Content uploaded by Marta Mesquita da Costa
Author content
All content in this area was uploaded by Marta Mesquita da Costa on Aug 21, 2015
Content may be subject to copyright.
NBLAST: Rapid, sensitive comparison of neuronal structure
and construction of neuron family databases
Marta Costa
1,2
, James D. Manton
1
, Aaron D. Ostrovsky
1
, Steffen Prohaska
1,3
,
Gregory S. X. E. Jefferis
1*
1
Neurobiology Division, MRC Laboratory of Molecular Biology, Cambridge, CB2 0QH, UK
2
Department of Genetics, University of Cambridge, Cambridge, CB2 3EH, UK
3
Zuse Institute Berlin (ZIB), 14195 Berlin-Dahlem, Germany
Please note that this preprint is our second public draft. Current versions of the open source software
described in the manuscript are available by following links in the Experimental Procedures, which are
also summarised at http://jefferislab.org/si/nblast. Processed data derived from raw data generously
made publicly available by third parties (primarily flycircuit.tw) will be made available at least by
the time this paper is accepted, hopefully rather sooner; please contact Greg for details. We welcome
feedback, queries and suggestions on any aspect of the manuscript (including relevant prior art), code
or data to jefferis@mrc-lmb.cam.ac.uk.
Abstract
Neural circuit mapping efforts in model organisms are generating multi-terabyte datasets of
10,000s of labelled neurons. Such data demand new computational tools to search and organize
neurons. We present a general, sensitive and rapid algorithm, NBLAST, for measuring pairwise
neuronal similarity. NBLAST considers both position and local geometry and works by decom-
posing a query and target neuron into short segments; matched segment pairs are scored using a
log-likelihood ratio scoring matrix empirically defined by the statistics of real matches and non-
matches.
We validated NBLAST by processing a published dataset of 16,129 single Drosophila neurons.
NBLAST is sensitive enough to distinguish two images of the same neuron and can be used to
distinguish neuronal types without a priori information. Detailed cluster analysis of extensively
studied neuronal classes identified new neuronal types and unreported features of topographic
organization. NBLAST supports diverse additional query types including matching neurite tracts
with transgene expression patterns. We organize all 16,129 neurons into 1,052 clusters of highly
related neurons, further organized into superclusters, simplifying exploration and identification of
neuronal types including sexually dimorphic and visual interneurons.
1
this preprint is the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for; http://dx.doi.org/10.1101/006346doi: bioRxiv preprint first posted online August 9, 2014;
Introduction
Correlating the functional properties and behavioral relevance of neurons with their cell type is a
basic activity in the study of neuronal circuits. While there is no universally accepted definition of
neuron type, key descriptors include morphology, position within the nervous system, genetic mark-
ers, connectivity and intrinsic electrophysiological signatures (Migliore and Shepherd, 2005; Bota
and Swanson, 2007; Rowe and Stone, 1976). Despite this ambiguity, neuron type remains a key
abstraction helping to reveal organizational principles and enabling results to be compared and col-
lated across research groups. Furthermore there is increasing appreciation that highly quantitative
approaches are critical to generate the most efficient cell type catalogues in support of circuit re-
search (Petilla Interneuron Nomenclature Group et al., 2008; Nelson et al., 2006; Kepecs and Fishell,
2014)(http://www.nih.gov/science/brain/11252013-Interim-Report-Final.pdf).
Since neuronal morphology and position strongly constrain (and are partially defined by) connec-
tivity, they have been mainstays of studies of circuit organization for over a century. Classic techniques
to reveal neuronal morphology include the Golgi method made famous by Cajal, microinjection, and
filling of cells during intracellular recording. Recently these have been supplemented by genetic ap-
proaches to sparse and combinatorial labeling enabling increasingly large-scale characterization of
single neuron morphology (Jefferis and Livet, 2012).
Classically, the position of neuronal somata or arbors was established by comparison with anatomi-
cal landmarks, often revealed by a general counterstain; this approach is especially effective in brain re-
gions with strong laminar organization e.g. the mammalian retina (Badea and Nathans, 2004; Coombs
et al., 2006; Kong et al., 2005; Sümbül et al., 2014), fly optic lobe (Fischbach and Dittrich, 1989;
Morante and Desplan, 2008) or cerebellum (Cajal and Azoulay y, 1911). Recently, 3D light microscopy
and image registration have enabled direct, automated image fusion to generate digital 3D atlases of
brain regions or whole brains (Jefferis et al., 2007; Lin et al., 2007; El Jundi et al., 2009; Rybak et al.,
2010; Cachero et al., 2010; Yu et al., 2010b; Sunkin et al., 2013; Zingg et al., 2014; Oh et al., 2014).
Such atlases can generate specific, testable hypotheses about circuit organization and connectivity at
large scales. The largest study to date Chiang et al. (2011) combined genetic mosaic labeling and image
registration to produce an atlas of over 16,000 single cell morphologies embedded within a standard
Drosophila brain at http://flycircuit.tw.
Neuronal morphologies can be represented as directed graph structures embedded in 3D space.
However this is typically the (arbitrary) physical space of the imaging system used to reconstruct the
neuron, rather than a brain atlas. For this reason, databases such as NeuroMorpho.org (Parekh and
Ascoli, 2013) contain > 27,000 neurons, but do not include precise positional information. Data on
this scale presents both an acute challenge, finding and organizing related neurons, but also an oppor-
tunity: quantitative morphological classification may help solve the problem of cell type. However, a
key requirement is a tool enabling rapid and sensitive computation of neuronal similarity within and
between datasets. This has clear analogies with bioinformatics: the explosion of biological sequence
information from the late 80s motivated the development of sequence similarity tools such as FASTA
(Pearson and Lipman, 1988) and BLAST (Altschul et al., 1990). These, and related algorithms, enabled
pairwise similarity scoring, alignment and rapid database queries as well as hierarchically organized
databases of protein families (Sonnhammer et al., 1997).
Several existing strategies for measuring neuronal similarity exist, each with distinct target appli-
cations and depending on a particular data structure. Mayerich et al. (2012) have applied general graph
2
this preprint is the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for; http://dx.doi.org/10.1101/006346doi: bioRxiv preprint first posted online August 9, 2014;
similarity metrics (reviewed by Conte et al., 2004) to compare neuronal reconstructions represented
as fully-connected graphs with a ground truth reconstruction. Basu et al. (2011) decomposed branch-
ing neuronal trees into a family of unbranched paths for which they proposed a geometric measure
of similarity which could include positional information. Further simplifying the neuronal represen-
tation, Cardona et al. (2010) decomposed single unbranched neurites into sequences of vectors and
used dynamic programming to find an optimal 3D alignment. Critically, they validated this approach
on a database of a few hundred traced structures, achieving very high classification accuracy. How-
ever, while this algorithm could be modified for use with branched neurons, it treats each unbranched
neuronal segment as a separate alignment problem and so there is no natural way to handle trees with
many such segments.
The choice of data structure remains important: fully automatic 3D tracing of single neurons re-
mains unsolved (Brown et al., 2011), while expression patterns containing multiple neurons cannot
be represented as a single binary tree. We previously developed an image segmentation pipeline that
represents expression patterns (consisting of up to 100 neurons) as point clouds with tangent vectors
defining the local heading of the neurons. We used this simplified representation in a supervised learn-
ing approach to the challenging problem of recognizing groups of lineage-related neurons (Masse
et al., 2012). Ganglberger et al. (2014) have recently applied a related approach directly to unseg-
mented expression pattern image data at the expense of much higher memory demands (order 100 MB
per specimen), severely limiting throughput.
Combining the data representation of Masse et al. (2012) with a very large single neuron dataset
(Chiang et al., 2011) allowed us to test and validate a new algorithm, NBLAST, that is flexible, ex-
tremely sensitive and very fast (pairwise search times < 5 ms). Critically, the algorithm’s scoring pa-
rameters are defined statistically rather than by expert intuition.
We first describe the NBLAST algorithm, providing an open source command line implementa-
tion and a web query tool (see jefferislab.org/si/nblast/clusters). We validate NBLAST
for applications including neuron database search and unsupervised clustering of neurons. NBLAST
can identify well-studied neuronal types in Drosophila with sensitivity matching domain experts, in
a fraction of the time of manual classification. NBLAST can also identify new neuronal types and
reveal undescribed features of topographic organization. Finally, we apply our method to 16,129 neu-
rons from the FlyCircuit dataset, reducing this to a non-redundant set of 1,052 morphological clusters.
Manual evaluation of a subset of these clusters show that they closely match expert definition of cell
types. These clusters, which we also organize into an online supercluster hierarchy, therefore represent
a preliminary global cell type classification for the Drosophila brain.
Results
Algorithm
Our principal design goals were to develop a neuron similarity algorithm that included aspects of both
spatial location (within a brain or brain region) and neuronal branching pattern, and that was both
extremely sensitive and very fast. The applications that we had in mind were the searching of large
databases of neurons (10,000–100,000 neurons), clustering of neurons into families by calculating all-
against-all similarity matrices, and the efficient organization and navigation of datasets of this size.
We eventually selected an approach based on direct pairwise of comparison of neurons pre-registered
3
this preprint is the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for; http://dx.doi.org/10.1101/006346doi: bioRxiv preprint first posted online August 9, 2014;
to a template brain and represented as vector clouds. Further details are provided in Supplemental
Information.
The starting point for our algorithm is a representation in which neuron structures have been re-
duced to segments, represented as a location and an associated local tangent vector. This retains some
local geometry but does not attempt to capture the topology of the neuron’s branching structure. We
have found that a simplified representation of this sort can be constructed for image data that would
not permit automated reconstructions. In order to prepare data of this sort in quantity, we developed an
image processing pipeline summarized in Figure 1A and detailed in Experimental Procedures. Briefly,
brain images from the FlyCircuit dataset (Chiang et al., 2011) were subjected to non-rigid image reg-
istration (Jefferis et al., 2007) against a newly constructed intersex template brain. Neuron images
were thresholded and subjected to a 3D skeletonization procedure (Lee et al., 1994) implemented in
Fiji (Schindelin et al., 2012). These thresholded images were then converted to the point and tangent
vector representation (Masse et al., 2012) using our R package nat (Jefferis and Manton, 2014); the
tangent vector (i.e. the local heading) of the neuron at each point was computed as the first eigenvector
of a singular value decomposition (SVD) of the point and its 5 nearest neighbors.
After pre-processing, 3D data could be visualized and analyzed in R using nat (Figure 1B). Neu-
rons were represented by median 1070 points/vectors; the 16,129 neurons occupied 1.8 GB, fitting
comfortably into a laptop’s main memory. Since the fly brain is almost completely symmetric, but
neurons were labelled randomly in both hemispheres, we mapped all neurons to the left hemisphere
(defined primarily by cell body location, see Experimental Procedures and Figure 1B) using a non-rigid
mirroring procedure (Manton et al., 2014).
With a database of aligned neurons in an appropriate representation, we were then able to calculate
NBLAST pairwise similarity scores. One neuron is designated the query and the other the target. For
each query segment (defined by a midpoint and tangent vector) the nearest neighbor (using straight-
forward Euclidean distance) is identified in the target neuron (Figure 1C–D). A score for the segment
pair is calculated as a function of two measurements: , the distance between the matched segments
(indexed by ), and
, the absolute dot product of the two tangent vectors; the absolute dot prod-
uct is used because the orientation of the tangent vectors has no meaning in our data representation
(Figure 1C). The scores are then summed over each segment pair to give a raw score, :
(1)
The question then becomes: what is an appropriate function
? We developed an approach
inspired by the scoring system of the BLAST algorithm (Altschul et al., 1990). For each segment pair
we defined the score as the log probability ratio:
(2)
i.e. the probability that the segment pair was derived from a pair of neurons of the same type, versus a
pair of unrelated neurons . We could then define
empirically by finding the joint distribution
of and
for pairs of neurons of the same type (Figure 1E–G). For our default scoring matrix,
we used a set of 150 olfactory projection neurons innervating the same glomerulus, unambiguously
the same neuronal type (Figure 1F).
was calculated simply by drawing 5,000 random pairs of
4
this preprint is the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for; http://dx.doi.org/10.1101/006346doi: bioRxiv preprint first posted online August 9, 2014;
neurons from the database, assuming that the large majority of such pairs are unrelated neurons. Joint
distributions were calculated using 10 bins for the absolute dot product and 21 bins for the distance to
give two 21 row × 10 column matrices. The 2D histograms were then normalized to convert them to
probabilities and the log ratio defined the final scoring matrix (Figure 1G). Plotting the scoring matrix
emphasizes the strong distance dependence of the score but also shows that for segment pairs closer
than ~10 µm, the logarithm of the odds score increases markedly as the absolute dot product moves
from 0 to 1 (Figure 1H).
We implemented the NBLAST algorithm as an R package (nat.nblast) built on top of a high-
performance k nearest neighbor library (http://cran.r-project.org/web/packages/RANN/index.html), that
immediately enables pairwise queries, searches of a single query neuron against a database of target
neurons (Figure 2) and all-by-all searches. Runtimes on a single core laptop computer were 2 ms per
comparison or 30 s for all 16,129 neurons. In order to enable clustering of neurons on the fly, we also
pre-computed an all-by-all similarity matrix for all 16,129 neurons (2.6 × 10
8
scores, 1.0 GB). We also
developed a simple web application (linked from jefferislab.org/si/nblast) to allow online
queries for this test dataset.
NBLAST can find whole or partial matches for diverse query objects
The NBLAST algorithm is flexible, identifying both global and partial matches for multiple classes of
query object (Figure 2). The only requirements are that query objects (or fragments) must be registered
against a template brain and can be converted to a point and vector representation.
As a first example we query a (whole) FlyCircuit neuron against the 16,129 FlyCircuit neurons. The
top hits are all very similar neurons with small differences in length and neurite position (Figure 2B). A
second example uses a neurite fragment, corresponding to part of the axon; top hits all follow the same
axon tract, although their variable axonal and dendritic arbors indicate that they are distinct neuron
types (Figure 2C).
User tracings can also be used as queries. We traced the characteristic bundle of 20-30 primary neu-
rites of the fruitless neuroblast clone pMP-e (which generates male-specific P1 neurons, (Kimura et al.,
2008; Cachero et al., 2010)). Our query returned many single P1 neurons from the FlyCircuit database
(Figure 2D) (for more details see Figure 7). A similar approach can be used to identify candidate neu-
ronal types labelled by genetic driver lines where the detailed morphology of individual neurons cannot
be ascertained. As an example we take the GAL4 line R18C12 (Jenett et al., 2012) (Figure 2E). The
expression pattern includes an obvious bilateral dorsal tract associated with a specific cluster of cell
bodies (Figure 2E). We traced the main neurites of this cell cluster; NBLAST identified three very
similar FlyCircuit neurons, which completely overlapped with the expression of R18C12. These three
neurons appear to be different subtypes, each varying in their terminal arborizations. Conversely, we
used one tracing from a recently published projectome dataset containing >9000 neurite fibers (Peng
et al., 2014) to find similar FlyCircuit neurons (Figure 2F).
NBLAST scores are sensitive and biologically meaningful
A good similarity algorithm should be sensitive enough to reveal identical neurons with certainty,
while having the specificity to ensure that all high scoring results are relevant hits. We used the full
FlyCircuit dataset to validate NBLAST performance.
