The cell adhesion molecule Sdk1 shapes assembly of a retinal circuit that detects localized edges

Nearly 50 different mouse retinal ganglion cell (RGC) types sample the visual scene for distinct features. RGC feature selectivity arises from their synapses with a specific subset of amacrine (AC) and bipolar cell (BC) types, but how RGC dendrites arborize and collect input from these specific subsets remains poorly understood. Here we examine the hypothesis that RGCs employ molecular recognition systems to meet this challenge. By combining calcium imaging and type-specific histological stains, we define a family of circuits that express the recognition molecule Sidekick-1 (Sdk1), which include a novel RGC type (S1-RGC) that responds to local edges. Genetic and physiological studies revealed that Sdk1 loss selectively disrupts S1-RGC visual responses, which result from a loss of excitatory and inhibitory inputs and selective dendritic deficits on this neuron. We conclude that Sdk1 shapes dendrite growth and wiring to help S1-RGCs become feature selective.


25
The factors that guide developing arbors to synapse with appropriate targets within a layer are not well understood. An 26 initial idea, called Peter's Principle, posited that developing neurons synapse with nearby cells according to how often 27 they make contact (Binzegger et al., 2004;Peters and Feldman, 1976;Shepherd et al., 2005). However, recent 28 connectomic studies of the IPL demonstrate no obvious relationship between contact frequency and synapse number 29 (Briggman et al., 2011;Helmstaedter et al., 2013). Instead, these connectomic data support a model in which neurons 30 recognize targets in their immediate vicinity and synapse specifically with them. Key molecules in this recognition process 31 are thought to be members of the immunoglobulin superfamily (IgSFs).

33
Briefly, IgSFs are adhesion molecules that bind to themselves (homophillic) or compatible IgSFs (heterophilic) across cell-34 cell junctions. It has been proposed that selective IgSF expression within synaptic partners allows their neurites to adhere 35 and synapse when they encounter each other (Sanes and Zipursky, 2020). A recent study in mouse retina provides 36 support for this view (Krishnaswamy et al., 2015). In this study, the IgSF Sidekick-2 (Sdk2) in VG3-ACs and W3B RGCs, 37 drives these neurons to synapse with each other far more than they synapse with the Sdk2 negative neurons they 38 contact. Loss of Sdk2 ablates the enhanced VG3-W3B connectivity but does not alter the gross structure or overlap of 39 (Nefh), osteopontin (Ost) and the intracellular calcium buffer calbindin (Calb); (2) M2-RGCs should express high levels of 79 Ost+, low levels of Nefh, and low levels of Calb; (3) two predicted novel RGCs should express the steroid hormone

84
To validate these predictions, we took advantage of the mosaic arrangement of retinal types. Briefly, retinal neurons of the 85 same type are spaced apart at a characteristic distance, whereas neurons of different types are spaced randomly. Thus, 86 the density of cells labelled with a candidate marker will drop at short distances from a reference cell if the candidate 87 labels a single type. Sdk1 CG wholemounts stained with antibodies to Nefh and Ost showed a pair of Nefh+/Ost+ and Nefh-88 /Ost+ mosaics spaced at distances expected for the ON-RGC ( Figure 1H) and M2-RGC (Figure 1 -figure supplement 89 1C), respectively. A high-density mosaic was labelled by Brn3c ( Figure 1I), and a final pair of mosaics was found to be 90 Nr2f2+/Calb-and Nr2f2+/Calb1+ ( Figure 1J-K). We were unable to find an RGC that corresponded to the sixth Sdk1 91 cluster, as a stain with all these RGC markers and Ap2- labelled all GFP+ neurons in the GCL (Figure 1figure   92 supplement 1E-G). Thus, we provide molecular definition for 5 Sdk1-RGCs, which include 3 predicted novel types.

94
We used two approaches to define the anatomy of Sdk1-RGCs. In one approach, retrogradely infecting AAVs bearing 95 Cre-dependent reporters were delivered to the lateral geniculate nucleus (LGN) or superior colliculus (SC) of Sdk1 CG 96 mice. In the other, tamoxifen was used to drive reporter expression in a related strain (Sdk1 CE ) whose Sdk1 gene is   Figure 1M,Q). We refer to these Brn3c+ 102 neurons S1-RGCs. Ost+/Nefh-and Nr2f2+ RGCs were rarely observed using this method. These results indicate that only 103 a subset of Sdk1-RGCs project strongly to the LGN and SC.

