SNT: A Unifying Toolbox for Quantification of Neuronal Anatomy
Cameron Arshadi
1
, Ulrik Günther
2,3,4
, Mark Eddison
1
, Kyle I. S. Harrington
1,5,6
, Tiago A. Ferreira
1
1
Janelia Research Campus, Howard Hughes Medical Institute, Ashburn, Virginia, USA
2
CASUS – Center for Advanced Systems Understanding, Görlitz, Germany
3
Max Planck Institute of Molecular Cell Biology and Genetics, Dresden, Germany
4
Center for Systems Biology, Dresden, Germany
5
Virtual Technology & Design, University of Idaho, Moscow, Idaho, USA
6
Max Delbrück Center for Molecular Medicine in the Helmholtz Association, Berlin, Germany
Summary
Quantification of neuronal morphology is essential for understanding neuronal connectivity and many
software tools have been developed for neuronal reconstruction and morphometry. However, such tools
remain domain-specific, tethered to specific imaging modalities, and were not designed to accommodate the
rich metadata generated by recent whole-brain cellular connectomics. To address these limitations, we created
SNT: a unifying framework for neuronal morphometry and analysis of single-cell connectomics for the widely
used Fiji and ImageJ platforms.
We demonstrate that SNT can be used to tackle important problems in contemporary neuroscience,
validate its utility, and illustrate how it establishes an end-to-end platform for tracing, proof-editing,
visualization, quantification, and modeling of neuroanatomy.
With an open and scriptable architecture, a large user base, and thorough community-based
documentation, SNT is an accessible and scalable resource for the broad neuroscience community that
synergizes well with existing software.
Quantification of neuronal anatomy is essential for mapping information flow in the brain and classification of cell types in
the central nervous system. Although digital reconstruction (“tracing”) of the tree-like structures of neurons —axons and
dendrites— remains a laborious task, recent improvements in labeling and imaging techniques allow faster and more
efficient reconstructions, with neuroscientists sharing more than 140,000
1
reconstructed cells across several databases.
Powerful toolboxes have been developed for neuronal morphometry (Sup. Information). However, such tools can be
tethered to specific imaging modalities or remain specialized on specific aspects of neuroanatomy workflows. To address
these limitations we established a unifying framework for neuronal morphometry and analysis of single-cell connectomics
for the widely used Fiji and ImageJ platforms
2,3
.
We re-invented the popular Simple Neurite Tracer program
4
to create an open-source, end-to-end solution for semi-
automated tracing, visualization, quantitative analyses and modeling of neuronal morphology. All aspects of our software,
named SNT —Simple Neurite Tracer’s popular moniker— can be controlled from a user-friendly graphical interface or
programmatically, using a wide variety of scripting languages (Fig. 1a).
For semi-automated tracing we implemented a host of new features (described in Sup. Information), including support for
multi-channel, and time-lapse images, optimized search algorithms and image processing routines that better detect
neuronal processes, and made possible to reconstruct simple morphologies directly from thresholded images. With time-
lapse sequences, traced paths can be automatically matched across frames so that growth dynamics of individual neurites
can be monitored across time. (Fig. 1b). To expedite the proof-editing of traced structures, SNT allows users to edit, tag,
sort, filter, and rank traced segments, using either ad-hoc labels or morphometric traits. Altogether, these features improve
reconstruction accuracy and tracing efficiency.
For visualization, SNT features an interactive 3D viewer dedicated to neuron morphology —Reconstruction Viewer— that
is hardware accelerated, supports rendering of meshes and detailed annotation of morphometry data. In addition, SNT also
integrates with sciview
5
,
a visualization tool for mesh-based data and arbitrarily large image volumes, supporting virtual,
and augmented reality (Sup. Information).
.CC-BY-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted December 7, 2020. ; https://doi.org/10.1101/2020.07.13.179325doi: bioRxiv preprint
SNT: A Unifying Toolbox for Quantification of Neuronal Anatomy Page 2 of 18
Figure 1 | SNT as an end-to-end platform for data retrieval, visualization, quantification, and modeling of neuroanatomical data.
