1. Cell Biology
  2. Neuroscience
Download icon

Reconstruction of genetically identified neurons imaged by serial-section electron microscopy

  1. Maximilian Joesch
  2. David Mankus
  3. Masahito Yamagata
  4. Ali Shahbazi
  5. Richard Schalek
  6. Adi Suissa-Peleg
  7. Markus Meister
  8. Jeff W Lichtman
  9. Walter J Scheirer
  10. Joshua R Sanes Is a corresponding author
  1. Harvard University, United States
  2. University of Notre Dame, United States
  3. California Institute of Technology, United States
Tools and Resources
Cited
7
Views
4,158
Comments
0
Cite as: eLife 2016;5:e15015 doi: 10.7554/eLife.15015

Abstract

Resolving patterns of synaptic connectivity in neural circuits currently requires serial section electron microscopy. However, complete circuit reconstruction is prohibitively slow and may not be necessary for many purposes such as comparing neuronal structure and connectivity among multiple animals. Here, we present an alternative strategy, targeted reconstruction of specific neuronal types. We used viral vectors to deliver peroxidase derivatives, which catalyze production of an electron-dense tracer, to genetically identify neurons, and developed a protocol that enhances the electron-density of the labeled cells while retaining the quality of the ultrastructure. The high contrast of the marked neurons enabled two innovations that speed data acquisition: targeted high-resolution reimaging of regions selected from rapidly-acquired lower resolution reconstruction, and an unsupervised segmentation algorithm. This pipeline reduces imaging and reconstruction times by two orders of magnitude, facilitating directed inquiry of circuit motifs.

https://doi.org/10.7554/eLife.15015.001

eLife digest

Neurons connect with each other to form complex circuits that underlie mental activities. Mapping these connections to obtain a so-called wiring diagram is an essential step in learning how the brain works. The only way to do this precisely enough is by using electron microscopy. However, this technique is so time-consuming that thousands of hours of work are typically required to image even the smallest of tissue samples.

Electron microscopes fire beams of electrons at tissue samples, and detect the scattering of the electrons. Stains are used to make specific neurons less permeable to electrons, or more “electron dense”. Labeled cells scatter more electrons, which increases the contrast of the images. In an approach called serial-section electron microscopy, a tissue sample is first cut into extremely thin sections. These are imaged individually, and the images are then pieced together to reconstruct the sample.

Joesch et al. have now developed a new procedure – named ARTEMIS – that uses a combination of multiple techniques to speed up the mapping of neurons and their connections. ARTEMIS makes use of genetic engineering, serial-scanning electron microscopy, an enhanced chemical staining procedure and a new image processing approach. First, gene technology is used to selectively stain specific types of neurons in mice and flies. Then, a tissue sample is collected and treated with a chemical that enhances the electron density of the stained neurons, without disrupting the tissue’s structure. Next, a labeled target neuron is imaged at relatively low resolution to reveal its overall structure. Small areas of that neuron are then re-imaged at higher resolution to map the connections between neurons. Lastly, an algorithm pieces together the individual images to produce a reconstruction of the cell.

This pipeline of steps reduces the time required to map the shapes and connectivity of neurons with electron microscopy by some two orders of magnitude. This should enable neuroscientists to obtain more rapid insights into the roles of specific neural circuits in the brains of healthy animals, and to identify cases where this wiring goes awry and leads to disease.

https://doi.org/10.7554/eLife.15015.002

Introduction

Relating neural circuits to the computations they perform requires mapping patterns of structural and functional connectivity among neurons. Innovative light microscopic methods such as GRASP, trans-synaptic viral tracing and super-resolution imaging enable visualization of synapses made on or by identified neurons (Wickersham and Feinberg, 2012; Tønnesen and Nägerl, 2013). At present, however, only electron microscopy (EM) provides sufficient resolution to visualize the complete complement of synapses that neurons form and receive. Indeed, large-scale reconstructions from serial sections have provided deep insights into neuronal circuit principles (Briggman et al., 2011; Bock et al., 2011; Takemura et al., 2013; Morgan et al., 2016) . The optimal strategy is to collect serial sections containing all circuit elements and image them at nanometer resolution. Processes in the imaged volumes are then segmented to reconstruct the neurons (or parts of neurons) they contain. Advances in sectioning, imaging and segmentation methods make so-called 'saturated' reconstruction of volumes around 1000 μm3 feasible (Kasthuri et al., 2015). Even these modest volumes remain challenging, however, and when multiple samples must be compared -e.g., controls vs. mutant or treated vs. untreated animals - this approach is currently out of reach (Plaza et al., 2014).

An attractive alternative is 'sparse' reconstruction of specific cells within a fully imaged volume. For example, neuronal activity can be monitored using calcium indicators, then neurons with particular patterns of activity can be relocated in thin sections and reconstructed (Briggman et al., 2011; Bock et al., 2011). This method is, however, technically demanding and infeasible in many tissues. We therefore devised an alternative approach to sparse reconstruction that relies on marking specific cells with an electron-dense tracer. We then exploit the high contrast provided by the tracer to speed up imaging and reconstruction, which are currently the rate-limiting steps in connectomic analysis. Our pipeline includes the following series of steps: (a) tagging a specific cell type with a genetically encoded EM tracer, (b) enhancing the electron-density of the stain without compromising ultrastructure of the surrounding tissue, (c) imaging the cell rapidly at relatively low resolution, (d) re-imaging selected small volumes at higher resolution to map connectivity and (e) segmenting the cell using a novel algorithm that is reliable, fast and does not require computationally intense pre-training. Together, the gains from targeted reimaging and unsupervised segmentation decrease the time required for reconstruction by over two orders of magnitude. We call the method ARTEMIS for Assisted Reconstruction Technique for Electron Microscopic Interrogation of Structure.

Results and discussion

A classical ultrastructural tracer is horseradish peroxidase (HRP), which catalyzes the formation of a 3,3’-diaminobenzidine (DAB) polymer; the polymer binds osmium and is thereby rendered electron-dense. Recombinant HRP is enzymatically inactive in the cytosol because it fails to form disulfide bonds or become glycosylated, but this limitation can be overcome by directing the protein to topologically extracellular compartments such as vesicles (Li et al., 2010; Atasoy et al., 2014; Schikorski et al., 2007). We therefore generated an HRP variant that was codon-optimized, mutated to increase activity, and fused to an endoplasmic reticulum targeting sequence (erHRP). We also tested derivatives of plant ascorbate peroxidases called APX and APEX2, which are active in the cytosol (Martell et al., 2012; Lam et al., 2015). Initial studies using cultured HEK293 cells confirmed that all three constructs generated active peroxidase in the transfected cells (Figure 1a,b and data not shown).

Figure 1 with 5 supplements see all
Enhanced staining of genetically encoded tags for EM.

