Connectivity establishes spatial readout of visual looming in a glomerulus lacking retinotopy

Visual systems can exploit spatial correlations in the visual scene by using retinotopy, the organizing principle by which neighboring cells encode neighboring spatial locations. However, retinotopy is often lost, such as when visual pathways are integrated with other sensory modalities. How is spatial information processed in the absence of retinotopy? Here, we focused on visual looming responsive LC6 cells in Drosophila, a population whose dendrites collectively tile the visual field, but whose axons form a single glomerulus—a structure lacking retinotopic organization—in the central brain. We identified multiple glomerulus neurons and found that they respond to looming in different portions of the visual field, unexpectedly preserving spatial information. Through EM reconstruction of all LC6 synaptic inputs to the glomerulus, we found that LC6 and downstream cell types form circuits within the glomerulus that establish spatial readout of visual features and contralateral suppression—mechanisms that transform visual information for behavioral control.


Introduction 28 29
In many animals, brain regions involved in processing visual information are large and well 30 organized, featuring retinotopy-an organizational plan that preserves the mapping of space 31 originating in the retina, such that neighboring neurons respond to visual signals at 32 neighboring spatial locations. Animals need visual-spatial information in order to direct their 33 escape away from predators, to find mates, or to catch prey. Classical studies in cats (Hubel 34 and Wiesel, 1962) and monkeys (Tootell et al., 1988), and recent work in mice (Garrett et al.,35 2014) demonstrated that the retinotopic organization of higher visual areas is a dominant 36 feature of the organization of the mammalian brain. While this organizing principle has 37 facilitated detailed analyses of visual pathways, it remains unclear how visual-spatial 38 information is processed in the absence of retinotopy. In particular, retinotopy must be 39 sacrificed where visual information is integrated with other modalities and as vision is 40 translated into behavioral actions. We propose that rapid progress can be made on 41 understanding these critical transformations in the Drosophila brain. resemble natural behaviors such as escape jumps, backward walking, courtship behavior, and 50 reaching (Wu et al., 2016). Furthermore, the cell types that elicited avoidance behaviors 51 (LC4, LPLC2, LC6 and LC16) also responded to visual looming stimuli (Ache et al., 2019; 52 LC6 in magenta, candidate downstream LC6Gs in green). Flies were dissected and whole 162 brains (not including the photoreceptors) were imaged using two-photon microscopy ( Figure  163 2A, right). To obtain an LC6 activity-dependent tuning curve for responses of downstream 164 neurons, we first conducted a "calibration" experiment to select a series of increasing light 165 stimulation intensities that would evoke activity in LC6s with monotonically increasing 166 responses. The stimulation light was spatially restricted to the glomerulus being imaged 167 (detailed in methods). We established a six-pulse stimulation protocol with ramping light 168 intensity with which we could observe a monotonically increasing LC6 GCaMP signal in Having established multiple cell types as connected to LC6, we next examined whether these 192 cells respond to the same visual looming stimuli that selectively activate LC6 neurons (Wu et 193 al., 2016). For further analysis we selected the strongly connected, bilaterally projecting Chen et al., 2013) using the same split-GAL4 driver lines used for the targeted functional 196 connectivity experiments, and measured calcium responses from the LC6Gs (summed 197 response from ~four cells per brain hemisphere in each cell type) to visual stimuli using in 198 vivo two-photon microscopy ( Figure 3A, left). As with LC6 neurons, both LC6G types 199 responded to visual looming stimuli (example response for LC6G2 recordings in Figure 3A, 200 middle). However, the initial measurements appeared to show substantial selectivity for 201 stimuli presented in different spatial locations. In light of the anatomy of the glomerulus, this 202 spatial selectivity was unexpected. To confirm this, we mapped the responses to small 203 looming disks (maximum size 