We describe here a collection of 3060 split-GAL4 lines targeting cell types across the adult Drosophila CNS. All lines were created in collaborations with the FlyLight Project Team at Janelia Research Campus from 2013 to 2023, based on examination of over 77,000 split combinations. 1644 lines were published previously and are drawn together here as part of the larger collection. Table 1 summarizes prior publications. The remaining 1416 lines in the collection are described in other in preparation manuscripts or newly reported here as shown in Figure 1—source data 1, which specifies proper citations and other line-specific information.

To confirm and consistently document the expression patterns of included lines, all were rescreened by crossing to a UAS-CsChrimson-mVenus reporter inserted in at a specific genomic location (attP18) (Figure 1; Klapoetke et al., 2014). At least one male and one female CNS were dissected, antibody labeled, and imaged per line. Expression patterns were validated by manual qualitative comparison to prior data, where available, and scored for specificity and consistency (see below and Methods). Male/female differences were observed in 109 of these 3060 lines, confirming previously reported sexual dimorphisms or suggesting potential areas for further study (Figure 1F; Figure 1—source data 1).

Previously published and newer lines were evaluated as a whole, taking advantage of new screening data to highlight existing lines or identify new lines that best label cell types. The lines were selected for inclusion based on several factors:

We examined the distribution of the cell-type collection across the male and female CNS. To visualize the distribution, we aligned each image to a unisex template and segmented neuron patterns from rescreening images into a heat map (Figure 3). The cell-type lines show neuronal expression across most of the CNS, with 96% of voxels having expression from at least 5 lines, 90% with 10 or more lines, and 66% with 20 or more lines (Figure 3—source data 1). Prominently labeled areas include the fan-shaped body, lobula, and superior medial protocerebrum in the brain, and the T1 medial ventral association center and intermediate tectulum in the VNC. The antennal lobe is very rarely labeled. Other relatively rarely labeled areas include the anterior ventrolateral protocerebrum and prow in the brain and the ovoid (accessory mesothoracic neuropil) in the VNC.

We compared the line distribution between female and male images (Figure 3C, D). We observed more female split-GAL4 expression in the epaulette and abdominal ganglion, and more male expression in the antennal mechanosensory and motor center and several regions of the mushroom body (Figure 3—source data 1). Two example sexually dimorphic split-GAL4 lines from the cell-type collection illustrate regions of increased split-GAL4 expression in females (Figure 3E–H). The lines were identified based on NeuronBridge searches of the ‘female minus male’ image (Clements et al., 2021; Clements et al., 2024). Examination of individual male images suggests that (in contrast to the female) the subtraction approach highlighted male artifacts of the reporter system: either background incompletely segmented away from neurons (e.g. in the metathoracic ganglion) or repetitive neuron signal unrelated to other line-specific expression (‘sparse-T’ neuron and projections into the antennal mechanosensory and motor center; Meissner et al., 2023).

The collection of lines described above has been deposited with Bloomington Drosophila Stock Center (Bloomington, IN) for availability until at least August 31, 2026. Images and line metadata are available at https://splitgal4.janelia.org. Anatomical searching for comparison to other light microscopy (LM) and EM data is available at https://neuronbridge.janelia.org.

In addition to the line collection described above, we describe a split-GAL4 image data resource. It consists of all good quality FlyLight split-GAL4 images and associated metadata generated in the project’s first 9 years of split-GAL4 characterization (Figure 1—source data 3). It includes additional data for the lines described above, together with data for many additional split-Gal4 combinations. Many of these represent lower quality lines that label multiple nearby cell types (Figure 2D and see below), but the image collection also includes additional high-quality lines that were not chosen for stabilization or are currently not maintained as stable lines (e.g. because of similarity to other lines in the collection).