5
this preprint is the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for; http://dx.doi.org/10.1101/006346doi: bioRxiv preprint first posted online August 9, 2014;
Our first example uses an auditory interneuron, fru-M-300198 as query (Figure 3A–C). Ordering
search results by NBLAST score, the first returned object is the query neuron itself (since it is present
in the database), followed by the top hit (fru-M-300174) which completely overlaps with the query
(Figure 3A’). A histogram plot of NBLAST scores, showed that the top hit score was clearly an outlier:
96.1 % of the self-match score of the query against itself (i.e. the maximum possible score) (Figure 3C).
Further investigation revealed that these “identical twins”, both derived from the same raw confocal
image. The next 8 hits are also very similar to the query but are clearly distinct specimens, having small
differences in position, length and neurite branching that are typical of sister neurons of the same type
(Figure 3A”).
The score histogram shows that only a minority of hits (3 %) have a score above 0 (Figure 3B–C).
A score of 0 represents a natural cutoff for NBLAST, since it means that, on average, segments from
this neuron have a similarity level that is equally likely to have arisen from a random pair of neurons in
the database as a pair of neurons of the same type. We divided the neurons with score>0 into 8 groups
with decreasing similarity scores (Figure 3C’). Only the highest scoring real hits (group II) appear of
exactly the same type, although lower scoring groups contain neurons that would be ranked as very
similar.
Although raw NBLAST scores correctly identify similar neurons, they are not comparable from
one query neuron to the next: the score depends on neuron size and segment number. This confounds
search results for neurons of very different sizes or when the identity of query and target neurons is
reversed. For example, a search with a large neuron as query and a smaller one as target (pair 1) will
have a very low forward score, because the large neuron has many segments that are unmatched, but
a high reverse score, since most of target will match part of the query (Figure 3D). One approach
to correct for this is to normalize the scores by the size of the query neuron. Although normalized
scores are comparable, unequal forward and reverse scores between large and small neurons remain
an issue. One simple strategy is to calculate the mean of the forward and reverse scores (mean score).
Two neurons of similar size have a higher mean score than two neurons of unequal size (Figure 3D).
Repeating the analysis of Figure 3C–C’ using mean scores (Figure S2) eliminated some matches due
to unequal size that could be considered false positives.
During our analysis, we sporadically noticed cases where two neurons in the database actually were
actually the same physical specimen (Figure S1). We tested if NBLAST could systematically reveal
such instances. We collected the top hit for each neuron and analyzed the distribution of forward (Fig-
ure 3E) and reverse scores (data not shown). A small tail (~ 1 % of all top hits) have anomalously high
scores (over 0.8). Given this distribution, we examined neuron pairs with both forward and reverse
scores over 0.8. We classified these 72 pairs into 4 different groups. From highest to lowest predicted
similarity, the groups are: same segmentation, a segmented image of a neuron has been duplicated
(Figure S1A); same raw image, corresponding to a different segmentation of the same neuron (Fig-
ure 3B’); same specimen, when two images are from the same brain but not from the same confocal
image (Figure S1B); and different specimen, when two neurons are actually from different brains,
(therefore suggesting that they are of the same type). The distribution of NBLAST scores for these
four categories matches the predicted hierarchy of similarity (Figure 3F). These results underline the
high sensitivity of the NBLAST algorithm to small differences between neurons.
Taken together these results validate NBLAST as a sensitive and specific tool for finding similar
neurons.
6
this preprint is the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for; http://dx.doi.org/10.1101/006346doi: bioRxiv preprint first posted online August 9, 2014;
NBLAST scores can distinguish Kenyon cell classes
We wished to investigate whether NBLAST scores can be used to cluster neurons by structure and po-
sition, potentially revealing functional classes. We decided to begin our investigation of this issue with
Kenyon cells (KCs), the intrinsic neurons of the mushroom body neuropil and one of the most exten-
sively studied category of neurons given their key role in memory formation and retrieval (reviewed
in Kahsai and Zars, 2011).
There are around 2,000 KCs in each mushroom body Aso et al. 2009, and they form the medial
lobe, consisting of the γ, β′ and β lobes, the vertical lobe, consisting of the α and α′ lobes, the calyx,
where dendrites are found and around which cell bodies are positioned, and the peduncle, formed by
the anterior projection of the axons before joining the lobes (Figure 4A). Three main classes of KCs
and a few subclasses of neurons are recognized: the γ neurons are the first to be born and innervate
only the γ lobe; the α′/β′ neurons are generated next and project to the α′ and β′ lobes; the α/β neurons
are the last to be born, and project to both the α and β lobes. Four neuroblast clones which differ in
their position in the calyx generate the KCs, with each one generating the whole repertoire of neuron
types (Lee et al., 1999).
We started with a dataset of 1,664 KCs, representing 10.3 % of the FlyCircuit dataset (for details
of dataset collection see Supplemental Results) and collected raw NBLAST scores of each KC against
all others. An iterative hierarchical clustering approach allowed us to identify the main KC types and
subsequent additional analysis for each of these distinguished several subtypes.
In the case of the γ neurons (Figure 4B’), we identified 2 subsets, one corresponding to the classical
morphology (Figure 4B”) (groups I and III) and another to previously described atypical neurons (γ
dorsal neurons, group II) (Aso et al., 2014). Analysis of the classical γ neurons revealed that there were
differences between the neurons in their medial to lateral position in the calyx (groups A-D). These
differences correlated to a certain degree with differences in the dorsal/ventral position of the projec-
tions in the γ lobe, with the most medial, being also the most dorsal (Figure 4B”). These observations
suggest that the relative position of the projections of classical γ neurons is maintained at the calyx and
γ lobe. We experimented with clustering the classic γ neurons based only on the scores of the segments
in the peduncle. The overall organization almost fully recapitulated the positioning of the neurites in
the whole neuron analysis (see Figure S3 and ). Thus, the stereotypical organization of the classical γ
neurons is maintained throughout the neuropil.
The atypical γ neurons extended neurites posteriolaterally in the calyx and projected to the most
dorsal region of the γ lobe (Figure 4B”’). We isolated a previously identified subtype –γd neurons
(group a) (Aso et al., 2009)– that innervates the ventral accessory calyx (Aso et al., 2014). In addition,
we identified previously uncharacterized types (Groups b-c).
Analysis of α′/β′ neurons highlighted the characterized subtypes of these neurons (Figure 4C–C’)
which differ in their anterior/posterior position in the peduncle and β′ lobe (Tanaka et al., 2008; Aso
et al., 2014). Although we were unable to unambiguously assign a α′/β′ subtype to our neurons, there
were clear trends, with a subset of neurons displaying neurites more anteriorly than others (groups ii,
iii) in both the peduncle and β′ lobe.
The largest subset of KCs corresponds to α/β neurons (Figure 4D). We identified neurons from each
of the four neuroblast lineages (Figure 4D’) (Zhu et al., 2003) and for each of these, we distinguished
morphological subtypes that correlate to their birth time (Figure 4D”). There was a clear distinction
between the late born core (α/β core, α/β-c), on the inside stratum of the α lobe, and early born peripheral
7
this preprint is the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for; http://dx.doi.org/10.1101/006346doi: bioRxiv preprint first posted online August 9, 2014;
neurons (α/β surface, α/β-s), on the outside stratum of the α lobe. We also identified the earliest born
α/β neurons– α/β posterior or pioneer (α/βp)– that innervate the accessory calyx and run along the
surface of the posterior peduncle into the β lobe but stop before reaching the medial tip (Tanaka et al.,
2008). A new clustering based on peduncle position of the neuron segments did not recapitulate the
relative positions of the calyx neurites for each of the neuroblast clones observed in the whole neuron
analysis suggesting that the relative position of the α/β neurons in the peduncle does not completely
reflect their stereotypical organization in the calyx (see Figure S3 and ).
In summary, the hierarchical clustering of KCs using the raw NBLAST scores resolved the neurons
into the previously described KC types and some of the subtypes, and isolated uncharacterized subtypes
in an extensively studied cell population. In addition, it revealed organizational principles that have
been previously described (Tanaka et al., 2008). These observation support our claim that the NBLAST
scores retain enough morphological information to accurately search for similar neurons and organize
large datasets of related cells.
NBLAST identifies classic cell types at the finest level: olfactory projection neurons
We have shown that clustering based on NBLAST scores can identify the major classes and subtypes
of Kenyon cells. However it is rather unclear what corresponds to an identified cell type, which we
take to be the finest classification of neuron in the brain. We therefore analyzed a different neuron
family, the olfactory projection neurons (PNs), which represent one of the best defined cell types in
the fly brain.
PNs transmit information between antennal lobe glomeruli, which receive sensory input, and higher
olfactory brain centers, including the mushroom body and the lateral horn (Masse et al., 2009). Uniglomeru-
lar PNs (uPNs), whose dendrites innervate just one glomerulus, are highly stereotyped in both mor-
phology and developmental origin. They are classified into individual types based on the glomerulus
they innervate and the axon tract they follow; these features show fixed relationships with their axonal
branching patterns in higher brain centers and their parental neuroblast (Marin et al., 2002; Jefferis
et al., 2001; Wong et al., 2002; Jefferis et al., 2007; Yu et al., 2010a; Tanaka et al., 2012).
We manually classified the 400 uPNs in the FlyCircuit dataset by glomerulus, and defined the
manual gold standard annotations in an iterative process that took several days (for details see ). We
found a very large number of DL2 uPNs (145 DL2d and 37 DL2v neurons), in a total of 397 neurons.
Nevertheless, our final set of uPNs broadly represents the total variability of described classes and
contains neurons innervating 35 out of 56 different glomeruli (Tanaka et al., 2012), as well as examples
of the three main lineage clones and tracts.
We computed mean NBLAST scores for each uPN against the remaining 16,128 neurons and
checked whether the top hit was exactly the same type of uPN, another uPN or a match to another
class of neuron (Figure 5A). We restricted our analysis to types with at least two examples in the
dataset and to unique pairs (i.e. if PN A was the top hit for PN B and vice versa, we only counted them
once) (n=327). There were only 8 cases in which the top hit did not match the class of the query. Of
these, four had matches to a uPN innervating a neighboring glomerulus with identical axon projections
(DL2d vs DL2v, VM5d vs VM5v) that are challenging even for experts to distinguish. There were a
further four matches to neurons that were not uPNs, but corresponded to other PNs that innervated the
same glomerulus.
We also compared how the top 3 hits matched the query type (Figure 5B). For uPN types with
8
this preprint is the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for; http://dx.doi.org/10.1101/006346doi: bioRxiv preprint first posted online August 9, 2014;
more than three examples (non DL2, n=187), we collected the top three NBLAST hits for each of
these neurons. We achieved very high matching rates: in 98.9 % of cases (i.e. all except two) at least
one of the top hits matched the query type, and all three hits matched the query type 95.2% of cases.
Given the very high prediction accuracy, we wondered if an unsupervised clustering based on
NBLAST mean scores would group uPNs by type. To test this, we clustered uPNs (non DL2, n=214)
and divided the dendrogram at a height of 0.725, as this level was found to be the one at which most
groups corresponded to single and unique neuron types. For types with more than one representative
neuron, all neurons co-clustered, with three exceptions (Figure 5C). The cluster organization also re-
flects higher level features such as the axon tract / neuroblast of origin (Figure 5C’). Thus, unsupervised
clustering of uPNs based on NBLAST scores gives an almost perfect neuronal classification.
In conclusion, these results demonstrate that morphological comparison by NBLAST is powerful
enough to resolve differences at the finest level of neuronal classification. Furthermore, they suggest
that unsupervised clustering by NBLAST scores could help to reveal new neuronal types.
NBLAST can be used to define new cell types
Visual projection neurons
Visual projection neurons (VPNs) relay information between the optic lobes and the central brain. They
are a morphologically diverse group that innervate distinct optic lobe and central brain neuropils, with
44 types already described (Otsuna and Ito, 2006). We explored whether clustering of these neurons
based on NBLAST scores would find previously reported neuron classes and identify new ones.
We isolated a set of VPNs including 1,793 unilateral VPNs, 72 bilateral VPNs and 2892 intrinsic
optic lobe neurons. Hierarchical clustering of the unilateral VPNs (uVPNs) resulted in a dendrogram,
which we divided into 21 groups (I-XXI), in order to isolate one or a few cell types by group based
both directly on morphological stereotypy and on our reading of the previous literature (Otsuna and
Ito, 2006)(Figure 6A–A’ and Figure S4A–A’). We further investigated these groups to determine if
central brain innervation was a major differentiating characteristic between classes.
Lobula-, AOTU- and PVLP-innervating uVPNs We took neuron skeletons from groups I–III
uVPN and isolated only the axon arbors innervating the anterior optic tubercle (AOTU) and posterior
ventrolateral protocerebrum (PVLP) (Figure 6B). A new clustering based on the scores of these par-
tial skeletons, allowed us to identify seven different groups (1-7). A clear distinction between neurons
that innervated the PVLP (groups 1, 2) and those that extensively innervated the AOTU (groups 3–6)
was evident. Our analysis divides the LC10 uVPN class into 5 subgroups, four of them not previously
identified ( Table S1 and Figure 6C).
Lobula-, PVLP- and PLP-innervating uVPNs We performed a similar analysis with uVPN groups
that had dendritic innervation restricted to the lobula and axons projecting to the PVLP and posterior
lateral protocerebrum (PLP) (groups IV, VI, VII and XI) (Figure S4A’). Following the same strat-
egy, we re-clustered neurons based on the scores calculated only for the axon arbors that overlapped
with the PVLP or PLP (Figure S4B). We also obtained seven distinct types, including a new subtype
(Figure S4B–C and Table S2).
9
this preprint is the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for; http://dx.doi.org/10.1101/006346doi: bioRxiv preprint first posted online August 9, 2014;
Bilateral VPNs In addition to the analysis of the uVPNs, we also performed a hierarchical clustering
of the bilateral VPNs (Figure 6A”). Of the resulting 8 distinct groups (i–viii), we were able to match
one to the bilateral LC14 neurons (Otsuna and Ito, 2006).
VPN summary Our analysis of VPN neurons has demonstrated that similarity searches performed
with only part of a neuron are useful to highlight morphological features that might be most important
for defining neuron classes. We were able to match 11 of our defined groups to known VPN types,
and furthermore described two new subclasses and four subtypes of uVPNs, showing that this type of
analysis is able to identify new cell types even for intensively studied neuronal classes.
Auditory neurons
Auditory projection neurons (PNs) are characterized by their innervation of the primary or secondary
auditory neuropils, the antennal mechanosensory and motor center (AMMC) and the inferior ventrolat-
eral protocerebrum (IVLP or wedge). Several distinct types have been described based on anatomical
and physiological features (Yorozu et al., 2009; Lai et al., 2012; Kamikouchi et al., 2006, 2009). Just as
for the VPNs, we tested our ability to identify known and new cell types. In this case, we employed a
two-step search strategy using a previously identified FlyCircuit neuron named by Lai et al. (2012) as
the seed neuron for the first search (for details see ). For each of the 5 types we analyzed, hierarchical
clustering of the hits revealed new subtypes of known auditory PN types that differed mainly in their
lateral arborizations (Figure S5E and Table S3 ).
mAL neurons
The fruitless-expressing mAL neurons are sexually dimorphic interneurons that are known to regulate
wing extension by males during courtship song (Koganezawa et al., 2010; Kimura et al., 2005). Males
have around 30 neurons, but there are only 5 in females. Although the gross neuronal morphology
is similar in both sexes, both axonal and dendritic arborisations are located in distinct regions, likely
altering input and output connectivity. We investigated whether clustering could distinguish male and
female neurons and identify male subtypes. Clustering a set of 41 mAL neurons (for dataset collection
see ) cleanly separated male from female neurons (Figure 7A–B). Clustering analysis of the male neu-
rons using partial skeletons that only contained the axonal and dendritic arbors (Figure 7C), identified
3 main types and 2 subtypes of male mAL neurons. The 3 types differed in the length of the ipsilateral
ventral projection; this feature has previously been proposed as the basis of a qualitative classification
of mAL neurons (Kimura et al., 2005). However all types and subtypes also showed reproducible dif-
ferences in the exact location of their axon terminal arbors. Our analysis therefore suggests that the
population of male neurons includes types with correlated differences in input-output connectivity.
