To gain genetic access to the cell types that convey CS and US information to the MB and those mediating memory-based actions, we designed screening to examine the expression patterns of over 4,000 intersections of split-GAL4 hemidrivers (Figure 1D). We selected 1,183 split-GAL4 lines representing various cell types in the central brain and the VNC for further characterization. For these lines, we employed higher resolution confocal microscopy and visualized the individual neurons that compose each split-GAL4 pattern with the MCFO method. We eventually identified 828 useful lines based on their specificity, intensity and non-redundancy. These fly lines are now publicly available through the webpage of the Janelia Flylight team project (http://www.janelia.org/split-gal4), where we have deposited a total of 28,376 confocal images from 6,374 tissue samples to document their expression patterns. We included lines with off-targeted expression, as they can be valuable for anatomical, developmental or functional imaging experiments, even if not suitable for behavioral experiments. Additionally, we retained drivers that intersected unintended cell types from the screening. Examples of confocal microscopy images are shown in Figure 1F and Figure 1-figure supplement 1.

We have annotated our split-GAL4 lines by matching the labeled neurons to their counterparts in the hemibrain connectome volume (Scheffer et al., 2020). We utilized confocal images registered to a standard brain, and matched neuronal cell types in each split-GAL4 line with those present in other lines and with the EM-reconstructed neurons (Figure 2A-D, see Materials and Methods). This light microscopy (LM) to EM matching process allows us to locate the cell type of each driver line in the connectome map, enabling users to select driver lines for further functional investigations based on their upstream and downstream synaptic partners (Figure 2E; Figure 2-figure supplement 1-20).

Among the 828 lines, 355 lines exhibit relatively high specificity to a non-redundant set of at least 319 cell types. Figure 1E provides an overview of the categories of covered cell types. Detailed information, including genotype, expression specificity, matched EM cell type(s), and recommended driver for each cell type, can be found in Supplementary File 1. A small subset of these lines have been previously used in studies (Aso et al., 2023; Dolan et al., 2019; Gao et al., 2019; Scaplen et al., 2021; Schretter et al., 2020; Takagi et al., 2017; Xie et al., 2021; Yamada et al., 2023). All transgenic lines newly generated in this study are listed in Supplementary File 2.

Prior to the completion of the EM connectome map, we conducted parallel efforts to identify cell types downstream of MBONs or upstream of DANs using confocal images of GAL4 drivers registered to a standard brain (Bogovic et al., 2020). We searched for GAL4 drivers containing cell types with potential connections to MBONs or DANs by quantifying the number of voxels overlapping with MBON axons or DAN dendrites (Otsuna et al., 2018). We then built split-GAL4 intersections from selected pairs of drivers from the established hemidriver library (Dionne et al., 2018; Tirian and Dickson, 2017).

After matching with EM-reconstructed neurons, these efforts resulted in split-GAL4 drivers encompassing 111 cell types that connect with the DANs and MBONs (Figure 3). Several of the cell types originally selected by LM were found to be not directly connected with MBONs or DANs. Nevertheless, these lines can be valuable for investigating other behaviors. For example, one such line, SS32237, was found to exhibit robust female-female aggression when activated (Schretter et al., 2020).

In the hemibrain EM connectome, there are about 400 interneuron cell types that have over 100 summed synaptic inputs from MBONs and/or synaptic outputs to DANs. Our newly developed collection of split-GAL4 drivers covers 30 of these major interneuron cell types (Figure 3C). While this constitutes a small fraction, it includes cell types with remarkable connectivity patterns. For instance, CRE011, presented as a single neuron per hemisphere, integrates over 2,000 inputs from 9 types of MBONs. This is the highest number of synaptic inputs from MBONs among all interneurons (Figure 3C). CRE011 provides cholinergic input to reward DANs (Figure 2E) and neurons in the lateral accessory lobe, a premotor center housing dendrites of multiple descending neurons (Kanzaki et al., 1994; Namiki et al., 2018). Another notable example is SMP108, which receives inputs from multiple glutamatergic MBONs and makes the highest number of cholinergic connections to DANs involved in appetitive memory (Figure 3C). We recently reported on SMP108’s role in second-order conditioning (Yamada et al., 2023) and its upstream neurons labeled in SS33917 in transforming appetitive memory into wind-directed locomotion (Aso et al., 2023). Supplementary File 3 contains connectivity information of the MB major interneurons, along with the predicted neurotransmitters (Eckstein et al., 2023) and the available driver lines.

