Identifying Inputs to Visual Projection Neurons in Drosophila Lobula by Analyzing Connectomic Data

Visual Abstract

In addition to motion and color, the visual system of flies is equipped with neurons tuned to a variety of visual features. Recent studies have shown that a series of columnar output neurons of the lobula (Fig. 1A) collectively named lobula columnar (LC) and lobula plate-LC (LPLC) neurons are tuned to visual features resembling conspecifics and predators, and their activity can induce a variety of behavioral responses, depending on types (Panser et al., 2016;Wu et al., 2016;Keleş and Frye, 2017;Klapoetke et al., 2017;von Reyn et al., 2017;Ribeiro et al., 2018;Ache et al., 2019;Städele et al., 2020;Clark, 2020, 2021;Sten et al., 2021). Several of these neurons are thought to achieve their selective visual tuning by pooling the feature selectivity of their presynaptic partners (Klapoetke et al., 2017;Keleş et al., 2020;Clark, 2020, 2021). However, the presynaptic partners of these LC and LPLC neurons remain largely unknown. This is because of the lack of connectomic reconstruction encompassing the lobula and its upstream neuropils. For example, a recent connectomic dataset used to analyze the motion detection circuitry included most of the visual system, but the reconstruction was focused on T4 and T5 and their inputs, leaving out the most of the lobula circuitry (Shinomiya et al., 2019). Similarly, community-driven tracing and proofreading effort of the full adult fly brain (FAFB) dataset (Zheng et al., 2018) is still in progress. While the densely-reconstructed hemibrain dataset contains almost the entire lobula, since it lacks almost the entire medulla, neurons providing inputs into the lobula [e.g., transmedullar (Tm) neurons) are fragmented and unlabeled; Scheffer et al., 2020].
However, recent studies have shown that close inspection of these fragmented axon terminals of input neurons into the lobula in the hemibrain dataset can still offer potentially useful information about the lobula circuitry Clark, 2020, 2021). In the present study, we perform clustering analyses on those fragmented terminals based on their connectivity and morphology, defining putative input cell types for intrinsic and projection neurons of the lobula, including LC and LPLC neurons. This analysis comes with the caveat that it is difficult to conclusively identify cell types based on their axonal morphology alone and thus needs to be complemented by a more comprehensive connectome encompassing the entire optic lobe. Even so, the results presented here provide a previously unknown putative connectivity to physiologists and psychophysicists interested in visual processing in the lobula.  (LP). The lobula houses projection neurons such as lobula columnar (LC) and lobula tangential (LT), which send axon terminals to the central brain and are thought to be responsible for detecting various visual features for approach and avoidance behaviors. Neurons connecting the medulla and lobula, such as transmedullar (Tm) neurons, are likely providing major synaptic inputs into LCs and LTs, but their connectivity remains largely unknown. The teal shaded region roughly indicates the volume reconstructed in the hemibrain dataset. B, A schematic of the analysis pipeline. Small lobula intrinsic neurons in the hemibrain dataset were selected as cells of interest, and their connectivity to labeled downstream neurons were tabulated. The candidate cells were then rotated into the principal component (PC) space of LT1 postsynapses, which we used as a proxy for layer 2 of the lobula, from which we derived normal and tangential axes of the lobula layers. We then quantified the morphology of the cells of interest by synapse spread in the PC space and by the histogram of innervation depth relative to a parabolic surface fit to the LT1 synapses. Finally, we applied hierarchical clustering to the concatenated feature matrix, and visually annotated the resulting clusters. C, The morphology of reconstructed LT1 neurons, viewed from (top) normal and (bottom) tangential directions, visualized alongside with the boundary of the lobula neuropil. The dendrites of a single LT1 neuron cover the entire tangential extent of the lobula, rather than tiling the lobula as a population. The dendrites of LT1 form a single thin layer around layers 2 and 3 of the lobula. D, The morphology of a single LT1 neuron (green), with several cells of interest in cluster 1 (see also Figs. 4,6), viewed tangentially. E, The parabolic surface (blue) fit to the LT1 presynapses (black). For clarity, only a random 10% of the total LT1 presynapses are visualized.

Materials and Methods
The goal of present analysis is to categorize unlabeled putative columnar neurons in lobula into cell types, and examine their connectivity to downstream neurons. Typically, a columnar visual neuron type consists of many cells sharing stereotypic connectivity and morphology tiling the visual space. Thus, the problem boils down to grouping neurons by their individual connectivity and morphology while discounting their exact spatial location. To this end, we performed an agglomerative clustering on their connectivity and morphology. A schematic of the analysis pipeline is shown in Figure 1B.

Code accessibility
The code to perform the analysis presented in the paper is freely available online at https://github.com/ ClarkLabCode/LobulaClustering. The code is available as Extended Data 1.

Inclusion criteria
Cells without cell type assigned in the hemibrain 1.2.1 dataset that had synapses only in the lobula and had .50 and fewer than 500 total synaptic sites were selected. The upper bound of 500 synaptic sites was comparable to the size of the smallest LC neurons (e.g., LC12) or lobula intrinsic neurons (e.g., Li13). Cells with fewer than 50 synaptic sites were excluded because they were difficult to classify reliably because of the small amount of connectivity data available. We refer to these cells as "cells of interest." A total of 7465 cells of interest met the above criteria and were incorporated to the analyses.

