The complex synaptic pathways onto a looming‐detector neuron revealed using serial block‐face scanning electron microscopy

The ability of locusts to detect looming stimuli and avoid collisions or predators depends on a neuronal circuit in the locust's optic lobe. Although comprehensively studied for over three decades, there are still major questions about the computational steps of this circuit. We used fourth instar larvae of Locusta migratoria to describe the connection between the lobula giant movement detector 1 (LGMD1) neuron in the lobula complex and the upstream neuropil, the medulla. Serial block‐face scanning electron microscopy (SBEM) was used to characterize the morphology of the connecting neurons termed trans‐medullary afferent (TmA) neurons and their synaptic connectivity. This enabled us to trace neurons over several hundred micrometers between the medulla and the lobula complex while identifying their synapses. We traced two different TmA neurons, each from a different individual, from their synapses with the LGMD in the lobula complex up into the medulla and describe their synaptic relationships. There is not a simple downstream transmission of the signal from a lamina neuron onto these TmA neurons; there is also a feedback loop in place with TmA neurons making outputs as well as receiving inputs. More than one type of neuron shapes the signal of the TmA neurons in the medulla. We found both columnar and trans‐columnar neurons connected with the traced TmA neurons in the medulla. These findings indicate that there are computational steps in the medulla that have not been included in models of the neuronal pathway for looming detection.


| Immunocytochemical labeling of retinal wholemounts
For immunolabeling of retinal wholemounts, we used a marker (Penol) to place a spot of permanent ink (xylene free) on the dorsal part of the eye before removing it from the orbit. After enucleation, we made a small cut in the dorsal part of the retina and sclera before dissecting the retina from the eye cup. The retina was then flattened by making four radial incisions from the periphery almost to the center and transferred onto the non-gridded surface of a piece of nitrocellulose filter paper (Millipore, cat. number HABG01300). The filter paper with attached retina was then positioned on a piece of folded tissue paper (e.g., Kimwipe) and a few drops of HEPES-buffered extracellular solution was added on top and allowed to soak through. For fixation, a few drops of 4% paraformaldehyde (in .1 M PB) was added on top and allowed to soak through. After repeating this 2-3 times, the filter paper with retina was transferred to a larger volume of 4% paraformaldehyde in .1 M PB and fixed for 30 min at room temperature.
After fixation, the retina was washed six times (10 min each) in .01 M PBS and incubated in antibody incubation solution (identical to that used for slices) overnight at 4 C. The retina was then incubated for four nights (at 4 C) with primary antibodies (mouse anti-ankyrin-G and guinea pig anti-parvalbumin; Table 1) in antibody incubation solution identical to that used for slices, but with .2% Triton X-100. Afterwards, the retina was washed six times (10 min each) in PBS and incubated overnight (at 4 C) with secondary antibodies (goat antiguinea pig coupled to Alexa 488, #A11073 from Thermo Fisher Scientific, diluted 1:1000; goat anti-mouse coupled to Alexa 594, #A11032 from Thermo Fisher Scientific, diluted 1:1000) in antibody incubation solution (identical to that used for slices, i.e., with .2% Triton X-100).
Finally, the retina was washed six times (10 min each) in PBS and mounted in Vectashield between a microscope slide and a precision T A B L E 1 Primary antibodies Antibody name Immunogen Source, cat #, RRID Antibody type Dilution

