Multiplexed volumetric CLEM enabled by antibody derivatives provides new insights into the cytology of the mouse cerebellar cortex

Mapping neuronal networks that underlie behavior has become a central focus in neuroscience. While serial section electron microscopy (ssEM) can reveal the fine structure of neuronal networks (connectomics), it does not provide the molecular information that helps identify cell types or their functional properties. Volumetric correlated light and electron microscopy (vCLEM) combines ssEM and volumetric fluorescence microscopy to incorporate molecular labeling into ssEM datasets. We developed an approach that uses small fluorescent single-chain variable fragment (scFv) immuno-probes to perform multiplexed detergent-free immuno-labeling and ssEM on the same samples. We generated eight such fluorescent scFvs that targeted useful markers for brain studies (green fluorescent protein, glial fibrillary acidic protein, calbindin, parvalbumin, voltage-gated potassium channel subfamily A member 2, vesicular glutamate transporter 1, postsynaptic density protein 95, and neuropeptide Y). To test the vCLEM approach, six different fluorescent probes were imaged in a sample of the cortex of a cerebellar lobule (Crus 1), using confocal microscopy with spectral unmixing, followed by ssEM imaging of the same sample. The results show excellent ultrastructure with superimposition of the multiple fluorescence channels. Using this approach we could document a poorly described cell type in the cerebellum, two types of mossy fiber terminals, and the subcellular localization of one type of ion channel. Because scFvs can be derived from existing monoclonal antibodies, hundreds of such probes can be generated to enable molecular overlays for connectomic studies.