P1 neurons
P1 neurons are the most significantly dimorphic fruitless-expressing neurons. Male P1 neurons are
involved in the initiation of male courtship behavior while female P1 neurons degenerate during de-
velopment due to the action of doublesex (Kimura et al., 2008). There are around 20 P1 neurons that
develop from the pMP-e fruitless neuroblast clone. They have extensive bilateral arborizations in the
10
this preprint is the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for; http://dx.doi.org/10.1101/006346doi: bioRxiv preprint first posted online August 9, 2014;
PVLP and ring neuropil, partially overlapping with the male-specific enlarged brain regions (MER)
(Kimura et al., 2008; Cachero et al., 2010) (Figure 7D).
Despite their critical role, P1 neurons have been treated as an homogeneous neuronal class. We
therefore investigated whether clustering could identify anatomical subtypes. Hierarchical clustering
of the P1 neuron set (for dataset collection see ), distinguished 10 groups (Figure 7D’). Nine of these
(1–9) contain only male fru-GAL4 neurons as expected. The 9 male groups have the same distinctive
primary neurite and send contralateral axonal projections through the arch (Yu et al., 2010b) with
extensive arborizations in the MER regions (Cachero et al., 2010). However each group shows a highly
distinctive pattern of dendritic and axonal arborisations suggesting that they are likely to integrate
distinct sensory inputs and to connect with distinct downstream targets.
Intriguingly, group 10 consists only of female neurons, including two female fru-GAL4 and two
other drivers. Their morphology is clearly similar to but distinct from group 9 neurons, suggesting that
a small population of neurons that share anatomical (and likely developmental) features with male P1
neurons is also present in females.
Superclusters and Exemplars to Organize Huge Data
In the previous examples we have shown that NBLAST is a powerful tool to identify known and un-
cover new neuron types when analyzing specific neuron superclasses within large datasets. However,
subsetting the dataset in order to isolate the chosen neurons requires considerable time. We wished
to establish a method that would allows us to organize large datasets, extracting the main types au-
tomatically, and retain information on the similarity between types and subtypes, while providing a
quicker way to navigate datasets. We used the affinity propagation method of clustering (Frey and
Dueck, 2007), combined with hierarchical clustering to achieve this. Applying affinity propagation to
the 16,129 neurons in the FlyCircuit dataset resulted in 1,052 clusters (Figure 8A–B). Using hierarchi-
cal clustering of the exemplars and by manually removing a few stray neurons, we isolated the central
brain neurons (groups B–C) from the optic lobe and visual projection ones (group A) (Figure 8C). An-
other step of hierarchical clustering of central brain exemplars revealed large superclasses of neuron
types when we divided the dendrogram in 14 groups (I–XIV). Each one contained a distinct subset of
neuron types including, for example, central complex neurons (I), P1 neurons (II), 2 groups of KCs (γ
and α′/β′ and α/β) (IV–V) and auditory neurons (VIII) (Figure 8D–D’).
The affinity propagation clusters are also useful for identifying neuronal subtypes by comparing all
clusters that contain a specified neuronal type (Figure 8E). We present examples for the neuronal types
AMMC-IVLP projection neuron 1 (AMMC-IVLP PN1) (Lai et al., 2012), and the uVPNs LC10B and
LC4. For each of these, morphological differences are clear between clusters, suggesting that each one
might help to identify distinct subtypes.
We have shown that combining affinity propagation with hierarchical clustering is an effective
way to organize and explore large datasets, by condensing information into a single exemplar and by
retaining the ability to move up or down in the hierarchical tree, allowing the analysis of superclasses
or more detailed subtypes.
11
this preprint is the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for; http://dx.doi.org/10.1101/006346doi: bioRxiv preprint first posted online August 9, 2014;
Discussion
The challenge of mapping and cataloguing the full spectrum of neuronal types in the brain depends not
only on the ability to recognize similar neurons, by shape and position, but also on establishing methods
that facilitate unbiased identification of neuron types from pools of thousands or millions of individual
neurons. The comparison of neurons relies both on morphology and position within the brain, as this
is an essential determinant of their function and synaptic partners. A neuron search algorithm should
therefore be: (1) accurate, with hits being biologically meaningful; (2) fast and computationally inex-
pensive; (3) provide an interactive search method and (4) generally applicable. Here, we have described
NBLAST, a neuron search algorithm that satisfies all these criteria.
First, the algorithm correctly distinguishes closely related subtypes across a range of major neuron
groups, with an accuracy of 97.6 % in the case of olfactory projection neurons. Unsupervised clustering
of these neurons, based on NBLAST scores, correctly organized neurons into described types. We did
find, however, that the size of a neuron influences the accuracy of algorithm, especially for smaller
neurons, even when using the normalized score. One future research area will be to convert the raw
scores that we have used into an expectation (E) value (cf. BLAST), that would directly account for
the size of a neuron.
Second, NBLAST searches are very fast, with pairwise comparisons taking about 2 ms on a laptop
computer, with queries against the whole 16,000 neuron dataset taking about 30 s. Furthermore, for
defined datasets all-by-all scores can be pre-computed allowing immediate retrieval of NBLAST scores
for highly interactive analysis. With the amount of data available only expected to increase, strategies
to query and store these data need to be investigated. One effective approach to handle much larger
number of neurons will be to compute sparse similarity matrices, storing only the top hits for a
given neuron, an approach often taken for genome-wide precomputed BLAST scores. Alternatively,
queries could be computed only against the non-redundant set of neurons that collectively embody the
structure of the brain, similarly to the strategy employed by UniProt (Suzek et al., 2007). At most, this
set would not exceed 50,000 neurons (due to the strong bilateral symmetry in the fly brain) and we
expect that it would in practice not need to exceed 5,000. Our clustering of all ~16,000 neurons of the
FlyCircuit dataset identified ~1,000 exemplars providing a non-redundant data set that could be used
for rapid searches.
Third, our method permits a variety of different search strategies from a variety of objects. Searches
with whole or neuron fragments, or tracings can be used to distinguish closely related neuronal types
by their terminal arbors or to identify candidate neurons from a GAL4 line.
Finally, one important question is obviously the extent to which our approach can be generalized.
This issue largely reduces to the relationship between the length scales of neurons being examined
and their absolute spatial stereotypy. Our method implicitly assumes spatial co-localization of related
neurons; this is enforced in our input data by the use of image registration. Our search strategy should
therefore be appropriate for any situation in which neuronal organization is highly stereotyped at the
length scale of the neuron under consideration. There is already strong evidence that this is true across
large parts of the brain for simple vertebrate models like the larval zebrafish: indeed Portugues et al.
(2014) used exactly the same registration software to demonstrate highly spatially stereotyped visuo-
motor activity patterns. Preliminary analysis (GSXEJ and JDM, https://github.com/jefferis/
nat.examples/tree/master/05-miyasaka2014) suggests that our method can be applied directly
to olfactory projectome data (Miyasaka et al., 2014) from larval zebrafish. Mouse gene expression
12
this preprint is the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for; http://dx.doi.org/10.1101/006346doi: bioRxiv preprint first posted online August 9, 2014;
(Lein et al., 2007) and long range connectivity also show global spatial stereotypy as evidenced by
recent atlas studies combining sparse labeling and image registration (Zingg et al., 2014; Susaki et al.,
2014; Oh et al., 2014)). Our method should allow simple querying and hierarchical organization of
these datasets with relatively little modification beyond calculating an appropriate scoring matrix.
However there are clearly situations in which global brain registration is not an appropriate starting
point. For example the vertebrate retina has both a laminar and a tangential organization. Recently
Sümbül et al. (2014) have introduced a registration strategy that demonstrates that the lamination of
retinal ganglion cells in mouse retina is spatially stereotyped to the nearest micron. However retinal
interneurons and ganglion cells are organized in mosaics across the retinal surface (typically referred
to as the XY plane). Therefore global registration is not appropriate in this axis, rather it is necessary
to align neurons into a virtual column. The situation is similar for Drosophila columnar neurons in
the outer neuropils of optic lobe, for which neurons are organized into about 800 parallel columns
(reviewed in Paulk et al., 2013). There are two ways that we envisage this situation can be handled.
The first would be to carry out a local re-registration, that maps each column onto a single canonical
column. The second would be to amass sufficient data that neurons from neighboring columns would
tile the brain, enabling their identification as a related group by standard clustering or graph theoretic
approaches.
The aim of cataloguing all neuron types in the brain relies not only on an accurate algorithm to
find similar neurons, but also on having an easy and unbiased method to distinguish neurons types
and/or subtypes. This is a challenging problem, but morphological approaches may eventually provide
unambiguous automated classification. Sümbül et al. (2014) recently explored the issue of defining the
optimal cut height for morphological clustering of mouse retinal ganglion cells, establishing a reliable
approach for these specific neurons. We have demonstrated the applicability of NBLAST across a very
wide range of neuronal classes using hierarchical clustering and cutting the dendrogram at a specified
height. This process of identification relied on extensive data exploration and iteration. We believe that
the extent of morphological variability within a neuronal type precludes the existence of one unique
value for dendrogram cutting height. Instead, the range of heights we found, between 0.7 and 2, can
guide future exploration for other neuron types and datasets, although this process will still require
iterative analysis and manual verification.
Experimental Procedures
Image Preprocessing
The flycircuit.tw team supplied 16,226 raw confocal stacks in the Zeiss LSM format on a single 2
TB hard drive in April 2011. Each LSM stack was first uncompressed, then read into Fiji/ImageJ
(http://fiji.sc/Fiji) where the channels were split and resaved as individual gzip compressed
NRRD (http://teem.sourceforge.net/nrrd) files. Where calibration information was missing
from the LSM file metadata, we used a voxel size of (0.318427, 0.318427, 1.00935) microns as rec-
ommended by the FlyCircuit team. There were two important issues to solve before images could
be used for registration: 1) identifying which image channel contained the anti-Dlg (discs large 1)
counterstaining revealing overall brain structure and 2) determining whether the brains had been im-
aged from anterior to posterior, or posterior to anterior. The first issue could be solved by exporting
the metadata associated with each LSM file using the LOCI bioformats (http://loci.wisc.edu/
13
this preprint is the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for; http://dx.doi.org/10.1101/006346doi: bioRxiv preprint first posted online August 9, 2014;
software/bio-formats) plugin for Fiji and developing some heuristics to automate the identifi-
cation of the channel sequence; for a minority of images this metadata was missing and the channel
order was determined manually. The second issue, slice order, could not be determined automati-
cally from the image metadata. We therefore made maximum intensity projections (using the unu
tool, http://teem.sourceforge.net/unrrdu) along the Z axis of the channel corresponding to
the labelled neuron for each stack. Each projection was then compared with the matching thumbnail
available from the flycircuit.tw website. The correlation score between the projection and thumbnail
images was calculated both with and without a mirror flip across the YZ plane; a large correlation score
for only one orientation was used as evidence for a given slice ordering. A small number of ambiguous
results were verified manually. We successfully preprocessed 16,204/16,226 total images i.e. a 0.14
% failure rate. 12 failures were due to mismatches that could not be resolved between the segmented
neuron present in the LSM file and the thumbnail image for the neuron identifier on the flycircuit.tw
website; the remaining 10 failures were due to physical offsets between the brain and GFP channels or
corrupt image data.
Template Brain
The template brain (FCWB) was constructed by screening for whole brains within the FlyCircuit
dataset, and manually selecting a pool of brains that appeared of good quality when the stacks were
inspected. Separate average female and average male template brains were constructed from 17 and 9
brains, respectively using the CMTK (http://www.nitrc.org/projects/cmtk) avg_adm func-
tion which takes a single brain as a seed. After five iterations the resultant average male and av-
erage female brains were placed in an affine symmetric position within their image stacks so that
a simple horizontal (-axis) flip of either template brain resulted in an almost perfect overlap of
left and right hemispheres. Finally the two sex-specific template brains were then averaged (with
equal weight) to make an intersex template brain using the same procedure. Since the purpose of
this template was to provide an optimal registration target for the flycircuit.tw dataset, no attempt
was made to correct for the obvious disparity between the XY and Z voxel dimensions common to
all the images in the dataset. The scripts used for the construction of the template are available at
https://github.com/jefferislab/MakeAverageBrain.
Image Registration
Image registration of the Dlg neuropil staining used a fully automatic intensity-based (landmark free)
3D image registration implemented in the CMTK toolkit available at http://www.nitrc.org/projects/
cmtk (Rohlfing and Maurer, 2003; Jefferis et al., 2007). An initial linear registration with 9 degrees of
freedom (translation, rotation and scaling of each axis) was followed by a non-rigid registration that al-
lows different brain regions to move somewhat independently, subject to a smoothness penalty (Rueck-
ert et al., 1999). It is our experience that obtaining a satisfactory initial linear registration is crucial. All
registrations were therefore manually checked by comparing the reformatted brain with the template
in Amira (academic version, Zuse Institute, Berlin), using ResultViewer https://bitbucket.org/
jefferis/resultviewer. This identified about 10 % of brains which did not register satisfactorily.
For these images a new affine registration was calculated using a Global Hough Transform (Ballard,
1981; Khoshelham, 2007) with an Amira extension module available from 1000shapes GmbH; the re-
14
this preprint is the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for; http://dx.doi.org/10.1101/006346doi: bioRxiv preprint first posted online August 9, 2014;
sult of this affine transform was again manually inspected. In the minority of cases where this approach
failed, a surface based alignment was calculated in Amira after manually aligning the two brains. Once
a satisfactory initial affine registration was obtained, a non-rigid registration was calculated for all
brains. Finally each registration was checked manually in Amira against the template brain. The result
of this sequential procedure was that we successfully registered 16,129/16,204 preprocessed images,
giving a registration failure rate of 0.46 %.
Image Postprocessing
The confocal stack for each neuron available at http://flycircuit.tw includes an 8 bit image
containing a single (semiautomatically) segmented neuron prepared by Chiang et al. (2011). This image
was downsampled by a factor of 2 in and , binarized with a threshold of 1 and then skeletonized using
the Fiji plugin ’Skeletonize (2D/3D)’ (Doube et al., 2010). Dot properties for each neuron skeleton
were extracted following the method in Masse et al. (2012), using the dotprops function of our new
nat package for R. This converted each skeleton into segments, described by its location and tangent
vector. Neurons on the right side of the brain were flipped to left by applying a mirroring and a flipping
registration as described in Manton et al. (2014). The decision of whether to flip a neuron depended on
earlier assignment of each neuron to a brain hemisphere using a combination of automated and manual
approaches. Neurons whose cell bodies were more than 20 µm away from the mid-sagittal YZ plane
were automatically defined as belonging to the left or right hemisphere. Neurons whose cell bodies
were inside this 40 µm central corridor were manually assigned to the left or right sides, based on the
position of the cell body (right or left side), path taken by the primary neurite, location and length
of first branching neurite. For example, neurons that had a cell body on the midline with significant
innervation from the first branching neurite near the cell body on the left hemisphere, with the rest of
the arborisation on the right, were assigned to the left side and not flipped. On the other hand, neurons
with similar morphology to these but in which the first branching neurite is small, compared to the total
innervation, were assigned to the right and flipped. The cell body positions used were based on those
published on the http://flycircuit.tw website for each neuron; these positions are in the space
of the FlyCircuit female and male template brains (typical_brain_female and typical_brain_male). In
order to transform them into the FCWB template that we constructed, affine bridging registrations
were constructed from the typical_brain_female and typical_brain_male brains to FCWB and the cell
body positions were then transformed to this new space. Since these cell body positions depend on two
affine registrations (one conducted by Chiang et al. (2011) to register each sample brain onto either
their typical_brain_female or typical_brain_male templates and a second carried out by us to map
those template brains onto our FCWB template) these positions are likely accurate only to ±5 microns
in each axis.
