Identification and reconstruction of Drosophila LPT neurons in an Electron Microscopy dataset

A. The Lobula Plate Tangential Neurons were manually reconstructed on one side of Full Adult Female Brain (FAFB) EM volume (Zheng et al., 2018). The skeleton of one such LPT, the Vertical System number 1, or VS1 neuron, is shown (in black) superimposed on a single slice of the EM volume (Zheng et al., 2018). The data and this image were generated using the CATMAID environment (Saalfeld et al., 2009). The inset shows the very large caliber processes of several VS neurons as well as the much smaller processes of nearby cells, and the somata of neurons in the Lobula Plate cell body rind (bottom of inset). 58 LPT neurons were found on one side of the FAFB brain. The complete inventory of LPT Neurons is detailed Supplementary File 1, and a gallery of their morphologies is in Supplementary File 2. B. Examples of light microscopy images of LPT neurons that were used to guide the search, and as a comparison for, reconstructed neurons in the EM volume. Images show manually segmented MulticolorFlpOut (MCFO) labeled LPT neurons (green) together with a neuropile marker (anti-Brp, gray). The names of each cell, listed in the lower left corner, are based on our standardized nomenclature, summarized in the text, for the matching EM-reconstructed LPT neurons.

Predicting visual motion response from anatomy.

A. An EM-reconstructed lobula plate tangential (LPT) neuron, VS1, shown with the outline of the FAFB brain volume and the indicated optic lobe neuropils. B. Four subtypes of directionally selective T4 neurons receive input from the proximal layer of the medulla and each project to one of four layers in Lobula Plate (LOP). C. Top: a fly head model with lenses color-coded to illustrate retina positions. Bottom: viewing directions of all ommatidia in our reference compound eye (Zhao et al., 2022), presented as a Mollweide 2D projection. D. Anatomical positions in LOP mapped to ommatidia directions in (C) via set of reconstructed T4 neurons (Figure 2—figure supplement 1A). E. Local preferred directions of motion for four T4 subtypes mapped to the eye coordinates, as a Mollweide projection, down-sampled by a factor of 9 (Figure 2—figure supplement 1B,C). F. Top: 2 views of the VS1 neuron, whose dendritic branches are colored by the 2 innervated LOP layers, using the same color code as in (E). Bellow: two cross sections, corresponding to slices (1) and (2) above, showing several reconstructed T4b and T4d neurons and the 4 LOP layers. G. Layer coverage by the VS1 dendrite, displayed in eye coordinates. H. Predicated Motion Pattern Map (PMPM) estimated from the dendritic layer coverage and T4 neurons’ preferred directions in (E). J. Motion response of a Calliphora VS1 neuron, measured with electrophysiology (replotted in eye coordinates from (Krapp and Hengstenberg, 1996)). K1. Top 3 plots: comparisons of the PMPM to ideal optic flow fields induced by rotations along the 3 indicated axes. Bottom: heat map shows the average angular difference (αR) between the PMPM and all sampled ideal flow fields. K2. Similar to K1, but for translation (αT). The directions with minimal average angular difference for rotational and translational flow fields are indicated by ‘×’ and ‘+’. The black line shows the contour that contains the directions of tested ideal flow fields with the 10% smallest differences.

The Horizontal and Vertical System Neurons—morphology, LOP coverage, PMPMs, and optic flow tuning.

A. The HS and VS neurons reconstructed on one side of the FAFB brain. From left to right: EM reconstruction, LOP layer coverage, color-coded to match Figure 2, PMPM, optic flow tuning as heatmaps of average angular difference (αR for rotation; αT for translation; coordinates [elevation, azimuth] and error value: αR ≡ αR0 and αT ≡ αT0 for minimal difference). The heatmap coordinate system has positive elevation above the equator, positive azimuth in the right visual field, and [0°,0°] is the front. B. Optimal rotation axes for all VS neurons in the right-side optic lobe (in black) and assuming symmetry also for the left-side (in red), shown with a Mercator projection.

The other LPT neurons with T4 inputs—morphology, LOP coverage, PMPMs, and optic flow tuning.

Following the convention of Figure 3, the other LPT neurons are presented, grouped by morphological categories, described in the text. From left to right: EM reconstruction, LOP layer coverage, color-coded to match Figure 2, PMPM, optic flow tuning as heatmaps of average angular difference (αR for rotation; αT for translation; coordinates [elevation, azimuth] and error value: αR ≡ αR0 and αT ≡ αT0 for minimal difference).

Brain-spanning network connectivity between LPT neurons.

