Sexual Dimorphism in the Fly Brain

Summary Background Sex-specific behavior may originate from differences in brain structure or function. In Drosophila, the action of the male-specific isoform of fruitless in about 2000 neurons appears to be necessary and sufficient for many aspects of male courtship behavior. Initial work found limited evidence for anatomical dimorphism in these fru+ neurons. Subsequently, three discrete anatomical differences in central brain fru+ neurons have been reported, but the global organization of sex differences in wiring is unclear. Results A global search for structural differences in the Drosophila brain identified large volumetric differences between males and females, mostly in higher brain centers. In parallel, saturating clonal analysis of fru+ neurons using mosaic analysis with a repressible cell marker identified 62 neuroblast lineages that generate fru+ neurons in the brain. Coregistering images from male and female brains identified 19 new dimorphisms in males; these are highly concentrated in male-enlarged higher brain centers. Seven dimorphic lineages also had female-specific arbors. In addition, at least 5 of 51 fru+ lineages in the nerve cord are dimorphic. We use these data to predict >700 potential sites of dimorphic neural connectivity. These are particularly enriched in third-order olfactory neurons of the lateral horn, where we provide strong evidence for dimorphic anatomical connections by labeling partner neurons in different colors in the same brain. Conclusion Our analysis reveals substantial differences in wiring and gross anatomy between male and female fly brains. Reciprocal connection differences in the lateral horn offer a plausible explanation for opposing responses to sex pheromones in male and female flies.

Sex specific regions could also be observed in the ventral nerve cord. The largest FER covered a superficial layer of the abdominal ganglion that may relate to female specialisations for oviposition. A second region runs along the most lateral part of the wing neuropil of the mesothoracic ganglion. The male-enlarged region had more complex morphology restricted to the interior of the neuropil, located primarily in the leg neuropil of the prothoracic ganglion and the wing neuropil of the mesothoracic ganglion. This may relate to the sensory role of the forelegs during the tapping stage of courtship and wing circuits regulating courtship song [1][2][3].

Overlap of fru+ Neurons and MER
In males, the highest intensity fru+ processes occupy 6.5% of the neuropil while the Male Enlarged Regions occupy 4.5% with 1.7% of neuropil voxels present in both groups; random overlap would predict only 0.29% of neuropil voxels would be double positive. There is a highly significant association between these two groups of voxels (Fisher's exact test, odds ratio 11.4, p<2.2e-16). We calculated the overlap between voxels containing at least two male-specific arbours (6.7% of the neuropil) and the MER (4.5% of the neuropil), finding that double positive voxels occupied 2.7% of the neuropil (odds ratio 19.9, p<2.2e-16, Fisher's exact test). Similarly in females, the overlap between voxels containing at least two fru+ female-specific arbours (0.8% of the neuropil) and the female-enlarged region (3.5%) identifies double positive voxels occupying 0.12% of the neuropil. Although a relatively small overlap, this is still significantly above chance (odds ratio 5.1, p<2.2e-16, Fisher's exact test). about the projection patterns or developmental origin of the neurons in question. We have coordinated a new scheme with [5] where the fly's brain is divided into the deutocerebrum (DT), the anterior and posterior superior protocerebrum (aSP and pSP respectively), the posterior medial protocerebrum (pMP), the anterior and posterior inferior protocerebrum (aIP and pIP respectively), the anterior and posterior suboesophageal ganglia (aSG and pSG respectively), and the mushroom body (MB) (see Figure 1A). Clones are placed in a region depending on the location of the cell bodies and primary arborisation, the great commissure is used as the dividing line between the two (adapted from [6]). Where this is ambiguous the area of primary innervation is used to select a position along the A-P axis. After all clones had been identified, each clone received a unique lower case letter appended to the name stem; the first clone was the one closest to the midline in the dorsal-most part of the region. Starting with the first clone the next clone clockwise when looking at frontal sections of the brain was given the next letter and so on. See Table S2 on page 39 for correspondences between names of fru clusters from different studies.
Naming in the VNC was likewise changed from previous schemes; cell body position dictated the position along dorsal-ventral axis and the region of primary innervation dictated the stem of the name (i.e. pro-, meso-, or metathoracic, Pr, Ms, and Mt respectively). The first clone of a region was the one with cell bodies closest to the midline and most anterior, the next was the one clockwise when looking at a VNC with the ventral side facing the observer and anterior 'up'. The names for the fru+ regions were suggested in our companion manuscript [5] and adapted in this study based on correspondences with the male-enlarged neuropil regions.