Neuron Search
The neuron search algorithm is described in detail in Results and Figure 1. The reference implemen-
tation that we have written is the nblast function in the R package nat.nblast, which depends on our
nat package (Jefferis and Manton, 2014). Fast nearest neighbor search, an essential primitive for the
algorithm uses the RANN package (Jefferis, 2014), a wrapper for the Approximate Nearest Neighbor
(ANN) C++ library (Mount and Arya, 2006). The scoring matrix that we used for FlyCircuit neurons
15
this preprint is the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for; http://dx.doi.org/10.1101/006346doi: bioRxiv preprint first posted online August 9, 2014;
was constructed by taking 150 DL2 projection neurons, which define a neuron type at the finest level,
and calculating the joint histogram of distance and absolute dot product for the 150 × 149 combina-
tions of neurons, resulting in 1.4 × 10
7
measurement pairs; the number of counts in the histogram was
then normalized (i.e. dividing by 1.4 × 10
7
) to give a probability density,
. We then carried
out a similar procedure for 5,000 random pairs of neurons sampled from the FlyCircuit dataset to give
. Finally the scoring matrix was calculated as
where ε (a pseudocount to avoid
infinite values) was set to 1 × 10
−6
.
Clustering
We employed two different methods for clustering based on normalized NBLAST scores. We used
Ward’s method for hierarchical clustering, using the default implementation in the R function hclust.
This method minimizes the total within-cluster variance, and at each step the pair of entities or clusters
with the minimum distance between clusters are merged (Ward Jr, 1963). The resulting dendrograms
were cut at a single selected height chosen for each case to separate neuron types or subtypes. This
value is shown as a dashed line in all dendrograms. By default, R plots the square of the Euclidean
distance as the axis, but in the plots shown, the height of the dendrogram corresponds to the unsquared
distance.
For the analysis of the whole dataset, we used the affinity propagation method. This is an iterative
method which finds exemplars which are representative members of each cluster and does not require
any a priori input on the final number of clusters (Frey and Dueck, 2007) as implemented in the R
package apcluster (Bodenhofer et al., 2011). The input preference parameter () can be set before
running the clustering. This parameter reflects the tendency of data samples to become an exemplar,
and affects the final number of clusters. In our analysis, we used , since this is the value where on
average matched segments are equally likely to have come from matching and non-matching neurons.
Empirically this parameter produced clusters that, for the most part, grouped neurons of the same type
according to biological expert opinion.
Neuron Tracing
Neuron tracing was carried out in Amira (commercial version, FEI Visualization Sciences Group,
Merignac, France) using the hxskeletonize plugin (Evers et al., 2005) or in Vaa3D (Peng et al., 2014)
on previously registered image data. Traces were then loaded into R using the nat package. When
necessary, they were transformed into the space of the FCWB template brain using the approach of
Manton et al. (2014).
Computer Code and Data
The image processing pipeline and analysis code used two custom packages for the R statistical en-
vironment (http://www.r-project.org) https://github.com/jefferis/nat and https://
github.com/jefferis/nat.as that coordinated processing by the low level registration (CMTK)
and image processing (Fiji, unu) software mentioned above. NBLAST neuron search is implemented
in a third R package available at https://github.com/jefferislab/nat.nblast. Analysis code
specific to the flycircuit dataset is available in a dedicated R package https://github.com/jefferis/
16
this preprint is the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for; http://dx.doi.org/10.1101/006346doi: bioRxiv preprint first posted online August 9, 2014;
flycircuit, with a package vignette showcasing the main tools that we have developed. Further de-
tails of these supplemental software and the associated data are presented at http://jefferislab.
org/si/nblast. The registered image dataset can be viewed in the stack viewer of the http://
virtualflybrain.org website and all 16,129 registered single neuron images will be available at
https://jefferislab.org/si/nblast or on request to GSXEJ on a hard drive; the unregistered
data remain available at http://flycircuit.tw.
Acknowledgments
We first of all acknowledge the flycircuit.tw team for generously providing the raw image data associ-
ated with Chiang et al. (2011). Images from FlyCircuit were obtained from the NCHC (National Center
for High-performance Computing) and NTHU (National Tsing Hua University), Hsinchu, Taiwan. We
thank members of the Jefferis lab for comments on the manuscript, Jake Grimmett and Toby Darling
for assistance with the LMB’s compute cluster and Torsten Rohlfing for discussions about image anal-
ysis and registration. We thank the Virtual Fly Brain project for their help in linking and incorporating
some of the results of this study in the http://virtualflybrain.org website.
This study made use of the Computational Morphometry Toolkit, supported by the National Insti-
tute of Biomedical Imaging and Bioengineering. This work was supported by the Medical Research
Council [MRC file reference U105188491] and a European Research Council Starting Investigator
Grant to GSXEJ, who is an EMBO Young Investigator.
17
this preprint is the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for; http://dx.doi.org/10.1101/006346doi: bioRxiv preprint first posted online August 9, 2014;
References
Altschul, S.F., Gish, W., Miller, W., Myers, E.W., and Lipman, D.J. (1990). Basic local alignment
search tool. Journal of Molecular Biology 215, 403–410.
Aso, Y., Grübel, K., Busch, S., Friedrich, A.B., Siwanowicz, I., and Tanimoto, H. (2009). The mush-
room body of adult Drosophila characterized by GAL4 drivers. Journal of neurogenetics 23, 156–
172.
Aso, Y., Hattori, D., Yu, Y., Johnston, R.M., Iyer, N.A., Ngo, T.T., Dionne, H., Abbott, L., Axel, R.,
Tanimoto, H., et al. (2014). The neuronal architecture of the mushroom body provides a logic for
associative learning. eLife 3, e04577.
Badea, T.C., and Nathans, J. (2004). Quantitative analysis of neuronal morphologies in the mouse
retina visualized by using a genetically directed reporter. Journal of Comparative Neurology 480,
331–351.
Ballard, D.H. (1981). Generalizing the Hough transform to detect arbitrary shapes. Pattern Recognition
13, 111–122.
Basu, S., Condron, B., and Acton, S.T. (2011). Path2Path: hierarchical Path-Based analysis for neuron
matching. In Biomedical Imaging: From Nano to Macro, 2011 IEEE International Symposium on
(IEEE), pp. 996–999.
Bodenhofer, U., Kothmeier, A., and Hochreiter, S. (2011). APCluster: an R package for affinity prop-
agation clustering. Bioinformatics 27, 2463–2464.
Bota, M., and Swanson, L.W. (2007). The neuron classification problem. Brain research reviews 56,
79–88.
Brown, K.M., Barrionuevo, G., Canty, A.J., De Paola, V., Hirsch, J.A., Jefferis, G.S.X.E., Lu, J.,
Snippe, M., Sugihara, I., and Ascoli, G.A. (2011). The DIADEM Data Sets: Representative Light
Microscopy Images of Neuronal Morphology to Advance Automation of Digital Reconstructions.
Neuroinformatics .
Cachero, S., Ostrovsky, A.D., Yu, J.Y., Dickson, B.J., and Jefferis, G.S.X.E. (2010). Sexual dimor-
phism in the fly brain. Curr Biol 20, 1589–601.
Cajal, S.R., and Azoulay y, L. (1911). Histologie du système nerveux de l’homme et des vertébrés (A.
Maloine).
Cardona, A., Saalfeld, S., Arganda, I., Pereanu, W., Schindelin, J., and Hartenstein, V. (2010). Identi-
fying neuronal lineages of Drosophila by sequence analysis of axon tracts. J Neurosci 30, 7538–53.
Chiang, A.S., Lin, C.Y., Chuang, C.C., Chang, H.M., Hsieh, C.H., Yeh, C.W., Shih, C.T., Wu, J.J.,
Wang, G.T., Chen, Y.C., Wu, C.C., Chen, G.Y., Ching, Y.T., Lee, P.C., Lin, C.Y., Lin, H.H., Wu, C.C.,
Hsu, H.W., Huang, Y.A., Chen, J.Y., Chiang, H.J., Lu, C.F., Ni, R.F., Yeh, C.Y., and Hwang, J.K.
(2011). Three-dimensional reconstruction of brain-wide wiring networks in Drosophila at single-cell
resolution. Curr Biol 21, 1–11.
18
this preprint is the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for; http://dx.doi.org/10.1101/006346doi: bioRxiv preprint first posted online August 9, 2014;
Conte, D., Foggia, P., Sansone, C., and Vento, M. (2004). Thirty years of graph matching in pattern
recognition. International journal of pattern recognition and artificial intelligence 18, 265–298.
Coombs, J., Van Der List, D., Wang, G.Y., and Chalupa, L. (2006). Morphological properties of mouse
retinal ganglion cells. Neuroscience 140, 123–136.
Doube, M., Kłosowski, M.M., Arganda-Carreras, I., Cordelières, F.P., Dougherty, R.P., Jackson, J.S.,
Schmid, B., Hutchinson, J.R., and Shefelbine, S.J. (2010). BoneJ: Free and extensible bone image
analysis in ImageJ. Bone 47, 1076 – 1079.
El Jundi, B., Heinze, S., Lenschow, C., Kurylas, A., Rohlfing, T., and Homberg, U. (2009). The locust
standard brain: a 3D standard of the central complex as a platform for neural network analysis.
Frontiers in systems neuroscience 3.
Evers, J.F., Schmitt, S., Sibila, M., and Duch, C. (2005). Progress in functional neuroanatomy: pre-
cise automatic geometric reconstruction of neuronal morphology from confocal image stacks. J
Neurophysiol 93, 2331–42.
Fischbach, K.F., and Dittrich, A. (1989). The optic lobe of Drosophila melanogaster. I. A Golgi analysis
of wild-type structure. Cell and Tissue Research 258, 441–475.
Frey, B.J., and Dueck, D. (2007). Clustering by passing messages between data points. science 315,
972–976.
Ganglberger, F., Schulze, F., Tirian, L., Novikov, A., Dickson, B., Bühler, K., and Langs, G. (2014).
Structure-Based Neuron Retrieval Across Drosophila Brains. Neuroinformatics .
Ito, K., Shinomiya, K., Ito, M., Armstrong, J.D., Boyan, G., Hartenstein, V., Harzsch, S., Heisenberg,
M., Homberg, U., Jenett, A., et al. (2014). A systematic nomenclature for the insect brain. Neuron
81, 755–765.
Jefferis, G. (2014). RANN k nearest neighbour search v2.3.0. Zenodo .
Jefferis, G.S.X.E., and Manton, J.D. (2014). nat: NeuroAnatomy Toolbox R package. Zenodo .
Jefferis, G.S.X.E., Potter, C.J., Chan, A.M., Marin, E.C., Rohlfing, T., Maurer, C.R.J., and Luo, L.
(2007). Comprehensive maps of Drosophila higher olfactory centers: spatially segregated fruit and
pheromone representation. Cell 128, 1187–1203.
Jefferis, G.S.X.E., and Livet, J. (2012). Sparse and combinatorial neuron labelling. Curr Opin Neuro-
biol 22, 101–10.
Jefferis, G.S., Marin, E.C., Stocker, R.F., and Luo, L. (2001). Target neuron prespecification in the
olfactory map of Drosophila. Nature 414, 204–208.
Jenett, A., Rubin, G.M., Ngo, T.T., Shepherd, D., Murphy, C., Dionne, H., Pfeiffer, B.D., Cavallaro,
A., Hall, D., Jeter, J., et al. (2012). A GAL4-Driver Line Resource for Drosophila Neurobiology.
Cell reports 2, 991–1001.
19
this preprint is the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for; http://dx.doi.org/10.1101/006346doi: bioRxiv preprint first posted online August 9, 2014;
Kahsai, L., and Zars, T. (2011). Learning and memory in Drosophila: behavior, genetics, and neural
systems. Int Rev Neurobiol 99, 139–67.
Kamikouchi, A., Inagaki, H.K., Effertz, T., Hendrich, O., Fiala, A., Göpfert, M.C., and Ito, K. (2009).
The neural basis of Drosophila gravity-sensing and hearing. Nature 458, 165–171.
Kamikouchi, A., Shimada, T., and Ito, K. (2006). Comprehensive classification of the auditory sensory
projections in the brain of the fruit fly Drosophila melanogaster. Journal of Comparative Neurology
499, 317–356.
Kepecs, A., and Fishell, G. (2014). Interneuron cell types are fit to function. Nature 505, 318–26.
Khoshelham, K. (2007). Extending Generalized Hough Transform to Detect 3D Objects in Laser
Range Data. In ISPRS Workshop on Laser Scanning, Proceedings, LS 2007. pp. 206–210.
Kimura, K.i., Hachiya, T., Koganezawa, M., Tazawa, T., and Yamamoto, D. (2008). Fruitless and
doublesex coordinate to generate male-specific neurons that can initiate courtship. Neuron 59, 759–
769.
Kimura, K.I., Ote, M., Tazawa, T., and Yamamoto, D. (2005). Fruitless specifies sexually dimorphic
neural circuitry in the Drosophila brain. Nature 438, 229–233.
Koganezawa, M., Haba, D., Matsuo, T., and Yamamoto, D. (2010). The Shaping of Male Courtship
Posture by Lateralized Gustatory Inputs to Male-Specific Interneurons. Current Biology 20, 1–8.
Kong, J.H., Fish, D.R., Rockhill, R.L., and Masland, R.H. (2005). Diversity of ganglion cells in the
mouse retina: unsupervised morphological classification and its limits. Journal of Comparative
Neurology 489, 293–310.
Lai, J.S.Y., Lo, S.J., Dickson, B.J., and Chiang, A.S. (2012). Auditory circuit in the Drosophila brain.
Proc Natl Acad Sci U S A 109, 2607–12.
Le, Q., Ranzato, M., Monga, R., Devin, M., Chen, K., Corrado, G., Dean, J., and Ng, A. (2012).
Building high-level features using large scale unsupervised learning. In International Conference in
Machine Learning.
Lee, T.C., Kashyap, R.L., and Chu, C.N. (1994). Building Skeleton Models via 3-D Medial Sur-
face/Axis Thinning Algorithms. CVGIP: Graph. Models Image Process. 56, 462–478.
Lee, T., Lee, A., and Luo, L. (1999). Development of the Drosophila mushroom bodies: sequential
generation of three distinct types of neurons from a neuroblast. Development 126, 4065–4076.
Lein, E.S., Hawrylycz, M.J., Ao, N., Ayres, M., Bensinger, A., Bernard, A., Boe, A.F., Boguski, M.S.,
Brockway, K.S., Byrnes, E.J., et al. (2007). Genome-wide atlas of gene expression in the adult
mouse brain. Nature 445, 168–176.
Lin, H.H., Lai, J.S.Y., Chin, A.L., Chen, Y.C., and Chiang, A.S. (2007). A map of olfactory represen-
tation in the Drosophila mushroom body. Cell 128, 1205–1217.