FlyLight’s analysis workflow consisted of multiple related image-generating pipelines (Figure 4A, B). Typically, we screened an ‘Initial Split’ (IS) cross temporarily combining the two candidate split hemidrivers and reporter before building (stabilizing) a genetically stable ‘Stable Split’ (SS) line. IS and SS data with the same five to six digit code reflect the same combination of split hemidrivers. Some split-GAL4 lines use region-specific nomenclatures, with names prefixed by ‘MB’ (mushroom body), ‘OL’ (optic lobe), or ‘LH’ (lateral horn), and in most cases label neurons within or nearby those regions. We examined about 77,000 IS crosses and after quick visual evaluation chose to image a fly CNS sample from about half of them. 36,033 such IS samples (individual flies) are included in this image collection (Table 2). About 10% of IS crosses were made into SS lines and imaged again, with 8273 such lines included.

Further documentation of the full SS pattern at higher quality with varying levels of pre-synaptic labeling was generated by the Polarity pipeline (Aso et al., 2014; Sterne et al., 2021), followed by full 20× imaging and selected regions of interest imaging at 63×, with 37,409 such samples from 7039 lines included. Single-neuron stochastic labeling by MultiColor FlpOut (MCFO; Nern et al., 2015) revealed single-neuron morphology and any diversity latent within the full SS pattern, with 54,807 such samples from 7679 lines included in the image collection.

In total the raw image collection consists of 46,653 IS/SS combinations, 129,665 samples (flies), 612,124 3D image stacks, and 4,371,364 secondary (processed) image outputs, together 192 TB in size. The IS data may be particularly valuable for work on understudied CNS regions, as it contains biological intersectional results (as opposed to computational predictions) that may have specific expression outside regions focused on in prior publications (Table 1).

Due to the size of the raw image collection, it has not undergone the same level of validation as our other image collections, and caution is recommended in interpreting the data therein. We believe it is nonetheless of overall good quality. The images are available at https://flylight-raw.janelia.org and are gradually being made searchable via NeuronBridge due to the size of the dataset. At the time of this version of record, the 36,000 IS images have been added, with the remaining data planned to be added by the end of 2024. For a further guide to interpreting the image data, see Supplementary file 2.

We include 982 split-GAL4 lines, 31 Generation 1 GAL4 lines, and 353 Generation 1 LexA lines selected for specific expression in the third-instar larval CNS (see Figure 1—source data 1). Seven more split-GAL4 lines were selected based on both larval and adult expression.

We selected larval lines based on the following criteria: (1) the projection pattern had to be sparse enough for individual neurons in a line to be identifiable; (2) the line had to be useful for imaging of neural activity or patch-clamp recording; and (3) the line had to be useful for establishing necessity or sufficiency of those neurons in behavioral paradigms. We thus selected lines with no more than 2 neurons per brain hemisphere or per subesophageal zone (SEZ) or VNC hemisegment. When there are more neurons in an expression pattern, overlap in projections makes it hard to uniquely identify these neurons. Also, interpreting neural manipulation results in behavioral experiments becomes much more challenging.

If lines had sparse expression in the brain or SEZ but also had expression in the VNC (sparse or not), we nevertheless considered them useful for studying the brain or SEZ neurons because (1) the brain neurons could be identified; (2) the line could be used for imaging or patch-clamp recording in the brain; and (3) the split-GAL4 line could be cleaned up for behavioral studies, using teashirt-killer-zipper to eliminate VNC expression (Dolan et al., 2017). Similarly, lines with specific VNC expression but with undesired additional expression in the brain or SEZ were still useful for imaging and behavioral studies in the VNC, because a loss of phenotype following teashirt-killer-zipper could establish causal involvement of VNC neurons. For LexA lines, we included a few with 3 or more neurons per hemisphere or per hemisegment because there are fewer LexA lines overall and they could still be used for functional connectivity studies.

Images of all of the selected lines are available at Virtual Fly Brain (VFB; https://raw.larval.flylight.virtualflybrain.org/) and will be further integrated into the main VFB website (https://virtualflybrain.org/). We anticipate these lines will be useful tools in conjunction with EM volumes of the larval nervous system.

