Neuroinform (2018) 16:3–13

https://doi.org/10.1007/s12021-017-9341-1

ORIGINAL ARTICLE

A Topological Representation of Branching Neuronal

Morphologies
Lida Kanari1 · Paweł Dłotko2 · Martina Scolamiero3 · Ran Levi4 ·
Julian Shillcock1 · Kathryn Hess3 · Henry Markram1

Published online: 3 October 2017

© The Author(s) 2017. This article is an open access publication

Abstract Many biological systems consist of branching distinct morphological groups. The use of this mathemati-
structures that exhibit a wide variety of shapes. Our under- cally rigorous method will advance our understanding of the
standing of their systematic roles is hampered from the start anatomy and diversity of branching morphologies.
by the lack of a fundamental means of standardizing the
description of complex branching patterns, such as those of Keywords Topological data analysis · Neuronal
neuronal trees. To solve this problem, we have invented the morphologies · Branching morphology · Clustering trees
Topological Morphology Descriptor (TMD), a method for
encoding the spatial structure of any tree as a “barcode”, a
unique topological signature. As opposed to traditional mor- Introduction
phometrics, the TMD couples the topology of the branches
with their spatial extents by tracking their topological evo- The analysis of complex branching structures, such as
lution in 3-dimensional space. We prove that neuronal trees, branched polymers (Alexandrowicz 1985), viscous finger-
as well as stochastically generated trees, can be accurately ing (Agam et al. 2002), and fractal trees (Mandelbrot and
categorized based on their TMD profiles. The TMD retains Freeman 1983), is essential for understanding a great vari-
sufficient global and local information to create an unbi- ety of physical and biological processes. For example, the
ased benchmark test for their categorization and is able to fundamental units of the nervous system, neurons, possess
quantify and characterize the structural differences between highly ramified arborizations (Jan and Jan 2010) that are
thought to reflect their involvement in different compu-
tational tasks (Cuntz et al. 2007; Zomorrodi et al. 2010;
Van Elburg and Van Ooyen 2010; Ferrante et al. 2013). In
Electronic supplementary material The online version of this order to understand the properties of branching morpholo-
article (https://doi.org/10.1007/s12021-017-9341-1) contains sup- gies we must study the differences between distinct arbor
plementary material, which is available to authorized users.
types. Much effort has therefore been devoted to grouping
morphologies into distinct classes (DeFelipe et al. 2013;
 Lida Kanari Markram et al. 2004; The Petilla Interneuron Nomenclature
[email protected] Group P 2008), a categorization process that is important
in many fields (Lyons et al. 1999). However, an efficient
1 Blue Brain Project-EPFL, Geneva, Switzerland method for quantitatively analyzing the morphology of such
2 Department of Mathematics, Swansea University, structures has proved difficult to establish.
Swansea, UK In general, the properties of branching morphologies
have been rigorously studied in two extreme cases: in the
3 Laboratory for Topology and Neuroscience at the Brain Mind limit of the full complexity of the structures (Carlsson
Institute, EPFL, Geneva, Switzerland
2009), where the entire set of points is used, and in the
4 Institute of Mathematics, University of Aberdeen, Aberdeen, opposite limit of feature extraction (DeFelipe et al. 2013;
UK Gomez-Gil et al. 2008; Blackman et al. 2014), where a
4 Neuroinform (2018) 16:3–13

