Neuronal circuits integrating visual motion information in Drosophila melanogaster

The detection of visual motion enables sophisticated animal navigation, and studies on ﬂies have provided profound insights into the cellular and circuit bases of this neural computation. The ﬂy’s directionally selective T4 and T5 neurons encode ON and OFF motion, respectively. Their axons terminate in one of the four retinotopic layers in the lobula plate, where each layer encodes one of the four directions of motion. Although the input circuitry of the directionally selective neurons has been studied in detail, the synaptic connectivity of circuits integrating T4/T5 motion signals is largely unknown. Here, we report a 3D electron microscopy reconstruction, wherein we comprehensively identiﬁed T4/T5’s synaptic partners in the lobula plate, revealing a diverse set of new cell types and attributing new connectivity patterns to the known cell types. Our reconstruction explains how the ON- and OFF-motion pathways converge. T4 and T5 cells that project to the same layer connect to common synaptic partners and comprise a core motif together with bilayer interneurons, detailing the circuit basis for computing motion opponency. We discovered pathways that likely encode new directions of motion by integrating vertical and horizontal motion signals from upstream T4/T5 neurons. Finally, we identify substantial projections into the lobula, extending the known motion pathways and suggesting that directionally selective signals shape feature detection there. The circuits we describe enrich the anatomical basis for experimental and computations analyses of motion vision and bring us closer to understanding complete sensory-motor pathways.

In brief Shinomiya et al. report the comprehensive identification of neurons in the Drosophila melanogaster lobula plate that receive inputs from the directionally selective T4 and T5 cells, using a reconstruction of 3D electron microscopy data. The two pathways share common synaptic partners, and multidirectional motion signals are integrated to encode new directions.

INTRODUCTION
Flying flies react to complex visual stimuli with acrobatic maneuvers in a matter of milliseconds, 1 which has inspired many to use the Drosophila melanogaster visual system to uncover circuit mechanisms of many neural computations. [2][3][4][5][6][7][8][9] Genetic driver lines enable studies of these computations, [10][11][12][13][14] often testing circuit hypotheses suggested by connectomes based on threedimensional electron microscopy (3D-EM). The fly optic lobe has four major neuropils (lamina, medulla, lobula, and lobula plate; Figure 1A) that are characterized by neurons connecting these structures and layer patterns housing these connections. The diversity of optic lobe neuron types has been well documented using classic methods 16,17 and more recently genetic driver lines for cell-type-specific expression. 12,18,19 The synaptic connectivity of neurons in the lamina, medulla, and lobula 15,[20][21][22][23][24] have been described by 3D-EM reconstructions. Together with functional studies, these data have revealed the detailed circuitry and likely mechanism(s) of motion detection by T4 and T5 neurons. T4 are the ON directionally selective neurons: they encode the direction of a bright edge's motion, while none of their dendritic inputs do. 25 T5 are OFF directionally selective neurons that encode the direction of moving dark edges by integrating inputs onto their dendrites in the first lobula layer (Lo1). 15,26 Both cells have four distinct subtypes, a, b, c, and d, that each project axons, retinotopically arranged, to one of the four layers of the lobula plate ( Figure 1B). 9,16 The evolutionary origin of the lobula plate may relate to the origin of insect flight. 27 The neuropil is best known for containing the dendrites of the ''giant'' lobula plate tangential cells (LPTCs). 17,[28][29][30][31] Many of these cells encode the optic flow-the pattern of visual motion-seen by flies during flight maneuvers such as body rotations, 32 suggesting that they play a key role in regulating flight behaviors. The vertical system (VS) and horizontal system (HS) cells, which have been identified in both larger flies and Drosophila, 16,29 are the best-studied LPTCs. The morphology of these cells in Drosophila was examined using genetic single-cell labeling, 33 and the electrophysiologically measured responses of these cells to visual motion patterns match those of larger flies. 34,35 The 2-photon calcium imaging of T4 and T5 responses in the lobula plate 9 has revealed that these inputs into the lobula plate mainly respond to motion along one cardinal direction: front-toback (Lop1), back-to-front (Lop2), upward (Lop3), and downward (Lop4), near the center of the eye. However, recent work has shown that the population of T4/T5 cells encodes a global pattern that is not consistent with these cardinal directions throughout the eye. 36 Anatomical and physiological data suggest a correlation between an LPTC's visual motion responses and its lobula plate layer pattern, 29,37 although each neuron's selectivity for visual motion patterns can be influenced by network connections. 38,39 Many details of the lobula plate circuitry have not been thoroughly investigated, with the noteworthy exception of two bilayer lobula plate-intrinsic (LPi) cells: LPi3-4 receives input in Lop3 and provides output to Lop4, whereas LPi4-3 sends signals from Lop4 to Lop3. 16,40,41 These cells inhibit their target LPTCs in response to motion in the opposing direction, sharpening the flow-field selectivity of the tangential cells, in a computation termed ''motion opponency.'' 40 Functional studies 4,40,42 suggest this occurs through the integration of excitatory, cholinergic T4/T5 input and inhibitory, glutamatergic LPi inputs by LPTCs, but the synaptic connectivity proposed by this parsimonious circuit hypothesis had not been verified.
