects (possibly linked with recent invasion) and species divergence. We observed a population structure in D. simulans, and consistent expression divergence between D. simulans and D. sechellia. We notably observed a strong involvement of the Cytochrome P450 gene family, as well as genes regulating the juvenile hormone (JH).

Materials and methods

Drosophila stocks

We studied four populations of D. simulans. Each population was represented by four isofemale lines (biological replicates). The populations came from France (the Rhone Valley, collected in 2003), Kenya (Nairobi, collected in 2001), Zimbabwe (Mazoe, collected in 1997) and the Seychelles Islands (Mahé and Praslin, collected in 2003). The four D. sechellia lines also originated from Mahé and Praslin in the Seychelles archipelago (collected in 2003).

RNA samples

The lines were mass reared in uncrowded culture, on axenic standard medium at 25°C, with a natural light cycle. For each isofemale line, ≥6 replicate cultures were raised in vials containing each 8 males and 8 virgin females (10 of each for D. sechellia). For a given D. simulans population, four different crosses were performed, each between females of one of the line and males from a different line of D. sechellia (at least 3 replicates per cross). We thus obtained four “populations” of hybrids corresponding to the four D. simulans populations (Table 1). The experimental design thus included a total of nine “populations”, each consisting of four biological replicates. Virgin male offspring from at least 3 replicate vials were collected within a few hours of emergence to create pools of 25 individuals that were transferred into fresh vials. Seven days later, each pool was frozen at −80°C. We used the Nucleospin RNA II kit from Macherey–Nagel to extract the RNA from the pools of 25 whole-body adult males, yielding to a total of about 3 μg RNA to hybridize on the arrays. RNA was then reverse transcribed in presence of alpha d’CTP p33.

Table 1 Experimental design to obtain hybrid males
Full size table

Arrays

The arrays were nylon filters spotted with long amplicons from the species D. melanogaster. They were hybridised in the TAGC platform in Marseille. There were 7,041 spots: 5,931 different whole cDNA of D. melanogaster and 1,100 control spots, either negative, or positive controls (a cDNA of Arabidopsis thaliana: chlorophylle synthetase). cDNAs were cloned into a vector, amplified and spotted on the array. Each spotted fragment contained both the cDNA and a specific part of the vector for spotting normalization (first hybridisation). These arrays were hybridised twice. Firstly, hybridisation was performed with a P33 labeled oligonucleotide probe specific to the vector sequence spotted. As every molecule spotted contained this sequence, the radioactive signal (vector hybridisation signal) read was proportional to the quantity of spotted cDNA. Secondly, after deshybridisation of the vector probe, we proceeded with the hybridisation of the cDNA samples. A second radioactive signal was read. It will be further designated as complex hybridisation signal.

Data normalization

The normalization procedure was defined by the manufacturer of the arrays. Every spot for which the vector signal was smaller than 5 times the negative spots’ median was eliminated from the analysis. For both signals, background (measured by the median of negative spots) was subtracted. Inter-spots/intra-arrays normalization was then performed by dividing the complex hybridisation signal by the vector hybridisation signal. The last normalization step was to divide the obtained expression value in each array by the corresponding median of all spots (or by the median of positive controls), effectively normalizing the signal between arrays. The two approaches (median of controls/median of all spots) led to the same results. Normalization quality was assessed visually by MA plot and box-plot of normalized expression values for each array (Supplementary Fig. 1). All genes with four or more missing data throughout the 36 arrays were discarded.