Connectivity
For each cell of interest, we counted the number of synapses it has onto labeled cell types in the lobula. Connections to multiple cells of a single cell type were summed together to generate a single feature, rather than being treated separately. This was intended to facilitate grouping of cells belonging to the same cell types with a shared pattern of connectivity regardless of their retinotopic location. A total 237 labeled cell types were found to be downstream of at least one of the cells of interest in the present analysis.

Morphology
To summarize features of the axonal morphology of the cells of interest, we calculated (1) the standard deviation of synapse position along the two tangential axes and the normal axis of the lobula layers, as well as (2) the histogram of synapses along the depth of the lobula. The native coordinate system of the hemibrain dataset is a Cartesian system where X points left, Y points anteriorly, and Z points ventrally (Scheffer et al., 2020). To identify the normal and tangential axes of the lobula layers, we first performed principal component analysis (PCA) on the postsynapses of LT1 neurons. LT1 is a tangential neuron with wide, dense, monostratified dendrites in layer 2 of the lobula (Fischbach and Dittrich, 1989;Fig. 1C,D). The hemibrain dataset contains four LT1 neurons, each covering the entire tangential extent of the lobula. The resulted first two principal component (PC) axes were approximately aligned to the ventrolateral and anteroposteior axes of the lobula, respectively, and the third PC axis was normal to that plane. Then, for each cell of interest, we calculated the standard deviations of the distributions of its presynapses along the three PC axes, which quantified tangential extents and vertical diffuseness of its axon terminal. We refer to these three features as "synapse spread" hereafter.
Next, we aimed to calculate the depth of each synapse along the layers of the lobula. Since the layers of the lobula are curved, the raw position of the synapses along the third PC axis do not accurately correspond to their layer affiliation. To solve this issue, we first fit a parabolic surface:ẑ ¼ a 0 1 a 1 x 1 a 2 y 1 a 3 x 2 1 a 4 y 2 1 a 5 xy; to the postsynapses of the LT1 neurons with the leastsquare method, where x and y represent position along the first and second PC axes (mm),ẑ represents the predicted position along the third PC axis (mm), and a i are coefficients (Fig. 1E). The fit coefficients were as follows: a 0 = À5.21, a 1 = 4.04 Â 10 À3 , a 2 = 4.04 Â 10 À3 , a 3 = 3.17 Â 10 À3 , a 4 = 7.65 Â 10 À4 , a 5 = À1.22 Â 10 À3 . The goodness of fit was R 2 = 0.744. We then calculated the depth of each synapse relative to layer 2 as z Àẑ, where z represents the true position of the synapse along the third PC axis. Note that this approximates the lobula layers to be parallel to each other, rather than concentric curved surfaces. Deviations from this approximation will be most severe far from the center of this PC coordinate system. For each cell of interest, we counted the number of presynapses in 5 mm wide bins ranging from À20 to 150 mm of relative depth, resulting in 14 features. This range covered the entirety of the lobula. We refer to these features as "innervation depth" in the following. Similar analyses of innervation depth have been performed previously on medulla neurons (Bausenwein et al., 1992;Zhao and Plaza, 2014). We validated that the synapse spread and innervation depth calculated this way appropriately reflect the neuronal morphology by applying the same analysis to postsynapses on the dendrites of multiple LC types with well characterized morphology (Wu et al., 2016).

Agglomerative clustering on cells of interest
We concatenated the connectivity, synapse spread, and smoothed innervation depth feature matrices into a single feature matrix, and performed an agglomerative clustering with Ward's variance minimization method (Ward, 1963) with SciPy (Virtanen et al., 2020). Before being concatenated, connectivity, synapse spread, and innervation depth matrices were, respectively, normalized by its total dispersion (sum of variances) and then weighted by the factors of 5, 1, and 3, respectively. This weighting reflected our relative confidence on the accuracy of the three feature sets. We set the number of clusters to be 40, which is approximately comparable to the number of Tm and TmY cell types reported in Fischbach and Dittrich (1989). The complete results of the clustering analysis are presented in Extended Data 2.

Annotation and visualization
Next, we visualized the cells of interest in each resultant clusters on the neuPrint website (Clements et al., 2020a) and compared their terminal morphology with existing anatomic literature (Fischbach and Dittrich, 1989) to determine cell type identity. We also performed a nonlinear dimensionality reduction on the concatenated feature matrix with uniform manifold approximation (UMAP; McInnes et al., 2018) for visualization.

Quantification of connectivity between cells of interest and visual projection neurons
Lastly, we analyzed the connectivity between the 40 clusters of cells of interest and 21 LC and LPLC types (Wu et al., 2016). For each LC or LPLC type, we calculated relative number of synapses they receive as a population from each of the 40 clusters. Additionally, an agglomerative clustering was performed on the relative connectivity matrix between the 40 clusters and the 21 LC/LPLCs.