Introduction
Mapping neuronal networks that underlie behavior is a central focus in neuroscience. Techniques that reveal the structure and molecular components of neurons and their connections have expanded over the last several decades. Automated high resolution serial section electron microscopy (ssEM) provides detailed structural information about neuronal networks but omits other important features of neuronal ensembles. For example, the molecular markers of cell or synapse types, the localization of specific molecules related to neuronal physiology, or the functional properties of neural circuits cannot be provided by ssEM alone. Therefore, knowledge gained from neuronal networks mapped by ssEM is still limited. There is a need for new technology to incorporate molecular and functional information into ssEM.
Several techniques that are useful for obtaining both structural context and molecular information in the same sample use light and electron microscopy. For light microscopy (LM), methods to colorize cells, such as Brainbow, take advantage of combinatorically expressed fluorescent proteins to enable dense labeling of neurons with multiple, distinct colors 1,2 . However, the resolution of fluorescence imaging is limited by diffraction. This limitation makes it challenging to differentiate among densely packed neurites or to assign the pre-and postsynaptic partners of each synapse. Expansion microscopy (ExM) can overcome this limitation 3 by physically expanding the dimensions of a sample. When combined with fluorescent lipid labeling 4 , ExM alone or combined with super resolution approaches, have the potential to yield both structural and functional information in the same neural sample 5,6 . Whether these all-fluorescence approaches will ultimately have comparable speed and resolution as electron microscopy is still uncertain. At present nanoscale electron microscopy is the gold standard for imaging cellular structure and any approaches that combine it with multi-molecular labeling would add greater interpretive power to this tried-and-true technique. To label specific molecules directly in electron microscopy images antibodies can be applied before either before or after resin embedding. The antibodies are visible by virtue of being tagged with metal beads or enzymes that chemically create electron-dense deposits visible by EM [7][8][9] . Antibodies conjugated to metal beads of different sizes allow identification of up to three different molecular labels 9 . However, these immuno-labeling strategies have several technical challenges. Pre-resin embedding immunolabeling requires membrane permeabilization for antibody access that perforates cell membranes, causing them to appear discontinuous in EM images.
Discontinuous membranes are incompatible with neural circuit tracing. Post-embedding methods (typically after sectioning) suffer from both the loss of antibody binding due to denatured antigen epitopes by the heavy metal staining and the challenges of using aqueous immunoreagents with hydrophobic resin embedded sections.
One way to circumvent the access problem associated with pre-or post-embedding is to avoid antibodies altogether. Molecular labeling via transgenically engineered probes with EM-visible molecular tags like miniSOG 10 or APEX 11 can be used to label one or more cell-types with ssEM 12,13 , the current maximum number of labels being four 13 . This technical limitation in the number of different labels that can be combined in transgenic approaches suggests that more scalable methods would be of value.
One such approach is volumetric correlated light and electron microscopy (vCLEM) in which fluorescence LM and ssEM are performed sequentially on the same sample and subsequently coregistered. Because multiplex labeling (e.g. with many colors) can easily be achieved by fluorescence LM sing spectral unmixing 14,15 , vCLEM can potentially incorporate more molecular and functional information into ssEM datasets than by directly tagging within the electron microscopy images. Moreover, vCLEM performed on animals that express a fluorescent protein or have a calcium indicator can identify certain cell types 16,17 or show the connectivity patterns of functionally identified cells [18][19][20][21][22] . The challenge of vCLEM with using conventional antibodies for immunofluorescence is, as mentioned above, the requirement of permeabilizing detergents for the antibodies to gain access to intracellular sites. These detergents compromise membrane structure 23 . Attempts to circumvent the need for permeabilization include the post-resin embedding techniques used in array tomography in which immunolabeling of partly hydrophilic resin serial sections is followed by modified heavy metal staining (to maintain antigenicity of the tissue) and EM imaging 24 . This CLEM approach is effective, but the modified heavy metal staining compromises ultrastructure and the water-permeant resin is too soft for the ultrathin (30-40nm) sectioning needed for connectomics.
Detergent-based permeabilization is also avoided by use of nanobodies as immunolabels 23 .
Nanobodies are single domain immuno-binders derived from camelid heavy chain only antibodies 25 .
Because nanobodies are 1/10 the size of conventional antibodies, they diffuse into tissue samples without permeabilizing agents like Triton X-100. In the presence of preserved extracellular space (ECS), 5 nanobodies diffused over a distance of ~100 micrometers into brain samples within 48 hours of their application 23 . Nanobodies, however, for brain markers are uncommon and immunizing camelids is timeconsuming to do. ECS preservation also facilitates standard antibody (IgG) diffusion into brain tissue samples without detergent treatment 26,27 . However, fluorescent labeling with IgGs is challenging if one desires multiple color probes in the same sample. The traditional approach of using fluorescent secondary antibodies to amplify the signal requires that each primary antibody originates from a different species. This limits the number of standard antibodies that can be disambiguated in one tissue sample.
Alternatively, the full-length antibody can be tagged directly with fluorophores that are covalently linked to a particular amino acid moiety. However, given the likelihood of many binding sites this approach can hinder the avidity of the probe and cause different antibodies to the same epitope to have different intensities obviating quantitative assays of fluorescence intensity. The goal of this study is to develop a larger number of immuno-probes that diffuse across cell membranes in the absence of detergents and combine them with multiplex imaging techniques like spectral unmixing for both cell wide and subcellular localization of epitopes.
Our CLEM approach is to generate small immuno-probes from full-size IgG monoclonal antibodies. Monoclonal antibodies (mAbs) are homogeneous with respect to their amino acid sequence and the epitope recognized 28 . Creating smaller probes from existing mAbs requires knowledge of a particular mAb's amino acid sequence. These sequences can be obtained by cDNA cloning from the hybridoma that produces the mAb [29][30][31][32] . Alternatively, if the hybridoma cells are not available, the amino acid sequence can be derived from the mAb protein itself 33,34 . With the knowledge of the amino acid sequence of a mAb, it is possible to synthesize a smaller single-chain variable fragment (scFv) that retains the binding specificity of the progenitor mAb. ScFvs are built by recombinantly linking the VH and VL domains of mAbs via a flexible peptide linker (Figure 1 a) 35 . Only if the VH and VL pair correctly (~60% of the time) do they bind to the antigen 30,35 . Because a scFv consists of only VH and VL domains, they are 1/5 the size of conventional antibodies. We surmised that their small size would allow them to diffuse into tissue samples without detergent-based permeabilization, as is the case for nanobodies 23 . Because there exist extensive collections of well-characterized mAbs that selectively label neuronal cell types or signaling molecules 36 , it is possible to develop a large collection of the corresponding scFvs for use in the 6 nervous system. Because scFvs are produced recombinantly 30 , they can be engineered so that different fluorescent dyes can easily be conjugated to them for multiplex imaging.
To test this idea, we developed eight scFvs based on eight well-characterized mAbs and conjugated them with various fluorescent dyes (see Table 1). Each scFv in this panel proved effective as a detergent-free immunofluorescent probe. We then used linear unmixing with confocal microscopy to visualize six different functional molecular markers in the same brain sample and co-registered these labels to ssEM volumes with pristine preservation of the ultrastructure of the same samples. Because the volumetric fluorescent and electron microscopy image data were of excellent quality, we believe this approach holds promise for routine linking of molecular information to connectomic information obtained from the same tissue samples.

