The comprehensive connectome of a neural substrate for ‘ON’ motion detection in Drosophila

1A classic problem in visual neuroscience, how the visual system detects image motion, has recently undergone resurgent interest in the fruit fly Drosophila. The problem has long been analysed using mathematical models that compare two adjacent viewpoints, of which the Hassenstein-Reichardt (HR) elementary motion detector (EMD;; Figure 1A) or Barlow-Levick (BL) model (; Figure 1B) are both widely canvassed. Both models compare two visual inputs separated in time and space and simulate motion detection well. However, the exact neural mechanism underlying the detection of direction by motion-sensing cells is still understood only imperfectly.

2Visual processing in flies occurs consecutively in four retinotopically organised neuropile relays beneath the photoreceptor layer (Figure 1C). Each neuropile comprises an array of repeating modules corresponding to the hexagonal lattice of ommatidia in the overlying compound eye. Visual signals generated by photoreceptors, R1-R6, are transmitted to two matched classes of large paired interneurons, L1 and L2, and a third, L3, in the first neuropile, or lamina (Figure 1C). Like ON and OFF bipolar cells in the vertebrate retina, the visual pathways split at the first synapse, with L1 and L2 as the main inputs, to separate ON- and OFF-edge motion pathways, respectively.

3ON-edge signals are then locally computed on the dendritic arbours of columnar cells, T4, in the most proximal stratum M10 of the second neuropile, or medulla. These cells respond to visual motion in a directionally-selective manner, and have four subtypes, T4a-T4d. Each subtype is tuned to one of four cardinal directions of stimulus motion, front-to-back (T4a), back-to-front (T4b), upward (T4c) and downward (T4d), approximately one each per unit column of the medulla. Each in turn projects its terminal to a specific stratum, one of four in the motion-sensing fourth neuropile, the lobula plate (Figure 1C). There, T4 signals from hundreds of columns sum to generate wide-field motion responses in target, directionally-selective tangential cells, aided by various lobula plate interneurons.

4How do T4 cells compute the direction of motion? A recent connectomic study using serial-section EM (ssEM) identified two medulla cell types, Mi1 and Tm3, as the direct pathways between L1 and T4. Detailed analyses revealed that the weighted centre of the Mi1-mediated receptive field is displaced from that of the Tm3-mediated subfield, and that the direction of displacement between the two subfields aligns in ~70% of cases with T4’s preferred direction. Electrophysiological recordings demonstrate a temporal offset between the responses of the two cell types: Mi1 exhibits a peak light response later than that in Tm3, although the time difference (~15 ms) is probably too small to account for T4’s measured temporal frequency optimum of ~1 Hz. The direction of spatial offsets from Tm3 to Mi1, generally consistent with T4’s preferred direction, matches the requirements of non-delay and delay input channels of a BL-type detector.

5Despite this evidence, there are multiple indications that Mi1 and Tm3 inputs are not the only players in T4’s motion detection. First, a BL-type detector requires Mi1 and Tm3 inputs to T4 to have different signs (Figure 1B); this is not supported by available neurotransmitter information (see below). Second, whereas each T4 dendrite is orientated primarily in a direction aligned with its directional preference, Mi1 and Tm3 synapses on those dendrites are distributed with substantial overlap. Third, although recent experiments confirm a general role for Mi1 in ON-motion detection, they suggest that Tm3 is required for responses only to fast-motion stimuli, so that other inputs must be active under slow-motion stimulus conditions.

6Here, we use a superior EM technique to reconstruct T4 cells and their presynaptic partners, and provide a comprehensive anatomical view of this critical step in fly motion vision. Our results suggest new candidate circuits for motion detection, including previously unidentified inhibitory pathways.

7To reconstruct T4’s input circuits comprehensively we used a new method of focussed ion-beam milling scanning electron microscopy (FIBSEM), which enabled a more comprehensive dense connectome with greater accuracy (see Methods). Briefly, we imaged a ~40 × 40 x 80 µm volume at an isotropic resolution of 10 nm per voxel, annotated ~53,500 presynaptic sites and ~315,500 postsynaptic sites, and reconstructed >900 neuron bodies including >60 T4 cells. The dataset volume comprised seven medulla columns: a central ‘Home’ column and its six immediate neighbours (Figure 1D). These seven columns have been partially analysed elsewhere; here we focus on the inputs to T4.

