An ON-type direction-selective ganglion cell in primate retina – Nature

Tissue preparation

Macaque (M. mulatta) eyes from animals of either sex (6 males, 5 females, age range 1.17–19.17 years) were obtained immediately post-mortem from the Oregon and California National Primate Research Center biospecimen distribution programs or from UC Berkeley from animals euthanized for unrelated studies. Eyes from UC Berkeley were enucleated under terminal anaesthesia in accordance with procedures approved by the Animal Care and Use Committee of the University of California, Berkeley and as specified in the National Research Council guidelines. The anterior segment was promptly removed before transferring posterior eyecups to Ames’ medium equilibrated with 95% oxygen/5% carbon dioxide. Eyecups were stored in the dark at room temperature (20–22 °C) until use (up to 48 h post-mortem). All locations in nasal retina are specified as equivalent eccentricities using the formula \(\sqrt{{(0.61x)}^{2}+{y}^{2}}\) where x is the distance in mm nasal to the foveal centre and y is the distance in mm superior or inferior to the foveal centre.

Two human retinal samples (male, age 51 and 72 years) were obtained from eyes exenterated for management of orbital tumours at Oregon Health & Science University. No retinal pathology was noted in either case. All samples were de-identified before receipt by the investigators and tissue use was thus deemed non-human subject research by the institutional review board at Oregon Health & Science University.

For calcium imaging experiments in mouse retina, we used either C57BL/6 J (JAX strain: 000664; RRID: IMSR_JAX: 000664) or crossbred Ai95(RCL-GCaMP6f)-D (JAX strain: 028865; RRID: IMSR_JAX:028865) homozygote mice with homozygous Vglut2-ires-Cre knock-in mice (JAX strain: 028863; RRID: IMSR_JAX: 028863), to obtain offspring with GCaMP6f expression in RGCs. Mice were housed under a standard 12:12 h day:night cycle. Ambient temperature and humidity were maintained at 20–22 °C and 50–60%. Mice were of either sex and were 6–26 weeks old. Mice were dark adapted for 1.5 h prior to euthanasia. Animals were anaesthetized with isoflurane and euthanized with cervical dislocation per procedures approved by the Animal Care and Use Committee at University of California, Berkeley. Eyes were then immediately enucleated, the anterior eye removed and the retina isolated for subsequent experiments.

Immunohistochemistry

The primary and secondary antibodies that were used in this study are detailed in Extended Data Table 1. The specificity of the BNC2 antibody was confirmed using antibodies raised against different epitopes of the same protein and by confirming selective expression of BNC2 in VGLUT3⁺ amacrine cells in the INL (a pattern predicted by the transcriptome) (Extended Data Fig. 7). Pieces of macaque or human retina were fixed for 30–120 min in 4% paraformaldehyde (PFA) in 0.1 M phosphate buffer at ~22 °C, rinsed in PBS and stored in PBS containing 0.025% NaN₃ (PBS-NaN₃) at 4 °C or cryoprotected in graded sucrose solutions and frozen at −20 °C until use. After washes in PBS, retinas were blocked for 1 h in 1% bovine serum albumin (BSA; Sigma A7030), 150 mM maleimide (Sigma 129585) in PBS. Primary antibodies were diluted in 3% normal horse serum (NHS), 1% Triton X-100, 0.025% NaN₃ in PBS (pH 7.4) and incubated for ~3 days at ~22 °C. Secondary antibodies were diluted in 3% NHS, 0.025% NaN₃ in PBS, and applied overnight at 22 °C. Retinas were counterstained for 20 min in Hoechst 33342 (Invitrogen 33342), then mounted in SlowFade Gold antifade reagent (Invitrogen S36936).

