Synthesis and 5α-Reductase Inhibitory Activity of C21 Steroids Having 1,4-diene or 4,6-diene 20-ones and 4-Azasteroid 20-Oximes
A 16 terabyte retinal connectome dataset computationally assembled from automated transmission electron microscope imaging has been explored with large-scale, multi-user annotation tools, revealing novel retinal signaling modes, new network architectures, and microglial activity sensing. Abstract: 112 words Text: 2199 words Abstract A connectome is the complete description of synaptic connectivity for a defined neural domain. A 16 terabyte mammalian retinal connectome dataset (RC1) has been
... (RC1) has been assembled at 2 nm resolution using automated transmission electron microscope (TEM) image acquisition, large-scale mosaicking and volume registration. RC1 spans the retinal inner nuclear, inner plexiform and ganglion cell layers, and is augmented by multichannel small molecule and activity markers. Exploration of RC1 with our multi-user annotation software (Viking) reveals novel signaling modalities such as vetting synapses, nanoribbon synapses, extreme network complexity, microglial activity sensing and new glial organelle patterning. Connectome assembly and analysis is now practical for wiring discovery in normal and disordered neural systems. Connectome analysis has the potential to be a Rosetta Stone for neuroscience by decoding the wiring of any brain region (1, 2). It is motivated by the idea that existing models of biological neural networks and heterocellular organization are incomplete. But few neuroscientists have had access to practical tools for acquiring, assembling, visualizing and analyzing connectomes. We recently described a framework for TEM imaging of large-scale neural samples (3) and here demonstrate the assembly and navigation of a representative connectome dataset. True connectome analysis requires a canonical sample of neurons (3), cell classification with complete coverage (4), and imaging resolution sufficient to track all connections including small synapses and gap junctions. This sets imaging resolution at ≈2 nm/pixel. The large size of a such datasets (3, 5) also demands high-speed acquisition. Further, analysis of complex neuronal populations requires phenotyping by molecular markers (6). All these needs are met by automated TEM imaging. We assembled a connectome for the rabbit retina by combining TEM imaging (3), computational molecular phenotyping (CMP) (4, 6) and excitation mapping (6-10). CMP provides robust Anderson et al. Exploring the Retinal Connectome :: Page 2 of 26 segmentation of mammalian neuronal classes and glia based on established molecular signatures (11), and excitation mapping embeds activity-driven reporter signals into the dataset. A 0.25 mm diameter serial-section tissue column (3) spanning the retinal inner nuclear, inner plexiform and ganglion cell layers was automatically imaged by TEM at a resolution of 2.18 nm/pixel, yielding over 341,000 image tiles in a 16.4 terabyte volume captured over 5 months at 3000 images/day. The tiles were automatically mosaicked into sections and section mosaics were automatically aligned into a volume using the NCRToolset (3). The volume was bounded by ultrathin, multichannel CMP datasets probed for the small molecules glutamate, glutamine, glycine, taurine, γ aminobutyrate (GABA) and the excitation marker 1-amino-4-guanidobutane (AGB). These channels were superimposed onto the TEM imagery to classify neurons, glia and microglia (Fig. 1) . The image column was also intercalated with sections probed for single markers every 30 slices (Fig. S1 ). The sample for the volume RC1 was chosen from a library of rabbit retinas that had been excitation mapped with AGB, a channel permeant organic cation strongly selective for glutamate receptor-channel complexes (7-9). AGB mapping provides single-cell synaptic excitation histories following in vivo visual stimulation (10). Volume RC1 was activated by a photopic 3 Hz flickering uniform field of three yellow pulses followed by one blue pulse for 90 minutes in vivo, yielding a wide range of neuronal activation strengths in different classes of neurons ( Fig . 1D,E,F) (6). Combining excitation mapping with CMP of intrinsic signals allows segmentation of neuronal populations into distinct biological classes, which simultaneously aids neuronal process tracking and provides independent validation of cell classifications based on connectivity. Thus we determined that RC1 contains 284 bipolar cells, 167 GABAergic and 118 glycinergic amacrine cells, over 350 Müller glia, 18 validated ganglion cells, 19 horizontal cells and 76 putative microglia. Further segmentation of RC1 with k-means clustering and PCA yields 37 molecular classes of cells: 2 hori-Anderson et al. Excitation mapping, tissue harvest and processing. The retinal sample for the transmission electron microscope (TEM) image volume RC1 was taken from a light-adapted female Dutch Belted rabbits (Oregon Rabbitry, OR) after in vivo excitation mapping (1-3). The animal was tranquilized with intramuscular ketamine/xylazine and deeply anesthetized with intraperitoneal 25% aqueous urethane. The eye was topically anesthetized with 1% lidocaine in 0.1% NaCl 10 