Mapping synaptic input fields of neurons with super-resolution imaging

Summary

As a basic functional unit in neural circuits, each neuron integrates input signals from hundreds to thousands of synapses. Knowledge of the synaptic input fields of individual neurons, including the identity, strength and location of each synapse, is essential for understanding how neurons compute. Here we developed a volumetric super-resolution reconstruction platform for large-volume imaging and automated segmentation of neurons and synapses with molecular identity information. We used this platform to map inhibitory synaptic input fields of On-Off direction-selective ganglion cells (On-Off DSGCs), which are important for computing visual motion direction in the mouse retina. The reconstructions of On-Off DSGCs showed a GABAergic, receptor subtype-specific input field for generating direction selective responses without significant glycinergic inputs for mediating monosynaptic crossover inhibition. These results demonstrate unique capabilities of this super-resolution platform for interrogating neural circuitry.

Introduction

Mapping synaptic connectivity at multiple scales, ranging from the synaptic fields of individual neurons to the wiring diagram of the whole brain, is important for understanding how neural circuits function and how circuit defects contribute to mental illness (Alivisatos et al., 2013; Morgan and Lichtman, 2013). An ideal platform for imaging synaptic connectivity should provide i) high resolution structural information for reliable identification of synaptic connections and accurate assignment of synapses to neurons, ii) the ability to image specific molecules, such as neurotransmitter receptors, important for determining synapse identity and properties, and iii) automated image segmentation capability for efficient analysis of large-volume reconstructions that capture entire neurons or circuits.

Both fluorescence microscopy and electron microscopy (EM) have been used for volumetric neural circuit reconstruction (Helmstaedter, 2013; Kleinfeld et al., 2011). EM provides exquisite spatial resolution and membrane contrast for accurate synapse identification, and the high imaging speed of modern EM instruments allows increasingly larger volume reconstructions (Helmstaedter, 2013; Kleinfeld et al., 2011). However, because of the stringent fixation and sample preparation conditions required for high-quality EM imaging, labeling of endogenous synaptic proteins for determining the molecular identities and functional properties of synapses remains a difficult task for large-volume EM reconstructions. In addition, automated segmentation of EM images is still challenging and remains a bottleneck for scaling up neural circuit analysis, though substantial progress has been made on the development of automated EM image analysis and crowd-sourcing methods (Chklovskii et al., 2010; Helmstaedter, 2013; Jain et al., 2010). In comparison, fluorescence microscopy is compatible with immunohistochemistry and imaging of endogenous proteins over large volumes (Kleinfeld et al., 2011; Miyawaki, 2015), and multi-colored fluorescence signals can also help simplify the task of automated image segmentation for efficient data analysis. However, the diffraction-limited resolution of fluorescence microscopy can lead to substantial errors in the identification and assignment of synapses within reconstructed circuits.

Super-resolution fluorescence imaging overcomes the diffraction limit (Hell, 2007; Huang et al., 2010) and may enhance our ability to reconstruct neural circuits by integrating high image resolution for synapse identification and assignment, protein-specific labeling for determining the molecular properties of synapses, and multi-color imaging for efficient data analysis. Here we developed a super-resolution reconstruction platform by combining Stochastic Optical Reconstruction Microscopy (STORM) (Huang et al., 2010; Rust et al., 2006) with serial ultrathin sectioning for large-volume reconstruction of endogenous molecular targets in tissues, and used this platform to image entire neurons and their synaptic inputs. We focused our studies on the inner plexiform layer (IPL) of the mouse retina where diverse classes of retinal ganglion cells (RGCs) integrate synaptic inputs (Anderson et al., 2011; Helmstaedter et al., 2013) to generate unique spatiotemporal representations of the visual scene (Gollisch and Meister, 2010). A classic example of such a computation is the determination of visual stimulus motion direction by On-Off direction-selective RGCs (On-Off DSGCs) (Vaney et al., 2012). The substantial prior knowledge of the structure and function of this cell type allows validation of our method, while unresolved structural questions in this system provide an opportunity to test the ability of our approach to extract novel biological information. For example, On-Off DSGCs are known to receive asymmetric inhibitory GABAergic inputs from presynaptic starburst amacrine cells (SACs) during null-direction stimulus movement (Briggman et al., 2011; Fried et al., 2002; Wei et al., 2011), and the α2 subunit of GABA(A) receptor plays an important role in this direction selectivity (Auferkorte et al., 2012). In addition to GABAergic synapses, glycinergic signaling also impacts the response of On-Off DSGCs to the edges of moving stimuli (Caldwell et al., 1978; Jensen, 1999), likely reflecting crossover inhibition between the on and off sublaminae mediated by glycinergic amacrine cells (Kittila and Massey, 1995; Stasheff and Masland, 2002; Werblin, 2010). However, the structural basis of this crossover inhibition in On-Off DSGC circuits is incompletely understood and it is unclear whether glycinergic interneurons make direct synaptic contacts onto On-Off DSGCs. To demonstrate the capabilities of our super-resolution platform, we reconstructed the inhibitory synaptic input fields of individual On-Off DSGCs and determined the spatial distribution and neurotransmitter receptor identity of the synapses therein. We also reconstructed the inhibitory input fields of two other types of retinal neurons, a small-field On-center RGC and a narrow-field amacrine cell, for comparative demonstration.

Results

Volumetric, multi-color super-resolution reconstruction

We labeled neurons and synaptic proteins with spectrally distinct photoswitchable dyes for multi-color STORM imaging (Dempsey et al., 2011). For neuron labeling, we used mice expressing GFP or YFP in the cytoplasm of a sparse subset of retinal neurons (Feng et al., 2000), and labeled the dissected retinal tissue with anti-GFP antibodies. For marking inhibitory synapses, we used an antibody against an inhibitory synapse scaffolding protein, gephyrin, which anchors glycine and/or GABA receptors at postsynaptic terminals (Tyagarajan and Fritschy, 2014). For presynaptic counter-staining, we used a cocktail of antibodies against several active zone proteins, bassoon, piccolo, munc13-1, ELKS, for dense labeling of all presynaptic terminals. Table S1 shows all of the antibodies tested in this work. We also included a general neuropil stain, wheat germ agglutinin (WGA), in a fourth color channel to produce images with dense information content to assist serial-section alignment.