Fluorescence in situ hybridization

Fluorescence in situ hybridization was performed on horizontal cryosections of the macaque GCL using the RNAscope Multiplex Fluorescent v2 Assay combined with Immunofluorescence—Integrated Co-Detection Workflow (Advanced Cell Diagnostics) according to the manufacturer’s instructions. In brief, a macaque retina was fixed with 4% PFA for 24 h at 4 °C, cryoprotected with 10%, 20%, and 30% sucrose and embedded in Cryo-Gel (Leica). A region of peripheral retina was cryo-sectioned through the plane of the GCL at 10 µm thickness. Sections were retrieved for 5 min at ~100 °C in Co-detection Target Retrieval solution and then incubated overnight with rabbit anti-RBPMS antibody (Phosphosolutions 1830-RBPMS) to label RGCs. Sections were then digested with Protease III for 30 min at 40 °C and hybridized with probes for macaque BNC2 (ACD 1238401-C1) and FSTL4 (ACD 1238411-C2) mRNA. Signals were amplified and fluorescently tagged with 1:1500 TSA Vivid fluorophore 520 (ACD 323271) and 1:1500 TSA Vivid fluorophore 570 (ACD 323272), respectively. Finally, sections were incubated in goat anti-rabbit Alexa Fluor Plus 647 (1:200; ThermoFisher A32733) overnight, counterstained with DAPI and mounted in Slowfade Gold.

Pre-immunolabelling for DiI filling

Pieces of macaque retina were fixed for 30 min in 4% PFA, rinsed in PBS and stored in PBS-NaN₃ at 4 °C until use. For pre-labelling, retinas were incubated on a shaker in primary antibodies for BNC2 and ChAT diluted in incubation buffer containing 3% NHS in PBS-NaN3 for 3 days at 22 °C. After washes in PBS, secondary antibodies were applied in incubation buffer overnight. Some retinas were permeabilized in incubation buffer containing 0.3% Tween 20 for 1 h at 22 °C (3 of 7 cells) before primary antibody incubation. In some cases, secondary detection of ChAT antibodies was performed after dye loading. Pre-labelled retinas were mounted ganglion cell side up on an Olympus BX51WI microscope in PBS-NaN₃. BNC2⁺ RGCs were located with epifluorescence, then impaled with sharp borosilicate microelectrodes filled with 1% DiI in 100% ethanol under infrared illumination with Dodt or DIC contrast^51,52. Current pulses (10 s) of ~+10 nA were applied for 3–5 min using a patch-clamp amplifier (HEKA EPC10) for dye injection. Since pre-labelling was with BNC2 antibodies only, we could not unambiguously distinguish pRGC10 and pRGC16 cells. However, the majority of BNC2⁺ cells are pRGC10 (Fig. 1e) and the most strongly labelled BNC2⁺ cells are within the pRGC10 cluster (Fig. 1d). On occasion, bistratified cells were recovered that could represent pRGC16 (Extended Data Fig. 8). After cell injections, retinas were post-fixed in PFA overnight at 4 °C, counterstained with Hoescht, mounted in Vectashield (VWR 101098-042) and imaged within one week.

Fluorescence imaging and image processing

For reconstruction of neuronal morphology, confocal laser scanning and Fast Airyscan images were acquired on a Zeiss LSM 880 microscope with Plan-Apochromat 20×/1.0 DIC water, 20×/0.8 air, and 63×/1.4 oil DIC objectives. Images were acquired at 1.93–2.41 pixels per µm (20×) and 7.6–11.7 pixels per µm (63×) resolution, with optimal axial resolution to permit 3D reconstruction. Z-stacks extended from the GCL to the INL, and z-step size ranged from 0.65 to 0.86 µm for the 20× objective and 0.16–0.19 µm for the 63× objective. Excitation laser lines were 561 nm for DiI, 488 nm for Alexa-488 (ChAT), 405 nm for Hoescht and 594 nm for Alexa-594 (BNC2). DiI and ChAT were imaged in separate tracks to prevent signal cross-talk.

For confocal imaging of IMHC after calcium imaging or for mosaic analysis, tile scans were acquired with a 20×/1.0 N.A. water or 20×/0.8 N.A. air objective on a Zeiss LSM 880. Tiles of 425.1 × 425.1 µm (2.41 pixels per µm) or 472.33 × 472.33 µm (1.63 pixels per µm) were acquired with 10% overlap and stitched with Zen 2 software (Zeiss) or the Grid/Collections plugins in ImageJ. All modifications to the original images are indicated in the figure legends. A lateral median filter was applied to some images where indicated to remove non-specific signal in the BNC2 channel. In such cases, the filter was applied uniformly to the entire image. Other linear changes to brightness and contrast were applied uniformly to images in ImageJ.