Statistical analysis

The statistical analysis was based on the whole set of 36 arrays. To determine differentially expressed genes, an Analysis of Variance (ANOVA) model was fitted for each gene (Kerr et al. 2000, 2002). The fitted model was the following:

$$ {\text{Y}}_{\text{ij}} = {{\upmu}}_{\text{i}} + {\text{E}}_{\text{ij}} , $$

where i is the population index (i = 1,…,9: 4 D. simulans, 1 D. sechellia and 4 hybrid “populations”), j is the biological replicate index (j = 1,…,4), Yij is the normalized signal (log-transformed), μi is the mean expression for the gene in population i and Eij is the residual variability. This model assumes a common variance for all populations, that is consistently estimated with 36 − 9 = 27 degrees of freedom for most genes (a few genes have missing values due to normalization). The homogeneity of variance between groups was verified using Levene’s test (Levene 1960). We showed a variance homogeneity for all genes but 20 (FDR = 0.1, see Supplementary Fig. 2 for the distribution of P-value). We tested the equality of mean expression between D. sechellia and D. simulans populations as well as between D. simulans populations. As for the comparisons between a population of hybrids and its two parental populations, we performed tests to compare the mean expression of the hybrid to the mean expression of each parental population, and we also compared the mean expression of the hybrid to the average of the mean expressions of the two parental populations. All these comparisons were performed using usual contrast t tests within the ANOVA model. For each comparison, raw P-values were adjusted by the Benjamini-Hochberg method, which controls the false discovery rate (FDR) (Benjamini and Hochberg 1995). We used a FDR of 0.1.

A mixed-effects model does not suit our analysis, since estimates of some parental effects would only have been from two values. However, we assessed the effect of our analysis assuming correlated data by simulating data with parental effect. We did not find any increased number of false positive even with a high biological/technical variability ratio, the only consequence was a decreased power (Supplementary Fig. 3).

Patterns of inheritance: additivity, dominance, overdominance

To assess patterns of inheritance, we analyzed the distribution of dominance effects using the ratio d/a, where a is half the difference in expression between the parental populations (D. sechellia and respectively each population of D. simulans), and d is the expression difference between F1 hybrid and the parental average. If d/a = 0, it means perfect additivity (d = 0), if |d/a|=1, complete dominance and if |d/a| >1, overdominance (Falconer and Mackay 1996). We performed this analysis for our four parental populations, only to those genes which were differentially expressed between the parents, to avoid any bias due to equality of expression between the parents.

Variance comparison

To compare the genomic variability in two populations A and B, we propose the following test. This test looks for an excess of genes with higher (or lower) variance in one population relative to another. For a given gene g, we note \( {\sigma_{\text{g,A}}^{2} } \) and \( {\sigma_{\text{g,B}}^{2} } \) the gene expression variances in populations A and B. If population A harbours more genetic variability than population B, then, for most of the genes, the ratio \( {R_{g} = {\frac{{\sigma_{\text{g,A}}^{2} }}{{\sigma_{\text{g,B}}^{2} }}}} \) will be higher than 1, whereas Rg should be higher than 1 for roughly 50% of the genes if the two populations are comparable. Therefore a test can be based on the number \( {{\text{N}}_{1}^{\text{AB}} } \) of genes for which the empirical ratio Rg of gene g is higher than 1. We note pAB the true proportion of genes for which \( {\sigma_{\text{g,A}}^{2} > \sigma_{\text{g,B}}^{2} } \). We test \( {{\text{H}}_{ 0} = \left\{ {{\text{p}}_{\text{AB}} = 1/2} \right\}} \) (A and B are comparable) versus \( {{\text{H}}_{1} = \left\{ {{\text{p}}_{\text{AB}} > 1/2} \right\}} \) (variability is higher in A). \( {N_{1}^{AB} } \) has a binomial distribution \( {{\text{B}}({\text{G}},{\text{p}}_{\text{AB}} )} \) with G the number of genes. The P-value of the test is thus:

$$ {{\text{P}}({\text{N}}_{1}^{\text{AB}} > {\text{n}}_{{1,{\text{obs}}}}^{\text{AB}} |{\text{p}}_{\text{AB}} = 1/2)}. $$

This analysis takes into account the number of lines available to estimate each variance. We simulated data with different number of lines (n = 2, 5, 10 and 50) with equal population variance, to measure the impact of a small number of lines on variance estimates. This did not affect the P-value distribution and thus the error. Significant tests were around 5% (Supplementary Fig. 4). Further simulation showed this only affects the power of the test.