(typically small) number of selected morphometrics (i.e.,

statistical features) are extracted from the morphology.
Topological data analysis (TDA) has been shown to reli-
ably identify geometric objects based on a sampled point
cloud when they are built out of well-understood pieces,
such as spheres, cylinders and tori (Carlsson 2009). It suf-
fers, however, from the deficiency that reliable grouping of
complex geometric trees by standard TDA methods, such
as Rips complexes (Edelsbrunner and Harer 2008), requires
thousands of sampled points, which is expensive in terms of
both computational complexity and memory requirements.
Feature extraction is thus the only currently feasible
solution to establishing a more quantitative approach to ana-
lyzing branching morphologies (Scorcioni et al. 2008; Ling
et al. 2012; The Petilla Interneuron Nomenclature Group P
2008). While this approach has been efficiently used in spe-
cific fields of image recognition (Schurer 1994), the extreme
diversity of the branching patterns of neurons (Markram
et al. 2004) makes it difficult to identify an optimal set of
statistical features that can reliably describe all their shapes. Fig. 1 Illustration of the separation of similar tree structures into
Neuronal classification has traditionally focused on visu- distinct groups, using topological analysis. The colored pie segments
ally distinguishing the shapes observed under a microscope show six distinct tree types: three neuronal types (upper half) and three
artificial ones (lower half). The thick blue lines show that our topolog-
(Masseroli et al. 1993), a method that is subject to large ical analysis can reliably separate similar-looking trees into groups. It
variation between experts (DeFelipe et al. 2013). is accurate both for artificially-generated trees and neuronal morpholo-
For this reason, experts generate a digital version of a gies. The dashed green lines show that classification using an improper
cell’s structure - a neuronal reconstruction (Dieter 2000) as set of user-selected features (number of branches, total length) cannot
distinguish the correct groups
a set of points in R3 sampled along each branch, together
with edges connecting adjacent pairs of points. This recon-
struction is a mathematical tree that represents the neuron’s
morphology and can be used for the extraction of its mor- As a result, neither using the full point cloud of the
phological properties. To avoid overfitting, which is a result trees nor performing expert-dependent feature selection are
of using a large number of features when few individual suitable to reliably study complex branching morphologies.
cells are available, feature selection is performed by experts In order to address this issue, we propose a standardized
who identify the relevant morphometrics for each group of topological descriptor, the Topological Morphology Des-
cells. Many sophisticated variants of the standard morpho- criptor (TMD), of any branching morphology. The TMD
logical features have been proposed over the years, such as algorithm encodes the branching pattern of the morphology
tree asymmetry (Van Pelt et al. 1991, 2001, 2005), centrifu- by discarding local fluctuations with little information con-
gal ordering (Van Pelt et al. 1989) and Strahler ordering tent, such as the position of the nodes between branch points
(Strahler 1952; Berry and Bradley 1976; Ledderose et al. and thus reduces the computational complexity of a tree.
2014), to describe the topology of branching structures. The TMD couples the topology of the branching structure
However none of those measurements preserves the correla- with the embedding in the metric space, encoding the over-
tions between distinct features. In addition, feature selection all shape of the tree. Note that the TMD is not a complete
is subjective, and alternative sets of morphometrics result invariant that fully describes the original tree, but a simpli-
in different classifications (DeFelipe et al. 2013), as illus- fication that retains enough information to perform well in
trated in Fig. 1 (see also SI: Figs. S1-S2), since the statistical the proposed discrimination tasks, by mapping the tree to
features commonly overlap even across markedly differ- a topological representation with less information loss than
ent morphological types. This is a direct consequence of the usual morphometrics.
the significant loss of information introduced by feature The TMD algorithm takes as input the partially ordered
selection, as the dimensionality of the data is substantially set of branch points (nodes with more than one child) and
reduced. leaves (nodes with no children) of the tree, where the order
Neuroinform (2018) 16:3–13 5

Fig. 2 Application of
topological analysis to a a b
neuronal tree (A) showing the
largest persistent component
(red). The persistence barcode
(B) represents each component
as a horizontal line whose
endpoints mark its birth and
death in units that depend on the
choice of the function f used
for the ordering of the nodes of
the tree. In our case, it is radial
distance of the nodes from the
root (R), so the units are
microns. The largest component c
is again shown in red together
with its birth (I) and death (II).
This barcode can be
equivalently represented as
points in a persistence diagram
(C) where the birth (I) and death
(II) of a component are the X
and Y coordinates of a point
respectively (in red). The
diagonal line is a guide to the
eye and marks points with the
same birth and death time

is given by the parent-child relation, and produces a multi- trees generation), and then to various groups of neuronal
set of intervals on the real line known as a persistence trees (see Information Sharing Statement). Our results show
barcode (Carlsson 2009), Fig. 2b. Each interval encodes that the TMD of tree shapes can be used effectively to
the lifetime of a connected component in the underlying assign a reliability measure to different proposed groupings
structure (see Glossary), identifying when a branch is first of random and neuronal trees (Fig. 1). Provided that the
detected (birth) and when it connects to a larger subtree available set of morphologies is representative of the bio-
(death). This information can be equivalently represented in logical diversity, we generate a diversity profile (Leinster
a persistence diagram (Carlsson 2009), Fig. 2c in which the and Cobbold 2012) that reflects the abundance of species as
pair of birth-death times determines a point in the real plane. well as their differences, in order to further investigate the
Either representation greatly simplifies the mathematical effects of different classification schemes (see SI: Diversity
analysis of the trees. Index).
This approach provides a simplified comparison pro-
cess, since distances inspired by persistent homology theory
(Carlsson 2009) can be defined between the outputs of Methods
the TMD algorithm (see SI: Distances between persis-
tence diagrams). Existing methods for computing distances The extraction of the barcode from an embedded tree T is
between trees, such as the edit distance (Bille 2005), the described by the TMD algorithm. Let T be a rooted, and
sequence representation (Gillette and Ascoli 2015; Gillette therefore oriented, tree (Knuth 1998), embedded in R3 . Note
et al. 2015), the blastneuron distance (Wan et al. 2015) that the operation described here is generalizable to trees
and the functional distortion distance (Bauer et al. 2014), embedded in any metric space. We denote by N := B ∪ L
are in general not universally appropriate, and therefore the set of nodes of T , which is the union of the set of branch
not biologically useful, and computationally expensive (see points B and the set of leaves L. In the case of a neuron, the
SI:Distances between trees). root R is the node representing the soma. Each node n ∈ N
Our method, in contrast, is applicable to any tree-like has references to its parent, i.e., the first node on the path
structure. We demonstrate its generality by applying it first toward the root, and to its children. Nodes with the same
to a collection of artificial random trees, (see SI: Random parent are called siblings.
6 Neuroinform (2018) 16:3–13

