Making the body plan: Precision in the genetic hierarchy of Drosophila embryo segmentation

¹ Department of Applied Mathematics and Statistics and The Center for Developmental Genetics, The State University of New York at Stony Brook, New York, 11794-3600, USA.
phone +1-631-632-8352; fax +1-631-632-8490
Email: spirov@kruppel.ams.sunysb.edu

² The Sechenov Institute of Evolutionary Physiology and Biochemistry, Russian Academy of Sciences, 44 Thorez Ave., St. Petersburg, 194223, Russia.
³ Mathematics Department, British Columbia Institute of Technology, 3700 Willingdon Ave., Burnaby, B.C., Canada, V5G 3H2; and Chemistry Department, University of British Columbia, 2036 Main Mall, Vancouver, B.C., Canada, V6T 1Z1.
phone +1-604-456-8199; fax +1-604-432-9173
Email: David_Holloway@bcit.ca

^* corresponding author

Edited by E. Wingender; received September 30, 2002; accepted December 12, 2002; published December 16, 2002

Abstract

We quantify fluctuations in protein expression for three of the segmentation genes in the fruit fly, Drosophila melanogaster. These proteins are representative members of the first three levels of a signalling hierarchy which determines the segmented body plan: maternal (Bicoid protein); gap (Hunchback protein); and pair-rule (Even-skipped protein). We quantify both inter-embryo and inter-nucleus (within a single embryo) variability in expression, especially with respect to positional specification by concentration gradient reading. Errors are quantified both early and late in cleavage cycle 14, during which the protein patterns develop, to study the dynamics of error transmission. We find that Bicoid displays very large positional errors, while expression of the downstream genes, Hunchback and Even-skipped, displays far more precise positioning. This is evidence that the pattern formation of the downstream proteins is at least partially independent of maternal signal, i. e. evidence against simple concentration gradient reading. We also find that fractional errors in concentration increase during cleavage cycle 14.

Key words: morphogens, pattern formation, gradient reading, genetic cascades, positional information, gene networks, signalling hierarchy, concentration noise, embryo segmentation, maternal genes, gap genes, pair-rule genes, embryo-to-embryo variability, nucleus-to-nucleus variability, genetic algorithms, stochastic, error transmission, axis specification, image processing, Drosophila melanogaster

Introduction

Positionally specific gene regulation is believed to be a fundamental process underlying embryological development, responsible for the patterned cell differentiation at the heart of many developmental events. Spatial pattern has long been thought to be specified by gradients [Boveri, 1904; 1910; reviewed in Sander, 1994], more recently identified as gradients in concentration of regulators of differentiation (e. g. signalling molecules or transcriptional regulators). Wolpert [1969, 1970, 1971] has elaborated greatly on this idea, that cells acquire positional information (knowing where they are in the organism) by reading the local concentration of gradient molecules.

In the early development of the fruit fly (Drosophila melanogaster), a gradient of maternally-derived protein, Bicoid (Bcd), is laid down in the egg [Driever and Nüsslein-Volhard, 1988a], with highest concentration at the anterior (head) end, and lowest concentration in the posterior (Fig. 1A). Bcd is a transcriptional regulator, and lies at the top of a genetic hierarchy which results in the proper segmentation of the body plan in the anteroposterior (AP) axis. Bcd is cited as a prime example of positional specification by gradients [Wolpert, 2002; Gilbert, 1994; Driever and Nüsslein-Volhard, 1988b; Struhl et al., 1989]: downstream segmentation genes are regulated by Bcd in a concentration dependent manner. The gap class of genes refine the exponential input of the Bcd gradient, being expressed in domains with sharp spatial boundaries (Fig. 1B). These gap genes, in turn, regulate pair-rule class genes, giving narrow stripes of expression (Fig. 1C). Pair-rule pattern is the first manifestation of the segmented body plan characteristic of insects. Local reading of concentration gradients has generally been the favoured mechanism within this patterning hierarchy, but there are a number of proposals for pair-rule patterning [e. g. Meinhardt, 1988; Lacalli, 1990; Reinitz and Sharp, 1995; Hunding, 1993; Bodnar, 1997] and the question remains quite open.