(a) Schematic diagram of the software. (clockwise): i) SNT is powered by the stack of ImageJ-based software, including: Fiji, ImageJ2,
sciview, SciJava, ImgLib2, TrakEM2 and pyimagej. ii) Reconstructions can be obtained directly from thresholded images or using semi-
automated procedures that support time-lapse and multi-channel light-microscopy imagery. iii) Once center-line reconstructions
(“tracings”) are obtained, they can be conveniently processed in subsequent image processing routines. iv) Dedicated neuroanatomy
viewers allow for effective quantitative visualization of complex data. v) In addition to single-cell morphometry, vi) circuit analyses are
facilitated through support of several online databases and reference brain atlases (Drosophila, mouse and zebrafish). vii) Biophysical
modeling of neuronal growth is achieved through Cx3D integration (Sup. Information). viii) Users may use SNT has a standalone
interactive program or as a multi-language scripting library.
(b—d) ImageJ interoperability allows for complex data retrieval. (b) Static frame from a non-fluorescent time-lapse video monitoring
the development of neuron polarity in a hippocampal neuron growing in vitro
19
. The 12 highlighted neurites were traced throughout the
video sequence and color coded across time as per color ramp. Insert details growth of neurite #1 at selected time-points. Plot depicts
growth dynamics of individual neurites across time. (c) Multichannel image of a hippocampal neuron stained in vitro for the presynatic
markers VGluT1-2 (green) and the postsynaptic NMDA receptor (magenta)
18
. Dendrites were traced (orange) and intensity profiles
obtained directly from the tracings. Profiled maxima from the marked region depict synaptic locations. ( d) Maximum intensity
projection of a three-channel 3D image of a gelled-brain section processed for expansion fluorescence in situ hybridization (ExFISH).
Dendrites of GFP-labeled neurons (green) were traced in SNT (center-lines for three cells displayed in orange, red, and yellow). Foci
reporting on Somatostatin (SST) mRNA (magenta) were detected on neighboring somata, segmented from a counterstain for total RNA
(cyan). Point ROIs reporting foci (circles) were labeled with the same hue of the closest traced cell. All procedures performed within
ImageJ. Right: Violin plot of SST expression for segmented cells in the sub-volume (N=147). Scale bars: (b,c): 10µm; (d): 20µm.
.CC-BY-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted December 7, 2020. ; https://doi.org/10.1101/2020.07.13.179325doi: bioRxiv preprint
SNT: A Unifying Toolbox for Quantification of Neuronal Anatomy Page 3 of 18
Figure 2 |Comprehensive analytical tools enable discovery biology. How many brain areas does a neuron connect to in the mouse
brain? The MouseLight (ML) database was programmatically parsed in SNT and the number of mid-ontology brain regions innervated by
individual axons retrieved under two criteria: (normalized) cable length, and number of axonal endings at target area.
(a) Frequency histogram of number of brain areas innervated by single cells (N=1094). ( b) Ranked examples of axonal
connectivity
rendered in SNT’s Reconstruction Viewer. Left: SNr neuron projecting exclusively to the CP (ML id: AA1044). Its single axonal tuft can
cover as much as 4% of the ipsilateral lobe of the target area. Center: SS neuron projecting to many areas in the isocortex (ML id:
AA0100). Right: Pyramidal-tract neuron projecting to many areas in the isocortex, midbrain and hindbrain (ML id: AA0788). Dendrites
are depicted in black and axonal arbors color-coded by “path distance to soma”, as per color ramp legend. Selected brain regions are
depicted according to color-coded abbreviations (MO: Somatomotor areas; SS: Somatosensory areas; CP: Caudoputamen; TH: Thalamus;
LHA: Lateral hypothalamic area; SNr: Substantia nigra, reticular part; MB: Midbrain; TEa: Temporal association areas; ECT: Ectorhinal
area; PIR: Piriform area; MY: Medulla). (c) Connectivity diagrams for the three chosen exemplars programmatically generated in
SNT’s Graph Viewer. In this “Ferris wheel” diagram, the neuron’s target areas are displayed around the brain area associated with the cell
soma (SNr and MOs [Secondary motor area]), with connecting edges indicating projection strength, and self-connecting edges depicting
local innervation. Here, edges were scaled and color-coded according to the two criteria used in a), as per color ramp legend. This
representation can also be extended to cell populations (Sup. Fig 7). When generating such diagrams, SNT automatically sorts target
areas by projection strength and groups them by parental ontology (labeled in external arcs). As a reference, the total axonal length of
each cell is: 28.598cm (AA1044), 44.649cm (AA0100), 18.454cm (AA0788). Abbreviations reflect Allen Mouse Brain Common
Coordinate Framework
51,52
nomenclature.