(ab) Bright-field images of HEK-cells rendered photon-dense by DAB polymerization, catalyzed either by APEX2 tagged with a nuclear export signal (APEX2NES) (a) or endoplasmic reticulum tagged HRP (erHRP) (b). (cd) Bright-field images of direction-selective retinal ganglion cells (ooDSGC) expressing APEX2NES (c) or erHRP (d) and rendered photon-dense as in (a). (e) Enhanced staining strategy. (fg) EM-micrographs of HEK-cells rendered electron-dense after DAB-polymerization as in (a). A standard EM staining protocol showed no detectable cytosolic DAB-polymer staining (f), whereas the addition of a reduction step dramatically enhanced DAB-polymer staining (g). (h) Ribbon synapses in the outer plexiform layer; ultrastructure is well preserved after reduction. (i) EM-micrograph of an ooDSGC soma (arrowhead) rendered electron dense after tissue reduction next to a non-expressing RGC (asterisk). (j) Close-up of cytosolic APEX2 staining. (k) Dendritic processes expressing APEX2 (asterisk) contacted by a presynaptic partner (arrowhead) in the inner plexiform layer. (l) Axonal long-range projections of an APEX2-expressing ooDSGC in the superior colliculus (asterisk) with a postsynaptic target (arrowhead). (m) erHRP-expressing (arrowhead) RGC next to a non-expressing RGC (asterisk). (n,o) Close-up of erHRP staining of a J-RGC soma (n) and dendrite (o) (arrowhead point to presynaptic partners). (c,d,i-l) are from Cart-cre mice; (mo) from Jam-B-creER mice. Scale bars: (ad) :25 μm; (fg): 10 μm; I, m : 5 μm; h, jl, n,: 500 nm;

https://doi.org/10.7554/eLife.15015.003

For selective expression in molecularly-identified cells, we generated adeno-associated viral (AAV) vectors in which expression of erHRP, APX or APEX2 required Cre-dependent recombination. These were used to infect retinas of transgenic mice in which specific retinal ganglion cell (RGC) types expressed Cre recombinase (ooDSGCs in Cart-cre [Kay et al., 2011]) or tamoxifen-activated Cre (J-RGCs in JAM-B-creER [Kim et al., 2008Joesch and Meister, 2016]). Two to four weeks after infection, retinas were reacted with DAB and H2O2 and examined light-microscopically, revealing intense labeling of RGCs (Figure 1c,d). Retinas were then processed for EM (see Methods). Unexpectedly, levels of electron-dense precipitate were so low that stained processes could not be traced reliably (Figure 1—figure supplement 1a,b). Numerous alterations to balance peroxidase activity and ultrastructural quality failed to improve matters: when ultrastructure was adequately preserved, staining for peroxidase was poor, and when reaction product was adequate, synaptic structures were poorly preserved. This difficulty may have been less apparent in previous studies using injected native HRP, which has substantially higher activity than the recombinant proteins we use.

The reason for the difference between light and electron microscopic results is likely that the opacity and electron density of the DAB polymer arise in different ways: its polycyclic structure renders it photon absorbent, but its electron-density results from redox reactions with osmium (Bahr, 1954). We reasoned that radicals produced during the long peroxidase reaction might oxidize the relevant functional groups in the polymer, leaving it photon-absorbent but inert to osmium. If this were true, reduction of functional groups on the polymer could restore reactivity to osmium (Figure 1e). We tested this hypothesis in transfected HEK cells. When cells were treated with the protocol we had used for retina, the precipitate was clearly visible by light but not electron microscopy. However, when the HEK cells were treated with a mild reducing agent (5 mM sodium hydrosulfite) between the peroxidase reaction and osmication, they were highly electron-dense (Figure 1f,g). This was the case using either conventional osmium staining or an enhanced 'double osmium' staining protocol (rOTO), although the latter showed a slight improvement probably due to the change in the redox state of osmium tetroxide (Figure 1—figure supplement 1c–f; see Methods). A similar effect was observed in erythrocytes, in which endogenous heme catalyzes the DAB reaction (Figure 1—figure supplement 1g–j).

When the reduction protocol was applied to retinas, we were able to visualize RGCs that had been tagged with APX, APEX2NES (APEX2 fused to a nuclear export signal) or erHRP (Figure 1i–o, Figure 1—figure supplement 2 and 3). As expected, APX and APEX2NES labeled the cytoplasm diffusely (Figure 1j) while erHRP labeled membrane-bound intracellular compartments (Figure 1n). Thus they could be used as orthogonal labels, although we have not yet pursued this application. The strength of the signal allowed us to identify small stained dendritic profiles (Figure 1k,o) and to view the terminals of RGCs in the superior colliculus, approximately 1 cm from the somata (Figure 1l). Remarkably, the reduction protocol actually improved the visualization of the ultrastructure irrespective of peroxidase expression or DAB treatment (Figure 1h). Part of the improvement resulted from an increase in the reactivity of membranes to osmium, thereby enhancing membrane-cytoplasm contrast (Figure 1—figure supplement 4). This improvement is also visible in the staining strength of synaptic densities (Figure 1—figure supplement 4a,b). When we asked blind-to-condition observers to judge the quality of synapses between both conditions, the reduced tissue was selected ~3 times more frequently than unreduced tissue (Figure 1—figure supplement 4e). Thus, rather than sacrificing ultrastructure for reactivity or vice versa, this protocol improved both.

For the approach to be useful, it is essential that peroxidase expression does not affect synapse formation and that synaptic partners of peroxidase expressing cells can be identified. When analyzing our datasets, we could not detect any structural differences between tissues expressing either of the peroxidases. The number of synapses counted in peroxidase expressing and control sections was also similar, 0.32 and 0.34 synapses / μm2, respectively. To test whether the electron dense precipitate hinders reliable synaptic identification, we characterized the synaptic connections received and made by APEX2 or erHRP expressing retinal ganglion cells (Figure 1k,o; Figure 1— figure supplement 5). Cells presynaptic to APEX2 or erHRP could be clearly identified based on ultrastructural details (Figure 1—figure supplement 5a,c). Postsynaptic partners of APEX2-expressing cells could also be identified. Moreover, synaptic vesicles of APEX2-expressing neurons could be detected in presynaptic terminals because they were unstained, and thereby contrasted with the electron-dense cytosol. However, the endoplasmic reticulum did not regularly extend to axonal terminals, making it difficult to identify postsynaptic partners of erHRP expressing ganglion cells.

To test the generality of the method we expressed APEX fused to GFP in direction-selective tangential cells of Drosophila melanogaster (Joesch et al., 2008) (DB331-Gal4 → UAS-APEX-GFP; see Methods). This driver line expresses mainly in 6 vertically sensitive and 3 horizontally sensitive tangential cells. This expression pattern was apparent using either GFP fluorescence or the polymerized DAB to mark the cells (Figure 2a,b). We were also able to visualize the electron dense precipitate of these processes in the fly’s optic lobes (Figure 2c,d). Although our method was optimized for mammalian tissue, ultrastructural detail was reasonable (Figure 2e) and allowed the identification of synaptic contacts made by and onto APEX expressing cells (Figure 2f–h).

APEX expressing interneurons in Drosophila melanogaster.

(a) Fluorescent image of a brain of Drosophila melanogaster (DB331-Gal4 → UAS-APEX-GFP) expressing APEX in direction-selective lobula plate tangential cells (LPTCs). (b) Bright field image of a similar plane from another (DB331-Gal4 → UAS-APEX-GFP) fly, labeled with DAB. (c) Electron micrograph of a frontal brain section (30 nm thickness) containing electron dense staining in axonal processes of LPTCs (arrowhead and framed box). (d) Enlarged view of the framed region in (c). Arrows point to electron dense processes. (e) Unstained synaptic terminal, showing quality of ultrastructural detail. (f) Synaptic nerve terminal in an APEX-positive process, identifiable by contrast-reversed vesicles. (g,h) APEX-positive postsynaptic processes, identifiable by the presence of adjacent vesicle-laden, T-bar-containing terminals. Scale bars: ac :100 μm; :10 μm; eg : 500 nm.

https://doi.org/10.7554/eLife.15015.009

To implement sparse reconstruction, we used the ATUM (Hayworth et al., 2014) (automated tape-collecting ultra-microtome) to serially section retinas containing either APEX2NES-expressing retinal ganglion cells (J-RGCs) or retinal interneurons (starburst amacrine cells (SACs)) labeled using Choline Acetyltransferase-Cre (Rossi et al., 2011). Because we could identify small peroxidase-stained neurites at overview resolution (20–30 nm per pixel), we tested the minimal requirements to reconstruct both cell types at these resolutions. Rapidly imaging every 10th section (270 nm separation), at 30 nm resolution was sufficient for mapping a J-RGC dendrite (144 sections covering > 1 × 107 µm3, imaged in 22 hr; Figure 3a,b). By comparison, imaging the same volume at high resolution (4 nm/pixel) would have taken ~2500 hr on the same microscope. Manual reconstruction of the J-RGC was straightforward and took 2 worker-hours (Figure 3c). To extend this result to another cell type, we imaged all 1260 sections from a bloc with APEX2NES-positive SACs and reconstructed SAC processes in this volume (Figure 3d–f). Thus, we can reliably find and reconstruct genetically identified cells.