| INTRODUCTION
The timely detection of an approaching predator can be critical for the survival of an animal. Throughout the animal kingdom, species have evolved different detection systems matching the constraints and needs of their respective environments (de Vries & Clandinin, 2012;Dunn et al., 2016;Klapoetke et al., 2017;Peek & Card, 2016;Wang & Frost, 1992).
Studies on the neuronal substrate of such detection systems identified several different types of looming-sensitive neurons: the giant fiber neurons in Drosophila (Peek & Card, 2016), which have counterparts in several other species, such as crabs, (monostratified lobula giant neurons of type 1,MLG1; Oliva & Tomsic, 2014), praying mantis (tangential projecting neurons in the lobula complex; Yamawaki, 2019), or in zebra fish (periventricular neurons of the optic tectum; Dunn et al., 2016). Physiological, structural, and ultrastructural investigations of these neurons and their circuits are necessary to highlight their similarities and differences. Considering the differences in behavior and distances in the animal tree, comparative studies of the neuronal circuits and their computational solutions will improve our understanding of the different ways evolution solved the task of predator/looming detection.
For locusts, visual cues allow the animal to discriminate between dangerously approaching objects and receding or translating ones (Gabbiani et al., 2002;Schlotterer, 1977). The underlying neuronal circuit in the locusts' visual system has been studied over several decades Gabbiani et al., 2004;Olson et al., 2021;Rind, 1987Rind, , 1996Sztarker & Rind, 2014). The physiology of the key neurons, the descending contralateral movement detector (DCMD) and the lobula giant movement detector (LGMD) is well documented and the whole system is used as the template for artificial collision detection models (Fu et al., 2019;Rind & Bramwell, 1996;Yue et al., 2006). There are two giant movement-detecting neurons located in the locust's lobula complex (LOX), the LGMD1, and LGMD2. For both neurons, it is characteristic that conspicuous dendritic arborizations are found in the LOX. In the case of LGMD1, this arborization can be divided into three distinct subfields, A, B, and C (O'Shea & Williams, 1974). The inputs to subfield A are of excitatory nature and involve acetylcholine release Rind & Leitinger, 2000). Subfield A receives synaptic input from afferent neurons covering almost the visual field of the locust's compound eye (Krapp & Gabbiani, 2005). The inputs to the LGMD1's subfields B and C are both inhibitory and phasic in nature mediated by GABA A release . The LGMD2 only has one dendritic field where it receives both excitatory and inhibitory input (Rind, 1987;. The input neurons of the LGMD1 subfield A originate from the second neuropil, the medulla (therefore, named trans-medullary afferent [TmA] cells) and connect the medulla with the outer region of the LOX (OLO, Rosner et al., 2017). Unfortunately, knowledge about these cells is sparse. In 1981, Strausfeld and Nässel showed using cobalt staining that the cells connecting the medulla with this subfield of the LGMD have two different types of branching patterns, indicating that there may be at least two different cell populations of TmA cells (Strausfeld & Nässel, 1981). Their connection with the LGMD was shown, in 2018, light microscopically using en masse staining . The ultrastructural features of LGMD1 input are well documented: the chemical synapses between TmA neurons and LGMD1 and 2 are always arranged in such a way that neighboring TmA neurons have output synapses both onto the LGMD and onto their neighbors (Rind & Simmons, 1998;Sztarker & Rind, 2014).
Each output synapse thus has two postsynaptic partners, the LGMD1 and the neighboring afferent neuron. Afferents make back-to-back synapses with one another and the LGMD1. The importance of this synaptic setup for the detection of looming stimuli is still debated with both lateral inhibitory and excitatory effects of one TmA on its neighbor proposed. Physiological experiments and modeling showed lateral inhibition is present and improves collision tuning (O'Shea & Williams, 1974;Pinter, 1977Pinter, , 1979Pinter, , 1983Rind et al., 2016;Rind & Bramwell, 1996). However, this role for lateral inhibition, mediated by muscarinic receptors (Rind et al., 2016;Rind & Leitinger, 2000;, 1998 has been questioned as blocking muscarinic receptors near the LGMD1 with receptor antagonist decreases excitation while potentiating their action leads to increased excitation not inhibition (Zhu et al., 2018). While the TmA neurons' terminal barbarizations in the LOX were previously well described, it was not known in which layer in the medulla they receive synapses from upstream neurons, nor whether they have dendritic branches in the medulla, which is commonly found in medulla interneurons, or receive input synapses along only one dendrite, which so-far was described only for Tm2 neurons in layer M10 of the Drosophila medulla (Shinomiya, Huang, et al., 2019). For a better and complete understanding of the whole computation process within the looming sensing pathway, a detailed description of the synapses made by the medullary cells' contributing to the circuit is needed. On electron micrographs of the LOX, the LGMD1 and 2 can be easily identified by their characteristic arrangement of the dendritic cross-sections within the neuropil with the LGMD2 directly posterior to the LGMD1 (Rind & Simmons, 1998;Simmons et al., 2013). This fact allows the use of the LGMD1 as a starting point of tracing the TmA neurons back to the medulla. The TmA neurons are themselves unambiguously identifiable due to their reciprocal synaptic arrangement. However, the tracing of a neuron is a challenging task. Over the last decades, the method of choice was to collect ultra-thin sections on grids for transmission electron microscopy investigations but this technique is very time consuming and prone to errors and therefore rarely used by researchers. Serial block-face scanning electron microscopy (SBEM) was (re-)introduced in 2004 ( Denk & Horstmann, 2004) and became a vital tool investigating large sample volumes in life and materials science (Zankel et al., 2009), but especially in neuroscience (Helmstaedter et al., 2013;Holcomb et al., 2013;Mukherjee et al., 2016;Scheffer et al., 2020). With SBEM, an ultra-microtome is installed in the specimen chamber of a scanning electron microscope and the block-face of the specimen is imaged after each cut with the diamond knife of the microtome. This technique makes it possible to identify TmA neurons at the LGMD by their synapses and to trace the neurons over a distance of over 300 μm to the medulla. Here, we present the first morphological description of two TmA neurons of fourth instar locusts, reconstructed along their whole length from serial electron micrographs obtained using SBEM. We also present the number and direction of their synapses and reconstructions of key parts of their pre-and postsynaptic neurons.
2 | MATERIAL AND METHODS