Generation of scFv-based immuno-probes and their use in detergent-free immunofluorescence
To test the hypothesis that scFvs could work as immuno-probes for detergent-free immunofluorescence, we first generated an anti-green fluorescent protein (GFP) scFv based on the amino acid sequence of the anti-GFP mouse mAb N86/38 36 . This mAb binds to both GFP and the GFP derivative YFP. The sequence (derived by PCR cloning from hybridoma cells 29 was used to construct an scFv in which the VH and the VL domains were connected via a 3x linker (GGGGS GGGGS GGGGS) (Figure 1 a). A sortase tag 23 was added to the C-terminus of the VL domain of the scFv for dye conjugation, followed by a 6 x His tag for purification. After the scFv was expressed in Expi 293 cells and purified by affinity chromatography, the fluorescent dye 5-TAMRA was linked (see Methods) (Figure 1 a).
We tested this red fluorescent anti-GFP scFv probe with our detergent-free immunofluorescence protocol (see Methods) on cerebral cortex from YFP-H mice 37 (n=3). The images showed colocalization of fluorescence signals from native YFP and red fluorescence from the scFv probe (Figure 1 b; Sup. Figure   1 a). The scFv probe thus retains the binding properties of the parental mAb and penetrates aldehydefixed brain tissue without the need for detergent treatment. Although the fluorescent signals colocalized in cell bodies and thick dendrites, a few thinner neuronal processes, possibly myelinated axons, were not well labeled with the scFv (arrows in Figure 1 b; but see below).

7
To study the depth of penetration, we immunolabeled two 300-μm thick samples from mouse cerebral cortex without detergent treatment (see Methods). After a free-floating incubation of seven days, this scFv penetrated to a depth of ~60 μm into the tissue (Sup. Figure 1 b). This was deeper than the 10μm penetration of a fluorescently labeled anti-GFP polyclonal antibody (pAb) (Sup. Figure 1 b).
Furthermore, we could improve the penetration to >100 μm by use of an ECS preserving perfusion protocol 27 (Sup. Figure 1 c). Fluorescently labeled nanobodies, which are half the size of scFvs, penetrate even deeper in the same time (Sup. Figure 1 b, c).  We then proceeded to produce an scFv probe for an endogenously expressed protein (calbindin, also known at calbindin-D 28k). This probe was based on a well-characterized mouse mAb (L109/57; 36 ).
Calbindin (CB) is a calcium-binding protein expressed by certain neuronal types in various brain regions, such as cerebellar Purkinje cells, neurons in layer 2/3 of the cerebral cortex, dentate gyrus granule cells, as well as a subpopulation of hypothalamic neurons 38  To evaluate the ultrastructure in the anti-CB scFv labeled cerebellum, we performed a routine EM staining protocol (see Methods) on a 120-μm thick cerebellar cortex sample that had been treated for 7-days with the free-floating immunofluorescence protocol both with and without the use of Triton X-100 for membrane permeabilization (see Methods). The ultrastructure of the sample without Triton X-100 was well-preserved, with continuous membranes and clearly visible synaptic vesicles (Figure 1 e, the first and second panels). The samples with Triton X-100, in contrast, generated EM images in which membrane structures were discontinuous and synaptic vesicles no longer identifiable (Figure 1 e, the third and fourth panels).