Relationship of This Study to Other Studies of fru+ Neurons
Two previous studies applied MARCM to NP21, a GAL4 enhancer trap inserted at the fru locus [7,8]. They described 38 NP21 clones in the central brain including one dimorphic and two male-specific clones. 26 of those 38 clones (353 neurons) have a homologue in our dataset; we cannot match the remaining 12 clones (45 neurons). See Table S2 on page 39 for detailed crossreferencing of our data against these earlier studies. In comparison with those studies, we describe the anatomy of 85 additional fru+ clones across the brain and VNC. All clones are registered in the same reference space, allowing us to compare male and female anatomy for all brain and most VNC clones. We identify 24 new dimorphisms. We also describe the complete arborisations of 13 clones that connect the brain and VNC. Finally all of our 3D data are freely available. Table S2 also details the correspondence between our 62 central brain neuroblast MARCM clones and the 60 clusters of [5] derived from an intersectional scheme using a fru Flp allele. Most large groups of neurons could be clearly identified in the two datasets, but a significant number of our clones remained without a corresponding identification. Some fru Flp clusters corresponded to as many as 3 MARCM clones (aDT3 vs aDT-a, aDT-e, aDT-f). Conversely clone pIP-c corresponds to cluster pIP3 and pIP7; these are two groups of neurons with distinct morphologies that are generated by the same neuroblast. Detailed analysis of the raw data using computerised pattern recognition may be the best means of finding all correspondences (Nicolas Masse, SC, AO, GSXEJ, unpublished observations). We expect that there will be some neurons present in only one data set. The fru Flp approach is limited by the availability of Gal4 lines to generate intersectional patterns that are sufficiently simple to visualise and distinguish every neuronal class. The fru GAL4 MARCM approach is more comprehensive in principle because it is possible to label all fru+ neurons in small reproducible groups (the neuroblast clones). However we chose to focus on postembryonic neurons by heat shocking experimental animals at larval hatching. Embryonic neuroblast clones were generated only sporadically in our data set, meaning that we likely missed some fru+ neurons that are born during embryonic development. For the 62 lineages in our central brain dataset we found a minimum of 3 examples of every clone (28 on average). Although there are differences in the likelihood that different neuroblasts generate MARCM clones, we expect that we have identified all postembryonic fru+ lineages. In the VNC there were multiple cases where we obtained only one or two examples of a given clone, so we cannot presume to have identified all possible lineages.   Figure 2) (A) Non-dimorphic clones expressing fru Gal4 in the central brain and suboesophageal ganglion. One example of each clone is show in a standard deviation z-projection overlaid on a similarly projected intersex template brain stained with nc82. Clones were extracted from raw stacks using a mask and the contrast has been increased in order to facilitate viewing of the fine processes of each clone. All clones shown are male. (B) Non-dimorphic clones expressing fru Gal4 in the VNC. Images were processed as in (A). (C) Comprehensive description of sexually dimorphic neuroblast clones in the central brain. One male (green) and one female (magenta) clone have been overlaid to demonstrate sexually dimorphic projections. For each clone the male and female image were extracted using an identical binary mask, a background subtraction and contrast adjustment (see methods) were used to make them comparable. White indicates strong overlap. Green and magenta arrowheads indicate male and female specific process respectively. Arrows indicate sex-specific arbours which are not evident in the image.    Figure 4) (A) One example of each clone is show in a standard deviation z-projection overlaid on a similarly projected intersex template brain or VNC stained with nc82. Clones were extracted from raw stacks using a mask and the contrast has been increased in order to facilitate viewing of the fine processes of each clone. An extractable VNC clone of vPr-a_F was not obtained so the central brain with the ascending process is shown in isolation.
(B and C) Example of sexually dimorphic overlap between ascending VNC clone dMs-b and brain clone aDT-b. In males the contralateral ascending projection of dMs-b has substantial overlap with aDT-b in the SOG while in females the ascending process largely misses the sexspecific arbour of aDT-b. Views are as in Figure 1B.   Table 1) Clonal analysis of fru+ neurons in the VNC. Names and location of fru+ clones in the VNC.