The lobula plate also houses processes of other columnar neuron types, including optic lobe-intrinsic neurons, such as Y, TmY (transmedulla Y), and Tlp (trans lobula plate) cells, which connect different optic lobe neuropils, and the LPC (lobula plate columnar), LLPC (lobula-lobula plate columnar), and LPLC (lobula plate-lobula columnar) cells, which are visual projection neurons (VPNs) into the central brain. 4,16,[43][44][45][46][47][48][49] Detailed connectivity information for these neurons is unknown and represents the last piece of the puzzle for the anatomical description of the motion information-processing circuit in the optic lobe. To close this gap, we reconstructed the neurons downstream of T4 and T5 in the lobula plate using a dataset imaged with focused-ion beam-aided scanning EM (FIB-SEM). 15,50 We exhaustively identified and cataloged T4/T5 synaptic partners and investigated complete synaptic profiles of LPi cells and the HS and VS cells. In the process, we identified new cell types as important components of the motion vision pathway and attributed new connectivity patterns to the known cell types, resolving several open questions about lobula plate connectivity.

RESULTS
EM reconstruction of the synaptic partners of T4 and T5 cells in the lobula plate Our FIB-SEM data volume 15,50 includes large parts of the lamina, medulla, lobula, and lobula plate ( Figure 1A), covering regions corresponding to the eye's equator. This volume contains many connected neurons corresponding to common retinotopic coordinates, enabling circuit reconstruction across neuropils. Medulla neurons, including Mi1, Tm1, and Tm2, relay signals to T4 in M10 and T5 in Lo1 ( Figure 1A). 15 The four subtypes of T4 and T5 send outputs to one of the four lobula plate layers (defined to encompass the terminals of groups of T4 and T5 cells; STAR Methods), where they synapse with optic lobe interneurons and neurons projecting to the central brain ( Figures 1B  and 1C). The major findings of our analysis are summarized in a series of figures and data tables (Data S1, S2, and S3; see resource availability for further data availability).
Connectivity of the seed T4 and T5 cells in the lobula plate We reconstructed and then identified many neurons in the FIB-SEM volume, focusing on T4 and T5 cells and their targets. In total, 277 T4s and 277 T5s were identified and at least partially reconstructed. Five cells of each subtype from a retinotopically overlapping region near the volume center were completely traced. 15 In the previous study, we detailed the dendritic inputs of these neurons, and here, we describe the connectivity of these same 40 cells in the lobula plate. All computationally predicted synapses (STAR Methods) of these cells were proofread to identify their presynaptic and postsynaptic partners.
The connectivity of the inputs and outputs of the representative T4 and T5 cells in the lobula plate is summarized in Figure 2A, including all neurons connected with R5 synapses to any of the seed neurons (detailed connectivity data are in Data S1; connectivity summary for R3 synapses is in Figure S1). We found 58 putative connected neuron types (mean of R5 synapses with any T4 or T5), including 16 unidentified fragments (Figure 2A; shown in gray). 74% of these communicate with the same subtype of T4 and T5, resulting in the four clusters in the connectivity diagram, each corresponding to synapses within one lobula plate layer. One noteworthy exception is LPLC2, the only neuron found to receive inputs from all T4/T5 subtypes, corroborating the observation that this cell type integrates spatially patterned inputs to detect looming objects. 4 How are the ON and OFF pathways integrated into the lobula plate? In nearly every case, neurons that are strongly connected to T4/T5 receive inputs from both cells but with a consistent bias for T5 (pooled across all downstream neurons: 45.4% T4 versus 54.6% T5; Figure 2B). This suggests that lobula plate neurons that primarily integrate inputs from T4 and T5 should respond to both bright and dark moving edges. However, some neurons may show differential sensitivity to dark or bright objects due to the greater percentage of inputs from T5 neurons as well as from other inputs. For example, LPLC2 responds more strongly to dark looming stimuli, 4 despite substantial, though T5-biased, inputs from both cells ( Figure 2B).
In mapping computational models onto the circuitry of the motion pathway, the T4/T5 axons are treated as output structures. 40 We find that T4 and T5 axon terminals are the primarily sites of synaptic output but have some inputs: 87% of T4's and 88% of T5's lobula plate synapses are presynaptic. T5 neurons feature higher synapse counts, which is consistent with the connectivity bias for T5 inputs ( Figure 2B). We find T4-T4, T5-T5, and T4-T5 connections within each layer that account for 27.8% of the total inputs to these cells (Data S1). These occur between neighboring axon terminals, and each inter-terminal connection is typically %3 synapses and are reminiscent of the within-subtype connections seen between dendrites of these cells. 15 Several cell types, including Y11 and Y12, provide strong input to T4/T5 terminals (indicated in magenta, Figure 2A), but it is not known whether these neurons inhibit or excite T4 and T5. T4 and T5 neurons with more output synapses tend to have more inputs ( Figure 2C). For example, T4a and T5a had more presynapses and postsynapses than the other subtypes, due ll OPEN ACCESS  Article Figure 2. Connectivity of the seed T4 and T5 cells in the lobula plate (A) The inputs and outputs of representative T4 and T5 cells (five cells per each subtype; see text for details and also Data S1) in the lobula plate were comprehensively identified. The input and output cells were grouped by the cell type, and the inputs and outputs corresponding to a mean of more than five (legend continued on next page) ll OPEN ACCESS to strong connections with several Lop1 neurons, including TmY20, HS, and LPi1-2 cells (Figure 2A; Data S1).
