Local processing of visual information in neurites of VGluT3-expressing amacrine cells

Synaptic inputs to neurons are distributed across extensive neurite arborizations. To what extent arbors process inputs locally or integrate them globally is, for most neurons, unknown. This question is particularly relevant for amacrine cells, a diverse class of retinal interneurons, which receive input and provide output through the same neurites. Here, we used two-photon Ca2+ imaging to analyze visual processing in VGluT3-expressing amacrine cells (VG3-ACs), an important component of object motion sensitive circuits in the retina. VG3-AC neurites differed in their preferred stimulus contrast (ON vs. OFF); and ON and OFF responses varied in transience and preferred stimulus size. Contrast preference changed predictably with the laminar position of neurites in the inner plexiform layer. Yet, neurites at all depths were strongly activated by local but not by global image motion. Thus, VG3-AC neurites process visual information locally, exhibiting diverse responses to contrast steps, but uniform object motion selectivity.


Introduction 24
Neurons receive most of their synaptic input on large intricately branched dendritic arborizations. 25 Traditionally, distributed inputs were thought to be summed linearly at the cell body (Yuste, 2011). 26 However, recent studies uncovered extensive local processing and clustered plasticity of synaptic inputs 27 that enhance the computational power of dendrites (Grienberger et al., 2015, Harvey and Svoboda, 2007, 28 Kleindienst et al., 2011, London and Hausser, 2005, Losonczy et al., 2008. Although less studied, similar 29 local processing appears to occur in terminal axon arbors, in which presynaptic inhibition and 30 inhomogeneous distributions of voltage-gated ion channels can diversify the output of a single neuron 31 (Debanne, 2004, Asari andMeister, 2012). 32 Amacrine cells (ACs) are a diverse class of interneurons in the retina (Helmstaedter et al., 2013, 33 MacNeil andMasland, 1998). Most of the approximately 50 distinct AC types lack separate dendrites and 34 axons, and, instead, receive input and provide output through the same neurites. The radially symmetric 35 arbors of starburst ACs receive synaptic input and release neurotransmitters near and far from the soma, 36 respectively (Ding et al., 2016, Vlasits et al., 2016. In a seminal study, Euler et al. (2002) discovered by 37 two-photon Ca 2+ imaging that the four to six primary neurites of starburst ACs with their daughter branches 38 function as independent centrifugal motion sensors. Another AC type (A17) was recently shown to process 39 converging inputs from rod bipolar cells separately (Grimes et al., 2010). For most AC types, however, 40 whether neurites process inputs locally or integrate them globally, what stimulus features are encoded by 41 neurites, and whether responses across neurite arbors are uniform or varied remains unknown. 42 VGluT3-expressing amacrine cells (VG3-ACs) stratify neurites broadly in the center of the inner 43 plexiform layer (IPL). In somatic patch clamp recordings, VG3-ACs depolarize to light increments (ON) 44 and decrements (OFF) constrained to a small area (i.e. receptive field center), but hyperpolarize to large 45 ON and OFF stimuli that include the receptive field surround (Kim et al., 2015, Lee et al., 2014. VG3-ACs 46 are dual transmitter neurons, which deploy their two transmitters in a target-specific manner. They provide 47 glutamatergic input to W3 retinal ganglion cells (W3-RGCs) driving object motion sensitive responses 48 (Kim et al., 2015, Krishnaswamy et al., 2015, Lee et al., 2014, and glycinergic input to  Contrast RGCs (SbC-RGCs) inhibiting responses to small OFF stimuli (Lee et al., 2016, Tien et al., 2016, 50 Tien et al., 2015. The target-specific use of excitatory and inhibitory transmitters (Tien et al., 2016, Lee et 51 al., 2016, and the observation that removal of ON-OFF responsive VG3-ACs selectively affects OFF 52 inhibition in SbC-RGCs (Tien et al., 2016) raise the question whether VG3-AC neurites process visual 53 information locally or integrate it globally. 54 Here, we used two-photon Ca 2+ imaging in a novel transgenic mouse line to analyze how stimulus 55 contrast, size, and motion are processed in VG3-AC neurites. 56