Included split-GAL4 driver stocks and references are listed in Figure 1—source data 1. See Table 2 for counts of driver stocks available from Bloomington Drosophila Stock Center.

Rescreening of the cell-type collection made use of a 20XUAS-CsChrimson-mVenus in attP18 reporter (Klapoetke et al., 2014). Additional reporters are listed in the Key Resources Table and Figure 1—source data 3.

Flies were raised on standard cornmeal molasses food typically between 21 and 25°C and 50% humidity. Crosses with hs-Flp MCFO were at 21–22°C, except for a 10- to 60-min 37°C heat shock. Other crosses were typically at 25°C for most of development and 22°C in preparation for dissection. Flies were typically dissected at 1–5 days old for the cell-type collection rescreening, 1–8 days old for other non-MCFO crosses and 3–8 days old for MCFO images reported in this study. Dissection and fixation were carried out as previously described (Aso et al., 2014; Nern et al., 2015). Protocols are available at https://www.janelia.org/project-team/flylight/protocols.

Immunohistochemistry, imaging, and image processing were carried out as previously described (Aso et al., 2014; Nern et al., 2015; Sterne et al., 2021; Meissner et al., 2023). See Key Resources Table for antibodies. Additional details of image processing, including source code for processing scripts, are available at https://data.janelia.org/pipeline. See Supplementary file 2 for a user guide to interpreting the image data.

If inconsistencies were observed in expression between rescreened images and older image data for a line, the cross was repeated. If still inconsistent, the other copy of the stock (all are kept in duplicate) was examined. If only one of the two stock copies gave the expected expression pattern, it was used to replace the other copy. Repeated problems with expression pattern, except for stochasticity within a cell type, were grounds for removal from the collection. If sex differences were observed, crosses were also repeated. Lines reported with sex differences showed the same difference at least twice. We could not eliminate the possibility of stock issues appearing as sex differences by coincidentally appearing different between sexes and consistent within them.

Cell-type lines were qualitatively scored first by visual examination of averaged color depth Maximum Intensity Projection (MIP) CNS images (similar to Supplementary file 1; Otsuna et al., 2018), and then by a second observer examining individual images. Scores were applied as described in Figure 2. As images were taken from a variety of reporters, an effort was made to discount background specific to the reporters, especially 5XUAS-IVS-myr::smFLAG in VK00005.

3029 cell-type split-GAL4 lines were analyzed, each including images of one male and one female fly sample from rescreening. 6058 brains and 6058 VNCs were included. Images were aligned to the JRC2018 Unisex template (Bogovic et al., 2020) and segmented using Direction-Selective Local Thresholding (DSLT; Kawase et al., 2015). DSLT weight value, neuron volume percentage, MIP ratio against the aligned template, and MIP-weight ratio were adjusted to refine segmentation. DSLT connected fragmented neurons with up to a 5-µm gap, enhancing neuronal continuity. Noise was eliminated and objects were separated into individual neurons using a connecting component algorithm. The algorithm distinguishes individual neurons by recognizing connected voxel clusters as distinct entities based on DSLT 3D segmentation, followed by removal of voxels below 45/4095 (12 bit) brightness and objects smaller than 30 voxels.

The segmentation produced 19,313 non-overlapping brain objects and 18,860 VNC objects, consisting of neurons, trachea, debris, and antibody background. Most non-neuronal objects and some very dim neurons were eliminated using color depth MIP-based 2D shape filters and machine learning filters, followed by visual inspection, yielding 10,778 brain and 9310 VNC objects. We observed 74 objects containing tissue background with neurons. The background was manually 3D removed using VVDViewer. Neuron masks were combined into a single volume heat map with a newly developed Fiji plugin ‘Seg_volume_combine_to_heatmap.jar’ (https://github.com/JaneliaSciComp/3D-fiber-auto-segmentation/tree/main; Otsuna, 2023). Voxel brightness indicates the number of objects (at most one per line) within each voxel.