20
this preprint is the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for; http://dx.doi.org/10.1101/006346doi: bioRxiv preprint first posted online August 9, 2014;
Manton, J.D., Ostrovsky, A.D., Goetz, L., Costa, M., Rohlfing, T., and Jefferis, G.S.X.E. (2014). Com-
bining genome-scale Drosophila 3D neuroanatomical data by bridging template brains. Bioarxiv
preprint .
Marin, E.C., Jefferis, G.S., Komiyama, T., Zhu, H., and Luo, L. (2002). Representation of the Glomeru-
lar Olfactory Map in the Drosophila Brain. Cell 109, 243–255.
Masse, N.Y., Turner, G.C., and Jefferis, G.S.X.E. (2009). Olfactory information processing in
Drosophila. Curr Biol 19, R700–13.
Masse, N.Y., Cachero, S., Ostrovsky, A., and Jefferis, G.S.X.E. (2012). A mutual information approach
to automate identification of neuronal clusters in Drosophila brain images. Frontiers in Neuroinfor-
matics 6.
Mayerich, D., Bjornsson, C., Taylor, J., and Roysam, B. (2012). NetMets: software for quantifying
and visualizing errors in biological network segmentation. BMC Bioinformatics 13 Suppl 8, S7.
Migliore, M., and Shepherd, G.M. (2005). An integrated approach to classifying neuronal phenotypes.
Nature Reviews Neuroscience 6, 810–818.
Miyasaka, N., Arganda-Carreras, I., Wakisaka, N., Masuda, M., Sümbül, U., Seung, H.S., and Yoshi-
hara, Y. (2014). Olfactory projectome in the zebrafish forebrain revealed by genetic single-neuron
labelling. Nat Commun 5, 3639.
Morante, J., and Desplan, C. (2008). The Color-Vision Circuit in the Medulla of Drosophila. Current
Biology 18, 553–565.
Mount, D.M., and Arya, S. (2006). ANN: A Library for Approximate Nearest Neighbor Searching.
Version 1.1.1.
Nelson, S.B., Sugino, K., and Hempel, C.M. (2006). The problem of neuronal cell types: a physiolog-
ical genomics approach. Trends Neurosci 29, 339–45.
Oh, S.W., Harris, J.A., Ng, L., Winslow, B., Cain, N., Mihalas, S., Wang, Q., Lau, C., Kuan, L., Henry,
A.M., et al. (2014). A mesoscale connectome of the mouse brain. Nature 508, 207–214.
Otsuna, H., and Ito, K. (2006). Systematic analysis of the visual projection neurons of Drosophila
melanogaster. I. Lobula-specific pathways. J Comp Neurol 497, 928–958.
Parekh, R., and Ascoli, G.A. (2013). Neuronal morphology goes digital: a research hub for cellular
and system neuroscience. Neuron 77, 1017–38.
Paulk, A., Millard, S.S., and van Swinderen, B. (2013). Vision in Drosophila: seeing the world through
a model’s eyes. Annual review of entomology 58, 313–332.
Pearson, W.R., and Lipman, D.J. (1988). Improved tools for biological sequence comparison. Proc
Natl Acad Sci U S A 85, 2444–8.
21
this preprint is the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for; http://dx.doi.org/10.1101/006346doi: bioRxiv preprint first posted online August 9, 2014;
Peng, H., Tang, J., Xiao, H., Bria, A., Zhou, J., Butler, V., Zhou, Z., Gonzalez-Bellido, P.T., Oh, S.W.,
Chen, J., Mitra, A., Tsien, R.W., Zeng, H., Ascoli, G.A., Iannello, G., Hawrylycz, M., Myers, E.,
and Long, F. (2014). Virtual finger boosts three-dimensional imaging and microsurgery as well as
terabyte volume image visualization and analysis. Nat Commun 5, 4342.
Petilla Interneuron Nomenclature Group, Ascoli, G.A., Alonso-Nanclares, L., Anderson, S.A., Bar-
rionuevo, G., Benavides-Piccione, R., Burkhalter, A., Buzsáki, G., Cauli, B., Defelipe, J., Fairén,
A., Feldmeyer, D., Fishell, G., Fregnac, Y., Freund, T.F., Gardner, D., Gardner, E.P., Goldberg,
J.H., Helmstaedter, M., Hestrin, S., Karube, F., Kisvárday, Z.F., Lambolez, B., Lewis, D.A., Marin,
O., Markram, H., Muñoz, A., Packer, A., Petersen, C.C.H., Rockland, K.S., Rossier, J., Rudy, B.,
Somogyi, P., Staiger, J.F., Tamas, G., Thomson, A.M., Toledo-Rodriguez, M., Wang, Y., West, D.C.,
and Yuste, R. (2008). Petilla terminology: nomenclature of features of GABAergic interneurons of
the cerebral cortex. Nat Rev Neurosci 9, 557–68.
Portugues, R., Feierstein, C.E., Engert, F., and Orger, M.B. (2014). Whole-brain activity maps reveal
stereotyped, distributed networks for visuomotor behavior. Neuron 81, 1328–1343.
Rohlfing, T., and Maurer, C. R., J. (2003). Nonrigid image registration in shared-memory multipro-
cessor environments with application to brains, breasts, and bees. IEEE Trans Inf Technol Biomed
7, 16–25.
Rowe, M., and Stone, J. (1976). Naming of neurones. Classification and naming of cat retinal ganglion
cells. Brain, behavior and evolution 14, 185–216.
Rueckert, D., Sonoda, L.I., Hayes, C., Hill, D.L., Leach, M.O., and Hawkes, D.J. (1999). Nonrigid reg-
istration using free-form deformations: application to breast MR images. IEEE Trans Med Imaging
18, 712–21.
Rybak, J., Kuß, A., Lamecker, H., Zachow, S., Hege, H.C., Lienhard, M., Singer, J., Neubert, K.,
and Menzel, R. (2010). The digital bee brain: integrating and managing neurons in a common 3D
reference system. Frontiers in systems neuroscience 4.
Schindelin, J., Arganda-Carreras, I., Frise, E., Kaynig, V., Longair, M., Pietzsch, T., Preibisch, S., Rue-
den, C., Saalfeld, S., Schmid, B., Tinevez, J.Y., White, D.J., Hartenstein, V., Eliceiri, K., Tomancak,
P., and Cardona, A. (2012). Fiji: an open-source platform for biological-image analysis. Nat Meth-
ods 9, 676–82.
Sonnhammer, E.L., Eddy, S.R., Durbin, R., et al. (1997). Pfam: a comprehensive database of protein
domain families based on seed alignments. Proteins-Structure Function and Genetics 28, 405–420.
Sümbül, U., Song, S., McCulloch, K., Becker, M., Lin, B., Sanes, J.R., Masland, R.H., and Seung, H.S.
(2014). A genetic and computational approach to structurally classify neuronal types. Nat Commun
5, 3512.
Sunkin, S.M., Ng, L., Lau, C., Dolbeare, T., Gilbert, T.L., Thompson, C.L., Hawrylycz, M., and Dang,
C. (2013). Allen Brain Atlas: an integrated spatio-temporal portal for exploring the central nervous
system. Nucleic acids research 41, D996–D1008.
22
this preprint is the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for; http://dx.doi.org/10.1101/006346doi: bioRxiv preprint first posted online August 9, 2014;
Susaki, E.A., Tainaka, K., Perrin, D., Kishino, F., Tawara, T., Watanabe, T.M., Yokoyama, C., Onoe,
H., Eguchi, M., Yamaguchi, S., et al. (2014). Whole-brain imaging with single-cell resolution using
chemical cocktails and computational analysis. Cell 157, 726–739.
Suzek, B.E., Huang, H., McGarvey, P., Mazumder, R., and Wu, C.H. (2007). UniRef: comprehensive
and non-redundant UniProt reference clusters. Bioinformatics 23, 1282–1288.
Tanaka, N.K., Endo, K., and Ito, K. (2012). Organization of antennal lobe-associated neurons in adult
Drosophila melanogaster brain. Journal of Comparative Neurology 520, 4067–4130.
Tanaka, N.K., Tanimoto, H., and Ito, K. (2008). Neuronal assemblies of the Drosophila mushroom
body. Journal of Comparative Neurology 508, 711–755.
Ward Jr, J.H. (1963). Hierarchical grouping to optimize an objective function. Journal of the American
statistical association 58, 236–244.
Wong, A.M., Wang, J.W., and Axel, R. (2002). Spatial representation of the glomerular map in the
Drosophila protocerebrum. Cell 109, 229–41.
Yorozu, S., Wong, A., Fischer, B.J., Dankert, H., Kernan, M.J., Kamikouchi, A., Ito, K., and Anderson,
D.J. (2009). Distinct sensory representations of wind and near-field sound in the Drosophila brain.
Nature 458, 201–205.
Yu, H.H., Kao, C.F., He, Y., Ding, P., Kao, J.C., and Lee, T. (2010a). A complete developmental
sequence of a Drosophila neuronal lineage as revealed by twin-spot MARCM. PLoS Biol 8.
Yu, J.Y., Kanai, M.I., Demir, E., Jefferis, G.S.X.E., and Dickson, B.J. (2010b). Cellular organization
of the neural circuit that drives Drosophila courtship behavior. Curr Biol 20, 1602–14.
Zhu, S., Chiang, A.S., and Lee, T. (2003). Development of the Drosophila mushroom bodies: elabora-
tion, remodeling and spatial organization of dendrites in the calyx. Development 130, 2603–2610.
Zingg, B., Hintiryan, H., Gou, L., Song, M.Y., Bay, M., Bienkowski, M.S., Foster, N.N., Yamashita,
S., Bowman, I., Toga, A.W., et al. (2014). Neural networks of the mouse neocortex. Cell 156,
1096–1111.
23
this preprint is the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for; http://dx.doi.org/10.1101/006346doi: bioRxiv preprint first posted online August 9, 2014;
Figure 1: Image preprocessing, registra-
tion and similarity score (NBLAST) algo-
rithm
(A) Flowchart describing the image prepro-
cessing and registration procedure. FlyCir-
cuit images were split into 3 channels. The
Dlg-stained brain (discs large 1) (channel 1)
images were registered against the FCWB
template. Registration success was assessed
by comparing the template brain with the
brain images after applying the registration
(reformatted). The neuron image in channel
3 was skeletonized and reformatted onto the
template brain. The neuron skeleton was con-
verted into points and vectors. (B) After im-
age registration, neurons on the right side
of the brain were flipped onto the left side.
For most neurons this was done automati-
cally. On the left, brain plot showing 50 ran-
dom neurons before and after flipping. On
the right, cases for which the neuron flipping
was assessed manually. These included cases
in which the cell body was on or very close
to the midline, with or without small pri-
mary ipsilateral neurites. (C) NBLAST algo-
rithm. The similarity of two neurons (query
and target), is given by a function of the dis-
tance and absolute dot product between the
nearest neighbor points of the query/target
pair. This distance function reflects the prob-
ability of a match between a pair of points
(p
match
), relative to any two random points
(p
rand
). (D) Diagram illustrating how nearest
neighbor points are calculated. For a query
(N1)/target (N2) pair, each point of N1 (u
i
)
is paired to the N2 point (v
i
) that minimizes
the distance (d
i
) between the points. (E) Cal-
culating the distance function. Two groups
of neurons were used to calculate the dis-
tribution probabilities of matching and non-
matching pairs. The first corresponds to a
known class of uniglomerular olfactory pro-
jection neurons (uPNs), DL2 uPNs that had
been previously identified in the dataset. The
second group corresponds to all remaining
neurons. Random pairs of neurons were com-
pared within each group. (F) Brain plot show-
ing all DL2 neurons in the dataset that were
used for this analysis. (G) Calculation of the
distribution for matching and non-matching
pairs of segments. For all segment pairs of
all neuron pairs of each group, the distance
and absolute dot product were plotted in a
distance histogram. The distribution proba-
bility for matching (p
match
) or non-matching
pairs (p
rand
) was calculated by normalizing
the distance histogram to 1. When calculat-
ing the distance function, 1 × 10
-6
was added
to both p
match
and p
rand
to avoid a 0 denomina-
tor. (H) Plot showing that the similarity score
depends on the spatial location of the points
(distance between points) and the direction of
the vectors (absolute dot product). The score
is the highest for a distance of 0 μm and an
absolute dot product of 1.
this preprint is the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for; http://dx.doi.org/10.1101/006346doi: bioRxiv preprint first posted online August 9, 2014;
Figure 2: NBLAST allows different types
of searches
(A) Searching for similar neurons with
NBLAST. Pair comparisons between the
chosen query neuron and remaining neurons
in the dataset return a similarity score, allow-
ing the results to be ordered by similarity. (B)
NBLAST search for matching neurons using
a whole neuron as a query against the FlyCir-
cuit dataset (16,129 neurons). The query neu-
ron, top hit and top 10 hits in anterior view are
shown. A forward search is shown (returns
the similarity score for the query compared
to the target) but a reverse search (returns the
score of the target against the query) could
also be used. Query neuron: fru-M-400121.
(C) NBLAST search for matching neurons
using a neuron fragment as a query against
the FlyCircuit dataset (16,129 neurons). The
query neuron and top 10 hits are shown. (C’)
Search with the mALT tract from an olfac-
tory projection neuron (Cha-F-000239). Lat-
eral oblique view is shown. The mALT tract
of the top 10 hits is shown as an inset. (C”)
Search with the lALT tract from an olfactory
projection neuron (Gad1-F-200095). Ante-
rior view is shown. The lALT tract of the top
10 hits is shown as an inset. (D) NBLAST
search for neurons with matching neurites to
a fragment from a fruitless neuroblast clone
(pMP-e). The target is the FlyCircuit dataset
(16,129 neurons). (D’) Volume rendering of
pMP-e clone with the selected fragment – the
characteristic stalk of P1 neurons. An ante-
rior and lateral view are shown. (D”) Query
fragment in lateral view. Top 10 hits in an-
terior and lateral view. (E) NBLAST search
for neurons with matching neurites to a frag-
ment traced from a GAL4 image (R18C12)
((Jenett et al., 2012)). The target is the Fly-
Circuit dataset (16,129 neurons). (E’) Max-
imum Z projection of FlyLight line R18C12
registered to JFRC2. The image was down-
loaded from the Virtual Fly Brain website.
(E”) The fragment used as query (traced in
Vaa3D) in anterior view. Top 3 hits in an-
terior and dorsal view. (F) NBLAST search
for GAL4 traces (Peng et al., 2014) match-
ing a selected FlyCircuit neuron. Query neu-
ron: VGlut-F-500818. The top 10 trace hits
are shown in anterior and dorsal view. Max-
imum Z projection of FlyLight line R18C12
registered to JFRC2. The image was down-
loaded from the Virtual Fly Brain website.
this preprint is the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for; http://dx.doi.org/10.1101/006346doi: bioRxiv preprint first posted online August 9, 2014;
Raw score
5551
7987
D
6438
-167916
Q1
Q2
T1
T1
A
Same raw images
,
different segmentation
Top hit
Query
Top hit
Top 8 hits
Top hit
Query
fru-M-300198 Score > 0
I II III IV
VI
VII
VIII
V
F
Query
C
Query
Top hit
2nd
3rd
hits
All hits
Score > 6500
Query
Query
Mean score
Q1
Q2
T1
T1
S
MeanQ,T
=(S
NormQ,T
+ S
NormT,Q
)/2
0.136
0.579
Normalized score to neuron siz
e
0.56
-0.42
Q1
Q2
T1
T1
0.69
0.60
S
NormQ,T
= S
RawQ,T
/S
selfQ
-1 > S
NormQ,T
> 1
Forward S
Q,T
Reverse S
T,Q
Mean
99%
A' A'' B
E
97%
C'
All hits
IIIIIIVVVIVIIVIII
Figure 3: NBLAST scores are accurate and
meaningful
(A) NBLAST search with fru-M-300198 as
query (black). Neuron plot of the query neu-
ron. (A’) Neuron plot of the query (black)
and top hit (red). The top hit corresponds to
a different segmentation of the query neuron,
from the same raw image. The differences be-
tween these two images is due to minor dif-
ferences during the segmentation. (A”) Neu-
ron plot showing the top 8 hits. There are
differences in neurite branching, length and
position. (B) All hits with a forward score
over 0, colored by score, as shown in C.