T4 and T5 provide strong inputs to diverse VPNs, including many tangential cells (identified cells named and indicated in green; Figure 2A). The connectivity of the well-known HS and VS cells 29,33 is described in Figure 3. Small-field VPN types (LPC, LLPC, and LPLC cells) 43,45,48,49 are also found with substantial T4/T5 inputs in each layer, and the morphology of connected VPNs is shown in Figure 5. T4 and T5 cells synapse onto bilayer LPi cells in each layer, which are further explored in Figure 4. We also identified many connections between T4/ T5 neurons and other optic lobe-intrinsic neurons, such as the TmY, Y, and Tlp cells (morphology shown in Figure 6) that interconnect different neuropils. For most newly described optic lobe-intrinsic cell types, we provide light microscopy (LM) images as additional validation ( Figures S2 and S3). We conclude with a summary of the core connectivity motifs at this output stage of the visual motion pathway ( Figure 7).

Synaptic connections of the horizontal system (HS) and vertical system (VS) lobula plate tangential cells
The HS and VS cells have been extensively studied [32][33][34][35]53 and are major T4/T5 targets in their respective layers ( Figure 2A). In Lop1, T4a and T5a provide strong inputs to the HS cells (T4a, mean of 44.6 synapses; T5a, mean of 53.4 synapses; Data S1). Each lobula plate houses 3 cells, HSN, HSE, and HSS (north, equatorial, and south). 33 In our imaged volume, we find identifiable fragments of all three cells ( Figures 3A and 3B), with dendrites that are almost purely postsynaptic. Based on the computational predictions, HSN, HSE, and HSS had 6,151, 4,514, and 2,066 postsynaptic densities and 7, 2, and 3 presynaptic T-bars, respectively, with most of the inputs supplied by T4a and T5a ( Figure 3F).
In Lop4, the VS cells receive a large fraction of T4d and T5d's synaptic outputs (Figures 2A and 2B). We identified 10 VS or VSlike cells in Lop4 ( Figures 3C and 3D), exceeding the expected number of VS cells in Drosophila based on earlier genetic labeling studies 33 but consistent with the number in larger flies 31,54,55 and a recent reconstruction in Drosophila. 37 These 10 cells have many common features: primary dendritic processes in Lop4 that are predominantly postsynaptic (typically >95% of the total synapses) and simple connectivity profiles, with $90% of the inputs supplied by T4d, T5d, and LPi3-4 ( Figure 3F; Video S1). Four of the 10 VS and VS-like cells also have dendritic branches in Lop2 ( Figure 3D). VS cells with dendritic arbors outside Lop4 have been previously described. 37,56 By integrating directionally selective inputs in other layers, these neurons should incorporate regional horizontal motion that accompanies body and head rotation around certain axes. 32,56,57 HS and VS cells are connected with a very small set of cell types in the lobula plate, outlining minimal circuit elements that could participate in the nonlinear summation of dendritic inputs. 58 To quantify the input connectivity of these large neurons, we selected small (%300 postsynaptic sites) branches of HS cells (one each from HSN and HSS) and VS cells (two Lop4 branches and one Lop2 branch; each from different cells) and proofread all synaptic sites. Figure 3E shows an HS branch and a Lop4 VS branch, and Video S1 shows a VS branch and its input neurons. The summary of these connectivity analyses ( Figure 3F) shows that $80% of the inputs are supplied by T4/ T5. The HS (Lop1), VS (Lop4), and VS (Lop2) branches receive 7.14%, 18.9%, and 10.7% of input synapses from bilayer LPi cells, respectively ( Figure 3F; Data S2). Intriguingly, the Lop2 VS branch receives inputs mainly from T4b, T5b, and LPi1-2 cells, suggesting that it indeed receives back-to-front local motion signals.