For volumetric reconstruction, we embedded tissues in resin and used serial ultrathin sectioning, in combination with STORM imaging, to generate large-volume super-resolution images. Serial ultrathin sectioning not only facilitates large-volume fluorescence reconstruction of tissue samples, but also allows the image resolution along the z direction (as defined by the section thickness) to be substantially higher than the diffraction limit, as has been demonstrated previously in array tomography and thee-dimensional (3D) STED reconstructions (Micheva and Smith, 2007; Punge et al., 2008). The partial exposure of epitopes in samples embedded in acrylic resin also allows many different synaptic proteins to be imaged through multiple rounds of post-embedding immunolabeling, which help identify synapses and characterize their molecular properties (Micheva et al., 2010; Micheva and Smith, 2007). However, the requirement of sample embedding for high-quality serial sectioning poses extra challenges for super-resolution imaging. Since STORM imaging relies on switching and localization of individual fluorophores to reconstruct super-resolution images (Huang et al., 2010; Rust et al., 2006), the resolution of a STORM image depends not only on the localization precision of individual fluorophores determined by their photon output, but also on the localization density determined by the labeling density. Achieving optimal STORM resolution thus requires the labeling and embedding conditions to simultaneously retain optimal fluorophore properties and high density labeling in resin-embedded samples. Resin embedding, however, substantially reduces the antigenicity of samples, which leads to a drastic reduction in antibody labeling density and severely compromises the image resolution achievable by STORM as we observed for tissues labeled after acrylic resin embedding. Such low label densities, which are also evident in previous STORM images of tissue samples prepared using a similar post-embedding labeling approach (Nanguneri et al., 2012), prevent accurate tracing of neurons and identification of synapses using super-resolution imaging. We therefore explored pre-embedding immunofluorescence labeling (Punge et al., 2008) to increase the labeling density. We further tested various embedding materials and found that epoxy resin was excellent for maintaining the photon output of the fluorescent dyes. Finally, since optimal photoswitching of dye molecules requires access to a switching agent, such as thiol, we chemically etched the resin-embedded tissue sections using sodium ethoxide solution to expose the dyes to the thiol-containing imaging buffer.

Experimentally, we immunolabeled retinal tissues, performed an additional fixation step to crosslink the antibodies, dehydrated the samples, and embedded them in UltraBed epoxy resin (Figure 1A). The resin-embedded tissues were cut into 70 nm ultrathin sections, arrayed onto glass coverslips, and etched with sodium ethoxide (Figure 1A). Coverslips were imaged using a microscope setup that allowed automated imaging of entire arrays of sections, and both STORM and conventional images were collected for the same tissue sections. The xy-resolution of the STORM images was ∼ 20 nm and that of the conventional images was diffraction-limited to ∼200-300 nm, whereas the z-resolution of both STORM and conventional images in this work was limited by the section thickness of 70 nm.

Figure 1. A super-resolution imaging and analysis platform.

(A) Tissues were dissected, fixed for immunohistochemical labeling, postfixed, dehydrated, and embedded in epoxy resin. Ultrathin sections were cut, arrayed on glass coverslips, and etched to expose fluorophores for STORM imaging. Individual serial sections were imaged and aligned to generate 3D reconstructions. (B) STORM maximum intensity projection of a volume (2.3 × 10⁵ μm³) of the mouse IPL containing an On-Off DSGC (blue) amidst presynaptic (magenta) and gephyrin (green) clusters imaged using the platform. (C) An enlarged image of synapses in a small region (1 μm thickness) of the IPL. For comparison, the corresponding conventional image of the upper right portion (to the right of the dashed line) is presented. See also Figure S1 and Figure S2.

We developed an automated image analysis pipeline for processing STORM and conventional images, which included corrections of chromatic aberration and lens distortions using bead fiducials, as well as montage and serial-section alignment using scale-invariant feature transformation (SIFT) followed by elastic registration (Saalfeld et al., 2012) to generate large-volume reconstructions (Figure 1A) (see Experimental Procedures for details). Volumetric STORM reconstructions of the IPL revealed efficient labeling throughout the sample, with neurons situated amidst hundreds of thousands of fluorescent clusters in each synaptic channel (Figure 1B, C; Figures S1 and S2).

Synapse Identification

Taking advantage of the multi-color super-resolution fluorescent signals, we developed image segmentation algorithms for automated neuron and synapse identification, and performed quantitative analysis of entire fields of molecularly identified synapses in our datasets. Labeled synaptic proteins appear as clusters of localizations in STORM images, but not all clusters in STORM images represent synapses (Dani et al., 2010; Specht et al., 2013). For synapse identification, we measured the volume and signal density of all fluorescent clusters. In both presynaptic and postsynaptic (gephyrin) channels, these two parameters separated fluorescent clusters into two distinct populations (Figure 2A). We assigned the population of clusters with larger volumes as putative “synaptic” (S) structures and the other population with smaller volumes as putative “non-synaptic” (NS) structures. The vast majority (∼91%) of the putative synaptic gephyrin clusters had closely apposed presynaptic clusters (Figure S3A, C), and example pairs of gephyrin and presynaptic clusters from this population clearly resembled synapses (Figure 2B), supporting our assignment. Of the putative non-synaptic population of gephyrin clusters, only a small fraction had a nearby presynaptic cluster (Figure S3A). Moreover, because some of these small gephyrin clusters were spatially close to the larger, paired gephyrin/presynaptic clusters and were thereby falsely identified as being paired, the automated pairing analysis of these small gephyrin clusters (Figure S3A) was less accurate than that for the larger synaptic clusters. Visual inspection showed that ∼90% of these small gephyrin clusters were unpaired and likely represent gephyrin-containing trafficking vesicles or background signals from non-specific antibody labeling, whereas the remaining small fraction of paired structures could represent small (potentially immature) synapses.

