Novel approach to quantitative spatial gene expression uncovers genetic stochasticity in the developing Drosophila eye

Abstract

Robustness in development allows for the accumulation of genetically based variation in expression. However, this variation is usually examined in response to large perturbations, and examination of this variation has been limited to being spatial, or quantitative, but because of technical restrictions not both. Here we bridge these gaps by investigating replicated quantitative spatial gene expression using rigorous statistical models, in different genotypes, sexes, and species (Drosophila melanogaster and D. simulans). Using this type of quantitative approach with molecular developmental data allows for comparison among conditions, such as different genetic backgrounds. We apply this approach to the morphogenetic furrow, a wave of differentiation that patterns the developing eye disc. Within the morphogenetic furrow, we focus on four genes, hairy, atonal, hedgehog, and Delta. Hybridization chain reaction quantitatively measures spatial gene expression, co-staining for all four genes simultaneously. We find considerable variation in the spatial expression pattern of these genes in the eye between species, genotypes, and sexes. We also find that there has been evolution of the regulatory relationship between these genes, and that their spatial interrelationships have evolved between species. This variation has no phenotypic effect, and could be buffered by network thresholds or compensation from other genes. Both of these mechanisms could potentially be contributing to long term developmental systems drift.

1 ∣. INTRODUCTION

Variation in gene expression is the rule rather than the exception (Albert & Kruglyak, 2015; Bothma et al., 2014; Marinov et al., 2014; Pai et al., 2015; Signor & Nuzhdin, 2018). Variation in gene expression can come from perturbations in the environment, for example due to changes in temperature or maternal nutrition (Arsenault et al., 2018; Chen et al., 2015; Signor & Nuzhdin, 2018), and from differences in transcription between cells due to the burstiness of transcription (Bothma et al., 2014; Felix & Barkoulas, 2015). Organisms also experience perturbation in the form of recombination and new mutations. However, despite genetic and environmental perturbations to gene expression, organisms are able to produce robust phenotypic outcomes, suggesting that development is not readily affected by differences in gene expression (Garfield et al., 2013; MacNeil & Wallhout, 2011; Raj & van Oudenaarden, 2008; Romero et al., 2012). In contrast, on evolutionary timescales organisms must produce adaptive heritable variation in response to selection pressures, tapping those same sources of variation to which they might otherwise be robust. Indeed, the genetic basis of adaptation is often inferred to be due to differences in gene expression (Chan et al., 2010; Gompel, Prud'homme, Wittkopp, Kassner, & Carroll, 2005; Hoekstra & Coyne, 2007; Jeong et al., 2008; Pai & Gilad, 2014; Signor, Liu, Rebeiz, & Kopp, 2016; Wray, 2007; Yassin et al., 2016). This balance between stability and lability is a question that has been addressed philosophically, but little experimental evidence exists to suggest the mechanisms underlying these phenomena (Casci, 2005; Gibson, 2009; Green et al., 2017; Heranz & Cohen, 2010; Hermisson & Wagner, 2004; Nijhout et al., 2017; Rutherford et al., 2007; Stern, 2000; True & Haag, 2001).

Genetic perturbations (differences in gene expression due to mutation or recombination) are the focus of this study, in contrast to other work such as that on cryptic variation that is interested in the interaction between mutations and the environment (Gibson & Reed, 2008; Paaby & Gibson, 2016; Paaby & Rockman, 2014). We use spatially resolved quantitative gene expression to gain a picture of variation in pathway components that cannot be understood from transcriptomics where many tissues have been combined, or single cell studies where gene expression variation due to genetic differences will be confounded with differences in transcription. From this we can understand how much genetically based gene expression variation we see in natural populations, where gene expression differences are either compensated by other genes in the pathway (Estes, Phillips, & Denver, 2011; Landry et al., 2005; Thompson, Zakon, & Kirkpatrick, 2015), mitigated by robust gene network structure (i.e., incoherent feed forward loops) (MacNeil & Walhout, 2011), or below threshold levels (Felix & Barkoulas, 2015). In addition, we focus on natural genetic variation within populations, which has long been the purview of evolutionary and population geneticists, while developmental biologists generally focus on the effect of large mutations in otherwise isogenic backgrounds (Paaby & Gibson, 2016, but see Lott et al., 2007).

The dearth of work on developmental variation in wildtype genetic backgrounds is in part because developmental approaches have long been restricted to data that is semi-quantitative (i.e., in situ hybridization, antibody staining). Indeed, gene expression studies are generally spatial or quantitative, but not both. In addition, the data is generally not interrogated from a quantitative perspective, including replication and rigorous statistical models. Without quantitative replication and statistical tests, one cannot effectively compare developmental processes among conditions, including health versus disease. This is especially important given the potential for complex interactions between conditionally neutral differences in gene expression to result in disease phenotypes. Here we use hybridization chain reaction (HCR) to bridge this gap between developmental and quantitative or population genetics by quantitatively measuring spatial gene expression in multiple genotypes from two sexes of two species (Drosophila melanogaster and D. simulans) (Choi, Beck, & Pierce, 2014). Furthermore, we introduce rigorous replication allowing for statistical hypothesis testing. Multiplexing of four genes simultaneously also allows straight forward analysis of gene co-expression, compared to other techniques that require inference across samples that have been individually stained. We use this substantial developmental dataset to focus on the well-known genes factors driving ommatidia specification the eye imaginal disc in Drosophila (Atkins et al., 2013; Li, Ohlmeyer, Lane, & Kalderon, 1995; Raj, van den Bogaard, Rifkin, van Oudenaarden, & Tyagi, 2008; Shah, Lubeck, Schwarzkopf, & He, 2016; Tsachaki and Sprecher, 2011).