Targeted interrogation of serial-section EM volumes in mouse retina.

(a). Overview-resolution EM volume containing J-RGCs expressing APEX2NES. Every 10th section was imaged at 30 nm/pixel. (b) Two consecutive sections containing APEX2NES expressing J-RGC processes (arrows). (c) Reconstructed J-RGC. Note the reduced complexity of the dendritic field due to the subsampling of the overview dataset. (d) Overview-resolution EM volume that contains SACs expressing APEX2NES. The sections were imaged at 20 nm/pixel. (e) Two consecutive sections showing a SAC-SAC contact (arrowheads). (f) Reconstruction of the SAC-SAC interaction in (e); arrowhead indicates contact shown in e. (gh). Comparison of ultrastructural detail visible at 20 nm/pix (g) or 4 nm/pix (h). Although both images contain ribbon synapses, the low-resolution image (g) lacks the required resolution to identify them. Arrowheads point to ribbons. (i) High-resolution (4 nm/pixel) EM subvolume selected from (d). (j) Two high-resolution EM-micrographs containing SAC-processes expressing APEX2NES receiving synaptic contacts (arrowheads). (k) Reconstruction of the SAC plexus with its characteristic dendritic fasciculation and net-like structure. Asterisk indicates contact and arrowheads the synaptic inputs shown in j (top panel). Scale bars: b: 5 μm; c: 50 μm; e, j: 1 μm; gh: 200 nm.

https://doi.org/10.7554/eLife.15015.010

Although 20 nm resolution is sufficient for viewing neurites, it does not allow optimal visualization of synaptic contacts (Figure 3g–h). However, sections generated using ATUM are collected on wafers, and can be re-imaged; a program, Wafer-Mapper (Hayworth et al., 2014), facilitates returning to the same place on a given section with micron-level precision. This feature allows targeted imaging at high resolution of areas chosen from the rapidly acquired, lower resolution reconstruction, thereby substantially reducing imaging time. To test this multi-scale approach, we focused on the SACs, whose dendrites can be less than 100 nm in diameter. We selected a heavily-labeled ~5 × 105 µm3 volume from the overview-resolution reconstruction of ~2 × 107 µm3 and acquired a high-resolution (4 nm/pixel) data set in 250 hr (Figure 3i). Compared to a completely high-resolution approach, this amounts to a 40-fold reduction in imaging time. More importantly, it expedited reconstruction, the current time-limiting step in connectomics (Plaza et al., 2014), by constraining efforts to defined regions of interest. We manually traced and reconstructed the SAC plexus in this volume, observing multiple SAC-SAC interactions, dendritic fasciculation in the SAC plexus and dendritic branching (Figure 3j,k and data not shown).

Finally, to further expedite the reconstruction pipeline, we developed an algorithm that detects and segments APEX2-positive processes, taking advantage of the high contrast rendered by the DAB polymer. Our algorithm is fast, parameterized and does not rely on the supervised machine learning training regimes currently required for lower contrast material (Kasthuri et al., 2015) (Figure 4a, Materials and methods). We tested the algorithm using 401 sections from the high-resolution dataset (Figure 3i). We hand-segmented a set of 444 APEX2 positive segments, and then compared these results to those obtained computationally. We calculated a recall statistic of 91.8% for the 2D segmentation portion of the algorithm (Figure 4b,c). All missed segments were small (on average 45 × 42 pixels), and were seldom necessary to establish connectivity. Consistent with human inferred segmentation, the analysis of segment adjacency in local neighborhoods can make connections in regions with small gaps. To test the accuracy of the algorithm for longer (>50 μm) processes, we segmented and reconstructed a dendritic branch of a starburst amacrine cell. Our algorithm could reconstruct the main characteristics of the dendritic processes and recapitulated most of the manually segmented details (Figure 4d,e). The primary advantage of this approach is processing time, being two orders of magnitude faster than approaches based on convolutional neural networks for membrane classification (Kasthuri et al., 2015; Kaynig et al., 2015). Training regimes in the neural network approaches require adjusting at least thousands of parameters separately for each new tissue; in contrast, our algorithm incorporates unsupervised components with a small fixed set of tunable parameters (see Methods and Supplementary Code). Importantly, the high contrast-to-noise ratio of the peroxidase-labeled cells is the enabling factor for the efficacy of this algorithm, which means that it could be used for reconstruction of DAB-stained structures in any tissue. Thus, although detailed synaptic properties may be obscured by the electron-dense stain, our methods are well-suited for rapid reconstruction of targeted neuron morphology and connectivity.

Automatic segmentation and reconstruction algorithm for APEX2 positive processes.

(a) The algorithm performs unsupervised 2D segmentation and 3D reconstruction without the need of pre-training. The procedure follows these steps: 1. Clustering-based image thresholding over pixel intensities from known contrast-enhanced ranges, modulated via quality assessment; 2. 2D segmentation based on unsupervised clustering; 3. Calculation of geometric properties for each identified segment, followed by noise and artifact pruning based on these properties; 4. Identification of large segments based on segment properties; 5. Graph-based segment search for reconstruction and labeling in 3D, followed by a final merging procedure to enforce consistency across the volume. (b) Unsupervised automatic segmentation of mouse SAC processes. Top: manual segmentation, ground-truth; bottom: automatic segmentation. (c).Reconstructions of the dendritic process in (b), comparing manual (top) and automatic reconstructions. (d) Manual reconstruction of a starburst amacrine cell process. En face (top) and side view (bottom). (e) Automatic reconstruction of the same process shown in d. En face (top) and side view (bottom). Scale bar in b: 1 μm; d, e: 10 μm.

https://doi.org/10.7554/eLife.15015.011

In summary, we have assembled a set of tools that enables rapid reconstruction of genetically identified neurons, so that their shapes and connections can be mapped at high resolution in much less time than required for conventional imaging and segmentation protocols. By expediting the analysis of neural motifs, ARTEMIS renders the interrogation of diverse samples feasible and holds a clear promise to unravel mechanisms ranging from neuronal development to computations.

Materials and methods

Animals

Animals were used in accordance with NIH guidelines and protocols approved by Institutional Animal Use and Care Committee at Harvard University. JAM-B-creER mice (Kim et al., 2008)were generated in our laboratory. ChaT-cre mice (Rossi et al., 2011), and Cart-cre mice (Madisen et al., 2010) were obtained from Jackson Laboratories. In the Chat-cre line, the Cre recombinase gene was targeted to the endogenous Chat gene; this line expresses Cre in starburst amacrine cells (SACs). In the Cart-cre transgenic line, Cre expression is controlled by regulatory elements from the Cartpt gene. In this line, Cre marks ON-OFF direction selective retinal ganglion cells. Mice were maintained on a C57/BL6J background. Both male and female mice were used in this study. Animals were 60 to 100 days old at the time of euthanasia.