Gene ontology

Lists of differentially expressed genes were examined for statistical over/under representation of Gene Ontology (Ashburner et al. 2000) terms using FuncAssociate (Berriz et al. 2003) with a reference background consisting of all genes in the arrays. Our array itself was compared to the whole genome, revealing several ontology biases in the construction of the array. This made essential the use of our array as background when examining differentially expressed genes for gene ontology bias. To further explore the terms and the corresponding genes, we used the Gene Ontology database provided by the Gene Ontology consortium (in May 2008).

Results

The experimental design allowed us to perform multiple comparisons, within D. simulans, and between D. simulans and D. sechellia, as well as their hybrids. By maximizing the biological source of variation (using biological and not merely technical replicates, Altman 2005) in populations (and species when applied), we revealed strongly significant variations. After all gene filtering during the normalization process, we assessed expression for 4,398 genes, which is about a fourth of the Drosophila genome. The differences observed are thus non exhaustive, but their consistency between all populations, through our large sampling, is supported by the power of the cross-design.

Comparison between populations of D. simulans

We did not detect any significant difference in gene expression between the three African populations. Contrasting with this result, all the comparisons between the French population and each of the three African populations showed differential expression. Respectively 6, 7 and 13 genes were found to be differentially expressed between the French population and the Kenyan, the Seychelles and the Zimbabwean populations (Supplementary Table 1). Six genes were differentially expressed in two pairwise comparisons. No gene was differentially expressed in all three pairwise comparisons. Out of the twelve genes which were over-expressed in the French population compared to at least one of the African population, four are cytochrome P450 genes (significantly over-represented, Fisher’s exact test, P < 0.05).

We compared variability using a binomial test based on the fact that variance ratios of genes are expected to be half of the time above 1 under the assumption of similar variances, using only pairwise comparisons. P-values of the binomial test are provided in Table 2. The variance from the French population is significantly higher than the variance from any other population but the Zimbabwean (P < 0.005, Bonferroni corrected threshold). In terms of variance, we observe a differentiation of the French and the Zimbabwean populations compared to other African populations (from the zone of origin of D. simulans). All other pairwise comparisons revealed significant differences, even though there is a wide range of P-values. It is important to note that this analysis is independent from the test of variance homogeneity gene by gene performed with Levene’s test. It is possible to have homogeneity gene by gene, whereas on a global scale, heterogeneity can be observed.

Table 2 Above the diagonal: P-values of binomial tests under the assumption of equality of variance between the populations
Full size table

Drosophila simulans vs. Drosophila sechellia

The comparisons between D. sechellia and D. simulans yielded to 347, 337, 353 and 518 genes differentially expressed with the populations of Zimbabwe, Kenya, Seychelles and France, respectively. Details of over/under-expressed genes are shown in Table 3. The striking result is that 304 genes are consistently differentially expressed between all four populations of D. simulans and D. sechellia (Supplementary Table 2). We can therefore assume that these genes present constitutive expression differences between the two species.

Table 3 Number of genes over-/under-expressed in D. simulans compared with D. sechellia
Full size table

Five terms were consistently over-represented in the subset of genes under-expressed in D. sechellia compared to every population of D. simulans. The molecular function “electron carrier activity” and the cellular components “vesicular fraction” and “microsome” refer to cytochrome P450 genes, as was assessed by examining the intersection of the genes with this annotation in D. melanogaster, and our set of differentially expressed genes.

The two other terms (namely “lipid metabolism” and “hormone catabolism”) refer to three genes: Juvenile hormone epoxide hydrolase 1 (Jheh1), Juvenile hormone epoxide hydrolase 3 (Jheh3) and Dopamine N-acetyltransferase (Dat). These three genes are highly pleiotropic as they are directly involved in the regulation of key hormones: Juvenile hormone and Dopamine (DA). Jheh1 and Jheh3 are involved in JH regulation by degrading it. According to Gruntenko and Rauschenbach (2008), the JH titre can be assessed by the JH degradation level. Thus, we can assume that an over-expression of Jheh1 and Jheh3 in D. simulans compared with D. sechellia implies a lower JH titre in D. simulans. The gene Dat is involved in the degradation of DA.