Results
The synapse spread and innervation depth features capture the dendritic morphology of LC neurons To gain confidence in our quantification of cellular morphology, we first calculated innervation depth and synapse spread features for the lobula postsynapses of 18 LC and LPLC projection neuron types with well characterized morphology (LC4, 6, 9, 11-13, 15-18, 20-22, 24-26, LPLC1 and 2; Wu et al., 2016;Fig. 2). Figure 2A,B shows previously reported light microscopy-based innervation patterns of LCs and LPLCs (Wu et al., 2016; Fig. 2A) side by side with their connectome-based innervation depth feature (Fig. 2B). The patterns of innervation captured by two methodologies agreed well. For example, the bistratified dendrites of LC4 in the layers 2 and 4 are clearly visible from the connectome-based innervation depth features ( Fig. 2A,B, the leftmost column). The synapse spread of LCs and LPLCs along the two tangential and the normal axis of the lobula are shown in Figure 2C,D. These features also captured well the known morphology of these neurons. For example, LC25 and 12, which, respectively, had large and small synapse spread along the both tangential dimensions (Fig. 2C) have been previously shown to have large and small isotropic dendrites (Wu et al., 2016, see their Fig. 6). LC22, which had large synapse spread along the long tangential axis of the lobula but small spread along the short axis, indeed has anisotropic, narrow dendrites (Wu et al., 2016, see their Fig. 6). In a similar vein, along the normal axis, vertically diffuse neurons such as LPLC1 and LC15 (Wu et al., 2016, see their Fig. 5) had large synapse spread, whereas monostratified LC25 had small spread (Fig. 2D). Overall, these results indicate that the two morphologic features we calculated accurately capture the morphology of neurites in the lobula neuropil.

Agglomerative clustering identifies clusters with homogeneous connectivity and morphology
The results of the clustering are summarized in Figures  3, 4 and Table 1. The detailed analyses of each cluster will be presented in the final part of Results. Clustering   Figure 2B. D, Synapse spread of the cells of interest along the tangential and normal axes of the lobula. E, The cells of interest on the UMAP embedding space, based on the normalized, weighted, and concatenated features. Clustering was performed directly on the feature matrix, so that the UMAP embedding is simply for visualization. a. u.: arbitrary unit.  Table 1. For each putative cell type, (left) a population of reconstructed neurons in the corresponding clusters viewed tangentially, (middle) individual examples of cells in the clusters, and (right) Golgi staining from Fischbach and Dittrich (1989) are shown. B, The same as A, but for clusters with less confidence (labeled as B in Table 1). The green and blue arrows, respectively, point toward anterior and ventral directions. successfully extracted groups of cells with similar connectivity and morphology, as can be seen from the blocklike appearance of the feature matrices ( Fig. 3B-D). Additionally, we visualized the cells of interest on the twodimensional nonlinear embedding by UMAP (Fig. 3E). Overall, the embedding agreed well with the hierarchical clustering, since neighboring clusters are often neighbors in the UMAP space. This was not without exceptions, such as clusters 15, 16, 29, and 31, which had clear substructures on the UMAP embedding. Table 1 summarizes the morphology, connectivity, and putative cell type affiliation of all identified clusters. Example clusters that we identified with previously documented types of neurons with reasonable confidence are shown in Figure 4. The input columnar cell types we identified most confidently included T3, T2, TmY9/Tm28, Tm20, and TmY11 (Fig.  4A).

Inputs into columnar visual projection neurons
To gain insight into the visual computations of columnar visual projection neurons in the lobula, we calculated connectivity between the 40 clusters of the cells of interest and 21 LC and LPLC neurons (Wu et al., 2016). Figure 5A summarizes major presynaptic clusters for each of the 21 LC/LPLC type, with putative cellular identities of the clusters labeled. Figure 5B also shows relative connectivity between 40 clusters and the 21 LC/LPLC neurons, but with a dendrogram from the agglomerative clustering analysis. The most fundamental divide in the dendrogram was between "shallow" LCs with prominent Lo2/3 innervation (LC4,11,12,17,18,21; Fig. 2A,B) and the rest. Among the six shallow LC types, LC11, 17, and 21 received more inputs from T3 (clusters 1-4), whereas LC4, 12, and 18 received more inputs from T2 (cluster 5). Among the remaining cell types, LC25 stood out for its The homogeneity of each cluster was ranked in three tiers: A, likely single cell type; B, several cell types; and C, mixture of many cell types. Cell types are labeled where identified, and a question mark is used in cases of cell types that appear homogeneous but are previously undocumented. The confidence of cell-type identification was also ranked in three tiers (A, B, and C for highest to lowest confidence) or marked NA for clusters where we were unable to find cell types.
almost exclusive connectivity to Tm5 neurons. LC20 and LC22 also had a relatively distinct connectivity pattern, with prominent connection to cluster 36 (putative TmY5). The remaining cell types were divided into two branches, one of which included LC6, 9, 13 and LPLC1, 2, and the other including LC10, 14, 15, 16, 24, 26, and LPLC4. The former branch stood out for their shared connectivity to clusters 26 and 32, putative Tm5Y. The clusters shared by the latter branch included clusters 19 (putative TmY10) and 40 (putative TmY5a). Additionally, we found some clusters that provide substantial inputs to only single LC/ LPLC type analyzed here. For example, LC15 was the only cell type that received .5% of its inputs from cluster 10 (putative TmY9/Tm28). Similarly, LC6 was the unique target of cluster 28 and LC22 was the unique target of cluster 18. We could not find morphologic matches for neither cluster 18 or 28, but they were morphologically homogeneous, likely representing genuine cell types.