Flies were raised on standard cornmeal-agar medium. The DB331 Gal4-line (Joesch et al., 2008) was kindly provided by Vivek Jayaraman (Janelia Research Campus) and the UAS-APEX-GFP (Chen et al., 2015) line was generated and kindly provided by Chiao-Lin Chen and Norbert Perrimon (Harvard Medical School and HHMI).

Peroxidases

The endoplasmic reticulum-targeted HRP (erHRP) was designed as follows. First, to improve expression in mammals, the nucleotide sequence of HRP (horseradish peroxidase from a plant, Armoracia rusticana) was codon-optimized (DNA2.0, Menlo Park, CA). Second, to regulate protein trafficking a signal secretion sequence from the human immunoglobulin kappa chain (from pDisplay, Invitrogen, Carlsbad, CA) as well as an endoplasmic reticulum (ER)-retention signal (-KDEL) were appended at the N- and C-termini, respectively. Finally, the N175S mutation was introduced to confer heat stability and resistance to H2O2 (Morawski et al., 2001). The sequence of this cDNA is available from GenBank #KU504630. Ascorbate peroxidase (APX, dimeric) from pea (Martell et al., 2012) , and the enhanced monomeric version APEX2 (derived from soybean APEX) (Lam et al., 2015) were codon-optimized for better expression in mammals. In APEX2NES, the nuclear export signal (NES) was appended to APEX2. The erHRP, APX, and APEX2NES plasmids under CMV promoter were transfected to HEK293T cells (ATCC) using the DMRIE-C transfection reagent (Life Technologies). Subsequently, the cDNAs were cloned into a plasmid encoding a Cre-dependent Adeno-associated virus (AAV) backbone with the CAG (CMV-beta actin promoter + beta-globin leader) promoter, woodchuck post-transcriptional element (WPRE), and the FLEX switch (Atasoy et al., 2008) (Supplementary file 1). AAV viruses were generated by transfecting these plasmids together with appropriate helper plasmids, and prepared using a chemical precipitation method (Guo et al., 2012).

Plasmids encoding the viral vectors will be sent to AddgenePlasmids described in this paper are available from Addgene (www.addgene.org).

AAV-mediated gene transfer

For viral-mediated gene transfer, adult Cre-mice were anaesthetized with ketamine/xylazine by intraperitoneal injection. A 30 1/2G needle was used to make a small hole in the temporal eye, below the cornea. 1 μl of vitreous fluid was withdrawn and then 1 μl of rAAV2 or rAAV2/9 Cre-dependent viruses (a titre of ~1 × 1011–12 genome copies per ml) was injected into the subretinal space with a Hamilton syringe and 33G blunt-ended needle. Animals were euthanized and retinas were dissected 4–6 weeks following injection.

Tissue preparation

Mouse retinas were dissected from eyecups in oxygenated Ames’ medium (Sigma) with constant bubbling (95% O2, 5% CO2) at room temperature. Four incisions were made to flat-mount the retina with ganglion cells facing up onto nitrocellulose filter paper. The tissue was drop-fixed, with 2% PFA and 2.5% glutaraldehyde followed by 2.5% glutaraldehyde (times specified in Supplementary file 2) then washed. HEK-cells were fixed with 2% PFA and 2.5% glutaraldehyde for 15 min followed by a 45 min fix in 2.5% glutaraldehyde. Flies were decapitated and dissected in oxygenated Ringer solution. A small incision was made on the back of the head and the posterior cuticle was separated from the head. This ensured that the fixative and staining solutions could penetrate into the brain while the rest of the cuticle protected brain tissue during processing. Flies were fixed with 2% PFA and 2.5% glutaraldehyde for 15 min followed by a 45 min fix in 2.5% glutaraldehyde. Following aldehyde fixation, cells and tissues were washed, reacted with DAB to reveal sites of peroxidase activity, washed again, reduced with 50 mM sodium hydrosulfite and stained with osmium. Osmium treatments included 2% aqueous osmium tetroxide (used only in HEK cell micrographs of Figure 1—figure supplement 1) or the reduced osmium tetroxide-thiocarbohydrazide (TCH)-osmium ('rOTO') (Willingham and Rutherford, 1984; Hua et al., 2015; Tapia et al., 2012) protocol (all other Figures). Due to the two consecutive osmication steps, the 'rOTO' protocol improves the signal of membranes compared to the aqueous osmium. This enables reasonable signal-to-noise ratios at high scanning, an essential requirement for our approach. Finally, the stained tissue was dehydrated and infiltrated with Durcupan resin. Sodium-cacodylate (cat. no. 12300), glutaraldehyde (16220 and 16120), paraformaldehyde (15710), osmium tetraoxide (19190), maleic acid (18150), acetone (glass distilled; 10015) and uranyl acetate (22400) were purchased at Electron Microscopy Sciences (EMS); AMES medium (A1420), 3,3’-diaminobenzidine (DAB; D5905), potassium hexacyanoferrate (II) (P9387), thiocarbohydrazide (88535), sodium hydrosulfite (157953) and durcopan resin (44610) were purchased at Sigma-Aldrich. Concentrations and incubations times, along with details on reagents are provided in Supplementary files 2 and .

Immunohistochemistry

For the image in Figure 2a, the fly brain was fixed in 2% paraformaldehyde in PBS for 30 min on ice, washed with PBS and blocked with 3% goat serum/1% Triton X-100/PBS. For staining, tissue was incubated with 3% goat serum/1% Triton X-100/PBS and rabbit anti-GFP Alexa Fluor 488 conjugate (dilution 1:1000, Invitrogen, A-21311) at 4°C for 1 days and washed with PBS. Brains were mounted on Vectashield mounting medium (Vectorlabs) and imaged in a confocal microscope (Olympus FVA). For the image in Figure 2b, brains were prepared for electron microscopic and imaged before osmication.

Electron microscopy

The cured blocks were trimmed to a 2 × 3 mm rectangle and a depth of 400 µm and then readied for automated serial sectioning. The automated, unattended collection of ~ 30 nm serial sections was accomplished using a custom tape collection device (ATUM) (Hayworth et al., 2014) attached to a commercial ultramicrotome. The sections were collected on plasma-treated carbon-coated polyamide (Kapton, Sheldahl) 8-mm-wide tape. Sections were post-stained with 1% uranyl acetate in maleate buffer for 30 s and with 3% Lead Citrate (Ultrostain II; Leica - cat. no. 16707235) for 30 s. An automated protocol to locate and image sections on the wafers was used (Hayworth et al., 2014) with a Sigma scanning electron microscope (Carl Zeiss), equipped with the ATLAS software (Fibics). Images were acquired using secondary electron detection.

For the medium- and high-resolution data sets, non-affine alignment was accomplished through the FijiBento alignment package (https://github.com/Rhoana/FijiBento) that enables the alignment of large data sets in a relatively short period of time. To this end, the alignment was performed on the Odyssey cluster supported by the FAS Division of Science, Research Computing Group at Harvard University. The aligned images were then manually segmented using a custom volume annotation and segmentation tool (VAST; http://openconnecto.me/Kasthurietal2014/Code/VAST). The segmented images were processed for data analysis and 3D modeling with Matlab scripts, and Persistence of Vision Raytracer (http://www.povray.org/) for rendering steps.