How are the inputs from T4, T5, and LPi neurons organized onto the dendrites of these LPTCs? The spatial distribution of T4a, T5a, and LPi2-1 inputs onto HS cells features synapses throughout the thickness of Lop1 without any apparent organization or substructure ( Figures S4A and S4A'). Likewise, VS and their input neurons display a patchy but mostly uniform synapse distribution in Lop4, whereas the Lop2 VS branch synapses are only found in the distal half of the layer ( Figures S4B and S4B'). The individual input synapses from T4, T5, and bilayer LPi onto individual HS and VS branches also feature a patchy distribution but with no obvious pattern in any of the three layers ( Figure S5, fragments same as those analyzed in Figure 3; Data S2). Overall, the connectivity between T4/T5, LPis, and the giant LPTCs is very similar across these layers ( Figures 3E and 3F). This relatively simple connectivity structure strongly supports the expectations of the functional studies of HS and VS-they mainly integrate inputs from directionally selective T4/T5 cells that are further sharpened by motion opponent inputs from LPi neurons. 40 Connectivity of the bilayer lobula plate-intrinsic (LPi) cells We identified four bilayer LPi neuron types as major T4/T5 targets ( Figure 2A). A previous study described LPi3-4 and LPi4-3 synapses per T4 or T5 cell are shown in the diagram (see Figure S1 for the same diagram but with the threshold lowered to three synapses per T4 or T5). The thickness of the arrows indicates the average number of synapses per T4 or T5 cell. Each rectangle indicates a cell type: colored rectangles correspond to uniquely identified cells, and gray rectangles represent neurons we could not uniquely identify due to incomplete reconstruction. For unidentified neurons, the main innervated layers are shown in italic letters. For example, Lop(1)34 means that the fragment has major arbors in Lop3 and Lop4 and minor arbors in Lop1. Several unidentified neuron fragments share a common designation (e.g., Lop3) but are likely to represent independent neurons (so far as can be determined within the limited data volume). LPi are lobula-plate-intrinsic cells; the TmY/Y/Tlp neurons connect the optic lobe neuropils, and LPTC and LPLC/LLPC/LPC cells are visual projection neurons (VPNs) that send outputs to the central brain. T4 and T5 receive far fewer synaptic inputs in the lobula plate than they provide output synapses. The cell types with more than five input synapses to T4 or T5 are highlighted with magenta frames. (B) Average numbers of output synapses from single T4 and T5 per postsynaptic cell type. Neurons are color coded by the layer where they receive inputs from T4 and T5. Generally, the outputs from T4 and T5 (and therefore inputs to their target neurons) are approximately evenly integrated by the postsynaptic cells, with a slight bias for T5. All the named neurons receiving more than an average of 20 synapses from both T4 and T5 are labeled (LPLC2 is labeled for all four layers). The dashed lines indicate a 25% and 50% difference from equal numbers of output from T4 and T5 to any target cell type. (C) Total numbers of input and output synapses of the representative T4 and T5 cells. Autapses (self-synapses) and synaptic contacts with glia are excluded from this quantification. Averaged synapse numbers of each cell type (five individual neurons per each cell type) are indicated as gray crosses. and speculated about the existence of all four types. 40 Here, we have reconstructed and identified LPi1-2 and LPi2-1 that bridge Lop1 and Lop2, confirming these predictions, although we are unable to describe their complete morphology. We found a strong candidate for LPi1-2 using LM ( Figure S2A), which suggests that LPi1-2, and perhaps also LPi2-1, may be considerably larger than LPi3-4 and LPi4-3. Confirming this proposal will require extensive reconstruction in a larger EM volume. All four LPi neuron types innervate neighboring layers, with a stereotypic synapse distribution (Figure 4, left). Each cell type has postsynaptic sites in one layer and presynaptic T-bars in the adjacent layer. At least 2/3 of the inputs are from layer-specific T4/T5 cells, whereas the outputs are shared by many neuron types (Figure 4, right; Data S3). The proportion of output synapses relative to all synaptic connections is very similar in LPi1-2, LPi2-1, and LPi4-3 and is close to 50%, whereas this number is lower in LPi3-4 (Data S3), implicating functional or developmental differences.
The LPi3-4 and LPi4-3 cells are glutamatergic, 40,46 and these cells provide directionally selective inhibition to target neurons. 40,42 Based on their similar morphology and connectivity, LPi1-2 and LPi2-1 are also likely inhibitory. This small circuit supports the Mauss et al. 40 mechanism: bilayer LPi cells integrate T4/T5 inputs in one layer and inhibit postsynaptic neurons integrating oppositely tuned T4/T5 signals in the adjacent layer, implementing motion opponency. The $1/3 of LPi inputs provided by various cells other than T4/T5 suggest that the lobula plate circuitry is more complicated, and perhaps more flexible, than the circuit models consider.
Lobula plate visual projection neurons (VPNs) that integrate T4 and T5 inputs In addition to HS and VS cells (Figure 3), we identified other VPNs as T4/T5 targets (Figures 2 and 5). In this study, we focused on identifying and quantifying T4/T5 target neuron connectivity rather than describing the complete synaptic profiles of the VPNs.