(C) Histogram of scores for a forward search
with fru-M-300198 as query. Only hits with
scores over −5,000 are shown. The left inset
shows the histogram of scores for all search
hits. The right inset shows a zoomed view of
the top hits (score > 6,500). For more exam-
ples see S1. (C’) Neuron plots correspond-
ing to the score bins in C. (D) Comparison
of the raw, normalized and mean score, for
two pairs of neurons: one of unequal (Q1,
T1) and one of similar size (Q2, T1). The
value of the raw score depends on the size
of the neuron, whereas the normalized score
corrects for it, by dividing the raw score by
the query self-score (maximum score). Nor-
malized scores are between −1 and 1. Mean
scores are the average between the normal-
ized forward and reverse score for a pair of
neurons. These scores can be compared for
different searches. (E) Histogram of the nor-
malized score for the top hit for each neu-
ron in the whole dataset. The mean and 99th
percentile are shown as a dashed red and
green lines, respectively. (F) Plot of reverse
and forward normalized scores for 72 pairs
of neurons for which both the forward and
reverse scores are higher than 0.8. These
pairs were classified into four categories, ac-
cording to the relationship between the two
images: images correspond to a segmented
image that is duplicated (‘Same segmenta-
tion’); images correspond to different neu-
ron segmentations from the same raw im-
age (‘Same raw image’); images correspond
to two different segmented images from the
same brain (‘Same specimen’); images cor-
respond to segmented images of the same or
similar neurons in different brains (‘Differ-
ent specimen’). The inset plot shows the nor-
malized reverse and forward scores for all top
hits. The threshold of 0.8 is indicated by two
black lines.
this preprint is the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for; http://dx.doi.org/10.1101/006346doi: bioRxiv preprint first posted online August 9, 2014;
α'/β' neurons
A
C'
C
B
B''
D''
B'''
D
γ neurons
D'
B'
α/β neurons
Figure 4: NBLAST search and classifica-
tion of hits reveals Kenyon cell subtypes
(A) Hierarchical clustering (HC) of Kenyon
cells (n=1664), divided into two groups. Bars
below the dendrogram indicate the neurons
corresponding to a specific neuron type: γ
(in green), α′/β′ (in blue) and α/β neurons (in
magenta), h=8.9. Inset shows the mushroom
body neuropil. (B) Neuron plot of the γ neu-
rons. (B’) HC of the γ neurons divided into
three groups (I–III), h=3. Inset on the den-
drogram shows the γ neurons (same as in
B). Neuron plots of groups I to III. A lateral
oblique and a posterior view of the neurons
are shown. There are differences between the
3 groups in the calyx in the medial/lateral
axis and in the dorsal/ventral axis in the γ
lobe: the more medial group 1 is the most
dorsal in the γ lobe. (B”) HC of the clas-
sic γ neurons, corresponding to groups I and
III in B’, divided into four groups (A–D).
Neuron plots of groups A–D, A–B and C–D.
There are differences between the 4 groups
in the calyx in the medial/lateral axis and in
the dorsal/ventral axis in the γ lobe. (B”’)
HC of the atypical γ neurons corresponding
to group II in B’, divided into three groups
(a–c). Neuron plots of groups a–c, a, and b–
c. Group a corresponds to subtype γd neu-
rons which innervate the dorsal most region
of the gamma lobe and extend dendrites lat-
erally. (C) Neuron plot of the α′/β′ neurons.
(C’) HC of the α′/β′ neurons, divided into
four groups (i–iv), h=1.43. The groups i and
iv take a more anterior route in the pedun-
cle and β′ lobe than groups ii and iii. Dor-
solateral view is shown. (D) Neuron plot of
the α/β neurons. (D’) HC of the α/β neurons,
divided into four groups (1–4), h=3.64. In-
set on the dendrogram shows the α/β neurons
(same as in D). Neuron plots of groups A to
D. Lateral oblique, posterior view and poste-
rior view of a peduncle slice of these groups
are shown. There are differences between the
4 groups in the calyx and in the medial/lateral
axis, with each group corresponding to the in-
dicated neuroblast clone (AM, AL, PM, PL).
(D”) HC of groups 1 and 2. Lateral oblique,
posterior oblique and a dorsal view of a pe-
duncle slice views are shown. HC of group
1 divided into 2 subgroups. This separated
the neurons into peripheral (cyan) and core
(red) in the α lobe. Peripheral neurons occu-
pied a more lateral calyx position and were
dorsal to core neurons in the peduncle and
β lobe. Similar analysis to groups 3 and 4 is
shown in Figure S3A. HC of group 2 divided
into 3 subgroups. The red and blue subgroups
match the core and peripheral neurons, re-
spectively; the green subgroup the α/β poste-
rior subtype (α/βp). These neurons innervate
the accessory calyx and their axons terminate
before reaching the most medial region of the
β lobe. AcCa: accessory calyx. Neurons in
grey: Kenyon cell exemplars.
this preprint is the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for; http://dx.doi.org/10.1101/006346doi: bioRxiv preprint first posted online August 9, 2014;
Query
A B
C
NBLAST search
Collect 3 t op hits
Classify
uPN
uPN class X
For each uPN
X Y Z
Compare to
query class X
Groups 1-5 Group 1(DM6) Group 2 (V M5v) Group 3 (V M5d) Group 4 (VM2) Group 5 (DM1)
VA7m
VC2
DP1l
DM1
VM3
VM2
DM5
DM6
VM5v
VM5d
VM7
VC4
VC3l
VM5v
VC3l
VL2p
VL2a
v
DA1
v
VA1lm
VP4
VM4
V
VM6
VM1
VC3m
VP1
VA4
VA1lm
VA1d
DC3
VA1d
DL3
DA1
DL1
DA2
DA4
DL4
DC2
C'
*
Matches to
query class
%
n
Top hit
98.3 184
Any one
98.9 185
Any two
96.2 180
All three
95.2 178
187
Figure 5: NBLAST search and classification of hits reveals uniglomerular olfactory projection neuronal types
(A) Plot of the reverse and forward normalized scores for the top hit in an NBLAST search using the uniglomerular olfactory projection neurons (uPNs) as queries. Only uPN types for which we have more than one
example and unique query/target pairs are included in this analysis (n=327). For each query neuron, we identified cases for which both the top hit and query were of the same class (True) (n=319); the top hit is a uPN but
does not match the class of the query (False) (n=4), or the top hit is not a uPN (Not uPN) (n=4). (B) Percentage of neuron type matches in the top hit and top 3 hits for each uPN. The top three hits for each uPN (mean
score) were collected and the neuron type of each hit and query was compared. Only non-DL2 uPNs for which we had more than three neurons examples or a type were used (n=187). (C) Hierarchical clustering of uPNs
(non-DL2s) (n=214) divided into 35 groups (1–35), h=0.725. Dendrogram showing the glomerulus for each neuron. The neuron plot inset shows the uPNs colored by dendrogram group. Below the leaves, the number of
neurons that innervate each glomerulus is indicated by the black rectangles. Neurons that innervate DA1 and VA1lm glomeruli but originate from the ventral lineage instead of the lateral or anterodorsal, respectively, are
indicated as vVA1lm and vDA1. The dendrogram groups correspond to single and unique neuron types except for DL1 and DA1 neurons which are split into 2 groups (12–13, 15–16, respectively) (red arrowhead) and the
outlier neuron VM5v in group 9 (red asterisk). (C’) Neuron plots corresponding to dendrogram groups 1–5 and to each of these individual groups, colored by dendrogram group. The antennal lobe is in green, the lateral
horn in purple.
this preprint is the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for; http://dx.doi.org/10.1101/006346doi: bioRxiv preprint first posted online August 9, 2014;
C
A
B
A'
Visual projection neurons
A''
Figure 6: NBLAST search and classifi-
cation of hits uncovers visual projection
neuronal types
(A) Clustering analysis of unilateral
(uVPNs) and bilateral visual projection
neurons (bilVPNs), defined as neurons
with segments that overlap one or two
optic lobes, respectively, and some central
brain neuropil. (A’) Hierarchical clustering
(HC) of unilateral visual projection neurons
(uVPNs), divided into 21 groups (I–XXI),
h=3.65. Inset on the dendrogram shows the
neuropils considered for the overlap. To the
right, neuron plots of groups I to III. The
neuropils that contain the most overlap are
shown. Other neuron plots are shown in
Figure S4. (A”) HC of bilVPNs, divided
into 8 groups (i–viii), h=1.22. Inset on the
dendrogram shows the neuropils considered
for the overlap. To the right and below,
neuron plots of dendrogram groups. Group
ii corresponds to the LC14 neuron type that
connects the 2 lobulas, with one outlier
terminating in the medulla. The neuropils
that contain the most overlap are shown.
(B) Reclustering of uVPN groups I, II and
III from A’. These neurons have dorsal
cell bodies, arborize in the lobula (LO) and
project to anterior optic tubercle (AOTU)
via the anterior optic tract or to the posterior
ventrolateral protocerebrum (PVLP). The
neuron segments that co-localize with
either the AOTU or PVLP were isolated,
followed by HC of the neurons based
on the NBLAST score of these neuron
segments. The dendrogram was divided
into seven groups (1–7), h=1.69. Neuron
plots corresponding to the dendrogram
groups. An anterior and a lateral view are
shown. Some of dendrogram groups were
matched to known uVPN types. Group 1
corresponds to LC6 neurons, group 2 to
LC9. These 2 groups innervate the PVLP,
and show some differences in the lobula
lamination in the anterior/posterior axis.
Groups 3 and 4 seem to correspond to two
new subtypes of LC10B, that innervate the
dorsal AOTU. They show a clear distinction
in AOTU and lobula lamination, with group
4 being the dorsalmost in the AOTU and
the most anterior in the lobula. Groups
5 and 7 are possible new subclasses of
LC10, that innervate the ventral AOTU.
They show a clear distinction in AOTU
and lobula lamination, with group 7 being
the dorsalmost in the AOTU (but ventral
to group 3) and the most anterior in the
lobula (but posterior to group 3). Group
6 corresponds to LC10A neurons that
project through the ventral AOTU and turn
sharply dorsally in the middle region. (C)
Overlay of Z projections of registered image
stacks of example neurons from the types
identified in B on a partial Z projection of
the template brain (a different one for each
panel). The white rectangle on the inset
shows the location of the zoomed in area.
LC: lobula columnar neuron.
this preprint is the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for; http://dx.doi.org/10.1101/006346doi: bioRxiv preprint first posted online August 9, 2014;
Figure 7: NBLAST search and classifica-
tion of hits reveals subtypes of fruitless-
expressing mAL and P1 neurons
(A) Analysis of the mAL neurons. Hierar-
chical clustering (HC) of the hits, divided
into 2 groups (h=1.25). The mAL neuron
used as the NBLAST query, fru-M-500159,
is shown in the inset. Hits with a normal-
ized score over 0.2 were collected. The leaf
labels indicate the gender of the neuron:
’F’ for female and ’M’ for male. (B) Neu-
ron plot of the 2 dendrogram groups corre-
sponding to male (in cyan) and female (in
magenta) mAL neurons. (C) Analysis of the
male mAL neurons. The neuron segments
corresponding to the terminal arbors (ipsi-
and contralateral) were isolated and the
neurons were clustered based on the score
of these segments. HC of neurons, divided
into 3 groups (groups I–III) (h=0.83), that
reflect differences in the length of the ven-
tral ipsilateral branch (arrowhead). Group I
can be further subdivided into two differ-
ent subtypes, which differ in the shape and
extent of their dorsal contralateral arbori-
sation (arrowhead). (D) Analysis of the P1
neurons. Neuron plot of a P1 neuron, fru-M-
400046. The male enlarged region (MER) is
shown in red. Anterior and posterior views
are shown. Volume rendering of the pMP-e
fruitless neuroblast clone, which gives rise
to P1 neurons. The distinctive primary neu-
rite was traced and used on a NBLAST
search for matching neurons. (D’) HC of
hits for a search against the P1 primary neu-
rite divided into 10 groups (1–10) (h=0.92,
indicated by dashed line). This group of
neurons corresponds to a subset of neurons
obtained after a first HC analysis. Hits with
a normalized score over 0.25 were collected
and further selected. The inset shows a neu-
ron plot with groups 1–10. The leaf labels
show the GAL4 driver used to obtain that
neuron; the colors follow the gender: cyan
for male and magenta for female. Below the
dendrogram, neuron plots of each group.
The MER is shown in grey for groups 9 and
10.
this preprint is the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for; http://dx.doi.org/10.1101/006346doi: bioRxiv preprint first posted online August 9, 2014;
Affinity
propagation
Clustering
by score
16129 neurons 1052 c
lusters
10 neurons/cluster
S
av
=0.559
A B
E
ExemplarsAll neurons
AMMC-IVLP PN1
Exemplar
C
I II III IV V
VI VII VIII IX X
XI XII XIII XIV
Central complex
P1 neurons, AOTU
SMP, SIP
KCs: γ,α'/β' KCs: α/β
SAD, WED
uPNs
AL LNs, LAL
SLP, LH
VLP neurons
octopaminergic
B,C
LC10B
All x All
NBLAST scores
D
A
B
C
All exemplars
D'
Central brain exemplars
LC4
n=3
n=121
n=11
n=98
n=11
n=82
wedge
AMMC
AOTU
PVLP
PLP
LO
LO
Figure 8: Organizing NBLAST scores by
affinity propagation clustering
(A) Clustering by affinity propagation.
This method uses the all-by-all matrix of
NBLAST scores for the 16,129 neurons.
This method defined exemplars, which are
representative members of each cluster. An
affinity propagation clustering of the dataset
generated 1,052 clusters, with an average of
10 neurons per cluster and a similarity score
of 0.559. (B) Plot showing the mean cluster
score versus cluster size. (C) Hierarchical
clustering (HC) of the 1,052 exemplars,
dividing them into three groups (A–C).
Group A corresponds mostly to optic lobe
and VPN neurons; groups B and C to central
brain neurons. The insets on the dendrogram
show the neurons of these groups. The main
neuron types or innervated neuropils are
noted. (D) HC of central brain exemplars
(groups B and C, inset on dendrogram),
divided into 14 groups, h=2.7. (D’) Neurons
corresponding to the dendrogram groups
in D. (E) Affinity propagation clusters of
defined neuron types. Neuron plot of exem-
plars (top row) or all neurons (bottom row)
for auditory AMMC-IVLP PN1 neurons
(compare with Figure S5D) and VPN types
LC10B (compare with Figure 6B) and LC4
(compare with Figure S4B). The number
of exemplars and neurons is indicated on
the top left corner for each example. The
AMMC is shown in green, the wedge in
magenta. AMMC: antennal mechanosensory
and motor center; AOTU: anterior optic
tubercle; LO: lobula; PVLP: posterior
ventrolateral protocerebrum; PLP: posterior
lateral protocerebrum.
this preprint is the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for; http://dx.doi.org/10.1101/006346doi: bioRxiv preprint first posted online August 9, 2014;
Supplemental Information
Supplemental Results
Algorithm Design Process
The algorithm design process was primarily motivated by a requirement for rapid and sensitive searches
of neuron databases. It was necessary to consider both the data structure and the similarity algorithm
jointly in the face of the design requirements.