Supplemental Data on the Web
Upon publication of the manuscript, the following supplementary data will be made available. This will be at http://flybrain.stanford.edu unless another location is indicated. • We will also copy all raw and processed image data for our study onto a 1.5TB external hard drive on request (LSM confocal files, registration files, registered images). These raw data are freely available for study by our colleagues. We would however request that 1) any future studies that analyse our data in new ways, freely release the derived data and 2) any future studies that combine our data with other new data should release those new data on a similar basis upon publication.

Image Registration
We used nc82 neuropil staining from multichannel confocal stacks as the input to a fully automatic intensity-based (landmark free) 3D image registration software [9,10]. An initial linear registration with 9 degrees of freedom (translation, rotation and scaling of each axis) was followed by a non-rigid registration that allows different brain regions to move somewhat independently, subject to a smoothness penalty [11]. The underlying deformation model is based on third order B splines; for the central brain the final grid of B spline control points was 51 x 51 x 35, for a total of 91035 control points and therefore 273105 degrees of freedom. The algorithm typically held about two thirds of the control points fixed because they contained no useful information ˜ usually locations outside the brain. The final control point grid had a mean spacing of 5 to 6 µm. The registration for the nc82 channel could then be applied to other channels containing labelled neurons.
We used an updated version of the core registration software which we have made available at http://flybrain.stanford.edu along with a control script used to coordinate multiple registrations.
In the later stages of the project, the core registration toolkit (CMTK) was made open source and is now available at http://www.nitrc.org/projects/cmtk/. Registrations were run either on an Apple Mac Pro (8 core) or on a Sun Gridengine 6.2 Cluster, consisting of 33 IBM 8 core Xeon blades, running Centos 5.2 Linux; registration jobs were run on 4 cores and took 1-4h to complete. The only pre-processing of raw confocal images before registration was rotation to match the orientation of the template to the nearest 90 degrees, followed by export in Biorad PIC format using a plugin that we have contributed to the Fiji distribution of ImageJ (http://pacific.mpi-cbg.de). We have also bundled plugins to read and write registration output in the NRRD image format (http://teem.sourceforge.net/nrrd/format.html) with Fiji.

Deformation-Based Morphometry
Deformation-Based Morphometry (DBM) is a sensitive and global approach to identifying structural differences. The basic approach is to look for systematic differences in the amount of expansion or contraction required to map each point in a sample brain onto the template. The relative volume change is measured by the Jacobian determinant; this has a value of 1 when there is no change in volume, with values above and below 1 indicating expansion and shrinkage, respectively. Before analysis the raw Jacobian determinants were first normalised to remove the effect of overall brain size by dividing by the mean Jacobian determinant for the whole brain [10].
We then carried out two types of analysis. The first was voxel-wise, generating a t-statistic parametric map for the difference in volume between two groups. For each voxel in the template brain, we calculated the normalised Jacobian determinants for the registration of each sample brain, log transformed those determinants [12], and then carried out a t-test. For example, if there were two groups of 40 sample brains each voxel in the resulting parametric map would contain a t-statistic resulting from an unpaired test with 78 degrees of freedom. The measurements were carried out on a grid of 2 x 2 x 2 µm voxels, but only inside the neuropil, a total of 606729 voxels. We used permutation testing [13] to estimate a null distribution for the t-statistic. To achieve a one tailed alpha of 0.05 for female and male enlarged regions across the whole neuropil we calculated the 5% and 95% quantiles, respectively, for the extreme values of 5000 permutations. This gave t values of -4.96 and 5.00, respectively. For simplicity we rounded these to t=±5. For comparison, the extreme t values for all 5000 permutations were -7.32 and 6.48.
We also carried out volumetric measurements by summing the normalised Jacobian determinants over a region of interest and then multiplying by the voxel size; this is equivalent to defining a region of interest (ROI) in template space, reverse transforming that ROI into the space of each sample brain and then measuring the volume of the transformed ROI in that sample brain.
We noticed two practical issues in the statistical analysis. The first was that voxel-wise DBM across the whole brain appears sufficiently sensitive to pick up structural variations that are likely to be due to genetic background. The second issue was that when carrying out DBM-based volumetric analysis, there were significant batch effects (in spite of the normalisation by total brain volume). This could be due to differences in antibody penetration or brain shrinkage during immunochemistry. Such batch differences would tend to reduce the effective number of degrees of freedom for any statistical test leading to false positives, so we therefore modelled them explicitly with a mixed effects model using the nlme package of R [14,15].