Generation of fluorescent scFv probes targeting various targets
To achieve multiplex labeling with scFvs, we included six additional scFvs (anti-vesicular glutamate transporter 1 (VGluT1), anti-glial fibrillary acidic protein (GFAP), anti-voltage-gated potassium channel subfamily A member 2 (Kv1.2), anti-parvalbumin (PV), anti-postsynaptic density protein 95 (PSD-95), and anti-neuropeptide Y (NPY)). These scFvs are based on six well-characterized mouse mAbs 36 (see Table 1). We conjugated them with various fluorescent dyes via the sortase reaction to create fluorescent immuno-probes (see Table 1). We then validated them by comparing scFv versus parental mAb immunofluorescence for each probe (Sup. Figures 2-4). Probes for CB, VGluT1, GFAP, Kv1.2, and PV generated detergent-free immunofluorescence patterns that were similar to or in some cases stronger showed results similar to those found with full size mAbs (Sup. Figure 4), but only when labeling was done on tissue samples where the ECS was preserved. spectral unmixing can be accomplished more efficiently by extracting reference spectra from the multilabeled sample 48 , but these approaches were unsatisfactory with this sample (Sup. Figure 6 b, c).
The spectrally unmixed labeling pattern for each probe was distinct and resembled what we found in the individually labeled samples ( Figure 2 c; Sup. Figure 6 a; Sup. Figure 7). In our sample cell nuclei stained with Hoechst dye) was acquired without spectral unmixing using 405-nm laser excitation. The short excitation and emission wavelength of Hoechst scattered strongly and lost intensity dramatically with depth requiring a 10-fold increase in laser power for the deepest parts of the volume but we found in a separate identically prepared sample, Hoechst could be combined with the other linear unmixing channels with equally good results (Sup. Figure 8). Overall, we used linear unmixing to acquire a multi- The sample was then cut into ~4000 serial 30 nm ultrathin sections using an automated tapecollecting ultramicrotome 49   a, The high-resolution ssEM volume acquired from the cerebellar lobule, Crus 1 with multi-color immunofluorescence from scFv probes separated by linear unmixing. The multi-color fluorescence data was co-registered with the high-resolution ssEM data. Numbers 1-4 indicate approximate regions where the ultrastructure was examined at high resolution. Owing to the absence of detergent in immunofluorescence labeling, fine ultrastructure was preserved throughout the ssEM volume, such as in the molecular layer (1), in the Purkinje cell layer (2) In order to facilitate visualization of the vCLEM dataset, we imported it into Neuroglancer 52 where each fluorescence channel and the EM channel can be visualized separately or simultaneously, and navigation through slices and different resolution levels can be carried out easily.

3D reconstruction of cells and subcellular structures identified by scFv labeling in the cerebellar cortex
An important question is whether the EM data, with six-color immunolabeling superimposed, is of sufficient quality to be successfully segmented by automatic means. Two different methods of automatic segmentation were successful. We used a flood-filling network 53   GFAP: This protein is expressed by astrocytes and Bergmann glial cells whose cell bodies reside in the Purkinje layer and whose processes (also known as Bergmann fibers) extend into the molecular layer 60,61 . As expected, labeling seen with the anti-GFAP scFv (red fluorescence signal in Figure 4   MLI a, the interneuron that is farthest from the Purkinje cell layer axon branched extensively in the volume and innervated a Purkinje cell's dendritic shaft at the arrowhead. This same axon was also part of the pinceau structure that surrounds a different Purkinje cell's axon initial segment (asterisk, also see Sup.  Figure 16 j, insets 1, 2). At least sometimes the glomeruli were associated with mossy fiber collaterals that reached the Purkinje cell layer or slightly above (Sup. Figure 16 j, insets 3, 4). This is expected based on previous descriptions of mossy fibers 59 . Surprisingly, the MGC dendrites were also innervated by parallel fibers ( Figure 5 h to j).
This has not been reported previously and is never the case for typical granule cells.

