Atlas-builder software and the eNeuro atlas: resources for developmental biology and neuroscience

To understand how embryos develop, it is essential to uniquely identify cells and cell populations. In addition, the recent advent of genetic tools for manipulating gene expression, expressing vital markers, and modifying/monitoring neuronal activity becomes more useful if the cells expressing these tools are uniquely identifiable. One approach for identifying distinct cell types in developing embryos is the generation of gene expression atlases. For example, the Eurexpress and the Allen Brain Atlas document thousands of gene expression patterns in the mouse embryo and brain, respectively (Diez-Roux et al., 2011; Lein et al., 2007). Similarly, FlyBase has a database of gene expression patterns in the developing fly embryo (Tomancak et al., 2007), and cell type-specific atlases have been generated for different stages of the Drosophila CNS (Beckervordersandforth et al., 2008; Broadus et al., 1995; Ito et al., 1995; Landgraf et al., 1997; Rickert et al., 2011; Wheeler et al., 2006, 2009). Despite the tremendous utility of these and similar databases, they typically have one or more limitations: a lack of single-cell resolution; an inability to register multiple expression patterns to determine their precise relationship; an inability to be upgraded by users; an inability to compare novel patterns with the database patterns; and a lack of a three-dimensional (3D) representation of the patterns.

Here, we develop new ‘atlas-builder’ software that overcomes each of these limitations. Our software can be used to create an atlas of any tissue in any organism with stereotyped cell positions. Newly-generated atlases are 3D, allow the precise registration of an infinite number of markers, are searchable and are open-ended because additional markers can be added. The ‘atlas-builder’ software is a publicly available resource (http://www.vaa3d.org). As a proof of principle, we generated an atlas of the late embryonic Drosophila CNS. We populated the atlas with eight transcription factor markers that allowed us to identify the major cell types in the CNS [motor neurons (MNs), glia, neurosecretory cells and interneurons]. To document the value of the atlas-builder software, we next added 75 Gal4 lines that are expressed in different, small subsets of neurons (fewer than 10 per hemisegment). This dramatically increases the number of identified interneurons in the CNS (to over 50%) and provides new tools for manipulating and monitoring neuronal activity. The resulting ‘eNeuro atlas’ is also a publicly available resource (http://www.doelab.org/eNeuro-Atlas.html).

We illustrate the utility of the atlas-builder software, using the late Drosophila ventral CNS as reference tissue. To generate an atlas of the Drosophila CNS, it was essential to have a single reference marker that labeled neurons distributed in all regions of the CNS; this ‘fiduciary’ marker can be used as a landmark to add all subsequent markers to the atlas. The Even-skipped (Eve) transcription factor is detected in the nuclei of a well-characterized subset of neurons distributed in all regions of the CNS, thus making it an excellent candidate for a fiduciary landmark (Doe et al., 1988; Frasch et al., 1987; Fujioka et al., 1999). Here, we focus on the second thoracic, third thoracic and first abdominal (T2, T3, A1) segments of the stage 16 embryonic CNS. Each of these segments has a set of unpaired midline cells plus bilaterally symmetric left and right ‘hemisegments’. We chose these three segments because they contain much of the known diversity in the ventral CNS: T2, T3 and A1 have important segment-specific differences; and A1 serves as a proxy for the closely related posterior A2-A7 segments (Schmid et al., 1999). We developed computational tools to map the position of Eve+ nuclei from multiple samples (see Fig. 1A-C). These tools generate a ‘single-marker’ atlas (supplementary material Data File S1) and a list of associated statistics (supplementary material Data File S2). Each neuron is represented by a sphere, the size and position of which represents the average size and position of all data collected on that neuron. We found some variability in neuron position between embryos. Importantly, however, the average position of an Eve+ neuron was highly reproducible with an average s.d. of up to 1.0 cell diameter in all three axes (Fig. 1D).