Figure 2. Automated inhibitory synapse identification within the IPL.

(A) Gephyrin (left) and presynaptic (right) clusters across the IPL can be separated into putative synaptic (S) and non-synaptic (NS) populations based on the volumes and signal densities of the clusters. Shown are the 2D distributions of cluster volume and signal density constructed from all gephyrin and presynaptic clusters identified in the image block, with the cluster volume plotted on the log scale. The signal density is defined as the fraction of the volume occupied by the cluster that is positive for the gephyrin or presynaptic signal. (B) Six example pairs of gephyrin (green) and presynaptic (magenta) clusters. The synapses are rotated to show side and en face views. (C) Projection images of the IPL showing gephyrin-paired and unpaired synaptic clusters in the presynaptic channel. GCL: Ganglion cell layer. INL: Inner nuclear layer. (D) The laminar distributions of the gephyrin-paired and unpaired presynaptic clusters divide the IPL into several sublaminae. Grey bars: Presynaptic pairing index as a function of IPL depth. ‘Presynaptic pairing index’ is calculated as the difference of the paired and unpaired presynaptic laminar intensity distributions after first standardizing each distribution to have a mean of zero and a standard deviation of one. Blue bars: Volume of the On-Off DSGC (μm³) per 0.5 μm bin as a function of IPL depth. (E) The total signal intensity (left), average cluster density (middle) and average cluster volume (right) for each cubic micron of imaged tissue measured as a function of depth within the IPL for gephyrin clusters that are paired with presynaptic clusters. (F) Similar to (E) but for presynaptic clusters that are paired with gephyrin clusters. The delineation of sublaminae S3 and S7 in (E) and (F) was determined based on (D). See also Figure S3.

For the presynaptic clusters, even the synaptic population contained a substantial fraction of clusters (∼70%) that were not paired with gephyrin clusters (Figure S3B, D). This is expected as the cocktail of antibodies against presynaptic active-zone proteins should label the presynaptic terminals of both excitatory and inhibitory synapses, and excitatory presynaptic terminals would not be expected to pair with gephyrin. The population of small presynaptic clusters did not show any appreciable pairing with gephyrin clusters (Figure S3B).

These analyses demonstrated that we could identify synapses based on the size and signal density of the fluorescent clusters observed in STORM images. In the following, we focus our analysis on the population of synaptic clusters with larger volumes. In contrast, similar analysis of either gephyrin or presynaptic clusters observed in the corresponding conventional images did not allow clear distinction between synaptic and non-synaptic clusters (Figure S4A).

Examination of the hundreds of thousands of automatically identified synapses in STORM images of the inner retina showed non-uniform distributions across the depth of the IPL (Figure 2C). The difference between gephyrin-paired (inhibitory) and unpaired (putative excitatory) presynaptic cluster intensities divided the IPL into several sublaminae, two of which coincided with On-Off DSGC stratification in sublaminae S3 and S7 (Figure 2D) (Vaney et al., 2012). Interestingly, the gephyrin and presynaptic signal intensities, and the density and volumes of these gephyrin-positive inhibitory synapses all peaked in S3 and S7 (Figure 2E, F). As the size of inhibitory synapses correlates with synaptic strength (Lim et al., 1999; Nusser et al., 1997; Nusser et al., 1998), this observation suggests that the inhibitory synapses subserving On-Off direction-selectivity may be among the strongest inhibitory connections in the mouse retina.

Identifying inhibitory synaptic inputs to labeled neurons

To demonstrate the ability of our super-resolution platform to segment and analyze synaptic inputs onto identified neurons, we reconstructed two types of retinal ganglion cells and their associated inhibitory synaptic fields. Each of these data sets consisted of both STORM images (Figure 3A, left) and, for comparison, the corresponding conventional images (Figure 3A, right). To identify synaptic inputs onto neurons, we measured the density of gephyrin clusters and associated presynaptic signals as a function of distance to the neuron surface. Both density functions derived from STORM images were sharply peaked near the neuron surface with the gephyrin peak slightly inside the neuron and the presynaptic signal slightly outside the neuron as expected for input synapses (Figure 3B). These density peaks, in particular the gephyrin peak, were followed by a depletion zone where the density dropped below the mean density of the surrounding IPL. For automated assignment of synapses to the neuron, we set a cutoff at the point where the gephyrin density dropped below the mean density of the surrounding IPL, and selected only those gephyrin clusters located at a distance below the cutoff as synaptic inputs to the neuron (Figure 3B, D left panel, and Movies S1).

Figure 3. Automated segmentation of synaptic inputs to neurons.