The current collection also contains 189 lines for cell types that have innervations within the MB (Figure 3-figure supplement 1). These lines offer valuable tools to study cell types that may have been overlooked previously. Notably, the SS48794 driver labels OA-VUMa2 octopaminergic neurons, which are the Drosophila counterparts to the honeybee OA-VUMmx1 neurons, the first neurons identified as mediating US signals in an insect brain (Hammer, 1993). Moreover, several drivers in this collection provide improved specificity. When combined with previous collections (Aso et al., 2014a; Rubin and Aso, 2023), we now have coverage for 7 types of Kenyon cells and 58 out of 68 cell types within the MB (excluding PNs) using split-GAL4 drivers. Overall, this amounts to about 87% coverage for non-PN cell types within the MB and about 10% coverage for MBON-downstream and DAN-upstream cell types (Supplementary File 3 and 4, Figure 3B-C).

In Drosophila, the primary CS pathway to the MB involves the antennal lobe PNs that convey olfactory signals. We have developed a set of driver lines for PNs and other cell types in the antennal lobe (Supplementary File 1). This set includes 191 lines, covering more than 48 of the approximately 190 PN types identified through EM connectome and LM studies (Bates et al., 2020; Li et al., 2020; Lin et al., 2007; Tanaka et al., 2004; Zheng et al., 2022), encompassing both uni- and multi-glomerular PNs (Figure 4 and 5; Supplementary File 5).

The antennal lobe, in addition to the 51 olfactory glomeruli, contains 7 glomeruli involved in processing thermo- and hygro-sensory information (Enjin et al., 2016; Frank et al., 2015; Gallio et al., 2011; Jenett et al., 2012; Liu et al., 2015; Marin et al., 2020; Stocker et al., 1990; Tanaka et al., 2012). We provide 8 lines that cover sensory neurons projecting into these non-olfactory glomeruli and 18 lines covering the projection neurons emanating from them (Figure 6; Supplementary File 1 and 5).

Our understanding of the neural pathways that encode the US has been greatly advanced by experiments that have tested the sufficiency of various neuronal cell types to substitute for the US (Aso et al., 2010; Aso and Rubin, 2016; Chiang et al., 2011; Claridge-Chang et al., 2009; Hige et al., 2015a; Huetteroth et al., 2015; Liu et al., 2012; Saumweber et al., 2018; Schroll et al., 2006; Yamagata et al., 2015). These experiments leveraged thermogenic or optogenetic tools expressed in specific neuronal cell types, especially DANs, to assess their functions in associative learning. The approach to directly stimulate DANs, although valuable, bypasses the earlier US processing pathways and potential feedback in the circuit. Because of this experimental caveat, it is preferable to activate neurons at the sensory level of reward or punishment pathways to faithfully replicate the natural activity of these DANs. In that way, DANs can be regulated by both US sensory pathways and feedback pathways from MBONs. That is likely to be essential for the successful execution of more continual learning tasks in which flies update memories based on the current and past experiences (Felsenberg et al., 2017; Jiang and Litwin-Kumar, 2021; McCurdy et al., 2021; Rajagopalan et al., 2022).

Our collection provides the starting point for generating the genetic tools needed for this approach. For example, we have generated drivers for the cell types in the thermo-sensory pathways, as well as PNs for odors with innate preference, such as CO2 and cVA (Datta et al., 2008; Lin et al., 2013; Suh et al., 2004). These cell types are candidates to convey the reinforcement signals to the MB and other brain regions for associative learning.

On the other hand, gustatory sensory neurons constitute the first layer of neurons that detect food-related sensory signals (taste), which are conveyed to the MB through largely uncharacterized pathways (Bohra et al., 2018; Burke et al., 2012; Deere et al., 2023; Kim et al., 2017; Miyazaki et al., 2015; Sterne et al., 2021). GAL4 driver lines that recapitulate expression patterns of gustatory receptors (GRs) have been generated and utilized for functional studies (Dahanukar et al., 2007; Harris et al., 2015; Miyamoto et al., 2012; Wang et al., 2004; Yavuz et al., 2014). However, these driver lines tend to contain a morphologically and functionally heterogeneous set of sensory neurons (see for examples: Chen et al., 2022; Thoma et al., 2016) and may contain off-targeted expressions. To address these limitations, we have developed split-GAL4 drivers specific to different subsets of gustatory sensory neurons by generating hemidrivers for GR-gene promoters and screening intersections with existing hemidrivers (Figure 7A).