The Drosophila eye is formed from an imaginal disc, which is initially patterned by a wave of differentiation marked by a visible indentation of the tissue, termed the morphogenetic furrow (MF). The MF passes from the posterior to the anterior of the disc over a period of two days (90 min per adjacent row of cells), giving each disc an element of both time and space in development (Figure 1) (Roignant & Treisman, 2009). A strength of the eye disc as a model is that within the MF all cells are arrested at G1, and there will be less additional variability introduced due to differences between cells in their stage of cell division (Baker & Yu, 2001; Escudero & Freeman, 2007; Firth & Baker, 2005; Firth, Bhattacharya, & Baker, 2010). Furthermore, by averaging across all cells within a row we reduce the impact of differences in transcriptional bursting across the eye disc (Bothma et al., 2014; Fukaya, Lim, & Levine, 2016; Tantale et al., 2016).

FIGURE 1.

(a) A summary of the eye patterning genes and pathway explored in this paper. The position of the MF is shown in red, and its direction of movement indicated below. Regulatory relationships are illustrated either as repression (bar) or activation (arrow). Regulatory relationships which are unclear are shown as gray dotted lines. Note that hairy is activated by hedgehog at short range, and repressed by it indirectly at long range. Only the latter interaction is illustrated here. (b) Example images from the dataset, illustrating gene expression patterns of each gene. On the left is D. melanogaster and on the right D. simulans. Note that due to the different shapes of individual imaginal discs more of the antennal disc is shown for D. melanogaster than D. simulans, manifesting as a separate expression pattern to the left of each image. These areas were not included in the analysis, and indeed these expression patterns are shown prior to any image registration and are merely for illustrative purposes. The composite image makes the additional point that we were able to analyze co-expression patterns of all four genes without needing to stain each gene in different samples and infer gene co-expression patterns. (c) An illustration of the general expression pattern of each of the four genes in the study, along the anterior–posterior axis of the eye disc

The furrow is initiated by hedgehog, which represses hairy within the MF and activates it indirectly in front of the MF through the long-range diffusion of dpp within the tissue (Figure 1) (Felsenfeld & Kennison, 1995; Fu & Baker, 2003; Strutt & Mlodzik, 1997). hairy represses atonal, preventing precocious neural development anterior to the MF (though this role has been recently contested) (Bhattacharya & Baker, 2012; Brown, Sattler, Markey, & Carroll, 1991; Brown, Sattler, Paddock, & Carroll, 1995). hedgehog activates the expression of atonal, driving the MF anteriorly (Figure 1) (Heberlein, Wolff, & Rubin, 1993; Ma, Zhou, Beachy, & Moses, 1993; Greenwood & Struhl, 1999). atonal is the proneural gene in Drosophila, establishing the competency to become photoreceptor cells (Jarman, Grell, Ackerman, Jan, & Jan, 1994). The relationship between Delta/Notch and the other members of the pathway is more complex, but it is clear that in cells posterior to the furrow Delta/Notch repress atonal (Figure 1) (Baker, Yu, & Han, 1996; Dokucu, Zipursky, & Cagan, 1996; Firth & Baker, 2005; Gavish et al., 2016). There is also some evidence that Delta/Notch repress negative regulators of atonal at the furrow, such as hairy (Baonza and Freeman, 2001; Bhattacharya & Baker, 2009; Brown et al., 1995). In addition, there is some evidence suggesting that Delta/Notch are involved in the early stages of atonal induction, and alternatively that atonal activates its own transcription (Baker & Yu, 1997; Baker et al., 1996; Jarman et al., 1994; Li & Baker, 2001; Spencer, Powell, Miller, & Cagan, 1998; Sun, Jan, & Jan, 1998). Delta/Notch are expressed in the MF under the redundant control of hedgehog and dpp (Baonza & Freeman, 2001; Baonza & Freeman, 2005; Parks, Turner, & Muskavitch, 1995). There are many other genes involved in the specification of the eye disc that will not be mentioned here, in favor of focusing on the genes we have assayed. We analyze the spatial quantitative expression of hedgehog, hairy, atonal, and Delta to understand the evolving regulatory logic of the gene network and changes in spatial dynamics between sexes and species.

We term differences in gene expression between species, genotypes, and sexes “genetic stochasticity,” as there are no phenotypic differences between these Drosophila eyes other than size and proportion of photoreceptor subtype (Hilbrant et al., 2014). We interpret genetic stochasticity in quantitative spatial patterns of gene expression in light of regulatory relationships among them. We approach a spatial and quantitative analysis of these gene expression patterns in three ways, first by explicitly creating a spatial gene expression profile and comparing between genotypes, sexes, and species. Second, we were interested in examining if the regulatory relationship between these genes had evolved between species or harbors variation within a species. Lastly, we investigated the possibility that the spatial relationship between these genes relative to the MF had evolved or harbors variation within populations. These analyses summarize developmental data using quantitative genetic models, to uncover gene expression differences across tissues that would otherwise be undetected using conventional qualitative methods.