For analysis of the macaque retina hemifield, images were acquired on a Zeiss Axioplan 2 epifluorescence microscope with a Plan Apochromat 10×/0.45 N.A. air objective and a 100 W Hg arc lamp excitation source. Tile scans were acquired using the MosaiX module in Axiovision software and stitched using the Grid/Collection stitching plugin for ImageJ.

Two-photon calcium imaging

For calcium imaging experiments, pieces of macaque retina (~4–8 mm²) with choroid and retinal pigment epithelium attached were isolated from the sclera under infrared illumination (850 nm). Given the higher density of pRGC10 on the horizontal midline of the retina (Extended Data Fig. 3), most recording fields were from retinal regions extending from the nasal margin of the optic nerve head to the far nasal periphery. Samples were mounted on an inorganic membrane disc (Anodisc, 13 mm, pore size 0.2 μm, GE Whatman) and stabilized with a slice anchor (Warner Instruments). Retinas were oriented in the recording chamber such that the naso-temporal axis was approximately aligned with the horizontal visual stimulation plane. The preparation was continually superfused with warm (36–37 °C) bicarbonate-buffered Ames’ medium at 5-6 ml min⁻¹. Samples were imaged with either an Olympus 20×/0.95 N.A. or a Nikon 16×/0.8 N.A. water dipping objective under infrared illumination (850 nm) with oblique contrast optics. Retinas were bolus loaded with the membrane-permeable calcium indicator, Cal-520 AM (dissociation constant (K_d) = 320 nM, AAT Bioquest) or Cal-590 AM (K_d = 561 nM, AAT Bioquest). Indicators were dissolved in DMSO containing 20% Pluronic F-127 (ThermoFisher P3000MP), then diluted to a final concentration of 0.91 mM in HEPES-buffered or bicarbonate-buffered Ames’ medium (pH = 7.4). For dye loading, a 2–5 MΩ borosilicate pipette was filled with this solution and applied by pressure application (~3 psi for ~10 × 5 s) to the GCL after penetrating the inner limiting membrane. Recordings commenced at least 45 min after dye loading. In some cases, sulforhodamine 101 (0.8 µM; Sigma S7635) was added to the Ames’ medium and perfused for 15 min before commencing recordings to visualize the vasculature.

Mouse retinas were dissected off the retinal pigment epithelium under infrared illumination and imaged similarly to macaque. Areas of dorsal hemiretina within 0.7–1 mm of the optic nerve were used to minimize topographical variance in DS tuning²⁹. Vglut2-GCaMP6f mice (n = 2) were imaged immediately whereas a wildtype mouse retina was loaded with Cal-590 AM as described above and recording commenced ~1 h after bolus loading.

Two-photon calcium imaging was performed with a modified Scientifica MP-2000 multiphoton microscope and a modified HyperScope multiphoton microscope (Scientifica) fitted with a mode-locked Ti-Sapphire laser tuned to 930 or 1040 nm (Chameleon Ultra II; Coherent). The laser intensity was adjusted so that the minimal power was used to visualize loaded cells (range 9–21 mW for 1,040 nm, 6–25 mW for 930 nm). An open-source CAD model, part list, and filter spectra for both microscope systems can be found at https://github.com/Llamero/Puthussery_Lab_2P_Setup.

MP-2000 setup

Two-photon emission light was split by a dichroic mirror (FF552-Di02, Semrock) to red (FF01-590/36, Semrock) and green (FF01-510/42, Semrock) detection channels. The scan field was 256 × 256 pixels or 128 × 128 pixels (220 × 220 µm) acquired at a frame rate of 2.67 Hz or 10.67 Hz. Images were acquired using Sciscan acquisition software (v1.3, Scientifica).

Visual stimuli were generated with a gamma-corrected LG LP097QX1 TFT-LCD retina display (2,048 × 1,536) (Adafruit Industries) that was projected onto the retina through the 20× objective. The RGB output of the display was triple band-pass filtered (FF01-466/555/687-25, Semrock) to separate the RGB visual stimulation bands from the photomultiplier detection channels.