Hybrids vs. parental populations

Expression in male hybrids was compared with males of both parental species, and the mean expression of the parents. By this process, the differentially expressed genes were those different from the parents and therefore not showing the dominance of a parental allele on the other. This also excluded genetic additive effects since their expression had to be significantly different from the parents’ mean. A significant difference could be due to underdominance, i.e. failed interaction between the two alleles, or misregulation through negative epistasis. We found few genes perturbed in hybrids. No gene disruption was detected in hybrids obtained with the populations of D. simulans from Kenya and from Seychelles. Significant over-expression was detected for four and five genes in hybrids offspring of the French and Zimbabwean populations, respectively (Table 4). X-linked genes were significantly over-represented in this set of genes (Fisher exact test, P < 0.001). One gene was common in both comparisons (Cp110). This gene is involved in centriole replication, although its precise function is unknown (Dobbelaere et al. 2008). Another gene (r-cup) is involved in male meiosis, and its disturbance could have a role in hybrid sterility (Barreau et al. 2008).

Table 4 Genes over-expressed in hybrids compared to parents
Full size table

We also examined patterns of dominance and additivity. For this analysis, we included only genes for which expression is different in the two parental populations. We used the ratio dominance over additivity (d/a). About 45% of genes showed additivity to partial dominance (0 < d/a < 0.5), about 26% showed partial to complete dominance (0.5 < d/a < 1), and about 29% showed overdominance (Fig. 1).

Fig. 1
figure 1

Distribution of the ratio d/a for hybrids crossed from our four different populations of D. simulans. When d/a = 0, d = 0, we are thus showing perfect additivity. From 0 to 1 (or −1), we shift from complete additivity to complete dominance. Over 1 (under −1), we show overdominance

Full size image

Discussion

Our multiple comparisons of transcriptomes revealed three main features: 1-geographic differentiation in D. simulans; 2- expression divergence between D. simulans and D. sechellia (about 7% of the genes), notably cytochrome P450 genes, and genes involved in hormone metabolism (JH and DA); 3- only eight genes misregulated in hybrids among which X-linked genes were over-represented.

Assessment of the use of interspecific array

We used arrays carrying cDNA from D. melanogaster. The use of interspecific arrays has been shown to introduce a bias in comparisons, due to differential hybridisation caused by sequence divergence (Gilad et al. 2005; Oshlack et al. 2007). However, D. melanogaster, one of the closest sister species of D. simulans and D. sechellia, is approximately as divergent from D. simulans and D. sechellia. The divergence between D. sechellia and D. simulans is thought to have occurred around 250.000 years ago, while the divergence between the two species and D. melanogaster dates back 2 to 3 million years (Lachaise et al. 1988; Hey and Kliman 1993; McDermott and Kliman 2008). We can thus assume the hybridisation signal to be similarly affected by the use of a D. melanogaster array. Moreover, in order to limit the possible bias, we chose a long cDNA array rather than oligonucleotide, which limited the effect of possible mismatches in the sequence. To assess whether there was a hybridisation bias, we looked for a correlation between: 1- the sequence divergence difference between each species and D. melanogaster and 2- the mean differences of expression between the two species. We assessed this in a large subset of genes included in the array (2,303 randomly chosen genes), using the sequences from the database (http://www.flybase.org). We observed no correlation (Spearman correlation coefficient: Rho = −0.0054; P = 0.8056). D. simulans and D. sechellia diverge by less than 2%, which is comparable with the level of divergence of disparate populations of D. melanogaster. 80% of divergent bases with D. melanogaster are shared between D. simulans and D. sechellia (Dworkin and Jones 2009). In this study, they found little evidence of any bias between D. sechellia and D. simulans. Although on a global scale, the influence of mismatch is likely negligible, we agree that individual genes can be affected. However, Mezey et al. (2008) showed that the effect of sequence divergence is mainly an increased variance leading to a decreased power of the test. Our tests are thus likely conservative. To assess a possible bias, we used sequences from Flybase to calculate the difference in coding sequence divergence (D. melanogaster vs. D. simulans minus D. melanogaster vs. D. sechellia). This was calculated first for differentially expressed genes between D. sechellia and D. simulans, and second for the random set of genes previously mentioned. Differentially expressed genes are represented by a subset of 232 genes taken from our list of 304 genes consistently differentially expressed between D. sechellia and D. simulans. We could not consider all genes because of incomplete/erroneous ortholog annotation in Flybase. However, in both sets (random and differentially expressed), genes were sorted the same way. Although the P-value is relatively low, the mean difference in divergence between random genes and differentially expressed genes is not significant (Mann–Whitney, W = 286,527, P-value = 0.06643). We assessed further the extent of the possible bias using a Q-Q plot of the distributions (Supplementary Fig. 5). This revealed that about 30 genes out of the 232 are biased toward a higher divergence compared to the random genes distribution. This bias can have two explanations. The first one is that sequence divergence affected hybridisation, leading to false positive. The second explanation is biological: genes with a higher coding sequence divergence can also be genes evolving faster in terms of expression changes. Although there might be some bias, cytochrome genes as well as Jheh1, Jheh3, and Dat show a divergence difference below 1.3%, situated in the unbiased part of the Q-Q plot. This bias may have two consequences: a decreased power for the test, and a slight increase in the false positive rate.