Results and discussion 57
We crossed VG3-Cre mice to a novel transgenic strain (Ai148) expressing the genetically encoded Ca 2+ 58 indicator GCaMP6f in a Cre-dependent manner enhanced by tTA-based transcriptional amplification. 59 Staining for VGluT3 confirmed that GCaMP6f labeling in the inner nuclear layer (INL) and the IPL of 60 VG3-Cre Ai148 retinas was mostly restricted to VG3-ACs ( Figure 1A-D) (Grimes et al., 2011, Kim et al., 61 2015. We imaged GCaMP6f signals in 23 x 23 µm scan fields in the IPL of flat-mounted retinas. 62 Recording depths of scan fields were registered by their relative distance to the outer and inner boundaries 63 of the IPL (0-100 %) detected by imaging transmitted laser light (Figure 1 -figure supplement 1). Visual 64 stimulation (385 nm) was spectrally separated from GCaMP6f imaging (excitation: 940 nm, peak emission: 65 515 nm); and recordings were obtained from the ventral retina, where S-opsin dominates (Haverkamp et 66 al., 2005, Wang et al., 2011. Ca 2+ responses were no different for scanning at 9.5 Hz vs 37.9 Hz (Figure 1 67 -figure supplement 2). To maximize spatial resolution (pixel size: 0.29 x 0.36 µm), we therefore acquired 68 all data in this study at 9.5 Hz. To identify processing domains in VG3-ACs neurites objectively, we 69 automated detection of regions of interest (ROIs) in each scan field based using a serial clustering procedure 70 ( Figure 1E-H, s. Materials and methods). 71 In somatic patch clamp recordings, VG3-ACs depolarize to small ON and OFF stimuli (Lee et al., 72 2014, Kim et al., 2015, Grimes et al., 2011. Somatic Ca 2+ transients exhibited similar ON-OFF profiles to 73 those observed in voltage recordings (Figure 2A,B). To test, whether these responses are uniformly 74 distributed across VG3-AC neurites or not, we recorded Ca 2+ transients elicited by contrast steps in a small 75 spot (radius: 50 µm) at different depths of the IPL (Figure 2A,B and Video 1). We quantified contrast 76 preference by a polarity index, ranging from -1 for pure OFF responses to 1 for pure ON responses (s. 77 Materials and methods). Polarity indices varied widely between ROIs (n = 1120) ( Figure 2C) and were 78 correlated with IPL depth (R 2 = 0.129, p < 10 -67 ), as neurites in the outer IPL (depths < 40 %) responded 79 more strongly to OFF stimuli, and neurites in the inner IPL (depths > 40 %) responded more strongly to 80 ON stimuli ( Figure 2D). Both ON and OFF responses were transient ( Figure 2E); and in the inner IPL 81 (depths > 40 %), ON responses were more transient than OFF responses ( Figure 2F). Thus, VG3-AC 82 neurites process visual information locally, not globally, and exhibit diverse contrast preferences and 83 response kinetics in different layers of the IPL. 84 A hallmark of VG3-ACs' somatic voltage responses is strong size selectivity. We explored how 85 VG3-AC neurites at different depths in the IPL respond to contrast steps in spots of different sizes. At all 86 depths, only small stimuli (radius: < 200 µm) elicited Ca 2+ transients ( Figure 3A and Video 1). OFF 87 responses preferred smaller stimuli than ON responses across most of the IPL ( Figure 3B,C). Interestingly, 88 variations in response center size between ROIs were correlated for ON and OFF responses (Figure 3 -89 figure supplement 1). Thus, stimulus size, like contrast, is encoded locally in VG3-AC neurites; and some 90 mechanisms that shape size selectivity appear to be shared between ON and OFF responses of a given VG3-91 AC neurite. 92 VG3-ACs participate in object motion sensitive circuits in the retina, amplifying selectively the 93 responses of W3-RGCs to local image motion (Krishnaswamy et al., 2015, Kim et al., 2015. We tested the 94 ability of individual VG3-AC neurites to distinguish local and global image motion, using a stimulus in 95 which square wave gratings overlaying center and surround regions of neurite receptive fields moved 96 separately or together (Kim et al., 2015, Olveczky et al., 2003, Zhang et al., 2012. Isolated motion of the 97 center grating elicited robust Ca 2+ transients in VG3-AC neurites at all depths, which remained silent during 98 simultaneous motion of gratings in center and surround (i.e. global motion) ( Figure 4A and Video 2). As a 99 result, local motion preference indices (s. Materials and methods) of > 83% of ROIs were > 0.9 ( Figure  100 4B,C). Thus, contrary to the diversity of responses to contrast steps, VG3-AC neurites exhibit uniform 101 object motion selectivity. 102 Our results show that VG3-ACs process visual information locally and exhibit diverse preferences 103 for stimulus contrast and size across their neurite arbors. Changes in contrast preference with increasing 104 IPL depth likely reflect a shift in input from OFF to ON bipolar cells and align with the stratification patterns 105 of bipolar cell axons (Franke et al., 2017, Helmstaedter et al., 2013. The consistent differences in receptive 106 field size of OFF and ON responses ( Figure 3B) indicate separate origins (e.g. in OFF and ON bipolar cells, 107 respectively), whereas correlated variations in ON and OFF receptive field sizes between neurites ( Figure  108 3 -figure supplement 1) hint at a shared underlying mechanisms (e.g. varying strength of surround 109 inhibition to VG3-AC neurites). In spite of their diverse responses to contrast steps, VG3-AC neurites show 110 uniform object motion selectivity, which relies on rectified excitatory input from bipolar cells and strong 111 surround inhibition from ACs (Kim et al., 2015). Local processing suggests that the output of VG3-ACs at 112 different release site may convey different visual information. Whether release sites with different Ca 2+ 113 response profiles connect to different ganglion cell targets and/or use different neurotransmitters (glutamate 114 vs. glycine) is an interesting question for future studies.