The analyses of individual clusters
The following describes the detailed morphologic analyses of the 40 clusters of the cells of interest resulting  Table 1. B, Relative connectivity from the 40 clusters to the 21 LC/LPLC types, with a dendrogram from the agglomerative clustering analysis of the LP/LPLC types. Connectivity is normalized for each LC/LPLC type, such that each row of the connectivity matrix adds up to 1. from the agglomerative clustering (Fig. 3), on which putative cell type identification in Table 1 and Figures 4, 5 are based.

Clusters 1 through 4
Clusters 1 through 4 were highly distinct from the other clusters, as can be seen from the dendrogram (Fig. 3A) and the UMAP embedding (Fig. 3E). Cells in these clusters had small monostratified axon terminals slightly elongated along the long tangential axis (synapse spread of ;2.5 mm along the long axis, and 1.5 mm along the other axes), which innervated the boundary between layers 2 and three of the lobula (i. e., Lo2 and Lo3; Fig. 6A-C). The major synaptic targets of these clusters were LC17 (13.7, 6.4, 25.0, and 9.9 synapses/cell for clusters 1, 2, 3, and 4, respectively), LC11 (15.8, 4.2, 6.3, and 2.5 synapses/cell for clusters 1, 2, 3, and 4, respectively), LT1 (8.1, 3.1, 22.4, and 12.3 synapses/cell for clusters 1, 2, 3, and 4, respectively). LPLC1 was also a major target for clusters 3 and 4 (8.7 and 2.5 synapses/cell for clusters 3 and 4, respectively). Visually, the four clusters appeared homogeneous, consisting of cells with small crumpled terminals (Fig. 6D,E). Their main neurites sometimes overshot and then came back to the layers 2 and 3, forming characteristic loops. Based on their extensive connectivity to LC11 Tanaka and Clark, 2020) and distinctive morphology (Fischbach and Dittrich, 1989; Fig. 6E,F), these clusters most likely correspond to T3. T3 is a cholinergic (Konstantinides et al., 2018) neuron that connects the proximal medulla (M9) and distal lobula, is an ON-OFF cell, and is tightly tuned to small moving objects Tanaka and Clark, 2020). T3 neurons receive cholinergic inputs from both ON (Mi1) and OFF (Tm1) cells , which likely implement their ON-OFF property.
Clusters 6 and 7 were somewhat heterogeneous clusters with small membership (Fig. 7G). They were slightly more vertically diffuse than cluster 5 (Fig. 7A,C). Cluster 6 consisted of at least two cell types: small vertical cells with Y-shaped ending at Lo4 and swelling at Lo2 (Fig.  7H), and thin monostratified cells in Lo4 with long, knobby neurites branching perpendicularly (Fig. 7J). Cluster 7 was similar to cluster 6 but contained horizontal spiny cells in Lo2, which appeared to be fragmented parts of larger cells (Fig. 7K). Their major targets after LC4 were LPLC1 (2.1 synapses/cell) for cluster 6 and LC12 (2.0 synapses/ cell) for cluster 7, but these synapse counts were about an order of magnitude lower than their connectivity to LC4. The terminal morphology of the vertical cells resembled Tm4 (Fischbach and Dittrich, 1989;Fig. 7I), although Tm4 appears to have more extensive swelling in Lo1. TmY2 also had similar Y-shaped ending (Fig. 7I), but they have swelling in Lo3 rather than Lo2 (Fischbach and Dittrich, 1989).
Cluster 9 was a highly anisotropic cluster with synapse spread of 12.2 and 4.8 mm along short and long tangential axes (Fig. 8B). The dominant cell type of this cluster had a long, knobby neurites branching at acute angles in Lo5B, which entered the lobula ventrally and then bent dorsally D, A population of cluster 5 cells, viewed from a tangential direction. E, A representative cluster 5 cell. F, Morphology of T2 axon terminals, from Fischbach and Dittrich (1989). G, A population of cluster 6 cells, viewed from a tangential direction. H, A typical "vertical" cell in cluster 6. I, Morphology of Tm4 and TmY2, from Fischbach and Dittrich (1989). J, A typical "monostratified" cell in cluster 6, viewed from a normal direction. K, A population of cluster 7 cells, viewed from a tangential direction. The blue arrows indicate the ventral direction throughout. (Fig. 8F,G). The major postsynaptic targets of this cluster after LC10 were LC13 (5.1 synapses/cell), Li19 (3.8 synapses/cell), LC6 (2.3 synapses/cell), and LT57 (2.0 synapses/cell).
Cluster 10 was another highly anisotropic cluster with synapse spread of 12.2 and 4.6 mm along short and long tangential axes (Fig. 8B). The major postsynaptic targets of cluster 10 other than LC10 were Li16 (5.5 synapses/ cell), Li12 (4.5 synapses/cell), LC15 (2.7 synapses/cell), and Li13 (2.4 synapses/cell). The dominant cell type of this cluster had a linear neurite with little branching that entered the lobula dorsally and then bent ventrally almost perpendicularly (Fig. 8H). While this cluster appears to be bistratified in Lo4 and Lo5B as a population (Fig. 8H), F, A population of cluster 9 cells, viewed from a tangential direction. G, Representative cluster 9 cells, viewed from a normal direction. H, A population of cluster 10 cells, viewed from a tangential direction. I, Representative cluster 10 cells, viewed from a normal direction. Four different subtypes are shown, as labeled. J, Morphology of TmY9 and Tm28 axon terminals, from Fischbach and Dittrich (1989). K, A population of cluster 11 cells, viewed from a tangential direction. The blue arrows indicate the ventral direction throughout. closer inspection revealed variations in the morphology of single cells (Fig. 8I): some cells were genuinely bistratified, but the branches were often longer in one layer than the other. Some other cells entirely lacked branching. Interestingly, some of these non-branching cells initially entered Lo5B and then switched to Lo4. This long, linear terminal morphology resembled TmY9 (Fischbach and Dittrich, 1989;Fig. 8J). TmY9 is known to have long, linear neurites similar to its lobula terminal in the medulla and lobula plate as well (Fischbach and Dittrich, 1989). Another known cell type with long, linear terminals oriented ventrally is Tm28 (Fischbach and Dittrich, 1989; Fig. 8J), although they appear to innervate a little deeper. Note that, although the drawing of TmY9 in Fischbach and Dittrich (1989) lacks branching, Musca Y17 cells, a putative TmY9 homolog, has a bistratified branching (Douglass and Strausfeld, 1998), similar to what is shown here.
Cluster 11 was a highly heterogeneous cluster with a variety of cell types in deeper layers of the lobula with elongated morphology along the long tangential axis (Fig. 8K). It appeared to contain cells similar to those in clusters through 8-10. The largest postsynaptic target of this cluster was LC10, but the mean synapse count per cell was only 1.4, likely reflecting the heterogeneity of the cluster.