In the present work, we take advantage of a large databank of segmentation protein expression data at the State University of New York, Stony Brook (http://flyex.ams.sunysb.edu/FlyEx/) [Kosman et al., 1998; 1999] to begin to do statistics that should reflect on some fundamental, and old, problems in the role of gradients in biological development. These are: size regulation, or scaling, i. e., pattern remains unaltered, despite high variability in embryo size; gradient stability in the face of fluctuations in temperature and other environmental factors; transmission of positional information by gradient concentration, in the presence of inherent (due to disorder of molecular motion) concentration fluctuations.

This last item, transmission of information, is an avenue for testing the Wolpert-style mechanism of positional specification by local gradient reading. Such a mechanism does not provide for error suppression. Amplification of errors during signalling would be compatible with local gradient reading down the hierarchy: suppression of errors is evidence against local gradient reading. In general, by quantifying segmentation errors, we can begin to offer experimental results to complement the recent upsurge in theoretical interest in the role of noise in gene networks [e. g. McAdams and Arkin, 1999; Hasty et al., 2000; Thattai and van Oudenaarden, 2001].

Figure 1: Confocal microscope images of a fruit fly embryo, stained with fluorescent antibodies to three different gene products (Bicoid, Hunchback and Even-skipped), in late cleavage cycle 14 (about 4 hours of development; see Foe and Alberts 1983, for Drosophila staging). Cellularization takes place in this cleavage cycle: at the beginning of the cycle, most of the nuclei (approximately 6000) are in a single syncytial layer at the surface of the egg; by the end of the cycle membranes have grown inwards, creating a cellular epithelium. Fluorescence intensity is proportional to protein concentration. Anterior left, posterior right, dorsal up, ventral down. Each dot is a nucleus (approximately 3000 in view at this orientation, but not all will have high fluorescence for a particular protein). A) Stain for Bicoid protein, maternal class. B) Stain for Hunchback protein, gap class. C) Stain for Even-skipped protein, pair-rule class.

Methods and algorithms

Images of Drosophila Gene Expression.
Gene expression (at the translational level) was measured using fluorescently-tagged antibodies as described in Kosman et al. [1998]. For each embryo a 1024 x 1024 pixel image with 8 bits of fluorescence data in each of 3 channels was obtained (Fig. 1 is from the same embryo: parts A, B, and C of the figure are from separate channels, visualizing three different proteins with one snapshot). Image processing [Kosman et al., 1999] transforms each image into an ASCII table containing a series of data records, one for each nucleus. About 2500-3500 nuclei are described for each image. Each nucleus is characterized by a unique identification number, the AP and dorsoventral (DV) coordinates of its centroid, and the average fluorescence levels of three gene products. The overall result is the conversion of an image to a set of numerical data which is then suitable for further processing. There are currently over 1000 embryo images, with data on 14 segmentation genes, in the databank (http://flyex.ams.sunysb.edu/FlyEx/). In this paper, we present data on three segmentation genes: Bicoid (Bcd), Hunchback (Hb), and Even-skipped (Eve). Bcd and Eve data are taken from the same embryos, Hb data is pooled from these and other embryos.

Temporal Classification.
All embryos under study belong to nuclear cleavage cycle 14 [Foe and Alberts, 1983]. It is in this cycle, about an hour long, that the segmentation patterns develop. Embryos were classified into eight time classes within the cycle, primarily by visual inspection of the (highly dynamic) expression pattern of the pair-rule gene Eve [Myasnikova et al., 2001]; all images were stained for Eve, as well as two other proteins. This classification was later verified by observation of the other patterns and by membrane invagination data. In this paper, we compare expression at early (time class 1 plus time class 2, pooled) versus late (time class 5 plus time class 6, pooled) cycle 14.

Anteroposterior (AP) Expression Profiles.
Because the expression of segmentation genes is largely a function of position along the AP axis, it is natural to use the AP profiles of gene expression as a first step towards characterization of embryo-to-embryo variability. Due to the discrete and irregular location of nuclei, and slight DV dependencies of AP expression (stripe bending, Fig. 1C), extraction of AP profiles is a non-trivial problem in image processing. In this paper, we present results for stripe-straightened images. The stripe-straightening procedure is based on a genetic algorithms approach for finding the coefficients for a polynomial deformation of the original AP, DV nuclear coordinates to new ones in which pair-rule stripes have been straightened, i. e. all DV dependence (e. g. for AP position along a stripe edge) removed. This transforms the data into columns (70-90), each 1% egg length (E.L.) wide. For details of the method, see Spirov et al. [2002].