Differentiation between D. simulans populations

Expression differentiation

D. simulans originates from Eastern Africa or Madagascar (Lachaise et al. 1988; Lachaise and Silvain 2004; Dean and Ballard 2004; Kopp et al. 2006). This ancestral area is consistent with several observations of our study. We would like to point out the difference between the African populations of D. simulans, and the French population which are revealed with all comparisons. Each of the African population revealed around 350 expression differences with D. sechellia, while the French one revealed 518, thus showing a significantly stronger differentiation compared with D. sechellia. The direct comparison revealed no differential expression between African populations of D. simulans, while we detected several differentially expressed genes between the French population and each African population. This is consistent with the weak but existing population structure observed on microsatellites by Schöfl and Schlötterer (2006) and on nuclear genes by Baudry et al. (2006). Gene flow should be higher between the three African populations than with the French population, as expected from an isolation by distance model and the evolutionary history of the species. A low level of existing geographic differentiation was also described for morphological traits (Gibert et al. 2004). Among African populations, the study of Schöfl and Schlötterer (2006) shows a differentiation between sub-Saharan populations and non sub-Saharan populations. Within the sub-Saharan population a differentiation was observed between the lines from Zimbabwe and Malawi on the one hand, and the lines from Uganda on the other hand, the latter being geographically closer to the likely region of origin of D. simulans. Our study suggests a similar differentiation between the Zimbabwean populations and the population from the ancestral zone. Indeed, when we compare F1 interspecific hybrids with their parents, the Zimbabwean population leads to 5 differentially expressed genes (Table 4). This result is comparable with what is observed for the French population (4 genes differentially expressed), but contrasts with the lack of differences observed for the two populations close to the likely origin of D. simulans: the Kenyan and the Seychelles populations. The Zimbabwean population shows a very specific pattern. It has been shown to be clearly apart from the eastern populations at the genetic level, but it still has quite high polymorphism (Baudry et al. 2006). Although results in terms of expression differences are contrasted, this population will be further considered as derived.

The study of Baudry et al. (2006) revealed significant differentiation between Eastern African populations and both the French and the Zimbabwean population using four X-linked loci. The authors did not detect any differentiation between populations from Madagascar, Mayotte, Kenya and Tanzania, which is consistent with our observation of similar expression profiles for D. simulans of the Seychelles Islands and the Kenya. However we did not detect gene expression differences between the population from Zimbabwe and the two other African populations with direct comparisons. It is likely the differentiation observed on DNA sequences (consistent with hybrid and variance analysis of the present study) is not strong enough to appear on a direct expression comparison, or does not affect expression. This result is consistent with what is known of D. simulans biogeographic history: a strong intra-population variation and a relatively weak differentiation between populations (Lachaise et al. 1988; Lachaise and Silvain 2004).