Our first application would be data acquired in Drosophila, where previous studies using image
registration have shown a high degree of spatial stereotypy (standard deviation of landmarks ~2.5 µm
in each axis for a brain of 600 µm in its longest axis, Jefferis et al., 2007). Therefore one key design
decision was to use co-registered data rather than calculating similarity using features of the neuron
that are independent of absolute spatial location.
On the algorithm side, the key initial design decision was whether to develop a direct pairwise
comparison algorithm or to use a form of dimensional reduction to map neuronal structure into a lower-
dimensional space. The major advantage of the latter approach is that the similarity between neurons
can be computed directly and almost instantaneously in the low dimensional space. However, the con-
struction of a suitable embedding function either requires existing knowledge of neuronal similarity
(likely supplied by experts in the form of large amounts of training data), huge amounts of unlabeled
data that enable direct learning of features (e.g. Le et al., 2012), or a strategy based on successful
extraction of key image features.
A number of considerations made us favor the approach of direct pairwise comparison. First, we
suspected that it would be possible to make a more sensitive algorithm by working with the original
data. Second, the amounts of image data available did not seem large enough to avoid a requirement
for extensive labeled training data. Third, we reasoned that our own intuitions about neuronal similar-
ity could be better expressed in the original physical space of the neuron than in a low dimensional
embedding. Our own exploratory analysis in which we summarized each neuron in different ways as
feature vectors of the same dimension and used a comparison function in the feature space (SP, GSXEJ
unpublished observations) confirmed that constructing a sensitive metric of this sort is challenging.
The selection of a pairwise similarity metric meant that we had to give particularly careful consid-
eration to performance issues in the design phase. We set two practical performance targets: 1) being
able to carry out searches of a single neuron against a database of 10,000 neurons in less than a minute
on a simple desktop or laptop computer. 2) Being able to complete all-against-all searches for 10,000
neurons (10
8
comparisons) in < 1 day on a powerful desktop computer. These targets meant that each
elementary comparison operation should take around 5 ms or less. Image pre-processing carried out
once per neuron would therefore be a good investment if it reduced the time taken for each pairwise
comparison. These considerations prompted us to generate a spatially registered, compact representa-
tion of each neuron as a separate pre-processing step for each neuron, rather than develop an algorithm
that simultaneously solved both the spatial alignment and similarity problem.
Algorithm Scoring
As described in the main results section, we defined NBLAST raw scores as the sum of segment pair
scores:
33
this preprint is the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for; http://dx.doi.org/10.1101/006346doi: bioRxiv preprint first posted online August 9, 2014;
(3)
We initially experimented with a function based on expert intuition:
(4)
This includes a negative exponential function of distance (related to the normal distribution), with a
free parameter based on our previous estimates of the variability in position within the fly brain of
landmarks after registration (Jefferis et al., 2007; Yu et al., 2010b) set to 3 µm. Although this provided
a useful starting point, we were unhappy with a scoring system that required parameters to be specified
rather than derived empirically from data. This then motivated us to investigate the statistical scoring
approach described in based on log probability ratios. This single parameter approach may still be
helpful in some situations where insufficient neurons are available in order to define a statistical scor-
ing matrix (we used 150 similar neurons and 5000 random pairs). It can also enable a bootstrapping
approach for new datasets in order to help idea similar pairs of neurons that can then be used to define
a full scoring matrix.
Kenyon cell analysis
Dataset collection and dividing into types We collected the dataset of initial KCs by performing
a forward and reverse search against all neurons using one identified KC (fru-M-500225). We then
selected neurons that had both raw scores above −2,500 (2,088 neurons). We performed affinity prop-
agation clustering (Frey and Dueck, 2007) of these neurons, obtaining 59 clusters, and manually veri-
fied each one, resulting in 1,562 neurons being identified as KCs. An additional search for high scorers
against these KC exemplars uncovered an extra 102 neurons, bringing the total number of KCs used
in our analysis to 1,664, representing 10.3 % of the FlyCircuit dataset.
We performed hierarchical clustering of the KCs, based on the NBLAST scores, and divided the
dendrogram into two groups (Figure 4A). Contrary to expectations, one group contained both the γ and
α′/β′ neurons, whereas the other group consisted exclusively of α/β neurons (Figure 4B–D), the largest
subset in our sample. We separated α′/β′ from the γ neurons in a subsequent hierarchical clustering of
this group. We performed additional analysis for each of the neuron types.
Analysis of γ neurons Hierarchical clustering of the 470 γ neurons resulted in a dendrogram which
we divided into three groups (I–III) (Figure 4B’). The number of clusters was chosen by visual inspec-
tion in order to reveal differences in morphology and organization between the groups. Groups I and
III corresponded to the classical γ neurons while group II matched atypical γ neurons. There were dif-
ferences in neurite positioning in the calyx, from medial to lateral, with group I being the most medial,
followed by groups II and III. There were also differences in the gamma lobe, with group II occupying
the anterodorsal region, while groups I and III were mostly mixed in the rest of the lobe. A subsequent
clustering analysis of the classical γ neurons divided into 4 groups (groups A-D) revealed that there
were differences between the groups in their medial to lateral position in the calyx. These differences
correlated to a certain degree with differences in the dorsal/ventral position of the projections in the
34
this preprint is the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for; http://dx.doi.org/10.1101/006346doi: bioRxiv preprint first posted online August 9, 2014;
γ lobe, with groups C and D, the most medial, being also the most dorsal (Figure 4B”). These obser-
vations suggest that the relative position of the projections of classical γ neurons is maintained at the
calyx and γ lobe.
In order to understand if the relative position of the classical γ neurites was maintained in between
the calyx and the γ lobe, we clustered the neurons based on the scores of the segments in the peduncle.
We took neuron skeletons from classical γ neurons and isolated the axon arbors that co-localized with
the peduncle volume (Figure S3B’). We then carried out a new clustering based on all-by-all NBLAST
scores of these partial skeletons, cutting the dendrogram at a level defined by visual inspection (4
groups). The overall organization almost fully recapitulated the positioning of the neurites in the whole
neuron analysis (compare Figure 4B’–B” with Figure S3B’). A clear and expected lamination was
found in the peduncle, with neurites occupying the most outer stratum. Differences in the medial to
lateral positioning of neurites in the calyx followed the previously observed organization, with the most
medial groups occupying the dorsal region of the gamma lobe. The overall organization almost fully
recapitulated the positioning of the neurites in the whole neuron analysis (for more information see
Figure S3 and Supplemental Results). Thus, the stereotypical organization of the classical γ neurons
is maintained throughout the neuropil.
Group II of the γ neurons matched atypical γ neurons (γ dorsal neurons) (Aso et al., 2014) with
neurons that extended neurites posteriolaterally in the calyx and projected to the most dorsal region
of the γ lobe (Figure 4B”’). Hierarchical clustering of these neurons resulted in a dendrogram that we
divided into 3 groups (a–c). This number of groups isolated the previously identified subtype of atypical
γ neurons – γd neurons (Aso et al., 2009) – into one group (group a). These neurons extend neurites
ventrolaterally at the level of the calyx (identified as ventral accessory calyx) (Aso et al., 2014). The
other 2 groups (b, c) correspond to uncharacterized types. Although they project to a similar region in
the γ lobe, their dendrites do not extend laterally and their calyx neurites are longer than γd neurons.
Analysis of α′/β′ neurons Hierarchical clustering analysis of the α′/β′ neurons and separation into
4 groups highlighted the characterized subtypes of α′/β′ neurons (Figure 4C–C’). They differ in their
anterior/posterior position in the peduncle and β′ lobe with three types described - α′/β′ anterior and
posterior (α′/β′ap) and α′/β′ medial (α′/β′m) (Tanaka et al., 2008, Y. Aso, personal communication)).
Although we were unable to unambiguously assign a α′/β′ subtype to each group i–iv because our
sample was too small, there were clear trends. Neurites of neurons in groups i and iv were more anterior
than the other 2 groups (ii, iii) in both the peduncle and β′ lobe. These relative positions were not
maintained in the calyx, with the two the anterior groups (i, iv) occupying either a medial or a lateral
position.
Analysis of α/β neurons The largest subset of KCs corresponds to α/β neurons (Figure 4D). During
the analysis of this group we found 18 neurons that did not correspond to α/β cells, since they innervated
either only the β or the α lobes, and they were removed from the analysis. We performed hierarchical
clustering on the remaining 1,091 cells and divided the resulting dendrogram into four groups (1–4)
(Figure 4D’), which matched the four neuroblast lineages from which they originate (Zhu et al., 2003).
The relative position of the neurites of the four groups within the calyx is somewhat maintained in the
peduncle, with the lateral neuroblast clones (group 1 AL; group 2 PL) extending along the dorsolateral
peduncle, while the medial clones (group 3 AM; group 4 PM) occupy a more ventromedial region.
35
this preprint is the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for; http://dx.doi.org/10.1101/006346doi: bioRxiv preprint first posted online August 9, 2014;
Hierarchical clustering of each group revealed an expected common organization for all neuroblast
clones (Figure 4D” and Figure S3A). For all groups there was a clear distinction between the late
born core (α/β core, α/β-c) and early born peripheral neurons (α/β surface, α/β-s). Core neurons are on
the inside stratum of the α lobe. They are also reported to occupy the inside stratum of the peduncle
and β lobe (Tanaka et al., 2008). We were unable to observe this, although the projections of α/β core
neurons were ventral to α/β surface neurons in both the peduncle and β lobe. There was also a trend
for α/β surface neurons to occupy a more medial position in the calyx in comparison to α/β core ones.
A subgroup of group 2 corresponded to the α/β posterior or pioneers neurons (α/βp). The α/βp neurons
are the earliest born α/β and they innervate the accessory calyx, run along the surface of the posterior
peduncle into the β lobe but stop before reaching the medial tip (Tanaka et al., 2008). A new clustering
based on peduncle position of the neuron segments did not recapitulate the relative positions of the
calyx neurites for each of the neuroblast clones observed in the whole neuron analysis suggesting that
the relative position of the α/β neurons in the peduncle does not completely reflect their stereotypical
organization in the calyx (for more information see Figure S3 and Supplemental Results).
In order to investigate the stereotypical organization of α/β neurites, we performed a similar analy-
sis as for the classic γ neurons, isolating the axon arbors that co-localized with the peduncle for groups
1 to 4 (Figure S3B”). The new clustering based on peduncle position of these partial neuron skeletons
did not recapitulate the relative positions of the calyx neurites for each of the neuroblast clones ob-
served in the whole neuron analysis (compare Figure 4D’ with Figure S3B”). In addition, there was no
clear organization of neurites in the α lobe that correlated with their position in the peduncle.
Olfactory projection neuron analysis
We started by manually classifying the 400 uPNs in the FlyCircuit dataset by glomerulus, neuroblast
lineage, and axon tract, using the original image stacks. The definition of the manual gold standard
annotations was an iterative process that took several days. The first round accuracy was about 95
%. Numerous discrepancies were revealed by subsequent NBLAST analysis and difficult cases were
resolved by discussion between two expert annotators before finalizing assignments. We excluded 3
neurons for which no conclusion could be reached. We found a very large number of DL2 uPNs, 145
DL2d and 37 DL2v neurons, in a total of 397 neurons. Nevertheless, our final set of uPNs broadly rep-
resents the total variability of described classes and contains neurons innervating 35 out of 56 different
glomeruli (Tanaka et al., 2012), examples of the three main lineage clones (adPNs, lPNs and vPNs)
in addition to one bilateral uPN, and neurons that follow each of the three main tracts (medial, medi-
olateral and lateral antennal lobe tracts). For subsequent analysis, we removed 3 neurons for which
registration failed.
Visual projection neuron analysis
We started with the 1,052 exemplars found by affinity propagation clustering of NBLAST scores (Fig-
ure 8). We then clustered those exemplars using hierarchical clustering and found that extrinsic and
intrinsic optic lobe neurons together formed a distinct “optic lobe” group within this (Figure 8C). We
then collected all neurons associated with those “optic lobe” exemplars and calculated the overlap of
neurons with each of the standard neuropils defined by Ito et al. (2014) (see Experimental Procedures
and Manton et al., 2014 for technical details). This then allowed us to separate neurons by innervation
36
this preprint is the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for; http://dx.doi.org/10.1101/006346doi: bioRxiv preprint first posted online August 9, 2014;
pattern into 3 groups: 1) ipsilateral optic lobe neuropils only (see Figure S6), 2) ipsilateral and central
brain neuropils (unilateral VPNs, uVPNs) or 3) both optic lobes and central brain neuropils (bilateral
VPNs). This selection procedure resulted in a set of 1,793 uVPNs and 72 bilateral VPNs.
Auditory neuron analysis
We employed a two-step search strategy. First, for 5 of the auditory types, we used the FlyCircuit
neuron named by Lai et al. (2012) as the seed neuron for the first search. Candidate neurons were
selected using strict anatomical criteria. A second search was then done using these candidates as
query neurons and collecting all high scorers (score over 0.5). These neurons corresponded to our set
for each of the types.
mAL neuron analysis
The set of maL neurons resulted from a search with a seed mAL neuron, fru-M-500159. We then
collected 41 hits with a mean NBLAST score greater than 0.2.
P1 neuron analysis
The set of P1 neurons was identified by searching the FlyCircuit dataset with a tracing of the distinc-
tive primary neurite of a pMP-e clone (Cachero et al., 2010). Hierarchical clustering of the top hits,
after manual verification, identified a subset consisting solely of P1 neurons which was used in the
subsequent analysis.
Online resources
The online resources provided in the paper are listed at http://jefferislab.org/si/nblast. In addition to
the open source software described in the Experimental Procedures we also provide:
• Code and instructions to generate some of the figure panels used in the paper. Instructions can
be found here. A video demo is also available here.
• The affinity propagation clustering of the flycircuit.tw dataset (as in Figure 8D), excluding in-
trinsic optic lobe neurons. This can be viewed online here, including interactive 3D rendering
of clusters powered by WebGL.
• The clusters identified by affinity propagation clustering (including intrinsic optic lobe neurons)
are all indexed by the Virtual Fly Brain website, which links to 3D WebGL renderings of each
cluster hosted at jefferislab.org. Clusters can be identified by VFB queries for the neuropil region
that they innervate. For example, search from the VFB homepage for “AMMC”. From the results
page, choose the query “Images of neurons with: some part here (clustered by shape)”. A list of
clusters, with thumbnail images is displayed; single exemplars are also displayed for each cluster,
hyperlinked to the original data at flycircuit.tw. The images of the individual neurons that are
part of this cluster can also be displayed in the stack browser from this page (“Show individual
members”). Clicking on a cluster thumbnail links to a page which includes a snapshot and 3D
rendering of the cluster, and information about the neurons that are part of this cluster, including
37
this preprint is the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for; http://dx.doi.org/10.1101/006346doi: bioRxiv preprint first posted online August 9, 2014;
links to the appropriate Neuron ID pages at flycircuit.tw; a second table provides links to result
pages for the most similar clusters. A video demo is available here
• A video demo showing how to search for FlyCircuit neurons similar to a GAL4 tracing using
the R packages detailed in Experimental Procedures.