We use processed (straightened) data from a strip, of width 50% DV height, centred on the AP axis. Average intensity at each AP position (each nucleus width, or 1% E.L.) is calculated, and this is plotted against AP direction, in percent egg length (% E.L.; Figs. 2, 3 and 4).

Results and discussion

We present AP profiles for three gene products: Bcd, Hb and Eve, each representative, respectively, of the maternal, gap, and pair-rule classes of segmentation genes. Errors are quantified at particular positions along the AP axis, and this is discussed in relation to two aspects of positional specification: a) error transmission down the signalling hierarchy (maternal to gap to pair-rule); and b) error dynamics during cleavage cycle 14 (early versus late time classes). Positional errors are reported here as the range in positions over which a particular intensity is crossed (i. e., in gradient reading, a particular concentration specifies a particular position: what range of positions actually occur?)

For Bcd, positional error is high, over the length of the embryo, and over cleavage cycle 14 (Fig. 2A, early; Fig. 2B, late). Houchmandzadeh et al. [2002] have reported that Bcd expression levels do not correlate with egg length (consistent with the exponential form of the Bcd pattern, having a characteristic length scale set by Bcd diffusivity and decay rates). Variability in embryo length may constitute part of the Bcd variability. Indeed, the positional ranges of 15-20% E.L. that we see in the anterior third of the Bcd profiles (Figs. 2A and B) are comparable to the approximately 15% range in egg length in Houchmandzadeh et al. In the central and posterior thirds of the embryo, however, Bcd positional error increases as the slope of the profiles flattens (see Lacalli and Harrison, 1991, for a theoretical discussion of this effect). In these regions, positional ranges are roughly 30% E.L. and more. If Bcd, as has been proposed, is a Wolpert-style gradient of positional information, we should expect to see this level of positional variability (or more, due to fluctuations in gradient-reading machinery) in the expression of downstream (gap and pair-rule) genes.

Figure 2: Anteroposterior (AP) profiles of Bicoid (maternal class gene) fluorescence intensity. A) early cleavage cycle 14, 82 embryos. B) late cleavage cycle 14, 41 embryos. Each line represents data from one embryo. The intensity at each AP position (in 1 % E.L. bins) has been averaged from 12-13 nuclei (data sampled from a strip of 50% dorsoventral height). Maximum intensities have been normalized to unity. Black arrows indicate the range around the 49% E.L. position (compare with Hb, Fig. 3). Striped arrows indicate the range around the 30% E.L. position (compare with Eve, Fig. 4).

The final patterns of the pair-rule genes (e. g. Fig. 1C), with their distinct and well-spaced boundaries between high and low concentration are crucial for specifying later cell differentiation and body segmentation. In the gap class (e. g. Fig. 1B), placement of high/low concentration boundaries has strong effects on downstream positioning of pair-rule concentration boundaries [Pankratz et al., 1990; Small and Levine, 1991; Rivera-Pomar and Jäckle, 1996]. For Hb, such a boundary exists (on average) at 49% E.L. (Figs. 1B, 3A and 3B). At this position, Bcd displays ranges of roughly (depending on outliers) 29% E.L. at early stage (Fig. 2A), and roughly 22% E.L. at late stage (Fig. 2B). (At the 49% E.L. position, Bcd appears to become slightly less variable over time.) Hb displays ranges of less than one-third of these: roughly 6% E.L. at early (Fig. 3A) and late (Fig. 3B) stages. As a positional gradient, Bcd is clearly incapable of specifying the Hb boundary with the observed precision, at either stage. Similar results were reported by Houchmandzadeh et al. [2002] for early cycle 14 embryos. The mechanism of Hb pattern formation is able to filter upstream variabilities in egg length and Bcd concentration, reducing positional error down the signalling hierarchy.

	Figure 3: AP profiles of Hunchback (gap class gene) fluorescence intensity. Lines and normalization as in Fig. 2. A) early cleavage cycle 14, 52 embryos. B) late cleavage cycle 14, 65 embryos. Black arrows indicate the range around the (average) position of 49% E.L.