HyperScope setup

Two-photon emission light was split by a dichroic mirror (FF552-Di02, Semrock) to red (FF01-590/36, Semrock) and green (FF01-515/30, Semrock) detection channels. The scan field was 512 × 512 pixels (246.71 × 246.71 µm) acquired at a frame rate of 30 Hz. Images were acquired using ScanImage Basic v2021.01.0 (MBF Biosciences).

Visual stimuli were generated with a fibre-coupled DLP LightCrafter 4500 (EKB Technologies) that was projected onto the retina through the 16x objective. The RGB illumination was generated using XLamp XP-E2 red, cool white, and blue LEDs respectively (CreeLED). The LED light was passed through band-pass filters (FF01-640/20, Semrock, ET550/20×, Chroma, FF01-465/30, Semrock) to separate the RGB visual stimulation bands from the photomultiplier detection channels. LED illumination was controlled via an open-source programmable LED driver (https://github.com/Llamero/Four_Channel_MHz_LED_Driver). Specifically, LED illumination was timed to occur only during the flyback of the resonant mirror to avoid cross-talk between LED induced excitation of the sample and the two-photon imaging, and the LED channels were synched to the DLP RGB output signal.

Stimulation protocol

Light stimuli were generated using custom software written in Igor Pro 9.0 and using PsychoPy toolbox (v2022.2.0 or earlier). Drifting bright bars (500 µm s⁻¹, 200 × 750 μm or 500 × 750 μm; 1.47 × 10⁵ photons s⁻¹ μm⁻² MP-2000, 2.083 × 10⁶ photons s⁻¹ μm⁻² Hyperscope) were presented on a dark background (4.47 × 10³ photons s⁻¹ μm⁻² MP-2000, 3.63 × 10⁴ photons s⁻¹ μm⁻² Hyperscope) and drifted orthogonal to the long axis of the bar at speeds of 500 µm s⁻¹ (corresponding to 2.24 degrees per second using the conversion factor of 223 µm per degree from ref. ⁵³). All angles indicate direction of stimulus motion on the retina. The total stimulation area was 750 µm² (MP-2000) or 1,000 µm² (Hyperscope). Bar direction order was pseudorandomized in blocks such that all directions were presented before repeating and there was an interstimulus interval of 5 s between bar stimuli. A subset of cells were centred in the scan field and probed with bright bars (500 × 750 μm) moving in the preferred and null directions at a range of velocities (125–2,000 μm s⁻¹, 0.57–9 degrees per second). One centred cell was also stimulated with bright spots of increasing diameter (25–750 μm). We allowed 35 s from the laser onset to first stimulus presentation to allow for laser and background adaptation⁵⁴. A translating lens was used to axially offset the visual stimulus from the imaging plane by ~200 µm (for macaque) or ~150 µm (for mouse) so the visual stimulus was focused on the photoreceptors while acquiring two-photon scans of the GCL. For some stimulus protocols, a recording was collected with no laser stimulation for background subtraction of any residual stimulus light reaching the detectors (MP-2000 recordings only).

Imaging procedures for mouse were as for macaque except the drifting bar stimulus was 500 × 750 µm, drifting at 250 µm s⁻¹ (8 degrees per second)⁵⁵, a speed that stimulates both ON-type and ON–OFF-type DSGCs in mouse retina^29,34,56,57.

Pharmacology

In some experiments, the GABA_A receptor antagonist gabazine (SR95531; HelloBio HB0901) was added to the bath solution (final concentration of 10 μM gabazine for 5 cells, 20 μM for 1 cell). Stock solutions were prepared in H₂O and stored at −20 °C until use. For pharmacological experiments, candidate ON-DSGCs were first identified online by pixel-based vector sum mapping and were centred in the scan field. Calcium responses were then recorded at baseline and at least 4 min after drug wash-in to ensure effects had reached steady-state.

Data analysis

Single-cell RNA-sequencing data analysis

For transcriptomic analysis, we mined existing single-cell RNA-sequencing datasets from macaque (GEO accession GSE118480¹⁴), human (GEO accession GSE148077⁴¹), and mouse (GEO accession GSE137400²⁵) retina. Cell cluster assignments were as reported in the original publications. Dot plot visualizations were generated using the Broad Institute Single-Cell Portal (https://singlecell.broadinstitute.org/single_cell) where dot size indicates the proportion of cells in the cluster that expressed the gene and dot colour indicates the relative gene expression level for each row. To identify a candidate ON-DSGC type, we compared expression of GABRA2 and GLRA2 between each peripheral RGC type and its corresponding foveal RGC type. Cell types were excluded as possible candidates if the difference in expression of GABRA2 and/or GLRA2 was >2 log fold and P < 0.05 by Wilcoxon rank-sum test.