22
VGluT1: VGluT1 is a member of the vesicular glutamate transporter family that is expressed by axonal terminals 70 . In the cerebellar cortex, VgluT1 is highly expressed in the boutons of parallel fibers and in a subgroup of mossy fiber arborizations 70,71 . In Crus 1, both in a previous study 71 Table 4 and 5 for the volumes of each terminal). We also performed automatic detection of synaptic vesicles (Sup. Figure   18 c, e) and mitochondria in these terminals (Sup. Figure 18 d, f) using machine learning algorithms (see Sup. Table 4

Discussion
Here we report a technique that visualizes multiple molecular labels superimposed on an electron microscopic volume, using detergent-free scFv-based immunofluorescence. We used this method to study a volume of cerebellum labeled with five scFv probes, acquired by serial section electron microscopy. This vCLEM dataset provided several new insights relevant for the cytology of cerebellar cortex. We deposited the expression plasmids of eight scFv probes in an open access repository 72 (see Table 1 for the links at Addgene). All of these probes possess a sortase tag to allow straightforward attachment of fluorescent or other labels for use in vCLEM studies or other applications.
Expanding this technique to more probes will require acquiring a larger collection of functional scFv probes to include additional commonly used molecular markers in bioscience. The scFv probes we generated were all based on mAbs from NeuroMab 36 . This facility has developed thousands of mAbs for neuroscience research, a subset of which have been converted into recombinant form 29 . Consistent with the overall track record of converting mAbs into scFvs 73 , current efforts to systematically convert NeuroMab mAbs into scFvs that retain the binding characteristics of the progenitor mAb is ~ 50%. All of the scFvs described here label brain tissue. To generate more scFvs, several strategies present themselves. First, de novo amino acid sequencing of a mAb can establish the sequences of the VH and VL domains of other commonly used mAbs or even individual components of polyclonal Abs for subsequent conversion into scFvs 74,75 . Second, artificial protein evolution using phage-display methods has been used to modify the structure of failed scFvs to render them functional 76 . Third, rational protein design via point-mutation or CDR-grafting has been used to improve the performance of failed scFvs 77,78 .
Finally, purely in vitro approaches using phage display or other methods can be used to screen naïve or immune scFv libraries to develop novel scFvs 79 . We therefore anticipate that many more functional scFv probes will become available, in particular through open access sources.
Multiplexability, i.e., multiple labels with different colors, can make vCLEM more cost-effective and more powerful. More resolvable color labels in a specimen means that more molecular information With the rapid progress in ssEM, very large datasets (~1 mm 3 ) 54,83 are being generated.
However, the generation of vCLEM datasets of such sizes remains technically problematic. However, the use of scFv labels, once associated with electron microscopy ultrastructure, may circumvent this problem.
For example, the cerebellar cortex dataset showed that despite their similar location, the cell bodies of Bergmann glia, MLIs and MGCs were distinct ( Figure 4 l; Figure 5 a, b, f, g) and could conceivably be used to train a classifier for determination of cell type. Indeed, using machine learning we could differentiate two molecularly defined mossy terminal types, despite them having the same neurotransmitter and being functionally and gross-anatomically similar (Figure 7). More recent machine learning algorithms like "SegCLR" 84 use data from small volumes and then identified cell types in large 26 volumes. We believe that the growing number of probes for vCLEM, combined with machine learning, will allow the classification of cellular and subcellular molecular types in unlabeled electron microscopy images.

Animals
Animals used in the study were adult C58BL/6J mice (Jackson Laboratory). All experiments using animals were conducted according to US National Institutes of Health guidelines and approved by the Committee on Animal Care at Harvard University.

