Investigation of heterocellular features of the mouse retinal neurovascular unit by 3D electron microscopy

Abstract The retina has a complex structure with a diverse collection of component cells that work together to facilitate vision. The retinal capillaries supplying the nutritional requirements to the inner retina have an intricate system of neural, glial and vascular elements that interconnect to form the neurovascular unit (NVU). The retina has no autonomic nervous system and so relies on the NVU as an interdependent, physical and functional unit to alter blood flow appropriately to changes in the physiological environment. The importance of this is demonstrated by alterations in NVU function being apparent in the blinding disease diabetic retinopathy and other diseases of the retina. It is, therefore, imperative to understand the anatomy of the components of the NVU that underlie its functioning and in particular the nanoscale arrangements of its heterocellular components. However, information on this in three spatial dimensions is limited. In the present study, we utilised the technique of serial block‐face scanning electron microscopy (SBF‐SEM), and computational image reconstruction, to enable the first three‐dimensional ultrastructural analysis of the NVU in mouse retinal capillaries. Mouse isolated retina was prepared for SBF‐SEM and up to 150 serial scanning electron microscopy images (covering z‐axes distances of 12–8 mm) of individual capillaries in the superficial plexus and NVU cellular components digitally aligned. Examination of the data in the x‐, y‐ and z‐planes was performed with the use of semi‐automated computational image analysis tools including segmentation, 3D image reconstruction and quantitation of cell proximities. A prominent feature of the capillary arrangements in 3D was the extensive sheath‐like coverage by singular pericytes. They appeared in close register to the basement membrane with which they interwove in a complex mesh‐like appearance. Breaks in the basement membrane appeared to facilitate pericyte interactions with other NVU cell types. There were frequent, close (<10 nm) pericyte–endothelial interactions with direct contact points and peg‐and‐socket‐like morphology. Macroglia typically intervened between neurons and capillary structures; however, regions were identified where neurons came into closer contact with the basement membrane. A software‐generated analysis to assess the morphology of the different cellular components of the NVU, including quantifications of convexity, sphericity and cell‐to‐cell closeness, has enabled preliminary semi‐quantitative characterisation of cell arrangements with neighbouring structures. This study presents new data on the nanoscale spatial characteristics of components of the murine retinal NVU in 3D that has implications for our understanding of structural integrity (e.g. pericyte‐endothelial cell anchoring) and function (e.g. possible paracrine communication between macroglia and pericytes). It also serves as a platform to inform future studies examining changes in NVU characteristics with different biological and disease circumstances. All raw and processed image data have been deposited for public viewing.


| INTRODUC TI ON
The retina is a multi-layered structure with a diverse collection of component cells that work together to produce a complex visual output (Hoon et al., 2014). Light is detected by rod and cone photoreceptors which convert it into an electrical signal for propagation through the retinal circuitry to the optic nerve and subsequently the visual cortex of the brain. It is well known that the retina is one of the most metabolically active tissues in the body, with a high demand for oxygen and nutrient delivery from the vasculature (Morrow, 2014).
To fulfil these needs, the retina is nourished by two distinct vascular beds: the intraretinal vasculature that supplies the inner retina and the choroidal vessels which serve the outer retina (Flammer & Mozaffarieh, 2008).
The blood vessels of the inner retina form part of the retinal neurovascular unit (NVU). The NVU refers to the physical and functional relationship among the capillary blood-vessel endothelial cells, pericytes, macroglia (Müller glia and astrocytes), microglia and neurons (ganglion cells, bipolar cells, horizontal cells and amacrine cells) of the retina (Simó et al., 2014). Collectively, the cells of the NVU play a key role in the maintenance of the inner blood retinal barrier, in mediating cell-to-cell survival via paracrine signalling, and in matching blood flow to the metabolic needs of the retina through a process known as neurovascular coupling (Duh et al., 2017;Metea & Newman, 2007). The latter is particularly important given the absence of autonomic innervation in the retina (Ye et al., 1990).
Disruption of the integrity of the retinal NVU has been implicated in the pathogenesis of several retinal diseases, including diabetic retinopathy, glaucoma and retinal neurodegenerative disorders (Ivanova et al., 2019;Lechner et al., 2017;Weinreb et al., 2014). Despite this, the three-dimensional ultrastructure of the retinal NVU remains to be fully characterised and quantitative methods for describing its key features have yet to be developed.
The retinal NVU in mice and humans has been previously analysed in two dimensions at the ultrastructural level using conventional transmission electron microscopy (TEM) (Fehér et al., 2017;van der Wijk et al., 2018). While this technique has provided valuable information on its structure, a greater understanding of the anatomical arrangements of the different cellular components, and their possible heterocellular interactions, requires analysis in three dimensions with nanometre spatial resolution. Serial block-face scanning electron microscopy (SBF-SEM), a technique developed by Denk and Horstmann (2004), allows for the automatic serial thin sectioning and scanning of an embedded tissue or sample in a scanning electron microscope. This allows for the collection of multiple serially registered ultrastructural images which, together with computational image reconstruction software, can provide a threedimensional view of tissue microanatomy, with resolutions as low as 3-5 nm (Cocks et al., 2018;Nian et al., 2021). In the retina, the use of SBF-SEM has so far been limited to studies focused on synaptic connectivity mapping (Behrens et al., 2021;Graydon et al., 2018;Helmstaedter et al., 2013;Wool et al., 2019), the packing geometry of retinal pigment epithelial cell granules (Agrawal et al., 2017) and neuronal changes associated with retinal degenerative phenotypes (Agrawal et al., 2017;Kerov et al., 2018).
In the present work, we have investigated for the first time the murine retinal NVU using SBF-SEM to define its ultrastructural characteristics in three dimensions. Image analysis tools have also been developed to quantitatively describe key aspects of the mouse retinal NVU, providing a basis in the future to better understand how this structure is altered during retinal disease.