Automatic 2D segmentation and 3D reconstruction

Full resolution EM images (100000 × 50000 pixels) underwent contrast adjustment via histogram equalization to normalize image intensities across slices. For each normalized image, a global threshold τ was computed via clustering-based image analysis (Bankman, 2008), modulated by the known contrast-enhanced pixel ranges of the ARTEMIS markers for a data set (e.g., for Suppl. Figure 4b: if τ < 0.6, τ = τ * 0.82, else τ = τ * 0.86). The normalized images were then converted to a binary representation, and clustering-based image thresholding was applied again, assuming two classes (positive and negative), to gather 2D candidate segments. To remove artifacts such as speckle noise, and holes in the tissue, only segments that satisfy the following constraints were stored: 3 pixels < area < 130,000 pixels and major axis length < 900 pixels. The remaining segments were assigned a unique label, and stored in a MySQL database, along with their coordinates and other geometric properties Based on the expected morphological progression of each process, irregularly large imaging artifacts that appear between adjacent segments in the volume were pruned. Graph-based segment search over the database entries established the connectivity between segments, including those with gaps between them, by finding the minimum distance between centroid points from all pairs of segments in a local neighborhood using a k-Nearest Neighbors-like (Hastie et al., 2009) algorithm (k = 2). Using the graph as a guide, 2D segments were merged into the final 3D reconstruction. We included the source code in the supplementary information, packaged as a Matlab live script, with example images and an animation.

Data

All relevant data is available on request. The electron microscopic data set of Figure 3k, which includes the segmented processes used for the reconstructions, has been deposited to Dyrad. Doi:10.5061/dryad.h67t6.

References

  1. 1
  2. 2
  3. 3
  4. 4
    Handbook of Medical Image Processing and Analysis (Second edition)
    1. I Bankman
    (2008)
    Academic Press.
  5. 5
  6. 6
  7. 7
  8. 8
  9. 9
  10. 10
    The Elements of Statistical Learning (Second edition)
    1. T Hastie
    2. R Tibshirani
    3. J Friedman
    (2009)
    Springer-Verlag.
  11. 11
  12. 12
  13. 13
  14. 14
  15. 15
  16. 16
  17. 17
  18. 18
  19. 19
  20. 20
  21. 21
  22. 22
  23. 23
  24. 24
  25. 25
  26. 26
  27. 27
  28. 28
  29. 29
  30. 30
  31. 31

Decision letter

  1. Kristen M Harris
    Reviewing Editor; The University of Texas at Austin, United States

In the interests of transparency, eLife includes the editorial decision letter and accompanying author responses. A lightly edited version of the letter sent to the authors after peer review is shown, indicating the most substantive concerns; minor comments are not usually included.

Thank you for submitting your article "Reconstruction of genetically identified neurons imaged by serial-section electron microscopy" for consideration by eLife. Your article has been reviewed by two peer reviewers, and the evaluation has been overseen by a Reviewing Editor and Eve Marder as the Senior Editor.

The reviewers have discussed the reviews with one another and the Reviewing Editor has drafted this decision to help you prepare a revised submission.

Summary:

Two complementary and timely advances are described for genetic tagging and reconstruction of subsets of retinal and Drosophila neurons by electron microscopy. Viral vectors are used that required Cre-dependent recombination and express erHRP, APX or APEX2. These were tested in mouse retinas, in which certain retinal ganglion cells expressed Cre-recombinase, or tamoxifen activated Cre, and also in Drosophila that expressed APEX fused to GFP in direction selective tangential cells. Listing the specimen preparation methods that led to the success or failure of APEX and HRP-dependent staining will benefit the field. Likewise, the combined application of rapid low resolution with selected high-resolution analysis will facilitate the time-consuming, but much-needed large volume reconstructions of genetically specified neurons by serial electron microscopy. The paper also describes other aspects of a workflow to reconstruct the labeled cells. These approaches are described as being quicker than other current methods. The entire method has considerable potential and will be of interest to the neuroscience community. Despite this overall favorable assessment of the study, several points should be addressed.

Essential revisions:

1) An important question concerns efficiency. Are all of the genetically specified neurons labeled or just a subset? For example, the Drosophila DB331 driver line should target a bundle of motion-sensitive cells with numerous spatially constrained projections, yet Figure 2 displays only a handful of processes. In both Figure 1 (C-D) and Figure 2, a low-resolution comparison of Cre-dependent fluorophore and DAB staining patterns are needed to address this concern.

2) The paper does a convincing job of showing how 'rapid' low-resolution images and reconstructions can be made of axons and dendrites. Nevertheless, to validate the method's use in circuit analysis, consistent high quality images are needed that show synapses identified by arrows with detailed descriptions in the figure legends and results (a few of the images are of sufficient magnification and quality but synapses should be identifiable (Figures 3h and Figure 1—figure supplements 3f and 4a, b)). Figure 1—figure – supplement 3f is the only image that actually identifies the synapses with arrows, yet still provides too little description.

3) All of the example images reveal suboptimal ultrastructure inside the labelled neurites, relative to other methods that have been used to label synapses in the literature. The dense reaction product obscures pre- and postsynaptic active zones, as well as presynaptic vesicles. This being the case, it is not clear how synapses will be identified on the specifically labeled cells. This issue must be addressed with images, labels, and clear examples of labeled synapses.

4) Both aqueous osmium and reduced osmium (rOTO protocol) were used. However, it is not completely clear which osmium staining was used in some of the figures. In addition, it is not entirely clear whether there were any significant differences between the single osmium or rOTO protocol methods for low magnification reconstructions or synapse identifications. The comparisons in Figure 1—figure supplement 1 need more explanation.

5) It appears that the findings are intended to be qualitative; however, some form of semi-quantitative comparisons are warranted to establish the claims. For example, blind scoring of synapses (such as 0-3) and ultrastructural quality under the different conditions would be welcomed to assess the stated improvements.

6) Similarly, a systematic (and blind as to condition) quantification of a subset of small volumes of synapses from control and reacted tissue is needed to rule out possible effects of peroxidase expression on the ability to identify synapses in either the labelled cells or among the associated or surrounding but unlabeled synapses.

7) The authors indicate that the technique provides improved preservation. Although in Figure 1, panels H to L are of acceptable magnification and resolution, they do not show particular evidence of improved preservation relative to other papers in the literature. Description and discussion are needed to clarify to what the improvement has been made.

8) It is noted that the fixation of the fly brain was with immersion, without dissecting it away from the cuticle. Only an incision was made in the back of the head. Some explanation is needed as to why the authors used this approach, instead of removing the whole cuticle as is standard in the literature. The tissue preservation does not look as good as that in other published images where the whole cuticle was removed, and fixative penetration was facilitated, for example with microwave irradiation.

9) Concerning labelling in the fly brain, the outcomes should be described in more detail in the Results section (Results and Discussion, fifth paragraph). Higher magnification images are needed in Figure 2 to identify synapses, which should be labelled on the images and described in the Figure legend. These higher resolution images could replace one or more of the panels in Figure 2e–g. It is important to show here the synaptic contacts with the labelling on either the pre or post-synaptic side. We recommend that sufficient magnification be shown so that synapses are readily identifiable, with arrows, and more detailed descriptions in the results and legends.

10) The description of the automatic 2D segmentation and 3D reconstruction explains that speckle noise is removed by pruning connected components smaller than 1500 pixels in area. As written, it is not clear how this removal was done. Based on the low resolution that was used for the EM imaging, the size of the removed component seems rather large. Description of how effective this process was when dealing with long thin neurites is needed.