minutes prior to intravitreal injection with 0.1 ml of 130 mM 1-amino-4-guanidobutane (AGB) sulfate with a 23 gauge pressure relief needle at the limbus. The rabbit was positioned between two LCD computer monitors and exposed to 90 minutes of flickering 3 hz square wave stimulation of 50% duty cycle in a pattern of one blue and three yellow pulses with a corneal flux density of 9.1 x 10 3 quanta/ sec/cm 2 at 440 nm for the blue stimulus; and dual peaks of 12.5 x 10 3 quanta/sec/cm 2 at 540 nm and 11.6 x 10 3 quanta/sec/cm 2 at 620 nm for the yellow stimulus. The rabbit received a final urethane overdose and was euthanized by thoracotomy in accord with University of Utah Institutional Animal Care and Use Committee guidelines, consistent with the Association for Research in Vision and Ophthalmology statement for the use of animals in research. The eyes were then immediately injected with 0.1 ml fixative and an additional 18 gauge needle pressure relief, enucleated, hemisected, and fixed 24 hours in 1% formaldehyde, 2.5% glutaraldehyde, 3% sucrose, 1 mM MgSO4, in 0.1 M cacodylate buffer, pH 7.4. The eye was dissected and isolated retinal pieces osmicated 60 min in 0.5% OsO4 in 0.1M cacodylate buffer, processed in maleate buffer for en bloc staining with uranyl acetate, and processed for resin embedding as described in Marc and Liu (4). The popular osmium-ferrocyanide method for enhancing TEM contrast was not used, as it quenches immunoreactivity and produced nonuniform staining. Retinal pieces were remounted in resin for serial sectioning in the horizontal plane through the inner nuclear layer, inner plexiform layer and ganglion cell layer (5, 6). Serial sections were cut at 70 nm on a Leica UC6 ultramicrotome onto carboncoated Formvar® films on gold slot grids. For volume RC1, optical thin sections were captured on multi-spot Teflon®-coated slides (Cel-Line; Erie Scientific Inc), and processed for computational molecular phenotyping (CMP) as previously detailed (7). After sodium ethoxide etching, the sections were probed with anti-hapten IgGs targeting γ aminobutyrate (GABA), glycine, glutamate, AGB, glutamine or taurine (Signature Immunologics Inc, Salt Lake City, UT) and visualized with silver-intensified 1.4 nm gold granules conjugated to goat anti-rabbit IgGs (Nanoprobes, Yaphank, Anderson et al. Exploring the Retinal Connectome :: Page 16 of 26 NY). Optical (8-bit 1388 pixel x 1036 line frames) images were captured, mosaicked, aligned and processed for classification as previously described (7) . The RC1 volume was initiated and terminated (capped) with 10-section optical CMP series and intercalated every 30 sections with one CMP section. The annotated grid list is available from: http://prometheus.med.utah.edu/marclab/connectome/gridlist.txt. Volume assembly. RC1 was created as described in . The center of a canonical field 250 μm in diameter was identified in each grid using SerialEM (8) and captured as an array of image tiles at roughly 950-1100 tiles/section. The capture took 5 months, over 341,000 individual captures and requires 16 terabytes of active storage. Each image has 15% area overlap with its neighbors. Image mosaics and volumes were generated using the NCR Toolset (http://www.sci.utah.edu/download/ncrtoolset). SerialEM metadata position information used by the NCR Toolset application ir-translate produces precise initial image mosaics which are refined by ir-refine-grid to correct for image aberrations. Slice-to-slice TEM image registration is automated by ir-stos-brute, ir-stos-grid. CMP-to-TEM registrations are operator-guided with ir-tweak. Volume integrity. As RC1 was manually sectioned and stained, it contains defective regions such as gaps, folds, cracks and stain artifacts. An unedited movie of those defects is available at most cells can be traced effectively even with gaps and defects. Concerns that manual sectioning might not effective in forming connectome datasets are not valid. Improvements may be achieved with automated sectioning, but absence of automated section tools should be no barrier to connectomics research. Image Viewing and Annotation. RC1 was viewed and annotated with Viking, originally named NGVV in Anderson et al. (7) . In developing Viking, we realized that we did not need (nor desire) full volume 3D viewing, but rather the ability to display and annotate one section at a time with guidepost markers from annotations above and below, paging through the data like a book. The Viking viewer is based on dynamically applying the computed slice-to-slice transforms to the image region desired by the user. This decreases storage substantially compared to creating a transformed volume and storing the raw data; allows repair and extension of the volume; and allows multiple users to access multi-terabyte dataset over a network, rather than requiring local copies. Further, users may toggle between between transformed and non-transformed views or choose dif-Anderson et al. Exploring the Retinal Connectome :: Page 17 of 26 ferent reference sections for any transform. On startup, Viking points at a website containing the desired volume data, downloads all section-to-section transforms, and generates section-to-volume transforms for