Pair-rule pattern is very dynamic over cleavage cycle 14. Eve pattern begins to emerge as a single broad peak, with anterior edge (on average) at 30% E.L. (Fig. 4A). This develops into the characteristic seven peak segmentation pattern (Fig. 4B), with the anterior edge of the first peak remaining at 30% E.L. At early stage (Fig. 4A), much of the Eve pattern is laid down at precision comparable to Bcd: with ranges of 30% E.L. and greater. However, the position of the anterior edge of the first peak displays much better precision, with a range of roughly 7% E.L. At this time and position, the Bcd range is roughly 15% E.L. Later (Fig. 4B), at this position, the Eve range is roughly 5%, while the Bcd range is roughly 13%. Both patterns show slight error reduction over time, but the Eve range is always less than half that of Bcd. Hb, also upstream of Eve, is at peak levels in a large region either side of 30% E.L., and cannot provide positional specificity for the localization of Eve�s first peak. The anterior third of the embryo is also under the influence of head-specific genes [Cohen and Jürgens, 1990; Pignoni et al., 1992; Vincent et al., 1997]. It would be instructive to quantify positional errors in these genes, and relate them to both Bcd and Eve errors. However, our present Eve data again shows independence of downstream precision from upstream signal: not only in the positional precision of the anterior edge of the first peak, but also in overall Eve variability at late stage, on the order of 5-10% E.L. at any position, compared to 13% E.L. to over 40% E.L. for Bcd. Temporally, we see that Eve variability decreases over cycle 14, so that overall Eve pattern is of comparable precision to the anterior edge of the first peak, rather than being comparable to Bcd precision (as in the early stage). This increase in precision matches the importance of the stripe boundaries for segmentation at the end of cycle 14.

	Figure 4: AP profiles of Even-skipped (pair-rule class gene) fluorescence intensity. Same embryos as in Fig. 2. Lines and normalization as in Fig. 2. A) early cycle 14, 82 embryos. B) late cycle 14, 41 embryos. Striped arrows indicate the range around the (average) position of 30% E.L.

To better quantify the inherent concentration fluctuations in segmentation proteins, we are beginning to do statistics on single embryos, eliminating between-embryo variability (due to natural and experimental sources). This allows us to study, experimentally, the stochastic behaviour of a gene network, an area which has gained increasing theoretical attention in recent years [Arkin et al., 1998; Elliott et al., 1995; Ko, 1991, 1992; McAdams and Arkin, 1997, 1999; Ross et al., 1994; Zlokarnik et al., 1998; Gardner et al., 2000; Hasty et al., 2000; Thattai and van Oudenaarden, 2001]. After stripe-straightening, all nuclei (12-13, in a 50% DV slice) at a particular AP position can be considered as a sample for which intensity fluctuations can be calculated. Since embryos can be stained for up to three proteins simultaneously, we can begin to look at signal transmission in the signalling hierarchy, e. g. from Bcd to Hb to Eve.

Fig. 5 shows scatterplots from two Bcd/Hb/Eve triple-stained embryos, one early stage (Fig. 5A, time class 1) and one late stage (Fig. 5B, time class 6, same embryo as Fig. 1). Each point shows intensity at a single nucleus. As in Figs. 2, 3, and 4, the widths of the plots indicate positional error for crossing any concentration threshold. In the same embryo, we can now compare positional ranges for the three gene products of this study. At early stage (Fig. 5A), Bcd and Eve have similar ranges across the embryo, though the anterior edge of the first Eve stripe displays somewhat lower error: 2% E.L. range compared with 3% E.L. for Bcd. At the mid-embryo Hb boundary, Hb has much lower error than Bcd: 2% E.L. vs. roughly 6% E.L. At late stage (Fig. 5B), Eve displays sharp boundaries on its seven stripes. At the 30% E.L. position of the anterior edge of the first stripe, Eve has a range of 1% E.L. and Bcd a range of 3%. Posteriorly, Eve boundaries display slightly larger ranges (roughly 2% E.L.) while Bcd ranges increase dramatically (up to 20% E.L.). At late stage, the mid-embryo Hb boundary is similar to early stage: 2% E.L. range for Hb versus 5% E.L. for Bcd. Overall, we see lower errors than in the multiple embryo images (Figs. 2, 3, and 4), reflecting the loss of inter-embryo variation. However, the relative trends down the signalling hierarchy remain. The mid-embryo Hb boundary (early and late stages) and the late stage Eve boundaries are controlled at better than twice the precision of the Bcd gradient. The embryo in Fig. 5A doesn�t display as dramatic an error drop (Bcd to Eve) at the 30% E.L. position as the multiple embryo Figures (2 and 4).