Next, we tested whether all CNS cell types have equally reproducible cell body positions. To identify distinct neural subtypes, we generated single-marker atlases for additional transcription factor (TF) patterns: phospho-Mothers against Dpp (pMad; all MNs), Bar (subset MNs), Dimmed (Dimm; most neurosecretory cells), Reversed polarity (Repo; glia), Apterous (Ap; subset interneurons, subset neurosecretory cells), Engrailed (En; subset interneurons, subset glia) and Hb9 (Exex – FlyBase; subset interneurons, subset MNs) (Fig. 2A-G; supplementary material Data File S3) (Benveniste et al., 1998; Bourgouin et al., 1992; Broihier and Skeath, 2002; Garces et al., 2006; Layden et al., 2006; Lundgren et al., 1995; Mccabe et al., 2003; O'Connor, 2005; Odden et al., 2002; Park et al., 2004, 2008; Patel et al., 1989; Xiong et al., 1994). Notably, we built our atlas at a time point slightly before a subset of neuropeptide (Dimm+) neurons would begin to project out of the ventral nerve cord (VNC; Allan et al., 2003; Miguel-Aliaga, 2004), thus rendering our Dimm+ neurons pMad–. We found that all neuronal markers had highly reproducible average nuclear positions (Fig. 2H). In contrast, glia often had highly variable nuclear positions (Fig. 2H), even though their cellular positions are predictable (Beckervordersandforth et al., 2008; Ito et al., 1995). We concluded that interneurons, MNs and neurosecretory cells have highly stereotyped nuclear positions. Importantly, the same steps can be used to create an atlas of any marker in any organism, as long as cell or nuclear positions are stereotyped.

A single-marker atlas, as we have generated for each of the eight TFs, is of limited usefulness. But if cell body positions are sufficiently stereotyped, we should be able to use Eve to register, or merge, all atlases into the same 3D volume. This would identify cells co-expressing two or more markers, as well as allow us to determine the percentage of cells in the CNS that express at least one marker (or ‘coverage’). We created computational tools to merge atlases, and used these tools to generate a single ‘TF atlas’ displaying all eight TFs expressed in 714 cells (Fig. 3A; supplementary material Data File S4). Importantly, the TF atlas merges overlapping cells that express two or more markers into a single cell.

To test the accuracy of the TF atlas – i.e. the number of false positives (two markers falsely predicted to co-express in a cell) or false negatives (two markers incorrectly assigned to separate cells when they are actually co-expressed) – we experimentally determined the actual overlap of all eight markers from our own stainings or from the literature (e.g. Fig. 3B,C). When we predicted co-expression solely based on proximity between cell body positions, we correctly identified 65% of all examples of in vivo co-expression (although the correct percentage rose to over 80% for neurons located in the medial CNS) (Fig. 3D; supplementary material Data File S5). When we added Rule 1, “if a pair of cells in one hemisegment is predicted to co-express a marker, then these pairs also co-express the marker in the other hemisegments”, we increased correct predictions to 96%, but this resulted in a high number of false positives (Fig. 3E). Ultimately, we settled on a set of conditions called Rule 2 that maximized true positives and negatives, while minimizing false positives and negatives (Fig. 3E; see Materials and Methods). Thus, we used the experimentally determined TF colocalization data as ‘ground truth’ to develop algorithms that allowed us to optimally determine whether two markers – stained in different embryos – are co-expressed or adjacent within the CNS.

We wanted to add a large number of new markers to generate a comprehensive eNeuro atlas. This would help us to validate our atlas-builder software and to provide new markers for previously uncharacterized neurons (mostly interneurons). We chose to add Gal4 lines expressed in 1-10 neurons per hemisegment, which gave us the dual benefit of increasing known molecular markers for interneurons and of providing functional access to these neurons for the first time. From a previous large-scale screen of 5000 Gal4 lines (Manning et al., 2012), we selected 75 lines with strong, regular expression patterns at late embryonic stages. We used each Gal4 line to express a nuclear-localized GFP, and created a single-marker atlas for each line (Fig. 4; supplementary material Data File S6). The 75 single marker atlases contained a total of 1998 cells (pre-merge) within the T2-A1 CNS, and 91% of the Gal4+ neurons had highly predictable positions (Fig. 4). We merged the Gal4 atlases with the TF atlas into a single, comprehensive ‘eNeuro’ atlas containing all 83 markers (supplementary material Data File S7). The eNeuro atlas contained fewer cells (1362) than the pre-merge numbers due to some neurons expressing multiple Gal4 lines, which were collapsed into a single neuron in the eNeuro atlas. Notably, we were able to experimentally validate 50% of the predicted co-expression in the eNeuro atlas (e.g. Fig. 5; supplementary material Data File S5). Thus, the eNeuro atlas integrated the co-expression of 83 molecular markers at single-cell resolution.