Let f be a real-valued function defined on the set of siblings can then be defined based on v: if n1 , n2 ∈ N, are
nodes of T . Any function f that is defined on the nodes of siblings and v(n1 ) < v(n2 ), then n1 is younger than n2 .
T can be used with the TMD algorithm, such as the radial The algorithm is initialized by setting the value of
distance, the path distance, the branch length, or the branch v(l), l ∈ L equal to the value of f (l). The leaves l ∈ L are
order (see SI, Fig. S4). Alternative functions should serve to added to a set of nodes, denoted A, which keeps a record
reveal shape characteristics that are independent from each of the active nodes. Following the path of each leaf l ∈ L
other and therefore be suitable for different tasks. For the toward the root R, all but the oldest (with respect to v) sib-
purpose of this study we define f to be the radial distance lings are killed, i.e., removed from A, at each branch point.
from the root R. For each n ∈ N, let Tn denote the sub- If siblings have the same value v it is equivalent to kill any
tree with root at the node n, and Ln the set of leaves of Tn . one of them. For each killed component one interval (birth-
We define a function v : N → R, computed by the TMD death) is added to the persistence barcode (Fig. 2). The older
algorithm, by v(n) = max{f (x) | x ∈ Ln }. An ordering of sibling cm is replaced by its parent in A and the value v(p)
Neuroinform (2018) 16:3–13 7

of its parent is set to f (cm ). This operation is applied iter- The most common topological metric that is used to
atively to all the nodes until the root R is reached. At this compare persistence diagrams is the bottleneck distance
point A contains only one component, the largest one. (Edelsbrunner and Harer 2008), denoted dB . Given a match-
ing (i.e., a bijection) between two persistence diagrams
D1 , D2 , we define the L∞ distance as the maximum dis-
tance between matched points. The bottleneck distance
dB (D1 , D2 ) is the infimum over all L∞ distances for the
possible matchings between the two persistence diagrams
(Edelsbrunner and Harer 2008).
We prove that TMD: (T , f )  → TMD(T , f ) is stable with
respect to the bottleneck distance (see SI: Stability of TMD).
For -small modifications of certain types in the tree T ,
the persistence diagram TMD(T , f ) is not modified more
than O(). In particular, the method is robust with respect
to small perturbations in the positions of the nodes and the
addition/ deletion of small branches.
However, none of the standard topological distances
between persistence diagrams is appropriate for the com-
parison of neuronal trees. The bottleneck distance as well
as distances stable with respect to it, such as the persis-
tence distortion distance (Dey et al. 2015) (see SI: Distances
between trees) cannot distinguish diagrams that differ in
their short components, which are nevertheless important
for the distinction of neuronal morphologies.
We therefore define in the space of the barcodes an
alternative distance dBar that we use to compare branch-
ing morphologies. For each barcode we generate a density
profile as follows: ∀x ∈ R the value of the histogram is
When all the branches are outgoing, i.e., the radial dis- the number of intervals that contain x, i.e., the number of
tance of the origin of a branch is smaller than the radial components alive at that point. The TMD-distance between
distance of its terminal point, the TMD algorithm is equiv- two barcodes TMD(T1 , f ) and TMD(T2 , f ) is defined as
alent to computing the barcode associated to a filtration the integral of the absolute differences between the den-
of concentric spheres of decreasing radii, centered at R sity profiles of the barcodes. This distance is not stable for
(Fig. 2). In this case, the birth time of a component is the a large number of -perturbations of the tree, but it is the
supremum of the radii of the spheres that do not contain only distance we are aware of that succeeds in capturing
the entire component. The death time is the infimum of the the differences between the short components of persistence
radii of the spheres that contain the branch point at which barcodes. This distance is similar to Sholl analysis (Sholl
the component merges with a longer one. 1953) with a few fundamental differences (see SI: Distances
The computational complexity of the TMD algorithm is between neurons). However, since this density profile col-
linear in the number of nodes. Note that the if statement lapses the barcodes into a one-dimensional distribution, it
in line 9 of the algorithm is critical for the linear complex- fails to capture the local differences between the branching
ity. The number of currently active children is saved at each structures of similar neuronal trees.
parent node to avoid quadratic complexity. For this reason, the persistence diagram was also con-
This process results in a set of intervals on the real line, verted into an unweighted persistence image, inspired by
each of which represents the lifetime of one component of persistence images introduced in Adams et al. (2016). We
the tree. The TMD algorithm that associates a persistence choose to use unweighted persistence images, since points
barcode TMD(T , f ) to a tree T is invariant under rotations close to the diagonal, which represent short components, are
and translations, as long as the function f is also. In this important for the discrimination of the neuronal trees, and
paper, f is the radial distance from R and as such it is these points are ignored in the weighted persistence images.
invariant under rotations about the root and rigid translations The unweighted persistence image representation allows
of the tree in R3 . the construction of an average image for groups of trees,
8 Neuroinform (2018) 16:3–13