(A) STORM maximum projection image of a region containing a dendritic branch of a reconstructed On-Off DSGC (left) and the corresponding conventional image (right). Neurite is in blue, gephyrin in green, and presynaptic channel in magenta. (B) The densities of the gephyrin clusters that are paired with presynaptic clusters (green trace), the unpaired gephyrin clusters (blue trace), and the gephyrin-paired presynaptic signal (magenta trace) measured as a function of the distance to the neuron surface. The distance at which the density peak of gephyrin clusters drops below the mean synapse density of the surrounding IPL (dashed green line) is used as a cutoff for defining gephyrin clusters on the neuron. (C) For each synapse, we measured the distances of the presynaptic and postsynaptic signal to the neuron surface and defined the difference between these two distances as the relative presynaptic-gephyrin distance from the neuron surface. All synapses assigned to the On-Off DSGC show positive relative distance values (solid line), consistent with these pairs being input synapses onto the neuron. In contrast, the spatial arrangement of nearby synapses within 500 nm of the neuron (dashed line) shows a broad distribution of both positive and negative relative distance values, indicating a random orientation of nearby synapses with respect to the neuron surface. (D) Assignment of synapses in the STORM image based on the cutoff selected in (B) reveals adjacent presynaptic and postsynaptic structures associated with the neuron (left panel). In contrast, assignment of synapses in the conventional images with a cutoff at 0 nm (middle panel) or 150 nm (right panel) show false negative (arrows) and false positive synapse assignments (arrow heads). (E) En face view (top) and side view (bottom) of the STORM maximum intensity projection of a reconstructed On-Off DSGC (blue) with associated synaptic gephyrin (green) and presynaptic (magenta) clusters. Although gephyrin and presynaptic clusters are clearly resolved in the original reconstruction (Figure 3D and Movie S1), they appear as overlapping white dots here due to image downsampling. See also Figure S4 and Movies S1 and S2.

Figure 3E and Movie S2 show the 1017 inhibitory synapses assigned to a reconstructed On-Off DSGC. The number of synapses that we identified by STORM reconstruction here was similar to that estimated by previous EM reconstructions of SAC inputs to On-Off DSGCs (Briggman et al., 2011). Moreover, more than 98% of the synaptic gephyrin clusters assigned to the neuron had an apposing presynaptic partner. All of the gephyrin-presynaptic pairs assigned to the neuron were spatially oriented with the presynaptic structure more distant from the neuron than the postsynaptic structure (Figure 3C), consistent with these structures being input synapses onto the neuron. Together, these results further demonstrated the high accuracy in our synapse identification and assignment.

In comparison, assignment of synapses to neurons based on the corresponding conventional fluorescence images was less precise as the diffraction-limited resolution made it difficult both to identify synaptic clusters and also to set a proper cutoff value for assigning clusters to the neuron (Figure S4). As a result, this analysis resulted in substantial error rates (up to ∼50%) depending on the selected cutoff distance (Figure 3D; Figure S4C).

Distribution of inhibitory inputs to On-Off DSGCs

We next evaluated the size and position of all gephyrin-positive synapses within the dendritic arbor of each reconstructed cell. On-Off DSGCs, such as that shown in Figure 4A, exhibited non-random synapse distributions on both local and whole-cell scales (Figure 4B-E). A Ripley's clustering analysis showed that synapses were significantly more depleted within ∼1 μm of another synapse than would be predicted by a random distribution on the dendritic arbor (Figure 4B, Figure S5A), likely reflecting a minimum inter-synapse spacing imposed by the finite size of each synapse, consistent with a previous observation (Bleckert et al., 2013). On the whole-cell scale, On-Off DSGCs exhibited sublaminar specificity with substantially higher synapse density in sublaminae S3 and S7 than in other sublaminae, even after normalization for the different surface areas of dendrites across the IPL depth (Figure 4C and Figure S5B). This pattern is consistent with the specific innervation of On-Off DSGCs by SACs (Vaney et al., 2012), which also stratify in S3 and S7.

Figure 4. Distributions of inhibitory synapses on an On-Off DSGC and a small field On-center RGC.

(A) Surface renderings of the On-Off DSGC (grey) shown in Figure 3E with all inhibitory synaptic inputs marked by circles whose color (blue to red) reflects gephyrin cluster intensity on a log scale. (B) A one-dimensional Ripley's clustering analysis along the path of the skeletonized neuron. Negative value of the Ripley's function K(t) – ū at short inter-synaptic distances indicate that near any given synapse, the density of other synapses is significantly lower than a random distribution (see Experimental Procedures for the definition of Ripley's K function). (C-E) The laminar (C) radial (D) and angular (E) distributions of the inhibitory synapse densities on the On-Off DSGC. (F) Surface renderings of a small-field On-center RGC (grey) with all inhibitory synaptic inputs marked by circles whose color (blue to red) reflects gephyrin cluster intensity. (G-J) Similar to (B-E) but for the On-center RGC. Pink regions in (B-E) and (G-J) reflect 5/95% confidence intervals of random distributions derived from 1000 randomizations of the synapse positions. See also Figure S5 and Movies S2 and S3 .

Distribution of inhibitory inputs to a small-field On-center RGC

For comparison, we examined the sizes and spatial distribution of inhibitory synapses (936 total) onto a small-field On-center RGC (Figure 4F, G-J and Movie S3), a putative type G6 as previously classified (Volgyi et al., 2009). Similar to On-Off DSGCs, synapses on this cell also exhibited a non-random spatial distribution on the local scale where Ripley's clustering analysis showed a ∼1-2 μm depletion zone in the vicinity of each synapse (Figure 4G). However, the inhibitory synaptic input field of this neuron exhibited less sublaminar specificity than On-Off DSGCs on a whole-cell scale (Figure 4H).

Receptor identity of inhibitory inputs to On-Off DSGCs

To demonstrate the capability of this super-resolution fluorescence reconstruction platform to determine the molecular identities of synaptic connections within neural circuits, we performed experiments to disambiguate different inhibitory synaptic input classes (GABAergic versus glycinergic) onto identified neurons. We labeled retinae with either an antibody against the α2 subunit of the GABA(A) receptor (GABA(A)Rα2) or an antibody cocktail against all alpha subunits of glycine receptors (GlyRα1-4), in addition to antibodies against GFP and gephyrin for marking neurons and inhibitory synapses, respectively. To determine whether each gephyrin-positive inhibitory synapse contained GABA(A)Rα2 or glycine receptors, we examined whether the corresponding synaptic gephyrin cluster was paired with a specific receptor cluster by using the same approach described above for pairing presynaptic and postsynaptic structures (Figure S6A, B).