We used this strategy to generate Gr64f-split-GAL4 lines and create clean drivers for reward sensory neurons. In fruit flies, sugar is detected by sensory neurons located on different taste organs of the body and also inside the brain (Fujii et al., 2015; Hiroi et al., 2002; Miyamoto et al., 2012; Rodrigues and Siddiqi, 1978). Gr64f-Gal4, in which Gal4 is expressed under the control of the promoter of the sugar receptor gene Gr64f, broadly labels sugar sensory neurons (Dahanukar et al., 2007 and Figure 7 – figure supplement 1). Gr64f-Gal4 expression can be found in heterogeneous sets of sensory neurons in the labellum, the tarsi and the labral sense organ (LSO) located along the pharynx. In addition, Gr64f-Gal4 also labels subsets of olfactory receptor neurons and neurons innervating the abdominal ganglion (Park and Kwon, 2011)(Figure 7-figure supplement 1), despite that whether these cells endogenously express Gr64f is yet to be confirmed. Such heterogeneity in Gr64f-Gal4 expression could limit its usage to substitute reward in complex associative learning.

To refine Gr64f-Gal4 expression, we intersected Gr64f-GAL4DBD with various AD lines selected based on the projection patterns of Gr64f-Gal4, and obtained 16 stable split-GAL4 lines that have expression in distinct subsets of Gr64f neurons (Figure 7). We examined the ability of these lines to serve as US in olfactory learning (Aso and Rubin, 2016) and their potency to drive local search behaviors (Corfas et al., 2019)(Figure 8 and Figure 8-figure supplement 1). Additionally, we measured the walking speed of flies, as flies decrease walking while feeding (Thoma et al., 2016).

Among the Gr64f-split-GAL4 lines, SS87269 emerged as the best driver for substituting sugar reward in olfactory learning given its specificity and the utility in long training protocols (Figure 8E). SS87269 expresses in the labellum and at least two types of tarsal sensory neurons, namely the ascending (atGRN) and the non-ascending segment-specific (stGRN) types (Figure 7E-F). The driver does not label pharyngeal sensory neurons, and importantly, it lacks expression in abdominal ganglion and olfactory sensory neurons that could be off-targeted neurons found in the original Gr64f-GAL4. When odor presentation was paired with the activation of SS87269 neurons, flies formed robust appetitive memories even after extended training with high LED intensity (Figure 8E, Figure 8-figure supplement 1 and Video 1). Furthermore, and consistent with the idea that this subset of sensory neurons encodes appetitive food-related taste information, flies with SS87269 activation showed proboscis extension and reduced locomotion (Figure 8F-I; Figure 8-figure supplement 2, Video 2 and 4). These flies also showed robust local search behavior during the post-activation period, i.e., an increased probability of revisiting the area where they experienced the activation (Video 3). Notably, the revisiting phenotype of SS87269 was stronger than the original Gr64f-GAL4 and any other Gr64f-split-GAL4 drivers (Figure 8F and Figure 8-figure supplement 2A).

Two other lines SS88801 and SS88776, which label stGRNs or stGRNs along with labial sensory neurons, respectively (Figure 7K-L,I), showed appetitive learning and reduced locomotion during activation (Figure 8E-F). Interestingly, however, the activation of stGRNs with SS88801 did not induce significant local search behaviors (Figure 8F and Figure 8-figure supplement 2A). This finding could be valuable for understanding circuits underlying local search behavior and invites further investigation to compare pathways from labial and tarsal sensory neurons to the MB and the central complex.

In contrast to SS87269, two other resulting lines, SS87278 and SS87279, express in cells that appear to convey aversive signals. Activation of these lines induced an increase in walking speed during activation and reduced the probability of return at the offset of LED activation (Figure 8F and Figure 8-figure supplement 2). Also, flies became progressively less mobile during the inter-trial interval period (Figure 8G). The reduced locomotion in the interval period was also observed with the original Gr64f-GAL4 (Figure 8G) and a bitter-taste driver, Gr66a-GAL4 (data not shown). With extended training using SS87278 and SS87279, the preference to the paired odor eventually became negative (Figure 8E). These drivers label distinct subsets of sensory neurons projecting to the abdominal ganglion (Figure 7O,P). The innervation of SS87278 inside the abdominal ganglion is similar to that of Gr66a-GAL4 (Figure 7P) (Dunipace et al., 2001), which is known to label multidendritic sensory neurons in the adult Drosophila abdomen (Shimono et al., 2009). Examining projection patterns in fly bodies with whole-animal agar sections, we found that sensory neurons in SS87278 also project to the abdominal body surface (Figure 7J), likely representing the class IV multidendritic neurons that detect multimodal nociceptive stimuli (Hwang et al., 2007; Ohyama et al., 2015). GAL4 expression in these aversive sensory neurons may thereby explain the compromised learning performance with the original Gr64f-GAL4.