which is useful for quantifying the differences between tree groups, no rigorous mathematical model has been pro-
types, since we are not aware of any unambiguous and posed for their objective classification. Finally, we used the
computationally feasible calculation of an average of per- TMD-distance to rank automatic reconstructions from the
sistence barcodes or diagrams. This method is based on the BigNeuron project (Peng et al. 2015). We thereby illustrate
discretization of a sum of Gaussian kernels (Scott 2008), the usefulness of the TMD across non-trivial examples.
centered at the points of the persistence diagram. This dis- Mathematical random trees are defined by a set of param-
cretization generates a matrix of pixel values, encoding the eters that constrain their shape: the tree depth Td , the branch
persistence diagram in a vector, called the unweighted per- length Bl , the branch angle Ba , the degree of randomness
sistence image. Machine learning tools, such as decision Dr , and the asymmetry of branches Ab (see SI: Random
trees and support vector machines can then be applied to trees generation). We defined a control group as a set of trees
this vector for the classification of the persistence diagrams. generated with predefined parameters (Td = 5, Bl = 10,
Note that the unweighted persistence images, unlike the Ba = π/4, Dr = 10%, Ab = 0.0) and independent random
persistence images defined in Adams et al. (2016), do not seeds. Each parameter was varied individually to generate
satisfy stability for the Euclidean distance between their groups of trees that differed from the control group in only
vectors with respect to the perturbations of trees that we one property. A tree is assigned to the group which is closer
consider (see SI: Stability of TMD). based on the comparison of the distances dBar between the
tree’s barcode and the barcodes of the trees in every group.
This distance is used to construct a classifier based on a
Results simple hierarchical clustering algorithm (Ward 1963). The
accuracy of this classifier is defined as the percentage of
We demonstrate the discriminative power of the TMD by successful trials.
applying it to four examples of increasing complexity. We prove that this classifier efficiently separates groups
The first application is the grouping of artificial random of random trees that differ in their tree depth (Fig. 3), with
trees that provide a well-defined test case to explore the an accuracy of 96% ± 3% (see SI: Random trees group-
method’s performance. The random trees are generated by ing). In Fig. 3 the distance matrix indicates the existence of
a constrained stochastic algorithm (see SI: Random trees three distinct groups, and the corresponding clustering. The
generation) and have properties that can be precisely mod- TMD of random trees generated by varying each of the other
ified. Next, we have analyzed datasets of more biological parameters Ba , Bl , Dr , Ab are grouped with an accuracy
relevance: neurons from different species, downloaded from of 88%, 96%, 99% and 100% respectively (see SI: Random
Ascoli et al. (2007), and distinct types of trees obtained trees grouping, Figs. S7-S11).
from several morphological types of rat cortical pyramidal Next, the TMD is used to quantify differences between
cells (Romand et al. 2011) (see Information Sharing State- neuronal morphologies. Neurons that serve distinct func-
ment). This last example is interesting because, although tional purposes exhibit unique branching patterns (Cuntz
there is biological support for their separation into distinct et al. 2007; Van Elburg and Van Ooyen 2010). In this study,

Fig. 3 Topological analysis of

artificial trees generated using a
stochastic process. Three sets of
trees are shown (only four
individuals out of twenty for
clarity). Each group differs from
the others only in the tree depth.
Each individual of the group is
generated using the same tree
parameters but a different
random number seed. The
TMD-distance of the trees allow
their accurate separation into
groups. The distance matrix
indicates the existence of three
groups which are identified with
high accuracy by a simple
dendrogram algorithm
Neuroinform (2018) 16:3–13 9

we used cat, dragonfly, fruit fly, mouse and rat neuronal different species, shown in Fig. 4. The neuronal trees of the
trees. The qualitative differences between the neuronal five different species are accurately (84%) separated into
tree types are evident from the individual geometrical the original groups. We note here that the performance of
tree shapes (Fig. 4A) as well as the extracted barcodes this process is reliable (> 70%) even for small training
(Fig. 4B). The regions of different branching density are vis- sets that contain only 25% of the whole dataset (see SI:
ible in the average unweighted persistence images of each Classification of neuronal trees).
group (Fig. 4D). Since branching density is thought to be We applied the TMD algorithm to a more challenging use
correlated with connection probability (Snider et al. 2010), case, because it is difficult for a non-expert to distinguish
we can identify the anatomical parts of the trees that are the different morphologies. While pyramidal cell (PC) mor-
important for the connectivity of different cell types. phologies (Fig. 5A) of the rat appear superficially similar,
The performance of a supervised classifier trained on the the unweighted persistence images (Fig. 5B) reveal funda-
unweighted persistence images (see SI: Supervised Classifi- mental morphological differences between them, related to
cation, Classification of neuronal trees) of the TMD results the existence and the shape of the apical tuft. The apical
is demonstrated by the grouping of neuronal trees from the tuft of PCs is known to play a key role in the integration of