.CC-BY-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted December 7, 2020. ; https://doi.org/10.1101/2020.07.13.179325doi: bioRxiv preprint
For data retrieval, SNT provides seamless integration with the ImageJ platform, and thus tracing and reconstruction analyses
can be intermingled with image processing workflows. To exemplify this, we used SNT to quantify challenging datasets:
expression of synaptic markers along dendrites (Fig. 1c), and fluorescent in situ hybridization (FISH) imaging of mRNA in
the same volume in which dendrites of pyramidal cells were reconstructed (Fig. 1d).
Another strength of SNT is that it can connect directly to all the major neuroanatomy databases, including FlyCircuit
41
,
InsectBrain
6
, MouseLight
7
, NeuroMorpho
1
, and Virtual Fly Brain
8
, supporting several multi-species brain atlases
(Drosophila, mouse, zebrafish, Fig. 1a, Sup. Information). While such online databases are highly queryable, they remain
constrained by website design limitations. Scripting frameworks that can programmatically parse their data bypass those
restrictions, facilitate data sharing, scientific reproducibility, bridge isolated data repositories, and promote the development
of new tools and features. Importantly, since SNT adopts the SciJava framework
9,10
, it can be scripted using popular
computer languages such as Python (and Jupyter notebooks) through pyimagej, Clojure, Groovy, JavaScript, Jython,
MATLAB
TM
,
R, Ruby, or Scala.
To demonstrate the analytical power of SNT, we parsed the MouseLight database, currently containing the most complex
reconstructions described in the literature
7
. In particular, we focused on quantifying the repertoire of strategies adopted by
individual cells to broadcast information across the brain. Since no synaptic strengths are currently known for MouseLight
neurons, projection strength to target areas must be inferred from morphometric surrogates. In a programmatic, unbiased
approach, we used two morphological criteria (normalized cable length and number of axonal endings) to retrieve the
number of anatomical brain areas innervated by individual axons (Fig. 2a). In doing so, we identified two extremes of
connectivity: cells that connect exclusively to a single projection brain area, and cells that project broadly over a multitude
of brain areas (Fig. 2b). A key feature of SNT is the ability to generate streamlined connectivity diagrams, holding
quantitative information determined from the intersection or union of multiple morphometric criteria that can be customized
using SNT’s interactive tool Graph Viewer. These type of diagrams can be generated at the single-cell level (Fig. 2c), or
from cell populations (Fig. S7), and are a valuable visualization tool for connectomics
11
.
SNT provides support for generative models of artificial neurons by utilizing the neurodevelopmental simulation
framework, Cx3D
12
. This not only provides capabilities for the algorithmic generation of neuronal morphologies, but
enables a new direction of image-based modeling for cellular neuroscience. On the latter, we provide a proof-of-concept
example, where artificial neurons are seeded in an image derived from an in vitro chemotaxis assay (Sup. Video 3). On the
former, we challenged SNT’s ability to morphometrically distinguish closely-related reconstructions. First, we generated
different mathematical gene-regulatory networks (GRNs)
13
capable of controlling neural growth by regulating extension,
branching, and directionality of neurites to define in silico morphologies. Second, we generated thousands of artificial
neurons constrained by these patterns. Third, we used built-in metrics
14,15
to statistically differentiate between these
computer-generated “neuronal types”. We found we could distinguish with high confidence all of the morphological classes,
including those closely related (Sup. Information). Altogether, these experiments demonstrate how SNT can bridge
experimental and modeled data to support model evaluation for both inference and predictive modeling.
In summary, SNT is a powerful tool for tracing, proof-editing, visualization, quantification, and modeling of neuroanatomy.