In GABA(A)Rα2-labeled samples, On-Off DSGC dendrites contained many GABA(A)Rα2-paired gephyrin clusters, but strikingly rare unpaired gephyrin clusters (Figure 5A, B, and Figure S6C). The gephyrin and GABA(A)Rα2 signal intensities in these synapses were strongly correlated with a Pearson coefficient of 0.82 (Figure S6E), suggesting that gephyrin intensity in these synapses correlates with synaptic strength, as is the case elsewhere in the nervous system (Lim et al., 1999; Nusser et al., 1997; Nusser et al., 1998). Quantitatively, 97 ± 1% of the gephyrin-positive synapses on On-Off DSGCs contained GABA(A)Rα2, suggesting a high labeling efficiency of the receptors. Compared with synapses on On-Off DSGCs, only ∼45% of all gephyrin-positive synapses analyzed across the IPL contained GABA(A)Rα2, demonstrating a strong enrichment of GABA(A)Rα2 in the synapses onto On-Off DSGCs. Although not all GABA receptor types are anchored at synapses by a gephyrin scaffold (Brickley and Mody, 2012; Tretter et al., 2012; Tyagarajan and Fritschy, 2014), gephyrin-independent GABA receptors are unlikely to contribute to direction-selectivity (Brickley and Mody, 2012; Massey et al., 1997). This, together with the similar synapse counts observed between our experiments and previous EM reconstructions of SAC synapses onto On-Off DSGCs (Briggman et al., 2011), suggests that the vast majority, if not all, of the inhibitory synapses onto On-Off DSGCs are gephyrin positive. Hence, our observations suggest that nearly all of the inhibitory synapses onto On-Off DSGCs contain the GABA(A)Rα2 subunit.

Figure 5. Receptor identity of the inhibitory synaptic inputs to On-Off DSGCs.

(A) Top panel: Surface rendering of a central cross-section of an On-Off DSGC (grey) with GABA(A)Rα2 positive (+) and GABA(A)Rα2 negative (-) inhibitory synapses marked as magenta and green circles, respectively. STORM image of the boxed region is shown in the bottom panel. Neuron: blue. Gephyrin: green. GABA(A)Rα2: magenta. (B) The GABA(A)Rα2-paired gephyrin (green), gephyrin-paired GABA(A)Rα2 (magenta), and unpaired gephyrin cluster (blue) densities as a function of the distance to the neuron shown in (A). (C) Top panel: Surface rendering of a central cross-section of an On-Off DSGC (grey) with GlyRα1-4 positive (+) and GlyRα1-4 negative (-) inhibitory synapses marked as magenta and green circles, respectively. STORM image of the boxed region is shown in the bottom panel. Neuron: blue. Gephyrin: green. GlyRα1-4: magenta. (D) The GlyRα1-4-paired gephyrin (green), gephyrin-paired GlyRα1-4 (magenta), and unpaired gephyrin cluster (blue) densities as a function of the distance to the neuron shown in (C). See also Figure S6 and Figure S7.

In stark contrast to the GABA(A)Rα2-labeled samples, in the GlyRα1-4 labeled samples we observed very few GlyRα1-4 positive synaptic gephyrin clusters on On-Off DSGCs (Figure 5C, D and Figure S6D). Quantitatively, only 8 ± 4% of the synaptic gephyrin clusters on On-Off DSGCs contained any GlyRα1-4 signal and even these synapses exhibited extremely sparse GlyRα1-4 labeling relative to nearby glycine-positive synapses not on the labeled On-Off DSGCs (Figure S6F, G). Since these nearby synapses contained substantial GlyRα1-4 signal, the lack of GlyRα1-4 in the On-Off DSGC synapses could not be attributed to low receptor labeling efficiency. Moreover, while previous work has shown a strong correlation between glycine receptor and gephyrin expression at synapses (Specht et al., 2013), we observed little correlation between the intensity of gephyrin and GlyR signals for these GlyR-positive gephyrin clusters on the On-Off DSGCs (Figure S6E). These results suggest that these sparse, low-intensity GlyR punctae probably reflect non-specific background labeling and even if they were specific synaptic labeling, they would contribute relatively little synaptic current due to the low receptor abundance. As gephyrin is required for clustering glycine receptors at synapses (Feng et al., 1998; Fischer et al., 2000; Kirsch et al., 1993), our results thus indicate that On-Off DSGCs in the mouse retina receive little glycinergic input.

In contrast to the STORM results, analysis of the corresponding conventional fluorescence images showed that a substantial population (20-30%) of the gephyrin-labeled “synapses” assigned to On-Off DSGCs were GABA(A)Rα2 negative (Figure S7). These errors arise primarily from two sources: 1) it is difficult to separate synaptic gephyrin clusters from non-specific background labeling or trafficking vesicles containing gephyrin based on conventional images (Figure S4A) and hence some of the gephyrin clusters assigned to the neuron may not correspond to synapses; 2) synapses near the neuron, but not on the neuron, can be mistakenly assigned to the neuron because of the limited resolution of the conventional images (Figure S4C).

Inhibitory inputs and outputs of a glycinergic interneuron

Last, we imaged gephyrin-positive inhibitory synapses associated with a subtype of narrow-field amacrine cell (NFAC) (Figure 6A and Movie S4), putatively a Type 7 based on previous characterization (Pang et al., 2012). NFACs mediate crossover inhibition between On and Off sublaminae of the IPL via glycinergic inhibition (Werblin, 2010). In contrast to On-Off DSGCs, the surface of this NFAC was highly enriched with paired GlyRα1-4 and gephyrin clusters, but largely depleted of unpaired, GlyRα1-4 negative gephyrin clusters (Figure 6A, B). The resolution of STORM allowed us to visualize the orientations of gephyrin-receptor pairs relative to the neuron surface and determine whether these structures were input synapses onto the cell or output synapses from the cell (Figure 6C). Unlike GABAergic synapses onto On-Off DSGCs, which were all input synapses (Figure 6D, 3C), the glycinergic synapses on the NFAC contain both input and output synapses (Figure 6C, D). Both synaptic inputs and outputs exhibited sublaminar specificity with enrichment in the Off sublaminae (Figure 6E), suggestive of this cell being an On-center responsive Type 7 glycinergic amacrine cell (Pang et al., 2012) providing crossover inhibitory output to the Off sublaminae (Werblin, 2010).