In the rodent spinal cord, over 20 different molecular subtypes of interneurons have been identified, but it is likely that interneuron diversity is much greater (Alaynick et al., 2011; Arber, 2012). Here, we used the eNeuro atlas to analyze the extent of interneuron diversity in the Drosophila embryonic ventral CNS. First, we assessed the coverage of the interneuron population. We defined interneurons as being negative for all motor neuron (pMad), neurosecretory (Dimm) and glial (Repo) markers (supplementary material Data File S8). The eNeuro atlas identified 59-60% of the interneuron population in T2-A1 depending on the method of counting total neuron numbers (Fig. 6A,B). Analysis of the eNeuro atlas shows interneurons to be extremely diverse. We describe interneuron diversity for A1, but similar results were observed in thoracic hemisegments (Fig. 7; supplementary material Table S1 and Data File S8). The interneurons in A1 can be subdivided into 80 molecularly distinct subtypes, where a subtype is defined by a unique combination of markers (Fig. 7A,C). Furthermore, 43 combinations of markers label a unique single interneuron (Fig. 7B,C). Most of the molecular subtypes found in A1 were also found in other segments (Fig. 7D). Some molecular subtypes were found only in A1, and some were unique to thoracic segments (Fig. 7D). This indicates that at different anterior-posterior positions along the nerve cord there are molecularly distinct interneuron subtypes. Thus, addition of new markers to the eNeuro atlas, which is possible when using our atlas-builder software (see Discussion), should rapidly increase the number of molecularly distinct interneurons. We conclude that interneurons are characterized by extreme molecular diversity, and predict that each interneuron will have a unique, reproducible, transcriptional profile.

Although our primary focus was on interneurons, we also used the eNeuro atlas to analyze midline cells and MNs. These populations are well characterized, which helps us to determine the accuracy of the eNeuro atlas. Midline cells contain the progeny of single-minded (sim)+ precursors, dorsal median cells and channel glia (Chiang et al., 1994; Ito et al., 1995; Kearney et al., 2004; Wheeler et al., 2006). We found that the position of midline cells in the eNeuro atlas strongly resembled the in vivo position of the midline cells (Fig. 8A,B; supplementary material Table S2 and Data File S8). Furthermore, we identified five Gal4 lines labeling subsets of midline cells that had not previously been distinguished at the molecular level (Fig. 8C; supplementary material Table S2). MNs include a set of 35 somatic MNs per hemisegment that can be molecularly subdivided based on Eve, Hb9 and Bar expression as well as visceral MNs (Landgraf and Thor, 2006; Miguel-Aliaga et al., 2008). The eNeuro atlas identified all Eve+, Hb9+ or Bar+ somatic MNs (Fig. 8D), and a group of poorly characterized Eve/Hb9/Bar triple-negative MNs (Fig. 8D). Furthermore, we found unique molecular profiles for the majority of MNs. In A1R alone we identified 23 subtypes of MNs (Fig. 8E; supplementary material Table S3 and Data File S8), whereas previously only seven molecularly distinct MN subtypes were known (Landgraf and Thor, 2006). We conclude that the eNeuro atlas (1) accurately identifies all known MNs and midline cells; (2) reveals unprecedented molecular diversity in MNs and midline cells; and (3) provides novel tools for studying cells that were previously genetically inaccessible (see Discussion).

To our knowledge, atlases generated with our ‘atlas-builder’ software are unique in allowing the integration of an unlimited number of new markers. For example, the Allen Brain Atlas has an infinitely expandable number of pages for gene expression profiles, but the expression patterns are not linked, and it is unclear which fine structures co-express different markers. In contrast, we use a single fiduciary marker to register new patterns to the existing atlas; thus, an unlimited number of markers can be co-stained with the fiduciary marker, added to the atlas and their overlap predicted with high accuracy. A second important feature of an atlas generated with our software is that users can perform a ‘BLAST’-like search to predict which known cells express a new marker. Although perfect accuracy is probably impossible due to existing biological variation, the ranking of nearby candidate cells will greatly facilitate identification of cells that express a new marker. Our tools are publicly available, and we provide detailed instructions to guide new users in creating and ‘BLASTing’ atlases for new markers (supplementary material Data File S9).