Calcium imaging

Cell somata ROIs of Cal-520 AM- or Cal-590 AM-loaded cells, or GCaMP6f-expressing cells were manually drawn in ImageJ and the average intensity of the ROI over time was extracted for each cell. As detailed above, for some retinas, a no-laser control recording was taken to subtract any low-level stimulus bleedthrough reaching the detectors. For experiments without the no-laser control, a ‘background ROI’ was drawn in an area containing a blood vessel, with no calcium indicator present, and subtracted from the raw recording.

Raw fluorescence intensity values were imported into Igor Pro (version 9.0.0.10, Wavemetrics) and ΔF/F values were calculated for each cell using the following equation:

where F is instantaneous fluorescence and F₀ is the mean fluorescence over a 23-s window prior to stimulus onset. For some ROIs where there was a slow background oscillation, a high pass filter was applied to restore the baseline. For analysis of responses to drifting bar stimuli, we extracted the time windows when visual stimuli were presented, sorted the directions and averaged the maximum ΔF/F amplitudes from three trials. The extent of direction selectivity was reported as either the normalized vector sum where the magnitude of the vector sum was divided by the scalar sum of responses to all of the recorded directions (Fig. 2f,g), or in the case of comparison of directional tuning before and after drug application, using the following formula:

The DSI ranges from −1 to 1 with values closer to 1 indicating higher direction selectivity. Negative values indicate a residual response in the opposite direction to the original preferred direction, which was seen in some cells after gabazine application. Only cells that were responsive to the light stimulus, (defined as having an average amplitude >1.5 above baseline s.d.), were included in analysis. Responsive cells were confirmed by manually examining each trace to omit traces where activity was not correlated with the light stimulus. The preferred angles were adjusted so that the superior, temporal, inferior and nasal corresponded to 90, 0, 270, and 180 degrees, respectively. The calcium responses were fit with the von Mises distribution (polar plots in Figs. 2 and 3 and Extended Data Fig. 5), which is the circular analogue of the gaussian distribution, given by:

where R max is the maximum response, μ is the preferred orientation in degrees, and κ is the width of the tuning curve.

For quantifying calcium responses during bar stimuli moving at different velocities, we measured the maximum ΔF/F amplitudes over a 0.5-s window during the preferred direction responses. For quantification of responses to different spot sizes, we measured the integral of the ΔF/F during the 2-s spot stimulation. For the mouse data, group polar graphs and histograms (Extended Data Fig. 5e,f) were generated using custom Python code (Google Colaboratory).

Pixel-based vector sum mapping

Potential DSGCs in the scan field were identified by pixel-based vector sum mapping using a custom macro written in ImageJ. In brief, the XYT series was first converted to a ΔF/F movie and gaussian filtered with a sigma value of 2. Substacks corresponding to each bar presentation period were extracted and z-projected (maximum). The x and y components of the response vectors were then calculated for each pixel in the image and the responses for each angle were averaged. A heat map visualization was used to identify pixels within the scan field showing the highest vector sum values.

Image registration

After immunostaining, we acquired an epifluorescence image of retinal pieces used for calcium imaging. The calcium imaging region was identified through matching of vascular landmarks and xy stage reference coordinates. Confocal z-projections of the 2 P scan areas were then acquired and stitched. Confocal images were registered to each 2 P imaging field using custom macros incorporating the bUnwarpJ 2.6.12 plugin in ImageJ (1.53q, NIH). For registration, we made an average z-projection of the 2 P scan field and upsampled the resolution to match the confocal image. To identify BNC2⁺ and FOXP2⁺ cells, we set an intensity threshold of 2× the background fluorescence. All scan fields were then manually checked to ensure detection of all immunostained nuclei.