Figure 6. Input and output inhibitory synapses of a NFAC.

(A) Surface rendering of a NFAC (grey) with GlyRα1-4 (+) input (purple circles), output (orange circles), and GlyRα1-4 (-) synapses (green circles) shown. (B) The GlyRα1-4-paired gephyrin (green), gephyrin-paired GlyRα1-4 (magenta), and unpaired gephyrin cluster (blue) densities as a function of the distance to the surface of the neuron. (C) Examples of input and output synapses distinguished by the positions of receptor and gephyrin signals relative to the surface of the neuron. Input synapses have the gephyrin clusters (green) in the dendrites (blue) and receptor clusters (magenta) on the surface. Output synapses have the receptor cluster immediately adjacent to the dendrite surface and the gephyrin clusters farther outside. (D) The relative displacement of receptor and gephyrin clusters from the neuron surface, with positive values indicating receptor being farther from the neuron (input synapses) and negative value indicating gephyrin being farther from the neuron (output synapses). The solid line shows the distribution for glycinergic synapses associated with the NFAC and the dashed line shows the distribution for the GABAergic synapses associated with an On-Off DSGC. (E) Laminar distributions of the input (purple) and output (orange) GlyRα1-4 positive synapses and the GlyRα1-4 negative synapses (blue) on the NFAC. See also Movie S4.

About 85% of the synapses on this neuron contained glycine receptors (Figure 6E), again indicating a high receptor labeling efficiency in our samples. Since NFACs are glycinergic cells, it is not surprising that the observed output synapses from this cell were mostly GlyRα1-4 positive. It is, however, interesting to observe that the majority of gephyrin-positive inputs onto this cell were also GlyRα1-4 positive, suggesting that this type of NFAC receives inhibitory input signals mainly from other glycinergic amacrine cells though our results do not exclude the possibility that this cell type also receives some GABAergic inputs.

Discussion

Mapping the spatial organization and molecular identity of synaptic connections within neuronal networks is important for understanding how the nervous system functions. Here we developed a super-resolution platform for volumetric reconstruction and automated segmentation of endogenous molecular targets in tissue, and demonstrated the ability of this platform to identify the spatial patterns and molecular identity of inhibitory synapses within neuropil as well as onto individual neurons using the mouse retina as a model system.

This method provides several benefits for reconstructing synaptic connectivity. First, the superior resolution of this approach, as compared to conventional fluorescence imaging, allows more accurate identification of synapses and assignment of synapses to neurons. Indeed, when comparing results from the same tissue samples, we found that conventional fluorescence imaging led to substantial errors both in the identification of synapses and in the assignment of synapses to neurons even with the improved z resolution afforded by ultrathin sectioning. These errors resulted in misidentification of inhibitory synaptic types onto On-Off DSGCs, which could lead to substantial misinterpretation of cellular physiology. In addition, the resolution provided by STORM also allowed us to quantitatively measure synapse size, which is often a good indicator of synaptic strength (Nusser et al., 1997; Nusser et al., 1998). This ability allowed us to map the relative strengths of inhibitory synapses at different locations both on identified neurons and across the IPL.

A second benefit of the super-resolution reconstruction platform is its ability to use standard immunohistochemistry for labeling multiple endogenous protein targets of interest, which allows the determination of the molecular identities of synapses. Such information is difficult to ascertain using the EM reconstructions alone, but is important for interpreting the function of specific synapses in neural circuits (Bargmann and Marder, 2013). Taking advantage of this capability, we showed that gephyrin-positive inhibitory synapses onto On-Off DSGCs were overwhelmingly GABAergic and each contained the GABA(A) receptor α2 subunit, suggesting that this receptor subunit is important for generating postsynaptic currents during motion detection. This result is consistent with previous data showing the enrichment of GABA(A)α2 in On-Off DSGC synapses and reduction in direction-selective responses in the GABA(A)α2 knockout mouse (Auferkorte et al., 2012). Our reconstructions also showed that On-Off DSGCs receive little, if any, monosynaptic glycinergic input. These structural data, together with the observations that blocking GABA receptors largely eliminates inhibitory currents in On-Off DSGCs (Stafford et al., 2014; Trenholm et al., 2011), suggest that glycinergic modulation of On-Off DSGCs does not occur via direct glycinergic inputs onto these neurons, but likely through glycinergic inhibition of bipolar cells or SACs that are presynaptic to On-Off DSGCs (Ishii and Kaneda, 2014; Majumdar et al., 2009; Zhang and McCall, 2012).

A third strength of this reconstruction platform is its ability to perform automated segmentation of synaptic connections in neural circuits without manual annotation. This automated analysis capability greatly speeds up the image processing required to extract biological information from individual reconstructions. For example, the image processing for volumetric reconstruction and segmentation of a whole On-Off-DSGC cell and associated synapses took <3 days of computation time without any need for manual segmentation or correction. In this work, the rate-limiting step of our reconstructions was the STORM image acquisition time as imaging an entire On-Off-DSGC of 2.3 × 10⁵ μm³ in four color channels took ∼3 weeks using a STORM setup equipped with an EMCCD camera. Our recent switch to a scientific CMOS (sCMOS) camera with a larger field of view and higher frame rate shortened the imaging time of a comparable volume to ∼3 days. We envision this automated imaging and segmentation pipeline to be beneficial for determining neural circuit properties in different genetic mutant and disease models, or at different time points during development, where a large number of reconstructions are needed.