a b c d

Fig. 4 Topological comparison of neurons from different animal representative barcode in B and its superimposed persistence diagram
species. Each row corresponds to a species: (I) cat, (II) dragonfly, (III) are shown in the third column (C). By combining the persistence dia-
fruit fly, (IV) mouse and (V) rat. Trees from several exemplar cells grams in (C) for several trees we can define an average unweighted
for each species are shown in the first column (A). Representative persistence image (D) in order to study and quantify the structural dif-
persistence barcodes for the cells in A are shown in the second col- ferences between distinct morphological groups. The trees in the first
umn (B). The structural differences of the trees are clearly evident in row (cat) are more tightly grouped than those in the second row (drag-
these barcodes. II, III and V have clusters of short components, clearly onfly), and two clusters are visible in the dragonfly trees. Considering
distinct from the largest component, while I and IV have bars of a rows 1 and 4, the extension of the elliptical peak perpendicular to
quasi-continuous distribution of decreasing lengths. Also, barcodes III, the diagonal line reflects the variance in the length-scale mentioned
and V show empty regions between dense regions of bars, indicating earlier for a single cell’s barcode. Note that the trees, barcodes, and
the existence of two clusters in the morphologies, while barcodes I unweighted persistence images are not shown to the same scale for
and IV are dense overall. The unweighted persistence image for each clarity: see the scale bar in each case
10 Neuroinform (2018) 16:3–13

c d e f

Fig. 5 Comparison of the TMD of apical dendrite trees extracted the expert-assigned groups of cells. The binary classification (C) and
from several types of rat pyramidal neuron. Four cell types are shown the confusion matrix (D) based on the TMD algorithm shows an over-
in (A): UPC, SPC, TPC-A, TPC-B (left to right). The morphologi- lap of TPC-A and TPC-B trees. When those two classes are merged (E,
cal differences between these cell types are subtle, but the unweighted F) the separation between the remaining types is evident. This result
persistence images (B) clearly reveal them, particularly the presence shows that the unweighted persistence images objectively support the
of two clusters in the TPC-A and TPC-B cell types. From these expert’s classification when the morphological differences between the
unweighted persistence images we train a decision tree classifier on classes are significant

neuronal inputs through their synapses in higher cortical Finally, the TMD algorithm can be used to assess the
layers, and is therefore a key indicator for the functional role quality of any manually or automatically reconstructed neu-
of the cell. ron if a reference morphology is available. The best use case
The separation of the PC trees into four groups cannot for this application is the datasets of BigNeuron (Peng et al.
be justified based on purely morphological grounds, since 2015), a community effort to advance single-neuron auto-
there is no coherent difference between the branching pat- matic reconstruction. The same stack of images of a scanned
terns of TPC-A and TPC-B (Fig. 5C, D). On the contrary, morphology is used for manual reconstruction (reference
the separation in three groups (UPC, SPC and TPC -the morphology) and for automatic reconstructions with a set of
superset of TPC-A and TPC-B- Fig. 5E, F) is supported by algorithms (test set). Due to the large number of reconstruc-
TMD-based classifiers, by detecting the fundamental dif- tions generated by the BigNeuron project, the analysis of the
ferences between their branching structures. Therefore, the data requires a high-computational-performance algorithm.
TMD provides a solid benchmark test to objectively support The linear complexity of the TMD makes it highly suitable
or disprove proposed classification schemes. for the analysis of this large dataset.
Neuroinform (2018) 16:3–13 11

a b

50 50

50 50
c

Fig. 6 Comparison of the TMD of BigNeuron neuronal morpholo- all the automatically reconstructed neurons (A, in blue), and the den-
gies. An image stack is used for the manual reconstruction (reference sity plot of the ten best automatic reconstructions (B, in red), ranked
neuron) and for the automatic reconstructions produced by a variety according to their TMD-distance from the reference neuron. A com-
of community supplied algorithms. The results of each algorithm are parison between panels A and B shows that the density plot of the ten
illustrated in panel C, from best (top left) to worst (bottom right). The highest ranked automatic reconstructions closely matches the structure
reference neuron (in black) is visualized against the density plot of of the reference morphology