• An online web-app allowing on-the-fly NBLAST queries of FlyCircuit neurons against other
FlyCircuit neurons, as well as queries of user-uploaded neurons against the FlyCircuit dataset,
available at http://jefferislab.org/si/nblast/online/.
38
this preprint is the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for; http://dx.doi.org/10.1101/006346doi: bioRxiv preprint first posted online August 9, 2014;
Figure S1: Neuron search with NBLAST
(A) NBLAST search with VGlut-F-000493 as query. Neuron plots of (from higher to lower score): the query (black) and top
hit (red), top 8 hits, hits with a score over 50,000 and hits with a score over 25,000. The top hit corresponds to a segmented
image that was duplicated. It perfectly overlays the query neuron. As the score decreases, so does the similarity of the hits
to the query. On the right, histogram of forward scores. Only hits with scores over −100,000 are shown. The score of the
query, top hit and top 8 hits are indicated. A dashed purple line marks 25,000. The left inset shows a zoomed view of the top
hits (score > 50,000) (dashed blue rectangle in main plot). The score of the query, top hit and top 8 hits are indicated. (B)
NBLAST search with Cha-F-600134 as query (black). Neuron plots of (from higher to lower score): the query and top hit,
top 8 hits, hits with a score over 5,000 and hits with a score over 0. The top hit corresponds to an image of a neuron from
the same brain but from a different raw image. It is very similar to the query neuron. As the score decreases, so does the
similarity of the hits to the query. On the right, histogram of forward scores. Only hits with scores over −8,000 are shown.
The score of the query and top hit are indicated. A dashed purple line marks 0. The left inset shows a zoomed view of the top
hits (score > 5,000) (dashed blue rectangle in main plot). The score of the query, top hit and second top hits are indicated.
this preprint is the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for; http://dx.doi.org/10.1101/006346doi: bioRxiv preprint first posted online August 9, 2014;
A
Query
fru-M-300198
Score > 0
Mean score
1
0.8
0.6
0.4
0.2
0
Query
Top hit
I II III IV V
VI VII VIII
A'' B
B'
A'
I
II
III
IV
V
VI
VII
VIII
Figure S2: NBLAST search using mean scores
(A) The query neuron fru-M-300198 (same as inFigure 3B). (A’) Neuron plot of the hits with a mean score over 0 for a search against the query. Hits are
colored by score bin (10), as in A”. (A”) Histogram of mean scores for hits against fru-M-300198 with a score over 0 divided into 10 bins (indicated in
the scale bar in B). (B) Hierarchical clustering of hits with a mean score over 0. The leaves of the dendrogram are colored by score (same as in A”), and
as shown in the scale bar. The dendrogram was divided into eight groups (I–VIII), with each one being assigned a color, shown on the colored rectangle
below the leaves. The query neuron is in group III, and the hits with the higher scores are in groups II and III. (B’) Neuron plots corresponding to the
dendrogram groups (I–VIII), following the colors assigned to each group. Groups II and III, corresponding to the highest scores, are the most similar
neurons to the query.
this preprint is the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for; http://dx.doi.org/10.1101/006346doi: bioRxiv preprint first posted online August 9, 2014;
Figure S3: NBLAST search and classification of hits reveals Kenyon cells subtypes
(A) Hierarchical clustering (HC) of the α/β neurons, divided into four groups (1–4) (h=3.64) (same as in Figure 4D’). The inset on the dendrogram
shows the α/β neurons. Groups 3 and 4 were clustered and divided into 2 groups each. This separates the neurons into peripheral (cyan) and core (red)
in the α lobe. Peripheral neurons occupied a more lateral calyx position and were dorsal to core neurons in the peduncle and β lobe. Similar analysis to
group 1 is shown in Figure 4D”. Lateral oblique, posterior oblique and a dorsal view of a peduncle slice (position indicated by dashed rectangle) are
shown. (B) Reclustering of Kenyon cells based on the co-localization of neurons segments in the peduncle. The neuron segments that co-localize with
the peduncle were isolated, followed by HC of the neurons based on the NBLAST score of the segments. (B’) HC of the neurons segments of classic γ
neurons, groups I and III (see Figure 4B”) divided into 4 groups, h=3.16. Neuron plot of the 4 groups. A posterior view of a slice of the peduncle shows
an expected clear organization. It correlates to the position of the neurons in the calyx, with more medial neurons (cyan and green) being dorsal and
ventral in the peduncle than more lateral neurons (red and purple). No clear structural organization is discernible in a lateral view of a slice of the γ lobe.
(B”) HC of the neurons segments of classic α/β neurons, groups 1 to 4 (see Figure 4D’) divided into 4 groups, h=4.16. Neuron plot of the 4 groups.
A posterior view of a peduncle slice shows an expected clear organization. It correlates to the position of the neurons in the calyx, with more medial
neurons (cyan and green) being ventrolateral in the peduncle than more lateral neurons (red and purple). No structural organization is discernible in a
dorsal view of a slice of the α lobe. For all neuron plots, the neurons in grey correspond to the Kenyon cell exemplars.
this preprint is the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for; http://dx.doi.org/10.1101/006346doi: bioRxiv preprint first posted online August 9, 2014;
Dendrogram group Neuron type Comments
1 LC6 Lobula innervation, dorsal cell bodies, axons follow
the anterior optic tract to the AOTU, turning
ventrally midway to innervate the lateral PVLP.
2 LC9 As group 1, but terminating in the medial PVLP
rather than extending laterally.
3 LC10B (possibly
different subtype to 4)
Dorsal AOTU innervation, ventral to group 4.
4 LC10B (possibly
different subtype to 3)
Dorsal AOTU innervation, dorsal to group 3.
5 New LC10 subclass The most ventral AOTU innervation. Similar to
group 7, but ventral to it in the AOTU.
6 LC10A Axons project through ventral AOTU, turn sharply
dorsally to terminate in the dorsal AOTU.
7 New LC10 subclass Ventral AOTU innervation, dorsal to group 5.
Table S1: Correspondences between hierarchical clustering groups of AOTU- and PVLP-innervating uVPNs via NBLAST scores and previously de-
termined neuron types (Otsuna and Ito, 2006).
Dendrogram group Neuron type Comments
A LC12 (possibly different
subtype to B)
Innervation in most lateral and anterior PVLP
glomerulus.
B LC12 (possibly different
subtype to A)
Innervation in more posterior and medial PVLP
glomerulus than group A.
C LC4 Innervates a more medial PVLP glomerulus than
LC12.
D LT12 Tentative match. Class was identified in Otsuna and
Ito (2006) based on a single neuron.
E LC11 Innervates more dorsal PVLP glomelurus than
LC12. Extends along the posterior PVLP, with a
sharp anterior turn. Terminates with a blunt-stick
like ending in the lateral PVLP.
F New LT subclass Similar to LT12, but with projections posterior to it,
terminating in the lateral region of the superior
posterior slope.
G Unmatched Do not correspond to a single type.
Table S2: Correspondences between hierarchical clustering groups of PVLP- and PLP-innervating uVPNs via NBLAST scores and previously deter-
mined neuron types (Otsuna and Ito, 2006).
this preprint is the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for; http://dx.doi.org/10.1101/006346doi: bioRxiv preprint first posted online August 9, 2014;
B
C
A
A'
Figure S4: NBLAST search and classi-
fication hits uncovers unilateral visual
projection neurons neuronal types
(A) Hierarchical clustering (HC) analy-
sis of unilateral visual projection neurons
(uVPNs), defined as neurons with segments
that overlap one optic lobe, and some cen-
tral brain neuropil. The dendrogram was di-
vided into 21 groups (I–XXI), h=3.65. In-
set on the dendrogram shows the neuropils
considered for the overlap. Below, neuron
plots of groups I to XXI. The neuropils that
contain the most overlap are shown. (B)
Reclustering of uVPN groups IV, VI, VII
and XI from A. These neurons arborize in
the lobula (LO) and project to the poste-
rior ventrolateral protocerebrum (PVLP) or
posterior lateral protocerebrum (PLP). The
neuron segments that co-localize with ei-
ther the PVLP or PLP were isolated, fol-
lowed by HC of the neurons based on the
NBLAST score of these neuron segments.
The dendrogram was divided into seven
groups (A–G), h=2.04. Neuron plots corre-
sponding to the dendrogram groups. An an-
terior, a lateral or lateral oblique views are
shown. Some of dendrogram groups were
matched to known uVPN types. Groups A
and B possibly correspond to two LC12
subtypes, that innervate the more lateral
PVLP glomeruli. Group B innervates a
more anterior and medial glomeruli than
group A (see also C). Group C corresponds
to LC4, that innervates a lateral PVLP
glomeruli and ventral to LC12. Group D
corresponds to LT12 neurons, that project
from the lateral to the medial PVLP, poste-
rior to LC4. Group E corresponds to LC11,
that innervates a lateral PVLP glomeruli,
dorsal to LC12. These neurons extends
along the posterior PVLP and make a sharp
anterior turn, terminating with a blunt-stick
like ending in the lateral PVLP. Group F
corresponds to a possibly new LT subclass,
with neurons projecting posteriorly to LT12
in the PVLP and extending into the supe-
rior posterior slope (SPS). (C) Overlay of
Z projections of registered image stacks of
example neurons from the types identified
in B on a partial Z projection of the tem-
plate brain (a different one for each panel).
The white rectangle on the inset shows the
location of the zoomed in area. LC: lob-
ula columnar neuron; LT: lobula tangential
neuron.
this preprint is the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for; http://dx.doi.org/10.1101/006346doi: bioRxiv preprint first posted online August 9, 2014;
A
B
C
D
E
Figure S5: NBLAST search and classifi-
cation hits reveals auditory neuron types
Searches were done using types identified
by Lai et al. (2012). A first search using
the neuron named by Lai for each type
(shown on the left panel with wedge in ma-
genta, AMMC in green or PVLP in pur-
ple) was done to collect the possible candi-
dates. A second search was then done us-
ing these neurons as queries and collect-
ing all the high scorers (over 0.5). The den-
drogram and neuron plots (anterior, poste-
rior and lateral views) showing either one or
all neurons from each clustering group are
shown on the middle and right panels. (A)
Hierarchical clustering (HC) of AMMC-
AMMC projection neuron 1 (PN1). The 34
top scorers were clustered and divided into
2 groups, h=0.72. Neuron plot of the query
neuron (black), and one neuron from each
group (red and cyan). To the right, neuron
plots of of each group. Differences in an-
terior length of the left projection are indi-
cated by an arrowhead. (B) HC of AMMC-
VLP PNs. The 34 top scorers were clus-
tered and divided into 2 groups, h=1.2. Neu-
ron plot of the query neuron (black), and
one neuron from each group (red and cyan).
On the left, neuron plots of each group. (C)
HC of AMMC-B1 PNs. Five hits of the 204
top scorers are shown on the left. The 204
top scorers were clustered and divided into
2 groups, h=3.46. Group I was matched to
an unidentified type of AMMC local neu-
rons (LN). It was clustered and divided
into 2 groups, h=0.77. Group II corresponds
to a mix of AMMC-B1 PNs and AMMC-
IVLP PN1. After selecting the AMMC-B1
PNs, the neurons were clustered and di-
vided into 3 groups, h=1.5. Neuron plot of
the query neuron (black), and one neuron
from each group (purple and green). Below,
neuron plots of the 3 groups. Arrows in-
dicate differences between groups. (D) HC
of AMMC-IVLP PN1. The 79 top scorers
were clustered and divided into 3 groups,
h=1.03. Neuron plot of the query neuron
(black), and one neuron from each group
(red, green and blue). Below, anterior, pos-
terior and lateral view neuron plots of the 3
groups. (E) HC of AMMC-IVLP PN2. Six
hits of the 170 top scorers are shown on the
left. The 170 top scorers were clustered and
divided into 6 groups, h=1.02. Neuron plot
of the query neuron (black), and one neu-
ron from each group (red, yellow, green,
cyan, blue, magenta). Below, anterior, pos-
terior and lateral view neuron plots of the
6 groups. Arrows and arrowheads indicate
differences between groups.
this preprint is the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for; http://dx.doi.org/10.1101/006346doi: bioRxiv preprint first posted online August 9, 2014;
Panel Neuron type Comments
A AMMC-AMMC PN1
(2 possible subtypes)
These neurons innervate both AMMCs, with a
ventral cell body. Group I extends more dorsally
than group II.
B AMMC-VLP
(2 possible subtypes)
These neurons innervate the ipsilateral AMMC and
the contralateral VLP. Group II extends more
laterally than group I.
C AMMC-B1 PN
(3 possible subtypes)
These neurons innervate the ipsilateral AMMC and
IVLP and the contralateral IVLP. Blue group
innervates more medial regions and has more
extensive innervation contralaterally; purple group
innervates more anterior and posterior regions
ipsilaterally, and the green group innervates the
dorsal regions in the ipsilateral AMMC.
C New AMMC-LN type
(2 subtypes)
A type of AMMC LN, with a dorsal cell body. Two
possible subtypes: the magenta group innervates
more dorsal regions than the orange group.
D AMMC-IVLP PN1
(3 possible subtypes)
These neurons innervate the ipsilateral AMMC and
the contralateral IVLP. Red group has a more
dorsomedial ipsilateral innervation more medial
regions, with some dorsal medial branches in the
contralateral hemisphere; some neurons extend a
long neurite ventrally ipsilaterally. The green group
innervates the more ventromedial regions
ipsilaterally. The blue group is similar to the green
one, although it does not extend as ventrally in the
ipsilateral side, and a few neurons extend a long
neurite ventrally (similar to red group).
E AMMC-IVLP PN2
(6 possible subtypes)
These neurons innervate the ipsilateral AMMC and
the contralateral IVLP, with a posterior cell body.
Group I innervates the more lateral regions in both
hemispheres. Group II has a long dorsal branch in
the contralateral hemisphere, at the lateral edge of
the neuron. Group III and IV are very similar, with
the latter innervating more dorsal regions in both
hemispheres. Group V corresponds to the strict
definition of the neuron type by Lai et al. (2012),
showing a short dorsal branch just medial to the
contralateral IVLP. Group VI are similar to group
V, with a few neurons showing a short dorsal
branch, and innervating a more ventral region in the
contralateral IVLP.
Table S3: Correspondence between hierarchical clustering of auditory neuron via NBLAST scores and previously determined neuron types (Lai et al.,
2012).
this preprint is the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for; http://dx.doi.org/10.1101/006346doi: bioRxiv preprint first posted online August 9, 2014;
Figure S6: NBLAST scores and classifica-
tion of hits highlights neuropil organiza-
tion of intrinsic optic lobe neurons
Hierarchical clustering of intrinsic optic
lobe neurons. This neuron set was defined
as any neuron that overlapped only one of
the optic lobes and with no arborization in
the central brain neuropils. Dendrogram of
the intrinsic optic lobe neurons, divided into
into 20 groups (I–XX) with the correspond-
ing heatmap calculated from the neuropil
overlap in the different neuropils: medulla
(ME), lobula (LO), lobula plate (LOP) and
accessory medulla (AME). The values were
log transformed. Neuron plots correspond-
ing to the dendrogram groups are shown be-
low. The neuropils for which the overlap is
more significant are plotted. Although some
organizational structure is seen, the den-
drogram groups to do not represent unique
types.
this preprint is the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for; http://dx.doi.org/10.1101/006346doi: bioRxiv preprint first posted online August 9, 2014;