Immunohistochemistry 138
Flat-mounted retinas were fixed for 30 min in 4% paraformaldehyde in mACSFHEPES at room temperature 139 (RT) and washed three times for 10 min in PBS at RT. The fixed tissue was cryoprotected with incubations 140 in 10%, 20%, and 30% sucrose in PBS for 1 hr at RT, 1 hr at RT, and overnight at 4°C, respectively, 141 followed by three cycles of freezing (held over liquid nitrogen) and thawing (in 30% sucrose in PBS). 142 Retinas were then washed three times in PBS for 1 hr at RT, and stained for VGluT3 (rabbit anti-VGluT3, 143 Cat. No. 1352503, Synaptic Systems) and GFP (chicken anti-GFP, 1:1000, Cat. No. A10262, 144 ThermoFisher) for three to five days at 4°C in PBS with 5% normal donkey serum and 0.5% Triton X-100. 145 Subsequently, retinas were washed three times for 1 hr in PBS, stained with Alexa 488-Alexa 568-146 conjugated secondary antibodies (Invitrogen, 1:1000) overnight at 4 °C, washed three times in PBS for 1 hr, 147 and mounted in Vectashield mounting medium (Vector Laboratories) for confocal imaging. 148

Confocal imaging 149
Confocal image stacks of fixed tissue were acquired through 20 X 0.85 NA or 60 X 1.35 NA oil immersion 150 objectives (Olympus) on an upright microscope (FV1000, Olympus). Confocal images were processed and 151 analyzed with Fiji (Schindelin et al., 2012). 152

Visual stimulation 153
Visual stimuli were written in MATLAB (The Mathworks) using the Cogent Graphics toolbox (John 154 Romaya, Laboratory of Neurobiology at the Wellcome Department of Imaging Neuroscience, University 155 College London). Stimuli were presented from a UV E4500 MKII PLUS II projector illuminated by a 156 385 nm LED (EKB Technologies) and focused onto the photoreceptors of the ventral retina via a substage 157 condenser of an upright two-photon microscope (Scientifica). All stimuli were centered on the two-photon 158 scan field and their average intensity was kept constant at ~ 1,600 S-opsin isomerizations / S-cone /s. To 159 measure the contrast and size preference, the intensity of spots of varying radii (10,25,50,100,200,300,160 and 400 µm) was square-wave-modulated (1.5 s ON, 1.5 s OFF) for 5 cycles. The order in which spots of 161 different size were presented was randomly chosen for each scan field. To test responses to local vs. global 162 motion stimuli, narrow square wave gratings (bar width: 50 µm) over the receptive field center (radius: 163 75 µm) and surround (150-800 µm from center of the image) were moved separately or in unison (Kim et 164 al., 2015, Zhang et al., 2012. A gray annulus was included in the spatial layout of the stimulus to reliably 165 separate movement in the center and surround. Each grating motion lasted 0.5 s, and movements were 166 separated by 1.5 s. 167