105
One possibility for the rare observation of Ost+/Nefh-and Nr2f2+ RGCs is that these RGCs innervate 'non-imaging 106 forming' brain regions more strongly than they innervate the LGN and SC (Martersteck et al., 2017). Consistent with this, 107 brain sections taken from intraocularly infected Sdk1 CG mice showed reporter-labelled RGC axons in the non-image  and wholemounts (D-F) from Sdk1 CG mice stained with antibodies to GFP and the BC marker Chx10 (A, D), AC marker AP2 (B,E), and RGC marker RBPMS (C,F). Scale = 25m. G. Bar   RGCs show that these cells target a common set of IPL laminae. Thus, we conclude that Sdk1 defines a family of neurons 129 whose circuits are contained within the inner 3 layers of the IPL.

133
Registering calcium imaged fields to posthoc stains assigns molecular identity to RGC response.

135
To characterize the functional properties of the Sdk1 RGC family, we devised a procedure to relate marker-gene

141
Next, we fixed and stained retinae for GFP, Sdk1-RGC markers, and vessels, and re-imaged these retinae with a confocal     S1-RGCs (Brn3c+) showed ON responses to a full field flash but responded to both the leading and trailing edge of the 187 bright moving bar ( Figure 2E). These results indicate that S1-RGCs can respond to both bright and dark stimuli that are 188 localized to their receptive field. Larger stimuli appear to recruit surround mechanisms, which strongly attenuate S1-

189
RGCs, particularly its OFF responses. Plotting S1-RGC bar responses against bar direction showed that many cells 190 respond to bars traveling along the same axis, suggesting that these neurons detect stimulus orientation ( Figure 2E).

191
Computing an orientation selectivity index (OSI) for these cells showed a higher orientational preference in this neuron as 192 compared with the other Sdk1-RGCs ( Figure 2I), however, these values are significantly weaker than the OSI found in 193 recently described orientation-selective RGCs Schwartz, 2016, 2017). Thus, we conclude that S1-RGCs are 194 ON-OFF cells that respond best to edges that fall within their receptive fields.

196
Sdk1 loss selectively impairs S1-RGCs visual responses.  Figure 3B). S1-RGCs were most affected by Sdk1 loss and showed a significant decrease in 219 response magnitude to bar stimuli ( Figure 3D). These results indicate that Sdk1 loss impairs S1-RGCs but leaves the 220 other Sdk1-RGCs relatively unaffected. Interestingly, single-cell sequencing data suggested that the other Sdk1-RGCs     . Selective loss of visual responses on S1-RGCs in the absence of Sdk1. A. Cartoon of S1-RGCs and ON-RGCs labelled by delivering retrogradely infecting AAVs bearing cre-dependent reporters in the lateral geniculate nucleus (LGN) or superior colliculus (SC) of Sdk1 CG mice. Targets of other Sdk1-RGCs that project to olivary pretectal nuclei (OPN) and medial terminal nucleus (MTN) are also shown. B. Sample two-photon image of a Sdk1 CG/+ retina labelled as described in A showing a large soma ON-RGC and a small soma S1-RGC. Sulphorhodamine101 labels vessels in the GCL. Scale = 25m. C. Polar plots of spike responses to a bar moving in 8 different directions recorded from example ON-and S1-RGCs in experiments like those shown in B. D-E. Raster of spike responses to an expanding flashing spot recorded from example S1-RGCs (D) and On-RGCs (E) in Sdk1 CG/+ (Het) or Sdk1 CG/CG retinae (KO). F-G. Average S1-RGCs firing rates versus bright (F) or dark (G) spot diameter measured from experiments like those shown in D. H. Average On-RGCs firing rate versus bright (ON) or dark (OFF) spot diameter measured from experiments like those shown in E. I-J. Raster of spike responses to centered dark or bright bar rotating through 8 orientations recorded from S1-RGCs (I) and On-RGCs (J) in Sdk1 CG/+ (Het) or Sdk1 CG/CG retinae (KO). K-L. Average firing rate versus bar orientation for S1-RGCs measured from experiments like those shown in I. M. Average firing rate versus bar orientation for ON-RGCs measured from experiments like those shown in J. N. Average orientation selectivity indices computed for S1-RGC and ON-RGC responses to the oriented bar stimulus in control (Het) and Sdk1-null (KO) retinae (n=7 for Sdk1 CG/+ ON RGCs, n=12 for Sdk1 CG/+ Brn3c RGCs, n=14 for Sdk1 CG/CG Brn3c RGCs, n=6 for Sdk1 CG/CG ON RGCs ; * = p <0.05). arbors, indicating that the receptive field center on these neurons is ~120-150m ( Figure 4F-G). Recordings from control 263 ON-RGCs displayed higher baseline spiking than S1-RGCs and sustained ON responses that were poorly suppressed 264 by large stimuli (Figure 4E), consistent with their full-field and moving bar responses in our calcium imaging experiments.