| Animals and sample preparation
The locusts were obtained from a gregarious culture with day/night cycle of 12/12 h at a mean temperature of 30 C at the Biosciences Institute (Biosciences Institute, Newcastle University, UK). Locusts are big insects so to reduce the distances to be reconstructed smaller fourth instar locusts were used (Figure 1a). At this stage, the LGMD1 and 2 have their mature form except at their branch extremities (Sztarker & Rind, 2014). Reconstructions were begun at dendritic subfield A in the mid region of a large branch where synapses should be mature. We scanned a data set that covered the main branches of one entire TmA in one sample, and another entire TmA in a second sample.
The animals were chilled on ice and brains were dissected and cut in half. During dissection, the brain was kept moistened with insect saline.  Sztarker & Rind, 2014). Continual manual block realignments were necessary to follow the TmAs into the medulla (adjustments 1-5 shown). The protocol was suitable for the SBEM imaging, though we experienced some problems with accumulations of a contrast chemical (presumably uranyl acetate; Figure 1c). Fortunately, neither the identification of the LGMDs in the LOX (Figure 1c, Sigma-Aldrich), .5% Triton X-100 (Sigma-Aldrich), and .05% NaN 3 .
Slices were then incubated for three nights (at 4 C) with primary antibody (mouse anti-ankyrin-G or guinea pig anti-ankyrin-G;

| Immunocytochemical labeling of retinal wholemounts
For immunolabeling of retinal wholemounts, we used a marker (Penol) to place a spot of permanent ink (xylene free) on the dorsal part of the eye before removing it from the orbit. After enucleation, we made a small cut in the dorsal part of the retina and sclera before dissecting the retina from the eye cup. The retina was then flattened by making four radial incisions from the periphery almost to the center and transferred onto the non-gridded surface of a piece of nitrocellulose filter paper (Millipore, cat. number HABG01300). The filter paper with attached retina was then positioned on a piece of folded tissue paper (e.g., Kimwipe) and a few drops of HEPES-buffered extracellular solution was added on top and allowed to soak through. For fixation, a few drops of 4% paraformaldehyde (in .1 M PB) was added on top and allowed to soak through. After repeating this 2-3 times, the filter paper with retina was transferred to a larger volume of 4% paraformaldehyde in .1 M PB and fixed for 30 min at room temperature.
After fixation, the retina was washed six times (10 min each) in .01 M PBS and incubated in antibody incubation solution (identical to that used for slices) overnight at 4 C. The retina was then incubated for four nights (at 4 C) with primary antibodies (mouse anti-ankyrin-G and guinea pig anti-parvalbumin; The resin blocks were mounted on an ultra-microtome Leica an accurate tracing and reliable identification of synapses. In the case of inconstant cutting, the thickness was increased to 50 nm. Side branches are absent within the region of the OCH2, so for most of the OCH2 of the first specimen 60 nm section thickness was chosen due to time considerations. Section thickness between 50 and 60 nm is suitable to identify and quantify synapses in the insect nervous system (Meinertzhagen, 1996). The total number of sections was 8400 for Sample 1 and 7500 for Sample 2. The region of the OCH2 was identified by the presence of tightly packed neurons together with glial cells and cell bodies.
The protocol was suitable for the SBEM imaging, though we experienced some problems with accumulations of a contrast chemical (presumably uranyl acetate; Figure 1c). Fortunately, neither the identification of the LGMDs in the LOX (Figure 1c,d), nor the tracing of TmA cells were compromised by this artifact.