Two-photon imaging 168
A custom-built upright two-photon microscope (Scientifica) controlled by the Scanimage r3.8 MATLAB 169 toolbox was used in this study; and images were acquired via a DAQ NI PCI6110 data acquisition board 170 (National Instruments). GCaMP6f was excited with a Mai-Tai laser (Spectra-Physics) tuned to 940 nm, and 171 fluorescence emission was collected via a 60 X 1.0 NA water immersion objective (Olympus) filtered 172 through consecutive 450 nm long-pass (Thorlabs) and 513-528 nm band-pass filters (Chroma). This 173 blocked visual stimulus light (peak: 385 nm) from reaching the PMT. Because we observed no qualitative 174 differences between Ca 2+ responses scanned at 9.5 Hz and 37.9 Hz (Figure 1 -figure supplement 2), we 175 acquired 23 x 23 µm images (pixel size: 0.29 x 0.36 µm) throughout this study at 9.5 Hz. Imaging depths 176 were registered by their relative distances to the borders between the IPL and the inner nuclear layer (0% 177 IPL depth) and between the IPL and the ganglion cell layer (100% IPL depth). Borders were detected in 178 transmitted light images (Figure 1 -figure supplement 1). Scan fields at different IPL depths were imaged 179 in pseudorandom order; and for each scan the retina was allowed to adapt to the laser light for 30 s before 180 presentation of visual stimuli. All images were acquired from the ventral retina, where S-opsin dominates 181 (Wang et al., 2011). Throughout the experiments retina were perfused at ~7 mL / min with 34°C 182 mACSFNaHCO3 equilibrated with 95% O2 / 5% CO2. 183

Image processing 184
Registration. Transmitted light images were acquired simultaneously with fluorescence images and were 185 used to detect z-axis displacements that resulted in rejection of the respective image series. Images of series 186 without z-axis displacements were registered to the middle frame using built-in functions in MATLAB. 187 Rigid transformations were applied to both transmitted and fluorescence images. The quality of registration 188 was confirmed by visual inspection, before transformed fluorescence images were used for further image 189 processing and analysis. 190 Denoising. Time series of each pixel were searched for outliers (> 10 SD). If outliers were isolated in time 191 (i.e. pixel value before and after outlier < 10 SD), they were replaced with the average of the value before 192 and after the outlier. This algorithm effectively removed PMT shot noise. 193 Segmentation. To identify functional processing domains in VG3-AC neurites with minimal assumptions 194 and user involvement, we developed a serial clustering procedure, in which a functional clustering 195 algorithm is successively applied to different image features. This procedure removed pixels of the image 196 not responding to visual stimulation and automatically assigned responsive pixels to functionally coherent, 197 spatially contiguous regions of interest (ROIs). The functional clustering algorithm was based on Shekhar 198 et al. (2016), beginning with principal components analysis to reduce the dimensionality of the input feature 199 to the minimum needed to explain 80% of its variance. This was followed by a K-nearest-neighbor (KNN) 200 algorithm, which generated a connectivity matrix. The connectivity matrix was then used in community 201 detection clustering (Le Martelot and Hankin, 2012). We first applied functional clustering to the raw data 202 of an image series and removed low-intensity pixels. Signals of remaining pixels were normalized to their 203 peak and fed back into the functional clustering algorithm to group pixels with similar response properties. 204 Groups of functionally similar pixels were divided into spatially contiguous ROIs within the image. The 205 average response traces of these ROIs were subjected to further rounds of functional clustering, in which 206 spatially adjacent ROIs that were grouped in the same cluster were merged. This process was repeated until 207 it converged on a stable solution (typically less than 15 iterations). Finally, ROIs identified in this procedure 208 were examined for signal correlation with the visual stimulus and size, to reject non-responsive and/or small 209 (< 5 pixels) ROIs. 210