8Since a T4’s dendrites spread across approximately seven columns, the dendrites of the T4 cells in the Home column should be completely contained in the reconstructed volume. We found Mi1-Home contacted >20 T4 cells, forming 2–43 synapses onto each. Among these, only four T4s received >35 such contacts. They included one of each subtype, T4a-T4d, as identified by the distinctive orientation of their dendrites (Figures 2A and 4C–4E).

9Next, we sought other cell types with input to the T4s. Reconstructed cell shapes revealed eight medulla neuron types that account for inputs to the T4 dendrites (Figure 2A). We confirmed previously reported inputs from neurons Mi1, Tm3, C3, Mi4, Mi9 and T4 itself. Comparing the remaining reconstructed cells with light microscopy data further identified a previously undescribed TmY neuron, which we call TmY15 (Figure 2A and B; Figure 2—figure supplement 1), and neurites of a new large tangential cell, which we call CT1 (Figure 2A and C–E; Figure 2—figure supplement 1). CT1 spans the opposing faces of medulla stratum M10 and lobula stratum Lo1 that contains dendrites of T4’s counterpart, T5, as a single giant neuron (Figure 2C). Its terminal arbours show a strikingly regular pattern in each medulla and lobula column (Figure 2D; Figure 2—figure supplement 1), which is also confirmed by EM reconstructions (Figure 2E). In each column, CT1 receives a few synaptic inputs from Mi1, Mi9 and all T4 subtypes (Figure 2F).

10We next examined the number of T4 inputs from each of these cell types (Figure 3). Each T4 receives >200 synaptic inputs in total, and Mi1 and Tm3 cells are indeed the largest contributors with 35.0% and 16.1% of the total, respectively (Figure 3). Endorsing improved completeness of the new FIBSEM reconstruction, total synaptic inputs from Mi1 and Tm3 to the four T4 subtypes (116, 109, 113, and 104, respectively), totalled 33% more than in the previous ssEM connectome (97, 69, 85 and 81).

11Additional unicolumnar cell types, Mi4, Mi9 and C3, all provide fewer synaptic inputs to the T4s than Mi1 and Tm3. However, we identified more than previously reported, primarily because many of these synapses are located in the periphery of a T4’s dendrites and were therefore excluded from an earlier account, which sampled densely only a single column. Secondarily, we benefit from greater image quality of FIBSEM. Their wider contribution is nevertheless significant: 13.3% (Mi9), 6.4% (Mi4) and 3.9% (C3), in aggregate ~24% of the total (Figure 3). The new FIBSEM reconstruction identified many more connections than the previous ssEM study which found 4 Mi4 inputs, 6 Mi9 inputs, and 5 C3 inputs.

12The newly identified TmY15 and CT1 cells together account for another ~15% of T4 inputs (Figure 3). In addition, a T4 cell also connects reciprocally to other T4s in neighbouring columns. Only very few synapses onto T4s were from other cell types or could not be identified: an Mi10 neuron makes a few input synapses (~0.6%), with the remaining 3.1% from several unidentified profiles that left the reconstructed volume, each contacting a T4 at ≤3 synapses (Figure 3B). At least some of these unidentified synapses are likely to be from branches of other T4s that reside outside the imaged region. We therefore reason that our dense FIBSEM reconstruction captures all input cell types to T4 and is now reliably definitive.

13All models of motion detection require comparison between inputs that view different points in visual space. To judge the potential relevance of T4’s input neurons to generating motion direction preference, we analysed the distribution of synapses over the T4 dendrites (Figure 4C–E) and, for cell types that could be unambiguously assigned to single columns, across visual columns (Figure 4A).