2 ∣. METHODS

2.1 ∣. Fly stocks

D. simulans were collected from the Zuma organic orchard in Zuma beach, CA in the spring of 2012 (Signor, New, & Nuzhdin, 2017). They were inbred by 15 generations of full sib crosses. D. melanogaster were collected in Raleigh, North Carolina and inbred for 20 generations (Mackay et al., 2012). Thus, any differences detected between species, sexes, or genotypes is due to natural variation.

2.2 ∣. Staging and dissection of larvae

All flies were reared on a standard medium at 25°C with a 12-hr light/12-hr dark cycle. Five replicate eye discs were isolated from each of three strains, two sexes, and two species, for a total of 60 discs. Vials were used a single time for collection to avoid pseudo-replication, such that every genotype and sex was collected from a separate replicate population. Density of the cultures was controlled, as a standard number of parents were given two hours to lay their eggs. 120 hr after hatching, wandering 3rd instar larvae were placed in phosphate buffered saline (PBS) and separated by sex. Their guts were carefully removed posteriorly and their body was inverted anteriorly to expose the brains, eye discs and mouth hooks. After fixation and labeling (described below), eye discs were isolated and mounted. The MF moves one row every 90 minutes; thus, the expectation is that the MF will be consistently located within the same approximate location. Further, replicates were not conducted concurrently meaning that variation due to differences in timing will be randomly distributed across genotypes and sexes. If there are systematic differences in the position of the furrow, for example because there is some evidence that D. simulans develops approximately half a day more slowly than D. melanogaster, we test explicitly for an effect of the position of the morphogenetic furrow on gene expression in the following sections, for both species and sex (Génétique, 1983; Kuntz & Eisen, 2014).

2.3 ∣. Hybridization chain reaction (HCR)

HCR is unique in that it produces gene expression patterns that are both quantitative and spatial. While other approaches such as FISH can be adapted to detect individual transcripts, HCR has a linear signal that is 20× brighter than FISH, it reduces non-specific background staining, and it can detect 88% of single RNA molecules in a cell with an appropriately low false discovery rate (Ma & Moses, 1995; Pan & Rubin, 1995). It is also highly repeatable, with different sets of probes targeted to the same gene showing correlations of .93–.99 (S. Fraser, pers. comm.). The DNA probes were designed and synthesized by Molecular Instruments (Choi et al., 2014) (S1 Table). Four genes were multiplexed in each preparation as orthogonally-designed hairpins allowed the simultaneous amplification of their target sequences (Figures 1 and S1). Each target mRNA was detected using five DNA probes to annotate the position and expression levels for each of the four assayed genes (hairy, atonal, Delta and hedgehog). Each probe contained two-initiator sequences (I1 and I2) that bound to a specific amplifier.

The protocol for HCR was modified from Choi et al. (2014) and is described briefly. The full protocol is available in S1 File. Inverted 3rd instar larvae were fixed in 4% paraformaldehyde diluted with PBS containing .2% Tween 20 (PBST). After fixation, larvae were washed with PBST, then increasing concentrations of methanol (30%, 70%, and 100%) at 25°C. Larva were stored in 100% methanol at −20°C. Methanol-fixed samples were thawed, washed with ethanol, re-permeabilized in 60% xylene, washed with ethanol, then methanol and rehydrated with PBST. Samples were permeabilized with proteinase K (4 mg/ml), fixed in 4% formaldehyde then washed with PBST at 25°C. Finally, at 45°C, samples were pre-hybridized for 2 hr before the addition of all the probes. The probe-hybridized larvae were placed in wash buffer (Molecular Instruments) at 45°C to remove excess probes. Fluorescently labeled hairpins were snap-cooled then added to the samples at 25°C and placed in the dark to amplify the signal. Afterwards, samples were washed in 5X SSCT solution, isolated in PBST, then placed in Prolong Gold anti-fade mounting medium (Molecular Probes).

2.4 ∣. Microscopy

Three dimensional images of mounted, HCR stained 3rd instar larva eye discs were acquired on a Zeiss LSM 780 laser scanning microscope (Carl Zeiss MicroImaging, Inc., Thornwood, NY) with Objective Plan-Apochromat 63×/1.40 Oil. The gain was adjusted to avoid pixel saturation.