Overall, the refinement of Gr64f-Gal4 expression with SS87269 now allows for specific manipulation of the rewarding subset of gustatory sensory neurons and thereby permits training with an extended number of trials. While we have not yet conducted anatomical screening, LexADBD lines generated with Gr64f, Gr43a, Gr66a, Gr28b.d promoters (Figure 7A) should enable similar refinement of these sensory neuron lines. We also made a small number of lines for cell types in the subesophageal zone (SEZ) area (Figure 8-figure supplement 3), which complement previous collections of drivers for gustatory interneurons (Otto et al., 2020; Sterne et al., 2021).

Lastly, we generated driver lines for putative ascending nociceptive pathways. We determined the activation preference for 581 combinations of ZpGAL4DBD and p65ADZp hemidrivers in the circular arena (Figure 9A, Supplementary File 1). We found one driver, SS01159, which showed the most robust avoidance of LED quadrants and demonstrated behavioral components that are characteristic to nociception, including backward walking, turning, increased speed, and jumping upon activation (Figure 9B-D, and Video 5). This driver labels a group of ascending neurons (Figure 9E), which likely carry nociceptive signals from the body and legs to the brain. We then generated drivers for subsets of these ascending neurons guided by single neuron morphologies of cells in SS01159 determined by MCFO stochastic labeling (Nern et al., 2015). Other than SS01159, the collection in total contains approximately 100 split-GAL4 lines covering ascending neurons. While not completely matched to EM cell types due to only a portion of their morphologies being available in the hemibrain volume, these lines serve as a valuable resource for querying the reinforcement pathways.

Despite the recent progress in the EM connectome field, EM-based methods remain prohibitively expensive and morphological variability of neurons has been compared only across two individuals and three hemispheres (Schlegel et al., 2023). As a part of Janelia Flylight team project to generate cell-type-specific driver lines, we have imaged over 6,000 fly brains for the present study. While annotating those confocal images of split-GAL4 lines, we occasionally encountered samples with atypical morphologies (Figure 10). For example, one V_l2PN neuron, which typically projects to the lateral horn of the ipsilateral hemisphere, exhibited a peculiar morphology in one brain sample, where its axon crossed the midline and projected to the contralateral hemisphere, yet it still reached the correct target area within the lateral horn of the opposite side (Figure 10A). Another instance involved a DPM neuron, the sole serotonergic neuron of the MB lobes. While typical brain samples contain only one DPM neuron per hemisphere, we found a brain with two DPM neurons in one hemisphere (Figure 10B). In this case, the DPM projections exhibited an atypical innervation of the calyx of the mushroom body. We also found examples involving MBONs. The dendrites of MBON- α1 are typically confined inside the α lobe of the MB, but we discovered a case of MBON-α1 in addition projecting to the ellipsoid body (Figure 10C). MBON-α3 displayed striking variability in soma positions, but only subtle variability in axonal projections (Figure 10D). The table in Figure 10E summarizes additional examples of the atypical morphologies of MBONs. Overall in 1241 brain hemispheres examined, we found mislocalization of dendrites and axons in 3.14% and 0.97% of MB cell types, respectively. If this rate of mislocalization is generalizable to other brain regions, a fly brain of ∼130,000 neurons (Dorkenwald et al., 2023) would have a few thousands of neurons with mislocalized dendrites or axons. These examples of atypical morphology were observed only once in dozens of samples, and thus can be considered as erroneous projections either resulting from stochastic developmental processes, or possibly caused by ectopic expression of reporter proteins on the plasma membrane at a high level.