| Sample collection and preparation
Tissue samples used in this study were from healthy mouse retina and all capillaries sampled were in the superficial vascular plexus.
An adult C57BL/6 mouse (3 months) was killed by Schedule 1 procedure according to the Animals Scientific Procedures Act (ASPA) 1986, the eyes extracted and retinas micro-dissected into fixative assess the morphology of the different cellular components of the NVU, including quantifications of convexity, sphericity and cell-to-cell closeness, has enabled preliminary semi-quantitative characterisation of cell arrangements with neighbouring structures. This study presents new data on the nanoscale spatial characteristics of components of the murine retinal NVU in 3D that has implications for our understanding of structural integrity (e.g. pericyte-endothelial cell anchoring) and function (e.g. possible paracrine communication between macroglia and pericytes). It also serves as a platform to inform future studies examining changes in NVU characteristics with different biological and disease circumstances. All raw and processed image data have been deposited for public viewing.

K E Y W O R D S
astrocytes, basement membrane, capillaries, endothelium, macroglia, neurons, Neurovascular Unit, Peg-and-socket, pericytes, retina, SBF-SEM (2% glutaraldehyde in 0.1 M sodium cacodylate buffer, pH 7.3) for a minimum of 12 h for electron microscopy. The samples were then processed using a heavy metal staining protocol to provide for an electron dense surface for the beam to interact with (Cocks et al., 2018;Wilke et al., 2013).
The samples were embedded into resin (Taab 812 epoxy resin) and left to polymerise for a minimum of 36 h. The resin blocks were trimmed using a razor blade to form a trapezoid block face. The blocks were then trimmed to approximately 0.75 mm × 0.75 mm and glued onto a pin.

| SBF-SEM settings-Image collection
Resin-embedded retinal tissue samples were imaged using a Zeiss Sigma SEM chamber (Zeiss) combined with Gatan 3View software (Gatan). SBF-SEM collects multiple serial-registered digitised images, following automated ultramicrotome sectioning of the block face in situ. The Gatan 3View software was used to identify retinal capillaries of interest and to ensure that they remained in the centre of the field of view as the images were collected. The diamond ultramicrotome was set to cut sections at either 100 or 120 nm and 100-150 consecutive micrographs were captured for each capillary covering a z distance of 12-18 μm. Image dimensions were set to 3000 × 3000 pixels, at a pixel size of 6 nm, with a 20 μs/pixel dwell time. All raw data analysed in the report are available at the

| Image processing
DM3 files collected from the SBF-SEM were opened in Microscopy Image Browser (MIB v2.1) for post-processing (http://mib.helsi nki. fi/) (Belevich et al., 2016). DM3 image files opened in MIB are displayed with low contrast, so cellular features cannot be easily distinguished. Image contrast was therefore optimised, and contrast normalisation was performed throughout the z-dimension for all images within the stack. Images were then aligned using the drift correction tool. To reduce the dataset size, and speed up dataset operations, the images were converted from 16-to 8-bit, and the processed dataset was saved as an Amira mesh binary (.am) file. An example of individual microscopic images obtained from a stack of digitally aligned image data are given in Figure 1. Video files of the image stacks for each capillary are provided in Files S1- S4 (10.25405/ data.ncl.19086755; 10.25405/data.ncl.19086752; 10.25405/data. ncl.19086617; 10.25405/ data.ncl.19086746).
Before segmentation, the cellular components of the NVU were inspected and assigned an identity based on either their location in the capillary or their morphology. The endothelium lines the lumen of the capillary and is thus the innermost layer of the vasculature. As   Figure 1, the endothelium is surrounded by pericytes, which appear to have a spindle-like morphology. The basement membrane (BM) has a lighter tone and surrounds the pericytes separating them from the endothelium. Macroglia were identified by their abundance of electron-dense endoplasmic reticulum in the cytoplasm (Bianchi et al., 2015;Wakakura & Foulds, 1988