| Segmentation of data
Reconstructions were done using the Amira ® 3D Software (Version 5.6.0 FEI™). The micrographs were imported as TIFF-files after converting them from the original GATAN file type dm3 using DM. The Movies were generated using Amira ® .

| Skeleton reconstructions
Additional neurons around TmA1 and 2 were traced in each animal using the filament editor module in Amira ® . Instead of displaying the neurons as a mesh, only the skeleton of the neurons is reconstructed.
In total, 9 neurons were traced in Animal 1 and 23 neurons in Animal 2. The neurons were chosen close to the traced TmA1 and 2 neurons, which appeared to be a bundle of neurons enclosed by a glia cell. In Animal 2, neurons from two other bundles which looked similar to the TmA bundle were also traced. As initial starting point for tracing, neuronal cross-sections were chosen in the area of the transition from OCH2 to medulla.
Slices were then incubated for three nights (at 4 C) with primary antibody (mouse anti-ankyrin-G or guinea pig anti-ankyrin-G;

| Immunocytochemical labeling of retinal wholemounts
For immunolabeling of retinal wholemounts, we used a marker (Penol) to place a spot of permanent ink (xylene free) on the dorsal part of the eye before removing it from the orbit. After enucleation, we made a small cut in the dorsal part of the retina and sclera before dissecting the retina from the eye cup. The retina was then flattened by making four radial incisions from the periphery almost to the center and transferred onto the non-gridded surface of a piece of nitrocellulose filter paper (Millipore, cat. number HABG01300). The filter paper with attached retina was then positioned on a piece of folded tissue paper (e.g., Kimwipe) and a few drops of HEPES-buffered extracellular solution was added on top and allowed to soak through. For fixation, a few drops of 4% paraformaldehyde (in .1 M PB) was added on top and allowed to soak through. After repeating this 2-3 times, the filter paper with retina was transferred to a larger volume of 4% paraformaldehyde in .1 M PB and fixed for 30 min at room temperature.
After fixation, the retina was washed six times (10 min each) in .01 M PBS and incubated in antibody incubation solution (identical to that used for slices) overnight at 4 C. The retina was then incubated for four nights (at 4 C) with primary antibodies (mouse anti-ankyrin-G and guinea pig anti-parvalbumin;

| RESULTS
A general description of the LOX, and medulla is given in (Rosner et al., 2017;Strausfeld, 1976;Strausfeld & Nässel, 1981), and a description of the LGMD1 and 2's ultrastructure is available in (Rind et al., 2016;Rind & Simmons, 1998). The anatomy of the LGMD1 and 2's reconstructed from silver-stained sections in a fourth instar locust is shown in Figure 1d. In this study, we used SBEM to trace two neurons, which synapsed with the LGMD1.
Each TmA was from a different animal. Identification of the LGMDs was possible because the LGMD1 subfield A forms a dendritic arbor that can be clearly recognized in sections of the optic lobe where it appears as a crescent of lightly stained profiles ( Figure 1c). Profiles of TmA neurons involved in looming detection are identified by their output synapses onto the LGMD1 (Figure 2a-d,f).
T A B L E 1 Synaptic connections of TmA1 and TmA2 neuron in the LOX. Connections were counted and categorized into: input synapses (IP), or output synapses (OP). The number of synapses that had a reciprocal back-to-back partner in a neighboring neuron is also stated. Synaptic partners were either the LGMD1, the LGMD1 but not the reconstructed branch or unclassified   Sigma-Aldrich), .5% Triton X-100 (Sigma-Aldrich), and .05% NaN 3 .
Slices were then incubated for three nights (at 4 C) with primary antibody (mouse anti-ankyrin-G or guinea pig anti-ankyrin-G;   LOX. A part of the main dendritic arbor of the LGMD1 was identified ( Figure 1c) and reconstructed in both animals (Figure 3a,d). In the LOX, both TmA1 and 2 neurons ramified extensively into smaller neurites (<1 μm in diameter) that exhibited multiple synapses with the LGMD1 (Table 1). The overall number of synapses made by TmA1 or 2 onto the reconstructed LGMD1 branch, was lower than those made onto unreconstructed LGMD1 processes (Table 1). For TmA2 particularly, there were more synapses at identified LGMD1 branches that were not reconstructed. The highest number of synapses was found with additional cells which we could not classify. Despite the high number of synaptic connections, there were only a small proportion of reciprocal synapses (RS) between TmA1 and 2 neurons and other, unidentified cells in the LOX (Table 1).
Slices were then incubated for three nights (at 4 C) with primary antibody (mouse anti-ankyrin-G or guinea pig anti-ankyrin-G;   Previously the medulla had been regarded as an input area for the TmAs only (Jones & Gabbiani, 2010;Rind & Bramwell, 1996) and it was assumed that they would receive input from LMC neurons, which terminate in ML1-3 (Elphick et al., 1996). When we mapped the number and distribution of input and output synapses along the entire TmA reconstructions, we found predominantly input synapses in the medulla but both input and output synapses were identified along the entire reconstruction of the TmA neurons, indicating a much more complex synaptic arrangement than previously assumed.

