All animals used in this study were adult mice (8–12 weeks) on a C57 Black 6 background, with equal numbers of males and females. Animals were housed at the Vanderbilt University Division of Animal Care facility and subjected to a 12-h light/dark cycle. Animals were provided with water and rodent chow ad libitum. Cre 77.6 mice (Jackson Labs #024098) were crossed with MORF3 mice (Jackson Labs #035403) to generate G-MORF mice. In G-MORF mice, a highly antigenic spaghetti monster fluorescent protein (smFP) with 20 V5 epitope tags is targeted to cell membranes through a farnesylation sequence (Hancock et al., 1991). Critically, this construct requires a stochastic frameshift mutation during development for the smFP to be properly translated. This feature results in around 1% of total retinal astrocyte labeling with smFP, enabling individual cells to be distinguished from one another (Supplementary Figure S1). Each mouse yields between 75 and 150 well-separated, membrane-labeled astrocytes per retina allowing for morphology visualization at a quality comparable to whole-cell dye filling (Supplementary Figures S1, S3). The smFP is not endogenously fluorescent. For whole-cell patch dye filling, we crossed Cre 77.6 mice with a floxed tdTomato reporter mouse (Jackson Labs #007914).

One week prior to sacrifice, mice were anesthetized with a solution of ketamine / xylazine (90 mg/kg ketamine, 5 mg/kg xylazine) and injected with a solution of 1% fluorescein in 1X PBS. An angiogram was then taken as a retinal fingerprint which would later allow for the determination of retinal orientation following immunohistochemistry. G-MORF mice were sacrificed by means of intraperitoneal pentobarbital injection followed by transcardial perfusion of 1X PBS. Eyes were immediately enucleated, and the retinas dissected fresh. Retinas were transferred to a solution of collagenase (LS005273, Worthington Biochemical) and hyaluronidase (LS002592, Worthington Biochemical) in Ame’s media (A1372-25, US Biologic) for 10 min at room temperature to digest the vitreous. Following this incubation, vitreous was removed as a single sheet using forceps. Relief cuts were made to divide the tissue into quadrants and the retinas then transferred to a solution of 4% paraformaldehyde to shake at room temperature for 1 h. Retinas were stored in 1X PBS with azide at 4°C until used.

Tissue was blocked in a solution of 5% normal donkey serum (NDS) in 0.1% Triton-X for 3 h, while shaking at room temperature. The tissue was then incubated in a solution of primary antibody in 0.1% Triton-X with 3% NDS for 3 days, while shaking at 4°C. Primary antibodies used in this paper include: rabbit anti-V5 (1:500, Bethyl Laboratories, A190-120A), goat anti-V5 (1,500, Abcam, ab95038), isolectin GS-IB4 biotin-XX conjugate (1,500; Invitrogen, I21414), goat anti-GFAP (1,500, Abcam, ab53554), rabbit anti-NG2 (1,500, Millipore, AB5320), mouse anti-SMI-31 (1,500, Biolegend, 801,601), rabbit anti-NLRP3 (1,250, Cell Signaling, 15,101), goat anti-IBA1 (1,250, Novus Biologicals, NB100-1028), and mouse anti-beta III tubulin (1,500, Millipore, MAB5564). Following three five-minute washes of the retinas in 1X PBS, tissue was incubated in a solution of secondary antibody. Secondary antibodies and probes used in this paper include donkey anti-rabbit conjugated to Alexa 488 (1,200, Jackson ImmunoResearch, 711–545-152), donkey anti-goat conjugated to Alexa 488 (1,200, Jackson ImmunoResearch, 705–545-003), donkey anti-rabbit conjugated to Alexa 555 (1,200, Invitrogen, A-31572), donkey anti-goat conjugated to Cy3 (1,200, Jackson ImmunoResearch, 705–165-147), streptavidin conjugated to Alexa 647 (1,200, Invitrogen, S21374), and phalloidin conjugated to Alexa 647 (1,200, Invitrogen, A22287). Retinae were washed three times for 5 minutes with 1X PBS and then mounted with Fluoromount-G on slides for imaging.

Animals were euthanized by cervical dislocation and eyes dissected under red light conditions (630 nm, 800 μW/cm2, FND/FG, Ushio, Cypress, CA). Retinas were transferred to a carbogen-saturated Ames’ medium (US Biologic, Memphis, TN) supplemented with 20 mM D-glucose and 22.6 mM NaHCO3 (pH 7.4, 294 Osm). Relief cuts were made into each retina and flat mounted onto a physiological chamber. Ames’ media perfused each retina at a flow rate of 2 mL/min at 35°C (Model TC-344C, Warner Instruments, Hamden, CT). Individual astrocytes were located using a 40X water immersion objective on an Olympus BX50 microscope. Cells were patched onto in a whole-cell configuration using a borosilicate pipette (I.D. 0.86 mm, O.D. 1.5 mm; Sutter Instruments, Novato, CA) filled with (in mM): 125 K-gluconate, 10 KCl, 10 HEPES, 10 EGTA, 4 Mg-ATP, 1 Na-GTP, 0.1 Alexa 488 and 0.1 Alexa 647 dye (Invitrogen, Carlsbad, CA). The intracellular solution pH was 7.35 and osmolarity was 285 Osm.