The automatic reconstructions were ranked based on system. A major challenge in neuroscience has therefore
their TMD-distance from the reference morphology. The been to reliably describe the shape of neurons. We have
TMD was able to accurately assess the quality of the introduced here the Topological Morphology Descriptor,
automatic reconstructions, as presented in Fig. 6, as the derived from principles of persistent homology. The TMD
similarity of the branching structure of the automatic recon- of a tree retains enough topological information to allow
structions to the reference neuron decreases with the TMD- the systematic comparison between branching morpholo-
ranking. The density plot of all the automatic reconstruc- gies. Therefore, it provides a topological benchmark for
tions Fig. 6A does not reproduce the shape of the reference the rigorous comparison of different structures and it could
morphology, as reconstruction errors are over-represented. advance our understanding of the anatomy and diversity of
On the contrary, the density plot of the ten TMD-best recon- the neuronal morphologies.
structions closely matches the structure of the reference This technique can be applied to any rooted tree equipped
morphology. with a function defined on its nodes. Further biological
examples include botanic trees (Lopez et al. 2010), corals
(Kruszyṅski et al. 2007) and roots of plants (Wang et al.
Discussion 2009). The method is not restricted to trees in R3 , but can
be generalized to any subset T of a metric space M, with a
The morphological diversity of neurons supports the com- base-point R. A persistence barcode can then be extracted
plex information-processing capabilities of the nervous using a filtration by concentric spheres in M centered at
12 Neuroinform (2018) 16:3–13

R, enabling us to efficiently study the shape of complex References

multidimensional objects.
While the static neuronal structures presented in this
Adams, H., Chepushtanova, S., Emerson, T., Hanson, E., Kirby, M.,
paper are biologically interesting themselves, our method Motta, F., Nevill, R., Peterson, C., Shipman, P., & Ziegelmeier,
could also be generalized to track the morphological evo- L. (2016). Persistence images: A stable vector representation of
lution of trees. The topological study of the growth of an persistent homology CoRR arXiv:1507.06217.
Agam, O., Bettelheim, E., Wiegmann, P., & Zabrodin, A. (2002).
embedded tree could be addressed through Multidimen-
Viscous fingering and the shape of an electronic droplet in the
sional Persistence (Carlsson and Zomorodian 2009), a new quantum hall regime.
area of TDA, for which computational tools are currently Alexandrowicz, Z. (1985). Growth and shape of branched polymers,
being explored (Lesnick and Wright 2015; Gäfvert 2016). aggregates and percolating clusters. Physics Letters A, 109(4),
In this case the spherical filtration identifying relevant topo- 169–173.
Ascoli, G.A., Donohue, D., & Halavi, M. (2007). Neuromorpho.org:
logical features of the tree could be enriched with a second a central resource for neuronal morphologies. Journal of Neuro-
filtration representing temporal evolution. This application science, 27(35), 9247–51.
could be useful in agriculture to study growing roots (Wang Badea, T., & Nathans, J. (2011). Morphologies of mouse retinal
et al. 2009) and trees, and in neuroscience, to study neurons ganglion cells expressing transcription factors brn3a, brn3b, and
brn3c: Analysis of wild type and mutant cells using genetically-
in the developing brain. directed sparse labeling. Vision Research, 51(2), 269–279.
Bauer, U., Ge, X., & Wang, Y. (2014). Measuring Distance Between
Reeb Graphs, SOCG’14. New York: ACM.
Information Sharing Statement The artificial random trees used in Berry, M., & Bradley, P.M. (1976). The application of network anal-
Figs. 1 and 3 were generated by software developed in BBP. The tree ysis to the study of branching patterns of large dendritic fields.
structures can be made available (in hdf 5 format) upon request. The Brain Research, 109(1), 111–132.