2.5 ∣. Extraction of gene expression profiles

The first steps in the image analysis is bringing each image to the same orientation and segmenting it. Image segmentation produces a mask in which pixels are assigned to objects or background. Here the objects are one or several mRNA molecules. Then the cellular structure of the imaginal disc is approximated using a hexagonal array. Though the real underlying cell structure of the imaginal disc is technically able to be recognized, this was unsatisfactory in our data due to imaging noise. Thus, at the second step using the R package hexbin we constructed a partition of the imaginal disc area into elements that represent pseudo-cells and have a biologically-relevant hexagonal shape (Brennan, Ashburner, & Moses, 1998). The number of pseudo-cells was selected by visual inspection of the combined image in which the hexagonal structure was overlaid onto the atonal channel to verify fit. We are primarily interested in expression profiles around the MF, providing us a convenient landmark to align images from different preparations, thereby assigning coordinates to the pseudo-cells. However, deformations of the eye disc during growth and preparation sometimes distorts the MF. We used splines to correct for any bending or deformation of the MF. We note that the use of pseudo-cells could potentially result in gene expression from single cells being segregated into multiple cells and vice versa, however as it effects all genes equally in a given locale, this should have little effect on the outcome of the analysis. Next, using the histograms of cumulative pixel intensities of objects in expression domains and non-expressing areas we estimated the typical intensity of a transcript and typical background signal, respectively. Consequently, the cumulative intensities greater than the background are divided by the intensity attributed to single mRNA molecule to yield counts of mRNA molecules. This normalizes the expression profiles and corrects for differences in microscope gain between images. Finally, the gene expression profiles are estimated for every pseudo-cell.

2.6 ∣. Morphological reconstruction and contrast mapping segmentation

To detect gene transcripts within the image stacks we applied a version of the MrComas method that was modified for processing 3D images (Please see S2 Figure for an illustration of the different steps in the process described below, in addition Kozlov et al., 2017 provides a more thorough description of the process). This approach first enhances contrast within the image and reduces noise. The images were enlarged by a factor of four with the nearest-neighbor algorithm. They were processed by morphological reconstruction using both opening and closing, where closing (opening) is dilation (erosion) that removes extraneous dark (bright) spots and connects bright (dark) objects (Vincent, 1993). The contrast mapping operator assigns each pixel the maximum value between the pixel-by-pixel difference of the reconstructed images and their pixel-by-pixel product and produces the rough mask for each channel. An image, I, is mapping from a finite rectangular subset L onto the discrete plane Z² into a discrete set 0, 1, …, N − 1 of gray levels. y denotes the pixel of an image I inside a structural element B.

Let the dilation δ_Band erosion ϵ_B by structural element B be defined as:

$${\delta }_B(I)={}_{\gamma {\epsilon }_B}I(y)=IB{ϵ}_B(I)={}_{\gamma {\epsilon }_B}I(y)=IB$$ (1)

where ∨ and ∧ denote infimum and supremum respectively. Then formulae:

$${\delta }_{I,B}^1(J)=(JB)I{\epsilon }_{I,B}^1=(JB)I$$ (2)

denote geodesic dilation ${\delta }_{I,B}^1$ and erosion ${\epsilon }_{I,B}^1$. Binary reconstruction extracts those connected components of the mask image which are marked on the marker image, and in grayscale it extracts the peaks of the masked image marked by the marker image. Using the dilated masks image I as the marker J : J = δ_B(I) defines closing by reconstruction:

$${\gamma }_B(I)={\epsilon }_{I,B}^1{\epsilon }_{I,B}^1\cdots {\epsilon }_{I,B}^1[{\delta }_B(I)]$$ (3)

Opening by reconstructions uses eroded mask I as a marker J:

$${\varphi }_B(I)={\delta }_{I,B}^1{\delta }_{I,B}^1\cdots {\delta }_{I,B}^1[{\epsilon }_B(I)]$$ (4)

Then the difference between closing and opening by reconstruction has the meaning of the gradient:

$${\nabla }_B(I)={\gamma }_B(I)-{\varphi }_B(I)$$ (5)

To create strong discontinuities at object edges and flatten signal with the objects the contrast mapping operator takes a maximum between the difference and the pixel-by-pixel produce of the reconstructed images and produces a rough mask for each channel:

$$R=\max \{{\nabla }_B(I),{\gamma }_B(I){\varphi }_B(I)\}$$ (6)

Subsequently, this mask was subjected to distance transform, which substituted each pixel value with the number of pixels between it and the closest background pixel. This operation creates “peaks” and “valleys” of intensity inside foreground objects. To split erroneously merged objects watershed transform was applied, which treats the whole image as a surface and intensity of each pixel as its height and determines the watershed lines along the tops of ridges separating the catchment basins (Meyer, 1994). The quality of segmentation is assessed visually by inspection of the object boarders overlaid with the original image. Finally, each mask is returned to its original size and quantitative measures are made of shape and intensity characteristics such as the number of pixels, as well as their mean and standard deviation in the detected object. MrComas is free and open source software available at http://sourceforge.net/p/prostack/wiki/mrcomas. Application of MrComas to synthetic images to evaluate performance and comparison to existing methods are described in Kozlov et al. (2017).

2.7 ∣. Approximating the MF

We defined the position of the MF as the location of maximum atonal expression. The shape of the MF was approximated with a spline using function smooth.spline in R. The degrees of freedom and other parameters were chosen to make the approximation coincide visually with a MF image.