Population structure in expression has been shown in other organisms. In humans, Storey et al. (2007) showed population structure in expression between European and Nigerian cell cultures, although these results are still controversial due to possible differences in the cell immortalisation process (Davis and Kohane 2009). Adaptation to local environment at the level of gene expression has also been shown in other organisms such as Saccharomyces cerevisae (Townsend et al. 2003) or on a teleost fish which showed adaptation to temperature (Oleksiak et al. 2002; Whitehead and Crawford 2006).

Population variance, an evidence for decanalization?

We observed a higher interspecific expression variance in derived populations compared with populations from the ancestral area (Eastern Africa/Madagascar). It is interesting to note that DNA sequence variation shows the opposite trend (Schöfl and Schlötterer 2004; Baudry et al. 2006). This resembles the pattern produced by the process of decanalization, i.e. the revelation in a new environment of existing cryptic variation (Gibson and Wagner 2000; Gibson and Dworkin 2004). This hypothesis predicts phenotypic constraints in the ancestral area due to stabilizing selection. At the genetic level, variation can accumulate and not be expressed in the phenotype, because of various buffering mechanisms such as epigenetic interactions, environmental constraints, etc. When the genotype is transferred in a new environment (for example when there is invasion of a new area), the environment changes, and so does the selective pressure. This can in turn result in the expression of the cryptic genetic variation, which increases the phenotypic variance (in our case, the variance in expression level). This was notably observed for mesosternal bristle number in a drosophilid (Yassin et al. 2007). This process happens despite a reduction in sequence polymorphism due to bottleneck and founding effects (Lachaise et al. 1988; Lachaise and Silvain 2004). However, decanalization has been only described with specific traits, and never at a general scale. Whether this process can be seen by assessing expression variance in all genes is still unknown. Moreover, studying different traits in several populations of D. simulans, Capy et al. (1993) revealed morphometric clines, but no difference in variance of traits between derived populations and those from the ancestral range. Adopting a similar approach, we checked expression data from D. melanogaster for a similar pattern (data from Hutter et al. 2008). We could not detect any significant variance difference between European and African populations. As D. melanogaster’s invasion of the world is older than D. simulans’s, it is possible stabilizing selection has acted on regulation, effectively dropping the expression variance in derived populations. The variance difference in D. simulans seems biologically significant since it matches the invasion pattern of the species. This tends to eliminate the hypothesis of a technical bias. However, we could not confirm our observation, either on other expression data, or on phenotypic data. This result therefore deserves further investigation.

Gene expression divergence between species

Our study shows 304 genes consistently differentially expressed between the two species. We compared the genes differentially expressed in our study with those revealed by Dworkin and Jones (2009). Thirty-four percent of their differentially expressed genes were present in our array. Out of these, 16% (60 genes) were also in our list of 304 genes differentially expressed between D. simulans and D. sechellia. The discrepancies can be explained by differences in the design (we compared males while they compared mixed sexes), lines choice (ancient laboratory lines versus recent wild lines), and power of the arrays.

Cytochrome P450 gene family: detoxification or hormone regulation

Cytochrome P450 is a large family of 83 functional genes in D. melanogaster (Tijet et al. 2001). The down-regulation of cytochrome P450 genes in D. sechellia compared with D. simulans (also observed by Dworkin and Jones 2009) can probably be explained by the role of this gene family in detoxification. It could be related to the specialization of the species on the toxic plant Morinda citrifolia (Dworkin and Jones 2009). The strict association between D. sechellia and its host could have reduced the variety of toxins D. sechellia is exposed to, releasing selective constraints on detoxification genes such as the cytochrome P450 gene family. It is also possible that some differences in the need of detoxification genes arose from different environmental conditions (not related to the specialization on M. citrifolia). However, we observe cytochrome P450 regulation divergences in comparisons of D. sechellia with all four populations of D. simulans, despite the fact that they come from four different geographic areas. If the latter hypothesis was supported, we would likely observe differences between all D. simulans populations.