Clusters 12 through 15
Clusters 12 through 15 are another set of clusters with large synapse spread along the long tangential axis (Figs. 3A, 9B,C). These clusters innervated shallower, around Lo2 to Lo4 (Fig. 9A), unlike the previous four clusters. Clusters 12 and 13 were homogeneous clusters dominated by vertically diffuse cells that shared connection to LT87 (10.9 and 19.0 synapses/cell for clusters 12 and 13, respectively), as well as to LC31 (2.9 and 12.2 synapses/ cell for clusters 12 and 13, respectively) and LC17 (2.4 and 9.9 synapses/cell for clusters 12 and 13, respectively; Fig. 9D). These cells had a meandering main neurite with many minor branches (Fig. 9E). In addition, they had a long secondary branch sticking out ventrally from the main neurite perpendicularly at around Lo3, which was also meandering and with many minor branches (Fig. 9E). Among previously documented cells, Tm19 appeared similar to these clusters in that it has an extensive secondary branch (Fischbach and Dittrich, 1989; Fig. 9F), but their innervation appears deeper.
Unlike clusters 12 and 13, clusters 14 and 15 were highly heterogeneous and it was difficult to make out the dominant cell types (Fig. 9G). The main postsynaptic targets of these clusters were LPLC1 (19.9 and 5.7 synapses/cell for clusters 14 and 15, respectively). Cluster 14 contained some T2 cells, as can be also seen from the UMAP embedding, where some cluster 14 cells are next to the T2-dominated cluster 5.

Clusters 16 and 17
The branch containing clusters 16 through 22 were characterized by their innervation being relatively restricted to the deep layers, Lo5B and Lo6 (Fig. 10A). Clusters 16 and 17 were monostratified clusters in Lo5B with small synapse spread around 3 mm in every direction (Fig. 10B,C). They also shared their connectivity to LT58 (8.4 and 12.8 synapses/cell for clusters 16 and 17, respectively), LC10 (4.3 and 4.2 synapses/cell for clusters 16 and 17, respectively), and LC24 (2.6 and 1.8 synapses/ cell for clusters 16 and 17, respectively). However, the two clusters appeared to have distinct terminal morphology. Cells in cluster 16 had more tangentially extensive terminals, which consisted of thin, knobby neurites branching perpendicularly, oriented posteriorly (Fig. 10D,  E). In contrast, the terminals of cluster 17 cells were thicker and branched less (Fig. 10G,H). The main neurite of these cells appeared to touch the bottom of Lo5B first and then curl upward, creating a distinctive looped morphology similar to those of T3 (Fig. 10H). In terms of connectivity, clusters 17 differed from cluster 16 in that it had a large number of synapses on mALC2 (9.6 synapses/ cell). The morphology of cluster 17 cells highly resembled that of Tm20 (Fischbach and Dittrich, 1989; Fig. 10I). The closest match we could find for cluster 16 was Tm11 (Fig.  10F), although we are not particularly confident about this identification (Fischbach and Dittrich, 1989). Tm20 is a cholinergic cell (Davis et al., 2020) that is directly postsynaptic to spectrally sensitive photoreceptor R8 (Gao et al., 2008;Takemura et al., 2015), necessary for color learning combinedly with Tm5 subtypes . Other major inputs to Tm20 include L2, L3, C3, Tm1, and Mi4 . LT11, a tangential projection neuron necessary for wavelength-specific phototaxis (Otsuna et al., 2014), was not among the strongest postsynaptic target of cluster 17, although LT11 was previously suggested to be downstream of Tm20 based on transsynaptic GFP reconstruction (GRASP; Lin et al., 2016).