It is based on recent technologies, supports modern microscopy data, integrates well with the ImageJ platform, interacts
with major online repositories, and synergizes with post-reconstruction analysis software, and recent data-mining
frameworks
16,17
. With a large user base and thorough community-based documentation (https://imagej.net/SNT), SNT is an
accessible, scalable and standardized framework for efficient quantification of neuronal morphology.
Methods
The figures and analyses from this manuscript can be generated programmatically. See https://github.com/morphonets/
SNTManuscript for details.
Programming. SNT was programmed with Eclipse Java IDE 4.4–4.16 (Eclipse Foundation), IntelliJ IDEA 2020
(JetBrains), and Fiji’s built-in Script Editor on an Intel i7 laptop running Ubuntu 18.10–20.04.
Cell Image Library Imagery. Analyses were performed manually from SNT’s GUI with the following modifications to the
SNT: A Unifying Toolbox for Quantification of Neuronal Anatomy Page 4 of 18
.CC-BY-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted December 7, 2020. ; https://doi.org/10.1101/2020.07.13.179325doi: bioRxiv preprint
original images: CIL810
18
(an RGB image) was converted to a multi-channel composite; CIL701
19
(an unannotated Z-series)
was converted to a time-series stack. Please refer to the original publications for details on the datasets.
ExFISH. Histology: Expansion fluorescence in situ hybridization (ExFISH) was performed using an optimized protocol not
yet published (manuscript in preparation). In short: 150µm-thick cortical slices of an heterozygous Thy-1-GFP-M
20
adult
mouse were embedded in a hydrogel using standard embedding procedures for thick tissue
21
. mRNA was detected using
HCR 3.0
22
. SST probes and fluorescent hairpins were obtained from Molecular Instruments (molecularinstruments.com).
Gelled sections (~2× expanded) were imaged in PBS on a commercial Zeiss Z1 lightsheet microscope. Spot quantification:
Signal from total RNA labeling was segmented using Labkit
23
. Individual cells were then masked using watershed filtering
and labeling of connected components using MorphoLibJ
24,25
. Ill-segmented somata were manually eliminated with the aid
of BAR tools
26
. Spot density (no. of spots per cell) of SST signal was determined by iteratively running
3DMaximaFinder
27,28
at locations of each connected component. It should be noted that this approach is rather elementary: it
was designed as a proof-of-principle image processing routine that can be performed mid-way through a tracing session
using accessible ImageJ tools.
MouseLight Single-cell Connectivity. MouseLight (ML) database was programmatically parsed to obtain the number of
brain areas (Allen Mouse Common Coordinate Framework (CCF) compartments) associated with individual axonal arbors
using two criteria: 1) normalized cable length and 2) number of axonal endings at target area). Only CCF compartments of
ontology depth 7 with public meshes available were considered. For 1) axonal length within an anatomical compartment
was measured by taking all nodes within the compartment and summing the distances to their parent nodes. To exclude
enpassant axons, a cell was considered to be associated with the compartment if such length would be at least 5% of the
compartment’s bounding-box diagonal. For 2) only neurons with at least two end-points were considered. Selected
examples in Fig. 1 were chosen by sorting cells by number of associated areas, and selecting those with the largest axonal
cable length (AA1044: 28.598cm, AA0100: 44.649cm, AA0788: 18.454cm).
Tracing and Path Fitting Benchmarks. Tracing benchmarks and fitting procedures were performed programmatically and
can be reproduced using the scripts available at https://github.com/morphonets/SNTmanuscript. DIADEM scores
29
were
computed with default thresholds and retrieved in “post-DIADEM competition” mode. For degradation of traces (Fig. S2),
each node in the reconstruction was displaced to a random position within a 1µm neighborhood around each axis.