Interestingly, the cytochrome P450 genes are also significantly over-represented in the twelve genes that are over-expressed in the French population compared with at least one African population (Supplementary Table 1). Four genes are cytochrome P450s, and one (Walrus) has a similar molecular function. It is possible the French population encounters a wider set of toxins due to anthropization processes (Dworkin and Jones 2009) leading to stronger constraints on the cytochrome P450 gene family. However, without data concerning pollutants at the sites of sampling, it is difficult to verify this hypothesis. Alternatively, the divergence of expression of these genes could be related with the involvement of these genes in hormone metabolism (Feyereisen 1999; Tijet et al. 2001), and especially JH regulation.

Divergence of hormonal regulation

We have observed divergence of expression for three genes involved in hormone (notably juvenile hormone and Dopamine) regulation: jheh1, jheh3, and dat. Although the role of JH has been widely described in females of D. melanogaster (Gruntenko and Rauschenbach 2008; Liu et al. 2008), it is still poorly known in adult males. However, a physiologic approach has shown a role of JH in seminal fluid protein accumulation in the male reproductive accessory glands (Wolfner et al. 1997), a role also supported by the mutant-based study of Wilson et al. (2003). Mutants with weak receptivity to JH show lower protein accumulation in these glands, and this can in turn affect male fertility. Mutant males also show very little interest in courtship, suggesting a role (either direct or indirect via the perturbation of accessory glands’ protein synthesis) of JH in courtship behavior. Dopamine (DA) also plays a role in courtship behavior, a role that could be consistent with the misregulation of Dat. The changes in the regulatory pathway of JH between the two species suggests a change of reproductive behavior in males, which could possibly correspond to a change in females, i.e. coevolution of both sexes via sexual selection. While highly speculative, this hypothesis could be a source (as much as a consequence) of the reproductive/behavioural isolation between the two species. DA has been shown to be involved in JH regulation in females, but has apparently little effect on JH in males as has been shown in D. virilis (Gruntenko and Rauschenbach 2008). It is however likely that a change in DA level will affect reproduction in males, but this won’t be via JH.

Sterile hybrids, yet weak gene expression differences

We observed 8 genes over-expressed in interspecific hybrids compared to their parental populations, and none under-expressed. Six of these genes are located on the X chromosome. Genes were differentially expressed only for offspring of the two most differentiated hybrid populations: Zimbabwe and France. One gene was common in the two comparisons (Cp110).

Impact of the X-chromosome

An interesting result of our analysis is that out of the 8 genes mis-regulated in hybrids, six are located on the X chromosome, a number significantly higher than expected under the assumption of random localization of the differentially expressed genes. This observation is consistent with the so-called “faster-X” effect, which is a commonly mentioned but still controversial characteristic of speciation (Betancourt et al. 2002; Thornton and Long 2002; Musters et al. 2006; Begun et al. 2007; Masly and Presgraves 2007; Presgraves 2008; Vicoso and Charlesworth 2009). According to this theory, X-linked genes evolve more rapidly than genes on autosomes, perhaps due to higher efficiency of selection on the hemizygous X in males. X-linked genes have a specific evolution, due to their presence two-thirds of the time in females and one-third in males and due to their smaller population size than autosomes. However, using introgression, Hollocher and Wu (1996) found no higher density of sterility factor in the X chromosome than on autosomes. This suggests that the X-linked disturbance causing sterility is linked to divergence in regulation and not directly to sequence divergence of X-linked genes.