11) A more direct statement is needed in Figure 3 legend, that 20 nm resolution is not sufficient to identify synapses, even in the unlabeled processes. Then in Figure 3h (4 nm resolution), synapses should be explicitly identified with arrows.

12) The serial sections selected for Figure 3j should show a synapse with one of the labelled cells, which is needed to validate the circuit reconstruction.

13) Supplementary files 13 were not included in the PDF for review and need to be submitted with the revision for proper review.

14) In addition, it would be helpful if the software could be prepared for expert review upon revision.

[Editors' note: further revisions were requested prior to acceptance, as described below.]

Thank you for resubmitting your work entitled "Reconstruction of genetically identified neurons imaged by serial-section electron microscopy" for further consideration at eLife. Your revised article has been favorably evaluated by Eve Marder (Senior editor), a Reviewing editor, and two reviewers.

In general, the authors have addressed prior concerns raised by the reviewers. There remain a few outstanding issues described below.

Essential further revisions:

1) Issue of overstatements on 'improved ultrastructure' needing more work:

A) Abstract – remove the phrase "and improves ultrastructural preservation". This method does not improve ultrastructural preservation; it enhances staining. Could substitute "while retaining ultrastructure" or something similar. The purpose of this method is not to improve ultrastructural preservation, and as noted in the prior review, the ultrastructure of the labeled cells is not particularly good with respect to other methods.

B) Introduction – second paragraph – (b) enhancing the electron-density of the stain without compromising ultrastructure. To be clear reword as "without compromising the surrounding ultrastructure" or in a similar way. As written, it sounds like the ultrastructure of the labelled cell would not be compromised, but of course it is compromised as the stain obscures most of the ultrastructure.

C) Consider the following concern raised by one of the reviewers and revise the Introduction and Discussion accordingly:

The narrative does not oversell the technique applicability for ultrastructure analysis. However, since the ability to examine synaptic ultrastructure in genetically designated neurons would represent a significant milestone, it may be more appropriate to characterize explicitly this current incarnation of peroxidase-based methods as most appropriate for rapid reconstruction of targeted neuron morphology and connectivity, rather than synapse properties.

2) Some specific figures still need to be improved:

A) Figure 2: The addition of panels showing complementary GFP and DAB staining is helpful. However, since a single APEX-GFP construct is used, the authors should address the differences between panels A and B either in Results or the figure legend. Specifically, while DAB produces symmetrical staining, GFP is quite a bit stronger in one lobe; furthermore, prominent puncta are visible in panel A that have no match in panel B. What are the puncta? Can the authors confirm that 2A is indeed GFP, despite the puncta and the general graininess of the image? Finally, Figure 2b legend is vague as to whether or not a section from the brain in 2A is shown following DAB polymerization – it probably is not the same brain, and this should be made clear.

B) Figure 1l and Figure 2e are not particularly convincing images: If these are from a series, then perhaps showing the subsequent images in a supplemental figure would help convince a reader about the nature of the synaptic contact.

C) Provide new Figure 1—figure supplements 1c and 1e – these were mentioned before, and un-necessarily make the standard look worse due to folds and knife marks in the images.

3) In each figure legend, state explicitly the mouse or fly sources of the images.

https://doi.org/10.7554/eLife.15015.018

Author response

Essential revisions:

1) An important question concerns efficiency. Are all of the genetically specified neurons labeled or just a subset? For example, the Drosophila DB331 driver line should target a bundle of motion-sensitive cells with numerous spatially constrained projections, yet Figure 2 displays only a handful of processes. In both Figure 1(C-D) and Figure 2, a low-resolution comparison of Cre-dependent fluorophore and DAB staining patterns are needed to address this concern.

As the reviewers noted, the DB331-Gal4 line expresses Gal4 in a set of motion-sensitive cells comprising 6 vertically and 3 horizontal motion sensitive lobula plate tangential cells, plus a sparse group of small interneurons. Confusion arose because in the original Figure 2, the plane of section revealed the axonal, but not the dendritic processes, which are posterior to the axons. Indeed, more posterior sections do show the expected complex pattern of dendritic processes expected from these tangential cells (data not shown). We have taken two steps to address the concern. First, we show a section in which 9 DAB-positive processes, the expected number, are visible. We have added arrows to point them out. Second, as suggested, we have added low- resolution panels to compare the Gal4 dependent expression of fluorophore (APEX-GFP) and DAB reaction product (Figure 2a–b). These micrographs document the correspondence between these two reporters.

2) The paper does a convincing job of showing how 'rapid' low-resolution images and reconstructions can be made of axons and dendrites. Nevertheless, to validate the method's use in circuit analysis, consistent high quality images are needed that show synapses identified by arrows with detailed descriptions in the figure legends and results (a few of the images are of sufficient magnification and quality but synapses should be identifiable (Figures 3h and Figure 1—figure supplements 3f and 4a, b)). Figure 1—figure supplement 3f is the only image that actually identifies the synapses with arrows, yet still provides too little description.

We have added several high-resolution images in which synapses (pointed out with arrows) can be identified. They are Figure 1k, l, o; Figure 1—figure supplement 4; and Figure 1—figure supplement 5. We have also provided more detailed descriptions in the text and figure legends.

3) All of the example images reveal suboptimal ultrastructure inside the labelled neurites, relative to other methods that have been used to label synapses in the literature. The dense reaction product obscures pre- and postsynaptic active zones, as well as presynaptic vesicles. This being the case, it is not clear how synapses will be identified on the specifically labeled cells. This issue must be addressed with images, labels, and clear examples of labeled synapses.

As the reviewers note, the cytosolic DAB-polymer deposited in APEX-expressing neurons obscures the ultrastructure of the labeled neurites. Nonetheless, key aspects of synaptic interactions can still be clearly identified. For example, cells presynaptic to APEX-expressing cells have unaffected synaptic machinery, including presynaptic densities, ribbons and vesicles. In addition, even within APEX-positive cells, synaptic vesicles can be visualized because they remain unstained. Moreover, erHRP is confined to membrane bound compartments, leaving the cytosol unstained; in cells expressing erHRP, both pre- and postsynaptic structures can be visualized. To emphasize these points, we have provided new images in Figure 1k, l, and O andalso added a new supplementary figure, Figure 1—figure supplement 5. These figures present images of pre- and postsynaptic partners to APEX2 positive cells in the retina and superior colliculus, as well as synapses onto erHRP expressing retinal ganglion cells. We have also revised the text and figure legend to emphasize that ultrastructural detail preserved in peroxidase expressing cells does allow identification of pre- and postsynaptic partners. That said, previous publications make clear that postsynaptic densities are not as prominent in the retina as in other brain regions such as cortex. We also note this in the text.

4) Both aqueous osmium and reduced osmium (rOTO protocol) were used. However, it is not completely clear which osmium staining was used in some of the figures. In addition, it is not entirely clear whether there were any significant differences between the single osmium or rOTO protocol methods for low magnification reconstructions or synapse identifications. The comparisons in Figure 1—figure supplement 1 need more explanation.

We used the “rOTO” protocol in all experiments on retinal tissue, and now include a sentence in the Materials and methods section clarifying this. This protocol improves the signal strength of membranes, dramatically improving the signal-to-noise ratio at short pixel dwell times, thereby allowing increased scanning speed. We used the single osmium protocol only in the HEK cell experiment in Figure 1—figure supplement 1, to provide additional evidence on the importance of the redox state of the reaction products. We edited the text to clarify this point.