One potential limitation of this super-resolution fluorescence platform, as compared with EM approaches, is the density of neuronal processes that can be reconstructed within a volume. In this work, we reconstructed the spatial distributions and molecular identities of synapses onto individual neurons in Thy1-GFP/YFP transgenic mice, in which only sparse subsets of neurons are labeled. We expect that our approach can be extended to the reconstruction of multiple, synaptically-coupled neurons using recently developed high-density, high-antigenicity, genetic labeling approaches (Cai et al., 2013; Loulier et al., 2014; Viswanathan et al., 2015) or by microinjection of probes to directly label multiple neurons. Although the image resolution here was limited in the z-direction by the 70 nm section thickness, we anticipate a substantial improvement in z-resolution by using 3D STORM (Huang et al., 2008). In particular, using high-precision z-localization approaches (Jia et al., 2014; Shtengel et al., 2009; Xu et al., 2012), the optical resolution can reach ∼10 nm in all three dimensions. However, this resolution is still lower than that achievable by EM and the labeling density may impose an additional limitation on resolution. Together, these may limit the density of neurites that can be reconstructed, and it remains to be determined whether this STORM platform can be used for dense reconstruction of all neurons in a volume.

With its unique capabilities complementary to existing reconstruction methods, we expect that this volumetric super-resolution reconstruction platform will enable a variety of synaptic connectivity analyses that will substantially enhance our understanding of the structural basis of nervous system function. The ability to reconstruct and identify endogenous molecular targets in large tissue volumes should also benefit the studies of many other biological systems.

Experimental procedures

Animals

Animal work was performed in accordance with protocols approved by the Institutional Animal Care and Use Committee at Harvard University. Adult transgenic mice (Tg(Thy-1-EGFP)MJrs/J or YFP (Tg(Thy1-YFP)HJrs/J, The Jackson Laboratory, Bar Harbor ME) (Feng et al., 2000), both male and female animals 6-24 weeks of age, were used in our experiments.

Retinal tissue preparation

Whole eye-cups were immersion fixed in 4% paraformaldehyde for 10-60 minutes at room temperature. Both whole-mount and vibratome-sectioned retinae were used for labeling. For whole-mount labeling, retinae were laid flat on nitrocellulose membranes and individual labeled neurons were excised in circular punches (diameter ∼500 μm, thickness ∼200 μm). For vibratome section labeling, retinae were immersed in 37°C 2-3% agarose, cooled on ice, and sectioned at 50-150 μm thickness in 1X DPBS.

Immunohistochemistry

Retinae were blocked in 10% normal donkey serum in 1X DPBS with 0.3% Triton X-100 and 0.02-0.05% sodium azide for 2-3 hours at room temperature and incubated in primary antibody solutions diluted in blocking buffer overnight for 3-4 nights at 4°C. A complete list of all primary antibodies tested in this work is provided in Table S1 with the antibodies selected for the STORM reconstructions highlighted. Following primary antibody incubation, retinae were washed 6 times for 20 minutes each in 2% normal donkey serum in 1x DPBS at room temperature and incubated in secondary antibodies (detailed in Supplemental Experimental Procedures) overnight at 4°C for 1-2 nights to label the neuron with photoswitchable dye Atto 488 and two synaptic targets (gephyrin and presynaptic proteins or gephyrin and receptors) with photoswitchable dyes Alexa Fluor 647 and DyLight 750 respectively. The antibodies for labeling synaptic proteins were also conjugated to Alexa Fluor 405 to facilitate photoactivation of Alexa Fluor 647 and DyLight 750. Retinae were then washed 6 times for 20 minutes each in 1x DPBS at room temperature and incubated overnight in Cy3B-labeled WGA.

Postfixation, dehydration, and embedding in epoxy resin

Labeled retinae were postfixed for 2 hours in 3% paraformaldehyde and 0.1% glutaraldehyde diluted in 1x DPBS. Postfixed retinae were dehydrated in a graded series of ethanol washes (50%/70%/90%/100% two times) for 10-20 minutes each and then incubated in UltraBed Epoxy Resin (Electron Microscopy Sciences, Hatfield PA) solutions of increasing concentration for two hours each: (75% ethanol/25% resin; 50% ethanol/50% resin; 25% ethanol/75% resin; 100% resin 2 times). Dehydrated resin blocks were then polymerized in UltraBed overnight for 16 hours at 70°C.

Ultrathin sectioning

Ultrathin sections were cut at 70 nm on a Leica UC7 Ultramicrotome (Leica Microsystems; Buffalo Grove IL) using an ultra Jumbo diamond knife (Diatome; Hatfield PA). The section thickness was verified in two independent ways, as described in the Supplemental Experimental Procedures. Sections were collected on glass coverslips coated with 0.5% gelatin/0.05% chromium potassium sulfate. Coverslips were dried at 60°C for 25 minutes.

Preparation of coverslips for imaging

Coverslips of tissue sections were immersed in 10% sodium ethoxide solution for 5-20 minutes to etch the embedding resin for optimal photoswitching of dyes. Fluorescent beads (mixture of 540/560 and 715/755 FluoSpheres from Life Technologies, detailed in the Supplemental Experimental Procedures) were spotted on the coverslips as fiducial markers. Coverslips were secured to glass slide flow channels, filled with STORM imaging buffer (10% glucose/17.5 μM glucose oxidase/708 nM catalase/10mM MEA/10 mM NaCl/200mM Tris), and sealed with epoxy.