ScFv production
The The sortase reaction was performed as described in 85 . In brief, GGGC-dye conjugates (500 uM) and sortase tag-containing scFvs (100 uM) were mixed in the sortase reaction buffer (50 mM Tris-HCl, 150 mM NaCl, 10 mM CaCl2, pH = 7.5~8.0), followed by adding sortase A pentamutant that has a 6 x His tag (2.5 uM). The mixture was incubated with shaking at 500 rpm at 12 °C for 3 h. Then Ni-NTA agarose was added into the mixture, followed by an incubation shaking at 500 rpm at RT for 20 min to remove sortase and any leftover GGGC-dye conjugates. After the incubation, the mixture was loaded into a microcentrifuge spin column (USA Scientific) and centrifuged briefly to remove the Ni-NTA agarose. The filtered solution that contains the scFv-dye conjugates was loaded into Amicon® Ultra-4 centrifugal filter unit (Millipore Sigma), washed with Tris buffer and then buffer exchanged into Tris buffer with 15% glycerol by centrifugation at 4000 rpm at 4 °C. The concentration of the purified scFv-dye conjugates was determined by A280. The purified scFv-dye conjugates were stored at -20 °C.

Perfusion and fixation
The mouse was anesthetized by isoflurane until there was no toe-pinch reflex. Mice were then For ECS-preserving perfusion, the detailed protocol was described in 27 . In brief, mice were anesthetized by isoflurane, transcardially perfused with aCSF at the flow rate of 10 ml/min for 2 min to remove blood, followed with 15 w/v% mannitol (Sigma-Aldrich) aCSF solution for 1 min, 4 w/v% mannitol aCSF solution for 5 min, and 4 w/v% mannitol, 4% paraformaldehyde in 1 x PBS for 5 min. Brains were dissected out and then post-fixed in the same fixative for 3 h on a rotator at 4 °C. Brains were sectioned into 50-µm coronal sections using a Leica VT1000 S vibratome, and then store in in 1 x PBS at 4 °C.

Immunofluorescence
Detergent-free immunofluorescence labeling was performed with scFv probes or nanobody probes (see Sup. Table 1 for the scFv and nanobody probes used in this study). The detailed protocol was described in 23 . In brief, 50-µm or 120-µm coronal sections were transferred into 3.  Figure 18) (see Sup. Table 2 for the primary antibodies and Sup. Table 3 for the secondary antibodies used in this study). The detailed protocol is available on the NeuroMab website 41  For spectral imaging, the detailed protocol was described in 48  caused the Hoechst signals to be unmixed into the channel of Alexa 488 in error, but we did find one spectrum that generated good unmixing results (Sup. Figure 5).
The brightness, contrast, and gamma of all fluorescent images were adjusted. Fluorescence image volumes were projected to a single plane by maximum intensity for visualization in 2D.

ScFv, nanobody, and antibody penetration experiment
Immunofluorescence with or without detergent by scFv probes, nanobody probes, primary antibodies with dyes directly conjugated, or with primary antibodies plus secondary antibodies was performed on 300-µm coronal sections from animals perfused with normal or ECS-preserving protocol.
See above for detailed protocol of immunofluorescence. For the sections labeled with scFv probes, nanobody probes, primary antibodies with dyes directly conjugated, the incubation lasted for 7 days. For the sections labeled with the primary mAb for calbindin, the sections were incubated with the primary antibody solution (see Sup. Table 2 for the dilution ratio of this mAb) for 4 days on a rotator at 4 °C. After the incubation, sections were washed with vehicle for 3 x 10 min, and then incubated with the secondary antibody solution (see Sup. Table 3 for the dilution ratio of the secondary antibody) for 3 days on a rotator at 4 °C. After the incubation, sections were washed with 1 x PBS for 3 x 10 min. Then the sections were sliced into 50-µm sections lengthwise using a VT1000 S vibratome. The section from the middle was mounted onto glass slides, and then imaged by confocal microscopy to evaluate the penetration depth.

EM preparation
After 120-µm sections were imaged, a small amount of 1 x PBS was added between the glass and coverslip to unseal the coverslip. Sections were transferred into 3.7 ml shell vials with 1 ml secondary fixative (2% PFA, 2,5% glutaraldehyde in 0.15M sodium cacodylate buffer with 4mM Ca 2+  Reconstructor software was used to convert the reconstructed files to .tiff files.