Cluster 23
Cluster 23 was highly distinct from the neighboring clusters, as can be seen both from the dendrogram (Fig.  3A) and UMAP embedding (Fig. 3E). Cluster 23 consists of a small monostratified cells in Lo3, with synapse spread of around 2 mm in every direction (Fig. 13A-C). Its main postsynaptic targets were Li11 (8.5 synapses/cell), LT1 (4.8 synapses/cell), LC11 (2.6 synapses/cell), LPLC1 (2.1 synapses/cell), and LC21 (1.8 synapses/cell). Visually, cluster 23 appeared to consist of highly homogenous cells with a knobby terminal and a small protrusion at a shallower location (Fig. 13D,E). The best morphologic match we could find was T2a, a cell type morphologically and transcriptomically related to T2 and T3 (Fischbach and Dittrich, 1989;Keleş et al., 2020;Özel et al., 2021;Fig. 13F). Indeed, this cluster shared its connectivity to LT1, LC11, and LPLC1 with T3, indicating that it could be functionally similar to T3. While functional properties of T2a have not been studied, it shares inputs from Mi1 and Tm1 with T3, implying it is also an ON-OFF cell . T2a lacks inputs from L cells unlike T2, despite its proximal medulla innervation . Another cell type with similar terminal morphology was Tm21 (Fischbach and Dittrich, 1989; Fig. 13F).

Clusters 24 and 25
Clusters 24 and 25 were highly heterogeneous clusters, and it was difficult to make out the dominant cell types (Fig. 14D,F). The low synapse counts to even their largest targets (cluster 24: 2.16 synapses/cell to mALC2; cluster 25: 1.2 synapses/cell to LC4) and the very diffuse innervation depth profile (Fig. 12A) likely reflect this heterogeneity. Of note, cluster 24 contained a set of neurites that entered the lobula from the dorsal proximal end as if hugging the bottom of Lo6 (Fig. 14D,E). Cluster 25 contained a large number of fragmented neurites around the dent in the lobula (Fig. 14F) which was likely caused by how it was mounted for scanning. Thus, this cluster may be  Fischbach and Dittrich (1989). G, A population of cluster 21 cells, viewed from a tangential direction. H, Representative cluster 21 cells. I, Morphology of a Tm8 axon terminal, from Fischbach and Dittrich (1989). J, A population of cluster 22 cells, viewed from a tangential direction. K, L, Representative (K) thick and (L) thin cells in cluster 22. The blue and green arrows, respectively, indicate the ventral and anterior directions. driven by the imaging methodology rather than by the biology.
Cluster 27 was somewhat heterogeneous (Fig. 15G), and the only major postsynaptic target with more than a single synapse/cell was LC10 (9.6 synapses/cell). The dominant cell type had thin processes and terminals whose short branches had knobby endings (Fig. 15H). These features were similar to the thin cells in cluster 22 (Fig. 12K), but cluster 27 cells were less tangentially extensive and positioned slightly shallower. Cluster 28 was also dominated by thin, monostratified cells in Lo5B (Fig.  15I,J), which had knobby endings, similar to cluster 27.

Clusters 30 and 31
Clusters 30 and 31 were the only clusters with substantial Lo1 innervation (Fig. 16A). Lo1 appears to be a special layer dedicated to OFF edge motion detection, housing T5 and its inputs but avoided by other cell types, including lobula VPNs (Fischbach and Dittrich, 1989;Shinomiya et al., 2014Shinomiya et al., , 2019Wu et al., 2016). This is similar to how M10, the putative evolutionary precedent of Lo1 (Shinomiya et al., 2015), is dedicated to T4 and their inputs but avoided by other cells (Fischbach and Dittrich, 1989). Cluster 30 was a highly homogeneous cluster of small monostratified terminals (Fig. 16D), whose only major postsynaptic target was CT1 (20.9 synapses/cell). CT1 receives inputs from Tm1 as well as Tm9 in the lobula (Shinomiya et al., 2019). The terminal morphology of  Fischbach and Dittrich (1989). The green arrows indicate the anterior direction.
Cluster 31 was a large heterogeneous cluster mostly confined to Lo1 (Fig. 16A,F). This cluster appears as multiple distant clusters on the UMAP embedding (Fig. 3E), also likely reflecting its heterogeneity. Upon visual inspection, we found oriented neurites with dense branching similar to T5 dendrites, as well as small terminals that resembled Tm1, 2, and 9 (Fig. 16G). On average, cluster 31 did not have more than single synapse/cell to any of the labeled lobula neurons included in the current analysis. This observation supports the idea that Lo1 is the dedicated layer for motion detection by T5 and is relatively independent from the rest of the lobula.