| Immunocytochemical labeling of retinal wholemounts
A total of 490 synapses were found in the TmA1 neuron, and 493 in TmA2, including the synapses counted in the LOX, the OCH2, and the medulla (Figure 3b, Table 2). A back-to-back reciprocal synapse involving two TmAs and the LGMD counts as two synapses,  Table 2). In the medulla, many more input  Sigma-Aldrich), .5% Triton X-100 (Sigma-Aldrich), and .05% NaN 3 .
Slices were then incubated for three nights (at 4 C) with primary antibody (mouse anti-ankyrin-G or guinea pig anti-ankyrin-G;   OCH2 (28 and 34, respectively; Figure 3b, Table 2), which are almost entirely input synapses, except in three cases for TmA1.

| TmA1 and 2 synaptic partner neurons in the medulla and OCH2 revealed
The description of the medulla interneurons focuses mainly on those associated with TmA2 as the image quality of this data set was better and allowed a more reliable tracing of finer neurites. The profiles that were found in synaptic contact with TmA2 varied in electron density and in the number of mitochondria and synaptic vesicles, indicating that several different neuron types may be involved (Figure 4e-p).
They included profiles that resembled photoreceptor terminals occurring in the lamina, with dark cytoplasm and a high number of mitochondria (data not shown), however they could not be photoreceptors as the only long photoreceptor in the locust, R7, does not extend beyond ML3 (Schmeling et al., 2015).
Previously, there were only speculations about the wiring of the TmA neurons in the medulla. We traced a total of 13 synaptic partner neurons: 3 neurons connected to TmA1 (Figures 5 and 6) and ten to TmA2 (Figures 7 and 8). For TmA1 and TmA2, we can distinguish two groups of synaptic partner neurons based on their morphology. The first group accompanies the TmA neurons over their entire length, or at least over most of the proximal medulla.  7 and 8a,b) and the synapse distribution there mapped ( Figure 8c). In morphology and types and location of synapses, TmAL1 and TmAL2 closely resembled TmA1 and TmA2 respectively, although the number of synapses found along the branches of TmAL1 is much lower than the number found at the other TmAL2-4 neurons ( Table 3). A small proportion of both input-and output synapses had a back-to-back reciprocal partner synapse in a neighboring profile.
ii. TmAL3, (Figures 7c and 8d, iii. TmAL4 (Figures 7d and 8f)  SCR_000034). With the pipette resistances used, we typically did not find it necessary to apply a retaining current to prevent leakage of dye. After successful impalement, cells were injected with a current of À500 pA for 3 min ($1 Hz; 900 ms on per cycle).
Slices were then incubated for three nights (at 4 C) with primary antibody (mouse anti-ankyrin-G or guinea pig anti-ankyrin-G; Table 1)