Masking of Registered Clones
In order to analyse individual clones from brains with high levels of superficial background fluorescence or additional non-overlapping clones, masks were generated with Fiji's Segmentation Editor plugin. Three registered female image stacks containing the same clone were selected and merged to a single RGB image. A paint brush tool was used to select regions that contained shared processes and these shared regions were exported as a binary mask. This image file was overlaid with a male stack having the corresponding clone and evaluated to determine if the mask needed expansion for male clones. If the mask needed expansion, three registered male brains expressing the clone of interest were then overlaid, opened in the Segmentation Editor and the 8-bit female mask was added to the plugin's label field using the Image Calculator in Fiji; note that the mask was never made smaller, as a result, the male mask functioned as a logical OR of the female and male patterns of the clone. This new mask was exported, saved and compared to a male stack to see if the mask covered all of the processes in a registered image. As MARCM clones are unilateral, all masks were made with the assumption that the cell bodies of the clone would be located on the left side of the fly's brain (the image's right side).
To extract clones from raw images, the mask file was multiplied with a registered stack containing the clone of interest using the Image Calculator in Fiji or by passing the mask file to the CMTK reformatx tool. If the clone of interest was on the fly's right, the image was first horizontally flipped and then further transformed using a bridging registration of the template brain to its horizontally flipped self. The end result was a registered image stack containing the clone of interest without any adjustments to voxels inside the masked region, but with all voxels outside the masked region set to zero.

Image Processing and Visualisation
In order to prepare images for further analysis or display, two additional processing steps were carried out. The anti-GFP/CD8 antibodies resulted in an appreciable level of background staining within the neuropil. We fitted a Gaussian distribution to these background voxel intensities and subtracted a threshold value of + 2 from all voxels. We then mapped the remaining intensities onto an 8 bit range from 0 to 255 using the unu quantize tool (http://teem.sourceforge.net/unrrdu). For overlap analysis, we downsampled images fivefold in xy and twofold in z, first applying a Gaussian convolution with = 1.0 to eliminate spatial frequencies above the new voxel size (2.06 x 2.06 x 2.14 µm).
For visualisation in Amira 5.2.2 (www.amiravis.com) images were cropped to the smallest cube containing all non-zero voxels using the CMTK convert tool; this reduced the image size by an average of >80% easing visualisation. Some visualisations depended on label fields to segment the arbours of a particular neuroblast clone into specific zones. Label fields were constructed in Amira using the smoothed images mentioned above and a combination of thresholding and the lasso tool in the 3D viewer. Label fields were then used in combination with the Amira Volren module for visualisation or used to multiply the original image data for overlap analysis. The same procedure was used to isolate just the dimorphic arbours from the 21 dimorphic clones in the central brain. The volume of the dimorphic arbours defined by these label fields is a reasonable approximation of the volume of neuropil that could be sampled by the processes of each clone.
Confocal images of twin-spot MARCM brains were deconvolved with Huygens Essential (Scientific Volume Imaging, Hilversum, The Netherlands) using the classic MLE algorithm, a theoretical point spread function that corrects for spherical aberration due to sample depth and assuming a signal to noise ratio of 20. The images were then opened in Fiji and the brightness and contrast were adjusted and regions of interested were selected and expanded for display purposes. The blue channel in Figure 6C-F was calculated as follows: First, green and magenta channels were manually thresholded. Second, voxels were set to blue if they were positive in both channels or if they were positive in one channel and a voxel within a cube of edge 3 centred in the voxel was positive in the other channel (i.e. contiguous pixels).
For display purposes, the full fru Gal4 expression pattern was averaged between two male (m) and two female (n) brains. Two brains of each sex were registered to our intersex template and averaged using Fiji. In addition, two masks have been applied to the averaged brains, one to remove the cell bodies on the surface of the brain and a second to remove the mushroom bodies. The RGB Rainbow lookup table from Fiji was used to enhance contrast.

Cluster and Overlap Analysis
Cluster analysis in Figure 5 was carried out using the R function hclust using Ward's method which aims to find compact, spherical clusters. The input was a distance matrix calculated using the dist function from Pearson cross-correlation scores of all raw LHN images cropped to a cube surrounding the lateral horn.
Overlap analysis of olfactory clones in Figure 6 was carried out using data that retained intensity information rather than using binary masks. Quantile based normalisation was first applied to all of the images of a particular clone and sex using the normalize.quantiles function of R package affy. This compensated for arbitrary differences in the intensity of different images of the same clone due to differences in staining intensity or confocal microscope acquisition parameters. We then calculated the dot product of each combination of PN input clone and LHN output clone.