T4 and T5 connect with columnar VPNs, smaller cells that as a population cover large parts of the lobula plate. These cells belong to three main groups (LPC, LLPC, and LPLC) that are distinguished by innervation patterns in the optic lobe and axonal path to the central brain (further explained in Figure 5 legend). Based on their arbor sizes in the lobula plate and T4/ T5 and LPi inputs, these cells likely respond to visual motion within small patches of the fly's field of view. We distinguished two LPC types and three LLPC types based on lobula plate layer patterns (Figures 5A-5E), in agreement with LM analyses. 49 LPC1 ( Figure 5A) receives inputs from T4b and T5b. This anatomy suggests that these cells integrate back-tofront motion signals, which has been confirmed by calcium imaging. 49 LPC2 ( Figure 5B) receives T4c/T5c inputs (Figure 2A) and is expected to encode upward motion. LLPC1 ( Figure 5C), a VPN responsive to front-to-back visual motion, 49 has dendritic arbors in Lop1 and Lop3, with stronger T4/T5 input in Lop1 (Figure 2A). The process in the lobula appears to be mainly presynaptic ( Figure 5C). LLPC2 and LLPC3 are similar cells with T4/T5 input in Lop3 and Lop4, respectively ( Figures 5D and 5E). LPLC1 and LPLC2 cells 43,45 are notable for receiving T4/T5 inputs in multiple lobula plate layers: T4/ T5 a, b, c, and d for LPLC2, in agreement with the described mechanism of looming sensitivity, 4 and T4/T5 b and d for LPLC1 (Figures 2A, 5F, and 5G).
T4 and T5 neurons connect with LPTCs, some of which we matched to known neurons, but in other cases, we name them based on their layer innervation patterns (Figures 5H-5O). As most LPTCs are morphologically unique, these cells could be matched to LM images or other EM reconstructions. 37,59 One cell we matched based on its unique morphology is LPT13a (EM, Figure 5J; LM, Figures S2D and S2D'). The dorsal centrifugal horizontal (DCH) cell ( Figure 5P) is a unique LPTC that is predominantly presynaptic to T4 and T5 (Figure 2A): 15.3% of T4a inputs and 12.7% of T5a inputs (excluding synapses between T4/T5 terminals) are from DCH, by far the largest input to T4/ T5 from a single LPTC. The terminals of DCH cover the dorsal half of Lop1, whereas the homologous VCH cell covers the ventral half. 37,52,60 The CH neurons innervate the ipsilateral inferior posterior slope in the central brain, are GABAergic, 52,61,62 and likely inhibitory. Although we did not find VCH (due to the restricted volume), our data suggest that these two cells are the only major LPTCs that feed signals from the central brain to T4a/T5a (Data S1).
H1 is a heterolateral LPTC directly connecting both lobula plates ( Figure 5Q) 29 and is sensitive to ipsilateral back-to-front visual motion, similar to H2. 39,51 We found profiles that likely correspond to the proximal and distal terminals of both H1 cells. The proximal terminal is predominantly postsynaptic and confined within Lop2, whereas the putative distal terminal branch is presynapse rich, with boutons mainly in Lop1 and Lop2. The proximal terminal receives inputs from T4b and T5b (mean is $4.8 synapses per terminal from both T4b and T5b; Figure S1). While this connection strength is lower than many T4/T5 target neurons, H1 is expected to integrate from many T4b/T5b throughout Lop2, which is consistent with the described motion preference. 29,39,51,60 The distal terminal of H1 has limited synaptic contacts with T4 or T5 cells (only accounting for $0.1% of H1's predicted outputs).
The H2 cell, another identifiable LPTC that is known from larger flies, has dense neuronal processes confined to Lop2 (Figure 5R), is a strong T4b/T5b target (Figure 2A), and projects to the inferior posterior slope in the contralateral brain hemisphere. 31,52 H2 branches in Lop2 feature mixed presynaptic and postsynaptic terminals ( Figure 5R, inset), as suggested by genetically driven synaptic markers. 52 The ''hemibrain'' connectome dataset shows H2 as the strongest input to DCH and VCH 48 in the central brain, thus contributing to processing motion information from both eyes.
Optic lobe-intrinsic neurons that integrate T4 and T5 inputs T4 and T5 target optic lobe-intrinsic cells other than the bilayer LPi neurons, including several types of LPi, TmY, Y, and Tlp neurons ( Figure 6). We identified both known and new optic lobeintrinsic cell types as T4/T5 targets. For most of the new cell types in this group, we further confirmed their morphology with LM matches ( Figure S3).
Am1 is a single, large, amacrine-like neuron innervating the medulla, lobula, and lobula plate with tree-like arborization. 28,50 Am1 receives inputs from T4b/T5b in Lop2 (Figures 2A and 6A) and has significant synaptic contacts with some LPTCs. The predicted synapses contain strong inputs from DCH and contralateral H1 and outputs to DCH and HS cells. Based on these connections, we expect that Am1 is inhibitory (unlikely to excite HS cells in response to ipsilateral T4b/T5b input) and participates in a bilateral circuit that integrates horizontal motion signals from both eyes 38,39,51,63 ( Figure 7C).
We find several putative LPi and LPi-like neuron types ( Figures 6B-6F) that all differ from the bilayer LPi types and further illustrate the diverse neuronal composition of each layer. A large cell we tentatively named LPT/LPi2a receives the strongest inputs from T4b and T5b among all neurons in our data (Figures 2A and 6B). LPT/LPi2a has a similar but distinct morphology from the bilayer LPi2-1 cell in the lobula plate, with main branches containing presynapses and postsynapses in Lop2 with additional sparser processes in Lop1. While T4/T5 supply >80% of LPi2-1's input, they supply <50% for LPT/ LPi2a, suggesting that it participates in circuits with more elaborate connectivity than the main bilayer LPis ( Figure 7A). Our best candidate for an LM match is a VPN with a central brain projection ( Figure S2B). LPi2b is another large Lop2 cell that appears to span the entire lobula plate, but with a more restricted layer pattern and fewer T4/T5 inputs ( Figures 6C and S2C). LPi34-12 ( Figures 6D and S3I) is named for its layer pattern, as it receives T4/T5 input in both Lop3 and Lop4 and has output synapses in Lop1 and Lop2 (Figure 2A). This cell likely implements an undescribed interaction between motion detected along different directions.