| Immunocytochemical labeling of retinal wholemounts
For immunolabeling of retinal wholemounts, we used a marker (Penol) to place a spot of permanent ink (xylene free) on the dorsal part of the eye before removing it from the orbit. After enucleation, we made a small cut in the dorsal part of the retina and sclera before dissecting the retina from the eye cup. The retina was then flattened by making four radial incisions from the periphery almost to the center and transferred onto the non-gridded surface of a piece of nitrocellulose filter paper (Millipore, cat. number HABG01300). The filter paper with attached retina was then positioned on a piece of folded tissue paper (e.g., Kimwipe) and a few drops of HEPES-buffered extracellular solution was added on top and allowed to soak through. For fixation, a few drops of 4% paraformaldehyde (in .1 M PB) was added on top and allowed to soak through. After repeating this 2-3 times, the filter paper with retina was transferred to a larger volume of 4% paraformaldehyde in .1 M PB and fixed for 30 min at room temperature.
After fixation, the retina was washed six times (10 min each) in .01 M PBS and incubated in antibody incubation solution (identical to that used for slices) overnight at 4 C. The retina was then incubated for four nights (at 4 C) with primary antibodies (mouse anti-ankyrin-G and guinea pig anti-parvalbumin; Table 1) in antibody incubation solution identical to that used for slices, but with .2% Triton X-100. Afterwards, the retina was washed six times (10 min each) in PBS and incubated overnight (at 4 C) with secondary antibodies (goat antiguinea pig coupled to Alexa 488, #A11073 from Thermo Fisher Scientific, diluted 1:1000; goat anti-mouse coupled to Alexa 594, #A11032 from Thermo Fisher Scientific, diluted 1:1000) in antibody incubation solution (identical to that used for slices, i.e., with .2% Triton X-100).
Finally, the retina was washed six times (10 min each) in PBS and mounted in Vectashield between a microscope slide and a precision and Figure 9). A back-to-back RS made by TmA neurons with an M neuron were only found once (shared with M7 and TmAL3).

| Neurons neighboring TmA1 and 2 in the medulla
In both animals, we traced neurons located close to each TmA neuron at the proximal border of the medulla (ML10). In Animal 1, we followed nine neurons within the same bundle as TmA1, in the OCH2,  Figure 10b,e respectively, and in relation to the overall neural composition of the medulla in the silver stained section (Blest, 1961) in Figure 10g. This can also be seen in 3D for cells surrounding TmA2 in Movie S2.