14Each T4 subtype receives input from multiple Mi1s, with the strongest from the Home Mi1 and weaker inputs from surrounding Mi1s (Figure 4A); these spread along the shafts of the T4 dendrites (Figure 4C). Likewise, Tm3 synapses appear to distribute evenly over the T4 dendrites (Figure 4D). However, while Mi1 cells can be readily assigned to specific columns, Tm3s occupy more variable positions in the column periphery, and cannot be so assigned (Figure 4—figure supplement 1). Calculating anatomical receptive fields, is also limited because wide collaterals of Tm3 cells in strata M1 and M5 (Figure 2A) extend beyond the 7-column region. The reconstructed Tm3s therefore receive some input from columns not contained in the FIBSEM image stack used here and which we therefore could not trace. This limits the accuracy with which potential spatial offsets between Tm3 and other cell types could be determined (see Figure 4—figure supplement 2A and Discussion).

15Inputs to T4 from Mi4 and Mi9 are noticeably segregated from each other (Figure 4A and C). Significantly, Mi9’s input synapses always cluster at the tips of the T4 dendrites while Mi4’s inputs localize to their bases (Figure 4C). Inputs from other T4 cells also clustered towards T4’s dendrite tips (Figure 4E). The T4-T4 synapses showed striking subtype selectivity connecting with cells with the same direction preference, apparently qualifying a group of T4s to signal each other about a particular stimulus direction. This subtype selectivity was observed without exception in every T4-T4 contact. The retinotopic positions of the presynaptic T4s lie relative to the postsynaptic T4 in a direction opposite to the cell’s preferred direction (Figure 4E).

16Like Mi4, C3 and CT1 also made synapses at the bases (Figure 4C). Even though CT1 is a wide-field neuron (Figure 2C; Figure 2—figure supplement 1), its connections are displaced from the Mi1 centre in a manner very similar to that of Mi4 and C3 (Figure 4B), suggesting that each columnar arbour may serve a local computational role. We therefore counted each column’s CT1 arbour as a separate unit per column (Figure 4A). Finally, TmY15 is presynaptic to T4s with a branching arbour extending more widely than the 7-column area in stratum M10 (Figure 2B; Figure 2—figure supplement 1). We have reconstructed three TmY15 cells in the 7-column volume, of which two are only partial, suggesting that this cell type is distributed less frequently than one cell per column. Its synaptic inputs were rather uniformly distributed over T4’s dendrites (Figure 4D).

17To confirm the stereotypy of the different columnar distributions of T4 inputs, we extensively traced another set of T4s in a neighbouring column; again using inputs from Mi1 and dendritic morphology to identify four T4 subtypes associated with this column. These analyses supported the results for the Home column T4 cells (Figure 4—figure supplement 3).

18To gain insight into the possible signs of T4’s inputs, we next sought information on their neurotransmitters. Immunolabelling studies have suggested that C3 is GABAergic while Mi1 expresses choline acetyltransferase, suggesting it is cholinergic. Our immunolabeling results using anti-GAD1, anti-ChAT and anti-VGlut confirm these results for C3 and Mi1 and further suggest that GABA, acetylcholine and glutamate are the respective transmitters for Mi4, Tm3 and Mi9 (Figure 5). These results are also supported by cell-type-specific transcript profiling studies (L Henry, F Davis and S Eddy, personal communication). In addition, anti-GAD1 immunolabelling reveals that CT1 and TmY15 appear to use GABA (Figure 5).

19These considerations indicate that T4 dendrites receive both excitatory and inhibitory inputs and, together with the EM data, allowed us to examine the distribution of inputs of different predicted signs: the synapses from C3/Mi4/CT1 onto the T4 dendrite base all appear to be GABA-ergic and therefore inhibitory, whereas those from Mi1 and Tm3 along the shaft of the T4 dendrite are putatively cholinergic, and therefore most likely to be excitatory (Figure 6B). Mi9’s input at the T4 dendrite tip, putatively glutamatergic, may be either excitatory or inhibitory, depending on receptor expression at this synapse, given that glutamatergic neurotransmission can be inhibitory in Drosophila. T4-T4 synapses, also at the dendritic tips, are primarily cholinergic, and thus most likely to be excitatory.