Image analysis 211
Polarity index. Responses of each ROI to contrast steps in small spots (radius: 50 µm) were divided into 212 ON and OFF periods (1.5 s each). The median peak response to 5 stimulus repeats during each periods was 213 then used to calculate a polarity index as follows: 214 is the time to peak, measured from stimulus onset, and is a 221 delay set to the fourth frame (~420 ms) after the peak frame. Transience was only calculated for ON and 222 OFF responses that exceeded 25 % of the peak amplitude of the respective ROI. The maximum value of 223 the transience index is 1, indicating that the GCaMP6f signal returned to baseline at time after the peak. 224 Response center size. Because difference of Gaussians fits were poorly constrained by our data, we 225 quantified stimulus size preference using a single nonparametric measure, the response center size. ON and 226 OFF responses of each ROI were analyzed separately and their amplitudes plotted as a function of stimulus 227 size. If responses to any size exceeded 25 % of the peak amplitude of the respective ROI, response center 228 size was calculated as the center of mass of the stimulus size -response relationship. 229 Local motion preference index. Median responses of each ROI to isolated grating motion in the receptive 230 field center (i.e. local motion) and to synchronous grating motion in receptive field center and surround (i.e. 231 global motion) were used to calculate a local motion preference index as follows: 232 A local motion preference index of 1 indicates that the respective ROI responded only to local and not to 234 global motion. 235

Statistics 236
Functional imaging data were obtained from retinas of five mice. All summary data and response traces are 237 presented as mean ± SEM. Differences between groups were statistically examined by two-way analysis of 238 variance (for transience and response center size) with post hoc comparisons applying Student's t-tests with 239 Turkey's HSD (honest significant difference) multiple comparison correction or by Kruskal-Wallis test (for 240 polarity and local motion preference). 241

Author contributions 242
The study was conceived and designed by J. one-way ANOVA). ROIs at 21% IPL depth were more polarized toward OFF than at other depths (p < 0.01 344 compared to 28% and 34%; p < 10 -7 for 42% -54%). ROIs from 42% -54% IPL depth were more polarized 345 toward ON than ROIs from 21% -34% (p < 10 -5 ).  ON (open circle) and OFF (filled circle) responses as a function of IPL depth. Preferred stimulus sizes 361 differed between IPL depths (p < 10 -10 , main effect of contrast, two-way ANOVA) and between ON and 362 OFF responses overall (p < 10 -10 , main effect of contrast, two-way ANOVA). Moreover, preferred stimulus 363 sizes differed between ON and OFF responses of ROIs at all IPL depths (p < 10 -6 from 21% -34%, 364 p < 0.001 at 54%) except for 42% (p = 0.21) and 47% (p = 0.59). The interaction between depth and contrast 365 was also significant (p < 10 -8 ). 366 data (mean ± SEM) of local motion preference indices as a function of IPL depth. Local motion preference 374 indices did not differ across IPL depths (p = 0.4427, Kruskal-Wallis one-way ANOVA). No ROI group at 375 any depth was significantly different from any ROI group at another depth. 376 Video 1. Ca 2+ imaging of VG3-AC neurite responses to contrast steps in spots of varying size recorded 377 at different IPL depths. Image series of GCaMP6f responses at 24% (middle) and 53% (right) IPL depth 378 to contrast steps in spots of different size (left). The video is sped up 2.5-fold relative to the image 379 acquisition. In the left panel, the area of the scan fields is indicated by a red box. Two average normalized 380 ROI traces are shown at the bottom of the middle and the right panel. 381 Video 2. Ca 2+ imaging of VG3-AC neurite responses to motion stimuli recorded at different IPL 382 depths. Image series of GCaMP6f responses at 24% (middle) and 49% (right) IPL depth to synchronous or 383 isolated motion of square wave gratings in the center and surround separated by a gray annulus (left). The 384 video is sped up 1.25-fold relative to the image acquisition. In the left panel, the area of the scan fields is 385 indicated by a red box. Two average normalized ROI traces are shown at the bottom of the middle and the 386 right panel. 387 the mean (± SEM) responses of VG3-AC neurites scanned at 9.5 Hz; and green traces (shaded areas) show 400 the mean (± SEM) responses of VG3-AC neurites scanned at 37.9 Hz. (40% at 9.5 Hz: n = 126; 40% at37.9 401 Hz: n = 67; 54% at 9.5 Hz: n = 62; 54% at 37.9 Hz: n = 28.) (B) Summary data (mean ± SEM) of polarity 402 indices as a function of IPL depth for responses scanned at 9.5 Hz (black) and 37.9 Hz (green). The polarity 403 of ROIs was different between IPL depths (p < 10 -10 , main effect of depth, two-way ANOVA), but did not 404 differ significantly between scan rates (p = 0.62, main effect of frequency). The interaction between 405 frequency and depth was not significant (p = 0.50). 406