In contrast to these rare, and seemingly erroneous, morphological variations, we observed much more frequent and reproducible variations in the composition as well as morphologies in the two MBONs labeled by MB083C, which may amount to “individuality”. This split-GAL4 driver line invariably labels two cells in each hemisphere in 169 brain samples examined with four different reporters and in both sexes (57 males and 112 females; Figure 11-figure supplement 1). In all samples, these MBONs arborized dendrites in the γ3 and βʹ1 compartments. An obvious mistargeting of the axon was observed in only one sample, suggesting highly consistent development of these MBONs (Figure 11A-B). However, MCFO method visualized two distinct morphologies of these MBONs: MBON08 arborizes dendrites in the γ3 compartment of both hemispheres, whereas MBON09 arborize dendrites in ipsilateral γ3 and contralateral βʹ1 (Figure 11C-H)(Aso et al., 2014a). βʹ1 compartment was always labeled in both hemispheres for all 169 samples, suggesting that MBON09 represents at least one of the two cells in each hemisphere. The second cell can be either MBON08 or MBON09. In MCFO experiments, we observed 21 instances of MBON08 (8 in the left and 13 in the right hemisphere) and 188 instances of MBON09 (Figure 11I). Based on the likelihood, we expect 65% of flies contain four MBON09, while the remaining 35% of flies likely have at least one MBON08. In 71 hemispheres, two cells were visualized in different colors of MCFO: 52 contained two MBON09 and 19 contained one MBON08 and MBON09. We never observed a brain with MBON08 on both hemispheres or two MBON08 in one hemisphere (Figure 11J). When MBON08 and MBON09 coexist, MBON09 arborized in the lateral part of the ipsilateral γ3 and MBON08 arborize in the medial part of the contralateral γ3 (Figure 11E-H). This seemingly extended γ3 compartment innervated by MBON08 is not part of γ4, because it did not overlap with DANs in the γ4 (Figure 11-figure supplement 2A-B).

Although MBON08 was not found in brain samples used for the hemibrain or FAFB EM connectome (Dorkenwald et al., 2023; Scheffer et al., 2020; Zheng et al., 2018), DANs in the γ3 could be subdivided to two groups that innervate the medial or lateral part of the γ3 (Figure 11-figure supplement 2C) (Li et al., 2020). Therefore, subdivision of the γ3 compartment may exist irrespective of heterogeneity on the MBON side. In larval brains, two MBONs that correspond to adult MBON08/09 exhibit identical morphology (Eichler et al., 2017; Saumweber et al., 2018; Truman et al., 2023). Therefore, one of these MBONs acquires distinct morphology as MBON08 during metamorphosis at 21/209 odds, resulting in an asymmetric output pathway of the MB; We have never observed a brain with MBON08 on both hemispheres, and therefore MBON08 likely to appear in only one hemisphere, if at all (Figure 11J). This asymmetry could be one source of turning handedness and idiosyncratic learning performance (de Bivort et al., 2022; Smith et al., 2022), given that MBON09 forms extensive connection with other MBONs and fan-shaped body neurons (Li et al., 2020) and the activity of MBON08/MBON09 has strong influence on turning (Aso et al., 2023, 2014b; Matheson et al., 2022).

Split-GAL4 lines can be converted into split-LexA lines by replacing the GAL4 DNA binding domain with that of LexA (Ting et al., 2011). To broaden the utility of our collection, we have generated over 20 LexADBD lines to test the conversions of split-GAL4 to split-LexA. The majority (22 out of 34) of the resulting split-LexA lines exhibited very similar expression patterns to their corresponding original split-GAL4 lines (Figure 12). For the failed cases, the expression level was either undetectable or too weak.

The ability to define and manipulate a small group of neurons is crucial for studying neural circuits. Here, we have generated and characterized driver lines targeting specific cell types that are likely to be a part of associative learning circuits centered on the MB. We have provided these driver lines with a comprehensive lookup table linking these cell types with the EM hemibrain connectome. These lines, together with preceding collections of drivers (Aso et al., 2014a; Aso and Rubin, 2016; Davis et al., 2020; Dolan et al., 2018; Rubin and Aso, 2023; Shuai et al., 2015; see for examples: Sterne et al., 2021; Strother et al., 2017; Truman et al., 2023; Tuthill et al., 2013; Wang et al., 2021; Wolff and Rubin, 2018; Wu et al., 2016), collectively constitute a powerful resource for precise functional interrogation of associative learning in adult Drosophila melanogaster, and will be a foundation to reveal conserved principles of neural circuits for associative learning.

Drosophila melanogaster strains were reared at 22°C and 60% humidity on standard cornmeal food in 12:12 hour light:dark cycle. The genotypes of all split-GAL4 and split-LexA driver lines released here are listed in the Supplementary File 1. The new collection of split-GAL4 drivers reported here was designed based on confocal image databases (http://flweb.janelia.org) (Jenett et al., 2012; Tirian and Dickson, 2017), and screening expression patterns of p65ADZp and ZpGAL4DBD combinations was performed as described previously (Aso et al., 2014a; Pfeiffer et al., 2010).