20The finding of additional T4 inputs and their synaptic arrangements allows us to compare the predictions from functional studies with the requirements of different models of motion detection. Our findings reveal potential substrates for both BL- and HR-type motion detectors. A recent imaging study reported that T4 cells indeed use both HR model preferred-direction enhancement and BL model null-direction suppression. Which individual inputs onto T4’s dendrites fulfill the requirements for enhancement and suppression of directional selectivity? Of the synaptic inputs to T4 that are anatomically qualified to serve as two channels of a motion detector, several could support BL motion detectors (Figure 1B). First, inputs from Mi4, C3 and CT1 onto T4’s dendrite bases, which could form one input channel, are putatively inhibitory and therefore opposite in sign to excitatory input from Mi1 and Tm3 which could form another channel. Second, the vectors of physiological direction preference of each T4, as inferred from dendrite orientation, match those measured in a direction from excitatory inputs (dendrite shaft, Mi1/Tm3) to inhibitory inputs (dendrite base, Mi4/C3/CT1) in accordance with a BL motion detector (Figure 6B; cf. Figure 1B). Third, the inhibitory inputs arrive via a di- or trisynaptic pathway from L1, via L5 or Mi1 and others (e.g. L1→L5→Mi4→T4) (Figure 6A; see also Figure 3a in ref 16). By contrast, putatively excitatory inputs from Mi1 and Tm3 onto T4’s dendrite shafts are monosynaptic (Figure 6) suggesting a possible time delay between excitatory and inhibitory channels, but one that would be too short alone to account for the temporal frequency optimum in T4 cells of about 1 Hz.

21Among the potential inputs that could serve as HR motion detectors we first consider Mi1 and Tm3, as previously suggested. A role for these cells in T4’s motion computation has been confirmed: Mi1 block flies are motion blind to ON-edge stimuli, while Tm3 input is specifically required to detect fast ON-edge motion in the preferred direction. The polarities of transmission at both Mi1 and Tm3 inputs are likely to be the same because both appear to use acetylcholine as a neurotransmitter. In fact, a recent calcium imaging study has shown an enhanced T4 response when both cells are activated. Thus Mi1 and Tm3 could function as two excitatory input channels in an HR EMD circuit (Figure 1A), potentially depending on visual stimulus conditions such as different visual contrasts or velocities.

22This interpretation, however, runs counter to two features. First, assuming the measured directions of receptive field offsets between Mi1 and Tm3, the relative timing of the two inputs recorded by is contrary to that in an HR EMD. Moreover, the magnitude of the spatial offset itself appears to be very small. While our new data do not deny a role for Mi1 and Tm3 cells in motion detection, uncertainties in the direction of their small anatomical offsets (see Figure 4—figure supplement 2) do make it difficult to uphold our previous conclusion that these offsets provide the core of an EMD.

23Two additional input cell types, Mi9 and T4-T4 connections, that make synapses onto the leading edges of T4’s dendrites (Figure 4C and E) could also function as excitatory input channels of an HR model. In this case they would work together with Mi1 and/or Tm3, qualified by their substantial spatial offsets (Figure 4B). However, possible roles of Mi9 depend on two properties of this cell type: the sign of the glutamatergic synapses and the nature of its light response – Mi9 might respond to OFF rather than ON-stimuli, with main synaptic input from L3 and Mi4 (Figure 6A). Indeed recent imaging results indicate that Mi9 does respond to OFF stimuli and is likely to provide inhibitory input to T4 dendrites. T4-T4 synapses are also qualified to contribute to an HR mechanism: input from other T4s is spatially offset against the cells’ preferred direction (Figure 4E) and predicted to be delayed since an additional synapse is involved (e.g. Mi1→T4→T4 vs Mi1→T4). However, blocking T4’s synaptic outputs does not abolish directionally selective responses in T4, suggesting that a mechanism based on T4-T4 synapses would, at least under some conditions, have to act in parallel with others. Such parallel contributions of several T4 input pathways could, for example, serve to broaden the T4 response curve for: stimulus velocity, which peaks at a temporal frequency of ~1 Hz; or the range of stimulus modalities, for example to different light intensities or stimuli of differing spectral composition.