TmY cells have cell bodies in the medulla cell body rind and terminals in both the lobula and lobula plate (Figures 6G-6L). TmY4, TmY5a, TmY14, and TmY15 have been previously described, 15,16,22,23 whereas TmY16 and TmY20 are reported here for the first time and confirmed with LM matches (Figures S3A and S3B). Among all T4a/T5a targets, TmY20 has the highest number of inputs (Figure 2A; Data S1). Unlike most TmY cells, the TmY20 neurite in the medulla apparently lacks synapses (reminiscent of LPi3-4, which also lacks synapses in the medulla 40 ). TmY20 has mostly presynaptic terminals in lobula layers Lo5 and Lo6 ( Figure 6L), suggesting that this neuron relays front-to-back motion information to lobula neurons. The other TmY cells have extensive arborizations outside the lobula plate, and a full inventory of their connectivity may be required for detailed predictions about their role in motion processing. Y cells (Figures 6M-6O) are columnar neurons with cell bodies in the rind posterior to the lobula plate that innervate the medulla, lobula, and lobula plate. 16 Tlp cells (Figures 6P-6S and S3E-S3H) are similar to Y cells but lack a medulla branch. We identify one known (Y3) and two previously undescribed Y neurons (Y11 and Y12) as T4/T5 targets and confirm their morphology using LM (Figures S3C and S3D). Tlp, Y, and TmY cells all provide paths for relaying different subsets of T4/T5 outputs to the lobula ( Figure 7D). Y11 and Y12 are notable for integrating T4/T5 input from different layers: Y11 from Lop1 and Lop3 and Y12 from Lop1 and Lop4. The two cells are otherwise morphologically very similar, with boutons in the same medulla and lobula layers. Both cell types have presynaptic and postsynaptic contacts with T4 and T5 (Data S1), integrating their signals in their respective layers. Since Y11 synthesizes front-to-back (Lop1) and upward (Lop3) motion signals and Y12 combines front-to-back and downward (Lop4) motion signals, the two cells are likely to each encode a preferred motion direction in between the preferred directions of their input T4s and T5s ( Figure 7B).

DISCUSSION
The giant tangential cells of the fly lobula plate have received considerable interest for decades, 31,64,65 but the circuit description of this ''final'' optic lobe stage of the motion pathway has been rather incomplete. In this study, we used EM reconstructions to inventory the synaptic partners of T4 and T5 neurons with completeness unmatched by other approaches. By focusing on the neurons with an indispensable role in motion information processing-T4, T5, and the LPi neurons-we find a stunning diversity of synaptic partners in the lobula plate and establish a substantial catalog of neurons implicated in motion processing for future functional investigations. Our work reveals an elaborate architecture for processing visual motion, with several major findings: (1) all the lobula plate neurons downstream of T4 and T5 integrate from both cell types, with a slight bias for T5 inputs; (2) each layer houses a unique ensemble of downstream neurons while sharing a core circuit motif composed of T4/T5, a bilayer LPi cell, and output VPNs; (3) the T4 and T5 axon terminals receive substantial inputs that could refine the local motion direction tuning; (4) new circuit elements Only the LPi3-4 cell is completely reconstructed, while the other cells are only partially reconstructed since single neurons cover larger lobula plate areas than the imaged data volume. A candidate light microscopy match for LPi1-2 ( Figure S2) suggests the possibility that the LPi1-2 reconstructions (and perhaps also the similar LPi1-2 fragments) may be parts of one or a few large cells. In the right two panels, the ratios of the input and output synapses are shown for each indicated cell type. These data are based on a single selected branch for each cell type for which the presynaptic and postsynaptic connected neurons were identified wherever possible. Cell types occupying less than 2% of the total input or output synapses are not shown and are included as ''other.'' Several tangential elements that have synapses with the LPi cells were only partially reconstructed due to the restricted data volume. These fragments of considerable size are grouped as ''unidentified LPis and LPTCs.'' Data summary is based on Data S3. Article (legend on next page) ll OPEN ACCESS combine motion signals for different directions, including the Y11 and Y12 cells; and (5) many neurons that convey motion signals from the lobula plate to the lobula, implicating lobula circuitry with a more significant role in motion processing.