Previously, only a handful of Drosophila embryonic interneurons were individually identifiable using a combination of molecular marker expression and cell body position – e.g. pCC, EW1-EW3, Apterous+ neurons and Eve+ lateral interneurons (Doe et al., 1988; Higashijima et al., 1996; Park et al., 2004; Thomas et al., 1984; Tsuji et al., 2008). The eNeuro atlas expands the number of identified interneurons to 958, each of which has been assigned a unique name, and each of which is described by quantitative information on cell body position, size and frequency of appearance, as well as its Gal4 line expression profile (supplementary material Data File S8). Nevertheless, the eNeuro atlas is not yet ‘saturated’ for interneurons: ∼40% of interneurons do not express any marker (Fig. 5). Importantly, it is simple to add new markers to the eNeuro atlas (see below). We anticipate that additional markers will allow the unique identification of every interneuron in the ventral CNS. In the future, we will provide updated marker atlas files and add atlas files from other users with their permission. In addition, we plan to add markers that identify individual neuroblast lineages.

Until now, few genetic tools existed to manipulate specific subtypes of embryonic interneurons. Here, we have mapped 75 Gal4 lines that each label small, reproducible subsets of embryonic interneurons (Fig. 4). Some lines show the same pattern in larvae while others change pattern over time (data not shown). Importantly, the use of ‘immortalization’ genetics (Buckingham and Meilhac, 2011) can make many of these Gal4 markers useful at later stages of development or in the adult. Furthermore, many interneurons in the eNeuro atlas were labeled by multiple Gal4 lines (Fig. 6A,C; supplementary material Fig. S1), thereby raising the possibility of using intersectional genetics to target single interneurons, because the vast majority of these lines were made with a defined cis-regulatory DNA fragment (Pfeiffer et al., 2010). Thus, this work provides a significant advance in our ability to characterize and test the function of individual identified interneurons.

We also provide important new tools for the study of MNs and midline cells (Figs 7,8; supplementary material Tables S2,S3). These can be used to trace the axon projection of a previously uncharacterized Eve/Hb9/Bar triple-negative motor neuron (supplementary material Fig. S2). In the future they could be used to study the specification, axon pathfinding, target recognition and role in motor behavior. We also made a novel discovery regarding the well-studied Hb9+ MNs: although most Hb9+ MNs project out of the intersegmental nerve (ISN), some have been reported to project out of the segmental nerve root (SN) to transverse somatic muscles, similar to the projection pattern of the Bar+ MNs (Broihier and Skeath, 2002; Garces et al., 2006). Our atlas did not contain any Hb9+Bar+ co-expressing neurons, which was confirmed by experiments showing no Hb9+ neurons projecting out of the SN to transverse muscles (supplementary material Fig. S3). We conclude that the Hb9+ MNs do not co-express Bar or project out of the SN, at least not at embryonic stage 16. Thus, the eNeuro atlas provides tools for the developmental and functional analysis of individual and poorly characterized groups of MNs.

Limitations of the atlas approach include potential annotation errors that exaggerate positional variability, and stochasticity in Gal4 expression that results in some cells being unrepresented or unilaterally represented. It is also possible that the eNeuro atlas contains errors in cell merging. Errors are likely to be rare, however, because we examined 171 of 243 predicted Gal4/TF co-expressing cells and found only one error, which was manually corrected (supplementary material Data File S5); the remaining Gal4/Gal4 merge predictions are likely to have a similarly high accuracy.

The genetic analysis of neural development in Drosophila has led to many fundamental discoveries that also apply to mammalian neural development. The study of interneuron specification and function is in its infancy in both Drosophila and mammals. We suggest that this atlas of interneuron diversity and tools for functional analysis will facilitate the study of Drosophila interneurons, and that this will have a direct impact on the study of mammalian interneurons. This is an important area of research, as interneuron dysfunction is thought to underlie autism, epilepsy and other debilitating human neural disorders.

We used standard methods for propagating fly stocks. Transcription factor patterns were determined in wild-type flies (yellow white). Gal4 patterns were determined by crossing Gal4 [BN-gal4 (Sato et al., 1999), Ap-Gal4 (Bloomington #3041); 9-58-Gal4 (Ulrike Heberlein, Janelia Farm Research Campus, JFRC); a227-Gal4 (Julie Simpson, JFRC); Hb9-Gal4 (Broihier and Skeath, 2002); and all other Gal4s (JRFC) (Pfeiffer et al., 2008)] males to females containing UAS-GFP with a nuclear localization sequence or membrane tether (Bloomington #6452, #32184, #32197).