Immunolabeled retinas were first imaged at 20X magnification on a Nikon Ti-E Spinning Disk confocal microscope en montage using the focus surface interpolation tool in Nikon Nis-Elements AR 5.21.03. This image served as a map to direct imaging of individual cells and also to mark their coordinate location in the flattened retina. Z-stacked images were taken of individual astrocytes at 60X magnification through the entirety of each cell’s volume at a step size of 0.3 μm. Images for each channel were taken serially at each Z-step.

Morphology was quantified for astrocytes which met the criteria of being V5-positive, GFAP-positive cells with cell bodies residing in the Nerve Fiber Layer (NFL) or Ganglion Cell Layer (GCL). Cells were not included in analysis if they were in contact with any other V5-labeled astrocyte or could not have their outline determined accurately. Using this exclusion criteria, just over 1,000 cells were analyzed.

Cartesian coordinates for cells on flat-mounted retinae were mapped to geographic coordinates in native retinal space using the open-source program RETISTRUCT (Sterratt et al., 2013). It and its dependencies were installed onto a machine running Windows 10 using a Microsoft Image of the CRAN from 02/11/2019, R version 3.4.2, and Rtools35. Using the fluorescein angiograms alongside the 20X IB4 montages, retinal orientation was determined. This information along with cartesian coordinates of each cell were input into RETISTRUCT. The φ0 angle used during reconstruction was 22 degrees. Picturing the retina as a hemisphere, the rim (far periphery) is located at 0 degrees latitude and the ONH is located at −90 degrees. Longitude is taken such that zero degrees is temporal for left eyes and nasal for right eyes. 90 degrees longitude is always dorsal. The raw data outputting latitude and longitudes of each cell was read into a custom Python program which translated coordinates for all right eyes into a common left-eye coordinate space.

Cell outlines were generated using a semi-automated, custom ImageJ macro. All functions referenced are available in Fiji ImageJ version 2.3.0 on Windows operating systems. This program flattened a V5 labeled astrocyte Z-stack using the standard deviation Z-stack function. This flattened image was then automatically thresholded and binarized. Outlines of cells were acquired using the Particle Analyzer Tool looking for particles with size bounds of 3 μm² to infinity. ROIs were then generated, and a user scanned through them to find ROIs that either contributed to the cell’s outline or its holes. These ROIs were then subjected to the XOR operation to get the final complex cell outline. Using this macro to determine outlines takes about 1 min or less per cell. After the complex cell outline was generated, its convex hull was also generated. Each of these outlines were subjected to shape analysis using the following parameters in the ‘Set Measurements’ window: [area, center of mass, shape descriptors, area fraction, centroid, perimeter, fit ellipse, and Feret’s diameter]. The GFAP channel was processed in a similar way, except that binarization was achieved using the Phansalkar Auto Local Threshold function with a radius of 15. This automatic thresholding was sufficient for all images. Throughout this paper, projection area is used to refer to the area of V5-positive labeling in a Z-projected astrocyte image. Cell outlines and binarized GFAP images were used to determine percent GFAP-positive projection area.

Distance from each cell’s convex hull centroid to the closest blood vessel was determined using a custom Python script. This script read in the convex hull information from the ImageJ macro results and displayed a standard deviation Z-projected image to a user showing the V5 and IB4 channels. A circle was overlaid on the image centered at the centroid, and an adjustable slider changed the radius of this circle. A user adjusted the slider until the edge of the circle touched the closest vessel and then this value was recorded. Similarly, the diameters of unique vessels contacted were determined using another Python script with the same V5/IB4 images. In this program a user noted the boundary points of each unique vessel and their diameters were displayed and recorded. Unique vessels are defined after branching events (where each branch results in one continuing vessel and one new one). Only the number of unique contacts was quantified- the total number of contacts which includes repeated contacts to the same vessel was not.