A. LPT-LPT network graph based on connectivity from the optic lobe in FAFB (reported here) and the central brain in the Hemibrain connectome (Scheffer et al., 2020). All connections ≥ 10 synapses between the LPT neurons are included, with the edges indicating where the connections are made, the strength of the connections, and whether they are expected to be excitatory (pre-synaptic cell predicted to be cholinergic) or inhibitory (pre-synaptic cell is either GABAergic or glutamatergic) based on Figure 3—figure supplement 2 (and Methods). Each LPT neuron is shown as a pie chart representing the layers of LOP innervation. B. Examples of two connected LPT neurons, LPT44-Nod5 and LPT35-dCH, with their independent reconstructions in the FAFB data and Hemibrain. C. Number of connections and connection weights between LPT neurons grouped by the predicted polarity of their output synapses.

Predicted motion pattern maps for central brain LPT-target neurons.

A. Excitatory and inhibitory input breakdown of 66 LPT-target cell types, indexed by their PMPM-based field of view (summing up the fields of view of input LPT neurons without double-counting overlaps). The 4 out of 66 cell types that only receive inhibitory inputs are excluded in the following analysis. B. An example of an LPT-target neuron in Hemibrain that receives inputs from three LPT types. Its PMPM is constructed from those of the input LPT neurons, weighted by connection weights (given as percentages in this example). We exclude inhibitory inputs in computing PMPMs. C. PMPM-based field of view (including inhibitory inputs) as a fraction of 4π solid angle. Also shown (in gray) are the fields of view of input LPT neurons. The field of view of one eye is 41.4%, and of both eyes is 68.6% (based on the eyemap used throughout, (Zhao et al., 2022)). D. Overlap of input LPT neurons’ fields of view. E. The PMPM for the example target neuron together with optimal rotation and translation axes. F. Rotation-translation Selectivity Angle, RTSA, for both LPT neurons and their central brain targets. Large positive values imply a better fit of the PMPM to an optic flow field induced by some translational movement. Large negative values for some rotational movement. Values near zero indicate that the PMPMs are ambiguous, not clearly selective for either translation or rotation.

GAL4 driver lines used to produce images of labeled LOP neurons

All images in Figure 1B show manually segmented cells from confocal stacks that are part of the public Janelia MCFO image collection (Meissner et al., 2023) and available online at

The individual original image stacks can be directly accessed using links formatted as follows:

For example, for the first image in the table this would be:

Analysis details supporting the computational predictions of LPT PMPS.

A. EM reconstructed T4b axon terminals in LOP2. B and C. To present down-sampled PMPMs, the field of view of the fly’s right eye is divided into 24 regions. The average preferred direction in each region is represented by arrows in the PMPMs, whose locations are denoted by red dots here. This averaging is carried out for all flow field plots in this manuscript. Here we show both Mollweide and Mercator projections. D. For the analysis presented in the manuscript, the LOP layers are constructed as a polynomial surface fit to the axon terminals of each T4 subtype. We use 5th order polynomials, and evaluated the quality of assignments of nodes in T4 axon terminals to layers to avoid over-fitting. The mis-assigned nodes are plotted as a function of the order of the polynomial surface fit. E. Distance between annotated synaptic connector and the nearest node of the reconstructed neuron skeleton, for all LPT neurons described in this paper. Note that 97.6% of connectors fall within 1.5 µm of the skeleton. F. Percentage of synapses captured by including varying levels of Strahler branches of the dendritic tree, within this 1.5 µm distance, for the main LPTs described in this manuscript (except for the feedback neurons that are not major T4/T5 targets). Note that SN <=3 captures most synapses, using SN>=4 may lead to overestimate of the dendritic coverage.

Supporting evidence for identification of LPT neurons

A. Axon and primary dendrite diameters for all VS neurons in FAFB. We note that 16 VS cells were eventually found on the right and left sides of the brain, but at the time these measurements were made, one of the LPT17 cells on the left side had not yet been located. B. Comparison of EM reconstruction of VS neurons, in the FAFB brain, left-side neurons from FlyWire (in shades of red) and right-side neurons manually reconstructed in CATMAID (in shades of black). C. Cosine-similarity comparison of left vs. right-side LPT neurons in Flywire. Some cells are treated as a group, as indicated, and the cosine similarity is computed on the group connectivity average. For this computation, an LPT neuron’s connectivity is characterized by its connection with all other LPT neurons only. All FlyWire data based on the March 2023 public release (version 630).