We used standard methods to stain Drosophila embryos (Manning et al., 2012). Primary antibodies were mouse Eve 2B8 (1:50), mouse En 4D9 (1:10), and rat Elav (1:10), all from the Developmental Studies Hybridoma Bank (DSHB) developed under the National Institute of Child Health and Human Development (NICHD) and maintained by the University of Iowa; guinea pig Dimm (1:100; a gift from S. Thor, Linkoping University, Sweden); rat Repo (Campbell et al., 1994) (1:1000); rabbit pMad (pS1 1:300; Peter Ten Dijke, Leiden University Medical Center, Leiden, The Netherlands); chicken GFP (1:1000; Aves, Tigard, OR); rabbit Hb9 (Odden et al., 2002) (1:1000); and guinea pig Fox (Nechipurenko and Broihier, 2012) (1:20). Secondary antibodies were from Invitrogen/Molecular Probes and were used according to the manufacturer's instructions. Stage 16 embryos were selected for imaging based on morphological criteria (Pereanu et al., 2007).

Images were acquired on a BioRad Radiance, Zeiss 700 or 710 confocal microscope with a 63× objective collected at 0.3 or 0.5 μm steps. Images were cropped in ImageJ (NIH) to include segments T2-A1. Some images were used to generate atlases, whereas others were used to validate co-expression or for figures.

We developed a Vaa3D (Peng et al., 2010) (http://Vaa3d.org/) plug-in to segment cells. First, a TIFF image was smoothed with a median filter and/or Gaussian filter to reduce noise. We then separated the image into background and foreground pixels to make the later segmentation step more robust and computationally efficient. We provided three thresholding methods: adaptive thresholding, global 3D thresholding and global 2D thresholding. After thresholding, we used a watershed algorithm to segment cells based on either the intensity information or the shape information of foreground pixels. For each transcription factor we segmented and annotated a median of 10 image stacks. We segmented a median of five image stacks per Gal4 line.

The segmented cells were annotated using VANO (Peng et al., 2009) software (http://vano.cellexplorer.org/). Where possible, neuron names were taken from the literature and had the following syntax: ‘name/marker hemisegment’ (e.g. pCC/eve T2R or A/19H09 T2R). Segment identity was based on neuron position at mid stage 16; the U1-U5 neurons defined the posterior boundary of each segment. Ruby scripts were used to detect annotation errors, such as typos (code is available upon request). To minimize annotation, we used only Eve+ medial neurons (aCC, pCC, RP2, U1-U5) in each segment, as this was sufficient for optimal registration (data not shown). For new markers, annotation could be carried out on just one cell of interest or on all cells.

A newly developed Vaa3D plug-in called ‘AtlasBuilder’ uses rigid registration to map the annotated image stacks into a standard coordinate space defined by a target stack. The target stack was automatically determined from a randomly selected subset of stacks as the one that minimized the summed cell position variations (supplementary material Data File S1). We computed an affine transformation that best aligned the EVE channel of a subject stack to that of the target stack. This transformation was then applied to all channels of an image stack. Note that this type of rigid registration does not use local warping to align the neurons. It calculates an average position for each nucleus, which acts as unique x,y,z coordinates in virtual space; it also calculates the s.d. of the variance of actual centers from the mean, which acts as a measure of the stereotypy of each cell. For a complete list of data produced by registration, see the ‘cellStatistics.xls’ file for each atlas (supplementary material Data Files S2, S3 and S5). ‘AtlasBuilder’ can build atlases with different confidence thresholds, which are calculated based on how often a given cell appears in image stacks. For transcription factor atlases, we included only cells that were found in 33% or more of the image stacks; for Gal4 atlases, we included all cells.

To establish in vivo co-expression between markers, we used data from previously published reports, or performed co-labeling experiments ourselves. We determined the individual identity of each cell based on the relative position of the cell within the marker pattern, as we did for manual annotation. We also scored for projection patterns/cell morphology and for nuclear shape where these characteristics had been reported in the literature (supplementary material Data File S5).