Figure 5: Single embryos, stained for Bicoid (squares), Hunchback (crosses) and Even-skipped (diamonds). Each point represents fluorescence intensity at a single nucleus. Data has been taken from 50% dorsoventral slices. A) early cycle 14 (time class 1). B) late cycle 14 (time class 6, same embryo as Fig. 1). Arrows indicate positional ranges for Bicoid and Even-skipped at 30% E.L. (striped), and for Bicoid and Hunchback at 49% E.L. (black). Intensity not normalized.

Figure 6 shows the intensity fluctuations at each AP position in single embryos. Fractional errors (standard deviation divided by mean, for each 2% E.L.) are plotted against AP position. Each line represents one embryo. Same embryos as in Figs. 2, 3, and 4. Figs. 6A and 6B are for Bcd, early and late cycle 14 respectively. Figs. 6C and 6D are for Hb, early and late respectively. Figs. 6E and 6F are for Eve, early and late respectively. A basal level of 0.2 fractional error appears in Figs. 6 A-C, E, but this jumps in late stage Eve and, to a lesser extent, Hb. In Figs. 6 D and F, regions of higher fractional error appear to correspond to concentration boundaries, at which there is a large range of intensity, and regions of low average intensity.

		Figure 6: Superimposed fractional error (standard deviation divided by mean, for each 2% E.L. bin) plots. Each line represents a single embryo, same datasets as in Figs. 2, 3, and 4. A) Bicoid, early cycle 14. B) Bicoid, late cycle 14. C) Hunchback, early cycle 14. D) Hunchback, late cycle 14. E) Even-skipped, early cycle 14. F) Even-skipped, late cycle 14.

Our overall results reflect directly on the mechanism of positional specification in segmentation. Theory has shown that positional specification from the exponential Bcd gradient should worsen towards the posterior [Lacalli and Harrison, 1991]. This is indeed seen with our data (Fig. 2). In addition, computational work [Holloway and Harrison, 1999] has shown that the local reading of a gradient is itself a noisy process, and will add fluctuations to an input signal: the output will always be noisier than the input, in the absence of an error reducing mechanism. Our data do not show these predicted trends for local gradient reading: Hb and Eve errors do not monotonically worsen to the posterior; and Hb and Eve error levels are less than those seen for Bcd, especially at crucial boundary positions. We demonstrate error-reduction in the signalling hierarchy: simple Wolpert-style local gradient reading is not sufficient for this. (Error-reduction is also evident in the more than 1000 embryos in our databank with normal pair-rule pattern, in the face of 30% E.L. Bcd positional ranges.)

It has been demonstrated that Turing [1952] style reaction-diffusion models are capable of amplifying gradients while reducing input error [Gierer and Meinhardt, 1972; Holloway and Harrison, 1999]. A number of reaction-diffusion models of segmentation have been proposed [e. g. Nagorcka, 1988; Lacalli, 1990; Hunding, 1993]. Other methods of error reduction may be relevant [Kerszberg, 1996; Thattai and van Oudenaarden, 2002], and other models for segmentation exist [e. g. Meinhardt, 1988; Reinitz and Sharp, 1995]. In order to evaluate the viability of the proposed models, stochastic modelling should be done, in order to simulate the response of model mechanisms to inherent concentration fluctuations. Our data analysis is beginning to provide quantification of the fluctuations in the segmentation network, with which to test such stochastic modelling.

Data analysis is continuing in two directions: first, to extend the present basic statistics to other segmentation genes, to see if the conclusions we gain here generalize when other segmentation genes are considered, and to begin to investigate possible gap-gap, gap-pair rule and pair-rule pair-rule interactions; and second, to broaden the statistical techniques to correlation and significance over all positions (and perhaps more time classes) in the embryo.

Acknowledgement

We thank J. Reinitz, C.E. Vanario-Alonso and L.G. Harrison for stimulating discussions. This work is supported (AVS) by US NIH grant RO1-RR07801, INTAS grant 97-30950 and RFBR grant 00-04-48515. DMH thanks BCIT for travel funding.