18°) presented at all locations on a grid covering a large region 204 of the right eye's field of view. The resulting spatial Receptive Field (RF) is shown in Figure  205 3A (right). We precisely measured the head orientation (  Given their strikingly different morphologies, we investigated whether these neurons 230 differentially process visual information from LC6s. We presented a panel of visual stimuli 231 (Table 2) to compare calcium responses of LC6G1s and LC6G2s to those of LC6. As 232 previously reported, LC6 preferentially responded to dark looming stimuli, and also 233 responded to the individual features of a looming stimulus (darkening in the luminance-234 matched stimulus and the edge motion in the looming annulus stimulus), but not to a receding 235 dark stimulus or a bright looming stimulus (Wu et al., 2016). LC6 neurons also responded to 236 the motion of non-looming objects ( Figure 4B; response time series shown for slowest speed, 237 tuning curves on the right for other speeds). The bilaterally projecting LC6G1 neurons 238 responded to the looming related stimuli as well as the bar and small object motion stimuli in 239 a manner that was very similar to the LC6 population ( Figure 4B, C). This similarity of LC6 240 and LC6G1 responses extended to all stimuli presented, which is well illustrated by the 241 scatter plot comparing the responses of the two cell types ( Figure 4C, left). The regression 242 line relating (normalized) LC6 to LC6G1 responses is close to the unity line (slope=1.05, 243 Pearson's correlation r=0.87), suggesting that LC6G1 may serve as a 'summary' of the LC6 244 populations' visual responses across all stimuli. By contrast, the ipsilaterally projecting 245 LC6G2s were much more selective for looming stimuli, responding to dark looming and to 246 the edge motion of the looming annulus, but not to the other stimuli ( Figure 4B). The 247 dissimilarity of responses can also be observed from the comparison scatter plot ( Figure 4C, 248 right), where the points are more dispersed away from the unity line, and the linear regression 249 shows a weaker correlation (slope=0.51, r=0.40). While LC6G1 relays the LC6s' visual 250 responses (presumably to the contralateral LC6 glomerulus) the LC6G2 neurons is more 251 selective for looming than non-looming stimuli. 252 253 LC6 neurons exhibit biased connectivity in the glomerulus 254 255 Our functional analysis establishes that LC6G neurons encode the visual-spatial information 256 of looming stimuli from their inputs in the glomerulus, but anatomical analysis of the LC6 257 glomerulus from light microscopy images does not account for this selectivity (Figure 1). 258 One parsimonious explanation is that this spatially selective readout of LC6 neurons could 259 result from patterned subsets of synaptic connections between LC6 neurons and their 260 downstream targets. To test this hypothesis, we mapped the LC6 neurons in a recently 261 page 8 / 67 completed serial section transmission electron microscopy volume of the full adult brain 262 (Zheng et al., 2018). From a single side (by convention, the right hand side, RHS) of the 263 brain, we found 65 LC6 neurons (single example in Figure 5A), which we confirmed based 264 on their morphology in the lobula, the distinctive axonal loop tract (shared with LC9), and the 265 orientation of the axons in the glomerulus ( Figure 5B). We then reconstructed all LC6 266 neurons, completing their axon terminals in the glomerulus, and tracing the major, but not the 267 finest neurites in the lobula. Tracing the major dendritic branches in the lobula enabled us to 268 computationally estimate the corresponding region of the visual field for each LC6 neuron 269 ( Figure 5C and methods). The coverage of the lobula is nearly uniform, but with a higher 270 density toward the visual midline ( Figure 5-figure supplement 1A). Using this mapping 271 procedure, we constructed an 'anatomical receptive field' of LC6s, an RF estimate based only 272 on the dendritic location and extent of each of the 65 reconstructed LC6 neurons ( Figure 5D). 273

274