5) It appears that the findings are intended to be qualitative; however, some form of semi-quantitative comparisons are warranted to establish the claims. For example, blind scoring of synapses (such as 0-3) and ultrastructural quality under the different conditions would be welcomed to assess the stated improvements.

Figure 1—figure supplement 4 was intended to show the improvements of the “signal-to-noise” ratio between the membrane and cytosolic when comparing a standard with a tissue reduction protocol. We have now added high-resolution images of representative synapses to this Figure. As an additional “semi-quantitative” comparison, we asked a blind-to-condition observer to select the better example between pairs of manually selected images of synapses (images as in Figure 1—figure supplement 1a, b). This randomized pairwise comparison shows that on average, the reduction improves synaptic recognition by 40% by reducing the tissue. This seems to be because presynaptic densities are more prominent in the reduced condition. Results of this comparison are now presented in Figure 1—figure supplement 4e.

6) Similarly, a systematic (and blind as to condition) quantification of a subset of small volumes of synapses from control and reacted tissue is needed to rule out possible effects of peroxidase expression on the ability to identify synapses in either the labelled cells or among the associated or surrounding but unlabeled synapses.

As discussed in our response to point #3, pre- and postsynaptic partners can be distinguished and characterized in peroxidase expressing neurons. Qualitatively, these synaptic connections do not seem to vary in APEX2 or erHRP expressing cells, compared to control tissue. Although we agree that an unbiased (blind to condition) comparison would be valuable, it is not feasible. Due to the prominence the peroxidase staining it is also not trivial to design an experiment that is blind to condition. In addition, the size and prominence of synaptic specializations varies markedly among retinal subtypes. For example, ribbons in rod bipolar cells are larger and more pronounced than those in cone bipolars. Therefore, cell type specific markers, which we did not use, would be needed for valid comparisons. As an approximation, we noted the number of synapses that we counted per unit area in peroxidase-expressing and control retinas. We identified 0.34 synapses per μm2 in control tissue 0.32 synapses per μm2 in peroxidase expressing tissue. The difference was not significant. We have added this comparison to the text.

7) The authors indicate that the technique provides improved preservation. Although in Figure 1, panels H to L are of acceptable magnification and resolution, they do not show particular evidence of improved preservation relative to other papers in the literature. Description and discussion are needed to clarify to what the improvement has been made.

We apologize for being unclear in the text. We did not intend to claim that we have the best preservation relative to other papers in the literature. High pressure freezing, for example, may avoid the tissue distortions of chemical fixation protocols. Our claim is that our reduction protocol improves the contrast difference between membranes and cytosol by comparison to conventionally fixed tissue, and also enhances the signal of the presynaptic machinery (Please also see our response to point #5). We believe that these effects might be important for automatic segmentation algorithms that are expected to segment large data sets in the future. We have revised the text and figure legend (Figure 1—figure supplement 4) to avoid confusion.

8) It is noted that the fixation of the fly brain was with immersion, without dissecting it away from the cuticle. Only an incision was made in the back of the head. Some explanation is needed as to why the authors used this approach, instead of removing the whole cuticle as is standard in the literature. The tissue preservation does not look as good as that in other published images where the whole cuticle was removed, and fixative penetration was facilitated, for example with microwave irradiation.

We apologize for an incomplete description of the method, and have expanded the description. After making the incision we separate the posterior cuticle from the head. This allows unhindered access of the fixative to the brain tissue. We reasoned that by completely removing the cuticle we might do more damage to the tissue. The remaining cuticle was left intact to protect the tissue during all the staining procedure. We agree with the reviewers that this is not the best fly brain preservation. Our protocol was optimized for mammalian tissue. Our intention here was to test the generality of the method by determining whether it could be applied to flies; we show than it can. There is a clear path of improvement regarding tissue preservation in flies. We clarified this in the text.

9) Concerning labelling in the fly brain, the outcomes should be described in more detail in the Results section (Results and Discussion, fifth paragraph). Higher magnification images are needed in Figure 2 to identify synapses, which should be labelled on the images and described in the Figure legend. These higher resolution images could replace one or more of the panels in Figure 2e–g. It is important to show here the synaptic contacts with the labelling on either the pre or post-synaptic side. We recommend that sufficient magnification be shown so that synapses are readily identifiable, with arrows, and more detailed descriptions in the results and legends.

As requested by the reviewers, we replaced panels in Figure 2e–g with high magnification images. One shows a presynaptic contact to an APEX-positive cell, a second one of an APEX- positive cell containing a number of contrast-inverted vesicles, similar to those observed APEX- positive cells in the superior colliculus. As described in our response to point #8, our protocol was optimized for mammalian retina tissue and we acknowledge that there is need of improvement in the preservation protocol of the fly brain. We make this point clear in the text.

10) The description of the automatic 2D segmentation and 3D reconstruction explains that speckle noise is removed by pruning connected components smaller than 1500 pixels in area. As written, it is not clear how this removal was done. Based on the low resolution that was used for the EM imaging, the size of the removed component seems rather large. Description of how effective this process was when dealing with long thin neurites is needed.

The algorithm’s parameters are set based on the regions under consideration from a specific data set (i.e., the histogram profile of the pixel intensities in that region). We have described the pruning process for noise and other artifacts in the Materials and methods section of the supplemental material (it incorporates segment area and major access length), and have provided an in-depth description of the algorithm’s reconstruction methodology in the attached Matlab live script walk-through.

11) A more direct statement is needed in Figure 3 legend, that 20 nm resolution is not sufficient to identify synapses, even in the unlabeled processes. Then in Figure 3h (4 nm resolution), synapses should be explicitly identified with arrows.

We added the following sentence: “Although both images contain ribbon synapses, the low resolution image (G) lack the required resolution. Arrows point to the location of synaptic ribbons.”

12) The serial sections selected for Figure 3j should show a synapse with one of the labelled cells, which is needed to validate the circuit reconstruction.

As noted above, synaptic specializations cannot be visualized when both presynaptic and postsynaptic cells have DAB-polymer in the cytosol. We base identification of the synapses between APEX-positive starburst amacrine cells (Figure 3j, k) on the morphological criteria used Briggman et al. (2011) who faced a similar problem owing to the staining protocol they used. We can, however, visualize all other unstained synaptic partners based on their ultrastructural features. We replaced Figure 3j with high magnification images showing additional synaptic contacts to the reconstructed cells and pointed these out with arrowheads.

13) Supplementary files 13 were not included in the PDF for review and need to be submitted with the revision for proper review.

The files were uploaded but were unfortunately omitted from the final PDF provided to reviewers. We apologize for failing to notice this when we approved the PDF. These supplementary files are now included.

14) In addition, it would be helpful if the software could be prepared for expert review upon revision.

We now include a step-by-step Matlab script with examples as supplementary information.

[Editors' note: further revisions were requested prior to acceptance, as described below.]

Essential further revisions:

1) Issue of overstatements on 'improved ultrastructure' needing more work:

A) Abstract – remove the phrase "and improves ultrastructural preservation". This method does not improve ultrastructural preservation; it enhances staining. Could substitute "while retaining ultrastructure" or something similar. The purpose of this method is not to improve ultrastructural preservation, and as noted in the prior review, the ultrastructure of the labeled cells is not particularly good with respect to other methods.

Reworded as requested.

B) Introduction – second paragraph – (b) enhancing the electron-density of the stain without compromising ultrastructure. To be clear reword as "without compromising the surrounding ultrastructure" or in a similar way. As written, it sounds like the ultrastructure of the labelled cell would not be compromised, but of course it is compromised as the stain obscures most of the ultrastructure.

Reworded as requested.