| DISCUSSION
To our knowledge, this is the first description of the anatomy and connectivity of columnar TmAs synaptically connected with the LGMD1, a well-known looming detector in the locust (Gabbiani et al., 2002;Schlotterer, 1977). We traced two different TmA neurons (TmA1 and 2) in excess of 300 μm from their synapses with the LGMD1 back along their length within a block-face, centered on a single medulla column ( Figure 1). Previously, only very few publications showed the neuroanatomy of TmA cells (Rind et al., 2016;Strausfeld & Nässel, 1981;Wang et al., 2018). Moreover, the work we present here is the first description of those cells that are presynaptic to the LGMD1 from their origin in the medulla to their terminals in the lobula complex. TmA1 and 2 have a similar overall shape to the two classes of neurons filled by trans-synaptic migration of cobalt stain from the main sub-field A of the LGMD1 (Strausfeld & Nässel, 1981). Like TmA1, one class had dendrites in ML 1-3 and the other class, had branches in ML 6-10 only (T cells in figure   67, Strausfeld & Nässel, 1981). Recently, Wang et al. (2018) also showed fluorescence images of en mass medullary interneurons in close association with the LGMD1 but individual afferents were not followed to their origin in the medulla. However, light activation of the afferents using optogenetic stimulation led to depolarization of the LGMD1 indicating a likely connection between them. Wang et al. (2018) showed that the side branches leading to the TmA cell bodies have lengths of up to approximately 100 μm, which is well beyond the width of our data set and may explain why we were not able to reconstruct the TmA's cell bodies.
The synapses of afferents onto the LGMD1 showed some differences with previous studies. A characteristic of the TmA neurons is that they exhibit output synapses both with the LGMD1 and with their neighbors in a reciprocal arrangement, with pairs of presynaptic terminals back-to-back with one another sharing a synaptic cleft (Rind & Simmons, 1998). Our study using SBEM showed that most TmA output synapses onto the LGMD1 were not reciprocal; at least we found no presynaptic densities back-to-back in many cases. This could be due to the immaturity of the locust. However, this is not a good explanation for two reasons; previous work has shown that there is already a functional, LGMD1 driven collision avoidance behavior in L4 and RS are in evidence even in L1 (Simmons et al., 2013;Sztarker & Rind, 2014). Plus, we chose a part of the LGMD1 dendritic tree that was not adding new synapses, we began our reconstruction at medium sized branches of the LGMD1 as it is known that newly formed facets of the eye are integrated into the anterior eye margin with every instar (Anderson, 1978). The anterior eye margin is connected with the finest branches of the LGMD1 , and it is these finer branches which are formed during postembryonic larval development (Sztarker & Rind, 2014). In this way, we aimed to avoid newly connecting synapses that are still in development. A different explanation could lie in our very conservative approach: we only included synapses if at least two out of three independent researchers agreed that they were clearly visible.
Moreover, the enhanced osmium staining protocol (Deerinck et al., 2010) applied here has been shown to stain certain proteinaceous structures like the mammal postsynaptic density less intensely than previous protocols (Capetian et al., 2020), so some presynaptic bars may have been excluded because of insufficient staining. This is a possibility, as when TEM methods with standard contrasting were used in L1 and L4 RS were abundant in the LGMDs (figure 9 in Simmons et al., 2013). This exclusion of presynaptic densities may lead to an underestimate of synaptic connectivity but will not lead to false synapse identification.
We also found chemical synapses between the TmA1 and 2 neurons and neighboring neurons in the OCH2. In the OCH2 the TmA neurons mostly received synaptic input and it would be of great interest to know what cell type is involved here. In Drosophila, synapses occur between neurons in the OCH2 but they are rare and involve amacrine cells (ME-OCH2 cell; see figure 5A,B in Shinomiya, Huang, et al., 2019). These observations, may suggest that the chiasm is not merely a bundle of axons going elsewhere, but is also functional neuropil modulating visual information dynamically.
The TmA's synaptology in the medulla gives new insight into the processing stages of the collision detection pathway. The presence of predominantly input synapses in the medulla fits well with assumptions made for early-simplified computational models of the LGMD's circuit, where it was assumed that the TmA neurons receive input from LMC neurons in the medulla (Rind & Bramwell, 1996;Yue et al., 2006). However, the LMC neurons terminate in the distal Sigma-Aldrich), .5% Triton X-100 (Sigma-Aldrich), and .05% NaN 3 .
Slices were then incubated for three nights (at 4 C) with primary antibody (mouse anti-ankyrin-G or guinea pig anti-ankyrin-G; medulla in ML1-3 (Elphick et al., 1996) and only TmA1 has input and output synapses there in ML1-3. TmA2 at least must interact synaptically with neurons other than the LMCs. An option that cannot be excluded is that TmA1 neurons receive direct synaptic input from the only long photoreceptor cell R7, which terminates in the distal medulla, in layer 3 (Nowel & Shelton, 1981;Schmeling et al., 2015).
Particularly, as some of the processes in contact with the TmA neu-  (Stuart, 1999). Off excitation must be delivered from neurons other than R7. TmA2 has many synapses in the medulla (ML6-10) both inputs and outputs. Curiously, apart from varicosities and a few protrusions that did not exceed 3 μm, we did not find side branches from the main neurite, a situation which was shown for the Tm2 neurons in layer M10 of the Drosophila medulla (Shinomiya, Horne, et al., 2019), but is not common among the trans-medullary neurons involved in the directionally selective EMD in insects (Nowel & Shelton, 1981;Fischbach & Dittrich, 1989;Takemura et al., 2017;Shinomiya, Huang, et al., 2019;bee Ribi & Scheel, 1981;summarized in Borst et al., 2020). Looking at the synaptic arrangement in 3D, we found a small proportion of back-to-back RS in the medulla. To our knowledge, such an arrangement has not previously been described for the medulla and its functional significance would need further studies.
Although we are confident in finding small side branches even in neurons that we only skeletonized, there is a possibility certain side branches were missed. Nevertheless, the fact that we found both input and output synapses concentrated on these naked neurites ( Figure 3) led us to assume that the LGMD pathway is quick which may lead to simplified anatomy in its input neurons. The neurons, such as TmA2, did have regular knobbly varicosities throughout the medulla that often contained mitochondria. TmA1 resembles Tm9 that provides excitation to T5, the OFF channel of the EMD for direction selective motion detection in Drosophila (Shinomiya et al., 2014;Shinomiya, Huang, et al., 2019;Takemura et al., 2017). The LGMD1 reacts most strongly to luminance decreases . But finding three neurons entering the medulla and terminating there without branching makes it more unlikely that branches were missed during the tracing process plus as mentioned previously synapses occurred on the neurite itself ( Figure 3).
Early investigations suggested that the TmA1 or 2 would be excited by luminance changes in both directions: darkening or lightening although they would have a preference for darkening (O'Shea & Rowell, 1976;. Candidate TmA responses were described by Osorio (1987) who recorded and stained neurons in the medulla most sensitive to movement which, because of their brief responses and preference for motion he termed "transient and non-linear." Osorio stained three examples that projected from the medulla towards the LOX. Two had arborizations in the ML1-3 that resemble TmA1 and TmAL1 in that they have a plume of dendrites, but also, like TmAL1-3 but not TmA1, they have two small fields of arborization within a single column in ML6-10. Their responses were very similar to the LGMD1: they gave phasic responses, typically a single spike, to any supra-threshold luminance change and a burst of spikes to bar motion; the timing of their responses was precise, typically the standard deviation of the spike latency was only 2 ms; their spiking responses adapted so that they did not give a maintained response to stimulus frequencies above 10 Hz.
They responded equally to light increments and decrements over the range tested. They had small receptive fields compared with the LGMD1 their sensitivity to a 2 spot dropped by 50% 2 -3 from the receptive field center, and by 90% 5 from the center suggesting a receptive field of <20 .
In summary, we were able to 3D-reconstruct two TmA neurons, each in a different fourth instar locust. In the medulla, they interact with similar-shaped neurons to themselves, neurons that are classified as TmA-like, as well as with more tangentially oriented neurites. There are significant differences in branching patterns between them in both the medulla and lobula, indicating that there may be several classes of TmA neurons. Their anatomy is consistent with them being the TmA neurons termed transient and nonlinear by Osorio (1987) because they responded briefly to local luminance changes and to motion in any direction. Further studies will be necessary to confirm this. The LGMD pathway is an example of an escape/hiding behavior that is not mediated by neurons dedicated to optic flow analysis but to object detection and motion analysis, and this contrasts to the giant fiber mediated escape responses to looming optic flow in Drosophila   SCR_000034). With the pipette resistances used, we typically did not find it necessary to apply a retaining current to prevent leakage of dye. After successful impalement, cells were injected with a current of À500 pA for 3 min ($1 Hz; 900 ms on per cycle).
Slices were then incubated for three nights (at 4 C) with primary antibody (mouse anti-ankyrin-G or guinea pig anti-ankyrin-G;