We found and labeled many synapses between LC6 neurons in the glomerulus (1902 total, a 275 mean of ~29 synapses to each LC6 neuron from all other LC6 neurons, Table 3). We used 276 this connectivity to look for a correlate of retinotopy in the glomerulus: do LC6 neurons 277 preferentially synapse onto LC6 neurons with nearby dendrites (and thus spatially closer 278 receptive fields)? When the distance between RF centers (measured across the lobula) was 279 weighted by the number of synapses between the LC6 cells, we found that the average 280 distance between connected neurons is significantly smaller than the distances after random 281 shuffling of these connections ( Figure 5E). Therefore, LC6 neurons bias their connectivity, 282 such that they connect to their lobula neighbors with a distribution significantly smaller than 283 produced by random connections. 284 285 Since LC6 axons preferentially synapse onto other LC6 cells with neighboring dendrites, we 286 next examined whether these synapses exhibit any spatial organization. To visualize this, we 287 divided the glomerulus into 10 compartments along its long axis, each containing the same 288 number of LC6 pre-synapses (across all postsynaptic cells; see 'computational analysis of 289 EM reconstruction' in methods; Figure 5F). We then examined the spatial distribution of the 290 LC6 input within each compartment as a contour plot in visual coordinates (see methods). 291 We found that the most lateral glomerulus compartment (#1) is biased towards representing 292 the anterior visual field and the most medial glomerulus compartment (#10) is biased towards 293 the posterior visual field. Most remaining compartments sample visual inputs more broadly, 294 with peak responses close to the center of the eye (consistent with broad inputs covering 295 much of the visual field; Figure 5-figure supplement 1B). This result was somewhat 296 surprising-even though we could not find any obvious retinotopy of the LC6 axons in the 297 glomerulus with light level anatomical analysis (Figure 1), the preference for synapsing onto 298 LC6 neurons whose dendrites are neighbors in the lobula, as well as the spatial organization 299 of these synapses in the glomerulus demonstrates that some retinotopy is maintained, but it is 300 only observable in the organization of synapses. 301 The biased distribution of LC6-LC6 synapses demonstrates that some retinotopic 306 organization exists in the glomerulus, but can the readout of LC6 inputs account for the 307 spatially selective responses of the LC6G neurons? We manually reconstructed neurons 308 within the LC6 glomerulus (by following, at random, neurites postsynaptic to LC6s), until we 309 could match reconstructed neurons to the bilateral LC6G1 and ipsilateral LC6G2 cell types 310 that were examined in Figure 3 and 4. We performed our analysis on the processes of four 311 bilateral LC6G1 neurons (two each, per side of the brain) and on the unilateral axonal arbors 312 of five LC6G2 neurons. These nine neuronal arbors were fully reconstructed and 313 independently proofread (in the glomerulus; examples in Figure 6A). The connectivity 314 between these nine arbors and the 65 LC6 neurons is summarized in Figure 6B (and detailed 315 in Table 3). As expected from the morphology of these LC6 target neurons, and the 316 functional connectivity (Figure 2), we found many synapses between the LC6 neurons and 317 the right hand side (RHS) LC6G1 and LC6G2 neurons (mean number of synapses: ~912 and 318 ~428, respectively). Perhaps more surprising is that we also found LC6 synapses onto the 319 axons of the left hand side (LHS) LC6G1 neurons (mean of ~207; LC6G1 (L) in Figure 6B Figure 6C) 332 across the field of view of the entire right eye. This analysis shows several features that agree 333 with our functional RF mapping of these cell types (Figure 3). The LC6G1 receptive field 334 appears broader and has large responses close to the frontal midline, while the LC6G2 335 neurons have their strongest responses in a smaller zone, away from the midline. To generate 336 a more direct comparison, we aligned the functional RFs measured through in vivo calcium 337 imaging (generated with stimuli that partially covered the visual field; Figure 3C) to the 70% 338 contour of the anatomical estimate for each reconstructed cell type ( Figure 6D). Considering 339 that these estimates of the RF of each cell type were generated with very different methods, 340 and accounting for the approximations required to align the two data sets, we find this level 341 of agreement between these RFs to be substantial. One prominent feature of the fly nervous system is that brain compartments are often 360 connected-directly or