2.8 ∣. Inferring counts of transcript number

Segmentation of the image provided a table of coordinates and the shape and intensity characteristics of detected transcripts. Here, we applied filtering steps to remove false positives and determine the count of mRNA transcripts. First, we assumed the object we detected as least intense but most frequent corresponds to a single mRNA molecule. Then, we inferred background intensity for objects outside of well-annotated domains of expression of the four genes. Assuming that the majority of true objects contain a single molecule, we compare the distribution of cumulative intensities of particles in expression domains and areas of known non-expression to obtain the typical intensity of a true single molecule and a false positive, respectively. All detected signals that were lower than the typical intensity of a false positive were removed from the dataset. The number of removed objects is typically less then 10%. All other cumulative pixel intensities were divided by the typical intensity of a true single molecule as normalization coefficient to yield an estimate of the number of mRNA transcripts.

2.9 ∣. Image registration

We applied an affine coordinate transformation to each eye disc to make the corresponding maxima and the width of expression patterns of four genes in different eyes coincide as closely as possible. To do so, we shifted the coordinate system of each eye to its center and also scaled them in the A-P direction. The center of the pattern in A-P direction is the MF. This aligns the position of maximum atonal expression (defined as the center of the MF), such that gene expression patterns can be compared across discs.

We mapped the expression patterns to a unified hexagonal structure in order to make comparisons between pseudo-cells from individual imaginal discs. The unified cell structure was constructed using the R package hexbin. Each cell in the unified grid represents an “average” cell from individual eyes. The size of a hexagon in the unified grid is greater or equal than the cell size in the individual eye. Thus, the number of molecules in each unified cell in the mapped pattern equals the mean over the cells from native pattern that are covered by this unified cell. After such coordinate transformation, the MF region is defined as 20 cells on either side of the MF, to focus the analysis on the area of interest (the MF).

2.10 ∣. Filtering and quality controls for each eye disc

Some eye imaginal discs were damaged or deformed in the process of dissection or mounting, resulting in regions of erroneous gene expression, such as disruptions to the MF. The expression profile of each disc was examined by eye and these regions were individually trimmed out of the final dataset. At the edges of each eye disc the pattern of the MF was also degraded, so each eye disc was trimmed dorso-ventrally prior to analysis. Five outliers were excluded from the dataset of 60 discs, determined as a single member of the five replicates per genotype/sex with more than a 3× difference in expression values. This resulted in a final dataset of 55 eye discs.

There are no phenotypic differences between the eyes of D. melanogaster and D. simulans other than size and proportion of photoreceptor subtype. Size is included as a co-factor in the relevant models discussed below, and results only in a larger area being patterned for ommatidia rather than a difference in pattern. In addition, photoreceptor subtype is not determined or affected by the genes expressed during the initial patterning phase in the MF (Cook, Pichaud, Sonneville, Papatsenko, & Desplan, 2003; Johnston and Desplan, 2014; Johnston et al., 2011; Wernet et al., 2006).

2.11 ∣. Analysis of individual spatial gene expression patterns

We were primarily interested in variation in gene expression profiles across the eye disc, that is using differences in expression averaged across rows along the x-axis. While the y-axis is of interest, variation in the shape, size, degree of deformation, and occasional damage to the disc made this analysis intractable. We fit curves to each gene expression profile using the mgcv package in R, using a generalized additive model with integrated smoothness estimation. Smoothing terms are represented using penalized regression splines. predict.gam was used to fit the curves to the original range of values and down sample the curves to eight points. MANOVAs were performed using the “Pillai” test for species × genotype × sex.

2.12 ∣. Modeling framework to understand variation and evolution of the eye patterning gene network

We wanted to understand if variation existed within the regulatory logic of hairy, atonal, Delta, and hedgehog, or if there had been evolution between species. To understand the regulatory logic between genes we focused on biologically relevant relationships, such as the regulation of hairy by Delta and hedgehog, but excluded such relationships as hairy and atonal. This was due to the low overlap between hairy and atonal expression domains, where including cells where only one or another was expressed would artificially create a relationship between expression levels. Both hairy and atonal are downstream, directly or indirectly, of Delta and hedgehog thus it was these relationships that were modeled. We limited the analysis to cells where all genes included were expressed in at least ten molecules.

In the previous analysis, we investigated variation in the cryptic spatial quantitative expression pattern of genes in the MF. Here, we will investigate the possibility that genes in the MF have evolved, or harbor variation, for how they affect each other in particular cells. For example, is high atonal expression associated with high expression of hedgehog, given that hedgehog activates atonal? We used the following equations to determine the relationship between the expression of these genes:

$$hairy(i,s)=k^{Dl}\times Dl(i,s)+k^{hh}\times hh(i,s)+\alpha$$ (7)

$$atonal(i,s)=k^{Dl}\times Dl(i,s)+k^{hh}\times hh(i,s)+\alpha$$ (8)

The coefficient k and constant α were fit using standard methods for multiple regression. Here hairy(i,s) and atonal(i,s) are the measured expression level of each gene in cell i in individual s. Dl(i,s) and hh(i,s) are vectors containing the corresponding expression levels of hairy and atonal's regulators Delta and hedgehog. To determine if the regulatory logic is the same between genotypes and species we can then use the regression coefficients from these models in a MANOVA. We note that we cannot exclude the possibility that other unmeasured genes are responsible for producing this variation.