| Image segmentation and 3D reconstruction
Each feature of interest of the NVU was segmented (delineated by colour-coding) throughout z-stacks of data using MIB (http://mib. helsi nki.fi/). Due to the complexity of the cellular morphology, a combination of semi-automatic interpolation and manual image segmentation was performed with the use of tools, filters and interpolation in XY, XZ or YZ planes. Manual segmentation was performed by selecting the MIB brush tool to trace a cell membrane or feature which was then filled (using F or shift + F for the whole dataset). The selection was added to a material (A or shift + A) depending on the cell type. Colour-coded area lists corresponding to specific features in the images were assigned as follows: the BM (brown), endothelial F I G U R E 4 Pericyte-endothelial interaction via direct contact. Close contacts of a pericyte (blue) and endothelial cell (aqua) in absence of a bordering basement membrane (brown) are indicated for the raw data (a and c) and segmented data (b and d) of two sections 1.2 μm apart.
To speed up segmentation, interpolation was performed.
Structures with a similar morphology throughout the slices, often including endothelial cells and neurons, were manually segmented every two to four sections, with gaps automatically segmented using MIB's interpolation tool. Any segments which did not outline the cellular membrane or feature accurately were manually corrected and retracted using the brush tool.

| Identification of Peg-and-socket formations
Inspection of the data revealed that pericyte-endothelial membrane  the intersection of these two materials) are labelled as potential pericyte-endothelial interaction sites. This analysis tagged all pericyte-endothelial peg-and-socket features identified manually as well as regions where the two materials were sufficiently close, but not in peg-and-socket formations (i.e, they did not fit the criteria listed above).

| Proximity analysis
In

| Morphological analysis
Several image analysis procedures were implemented in MATLAB to quantify the morphological properties of the NVU features of interest. The area (A) of each object was determined by counting the number of pixels comprising it. The perimeter (P) of each object was measured by first using MATLAB to determine boundary pixels, and then estimating the perimeter by analysis of its Freeman chain code (Freeman, 1961). Convex hull areas (CA) and perimeters (CP), which enclose the smallest convex area about each object, were calculated equivalently. Radii were measured by determining the central coordinate of the subject area and calculating the 2D Euclidean distance (i.e. r = √ Δ x 2 + Δy 2 ) from it to each of the pixels composing the boundary, taking the mean of this distribution.
Ratios of these key metrics, which are standard in morphological image analysis (Mingqiang et al., 2008) were also recorded F I G U R E 7 Segmented data and 3D reconstruction of macroglia surrounding the vasculature. (a, b) Macroglia in (red and purple shades) can be seen to wrap around the vasculature across different sections along capillary depth. Purple shading depicts a cell identified as an astrocyte. Displayed in a and b are sections 6 μm apart. 3D reconstruction views (c, d) show the complexity of the wrapping along 18 of vessel (the corresponding video file can be found in File S17 (10.25405/data.ncl.19086341).  is evident here and in the 3D digital reconstructions of a complete data stack (Figure 3; see also video files of stacks of segmented capillaries in Files S11-S13: 10.25405/data.ncl.19086398; 10.25405/ data.ncl.19086401; 10.25405/data.ncl.19086131).

| Ultrastructural features of the capillaries
As the BM both enveloped and interwove with the pericyte this influences the closeness of pericytes with underlying endothelium. Several TEM studies have attempted to report the relationship of pericytes to endothelial cells in retinal capillaries.
Estimates from examination of singular ultra-thin sliced images in the x-and y-axes suggest that pericyte coverage of the endothelial surface ranges between 41% and 58% in rodent retinal capillaries (Frank et al., 1987;Tilton et al., 1985). Using our proximity analysis on capillary 1 (see "Methods" section), we have quantified pericyte ensheathment of the endothelium along the z-axis, with a stack mean and standard deviation of 35.3 ± 23.7% of the abluminal endothelial surface lying within 200 nm of the pericyte layer (range 12%-100%).