265
In contrast, the same recordings from S1-RGCs in Sdk1 CG/CG retinae showed a dramatic loss of responsivity to dark 266 stimuli and significantly weaker responses to ON stimuli ( Figure 4D, F-G). Recordings from nearby Sdk1-null ON-RGCs 267 showed comparable responses to their control counterparts ( Figure 4E, H), indicating that Sdk1 loss selectively affects 268 S1-RGCs.

270
We found that most S1-RGCs showed responses to axial bar motion, similar to what we saw in our calcium imaging 271 experiments. We suspected the inability to align bars with S1-RGC receptive fields in the imaging studies could have 272 activated their strong surround and attenuated their responses to this stimulus. We revisited the idea that these neurons 273 might exhibit sensitivity to stimulus orientation, presenting stationary bars whose width matched the receptive field size of 274 S1-RGCs and rotated through 8 different orientations. As expected from their responses to moving bars, S1-RGCs

279
RGCs showed similar responses to those in controls ( Figure 4J, M). Computing OSI values for these cells and comparing 280 the mean OSI for S1-RGCs and ON-RGCs in controls and knockout retina showed a selective reduction of orientation 281 selectivity for S1-RGCs in the absence of Sdk1 ( Figure 4N). Thus, we conclude that Sdk1 is required for S1-RGCs to 282 develop normal responses to visual stimuli.

286
The deficits we observed on S1-RGCs in Sdk1 CG/CG retinae might result from a loss of excitatory inputs, a change in 287 inhibitory inputs, or both. To investigate this idea, we recorded synaptic currents from S1-RGCs and ON-RGCs in control 288 and Sdk1 null retinae and compared their responses to visual stimuli. We began with expanding spots, isolating excitatory 298 299 zFinally, we examined the synaptic currents evoked on S1-RGCs to our oriented bar stimuli. Controls showed inward and 300 outward currents to both bright and dark oriented bars, but the magnitudes of these currents varied with bar orientation Whole-cell recordings from S1-RGCs held at potentials to isolate excitatory (~ -60mV) or inhibitory (~0mV) currents to an expanding flashing spot in Sdk1 CG/+ (Het) or Sdk1 CG/CG retinae (KO). C-D. Average peak current versus expanding bright (C) or dark (D) spot diameter measured from control or knockout S1-RGCs held at -60mV in experiments like those shown in A-B. E-F. Average peak current versus expanding bright (E) or dark (F) spot diameter measured from control or knockout S1-RGCs held at -0mV taken experiments like those shown in A-B. (n=8 for Sdk1-Het, n=9 for Sdk1-KO) G. Retinal cross-sections showing S1-RGCs in control (Het) Sdk1 knockout (KO) retinae. Scale = 50m. H. Linescans through S1-RGC arbors in control in Sdk1 null retinae taken from experiments like those shown in G. I. Skeletonized S1-RGC dendrites from control (black) and Sdk1 null (red) retinae. (Scale = 50μm). J-L. Average branch number (J), branch length (K), and dendritic area (L) measured from control and Sdk1 null S1-RGC dendritic arbors. M. Sholl analysis of dendritic arbors measured from Het and KO S1-RGCs. (n = 8 for both Sdk1-Het and Sdk1-KO; * = p < 0.05; ** = p < 0.01) comparing the laminar position of S1-RGC dendrites in Sdk1 CG/+ or Sdk1 CG/CG retinal cross-sections showed no obvious 340 difference between controls and nulls ( Figure 5G-H). We next asked if Sdk1 loss impacts the lateral anatomy of S1-RGC 341 dendrites. We found that the loss of Sdk1 led S1-RGCs to grow dendritic arbors that were less complex than their control 342 counterparts ( Figure 5I). Sdk1-null S1-RGC arbors contained similar numbers of dendritic branches ( Figure 5J), but they 343 were approximately half as long on average, which led to fewer intersections across the entire dendritic arbor as assessed 344 by Sholl analysis (Figure 5K-M). These deficits arose with no major change in the overall structure of the IPL, as assessed 345 by staining with a variety of markers that label AC and BC subsets targeting specific sublayers ( Figure 5 -figure   346 supplement 3). Moreover, reconstructed ON-RGCs in the same Sdk1 CG/CG retina showed dendritic arbors that were 347 similar to their counterparts in control retinae ( Figure 5 -figure supplement 1C-H), indicating that Sdk1-loss selectively 348 impairs the dendritic arborization of S1-RGCs.