EM imaging
The resin-embedded sections were cut into 30 nm serial ultrathin sections using automated tapecollecting ultramicrotome (ATUM) 86 . Serial sections were collected onto carbon-coated and plasmatreated Kapton tape. The tape was cut into strips and affixed onto 150 mm silicon wafers (University Wafer).
A Zeiss Sigma scanning electron microscope was used to acquire overview images from the serial sections. Two overview images were taken per wafer, which were around 1.35 µm apart in z-axis.
Typical imaging conditions are 8-kV landing energy, 1.2-nA beam current, 3-µs dwell time, 150-nm pixel size, and 4k x 4k images. The images were captured using a below-the-lens backscatter detector and Zeiss' Atlas 5 software. The overview images were aligned using the "Linear stack alignment with SIFT" plugin in FIJI.
Prior to acquiring high-resolution images, the serial section sections on wafers were post-stained for 4 min with a 3% lead citrate solution. After staining, the sections were degassed for a minimum of 24 h at 1x10-6 Torr. A Zeiss MultiSEM 505 scanning electron microscope equipped with 61 electron beams was used to acquire high-resolution images from the serial sections. Images were collected using a 1.5-kV landing energy, 4-nm image pixel, and a 400-ns dwell time.

High-resolution EM image processing
The preparation of the ssEM data before it could be segmented/analyzed includes two steps: The aligned stack was rendered at full resolution (4 x 4 x 30 nm) and each section was cut into 4k

Co-registration of fluorescence and EM volumes
The .vsv files were also ingested into Neuroglancer and resampled in the z-axis to be overlaid with the ssEM volume at a pixel size of 8 nm in the x and y plane and 30 nm in the z-axis.

Automatic segmentation
The ssEM dataset was segmented in 3D using Flood-Filling Networks (FFNs) 53 . The FFN segmentation model was trained at 32 x 32 x 30 or 16 x 16 x 30 nm resolution on the H01 dataset 54 , and run here on CLAHE intensity normalized data 87 downsampled to match the trained model resolution. The resulting base supervoxels were assembled into larger per-cell segments via manual proofreading.
The ssEM dataset was also segmented in 2D at 8 x 8 x 30 nm resolution by a method developed in our lab. A description of this approach was given in 55,88 . In brief, a pre-trained algorithm was used to generate ground truth tiles of membrane predictions, which was corrected by a human annotator. Three rounds of ground truth correction were used to iteratively train deep neural network 56 with a UNET architecture 89 . Neural network predictions were done on a commodity GPU using MATLAB connected to a VAST 90 instance, serving the EM images. In each round the model was applied on tiles randomly 34 selected from the entire EM space and corrections were made to tiles that contained errors. Producing further training rounds halted when a sufficient accuracy was met. Finally, 2D segmentation using a region-growing algorithm was applied on the entire space according to the local minima of the membrane predictions.

Mitochondria detection
We implemented an anisotropic U-Net architecture, incorporating a combination of 2D and 3D convolutions, to predict binary mitochondrial masks and instance maps similar to the U3D-BC approach in

Vesicle segmentation and visualization
We first saturated segmented all vesicles on six images from a terminal, totaling around 8,400 instances. Then, we fine-tuned the Cytoplasm 2.0 model from Cellpose 93,94 on these annotations.
Specifically, as the model was pretrained on images with cells of a bigger size, we scaled up our image and segmentation maps of vesicles by 2x in both XY dimensions. For the 20 volumes of terminals, we obtained roughly 5 million vesicle instances.

Statistical analysis
After the brightness of contrast of the layer of the fluorescent signals of VGluT1 were fixed at a value in Neuroglancer, a human annotator randomly picked 10 locations of VGluT1 positive mossy fiber terminals and 10 locations of VGluT1 negative mossy fiber terminals on slice 512 of the ssEM dataset.
After 3D reconstruction of these terminals, the volume size of each terminal was generated using VAST.
Mitochondria and synaptic detections (see above) were performed on segmentation of each terminal.
None of these approaches were done blind to whether a terminal is VGluT1 positive or negative. Two-35 tailed, unpaired t-tests on volume, synaptic vesicle number, synaptic density, and mitochondria volume per terminal were performed in Prism-GraphPad.

Acknowledgement
We