each section. Viking navigates between full-field of view and maximal resolution views and pans the image by dynamically transforming only the necessary tiles from any section using GPU processing. The details of this are available from Anderson et al. (in review). Image tiles are loaded directly into texture memory as 8-bit greyscale images and downloaded over HTTP as needed. Viking calculates which tiles should be visible and the resolution for display. Multiple users concurrently annotate data volumes using Viking. The annotation database is stored on a Microsoft SQLExpress server and exposed via HTTP using a Windows Communication Foundation web service. The web service returns plain text XML strings with objects formatted in Java Script Object Notation for compatibility with different software clients. Users navigate and annotate with keystroke, mouse and menu options. While very rich neural ontologies have already been developed (9), annotation is best done with a small set of markup tags, as these can later be ported to tools based on rich schemata. Our annotation schema was designed to be flexible and not restricted to circuitry, allowing users to define their own ontologies. We achieved this by placing annotations in the context of both graph and tree data structures. Annotating RC1 is based on a parent structure of the 'Cell' type, with numerous child structures for cell-cell contacts (e.g. synapses, gap junctions, adherens junctions) and key intracellular features. Each structure instance receives a unique identification number and, as a structure may span many sections, the geometry and position of an instance is stored in a related location table. Tracking a cell involves populating this table with all of its locations in the volume. Importantly, locations are stored in section space, not volume space. This allows later refinements of the section-to-volume transformation while preserving existing annotations in the database. Finally, location links store adjacency information for graphing the physical layout and connectedness of a structure in the volume. This is essential as cell processes frequently bifurcate and travel wavelike paths, resulting in many location appearances of a structure instance on a single section. This also permits modeling queries such as travel distance between two locations in a cell and is used for 3D renderings. Links between cells via contacts are also stored in a structure link table for summarizing circuitry. Finally, 3D renderings are managed by VikingPlot, a compiled Matlab application that queries structure information from the annotation database and renders surfaces for display. Anderson et al. Exploring the Retinal Connectome :: Page 18 of 26 Data sharing. The primary software resources for this project are available as free (SerialEM) or open-source (NCRToolset) applications or via a free license (Viking) for educational use through the University of Utah Scientific Computing and Imaging Institute software site mentioned above. The RC1 dataset is also available through an educational license on user-provided storage media. The connectivity diagrams from annotation of RC1 are available on the Marclab website (http://prometheus.med.utah.edu/) and the data from the structure tables for modeling will be provided upon request. Image preparation. The procedures for preparing publication figures from raw image data followed those detailed in . All of the raw optical image data are available upon request. Multi-modal registered optical images were max-min contrast stretched and sharpened using unsharp masking at a kernel extent of roughly 540 nm. TEM images from the NCRToolset process tended to have high contrast and were softened by adjusting the gamma to 1.2-1.3. Contrasts were adjusted to match brightness histograms across sections and images were sharpened using unsharp masking with 1-3 pixel kernels (2-6 nm). Overlay methods for combining optical and TEM images generally computed HSB values for a new image using the TEM greyscale brightness (B) and hue and saturation from (H,S) from the rgb optical image. Occasionally, fourth or fifth channels were added using alpha blending (e.g. Fig. 1D ). Renderings of structures in Viking were created in Matlab 2009a. The annotation system stores a point and a radius to describe the largest circle which can fit inside a cell on each section. Annotations are linked to their neighbors on adjacent sections with a graph structure. To render we drew a cylinder between each pair of linked annotation circles using the Matlab patch function and Phong lighting. Annotation circles with only two links were tilted from the XY plane at an angle one half of the total angle between the linked locations on the axis normal to the plane described by all three annotations. Circles were chosen because of the speed of the annotation user interface. As a result the renderings are an approximation of the cell but do encapsulate the full morphology available from the EM volume. In practice, we found it preferable to display cells in an anisotropic fashion, with stretched scaling for inner plexiform layer depth relative to planar scaling. Thus the cells appear ovoid in renderings rather than round.