351
Here, we investigated the role of Sdk1 in retinal circuit development. In the first section of this study, we molecularly, 352 anatomically, and functionally defined 5 Sdk1 interneurons (ACs and BCs) and 5 Sdk1-RGCs that target an inner set of 353 IPL lamina. This family of RGCs include two ON-DSGC types and ON-OFF S1-RGCs; the latter shows selectivity for 354 edges located in their receptive field. In the second section, we found that Sdk1 loss caused a significant loss of S1-RGC 355 responsivity with little effect on the other Sdk1-RGC types. By comparing visual responses between S1-RGCs and On-

356
RGCs in control and Sdk1 nulls, we show that the loss of Sdk1 impairs S1-RGCs' responses to both bright and dark spot 357 stimuli, as well as oriented bars. Finally, we show that these deficits arise from a selective loss of excitatory and inhibitory 358 synaptic input on S1-RGCs and correlates with a selective loss of small branches in this neuron's dendritic arbor. We 359 conclude that Sdk1 is required for the dendritic and synaptic development of a local edge-detecting RGC type.

363
Transcriptomic studies indicate that RGCs can be divided into at least 46 different clusters, which include several clusters 364 that do not correspond to any known RGC and are predicted to be novel. In the process of studying Sdk1, we developed 365 molecular signatures for three of these orphan clusters and characterized them using calcium imaging.  376 377 S1-RGCs are a high-density, narrow dendritic field neuron, characterized by small (<200m) diameter dendritic arbors, 378 strong surround suppression, and responses to both bright and dark stimuli. Similar properties have been found in several 379 other RGCs that respond to stimuli falling in their receptive field center but are silenced when the same stimuli fall in their also show sensitivity to edge orientation, which could arise from their orthogonally tuned excitatory and inhibitory inputs.

386
This input arrangement has been observed in recently described horizontally and vertically selective OS-RGCs; however, 387 these neurons show significantly stronger orientation selectivity compared to S1-RGCs. We cautiously speculate that S1-

401
Our work here adds to these findings.

403
We previously showed that an IgSF called Sdk2 enriches connections between VG3-ACs and W3B-RGCs, permitting this

464
Mice were given at least two weeks to recover before experimental use.  508 pixels with a step sizes ranging from 0.45μm to 8μm. Following image acquisition, minor processing of images including 509 landmark correspondence transforms to match two-photon and stained retinal fields as well as stitching of large tilescans 510 was performed using ImageJ.
For density recovery profile analysis (DRP), RGC soma centers were selected manually on ImageJ and their x-y coordinates used to generate DRPs using code modelled after (Rodieck, 1991). The resulting DRPs were averaged 514 across a given condition and fitted with a sigmoid curve for visual clarity. The mean cell density obtained from this 515 analysis was used to assess the proportion of cells labelled by Sdk1 among BCs, ACs and RGCs.

518
RGCs and Onα RGCs taken from P30-60 retinae were traced manually through the z-stack using simple neurite tracer 519 (ImageJ). Path ROIs describing neuronal processes were converted to stacks and analyzed for morphological properties 520 using the Trees toolbox on MatLab (Cuntz et al., 2010).

522
To obtain mean IPL projection depth of each Sdk1+ type, we took linescans of pixel intensity across the IPL from images 523 stained with antibodies to VAChT (sublamina 2 and 4) or reporter, normalized these signals to the maximum intensity, and 524 averaged these traces across each condition. IPL depth is expressed as a percent and sublaminae judged by the position 525 of the peak intensities in the VAChT channel. We applied the straighten transform (ImageJ) to correct the VAChT bands 526 on a few curved sections and applied the same transform to reporter channels prior to linescan procedure. were detected in loose patch recordings using the peakfinder function and binned (50ms) over the entire length of the trial; firing rate histograms for each trial were then averaged and subjected to further processing based on each stimulus. Direction and orientation selective indices were computed from mean firing rate histograms as described for calcium 588 imaging data. For whole-cell currents, trials were averaged, peak amplitude measured, and integral were computed using