Tracing DiI-filled cells

The morphologies of DiI-filled cells were traced from confocal image stacks using the Simple Neurite Tracer (SNT) 4.1.9 plugin in ImageJ. Semi-automated tracing was performed with the A* search algorithm enabled. To fill the skeletonized cells, the distance threshold settings were adjusted manually to best match the confocal image. Somas were traced and filled separately, and their size constrained in cases of oversaturation in the confocal image. In some cases where the connection of the primary dendrites to the soma was not visible due to soma overexposure, the paths were extrapolated using the shortest path. Depth coding of the traced and filled cells was performed by dividing the stack into two layers (GCL to ON-ChAT border and ON-ChAT to INL). Each cell was also inspected to determine the classification of the dendrites in each layer, and manually adjusted to reflect the raw image. Images of cell traces were upsampled (bicubic interpolation) for improved visualization.

IPL depth and co-fasciculation analysis

To determine the IPL stratification depth of the DiI-filled cells, z-axis intensity profile plots were generated from regions of the dendritic arbour imaged at 63× (Airyscan). For each cell, 1 or 2 regions were selected, ranging in area from 745 μm² to 17,375 μm². The IPL borders were identified by Hoescht staining, or using the position of the ChAT somas. The average background fluorescence value was subtracted and values were normalized to the maximum fluorescence. Negative values were set to 0. The total IPL depth (average 32.8 ± 4.3 μm) was converted to a percentage value from 0 to 100% (where 0 is the INL border and 100% the GCL border).

The extent of co-fasciculation between the DiI cell fills and ChAT labelling was estimated from 20× images of the complete cell fills. The DiI channel contained the traced and filled binary mask of the cell, and the raw ChAT channel was median filtered and thresholded. The DiI channel was used to mask the ChAT channel to determine the overlapping contact area for each slice in the image stack. The overlapping areas were summed per image and expressed as a percentage of total DiI area. Contact area was measured with the ChAT channel in the normal orientation, and after rotating the ChAT channel about the z-axis by 90 degrees.

3D surface rendering

Complete DiI-filled cells were rendered in Imaris 9.8.2 (Oxford Instruments) from cells traced and filled using a manually adjusted threshold in SNT (see ‘Tracing DiI-filled cells’). ‘Wrap-around’ synapses were rendered in Imaris from 63× Airyscan image stacks. The images were first made isotropic in ImageJ by scaling the pixel height and width to match the voxel depth. For both complete DiI cells and wrap-around synapses, channels were surface rendered in Imaris using default surface settings and manually adjusted thresholds. A cropped region of the DiI dendrite was selected and the ChAT surfaces were filtered to include only those within 1 μm of the DiI dendrite.

Cell classification, density and mosaic analysis

We classified BNC2⁺ RGCs in tile-scanned confocal z-stacks of the macaque GCL taken from four retinas from different animals. Retinal pieces were from peripheral nasal retina between the optic nerve head and far nasal periphery (nasal equivalent eccentricity ~3.5–8.5 mm, average total RGC density 1,251 ± 217 cells per mm²). Scan fields were acquired at an xy resolution of 1.6–2.41 pixels µm⁻¹ and tile-scanned with 10% overlap. The stitched image regions that were used for analysis averaged 8.44 ± 5.16 mm². Image stacks were 2D median filtered to reduce pixel noise and improve visibility of labelled nuclei for subsequent segmentation. The nuclei of BNC2⁺RBPMS⁺ cells were manually segmented and checked for accuracy by an additional observer. Intensity measurements were then extracted for the BNC2⁺ and FOXP2⁺ channels and z-scored to permit comparison across samples that had been processed or imaged on different days. FOXP2⁺ cells were classified based on a z-score threshold intensity value which was cross-checked against manual classification. Cell clusters were independently confirmed using k-means clustering, implemented in Igor Pro 9.

For cell density analysis, RGCs were segmented based on RBPMS staining using a semi-automated approach. In brief, a 2D or 3D gaussian blur filter was applied, images were thresholded and objects of interest were detected using automated particle analysis. Alternatively, RGCs were counted by applying a gaussian filter and then using the Find Maxima function in ImageJ. Manual corrections to segmentation were made to correct false positive or false negative segmentation. For spatial cell density plots, images were divided into 500 × 500 µm square sampling regions. In sampling regions where image focus was poor or tissue was damaged or obscured, cell counts were estimated by averaging the surrounding regions. The coverage factor was given by the spatial density of cells (cells per mm²) multiplied by the dendritic field area of that cell type (mm² per cell). A coverage factor of 1.0 indicates complete tiling with no spaces between arbours or arbour overlap. Higher coverage factors are indicative of greater overlap of the dendritic arbours.