| Pericyte and endothelial connections
The BM wraps around the abluminal endothelium and pericytes for the most part forming a barrier separating the two cells. However, as indicated above, along the capillary length gaps in the BM were observed allowing pericytes and endothelial cells to come into closer contact. These interactions were of two types, namely, direct contact points ( Figure 4) and peg-and-socket type contact areas ( Figure 5; File S14: 10.25405/data.ncl.19087082).
Such features have been reported from 2D TEM studies of Toussaint &Dustin, 1963 andCarlson (1989). SBF-SEM imaging enables better visualisation of these features in 3D and our data reveals a characteristic shape of the peg-and-socket abutments of pericyte and endothelial cell membranes as displayed in more detail in Figures 5 and 6 (see also Files S15 and S16: 10.25405/data. ncl.19086197; 10.25405/data.ncl.19086335). The pericyte or endothelial cell membrane was observed to protrude towards, and be enveloped by, the neighbouring cell plasma membrane although pericyte protrusion was more common.
Pericyte-endothelial peg-and-socket formations were noted to occur at several points throughout the data stacks of each capillary analysed. Over the depth of 15 μm in each of the capillaries, pegand-socket formations appeared many times with each formation spanning across several sections. Their distributions are shown in Figure 6. This suggests the nature of pericyte-endothelial interactions commonly occur in peg-and-socket-like formations. Moreover, these heterocellular connections are likely to provide structural anchoring between the two cell types. As such, these structures may be important for maintaining the structural integrity of the inner blood-retinal-barrier. More generally, the pericytes and endothelium lying on either side of a shared BM, and encircling the capillary lumen along the z-axis, support a role in the integrated regulation of vascular diameter and thereby retinal blood flow and nutrient exchange (Geevarghese & Herman, 2014).
It is of interest that pericyte-endothelial cell relationships can take on several formations depending on the tissue/organ served by the capillary, for example, close alignment/contact of pericytes and endothelial cells in placenta, pericyte pegs approaching endothelial cells in the kidney and peg-and-socket arrangements in the brain, in addition to the retina (Harris et al., 2021;Ornelas et al., 2021;Stefanska et al., 2016).

| Macroglial ensheathment of retinal capillaries
Previous TEM studies have suggested that the macroglia, especially Like the BM separating the pericytes from the endothelium, gaps were also present in the BM between the pericytes and macroglia.
When such gaps occur at the outer edge of the BM, it allowed macroglia to come into direct contact with the underlying pericytes.
Examples of these macroglia-pericyte direct contacts are displayed in Figure 8 and Files S18 and S19: 10.25405/data.ncl.19087088; 10.25405/data.ncl.19087103). The closeness of the macroglia to the outer vascular BM and pericytes supports previous studies suggesting that macroglia may directly regulate retinal blood flow through paracrine signalling (Newman, 2015).

| Neuronal cell relationship to macroglia and the vasculature
Although macroglia predominantly enveloped the capillaries (Figure 9), on occasions some areas of the outer capillary BM were absent from such coverage, enabling the neurons to come into contact with the BM. Examples of this are shown in Figure 10 and Files S20 and S21 (10.25405/data.ncl.19086704;10.25405/data. ncl.19087112). This supports a previous TEM-based study of retinal capillaries in tree shrews (Ochs et al., 2000).

| Quantitative assessment of the intercellular and cellular-BM proximities
The proximity analysis MATLAB script (File S8: 10.25405/data. ncl.19087070) enabled the determination of the minimal distances between the boundaries of the segmented features of interest.
Representative results obtained from capillary 1 are shown in Figure 11. Figure 11a shows the stack mean population histogram of the percentage of pericyte boundary pixels to those of the endothelium as a function of distance. Figure 11b shows the cumulative distribution function derived from Figure 11a, enabling the total percentage of pericyte boundary pixels that lie within a given distance of the endothelium to be determined. Identical analyses were performed for other intercellular and cellular-BM feature pairs, with a summary table of cumulative distribution function readings at 10, 100, 250, 500 and 1000 nm presented in File S23 (10.25405/ data.ncl.19087133). As the pixel aspect ratio for our SBF-SEM images is 6 nm × 6 nm, distances lying within 10 nm are essentially in direct contact. Figure 11c depicts the percentage of boundary pixels in contact for selected intercellular and cellular-BM feature pairs.
Relevant to our preceding discussions, we discovered that on average, 10.8% of pericyte boundary pixels were in direct contact with F I G U R E 1 2 Column scatter graphs for convexity, solidity and sphericity of NVU cellular components and the BM for capillary 1. All 150 slices were analysed with each dot representing the area-weighted mean obtained for all segmented features within each slice. Panels (a-c) convexities (convex perimeter:perimeter, CP:P), solidities (area:convex area, A:CA) and sphericities (Rmin:Rmax) of NVU component cells and the BM. Further details of calculation can be found in "Methods" section.
the endothelium, whereas only 0.3% were in contact with the macroglia. Consistent with our visual inspection of the images, no direct contacts were observed between the pericyte boundary pixels and those of the neurons.

| Quantitative morphological assessment of NVU component cells and the BM
Another Matlab script was programmed (File S10: 10.25405/data.

DATA AVA I L A B I L I T Y S TAT E M E N T
The raw data supporting this paper are available at: 10.