Neurotransmitter predictions for the LPT neurons.

A. Neurotransmitter prediction for all LPT neurons, based on the method of (Eckstein et al., 2020) for the matched neurons in FlyWire. Only 2 cell types show right/left discrepancy in the top prediction, indicated with arrowheads. B,C,C’. Experimental evidence that MeLp2 neurons are glutamatergic. B. Expression pattern of a split-GAL4 line (SS02082) labelling LPT55-MeLp2 neurons. GAL4-driven expression of a membrane-targeted GFP is in green, a general neuropile label (anti-Brp) in magenta. The asterisks show the soma locations of the right and left MeLp2 neuron. C and C’. EASI-FISH labeling (Eddison and Ihrke, 2022; Wang et al., 2021) of transcripts indicative of cholinergic (ChAT; red), glutamatergic (VGlut; green) or GABAergic (GAD1; yellow) transmitter phenotypes. A MeLp2 cell body (indicated by the adjacent asterisk in both C and C’) is labeled in magenta in C. Images show different channel combinations of a single confocal section.

Additional supporting evidence for the neurotransmitter predictions.

A,A’,A’’. EASI-FISH labeling (Eddison and Ihrke, 2022; Wang et al., 2021) of ChAT (red), VGlut (green) and GAD1 (yellow) in the lobula plate cell body rind, in the region where the HS and VS somata are found. Nuclei are also labeled (blue). An LPi2b neuron labelled by a split-GAL4 driver line (SS 53141) is shown in magenta.

Images show different channel combinations of a single confocal section (A, all channels; A’ transmitter indicators only; A’’, nuclei only). Large cells in this region appear to be either GABAergic (GAD1 marker, orange) or cholinergic (ChAT marker, red). B. For comparison with (A), we use the soma locations from the FAFB LPT set: LPi2b (magenta), other large optic lobe intrinsic LOP neurons (orange; these include Am1, 2 LPi12, Lpi21, and an undescribed cell that appears to be a large layer four LPi), 4 HS neurons (maroon), and 8 VS neurons (red). The large optic lobe intrinsic neurons are only partially reconstructed and not included in the LPT survey reported in this study, but could be matched to known cell types (Shinomiya et al., 2022). Given that these are all the large cell bodies we found in this region, and several of the LPis are expected to be inhibitory, we suggest that the GABAergic cells in the EASI-FISH images are the large LPi and Am1 neurons and the cholinergic cells are the HS and large VS cells.

morphology and layer coverage of the LOP-input LPT neurons

The remaining LPT neurons receive little T4 inputs, instead they most likely provide inputs to the LOP, some as feedback from the central brain. As these neurons receive minimal T4 inputs, we do not predict their motion pattern maps.

Optimal rotation or translation axes predicted by PMPMs.

These two plots summarize the optimal rotation (top) and translation (bottom) axes for the LPTs of Figures 3 and 4, excluding the VS neurons, which are treated in Figure 3B. The size of each neuron’s marker indicates the size of the angular error.

LPT connectivity in the Hemibrain connectome

A. Improvement in connectivity completeness for the matched set of LPT neurons, after manual proofreading in Hemibrain. B. Comparison of central brain LPT-LPT connectivity in FAFB (left) and Hemibrain (right) data sets.

supporting data for LPT integration analysis

A. Composite field of view for the example LPT target neuron in Figure 6A. B. Fields of view and predicted synapse polarity for the LPT neurons (the indexing used throughout indicates the corresponding LPT neuron). C. Comparing the minimal average angular differences computed via the brute force method for rotation and translation separately, and a non-linear regression method treating them jointly (see Methods).

Patterns of LPT output integration in the central brain

A. Number of target cells vs. number of their LPT inputs for the central neurons considered in this analysis (based on Hemibrain connectivity). B. Pairwise “collaboration matrix” where each entry represents all the LPT-target neurons that receive inputs from the LPT neuron indicated in the row and the LPT neuron in that column. The matrix is asymmetric since the value of each entry is the total connection weight from the LPT neuron in the row. C. The eigen spectrum of the Laplacian matrix constructed from the pairwise collaboration matrix, showing a gap after the 5th largest eigenvalue. This was used to select the number of clusters (=5) for the spectral clustering in C. D. The pairwise collaboration matrix re-ordered based on spectral clustering. There are 5 clusters within the connected graph. E. Synapse distributions and the relevant neuropils capturing the majority of synapses for each cluster in the central brain. The synapses shown are restricted to those between LPT neurons and their targets (i.e. those in the collaboration matrix).