biological morphologies used in Figs. 1, 2 and 5 are provided from Bille, P. (2005). A survey on tree edit distance and related problems.
Laboratory of Neural Microcircuitry (LNMC), EPFL (Romand et al. Theoretical Computer Science.
2011). The biological morphologies used in Fig. 4 were downloaded Blackman, A., Grabuschnig, S., Legenstein, R., & Sjöström, P. (2014).
from Neuromorpho.org. In particular, cat neurons were provided by A comparison of manual neuronal reconstruction from biocytin
Rose et al. (1995), dragonfly neurons by Gonzalez-Bellido et al. histology or 2-photon imaging: morphometry and computer mod-
(2015), fruit fly neurons by Chiang et al. (2011), mouse neurons by eling. Front in Neuroanatomy, 8(65), 65.
(Badea and Nathans 2011) and rat neurons by Romand et al. (2011). Carlsson, G. (2009). Topology and data. Bulletin of American Mathe-
The automatic and manual reconstructions used in Fig. 6 are provided matical Society, 46, 255–308.
by BigNeuron (Peng et al. 2015). Carlsson, G., & Zomorodian, A. (2009). The theory of multidimen-
sional persistence. Discrete & Computational Geometry, 42(1),
71–93.
Acknowledgments Among others, we thank Athanassia Chal- Chiang, A.S., Lin, C.Y., Chuang, C.C., Chang, H.M., et al.
imourda and Katherine Turner for helpful conversations in various (2011). Three-dimensional reconstruction of brain-wide wiring net-
stages of this research and Jay Coggan for a critical reading of the works in drosophila at single-cell resolution. Current Biology,
manuscript. We also thank Hanchuan Peng and Xiaoxiao Liu for pro- 21(1), 1–11.
viding and curating the BigNeuron datasets. This work was supported Cuntz, H., Borst, A., & Segev, I. (2007). Optimization principles of
by funding for the Blue Brain Project (BBP) from the ETH Domain. dendritic structure. Theoretical Biology and Medical Model, 4(21),
P.D. and R.L. were supported part by the Blue Brain Project and by the 21.
start-up grant of KH. Partial support for P.D. has been provided by the DeFelipe, J., López-Cruz, P.L., Benavides-Piccione, R., et al. (2013).
Advanced Grant of the European Research Council GUDHI (Geomet- New insights into the classification and nomenclature of cortical
ric Understanding in Higher Dimensions). MS was supported by the GABAergic interneurons. Nature Reviews Neuroscience, 14(3),
SNF NCCR “Synapsy”. 202–216.
Dey, T., Shi D., & Wang, Y. (2015). Comparing graphs via persistence
distortion. In SOCG.
Dieter, J. (2000). Accurate Reconstruction of Neuronal Morphology
Author Contributions LK and KH conceived the study. LK devel- in Computational Neuroscience, Frontiers in Neuroscience. Boca
oped the topological algorithm with the contribution of PD. PD, KH Raton: CRC Press.
contributed topological ideas used in the analysis. HM suggested the
Edelsbrunner, H., & Harer, J. (2008). Persistent homology—a sur-
biological datasets analyzed. RL, MS and JCS gave conceptual advice.
vey. In Goodman, J.E., & Pach, J. (Eds.) American Mathematical
All authors discussed the results, wrote the main paper and commented
Society, (Vol. 453 pp. 257–282). Providence.
on the manuscript at all stages.
Ferrante, M., Migliore, M., & Ascoli, G. (2013). Functional impact
of dendritic branch point morphology. Journal of Neuroscience,
33(5), 2156–65.
Open Access This article is distributed under the terms of the Gäfvert, O. (2016). Algorithms for multidimensional persistence.
Creative Commons Attribution 4.0 International License (http:// Gillette, T.A., & Ascoli, G.A. (2015). Topological characterization of
creativecommons.org/licenses/by/4.0/), which permits unrestricted neuronal arbor morphology via sequence representation: I - motif
use, distribution, and reproduction in any medium, provided you give analysis. BMC Bioinformatics, 16, 216.
appropriate credit to the original author(s) and the source, provide a Gillette, T., Hosseini, P., & Ascoli, G. (2015). Topological characteri-
link to the Creative Commons license, and indicate if changes were zation of neuronal arbor morphology via sequence representation:
made. II - global alignment. BMC Bioinformatics, 16, 209.
Neuroinform (2018) 16:3–13 13