Brains and ventral nerve cords of 3-10 days old flies were dissected, fixed and immunolabeled and imaged with confocal microscopes (Zeiss LSM710, LSM780 or LSM880) as previously described (Aso et al., 2014a; Jenett et al., 2012; Meissner et al., 2023; Nern et al., 2015). The detailed protocols and videos are available at https://www.janelia.org/project-team/flylight/protocols.

For sample preparation, flies were anesthetized on ice and briefly washed with 70% ethanol. Small incisions were made in the flanks of the thorax and abdomen under 2% paraformaldehyde in PBS with 0.1% Triton X-100 (PBS-T), and the flies were fixed in this solution overnight at 4°C. After washing in PBS containing 1% Triton X-100, the samples were embedded in 7% agarose and sectioned on Leica Vibratome (VT1000s) sagittally in slices of 0.3 mm. The slices were incubated in PBS with 1% Triton X-100, 0.5% DMSO, 3% normal goat serum, Texas Red-X Phalloidin (1:50, Life Technologies #T7471) and anti-GFP rabbit polyclonal antibodies (1:1000, Thermo Fisher, #A10262) at room temperature with agitation for 24 hours. After a series of three washes in PBS-T, the sections were incubated for another 24 hours in the solution containing secondary antibodies (1:1000, goat anti-rabbit, Thermo Fisher #A32731). The samples were then washed in PBS-T and mounted in Tris-HCL (pH 8.0)-buffered 80% glycerol + 0.5% DMSO. For imaging and rendering, serial optical sections were obtained at 2 µm intervals on a Zeiss 880 confocal microscope with a pan-apochromat 10x/0.45 NA objective using 488 and 594 nm lasers. Images were processed in Fiji (http://fiji.sc/) and Photoshop (Adobe Systems Inc.).

For flies expressing CsChrimson (Klapoetke et al., 2014), the food was supplemented with retinal (0.2 mM all-trans-retinal prior to eclosion and then 0.4 mM). Two- to six-day old adult females were collected and sorted on a Peltier cold plate 2-4 days before testing in behavioral assays. Flies were starved for 40–48 hr on 1% agar before they were subjected to behavioral experiments. Olfactory conditioning and optogenetic activation experiments were performed as previously described using the modified four-field olfactory arena equipped with the 627 nm LED board and odor mixers (Aso and Rubin, 2016; Pettersson, 1970). The odors were diluted in paraffin oil: pentyl acetate (PA, 1:10000, v/v) and ethyl lactate (EL, 1:10000, v/v). Videos were taken at 30 frames per second and analyzed using Fiji and Caltech FlyTracker (Eyjolfsdottir et al., 2014).

The confocal microscopy images of different split-GAL4 lines were registered to a common template JRC2018_unisex (Bogovic et al., 2020) and organized in Janelia Workstation software (https://github.com/JaneliaSciComp/workstation). Color depth MIP mask search (Otsuna et al., 2018) was used to search through the EM neuron library (hemibrain 1.2.1) for matching candidates. The searching target was obtained by either creating a mask on the full confocal image or using neurons of interest manually segmented in VVD viewer (https://github.com/takashi310/VVD_Viewer)(Wan et al., 2012). SWC files of top-matching EM neuron candidates were loaded into VVD viewer together with the confocal microscopy images in the same standard brain coordinates. By rotating the 3d images and manually scrutinizing the branching patterns, we picked the best matching candidate. Typically, we had high confidence of LM-to-EM matching for the line that labels only one cell per hemishere. For instance, we could unambiguously match the cell in SS67721 with SMP108 in the EM hemibrain volume. Our confidence of LM-to-EM matching tended to be lower for the lines that label multiple cells, because neurons of similar morphologies could be labeled together in those lines.

Connectivity information was retrieved from neuPrint (neuprint.janelia.org), a publicly accessible website hosting the “hemibrain” dataset (Scheffer et al., 2020). For cell types, we used cell type name assignments reported in Sheffer et al., 2020. Only connections of the cells in the right hemisphere were used due to incomplete connectivity in the left hemisphere. The 3D renderings of neurons presented were generated using the visualization tools of NeuTu (Zhao et al., 2018) or VVD viewer.

Statistical comparisons were performed on GraphPad Prism 7.0 using one-way ANOVA followed by Dunnett’s test for multiple comparisons. Sample size was not predetermined based on pilot experiments.