To predict overlap between single marker atlases, we developed a ‘Detect coexpressing cells’ function in AtlasBuilder. This method takes ‘.apo’ files (single atlases and/or atlases) and determines which cells are within a user-specified distance (we used six pixels, ∼ average nuclear radius). We did not use symmetry or bilateral rules, although they are available (supplementary material Data File S9 for details). This plug-in generates two output files: a ‘. coexpress.apo’ where predicted co-expression is white, and a ‘.coexpress.txt’ file, which lists each candidate cell followed by all cells found within the specified radius.

To calculate prediction accuracy we calculated three parameters: ‘true positives’ – co-expression found both by atlas prediction and in vivo analysis; ‘false negatives’ – co-expression found only by in vivo analysis; ‘false positive’ – co-expression found only by atlas prediction. True positives and false negatives were scored as a percentage of all expected (n=133) in vivo co-expressions (supplementary material Data File S5). False positives were scored as a percentage of total predictions.

We used the following three rules.

Rule 1: if a cell in one hemisegment is predicted to co-express, then all examples of that cell also co-express; 96% true positive rate, 4% false negative rate and 51% false positive rate.

Rule 2: if the same cell in two hemisegments is predicted to co-express, then all examples of that cell also co-express; 84% true positives, 16% false negatives and 30% false positives.

Rule 3: if the same cell in three hemisegments is predicted to co-express, then all examples of that cell also co-express; 40% true positives, 60% false negatives and 6% false positives.

The ‘Merge coexpressing cells’ plug-in for Vaa3d ‘AtlasBuilder’ takes two types of input: ‘.apo’ files and a user-generated look-up table text file (supplementary material Data File S10). The merging plug-in produces a new atlas with both ‘.apo’ and ‘cellstatistics.txt’ files containing positional information for the new cells; these new positions are calculated from the average positions of all original cells, even if they come from two separate ‘.apo’ files. In the resulting atlas, all markers expressing in a given cell are listed in the comments column of the ‘.apo’ files.

To create the TF atlas, we assigned each cell a new name, which was a truncated version of the original names (i.e. aCC/eve T2R and aCC/pMad T2R became aCC T2R).

To create the eNeuro atlas, we discarded any low-quality cells: cells appearing in less than 33% of the annotated Gal4 line image stacks, or cells with an average s.d. of the nuclear position that was greater than a single cell diameter. We used the ‘Detect coexpressing cells’ function to generate a list of predicted co-expressions among all 83 single atlases. We used custom Ruby scripts to apply ‘Rule 2’ (see above). We then used Ruby scripts and MATLAB to view the matrix of co-expression as biograph objects (code is available upon request). We merged cells in the following order: MPs (dMP2, vMP2, MP1), Eve+ cells, Bar+ cells, Hb9+ cells, pMad+ cells, cells predicted to label with other transcription factors, all other cells. Each subtype was assigned a name; if the cell appeared in the TF atlas, we used the same name; all others (646 cells) were assigned names using ‘IN’ as a prefix followed by a three-digit number, such as ‘IN094’ (see supplementary material Data File S10 for complete list of merged cells).

Cell diameter was calculated using the mean volume of all cells and then solved for r (volume=4/3πr³). The s.d. of cell position was calculated as x,y,z s.d. (std_x, std_y, std_z)/3 in Fig. 1. In Figs 2 and 4, average s.d. was calculated similarly, but we combined data from cell bodies in all hemisegments (e.g. aCC represents data from aCC/eve T2R, T2L, T3R, T3L, A1R and A1L).

We used two methods to estimate the number of interneurons in each segment: 540 interneurons, based on single interneuron DiI labeling (Campbell et al., 1994), which does not include unpaired midline neurons; or an average of 549 interneurons per segment (515 in T2, 542 in T3, 589 in A1), based on our own counts of Elav+ neurons minus pMad+ MNs and Dimm+ neurosecretory cells. To determine the interneuron coverage in the eNeuro atlas, we defined interneurons as triple negative for pMad (MNs), Dimm (neurosecretory cells) and Repo (medial glia), as well as having low positional variability (thus excluding lateral glia). This gave us an average of 324 interneurons per segment (Fig. 6; 337 in T2, 339 in T3, 295 in A1).

We thank Matt Clark, Sen-Lin Lai and Chris Wreden for comments on the manuscript; Maria Purice, Art Li, Andrew Reichold and Jimmy Kelly for help annotating marker patterns; and the Drosophila Stock Center (Bloomington, IN, USA) and Todd Laverty at the JFRC (VA, USA) for fly stocks.