Cluster 32 through 34
Clusters 32 through 40 were characterized by their tangentially compact and vertically diffuse morphology (Fig.  17B,C). The axon terminals of clusters 32 through 34 were in Lo5A/B, but they also had synapses in shallower layers (Fig. 17A). The major postsynaptic targets of cluster 32 were Li19 (12.9 synapses/cell), LC17 (9.3 synapses/cell), LT79 (8.4 synapses/cell), LPLC2 (8.1 synapses/cell), and LPLC1 (7.4 synapses/cell). These cell types overlapped with the main synaptic targets of cluster 26. In fact, the morphology of cluster 32 cells, with its thick neurites with small protrusions and somewhat densely branching terminals (Fig. 17D,E), resembled that of cluster 26 cells (Fig.  15D,E). Continuity of clusters 26 and 32 can also be seen from the UMAP embedding (Fig. 3E). The best morphologic match we could find for these clusters was Tm5Y (Fischbach and Dittrich, 1989;Fig. 17F).
Clusters 33 and 34 shared their connectivity to LT11 (20.7 and 11.2 synapses/cell for clusters 33 and 34, respectively), LC25 (7.2 and 4.2 synapses/cell for clusters 33 and 34, respectively), LC11 (5.8 and 1.4 synapses/cell for clusters 33 and 34, respectively), LC15 (2.3 and 1.1 synapses/cell for clusters 33 and 34, respectively), LC26 (1.6 and 1.1 synapses/cell for clusters 33 and 34, respectively). These were the only clusters with strong LT11 connectivity, which most likely made these clusters highly distinct from everything else on the UMAP embedding (Fig. 3E). Both clusters were visually homogeneous, consisting of cells characterized by modestly branching terminals in Lo5A/B and straight main neurite with small protrusions in shallower layers (Fig. 17G-I). Given their extensive connectivity to LT11, these clusters likely represent Tm5 neurons (Lin et al., 2016). Tm5 neurons, consisting of three subtypes Tm5a/b/c, are glutamatergic neurons postsynaptic to spectrally sensitive photoreceptors R7 and R8 (Gao et al., 2008). Tm5 subtypes, alongside with Tm20, have been shown to be combinedly necessary for color learning in flies Melnattur et al., 2014). Among these subtypes, Tm5c, resembled the cluster 33 and 34 cells the most in that it had a small protrusion in a shallower layer in addition to Lo5 (Gao et al., 2008).

Clusters 35 through 37
Clusters 35 through 37 were another set of vertically diffuse clusters (Fig. 18A), which on average appeared to be bistratified in Lo4/5A and Lo5B/6, but the pattern was less obvious compared with other bistratified clusters. The major postsynaptic targets of cluster 35 were Li16 (7.0 synapses/cell), Li12 (6.1 synapses/cell), LC15 (2.8 synapses/cell), LC10 (1.3 synapses/cell), and Li19 (1.1 synapses/cell). The dominant cell type of cluster 35 had branches sticking out of the main neurite perpendicularly, which aligned with the short tangential axis (Fig. 18D,E). This morphology strongly resembled that of TmY11 (Fischbach and Dittrich, 1989;Fig. 18F). The major postsynaptic targets of cluster 36 were LC20 (6.3 synapses/ cell), Li12 (2.3 synapses/cell), LC22 (1.2 synapses/cell), LC10 (1.2 synapses/cell), and Li19 (1.0 synapses/cell). The dominant cell type of this cluster appeared more vertically diffuse than cluster 35 (Fig. 18G,H), with many short protrusions along its main neurite ranging from L4 to L6. The best morphologic match we could find for this cluster was TmY5 (Fischbach and Dittrich, 1989; Fig. 18I). Cluster 37 was another large, heterogeneous cluster consisting of a variety of vertically diffuse neurons resembling cells from clusters 32 through 36 (Fig. 18J,K). Similar to cluster 25, cluster 37 contained many fragments around the dent in the sample. The postsynaptic cell types with more than a single synapse per cell were LC10 (1.8 synapses/cell) and LC20 (1.4 synapses/cell).

Clusters 38 through 40
Clusters 38 through 40 were vertically diffuse clusters with shallower layer innervations (Fig. 19A). Cluster 38 mainly innervated Lo1/2 and Lo3/4 (Fig. 19A). Their major postsynaptic targets were LC4 (4.8 synapses/cell), Li17 (3.9 synapses/cell), mALC1 (3.4 synapses/cell), LPLC1 (1.9 synapses/cell), and LC12 (1.6 synapses/ cell). Morphologically, cluster 38 appeared to contain at least two dominant cell types, which is likely why this cluster showed up as two distant clusters on the UMAP embedding (Fig. 3E). The first cell type was a short, vertical cell with Y-shaped ending at L4 and swelling in a shallower layer (Fig. 19D,E). This morphology, as well as its connectivity, made it resemble cells in cluster 6, which we guessed to be either Tm4 or TmY2 (Fig. 7G-I). The continuity of clusters 6 and 38 can be also seen from the UMAP embedding (Fig. 3E). The other cell type in cluster 38 reached a deeper layer, likely Lo5A (Fig. 7F), and had a knobby appearance with more branching than the first cell type. The best morphologic match we could find for this cell type was TmY7 (Fischbach and Dittrich, 1989;Fig. 19G).
Cluster 39 was a small cluster of homogenous cells innervating only the ventral rim of the lobula (Fig. 19H). These cells had branching terminals with knobby endings (Fig. 19I). The major postsynaptic targets of this cluster were Li14 (13.1 synapses/cell), LC10 (7.3 synapses/cell), PS179 (3.9 synapses/cell), LPLC2 (3.3 synapses/cell), and LC14 (2.7 synapses/cell). Extensive connectivity to Li14 as well as PS179 made this cluster unique. It is unclear whether this cluster constitutes a retinotopically specialized cell type by itself. Alternatively, this could be a subset of a cell type whose other cells are lost in large clusters like cluster 37, clustered separately because of errors in innervation depth estimation caused by high curvature of the lobula layers at the rim.
Cluster 40 was the most vertically diffuse cluster with the synapse spread of 11.0 mm along the normal axis (Fig.  19C). This cluster was a homogeneous cluster of bistratified cells in Lo3/4 and Lo6, with minor protrusions in the layers in between (Fig. 19A,J,K). The major postsynaptic targets of this cluster were LC10 (4.3 synapses/cell), Li11 (3.1 synapses/cell), LPLC2 (2.2 synapses/cell), LT51 (2.2 synapses/cell), and LC4 (1.8 synapses/cell). The morphology of these cells resembled that of TmY5a (Fischbach and Dittrich, 1989; Davis et al., 2020; Fig. 19K,L), for example, although we are not particularly confident about this identification.