2.13 ∣. Model for understanding overall variation and evolution of MF structure

Lastly, we wanted to understand if there is variation in the relationship between the MF and gene expression, or if variation existed for the size of the MF overall. Two processes occurred to make the MF comparable between samples, the MF was shifted to occupy the same position depending upon the position of maximum atonal expression, and the MF was scaled to occupy the same total area. If the MF was consistently in a different position for different species, sexes, or genotypes, meaning that the amount it needed to be shifted such that the center of the MF occupied the same location was consistently different for different factors, this might imply genetically based differences in the rate at which the MF proceeds. We performed an ANOVA in R using the amount that each disc was shifted to test for systemic differences in MF position. To test for differences in the amount that each disc was scaled, we need to control for differences in size in the original disc, as the scaling factor will depend both on the size of the original disc and the width of the MF relative to the disc. To account for differences in size we include the number of rows of cells in the original eye disc as a cofactor and perform ANOVA in R (Figure S2). The number of rows of cells is estimated along the D-V axis in the entire eye disc by the R program hexbin, and is not restricted to the region surrounding the MF.

3 ∣. RESULTS

3.1 ∣. Individual spatial gene expression patterns

First, to characterize the spatiotemporal dynamics of transcriptional activity along the anterior-posterior axis, we took the spatial average of signal across the dorsal-ventral axis and compared between genotypes, sexes, and species (Figure 2a, we note that the smoothed curves in the figures were created using smooth.spline in ggplot2, which is not the same method for curve fitting as in the analysis). We found abundant spatial quantitative variation in expression profiles (Figures 2-3). The expression profile of hairy around the MF harbors variation between genotypes and there is an interaction between genotype and sex (Table 1, p = 2 × 10⁻³, p = 0.02). There has also been evolution between species for hairy (Table 1, p = 3 × 10⁻⁴). While atonal has not evolved between species, there is variation in expression profile between genotypes, sexes, and there is an interaction between genotype and sex (Figures 3a-3c, Table 1, Figure S2, p = 4 × 10⁻⁴, p = .02, p = .02). Surprisingly, given the conservation of Delta in general, Delta harbors variation in spatial quantitative expression behind the MF between genotypes and sexes (Table 1, p = 2 × 10⁻³, p = 7 × 10⁻⁴) and there are significant interactions between genotype and sex (Figures 2b and 2c, Table 1, p = 2 × 10⁻³). There has also been evolution of Delta between species, and evolution of the interaction between species and sex (Table 1, Figure S2, p = .03, p = 3 × 10⁻⁴). hedgehog is not different between species but is significantly different between genotypes, sexes, and there is an interaction between the two (Table 1, p = 5 × 10⁻⁴, p = .05, p = 1 × 10⁻³). There is also a significant interaction between species and sex (Table 1, p = .01).

FIGURE 2.

(a) This is an example of a curve being fitted to the gene expression profiles, though note that the curve corresponds to the average in a given row (x-axis). The hexagons are intended to represent cells with varying amounts of hairy expression, from the highest (red) to the lowest (white). (b) An illustration of variation in Delta expression between species and sexes. Curves shown are fitted to all genotypes within a sex and species. (c) An illustration using the imaginal disc of how Delta expression varied between species and sexes, with lower expression in D. melanogaster females and D. simulans males (d) Evolution of Delta illustrated within the context of the gene network, illustrating how changes in Delta expression are not perturbing the gene network and result in phenotypically normal Drosophila

FIGURE 3.

(a) An illustration of variation in atonal expression between genotypes and sexes. Curves shown are fitted to each genotype and sex. (b) An illustration using the imaginal disc of how atonal expression varied between genotypes, with lower expression in females of D. melanogaster R153 and males of D. simulans Sz208. D. simulans Sz173 has lower expression than females of Sz208 but it is not sexually dimorphic. (c) Evolution of atonal illustrated within the context of the gene network, illustrating how changes in atonal expression are not perturbing the gene network and result in phenotypically normal Drosophila

TABLE 1.

The results of the MANOVA for species × genotype × sex for each gene

hairy						atonal
Effect	numDF	denDF	F-value	p-value		Effect	numDF	denDF	F-value	p-value
Species	8	35	5.1	3 x 10−4		Species	8	35	1.53	0.18
Genotype	32	152	2.07	2 x 10−3		Genotype	32	152	2.31	4 x 10−4
Sex	8	35	1.08	0.32		Sex	8	35	2.6	0.02
Species x Sex	8	35	1.94	0.08		Species x Sex	8	35	1.91	0.09
Genotype x Sex	32	152	1.73	0.02		Genotype x Sex	32	152	1.73	0.02
Delta						hedgehog
Effect	numDF	denDF	F-value	p-value		Effect	numDF	denDF	F-value	p-value
Species	8	35	2.43	0.03		Species	8	35	0.87	0.55
Genotype	32	152	2.07	2 x 10−3		Genotype	32	152	2.28	5 x 10−4
Sex	8	35	4.55	7 x 10−4		Sex	8	35	2.23	0.05
Species x Sex	8	35	5.03	3 x 10−4		Species x Sex	8	35	2.88	0.01
Genotype x Sex	32	152	2.03	2 x 10−3		Genotype x Sex	32	152	2.17	1 x 10−3

Significant p-values are indicated in bold with gray shading.