Imaging setup

Imaging was performed through Olympus UPlanSApo 100x 1.4 NA oil-immersion objectives mounted on Olympus IX71 inverted microscopes with back optics arranged for oblique incident angle illumination. The microscope contained a custom pentaband dichroic and pentanotch filter (Chroma Technology Corp, Bellows Falls VT) and laser lines at 488/561/647/750 nm (detailed in the Supplemental Experimental Procedures) for excitation of Atto 488, Cy3B, Alexa Fluor 647 and DyLight 750, respectively. A 405-nm laser was used for reactivation of dyes. Images were acquired on an Andor iXon3 897 or 897Ultra EMCCD camera through a QV2 quadview image splitter (Photometrics; Tucson AZ). Each camera pixel corresponded to ∼158 nm in sample space and the total imaging field size was ∼40 μm × 40 μm. Axial focus during imaging was maintained in an automated manner as described previously (Dempsey et al., 2011).

Automated image acquisition

Tissue sections and fiducial bead fields were initially located using a 4x objective. Regions of interest (ROIs) were subsequently identified with a 100x objective. The stage position coordinates for each ROI were determined, and the position list for all ROIs on a coverslip was then used to generate a master file that controlled laser illumination, camera activation, stage movement, AOTF control, and shutter sequences for automated STORM and conventional imaging. Each imaging session began with imaging of low-density bead fields by first exciting the 540/560 beads at 488 nm and detecting in the Alexa Fluor 647, Cy3B, and Atto 488 channels, and then exciting the 715/755 beads at 752 nm and detecting in the DyLight 750 and Alexa Fluor 647 channels. These low-density bead images were used to for chromatic aberration correction across different color channels.

Next, each ROI was imaged at the conventional resolution in each of the four color channels (DyLight 750, Alexa Fluor 647, Cy3B, and Atto 488). Next, images of the high-density bead field were acquired in each of the four color channels for: 1) flat-field correction to compensate for non-uniform illumination across the field of view and 2) lens distortion correction at image field edges.

STORM imaging of individual ROIs was next performed in four color channels. For each ROI, the DyLight 750 channel was imaged for ∼4K-4.5K frames at 30 Hz, the Alexa Fluor 647 channel was imaged for 6K-7K frames at 60 Hz, and the Cy3B and Atto 488 channels were each imaged for ∼10K frames at 60 Hz. To ensure that overlapping regions in each montage were not bleached, STORM movies were collected in two passes for each ROI, each consisting of half the total number of frames described above.

STORM image analysis

STORM movies were analyzed to determine the positions of individual molecules using a DAOSTORM algorithm (Babcock, 2012; Holden et al., 2011). Molecule lists were rendered as 2D images with 15.8 nm pixel size, which is close to both our ∼20 nm STORM image resolution and 1/10 of the camera pixel size. For consistency of analysis, the conventional images were up-sampled to 15.8 nm/pixel. Chromatic aberrations were corrected using the transformation maps generated from the low-density bead field images and lens-induced optical distortions were corrected using transformation maps generated from the high-density bead field images, as detailed in the Supplemental Experimental Procedures.

Alignment of multiple image tiles within individual sections

Each STORM image was aligned to the corresponding conventional image using two-dimensional cross-correlation (Guizar-Sicairos et al., 2008). For mosaic imaging, Scale-Invariant Feature Transformation (SIFT) (Lowe, 2004) was used to find points of similarity between overlapping regions in adjacent image tiles in the WGA channel and generate a rigid alignment transformation that was applied to the conventional and STORM images to stitch overlapping image tiles. On average the residual offset in alignment between SIFT points of similarity in two adjacent image tiles was < 40 nm.

Alignment of serial sections

Corresponding SIFT features between adjacent sections were used to determine a rigid linear transformation between sections, which was applied to all sections in the dataset to achieve a coarse, 3D rigid alignment of the data. Then, we applied elastic registration (Saalfeld et al., 2012) to further improve the alignment accuracy between adjacent sections while minimizing the global deformation of the entire image block. The warping transforms generated in these steps were applied to all conventional fluorescence and STORM channels.

Segmentation of STORM and conventional fluorescence images

STORM images were first filtered using a mask generated from the conventional images to remove background and signals from occasional debris on the coverslip. To generate this mask, the signals in the conventional images were thresholded using the lower threshold of a two-level Otsu threshold method (Otsu, 1979) that divided the signals in our images into three classes with the lowest-intensity class representing the background, the highest intensity class representing neuronal and synaptic features, and the middle class representing other low intensity signals above background. To identify the surface of the neuron, we smoothed the neuron signal with a Gaussian kernel with σ = 50 nm and then binarized the neuron signal using the lower threshold of the two-level Otsu threshold method. To identify fluorescent clusters in the gephyrin, presynaptic or receptor channels in the STORM images, we applied a 76 nm Gaussian convolution to the signal in the XY plane and an isometric Gaussian convolution (∼1 voxel) in Z, and used the lower threshold of the two-level Otsu threshold method to binarize the image and identify connected components in three dimensions. Additional separation of over-connected clusters was performed using a watershed transformation. Processing of conventional images was performed similarly, except that we binarized the conventional images based on the higher threshold of a two level Otsu threshold.

Two-dimensional (2D) analysis to separate different populations of gephyrin and presynaptic clusters

To determine whether a given cluster was synaptic, two parameters were considered for each cluster in the gephyrin and presynaptic channels: the volume of the cluster was calculated from the connected components within the segmented image. Second, the signal density was measured as the fraction of volume of the connected components that was occupied by signal-positive voxels in the raw data. For STORM images, plotting the distribution of these two parameters constructed from all clusters in the data set as a 2D histogram showed two peaks. Separation of the two populations is described in the Supplemental Experimental Procedures.

Ripley's K function

The Ripley's K function is calculated as K(t) = λ⁻¹ Σ_i≠j I(d_ij < t)/n, where t is the distance along neurites, λ is the average density of synapses on the neuron skeleton, I is the indicator function, d_ij is the distance between the i^th and j^th synapses, and n is the number of synapses on the neuron. ū is the average of K(t) derived from 1000 randomizations of synapse positions on the surface of the dendritic arbor.

A detailed complete description of the experimental procedures can be found in the Supplemental Experimental Procedures accompanying this paper.