Discussion
The lobula is one of the deepest neuropils in the fly visual system and is thought to detect behaviorally relevant visual features and send outputs to the central brain (Otsuna and Ito, 2006;Mu et al., 2012;Wu et al., 2016). While a number of cell types providing inputs to and sending outputs from the lobula have been characterized, connectivity among them, critical information to understanding the computations taking place in the lobula, have been largely missing. In the present study, we attempted to fill this gap by performing connectivity-based and morphology-based clustering on fragmented lobula axon terminals in the hemibrain dataset (Scheffer et al., 2020). Among the 40 clusters, we were able to find some morphologically matching cell types for 24 clusters, with varying degrees of confidence (Table 1). The cell types we identified with most confidence were as following: T3 (clusters 1 through 4), T2 (cluster 5), TmY29 and/or Tm28 (cluster 10), Tm20 (cluster 17), Tm9 (cluster 30), and TmY11 (cluster 35). These were mostly cell types that had distinctive terminal morphology in the lobula (Fischbach and Dittrich, 1989). The other identifications, Tm4 and/or TmY2 (clusters 6 and 38), Tm19 (clusters 12, 13), Tm11 (cluster 16), TmY10 (cluster 19), Li1 (cluster 20), Tm8 (cluster 21), T2a and/or Tm21 (cluster 23), Tm5Y (clusters 26 and 32), Tm5Y (cluster 32), Tm5 (clusters 33 and 34), TmY5 (cluster 36), TmY5a (cluster 40), were more speculative.
The current dataset lacks the information required to definitively identify cell types: the innervation patterns in the medulla and lobula plate. Future connectomic reconstructions encompassing the entire optic lobe will be necessary to unambiguously identify the putative cell types identified here. This can be in principle done by tracing in the FAFB EM volume (Zheng et al., 2018) on the FlyWire website (Dorkenwald et al., 2021). An alternative approach is to use molecular tools to map connectivity among genetically defined cell types, such as trans-synaptic GFP reconstitution (GRASP; Feinberg et al., 2008;Shearin et al., 2018), trans-Tango (Talay et al., 2017), and optogenetics (Simpson and Looger, 2018). This approach requires generating or identifying selective transgenic drivers to target new Tm and TmY neurons, and will be required to define functional connectivity, rather than anatomic connectivity. Neuroanatomical tools such as color depth maximum intensity projection mask search (Otsuna et al., 2018;Clements et al., 2020b) as well as marker gene identification with single cell RNA sequencing approaches (Özel et al., 2021) will be useful to this end. We found clusters consisting of apparently homogeneous cells whose morphology did not match any previously reported Tm or TmY cells (Fischbach and Dittrich, 1989;clusters 8, 9, 18, 22, 27-29, 39), likely representing new cell types. Given the existence of these novel neuron types as well as some mixture clusters, the 40 clusters we chose is likely smaller than the true number of Tm and TmY types.
While most of the putative cell types identified here have never been studied physiologically, there are some exceptions. For example, Tm4 (clusters 6/38) and Tm9 (cluster 30) are OFF cells presynaptic to T5 Serbe et al., 2016). Tm5 (clusters 33/34) and Tm20 (cluster 17) are spectrally sensitive cells downstream of the inner photoreceptors (Gao et al., 2008;Melnattur et al., 2014;Lin et al., 2016). T2 (cluster 5) and T3 (clusters 1-4) are known LC/LPLC inputs with selectivity for small moving objects Clark, 2020, 2021). Interestingly, the spectrally sensitive Tm neurons and object selective neurons appeared to converge on several LC neurons. For example, both LC11 and LC15 received sizable inputs from putative T3 and Tm5, suggesting that these cell types may prefer objects of specific colors. For lobula input neurons without known physiology, their connectivity to previously characterized LC/LPLCs can hint at their function. For example, cluster 32 neurons (putative Tm5Y) provide sizable inputs to LC6, 9, 13 and LPLC1, 2. LC6 (Wu et al.,  Fischbach and Dittrich (1989). The green arrows indicate the anterior direction. 2016; Morimoto et al., 2020), LPLC1 (Tanaka and Clark, 2021), and LPLC2 (Klapoetke et al., 2017;Ache et al., 2019) are sensitive to looming stimuli, suggesting that Tm5Y might play an important role in loom detection. Another example is cluster 10 neurons (TmY9/Tm28), which selectively connect to LC15. LC15 responds preferentially to moving, narrow vertical bars . The highly anisotropic dendrites of TmY9/Tm28 oriented along the dorsoventral axis (Fischbach and Dittrich, 1989) could be well suited to pool excitation along the vertical visual axis to detect long bars. Overall, the analysis presented here advances our knowledge of lobula connectivity and generates testable hypotheses for how lobula visual projection neurons become selective for specific visual features.