C) Consider the following concern raised by one of the reviewers and revise the Introduction and Discussion accordingly:

The narrative does not oversell the technique applicability for ultrastructure analysis. However, since the ability to examine synaptic ultrastructure in genetically designated neurons would represent a significant milestone, it may be more appropriate to characterize explicitly this current incarnation of peroxidase-based methods as most appropriate for rapid reconstruction of targeted neuron morphology and connectivity, rather than synapse properties.

We have included this limitation as requested.

2) Some specific figures still need to be improved:

A) Figure 2: The addition of panels showing complementary GFP and DAB staining is helpful. However, since a single APEX-GFP construct is used, the authors should address the differences between panels A and B either in Results or the figure legend. Specifically, while DAB produces symmetrical staining, GFP is quite a bit stronger in one lobe; furthermore, prominent puncta are visible in panel A that have no match in panel B. What are the puncta? Can the authors confirm that 2A is indeed GFP, despite the puncta and the general graininess of the image? Finally, Figure 2b legend is vague as to whether or not a section from the brain in 2A is shown following DAB polymerization – it probably is not the same brain, and this should be made clear.

We confirm that 2A does show GFP. The signal was enhanced by immunostaining with anti-GFP, which is our lab’s standard protocol. The puncta likely reflect antibody background rather than ectopic GFP.

As to the differences between A and B, they reflect the differences in tissue preparation and staining protocol rather than differences in which cells are labeled. We ascribe the lack of bilateral symmetry in A to imperfections in tissue clearing and preparation; it is not a consistent finding. By showing both panels at higher magnification, we emphasize the similarity.

The sections in A and B are from different animals. We have added this information to the legend.

B) Figure 1l and Figure 2e are not particularly convincing images: If these are from a series, then perhaps showing the subsequent images in a supplemental figure would help convince a reader about the nature of the synaptic contact.

We substituted better micrographs for Figure 1l and 2e. In Figure 2, we now include examples of both pre- and postsynaptic APEX-positive processes, as well as an unstained synapse from this material to demonstrate the quality of the ultrastructure.

C) Provide new Figure 1—figure supplements 1c and 1e – these were mentioned before, and un-necessarily make the standard look worse due to folds and knife marks in the images.

We have substituted new images for supplemental Figures 1c and 1e. 1C has no folds and 1E has far fewer folds than its predecessor. Since this is an issue of aesthetics rather than clarity, we do not believe it is worth spending more time preparing new tissue.

3) In each figure legend, state explicitly the mouse or fly sources of the images.

Added as requested.

https://doi.org/10.7554/eLife.15015.019

Article and author information

Author details

  1. Maximilian Joesch

    1. Center for Brain Science, Harvard University, Cambridge, United States
    2. Department of Molecular and Cellular Biology, Harvard University, Cambridge, United States
    Contribution
    MJ, Conception and design, Acquisition of data, Analysis and interpretation of data, Drafting or revising the article
    Competing interests
    The authors declare that no competing interests exist.
  2. David Mankus

    1. Center for Brain Science, Harvard University, Cambridge, United States
    2. Department of Molecular and Cellular Biology, Harvard University, Cambridge, United States
    Contribution
    DM, Acquisition of data, Analysis and interpretation of data
    Competing interests
    The authors declare that no competing interests exist.
  3. Masahito Yamagata

    1. Center for Brain Science, Harvard University, Cambridge, United States
    2. Department of Molecular and Cellular Biology, Harvard University, Cambridge, United States
    Contribution
    MY, Conception and design, Acquisition of data, Contributed unpublished essential data or reagents
    Competing interests
    The authors declare that no competing interests exist.
    ORCID icon 0000-0001-8193-2931
  4. Ali Shahbazi

    University of Notre Dame, Notre Dame, United States
    Contribution
    AS, Acquisition of data, Analysis and interpretation of data, Contributed unpublished essential data or reagents
    Competing interests
    The authors declare that no competing interests exist.
  5. Richard Schalek

    1. Center for Brain Science, Harvard University, Cambridge, United States
    2. Department of Molecular and Cellular Biology, Harvard University, Cambridge, United States
    Contribution
    RS, Acquisition of data
    Competing interests
    The authors declare that no competing interests exist.
  6. Adi Suissa-Peleg

    School of Engineering and Applied Sciences, Harvard University, Cambridge, United States
    Contribution
    AS-P, Acquisition of data
    Competing interests
    The authors declare that no competing interests exist.
  7. Markus Meister

    Division of Biology, California Institute of Technology, Pasadena, United States
    Contribution
    MM, Conception and design
    Competing interests
    The authors declare that no competing interests exist.
    ORCID icon 0000-0003-2136-6506
  8. Jeff W Lichtman

    1. Center for Brain Science, Harvard University, Cambridge, United States
    2. Department of Molecular and Cellular Biology, Harvard University, Cambridge, United States
    Contribution
    JWL, Conception and design
    Competing interests
    The authors declare that no competing interests exist.
  9. Walter J Scheirer

    University of Notre Dame, Notre Dame, United States
    Contribution
    WJS, Conception and design, Analysis and interpretation of data
    Competing interests
    The authors declare that no competing interests exist.
  10. Joshua R Sanes

    1. Center for Brain Science, Harvard University, Cambridge, United States
    2. Department of Molecular and Cellular Biology, Harvard University, Cambridge, United States
    Contribution
    JRS, Conception and design, Analysis and interpretation of data, Drafting or revising the article
    For correspondence
    sanesj@mcb.harvard.edu
    Competing interests
    The authors declare that no competing interests exist.
    ORCID icon 0000-0001-8926-8836

Funding

NIH Blueprint for Neuroscience Research (R01NS076467)

  • Maximilian Joesch
  • David Mankus
  • Masahito Yamagata
  • Richard Schalek
  • Markus Meister
  • Jeff W. Lichtman
  • Joshua R. Sanes

Human Frontier Science Program (LT000167/2010)

  • Maximilian Joesch

Howard Hughes Medical Institute (HCIA program - 1234)

  • David Mankus
  • Masahito Yamagata
  • Joshua R. Sanes

Intelligence Advanced Research Projects Activity (#D16PC00002)

  • Walter J. Scheirer

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Acknowledgements

We thank Alice Ting, Jeff Martell, and Stephanie Lam for advice and for providing the APX and APEX2 constructs. We also thank Chiao-Lin Chen and Norbert Perrimon for generating and providing the UAS-APEX-GFP fly and Benjamin de Bivort for support with the fly work. This work was supported by NIH grant NS76467 to MM, JL and JRS, an HHMI Collaborative Innovation Award to JRS, an IARPA contract #D16PC00002 to WJS and by The International Human Frontier Science Program Organization fellowship to MJ. The authors declare no conflicts of interest.

Ethics

Animal experimentation: Animals were used in accordance with NIH guidelines and protocols approved by Institutional Animal Use and Care Committee at Harvard University (Protocol 233 #92_19).

Reviewing Editor

  1. Kristen M Harris, Reviewing Editor, The University of Texas at Austin, United States

Publication history

  1. Received: February 4, 2016
  2. Accepted: June 27, 2016
  3. Accepted Manuscript published: July 7, 2016 (version 1)
  4. Version of Record published: July 25, 2016 (version 2)

Copyright

© 2016, Joesch et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.

Metrics

  • 4,158
    Page views
  • 1,223
    Downloads
  • 7
    Citations

Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.

Comments

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Download citations (links to download the citations from this article in formats compatible with various reference manager tools)

Open citations (links to open the citations from this article in various online reference manager services)