Visualization of all data was accomplished using the Python library Matplotlib and GraphPad Prism. BioRender was used to generate the image of the Micron IV and retinal vasculature in Supplementary Figure S2. All other figures were crafted in Adobe Illustrator and Python. UMAP (Uniform Manifold Approximation Projection) was run using the Python library UMAP-Learn with n_neighbors = 15. Running UMAP with a variety of n_neighbors values yielded similar embedding shapes. Thus, the default value of 15 was chosen. The 20 astrocyte morphology parameters that were subjected to the UMAP algorithm include: Full-Cell outline (FO) area, FO perimeter, FO major axis, FO minor axis, FO circularity, FO Feret diameter, FO roundness, FO solidity, area of the convex hull (CVH) of the FO, CVH perimeter, CVH Feret diameter, latitude (radians), longitude (radians), percent of FO projection area that is GFAP-positive, distance from the CVH centroid to the nearest blood vessel, the distance from the CVH centroid to the FO center of mass, mean vessel diameter for unique vessels contacted, minimum vessel diameter for unique vessels contacted, maximum vessel diameter for unique vessels contacted, and the number of unique vessels contacted. The UMAP embedding was clustered by density using Hierarchical Density-based Spatial Clustering of Applications with Noise (HDBSCAN). HDBSCAN parameters used were min_samples = 3 and min_cluster_size = 10. The embedding was also clustered according to each individual UMAP input parameter.

Data and code not explicitly shown in this manuscript are available upon request to the corresponding author.

Cells with stellate morphology have a central cell body which radiates processes uniformly in all directions, much like a star. We overcome limitations of current astrocyte visualization techniques using the G-MORF mouse line and find that this description does not accurately describe most retinal astrocytes (Supplementary Figures S1–S3). Even cells which appear stellate from V5 labeling alone commonly have their cell body in an unexpected location toward the periphery of the cell mass (Figure 1A). S100β prominently labels the soma, which is often far from the centroid of the cell’s V5 convex hull (Figure 1A, distancecentroid−soma=17.1±9.1μm,mean±stdev). Example cells from G-MORF mice exhibiting non-stellate morphology are shown in Figure 1B. Immunolabeling for NLRP3 inflammasome and microglial ionized calcium-binding adapter molecule-1 (IBA-1) did not show signs of immune activation from smFP expression which could affect astrocyte morphology in G-MORF mice (data not shown).

Astrocyte morphology revealed using immunolabeling against GFAP and S100β underestimates actual complexity as shown by V5. Some V5-labeled processes in G-MORF mice are GFAP-negative (Figure 2A), and many membrane protrusions are relatively planar rather than fibril-like (Figure 2B). Low levels of GFAP also obscure finer processes (Figure 2C). Retinal astrocyte spatial domains often overlap extensively (Figures 2D, 3A), indicating that pan-astrocyte markers alone cannot distinguish individual cells. GFAP signal represents a relatively small proportion of a cell’s projection area (21.1 ± 0.2%, Figure 2E), which depends on both cell size and retinal location (Figures 2F–G). Its content also varies with eccentricity, accounting for a larger percentage of cell projection area at the nerve head and in the far periphery and a smaller fraction as cell size increases. S100β similarly cannot reveal individual cell morphology. It does not uniformly label V5-positive area and is more prominently associated with the cytoskeleton than with cell membrane or cytosol (Figure 3B).

V5 labeling reveals recurring astrocyte structural motifs, which we now define. Beads are small, ellipsoid and triangular inflations of cell membrane along thin projections (Figure 4Ai). Sails are polygonal sheets of membrane that often have thin projections emanating from their vertices (Figure 4Aii). These can either be small motifs or encompass the majority of the cell volume. Tubes form as an entire cell or its processes envelop a blood vessel (Figure 4Bi). Bristles are short, fine projections on the order of 5 μm in length (Figure 4Bii). The smallest bristles often terminate in a roughly spherical shape and are between 0.7 and 2.5 μm in length. Pads and end-feet are morphologically similar but are distinguished based on which structures they contact. We define pads as rounded, flat projection endings which do not contact blood vessels, whereas end-feet terminate on blood vessels but do not envelop those vessels completely (Figures 4Cii,Di). We also observed that holes commonly form within large sheets of membrane (Figure 4Ci). These motifs align with morphology observed in whole-cell dye filling (Figure 5).