Gomez-Gil, P., Ramirez-Cortes, M., Gonzalez-Bernal, J., Pedrero, Schurer, T. (1994). An experimental comparison of different feature
A.G., Prieto-Castro, C.I., Valencia, D., Lobato, R., & Alonso, extraction and classification methods for telephone speech. pp.
J.E. (2008). A feature extraction method based on morphological 93–96.
operators for automatic classification of leukocytes. pp. 227–232. Scorcioni, R., Polavaram, S., & Ascoli, G. (2008). L-measure: a web-
Gonzalez-Bellido, P., Peng, H., Yang, J., Georgopoulos, A., & Olberg, accessible tool for the analysis, comparison, and search of digital
R. (2015). Eight pairs of descending visual neurons in the drag- reconstructions of neuronal morphologies. Nature Protocols, 3,
onfly give wing motor centers accurate population vector of 866–76.
prey direction. Proceedings of the National Academy of Sciences, Scott, D. (2008). Kernel density estimators, (pp. 125–193): Wiley.
110(2), 696–701. Sholl, D.A. (1953). Dendritic organization in the neurons of the visual
Jan, Y., & Jan, L. (2010). Branching out: mechanisms of dendritic and motor cortices of the cat. Journal of Anatomy, 87, 387–406.
arborization. Nature reviews. Neuroscience, 11, 316–28. Snider, J., Pillai, A., & Stevens, C.F. (2010). A universal property of
Knuth, D. (1998). The art of computer programming volume 2: axonal and dendritic arbors. Neuron, 66, 45–56.
Seminumerical algorithms. Massachusetts: Reading. Strahler, A.N. (1952). Hypsometric analysis of erosional topography.
Kruszyṅski, K.J., Kaandorp, J.A., & van Liere, R. (2007). A com- Bulletin of the Geological Society of Science, 63, 1117–1142.
putational method for quantifying morphological variation in The Petilla Interneuron Nomenclature Group P (2008). Petilla termi-
scleractinian corals. Coral Reefs, 26(4), 831–840. nology: nomenclature of features of gabaergic interneurons of the
Ledderose, J., Sencion, L., Salgado, H., Arias-Carrion, O., & Trevino, cerebral cortex. Nature Reviews Neuroscience, 9(7), 557–568.
M. (2014). A software tool for the analysis of neuronal morphol- Van Elburg, R., & Van Ooyen, A. (2010). Impact of dendritic size
ogy data. International Archives of Medicine, 7, 6. and dendritic topology on burst firing in pyramidal cells. PLoS
Leinster, T., & Cobbold, C. (2012). Measuring diversity: the impor- Computational Biology, 6(5), 1–19.
tance of species similarity. Ecology, 93(3), 477–489. Van Pelt, J., Verwer, R.W., & Uylings, H.B.M. (1989). Centrifugal-
Lesnick, M., & Wright, M. (2015). Interactive visualization of 2-d order distributions in binary topological trees. Bulletin of Mathe-
persistence modules. arXiv:1512.00180 [cs, math]. matical Biology, 51(4), 511–536.
Ling, C., Hendrickson, M., & Kalil, R. (2012). Morphology, classifica- Van Pelt, J., Uylings, H.B.M., Verwer, R.W.H., Pentney, R.J., & Wold-
tion, and distribution of the projection neurons in the dorsal lateral enberg, M.J. (1991). Tree asymmetry—A sensitive and practical
geniculate nucleus of the rat. PLoS ONE, 7, e49161. measure for binary topological trees. Bulletin of Mathematical
Lopez, L.D., Ding, Y., & Yu, J. (2010). Modeling complex unfoliaged Biology, 54(5), 759–784.
trees from a sparse set of images. Computer Graphics Forum, 29, Van Pelt, J., Van Ooyen, A., & Uylings, H.B.M. (2001). Modeling
2075–2082. dendritic geometry and the development of nerve connections.
Lyons, M.J., Budynek, J., & Akamatsu, S. (1999). Automatic clas- In de Schutter, E., & Cannon, R.C. (Eds.) (CD-ROM) Com-
sification of single facial images. IEEE Transactions on Pattern putational Neuroscience: Realistic Modeling for Experimentalist
Analysis and Machine Intelligence, 21(12), 1357–1362. (pp. 179–208). Boca Raton: CRC Press.
Mandelbrot, B., & Freeman, W. (1983). The fractal geometry of nature. Van Pelt, J., Van Ooyen, A., & Uylings, H.B.M. (2005). Natural
Earth Surface Processes and Landforms, 8(4), 406. variability in the geometry of dendritic branching patterns. In
Markram, H., Toledo-Rodriguez, M., Wang, Y., Gupta, A., Silberberg, Reeke, G.N., Poznanski, R.R., Lindsay, K.A., Rosenberg, J.R., &
G., & Wu, C. (2004). Interneurons of the neocortical inhibitory Sporns, O. (Eds.) Modeling in the Neurosciences: From Biological
system. Nature Reviews Neuroscience, 5(10), 793–807. Systems to Neuromimetic Robotics (pp. 89–115). Boca Raton:
Masseroli, M., Bollea, A., & Forloni, G. (1993). Quantitative morphol- CRC Press.
ogy and shape classification of neurons by computerized image Wan, Y., Long, F., Qu, L., Xiao, H., Hawrylycz, M., Myers, E.W., &
analysis. Computer Methods and Programs in Biomedicine, 41(2), Peng, H. (2015). Blastneuron for automated comparison, retrieval
89–99. and clustering of 3d neuron morphologies. Neuroinformatics,
Peng, H., Hawrylycz, M., Roskams, J., Hill, S., Spruston, N., Mei- 13(4), 487–499.
jering, E., & Ascoli, G.A. (2015). BigNeuron: large-scale 3D Wang, H., Siopongco, J., Wade, L., & Yamauchi, A. (2009). Frac-
neuron reconstruction from optical microscopy images. Neuron, tal analysis on root systems of rice plants in response to drought
87, 252–256. stress. Environmental and Experimental Botany, 65(2–3), 338–
Romand, S., Wang, Y., Toledo-Rodriguez, M., & Markram, H. (2011). 344.
Morphological development of thick-tufted layer v pyramidal Ward, J.H.J. (1963). Hierarchical grouping to optimize an objective
cells in the rat somatosensory cortex. Frontiers in neuroanatomy, function. American Statistics Association, 58(301), 236–244.
5, 5. Zomorrodi, R., Ferecskó, A., Kovács, K., Kröger, H., & Timofeev,
Rose, P., Jones, T., Nirula, R., & Corneil, T. (1995). Innervation of I. (2010). Analysis of morphological features of thalamocorti-
motoneurons based on dendritic orientation. Journal of Neuro- cal neurons from the ventroposterolateral nucleus of the cat. The
physiology, 73(3), 1319–1322. Journal of Comparative Neurology, 518(17), 3541–3556.