24In summary, we report the complete set of input cell types to direction-selective T4 cells, and reveal that these circuits are anatomically qualified to implement direction selectivity using both HR model preferred-direction enhancement and BL model null-direction suppression, as has indeed recently been reported from imaging studies in both T4 and T5 cells. While not precluding the contribution of specific inputs to directional selectivity, synaptic inhibition from specific medulla cells could also be required for other reasons, and could function to signal something entirely different. For example, in addition to acting locally, CT1 is anatomically qualified to modulate the gain of T4 and T5 neurons in response to overall levels of motion. Whatever mechanism(s) they support, our EM findings will be essential to understanding the detection of motion, in combination with specific functional tests.

25The heads of wild-type Canton-S female flies between 5 and 6 days post-eclosion were fixed and processed for EM by high-pressure freezing/freeze substitution, according to previously reported methods. The brain tissue was embedded in Durcupan epoxy resin (Fluka) after fixation. The left part of the brain was used for imaging.

26A block face series of 32,000 images was acquired by a Zeiss NVision FIBSEM at a resolution of 10 nm per pixel. A focussed ion-beam was used to remove 2.5 nm of material from the sample block face. The images were aligned using affine transforms and consecutive sets of four images were summed to generate isotropic 10 nm voxels.

27To obtain a dense reconstruction of seven medulla columns, we used a sequence of automated image processing followed by manual proofreading with Raveler (https://openwiki.janelia.org/wiki/display/flyem/Raveler). First, the medulla region of interest (ROI) was identified, which contained seven medulla columns. The image set of the ROI was then divided into 234 smaller overlapping cubes, each 5 × 5 × 5 µm (for a total reconstructed volume of ~30,000 µm3). In each cube, presynaptic T-bar ribbons were first detected automatically and then annotated by hand with Raveler, along with corresponding postsynaptic sites. Manual annotation took a total of ~6000 person-hours to identify both pre- and postsynaptic sites, with 8–10 proofreaders working in parallel on different non-overlapping subsets of cubes. Next, a watershed algorithm was applied to the pixel-wise predictions to generate the initial over-segmentation of the volume. The over-fragmented volume was then refined using a context-aware two-stage agglomeration framework to produce the final segmentation. The segmentation was improved through a manual process of ‘focussed’ proofreading in which we concentrated on the divisions between cells that are both biologically improbable and that produce important 3D neuron shapes, especially if these be doubtful. An additional inspection centred on features thought improbable, such as ‘orphan’ fragments that failed to touch the surface of any cube. This manual proofreading took a total of ~4000 person-hours, with 8–10 proofreaders working simultaneously on different non-overlapping subsets of cubes. The cubes were then stitched together using a combination of automatic and manual operations. Neuron 3D reconstructions were generated with software NeuTu-EM (https://github.com/janelia-flyem/NeuTu/tree/flyem_release, with a copy archived at https://github.com/elifesciences-publications/NeuTu), in which we examined synaptic locations and distributions, and T4’s dendrite orientations. Neuron reconstructions were closely examined for proofreading errors. Cell types of reconstructed neurons were identified by comparing the shapes of their arbours with the shapes of previously reported neuron reconstructions and from light microscopy. Connectivity matrices were generated by combining the results of neuron reconstructions with those for synapse mappings. The plots of spatial offsets for the columnar inputs (Mi4, Mi9, C3, and CT1) were calculated as the centre of mass of locations weighted by synapse count of the associated column in an ideal hexagonal array, referenced to the centre of the same calculation for Mi1.