Structural motifs are predictive of the underlying architecture in the Nerve Fiber Layer (NFL) and Ganglion Cell Layer (GCL). In Table 1 we quantify this predictive capacity by determining the probability of observing an underlying structure given the independent observation of an astrocyte motif, p(structure|motif). Sails are predictive of both RGC soma (41%) as well as bundles of RGC axons (65%) and weakly predict neighboring astrocytes (6%) (Figures 6A,C,D). As membrane sheets, they cover large regions of axon bundles but do not envelop them completely. They may also cover one or more RGC soma. Tubes are predictive of underlying blood vessels (100%), only occurring when one is present at its interior (Figure 6B). End-feet are morphologically similar to pads, with the distinction being their association with blood vessels and not RGC soma. As they are defined based on their vascular interaction, p(vessel|endfoot) = 100%. Their morphology can be roughly flat and circular like pads or can be more curved to complement the curvature of the vessel itself (Figures 4D, 6A). Similarly, they may terminate at a blood vessel as a larger mass of membrane that does not have a very distinct shape (Figure 6D). Beads are most predictive of axons (76%) but also contact RGC soma (22%). When contacting soma, we observed bead placement on the apical surface close to the hillock region as well as intercalating between the lateral surfaces of adjacent RGC soma (Figures 6A, 7A,B). When contacting axons, we observed beads often contacting varicosities on those axons (Figure 7C). Pads similarly contact RGC axons (43%), soma (43%), and other astrocytes (20%). Unlike beads, pads were only observed to contact the apical surface of RGC soma (Figure 7D). Bristles contact individual RGC axons (41%) and soma (34%) as well as neighboring astrocytes (20%) (Figure 8). They contact segments of axon with and without varicosities (Figure 8A) and often trace individual axons for the entire length of the motif (Figures 8A,C). Bristled processes can additionally span multiple axon bundles at the optic nerve head, making fine connections to axons in each fascicle (Figure 8B).

We used co-labeling of V5 and IB4 to determine whether morphological differences would be found between astrocytes exhibiting motifs interacting with vasculature and those which did not. To our surprise, every astrocyte contacted at least one blood vessel, with most cells contacting multiple unique vessels (Figures 9A,B). On average, astrocytes contacted 2.57 unique vessels (Figure 9B). We similarly found that every astrocyte contacted some neuronal element in the GCL or NFL, be it an axon or soma (Figure 9A). Additionally, individual astrocyte processes were not limited to making mutually exclusive contacts to blood vessels or neuronal elements. Rather, many processes contacted both (Figure 9C). In these instances, both axons and blood vessels appeared to act as tracks guiding processes to one another. We observe a similar interaction between astrocyte processes and axons undergoing fasciculation (Figure 9D).

All astrocytes interact with the vasculature through one of three mutually exclusive connection types (Figure 10A). Cells with enveloping connections appear as tubes which completely enwrap both major vessels as well as capillaries, with either few or no processes extended. They represent 8% of analyzed cells. To ensure they were not vascular pericytes, we co-labeled retinae for chondroitin sulfate proteoglycan (NG2), V5 and GFAP. V5-positive cells making enveloping connections are NG2-negative and GFAP-positive (Figure 11). The mural connection type is similar, with these cells only partially enveloping the vasculature. They represent 7% of analyzed cells and are commonly found at the junction of diverging vessels. Enveloping and mural contacts are made with the majority of the cell’s mass. Finally, 85% of cells make non-enveloping contacts. These include end-feet and small tube motifs.

These vascular connections describe overall cell shape and explain a significant portion of morphological heterogeneity. We quantified morphological features for over one thousand astrocytes and reduced that feature space to two dimensions using UMAP. Classes of vascular interactions, when applied to color the embedding, are predictive of plot structure with the caveat that distinguishing between cells making mural and enveloping connections did not provide additional prediction power (Figure 10B). Density-based clustering of the embedding revealed three populations of cells which do not directly correspond to vascular-defined classes (Figure 10B). We also colored the embedding according to each feature used as a UMAP input parameter (Figure 10B) which revealed that plot structure is explained in part by cell size and polarization (Feret diameter and minor axes). Cluster 1 contains cells with a large convex hull area and long Feret diameter, while drivers of Cluster 2 and 3 were not apparent. A dendrogram showing all correlations between UMAP input parameters is shown in Figure 10C.

Using the data collected for UMAP, we separately investigated whether astrocyte morphology varied with retinal location. We plotted individual morphological parameters against geographic coordinates for each cell in native retinal space. Latitude provides information regarding distance from the ONH whereas longitude provides information about retinal region (nasal, temporal, dorsal, ventral). Longitude was not predictive of any parameter that was quantified; however, many varied with latitude. When plotted against latitude, cell area and Feret diameters exhibit a slight negative curvature while solidity shows a slight positive curvature (Figure 12A). This indicates that the populations of astrocytes near the ONH and far periphery are compressed in shape relative to cells in the mid-retina (Figures 12A,B). Additionally, astrocytes in the mid-retina are more likely to contact larger numbers of unique blood vessels than their counterparts at the ONH and periphery (Figure 12C). These relationships are complementary in that cells with larger breadth can contact more vessels than those with compressed morphology. Cells which contact more unique vessels are also more likely to contact smaller vessels, whereas cells with compressed morphology tend to contact larger vessels (Figures 12A–C).