| Immunocytochemical labeling of retinal wholemounts
For immunolabeling of retinal wholemounts, we used a marker (Penol) to place a spot of permanent ink (xylene free) on the dorsal part of the eye before removing it from the orbit. After enucleation, we made a small cut in the dorsal part of the retina and sclera before dissecting the retina from the eye cup. The retina was then flattened by making four radial incisions from the periphery almost to the center and transferred onto the non-gridded surface of a piece of nitrocellulose filter paper (Millipore, cat. number HABG01300). The filter paper with attached retina was then positioned on a piece of folded tissue paper (e.g., Kimwipe) and a few drops of HEPES-buffered extracellular solution was added on top and allowed to soak through. For fixation, a few drops of 4% paraformaldehyde (in .1 M PB) was added on top and allowed to soak through. After repeating this 2-3 times, the filter paper with retina was transferred to a larger volume of 4% paraformaldehyde in .1 M PB and fixed for 30 min at room temperature.
After fixation, the retina was washed six times (10 min each) in .01 M PBS and incubated in antibody incubation solution (identical to that used for slices) overnight at 4 C. The retina was then incubated for four nights (at 4 C) with primary antibodies (mouse anti-ankyrin-G and guinea pig anti-parvalbumin; Table 1) in antibody incubation solution identical to that used for slices, but with .2% Triton X-100. Afterwards, the retina was washed six times (10 min each) in PBS and incubated overnight (at 4 C) with secondary antibodies (goat antiguinea pig coupled to Alexa 488, #A11073 from Thermo Fisher Scientific, diluted 1:1000; goat anti-mouse coupled to Alexa 594, #A11032 from Thermo Fisher Scientific, diluted 1:1000) in antibody incubation solution (identical to that used for slices, i.e., with .2% Triton X-100).
Finally, the retina was washed six times (10 min each) in PBS and mounted in Vectashield between a microscope slide and a precision