28Gal4 or Split-Gal4 driver lines used to visualize T4 input neurons were, SS02432 (R48A07AD, VT046779DBD) (Mi9), SS01019 (R48A07AD; R13F11DBD) (Mi4), SS00809 (R19F01AD; R71D01DBD) (Mi1), R26H09AD; R29G11DBD (C3), SS01001 (R65E11AD; R20C09DBD) (CT1) and VT048653 (TmY15). The CT1 split-Gal4 line was constructed using hemidrivers selected on the basis of GAL4 line expression data. VT048653, from the Vienna Tiles collection, predominantly labels TmY15 cells in the medulla based on stochastic labelling data (41/63 MCFO-labeled medulla cells in 17 brains were TmY15 neurons) and the overall expression pattern of this driver line. For Multicolor Flp-out (MCFO) labelling (Figure 2B; Figure 2—figure supplement 1A, B and D), driver lines were crossed to MCFO-1 or MCFO-7 and processed for immunolabelling, mounted in DPX and imaged as described. Imaging and processing of brains with (split-)GAL4 driven expression of presynaptic and membrane markers (Figure 2C and D; Figure 2—figure supplement 1C and E) using pJFRC51-3XUAS-IVS-Syt::smHA in su(Hw)attP1 and pJFRC225-5XUAS-IVS-myr::smFLAG in VK00005 was as described. pJFRC12-10XUAS-IVS-myr::GFP in attP2 was used for the neurotransmitter labelling experiments and the light microscopy data in Figure 4—figure supplement 1. Female flies were used for all experiments.

29Light microscopy panels (Figures 2B, C and D; Figure 2—figure supplement 1) show reconstructed views that were generated using the neuronannotator mode of Vaa3D and exported as TIFF format screenshots. Some panels (CT1 in Figure 2C and D; Figure 2—figure supplement 1, and TmY15 in Figure 2B and Figure 2—figure supplement 1) show different views of the same cells generated from the same confocal stack to illustrate different aspects of the anatomy. Additional images (not shown) of other cells of these types showed the same general anatomical features.

30Neurotransmitter immunolabellings used previously described antibodies against GAD1 (a gift from F.R. Jackson obtained via Y. Aso), ChAT (mAb ChAT4B1) obtained from the Developmental Studies Hybridoma Bank and VGlut (, a gift from A. DiAntonio). For these experiments, fly brains were dissected in PBS, fixed in 2% (v/v) PFA (prepared from a 20% stock solution, Electron Microscopy Sciences, #15713) in PBS for 1 hr, washed with PBT (0.5% v/v TX-100 in PBS), incubated with PBT-NGS (PBT with 5% v/v Normal Goat Serum) for 30 min and then incubated for ~2 days with primary antibodies (anti-GAD1 1:1000 dilution or anti-VGlut 1:5000 and anti-ChAT 1:50) in PBT-NGS on a nutator at 4°C, followed by further washes with PBT and incubation with secondary antibodies (DyLight 649 AffiniPure Donkey anti Rabbit IgG (H+L), Jackson ImmunoResearch Laboratories, Inc., #711-495-152,1:300 dilution and DyLight 549 AffiniPure Donkey anti Mouse IgG (H+L), Jackson ImmunoResearch Laboratories, Inc #715-505-151, 1:300 dilution) in PBT-NGS for ~2 days. After additional washes, brains were mounted in SlowFade Gold antifade (ThermoFisher Scientific) under a no. 1.5 coverslip using small amounts of modeling clay as flexible spacers. Brains were imaged with the anterior brain surface towards the coverslip, except C3 cells, which were imaged from the posterior to favour their different cell body locations. Native fluorescence was used to detect the GFP marker in these experiments. Images were acquired on an LSM 710 confocal microscope with a 63 × 1.4 NA objective. With the exception of the anti-GAD1 labeling of Mi9 and Tm3, all images in Figure 5 with labeling of a given neurotransmitter marker show brains processed together in the same tube. However, the labeled brains were mounted and imaged individually with some specimen specific adjustments of laser power and gain, and/or of brightness and contrast post-imaging. Multiple specimens were examined for each cell type with similar results. For CT1, 12/12 cells from six brains showed GAD1 immunoreactivity.

31Our 7-column medulla raw data have been published on the Janelia/Howard Hughes Medical Institute website, http://emdata.janelia.org/#/repo/medulla7column.