A low hybrid/parent regulation differentiation

The weak hybrid/parents differentiation observed in our study can be somewhat surprising, compared to results obtained in other studies. For example, using gene expression data on testes, Haerty and Singh (2006) found 241 genes differentially expressed between hybrid and parents. Three differences in the experimental design can explain this discrepancy. First, Haerty and Singh (2006) did not differentiate additive effects. Thus, it is possible that some of their differentially expressed genes represent in fact an averaged expression between the two parental genomes, a possibility we have excluded here. Second (and main) point, their study was on testes. They focused on an organ that is strongly affected as hybrid males between these two species are sterile (Cabot et al. 1994; Joly et al. 1997). Therefore they must have revealed genes involved in this sterility. Our goal was more to examine a global divergence. Therefore, we adopted a whole-body approach, which limited the detection of organ specific divergences, but highlighted more ubiquitous and global changes. Testes have a very specific expression pattern, likely perturbed in hybrids, but these particular differences would be hidden by our approach (Wolgemuth and Watrin 1991; Grimes 2004). Finally, the D. simulans line used by Haerty and Singh (2006) was a laboratory strain originating from Arizona, a population geographically far from the African native area of the species where we collected our samples.

The study of Michalak and Noor (2003) revealed 51 genes differentially expressed between hybrids and parents using D. mauritiana and D. simulans. None of these genes were found differentially expressed in our study. However, it has been previously shown that factors involved in hybrid sterility are different between simulans/sechellia vs. simulans/mauritiana hybrids (Coyne et al. 1991; Cabot et al. 1994). Another study detected 220 differentially expressed genes (Moehring et al. 2007). No gene is commonly differentially expressed between their study and ours. In this study, there is no correction for multiple testing, potentially allowing for a large number of false positives. If we adopted the same approach, our number of differentially expressed genes would jump to an average of about 67 (Supplementary Table 3), which is similar (Chi² = 1.97, df = 1, P = 0.16) to what is observed by Moehring et al. (2007). Interestingly, all these studies, as well as others, in Drosophila (Michalak and Noor 2003, 2004; Ranz et al. 2004; Landry et al. 2005; Haerty and Singh 2006) or various other organisms (Wang et al. 2006; Malone et al. 2007; Renaut et al. 2009; Mavarez et al. 2009) detected a large number of genes under-expressed in hybrids compared to parents, and only a few over-expressed. Contrasting with this observation, the present study only showed over-expressed genes in hybrids with FDR correction applied. Without the FDR correction, the representation of over/under-expressed genes is not consistent between the different hybrid populations. In fact, we have strong contrasts between the four different comparisons of hybrids with parents (Supplementary Table 3). This could be related with differences in allele profile between the parental D. simulans populations, therefore allowing for different misregulation patterns in hybrids. Furthermore, the populations we used may lead to hybrids with different properties, owing to the fact that we chose African populations, while other studies have chosen more recently derived American isofemale lines. This aspect remains to be more thoroughly explored.

Intermediate additivity of expression

We observed about 45% of genes showing additivity in expression in the hybrid. This suggests an intermediate pattern compared to what was observed in other studies, from a few percent (Gibson et al. 2004; Haerty and Singh 2006; Moehring et al. 2007) up to 71% (Hughes et al. 2006; Rottscheidt and Harr 2007). The possible reasons for these discrepancies are numerous, as detailed by Rottscheidt and Harr (2007): amount of inbreeding of the parental lines, methodological differences and phylogenic distance between the parents. The last argument probably explains why, using the same method, we observed more dominance than Hughes et al. (2006). Their parental lines are isofemale lines of the same species, while we used populations of two closely related species; it is therefore expected that we would observe more expression disturbances in the form of overdominance (Hughes et al. 2006). It is worth noting that this overdominance is probably still quite small for each individual gene, since we observed very few genes showing expression outside the range of their parental populations.

Cis-regulation, a major player?

Our observations of very few cases of misexpression in hybrids (=very little perturbation of trans-regulation) and a large set of expression changes between the two species suggest a major role of cis-regulation on the divergence between the two species. This is consistent with an evolutionary model of stronger constraints on trans-regulation: the pleiotropic role of transcription factors makes them likely to be more constrained than cis-factors, which usually affect only one locus, or even one allele of a given locus. This observation is consistent with other studies on Drosophila species (Wittkopp et al. 2004, 2008a, b).