indirectly-to their symmetric compartment across the hemispheres identified (Figure 1), only the LC6G1 neurons consist of a symmetric population of neurons 363 that cross the midline and innervate both LC6 glomeruli. The set of neurons whose 364 interconnections we completely reconstructed within the LC6 glomerulus are fully detailed in 365 Table 3 and summarized in Figure   In this study we investigated the circuitry downstream of the looming responsive LC6 visual 397 projection neurons. We used light-level anatomical analysis to identify candidate downstream 398 neurons (Figure 1 and 1-figure supplement 1) and then used intersectional genetic methods 399 to establish driver lines that precisely target each of these putative downstream target neurons 400 ( Figure 1-figure supplement 2). We used these driver lines to show that five of these cell 401 types are functionally connected to LC6 neurons ( Figure 2). We then selected two of these 402 cell types, the bilaterally projecting LC6G1 neurons and the ipsilateral LC6G2 projection 403 neurons for further functional studies. We found that these two cell types differentially 404 encode looming stimuli (Figure 3 and 4). The LC6G1 neurons responded to looming stimuli 405 over a large part of the visual field and showed nearly identical stimulus selectivity as the 406 LC6 neurons. By contrast, the LC6G2 neurons responded to looming stimuli within a more 407 restricted region of the visual field, and showed enhanced stimulus selectivity, preferentially 408 encoding looming stimuli, while being less responsive to non-looming stimuli than their LC6 409 inputs. 410

411
The organization of visual projection neuron axons into glomeruli in the fly brain has been 412 the subject of significant speculation. One popular proposal, inspired by the loss of retinotopy 413 within the glomeruli, is that these structures represent a transformation from a visual, spatial 414 . We tested this hypothesis using bilateral looming stimuli 516 while imaging from LC6G2 and found that the stimulus-evoked responses were consistent 517 with the proposed circuit mechanism. Contralateral suppression was already implied from 518 light level anatomy that showed the bilateral downstream cell type specifically contacted the 519 same glomeruli in both hemispheres. This type of connection between glomeruli is common; 520 based on our analysis of Janelia's GAL4 collection, we expect that most if not all optic 521 glomeruli are connected by one or more bilaterally projecting interneuron types. Also, 522 contralateral suppression has been observed in other contexts and brain regions in 523 Drosophila CNS. In addition to being common, it is also sensible (and is related to the 528 engineering principle of designing amplifiers with 'common mode rejection'). A major goal 529 of sensory systems appears to be the detection and localization of specific sensory cues, and 530 yet in the natural world cues are rarely discrete, they emerge in complex mixtures across localizable threat, such as an approaching object, is to compare whether this detection event 533 is restricted to one side of the animal, or whether it is prominent on both sides. We propose 534 that contralateral suppression is likely to be a basic strategy used in establishing stimulus 535 selectivity. 536 537 Outlook 538 539 Here we described three types of spatial organization within the LC6 glomerulus synthesizing 540 data from in vivo calcium imaging, functional connectivity, and light and EM level anatomy. 541 Using cell type specific driver lines (Wu et al., 2016)  to changes of brightness and contrast. Most other anatomy panels show composites of 586 multiple registered images, which were generated using a recently described template brain 587 (Bogovic et al., 2018). To more clearly display the cells of interest images used in these 588 panels were, in some cases, manually edited to exclude additional cells or background present 589 in the original images. Editing and assembly of these composites was mainly done using 590 believe this alignment to be accurate to within 5°-10°, or the spacing between 1-2 ommatidia. 720 Code for this transformation is posted at https://github.com/reiserlab/LC6downstream. 