Thus, hairy and Delta have evolved different spatial quantitative expression patterns between species, while Delta, atonal, and hedgehog harbor variation between genotypes and sexes. Given that there are regulatory relationships between these genes, it is interesting to see that they do not all harbor variation for the same factors. This could potentially be due to the influence of other unmeasured regulatory factors, or to variation in the relationship between these genes and other components in the gene regulatory network. However, whatever the source of “buffering” of the network, be it the effect of other genes or threshold effects on development, the fact that this information is not retained within the steps of the pathway supports our supposition that this variation does not ultimately have a phenotypic effect (see section 4).

3.2 ∣. Variation and evolution of the eye patterning gene network

The regulatory effects between genes can be discerned from the degree of quantitative co-localization, and variation in this co-localization between species, genotypes, and sexes. There has been evolution in the regulatory logic of hairy and its upstream regulators Delta and hedgehog between species (Figure 4, Table 2, p = .03, see S2 File for the r² values from the individual regressions). There is also variation between sexes in the regulatory logic of hairy and its upstream regulators Delta and hedgehog (Table 2, p = .03). There has been significant evolution of the regulatory logic of atonal, in a significant interaction between species and sex (Figure S3, p = 1 × 10⁻³). Furthermore, while there was no significant effect of genotype for hairy, there is for atonal, indicating that there is variation segregating in the population affecting the relationship between atonal, hedgehog, and Delta (p =1.6 × 10⁻⁵, see S3 File for the r² values from the individual regressions). There is also a significant interaction between genotype and sex (p = 1 × 10⁻³). Thus, the relationship between hairy and atonal and their regulators has evolved between species and sexes in hairy, and between genotypes and sex in atonal. We illustrate this difference between species in Figure 4, where a different relationship between hairy and hedgehog is visible between D. melanogaster and D. simulans. In brief, the frequency of cells with a given log transformed level of expression are plotted against one another for hairy and hedgehog. hairy is primarily expressed anterior to the MF and hedgehog posterior, and they have a different regulatory relationship in each region with hedgehog activating hairy indirectly at long range (anterior) and repressing it short range (posterior). This is reflected in the frequency of cells expressing both genes for D. melanogaster, where anterior to the MF there is a high frequency of hairy expressing cells and a low frequency of co-occurring high hedgehog expression. Posterior to the MF the opposite is true, with high expression of hedgehog lacking concordance with any expression of hairy. In D. simulans, posterior to the MF, this relationship is the same as in D. melanogaster. However, in anterior to the MF this is not the case. Expression of hairy and hedgehog both increase as the other increases, with widespread co-occurrence.

FIGURE 4.

An example of variation in regulatory logic between D. simulans and D. melanogaster for hairy and hedgehog. The heat map illustrates the density of points, and thus reflects the frequency of a given co-expression profile between hairy and hedgehog. Gene expression values were log-transformed to better illustrate lower values and split between anterior to the MF and posterior to the MF. The split between the two regions was to investigate the possibility that hedgehog had a different regulatory relationship with hairy depending upon its relationship to the MF, given that hedgehog is thought to activate hairy long range (indirectly) and repress hairy short range

TABLE 2.

The results of the MANOVA for species × genotype × sex for the regulatory relationship between hairy, Delta, and hedgehog, and atanol, Delta, and hedgehog

hairy
Effect	numDF	denDF	F-value	p-value
Species	2	41	4.03	0.025
Genotype	8	84	1.2	0.31
Sex	2	41	3.95	0.027
Species x Sex	2	41	0.78	0.46
Genotype x Sex	8	84	1,11	0.37
atonal
Effect	numDF	denDF	F-value	p-value
Species	2	41	2.64	0.08
Genotype	8	84	5.43	1.6 x 10−5
Sex	2	41	0.67	0.52
Species x Sex	2	41	8.09	1 x 10−3
Genotype x Sex	8	84	2.41	0.02

Significant p-values are indicated in bold with gray shading.

3.3 ∣. Variation and evolution of MF structure

The amount that the eye discs were shifted is not significant for genotype, sex, or species. This suggests that any differences in development time are not significant between species, genotype, or sex, and/or that the exact position of the furrow is not having an effect on these gene expression patterns. However, the amount that they were scaled is, after accounting for original differences in size, between species (p = 1.38 × 10⁻⁶). This suggests that the total relative width of the MF varies between species, but not between genotypes or sexes. This is also suggestive of evolving interrelationships among genes that could result in broader or narrower areas in which they enhance or suppress expression of one another.