We found that all lobula plate neurons that are strongly connected to T4 and T5 axon terminals integrate these inputs in the same layers ( Figure 2B). We find a modest but consistent asymmetry, where on average, neurons receive $20% more synapses from T5 than T4. This may indicate that T5's OFF signals are weighted more heavily than T4's ON signal. Future studies examining the entire lobula plate will be needed to explore broader patterns of T4/T5 connectivity, such as different local biases in different parts of the field of view or for different target neuron types. Nevertheless, we do not find any output neurons that are purely T4 or T5 selective. This is a significant finding that implies, at least for the motion pathway, that the ON/OFF separation is an internal feature. At the output stages of the pathway, ON-and OFF-motion signals are combined onto all prominent lobula plate targets.
Most of the identified LPTCs receive T4/T5 inputs in single layers; however, several VS cells (Figure 3), three columnar VPNs (LLPC1, LPLC1, and LPLC2), and some optic-lobeintrinsic neurons (e.g., LPi34-12) receive T4/T5 inputs in multiple layers (Figures 2A, 5, and 6). These connectivity patterns suggest that most LPTCs receive large-field motion information from T4/ T5 representing the preferred direction of one of the four layers, whereas small-field neurons may integrate signals from multiple layers and as a population could transmit more complex motion information to their downstream neurons. The best-explored example of this is LPLC2, whose looming sensitivity was attributed to T4/T5 and bilayer LPi inputs in all four layers, 3,4 a hypothesis that this study has substantively confirmed.

Bilayer LPi cells
The four bilayer LPi cells have a common distribution of synapses, with T4/T5 inputs in one layer and output synapses in a neighboring layer, where they presumably inhibit most or all of the neurons that also receive excitatory T4/T5 inputs in that layer (Figures 2A, 4, and 7A), implementing motion opponency. 40 Although these cell types likely serve similar functions in motion processing, there are also clear anatomical differences. The cell bodies of LPi1-2, LPi2-1, and LPi4-3 are in the lobula plate cortex, whereas LPi3-4s are in the medulla cortex 15,16,40 and, therefore, likely derive from different precursor cells. EM and LM data suggest substantial size differences among the bilayer LPis, with LPi3-4 likely being the smallest arborization and individual LPi1-2 cells arborizing across much of the lobula plate ( Figure S2A). Since the spatial coverage of individual LPi neurons differs between the types, the spatial integration of opponent signals may differ between layers for reasons that are unclear. All 4 bilayer LPi types provide inputs to both large-and small-field neurons (e.g., LPi1-2 is upstream of H2 and also LPC1; Figures 4, 5A, and 5R), so there is no simple relationship between the apparent receptive field size of the different LPi neurons and the receptive field sizes of their major targets. These differences raise questions about the evolution of the bilayer LPi cells. Are these LPis derived from a shared ancestral cell type, for example, via duplication, that later substantially diverged in some of their anatomical properties, or did the antiparallel inhibition mediated by the bilayer LPis evolve independently in different layers?

Y cells encoding oblique motion directions
We discovered that Y11 and Y12 integrate motion information in two layers and thus likely synthesize a preferred tuning for a new, oblique direction of motion (Figures 2A and 7B). These neurons effectively fill two gaps between the four directions represented by T4/T5 subtypes. Y11 and Y12 are also major sources of inputs to T4 and T5 neurons (Data S1), suggesting that they may alter the local directional preference of T4 and T5. Clarifying the role of these Y cells will require establishing whether they are excitatory or inhibitory (inhibition via glutamate is a strong possibility, as T4/T5 axons express GluCla , a glutamate-gated chloride channel, 66 yet are not strong targets of the bilayer LPi neurons; Figure 4; Data S1). These recurrent connections could provide a possible mechanistic explanation for the functionally identified subtypes of T4/T5; 36 however, this recent study found subtypes in layers 1 and 2, whereas Y11 and Y12 collect outputs and provide inputs to layers 1, 3, and 4. Both neurons combine a vertical motion signal with front-to-back motion, but we did not find complementary neurons for the oblique motion directions integrating Lop2/back-to-front motion. Beyond this role in the lobula plate, these Y cells are presynaptic in the lobula and medulla ( Figures 6N, 6O, and 7D). Identifying the targets of Y11 and Y12 will be an important goal of future connectomes.