721 722 Looming stimuli were used to map the receptive fields of the LC6G neurons (stimulus type 1, 723 Table 2). Dark discs expanding from ~4.5° to ~18°, at constant velocity of 10°/s, were 724 presented in a grid of 7 × 14 positions, with the centers of the looming stimuli separated by 725 ~9° (Figure 3). The grid in Figure 3A  Surprisingly, the LC6\LC9 tract was shifted somewhat medially in this brain, a feature that 804 has also occasionally been observed with light microscopy (http://flycircuit.tw). The 805 morphology of each neuron's dendrites was only approximately reconstructed, favoring 806 completion of the largest branches but not completing the finest dendritic branches. We 807 believe this is adequate for determining the retinotopic location of the center of each neuron's 808 anatomical receptive field, but the boundary of each receptive field might be underestimated. 809 The axon terminals of each LC6 neuron were completely reconstructed and all identifiable 810 synapses on the axons were tagged. The volume demarcated by all LC6 pre-synapses was 811 used to generate a mask identifying the LC6 glomerulus. We then followed the LC6 synapses 812 within the glomerulus to identify synaptic partners. By comparing the synaptic partners with 813 the morphology of the expected target neurons from light microscopy data (Figure 1 and 814 Figure 1-figure supplement 1), we identified the target neurons which we further 815 reconstructed for the analysis in Figure 6. We reconstructed 2 LC6G1 bilateral neurons with 816 cell bodies on the right-hand side and two with cell bodies on the left-hand side. We initially 817 reconstructed 2 of the ipsilateral LC6G1 neurons as well, but after finding variable receptive 818 field structures we searched extensively and found 3 more neurons of this same morphology.
All target neurons' morphologies (except for the left-hand-side LC6G1 neurons) as well as 820 the connections between the target neurons and LC6 neurons within the right-hand side 821 glomerulus were traced to completion and reviewed by an independent team member. 822 823

Computational analysis of EM reconstructions 824
All EM analyses and visualizations were carried out in the R language (R Core Team, 2013) 825 using open-source packages, mainly the "natverse" (http://natverse.org/) (Bates et al., 2019). 826 In order to estimate the retinotopic correspondence for each reconstructed LC6 neuron we 827 mapped the visual field of one eye onto a layer of the lobula. We fit a 2 nd order curved 828 surface through the dendrites of all LC6 neurons ( Figure 5B, C). In an ongoing effort we are 829 reconstructing multiple neurons types throughout the medulla to precisely identify the 830 location of every column in the right optic lobe. While this effort is beyond the scope of the 831 work we describe in this manuscript, we used these data to identify two medulla columns that 832  Figure 6D). All reconstructed neurons described in the manuscript will be 877 available at https://fafb.catmaid.virtualflybrain.org/, and all data analysis code is available at 878 https://github.com/reiserlab/LC6downstream. 879 880 Acknowledgements 881 We thank the Janelia Fly Light Project Team for help with imaging driver lines and 882 processing of FISH samples. We thank Jasmine Le for training and assistance with the 883 functional connectivity experiments. We thank members of the Bock lab, especially Scott 884 Lauritzen, for supporting our early EM reconstruction efforts and Gregory Jefferis for 885 introducing us to his tools for computational neuroanatomy. We thank Janelia's Connectome    Figure 1H-M rotated by 90° around the mediolateral axis. 1101 Figure 1  1102   H-M (G, G') and Figure 1-figure supplement 1A-F (H, H'). The LC26 glomerulus outline 1103 is based on syt-HA expression in LC26 driven by an LC26 split-GAL4 driver (Wu et al., 1104(Wu et al., 2016 Table 1     pre-stimulus baseline (p<0.1, two-sample t-test). See Table 1 Table 1 for detailed summary of fly lines used in this study and Table 2 Table 1 for detailed summary of driver lines used in this study.     Table 1 for detailed summary of fly lines used in this study and Table 2 for 1204 summary of all visual stimuli presented. 1205          2   38  39  18  29  34  38  31  36  33  40  23  12  27  31  27  13  26  21  17  38  23  32  33  22  23  27  18  20  33  30  31  15  33  38  39  28  21  21  30  19  30  38  22  31  30  39  27  36  39  35  44  20  16  22  32  32  32  26  14  28  24  13  33  31  23  4  1 1 4 # connections Table 3: Connectivity matrix of LC6 and target neurons from EM data, related to Figure 6