Mosaic analysis was performed on the same regions as described above. Voronoi domain areas were measured for each cell using a custom macro utilizing a built-in ImageJ function (http://biii.eu/voronoi-imagej). The Voronoi domain regularity index (VDRI) was used to assess mosaic regularity and is given by the average Voronoi domain area divided by the standard deviation. Real VDRIs were compared to VDRIs from random simulations of cells at the same density. VDRI values of 1.9 and lower indicate a random array of points²⁶. Voronoi domains intersecting with the image boundaries were excluded from analysis. The average VDRI for each retina was calculated before averaging across animals. The density recovery profile (DRP) was determined as described previously⁵⁸ and implemented using the sjedrp package in R (sjedrp v0.12, https://github.com/sje30/sjedrp/blob/master/DESCRIPTION; Stephen Eglen). The coordinates of the pRGC10 and pRGC16 centroids were used to calculate the DRP using radii of 0–1,000 µm in 20-µm increments. The DRP profiles from each retinal piece were pooled and mirrored on the x axis to better observe the ‘well-like’ exclusion zone²⁷. DRPs were also calculated for the random simulations. To calculate the approximate convergence, the bottom of the ‘well’ (average of 3 lowest bins) was taken as a percentage of the average density from 390−1,000 µm (the plateau of the density profile). Data from the first bin were excluded due to elevated variability associated with the smaller area of the innermost annulus⁵⁹.

Fluorescence in situ hybridization analysis

We identified BNC2⁺FSTL4⁺ RGCs in a horizontal section from one macaque retina (superior-nasal region, nasal equivalent eccentricity ~7.4 mm). 16 bit confocal tile scan stacks were acquired with a Plan-Apochromat 20×/0.8 air objective at an xy resolution of 1.93 pixels per µm. For analysis, an area approximately 1.2 mm² was z-projected (sum slices). RGCs were segmented from RBPMS staining using the semi-automated approach described above. The number of puncta per cell was quantified using ACD guidelines (SOP-45-006, ACD). In brief, the average background intensity per fluorophore was determined from a region without cells. This was comparable to the negative control background level. Average intensity per single dot was determined by an average of at least 21 single dots. Then, number of dots were determined per RGC by this formula:

After excluding artefacts such as blood vessels or high cell background, cells were considered BNC2⁺ if they had more than 40 dots and FSTL4⁺ if they had more than 80 dots.

Statistics and reproducibility

All data in the figures and text are presented as mean ± s.d. or median and interquartile range unless otherwise indicated. Non-parametric tests were used due to non-normal distribution of data (as determined by the Shapiro–Wilks test) or when sample size was small. For comparisons of independent samples, the Wilcoxon rank-sum test was used. For paired comparisons, the Wilcoxon signed rank test was used. All tests were two-sided and used an alpha level of 0.05 except where multiple comparisons were made. No methods were used to determine sample size a priori. No experimental groups were assigned in this study. Data acquisition and analyses were not performed with blinding to the experimental conditions as most experiments did not involve a treatment or perturbation and analyses were automated. For box plots, interquartile ranges were calculated using the Tukey method. Where representative micrographs are shown, experiments were replicated in multiple retinas as follows: Fig. 1b (n = 9 retinas), Fig. 2a (n = 6 retinas), Fig. 3a,b (n = 3 retinas), Fig. 3e (n = 3 retinas), Fig. 4b (n = 2 retinas), Extended Data Fig. 1a (n = 4 retinas), Extended Data Fig. 5b (n = 3 retinas), Extended Data Fig. 6a,b (n = 6 retinas, as in Fig. 2), Extended Data Fig. 7a,b,d–f (n = 2 retinas) and Extended Data Fig. 8 (n = 4 cells from 2 retinas). Micrographs from a single retina are shown in Extended Data Fig. 2d,e.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.