Synthetic Morphologies. Chemoatraction assay (Video S3): Code accessible from github.com/morphonets/
SNTManuscript. GRNs (Fig. S6): The code for generating GRNs is available at github.com/morphonets/cx3d/, and the five
GRNs used in this study are made available at github.com/morphonets/SNTManuscript (together with remaining analysis
scripts). Tools for inspecting GRNs are available at github.com/brevis-us/grneat. Morphometric analysis: Default metrics
provided by SNT (41 as of this writing
1
) were retrieved for all artificial neurons. Data was normalized and analyzed using
PCA (Principal Component Analysis), t-SNE
30
(t-Stochastic Neighbor Embedding), and UMAP
31
(Uniform Manifold
Approximation and Projection). Group comparisons on principal components, t-SNE features, and UMAP components were
performed using two-sample Kolmogorov-Smirnov tests adapted for multivariate data. p–values were combined using
Fisher’s combined probability test. Density maps and examplars: soma-aligned cells were skeletonized, and their skeletons
projected into the XY plane using SNT’s core functionality. Binary masks of skeletons were then summed up and resulting
image normalized to the number of cells. Exemplars were chosen from a random pool of 10 cells.
Code Availability
SNT source code is available at github.com/morphonets/SNT. The source code for the figures and analyses described in this
manuscript is available at github.com/morphonets/SNTManuscript. Both are released under the GNU General Public
License v3.0.
1 A subset of all available metrics: Average branch length, Average contraction, Average fractal dimension, Average fragmentation, Average partition asymmetry, Average
remote bif. angle, Cable length, Depth, Height, Highest path order, Horton-Strahler bifurcation ratio, Horton-Strahler number, Length of inner branches (sum), Length of
primary branches (sum), Length of terminal branches (sum), Mean radius, No. of branch points, No. of branches, No. of inner branches, No. of nodes, No. of primary
branches, No. of terminal branches, No. of tips, Width; [Sholl-based metrics]: Centroid, Centroid radius, Decay, Degree of polynomial fit, Enclosing radius, Intercept,
Kurtosis, Max, Max (fitted), Max (fitted) radius, Mean, Median, No. maxima, No. secondary maxima, Skeweness, Sum, Variance. Refer to user documentation for details.
SNT: A Unifying Toolbox for Quantification of Neuronal Anatomy Page 5 of 18
.CC-BY-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted December 7, 2020. ; https://doi.org/10.1101/2020.07.13.179325doi: bioRxiv preprint
![Figure 2 |Comprehensive analytical tools enable discovery biology. How many brain areas does a neuron connect to in the mouse brain? The MouseLight (ML) database was programmatically parsed in SNT and the number of mid-ontology brain regions innervated by individual axons retrieved under two criteria: (normalized) cable length, and number of axonal endings at target area. (a) Frequency histogram of number of brain areas innervated by single cells (N=1094). ( b) Ranked examples of axonal connectivity rendered in SNT’s Reconstruction Viewer. Left: SNr neuron projecting exclusively to the CP (ML id: AA1044). Its single axonal tuft can cover as much as 4% of the ipsilateral lobe of the target area. Center: SS neuron projecting to many areas in the isocortex (ML id: AA0100). Right: Pyramidal-tract neuron projecting to many areas in the isocortex, midbrain and hindbrain (ML id: AA0788). Dendrites are depicted in black and axonal arbors color-coded by “path distance to soma”, as per color ramp legend. Selected brain regions are depicted according to color-coded abbreviations (MO: Somatomotor areas; SS: Somatosensory areas; CP: Caudoputamen; TH: Thalamus; LHA: Lateral hypothalamic area; SNr: Substantia nigra, reticular part; MB: Midbrain; TEa: Temporal association areas; ECT: Ectorhinal area; PIR: Piriform area; MY: Medulla). (c) Connectivity diagrams for the three chosen exemplars programmatically generated in SNT’s Graph Viewer. In this “Ferris wheel” diagram, the neuron’s target areas are displayed around the brain area associated with the cell soma (SNr and MOs [Secondary motor area]), with connecting edges indicating projection strength, and self-connecting edges depicting local innervation. Here, edges were scaled and color-coded according to the two criteria used in a), as per color ramp legend. This representation can also be extended to cell populations (Sup. Fig 7). When generating such diagrams, SNT automatically sorts target areas by projection strength and groups them by parental ontology (labeled in external arcs). As a reference, the total axonal length of each cell is: 28.598cm (AA1044), 44.649cm (AA0100), 18.454cm (AA0788). Abbreviations reflect Allen Mouse Brain Common Coordinate Framework51,52 nomenclature.](/figures/figure-2-comprehensive-analytical-tools-enable-discovery-2tzw47uk.webp)