4 ∣. DISCUSSION

We present here a replicated, quantitative, and spatially explicit analysis of the expression of genes driving cell differentiation in the Drosophila eye. Our results summarize a complicated pattern of variation sorting in the gene network involved in patterning the MF. For example, the overall shape of the expression of hedgehog across the eye disc is different between genotypes, sexes, and there is an interaction between species and sex and genotype and sex. hedgehog upregulates hairy, but hairy has differences in expression between species (which hedgehog does not), genotypes, and there is an interaction between genotype and sex. Thus, the differences seen in upstream regulators, such as hedgehog, are not recapitulated in their downstream targets. In another example, Delta/Notch is expected to repress atonal, but while Delta/Notch is significant for all categories tested atonal is only significant for genotype, sex, and their interaction. It is possible that this variation is being mitigated or dampened by other regulatory factors not assayed here, or that certain aspects of genetic background are more or less sensitive to variation. For example, fixed variation between species could dampen variation at Delta/Notch while sorting variation remains sensitive between genotypes, which propagates to atonal.

It may be that all of this variation is within levels tolerated by the network, as it has been shown that gene networks can have thresholds of variation, below which differences in expression are effectively neutral. These thresholds can also be two sided, creating a sigmoidal curve the center of which is neutral phenotypic space (Felix & Barkoulas, 2015). Many studies have shown a relative insensitivity to variation in gene dosage, for example, in Drosophila early embryos the bicoid gradient results in normal development at one to four dosages of the gene, but markedly abnormal development at six or more (Liu, Morrison, & Gregor, 2013; Lucas et al., 2013; Namba, Pazdera, Cerrone, & Minden, 1997). It is also possible that the “genetic stochasticity” documented in these genes is in fact deleterious, and is being compensated for elsewhere in the network. While most deleterious mutations are purged by selection, they may rise in frequency due to genetic drift or hitchhiking, among other possible causes (Burch & Chao, 1999; Chun & Fay, 2011; Estes & Lynch, 2003; McKenzie & Clarke, 1988). This type of compensatory mutation has been documented in microbial and animal systems (Burch & Chao, 1999; Brown et al., 2010; Charusanti et al., 2010; Estes & Lynch, 2003; Estes et al., 2011; Maisnier-Patin & Andersson, 2004; McKenzie, 1993; McKenzie, Whitten, & Adena, 1982; Moore, Rozen, & Lenski, 2000; Stoebel, Hokamp, Last, & Dorman, 2009; Szamecz et al., 2014). Recently cell cycle heterogeneity has been implicated in the appearance of widespread noise in development, however this is likely not responsible for the genetic stochasticity observed here as the majority of MF cells assayed here are arrested at G1 (Keren et al., 2015; Kumar, 2013).

The differences in the co-expression patterns of hairy, atonal, and their upstream regulators hedgehog and Delta/Notch may represent the early stages of developmental systems drift, where pathway components can evolve different relationships while maintaining the same phenotypic output (True & Haag, 2001). This type of “quantitative developmental systems drift” has been proposed to be a pervasive feature of evolving developmental systems (Crombach et al., 2016; Wotton et al., 2015). It may also be that the overall relationship between the genes has not evolved, but that variation between the network components is beneath some critical threshold that would results in differences in downstream regulation—that is that co-expression patterns may show evolution and variation, but the basic inputs and outputs of the network are conserved. These are not fundamentally different ideas, rather they may represent different points on an evolutionary continuum. If developmental systems drift is at work, it could potentially be due to changes in these critical thresholds, for example if the threshold is lower in one species then less of an upstream gene is needed to create the same output. The evolution and maintenance of expression thresholds and developmental systems drift is potentially an interesting area of future research.

There have been other semi-quantitative approaches to studying spatial gene expression patterns. In another study on orthodenticle, the authors found that the spatial and temporal pattern of gene expression was conserved but the amount of gene product was not, though this work was not strictly quantitative given that measurements were from in situ hybridization and reporter constructs and there was no rigorous statistical testing (Goering et al., 2009). This is in contrast to our results which showed significant differences in the spatial relationship between gene expression patterns between species. Other semi-quantitative works on the Drosophila embryo using in situ hybridization found that the regulatory relationship between genes in the anterior-posterior blastoderm patterning network were conserved, despite differences between species in their spatio-temporal pattern (Fowlkes et al., 2011; Wunderlich et al., 2012). Here we find that the regulatory relationship between atonal and hairy, and their regulators hedgehog and Delta, has evolved between species, sexes, and genotypes.

One of the important messages from this work is that rigorous statistical testing can uncover molecular variation in spatial and quantitative developmental gene expression. Using the type of replication applied in quantitative genetics with developmental data we were able to apply rigorous statistical models to micro-evolutionary variation in development. Despite this variation, observed with repeatable observations of developmental patterns among natural genotypes, the phenotypes of all flies are normal. This points to a potential abundance of hidden noise in spatial and quantitative gene expression. The evolutionary approach to development generally targets large changes that have occurred over broad phylogenetic distances (Ito et al., 2013; Jeong et al., 2008; Kopp, Duncan, Godt, & Carroll, 2000; Reed et al., 2011; Rosenblum, Römpler, Schöneberg, & Hoekstra, 2010; Signor et al., 2016; Yassin et al., 2016). Accordingly, the presence of abundant underlying variation is perhaps not a huge surprise. But it does have large implications, as it suggests that developmental models should be modified so that such abundant genetic variation is buffered from perturbing the final phenotype. In the future, application of this type of replicated, quantitative, spatially resolved data will have unique insights into the penetrance of disease phenotypes and the origin of developmental defects.