Expanding the horizontal motion detection circuit with new cell types Our detailed analysis suggests that several new connections should be added to the existing models of binocular integration of rotational optic flow derived from work on blowflies. 38 The Am1 cell, which receives inputs from ipsilateral T4b/T5b and contralateral H1, likely combines optic flow across both eyes. H1 expresses a marker for glutamatergic neurons. 52 In Drosophila, glutamate could function as either excitatory or inhibitory, whereas in the blowfly, H1 seems to provide excitatory   39 T4b and T5b detect back-to-front movement and, via (putative) inhibitory LPi2-1 cells, suppress the activity of Lop1 neurons, including HS cells ( Figure 7C). Am1 may represent two more pathways for suppressing the activity of HS cells in response to back-to-front motion inputs directly and through DCH, which is also electrically coupled with HS in Calliphora. 38  (Figures S2 and S3). Based on these matches, LPT/LPi2a may be a type of VPN with a central brain projection. Presynapses and postsynapses are shown in magenta and gray, respectively. Scale bars, 30 mm. Figure 7. Summary of the motion pathway circuitry revealed by the lobula plate reconstruction (A) The primary connections between the T4/T5 cells, bilayer LPi cells, and output VPNs. FTB, front-to-back; BTF, back-to-front. The LPi cells, indicated in blue, are likely all inhibitory cells. 40 Each layer has VPN outputs that are predicted (or known for the few supported by functional studies) to encode motion with the same directional selectivity as the T4/T5 subtypes in that layer. Some VPNs, like the VS cells of Figure 3, integrate inputs from multiple T4/T5 subtypes in different layers. (B) The Y11 and Y12 cells and their lobula plate inputs. The two Y cell types receive excitatory inputs from T4 and T5 in two layers and (putative) inhibitory inputs from T4 and T5 neurons in the other layers, via bilayer LPi cells. By integrating these inputs, it is expected that these neurons become most sensitive to the direction of overlapping sensitivity of their inputs, and thus, Y11's preferred direction would be oblique-upward motion, and Y12 would prefer oblique-downward motion. (C) The bilateral circuitry comprised horizontal-motion-sensitive neurons, including H1, H2, DCH, and Am1 cells, integrating motion from both eyes. Connections within the optic lobe are based on the observation of this dataset, whereas contralateral projections and synaptic contacts in the central brain are also based on previous studies and datasets, including Farrow et al. 38,51 and Scheffer et al. 48 Connections from T4b/T5b to H1, indicated with asterisks (*), are shown in Figure S1. H1 is considered to be a glutamatergic cell, 52 and it is not known whether the signal from H1 is excitatory or inhibitory. Since electrical synapses cannot be directly observed in the FIB-SEM dataset, the indicated connections are based on prior work and the physical proximity of the axons of these cells in this dataset. The diagrams do not exhaustively list the inputs and outputs of the shown neurons. (D) Neurons relaying T4/T5 outputs to the lobula. The paired parallel lines indicate T4/T5 inputs in the lobula plate, whereas the dots represent the locations of the output (presynaptic) terminals in the lobula and the medulla. Input (postsynaptic) terminals in the lobula and the medulla, as well as terminals not connected to T4/ T5 in the lobula plate, are omitted. The central brain projections of TmY14 and the LLPC cells are shown as broken lines.
Both excitatory and inhibitory motion signals are integrated at different scales: the scale of bilayer LPi neurons and the CH neurons, and over the entire field of view by combining contralateral optic flow transmitted by H1 and H2 ( Figure 7C). These proposals also highlight a limitation of connectomic studies for predicting the selectivity of these neurons. We have mainly described VPN selectivity as a function of their feedforward T4/ T5 input; however, the resultant selectivity of any neuron to visual motion patterns depends on feedback connections from other neurons in the lobula plate, as well as inputs received in the central brain. 38,39 Connectomic studies establish chemical synapses but do not generally reveal electrical synapses (gap junctions). Gap junction predictions from light microscopic analyses 67 often require further confirmation, and determining how electrical connections shape the functional logic of these circuits will require detailed neurophysiological studies.
Multiple neuron types convey T4/T5 signals to specific lobula layers We found that T4/T5 are connected to several neuron types that likely relay these signals within the optic lobe. For example, TmY20 cells ( Figure 6L) receive the largest share of T4a/T5a output synapses (Figure 2A). Although the standard circuit models of the motion pathways, which comprise T4/T5, LPTCs, and bilayer interneurons ( Figure 7A), have remained compact, evidence for additional, strong pathways reveal that motion signals are integrated into a broader set of downstream circuits than previously suggested. Many T4/T5 downstream cells, including Tlp, LLPC, and TmY neurons (Figures 2, 5, and 6), project to the lobula, where they mainly target layer Lo4 (Figure 7D). The circuits of the lobula, outside of T5 inputs in Lo1, have been scarcely examined. Motion signals from the lobula plate should significantly contribute to visual processing in the lobula, and many VPNs could inherit motion signals from the lobula plate without any input sites there. The complete description of these pathways and their extended circuits will require an EM dataset that covers nearly the entire brain.
Toward the complete reconstruction of sensory-tomotor pathways The connectivity profile of T4/T5 neurons in the lobula plate details a large missing part of the motion pathways, the link between motion detection and output neurons of the optic lobe. With this stage finally reconstructed, the motion pathway from the photoreceptor cells to the central brain can now be traced neuron by neuron by combining results across EM reconstructions. 15,[20][21][22]24,68 Many of the VPNs we reconstructed are also identified in the hemibrain dataset 48 that contains much of the central brain, enabling the comprehensive identification of downstream circuits to extend the described pathways further. The strength of our study-the detailed analysis of a central region of the lobula plate-is also a limitation. The diversity of cells downstream of the lobula plate is certainly larger than we described here, and we hope that future studies in larger EM volumes 59,69 will cover the entire lobula plate. Many of our new discoveries suggest a more integrative picture of optic lobe processing, where the lobula plate is no longer seen as the sole substrate for motion processing but rather is understood to organize ON and OFF directionally selective signals for a variety of as-yet-unexplored roles in visually guided behaviors.

STAR+METHODS
Detailed methods are provided in the online version of this paper and include the following: