Content uploaded by Shin-ya Takemura
Author content
All content in this area was uploaded by Shin-ya Takemura on Feb 06, 2015
Content may be subject to copyright.
Review
Connectome of the fly visual circuitry
Shin-ya Takemura*
Janelia Research Campus, Howard Hughes Medical Institute, 19700 Helix Drive, Ashburn, VA 20147, USA
*To whom correspondence should be addressed. E-mail: takemuras@janelia.hhmi.org
Received 7 October 2014; Accepted 13 November 2014
Abstract
Recent powerful tools for reconstructing connectomes using electron microscopy (EM)
have made outstanding contributions to the field of neuroscience. As a prime example,
the detection of visual motion is a classic problem of neural computation, yet our under-
standing of the exact mechanism has been frustrated by our incomplete knowledge of the
relevant neurons and synapses. Recent connectomic studies have successfully identified
the concrete neuronal circuit in the fly’s visual system that computes the motion signals.
This identification was greatly aided by the comprehensiveness of the EM reconstruction.
Compared with light microscopy, which gives estimated connections from arbor overlap,
EM gives unequivocal connections with precise synaptic counts. This paper reviews the
recent study of connectomics in a brain of the fruit flyDrosophila and highlights how con-
nectomes can provide a foundation for understanding the mechanism of neuronal func-
tions by identifying the underlying neural circuits.
Key words: EM reconstruction, synaptic circuits, neural computation, motion detection, insect vision, medulla
Introduction
In neuroscience research, discoveries over the past few
decades in the fruit flyDrosophila have made great contri-
butions to our understanding of the mechanisms behind
nervous system function [1–3]. Recently, a unique combin-
ation of genetic, anatomical and physiological tools avail-
able in Drosophila [4] has accelerated research regarding
functions of neural circuits. For instance, mechanisms of
detecting visual motion are currently subject to intense
behavioral, physiological and anatomical investigations in
this species (reviewed in [5,6]).
It has been nearly 60 years since an influential model of
visual motion detection, the Hassenstein–Reichardt elemen-
tary motion detector (EMD) (Fig. 1a), was proposed by two
German scientists [7,8]. This computational model was ori-
ginally proposed based upon experiments of the optomotor
behavior in insects, and thereafter had also been applied to
explain motion detection in different species, including man
(for review, see [9]). The idea was that the nervous system
contains many of these local computational units, which
collectively cover the entire visual field, each extracting
locally the direction of image motion. The circuits shown in
Fig. 1a respond to rightward motion because light input
into the left channel of two adjacent facets is transmitted
with an additional delay relative to that into the right
channel. Thus the signals from both channels will be closely
aligned in time when they arrive at the downstream multipli-
cation unit so that the signals become amplified nonlinearly.
As a result, the model responds preferentially to the right-
ward motion, and a similar computation detects leftward
motion.
Since the initial EMD model was released, the local
motion detection has been a focus of successive theoretical
[10–13] and experimental [14–24] studies for more than
Microscopy, 2015, 37–44
doi: 10.1093/jmicro/dfu102
Advance Access Publication Date: 17 December 2014
© The Author 2014. Published by Oxford University Press on behalf of The Japanese Society of Microscopy. All rights reserved.
For permissions, please e-mail: journals.permissions@oup.com 37
at Janelia Farm Research Campus - Howard Hughes Medical Institute on February 6, 2015http://jmicro.oxfordjournals.org/Downloaded from
half a century. These conclusions were largely consistent
with the original proposal of the EMD model. However,
until recently, the EMD circuit had never been fully mapped
onto the concrete neural elements. Thus, for many decades,
we have not had conclusive evidence that those computa-
tions indeed occur in the actual nervous system.
A visual system in Drosophila
In the insect visual system, the image of the environment
formed by retina’s optics is then transmitted to the visual
centers or optic lobe. Both the retina and optic lobe have
specialized morphologies and functions to perform their
tasks. In the fly, the optic lobe is composed of four retino-
topically organized neuropils; the lamina, medulla, lobula
and lobula plate (Fig. 1b). Where is the EMD circuit present
in the optic neuropils? EMD circuits should lie at least
partly within the medulla, the second optic neuropil, as
shown by the following two lines of experimental evidence.
First, neither photoreceptors nor cells in the upstream
lamina show any directional selectivity [25–28]. Second, the
immediate downstream neuropil, the lobula plate, contains
large ‘lobula plate tangential cells’which exhibit strong direc-
tional selectivity [29]. These lobula plate cells integrate local
motion signals to produce wide-field motion response [29,
30]. These results suggest that the motion signals are com-
puted somewhere in between the lamina and the lobula plate.
Therefore it is essential to identify the neural circuits in the
medulla, but the medulla neuropil has been an impenetrable
region for a long time because of its size, neural complexity
and difficulty of neurophysiological approaches.
In the fly, each optic neuropil is an array of repeating
modules corresponding to the hexagonal lattice of omma-
tidia in the compound eye [31–34]. In the medulla, each
unit module, called a medulla column, receives stereotypic
columnar projections of lamina neurons [32]. These input
projections contain a pair of lamina cell types, L1 and L2,
together required for a fly to sense motion [23,35,36], after
receiving direct input from photoreceptors in the lamina
[37]. However, L1 and L2 themselves do not possess any
directional selectivity [38,39]. Thus, we should examine the
downstream medulla neurons to identify the circuits com-
puting motion signals. The medulla houses more than 60
different cell types, reported in the previous Golgi impregna-
tion study [40].
Connectome reconstruction using
serial-section transmission electron
microscopy
To identify the target medulla neurons postsynaptic to L1 and
L2 and provide the reliable foundation for further comprehen-
sive analysis on visual motion processing, a complete, dense
connectome of a single medulla column was reconstructed
using serial-section electron microscopy (serial-EM) [41]. To
date, serial-EM reconstruction remains the only reliable
method to determine neural circuits at the level of synapses
[42]. There are three established imaging techniques for
serial-EM reconstruction: (i) serial block-face electron micros-
copy (SBEM) [43], (ii) focused ion beam scanning electron
microscopy (FIB-SEM) [44] and (iii) serial-section transmis-
sion electron microscopy (ssTEM) (e.g. [31,42]). The method
of ssTEM was employed for the medulla reconstruction since
it provides the best x−yresolution of images; that seemed
advantageous for identification of tens of thousands of
synapses.
However, one obstacle to serial-EM reconstruction is the
need for processing massive amounts of EM data. To
Fig. 1. Visual motion detection in the Drosophila visual system. (a) Rightward motion component of Hassenstein–Reichardt EMD model. Light
input into the left channel (magenta) is transmitted with an additional delay, relative to the input to the right channel (cyan). The signals from
both channels therefore will arrive at the downstream multiplication unit closer in time to each other, and accordingly the signals become
enhanced nonlinearly. (b) Diagram of horizontal section of the fly’s brain and visual system. Fly’s optic lobe consists of four retinotopically
organized neuropils: the lamina, medulla, lobula and lobula plate. Scale bar =100 µm.
38 Microscopy, 2015, Vol. 64, No. 1
at Janelia Farm Research Campus - Howard Hughes Medical Institute on February 6, 2015http://jmicro.oxfordjournals.org/Downloaded from
reconstruct the volume of interest in an appropriate time
scale, a high-throughput, semi-automated reconstruction
pipeline was deployed for ssTEM [45], then applied to the
medulla reconstruction.
Semi-automated reconstruction pipeline
Connectome reconstruction using ssTEM is a multi-stage
process, starting with tissue preparation. The medulla tissue
was fixed by high-pressure freezing followed by freeze-
substitution, which preserved the native structure of speci-
mens to a greater extent than conventional methods [46].
The well-preserved tissue provides greater accuracy at each
stage of the reconstruction pipeline. Then, more than 2000
serial sections were cut through the entire depth of the
medulla, each roughly 40 nm thick (total sectioning time,
144 h). Three sections were lost during collection. Within
each section, a 90 × 90 µm field was imaged at the resolution
of 3.1 × 3.1 nm per pixel, taking 65 days. The image reso-
lution was sufficient to identify chemical synapses for both
pre- and postsynaptic sites.
The acquired EM image stack was then processed in the
reconstruction pipeline that consisted of five main tasks:
alignment, 2D segmentation, 3D linkage, synapse annota-
tion and proofreading.
Alignment
The separately acquired EM images of different sections and
different parts from the same section need to be registered.
First, all overlapping image pairs, both in a single section and
between consecutive sections, are aligned pairwise. Next, the
transformed images are fit into a global coordinate system via
a least-squares fit, then slightly warped to avoid discontinu-
ities at image seams (see [47] for details). An accurate image
alignment is critical for the rest of the automated processes as
well as the later manual proofreading. Although most image
defects can be handled automatically, some images have arti-
facts, such as section folds, that need extra care. In retrospect,
because of these artifacts and image distortions, ssTEM has
little or no advantage over the other SBEM or FIB-SEM tech-
niques, despite its higher resolution.
2D segmentation
Segmentation is the task that partitions an image into ‘mean-
ingful’sets of pixels, which in this case are distinct neural pro-
cesses. In a TEM dataset, given that the resolution of the
image stack is anisotropic (40 nm in z, 3.1 nm in x–y), 2D
image segmentation is executed first, followed by cross-section
linkage in 3D. The boundaries of each neuronal profile were
segmented by automated algorithms developed using different
segmentation techniques [48–52]. This auto-segmentation
produces two kinds of errors: (i) false boundaries or
over-segmentations, and (ii) missing boundaries or under-
segmentations. Decreasing either of these errors leads to an
increase of the other, resulting in a trade-off. Given that the
errors of over-segmentation are easier to fixinthemanual
proofreading than those of under-segmentation, the auto-
matic algorithms were tuned to produce an over-segmented
image volume. Such settings are also beneficial since they
segment small and thin processes more precisely. Those pro-
cesses are particularly crucial to the accurate connectivity map
because there are many small profiles at synaptic sites.
3D linkage
3D linkage identifies pairs of segments in adjacent sections
that are likely to belong to the same neuron, and links all
the pairs to assemble 3D shapes of neurons. Our automated
3D linkage constructed a linkage graph of consecutive seg-
ments based on evaluating their proximity and similarity
[53]. The accuracy of linkage is highly dependent on the
quality of section alignment and 2D segmentation.
Synapse annotation
To map all the chemical synapses in the volume, presynaptic
active zones must be identified. In the Drosophila visual
system, these are composed of presynaptic T-bar ribbon, often
seen as T-shaped structure. In invertebrates such as Drosoph-
ila and Caenorhabditis elegans,itiscommonforonepre-
synaptic site to signal several postsynaptic sites. In the
medulla, each T-bar is typically associated with three to eight
postsynaptic partners. These postsynaptic terminals are recog-
nized by their postsynaptic densities (PSDs), and then assigned
to their corresponding T-bar. A total of more than 10 000
T-bars and 38 000 postsynaptic sites were annotated in the
reconstructed volume, taking ∼300 person-hours. Although
the existence of gap-junctional coupling between particular
cell types was proposed previously [36], no such structure was
found in this EM preparation. The ultrastructure of candidate
gap junctions is perhaps fewer clear than chemical synapses.
More examination should validate the presence of gap junc-
tions in the future study.
Proofreading
Finally, the errors resulting from the above automatic seg-
mentation and linkage processes are corrected manually.
There are two major tasks in this step: (i) fixing errors of 2D
segmentation and 3D linkage to generate precise neuron
shape reconstructions, and (ii) tracing postsynaptic term-
inals sparsely. As described above, postsynaptic processes at
synaptic sites are often extremely small, and hence it is not
an easy task to segment and link them fully automatically.
Microscopy, 2015, Vol. 64, No. 1 39
at Janelia Farm Research Campus - Howard Hughes Medical Institute on February 6, 2015http://jmicro.oxfordjournals.org/Downloaded from
To provide meaningful synaptic connectivity, each postsy-
naptic terminal must be checked to be sure it is assigned to
its parent neuron correctly.
Although effort continues to improve the accuracy of each
automation step, the proofreading tasks are still fairly time-
consuming. It is therefore essential to have technical support
from a group of professional editors, referred to as proofrea-
ders, whose work is supervised by expert biologists. To
increase the reliability of the sparse tracing, two proofreaders
are assigned to each synapse. Then two connectomes from
the dual proofreader results were generated: an inclusive
version that included connections found by either proof-
reader, and a consensus version in which connections were
accepted only when both proofreaders agreed. Although the
inclusive connectome had ∼16% more connections, all the
additional connections had only one synaptic contact. All
connections with two or more synapses were present in both
connectomes. The analysis of medulla circuits used the con-
sensus connectome, but the conclusions remained unchanged
when using the inclusive connectome.
Cell identification from the reconstruction
Given the time-consuming nature of EM reconstructions,
the smallest possible medulla volume is required. Thus, an
initial reconstruction target was chosen as 10 × 10 × 50 µm
3
volume that permits a reconstruction of periodic module of
the medulla, a single medulla column. Within this volume,
the target column is densely reconstructed. The reconstruc-
tion volume is then expanded in order to trace some specific
connections between columns. Accordingly, the final recon-
struction volume becomes 37 × 37 × 50 µm
3
, in which the
surrounding 18 columns were sparsely reconstructed.
A total of 379 neurons were reconstructed in the medulla
volume (Fig. 2a). To map the reconstruction onto the existing
body of knowledge, the reconstructed bodies are assigned to
specific cell types previously proposed by light microscopy.
By matching the shapes of arbors with those in the previous
studies using Golgi impregnation [39] or more recent con-
focal microscopy of genetically labeled single cells (GSC)
[41], 56 distinct cell types are identified unequivocally among
the reconstructed neurons (Fig. 2b). Each of them is morpho-
logically unique. This includes the cell types for which a
Golgi counterpart could not be found but which are vali-
dated using isomorphs from GSC labeling. Each of these cells
is newly named and added to a medulla cell type library.
Candidate motion detection circuit in the
medulla
By combining the neural body identifications with synapse
annotation, a medulla connectome module was generated
(Fig. 3a). The connectome showed two medulla cell types,
Mi1 and Tm3, as postsynaptic targets of L1. It also revealed
the downstream connections of Mi1 and Tm3 to columnar
T4 neurons (Fig. 3a and b).
Several lines of evidence indicate that the motion signals
are relayed to the lobula plate via T4 cells [38,55–58]. A T4
cell has a unique morphology, having the dendrites in the
deepest layer of the medulla (stratum M10), and extending
two axonal projections into the lobula plate where one
makes terminal arbors within the neuropil, with another
process extending further to the end at its cell body location
outside the neuropil. A study using optical recordings on T4
cells demonstrated that specific subpopulations of T4s are
directionally tuned to one of the four cardinal directions:
upward, downward, forward and backward [59]. Those
direction preferences of T4s are also differentiated by their
terminal arbors in the lobula plate’s four sublayers, i.e. Lp1,
Lp2, Lp3 and Lp4 are responsible for motion front-to-back,
Fig. 2. Neuron reconstructions in the medulla. (a) A total of 379 neurons
reconstructed in the medulla volume. The dense reconstruction in the
initial volume, with the additional sparse reconstruction in the expanded
volume, was acquired for comprehensive analysis of medulla circuits.
(b) Cell types identified from the medulla reconstruction. The shapes of
arbors of reconstructed neurons are compared with those reported from
light microscopy. In many cases, a reconstructed cell is nicely matched
to the previously proposed cell type from Golgi impregnation study [39].
Some other reconstructions do not have any Golgi counterpart but they
are validated using isomorphs from more recent confocal microscopy
using a genetic approach (see [41] for details). Scale bar = 10 µm.
40 Microscopy, 2015, Vol. 64, No. 1
at Janelia Farm Research Campus - Howard Hughes Medical Institute on February 6, 2015http://jmicro.oxfordjournals.org/Downloaded from
back-to-front, down-to-up and up-to-down, respectively
(Fig. 3b).
Since the medulla connectome accurately captures not
only the presence but also the absence of strong connections
between any two cell types, Mi1 and Tm3 inputs to T4 are
known to be the two strongest paths bridging between L1
and T4. It is thus hypothesized that Mi1 and Tm3 inputs to
T4 implement the two arms of a motion detector (Fig. 3b,
details are as given below).
Anatomical receptive fields of T4 cells
To explore whether Mi1 and Tm3 converging onto T4 cells
could constitute the two arms of a motion detector, connec-
tions from individual Mi1 and Tm3 neurons onto the T4
cell in question were analyzed. It revealed that each T4
receives input from multiple Mi1s and Tm3s in the medulla,
suggesting that, unlike the circuit in the EMD model, several
points in the visual field constitute each arm of the motion
detector (Fig. 3b). To determine the visual receptive fields of
T4 inputs more in detail, synaptic connections between all
the relevant cells of each T4 were traced, i.e., connections
between L1 and Mi1/Tm3 and the downstream connections
from Mi1/Tm3 to T4. In this way, the T4 inputs can be
mapped as if onto a columnar array of L1s, and hence into
the visual field. These tracings were performed in the
expanded reconstruction volume of 19 medulla columns
(one central and 18 surrounding columns).
The T4 receptive fields of Mi1- and Tm3-mediated com-
ponents were calculated as follows: the number of synaptic
contacts from each L1 to an intermediate Mi1 is multiplied
by the number of contacts from the Mi1 to T4 (Fig. 3b). Simi-
larly, the number of contacts from each L1 to intermediate
Tm3s is multiplied by the number of contacts from the Tm3s
to T4, and then summed over all the Tm3 neurons that
receive input from the same L1. This multiplication is equiva-
lent to counting the number of independent synaptic routes
from each L1 to each T4, in which each route must use a dif-
ferent pair of the synaptic contacts between the L1 and the
intermediate target neurons, and the intermediate targets and
the T4. These calculations revealed that, although the ana-
tomical dendritic map of Mi1 and Tm3 shows substantial
overlaps, T4 receptive fields of Mi1- and Tm3-mediated
components are significantly displaced (Fig. 3c).
Fig. 3. Motion detection circuit revealed by the medulla connectome. (a) 3D graph representation of medulla connectome module. Cell types with
stronger connections are positioned closer to each other, using the visualization of similarities layout algorithm [54]. Clear separation is observed
between the pathways of L1 (spheres in purple) and L2 (spheres in green). The medulla connectome reconstruction identified two medulla cell
types, Mi1 and Tm3 (red arrows), as the two strongest paths bridging between L1 and L4. (b) Schematic representation of inputs to a single T4 cell
through Mi1 and Tm3 cells. Mock synaptic weights illustrate how the receptive fields are calculated (see the main text for details). The
connectome-based T4 receptive field strongly suggests that inputs to T4 from Mi1 and Tm3 cells implement the two arms of a motion detector. (c)
Cross-section view of dendritic arbors of Mi1 (cyan) and Tm3 (magenta), overlaid on the arrayof L1 terminals (yellow). The color saturation reflects
the number of synaptic contacts made onto the T4. T4 receptive fields of Mi1 component and Tm3 component overlap substantially with one
another, but these receptive fields are significantly displaced when they are calculated as illustrated in b. The direction of displacements (arrow, the
vector is fly’s backward in this case) correlates with the direction preference of the T4 on the assumption that the displacement vector is from the
Tm3 center of mass to the Mi1 center of mass. Scale bar in c= 8 µm. Figure panels are modified after Takemura et al. [41].
Microscopy, 2015, Vol. 64, No. 1 41
at Janelia Farm Research Campus - Howard Hughes Medical Institute on February 6, 2015http://jmicro.oxfordjournals.org/Downloaded from
T4 receptive fields offset correlates with the
direction preference
Is the direction of displacement between the Mi1 and Tm3
receptive field components for a T4 consistent with the direc-
tion preference of the T4? First, to determine the direction
preference of T4 cells, 19 T4s were selected for their locations
close to the central column, and then their axons were traced.
Of these, 16 were successfully traced down to the lobula
plate. These axon tracings determined the direction prefer-
ences of those T4s by their terminal arbor depth in the lobula
plate (Fig. 3b). Then, in the absence of evidence for synaptic
polarity onto T4s, the vector of receptive field offset is arbi-
trarily defined as the direction from the center of mass of the
Tm3 component to that of the Mi1 component. This assump-
tion adequately leads to the spatial displacements consistent
with the direction preference of the T4s (Fig. 3b and c).
Assuming that the Mi1 and Tm3 inputs to T4 are the
same excitatory synapses, as in the Hassenstein–Reichardt
EMD model, the Tm3 arm of the motion detector should
introduce a longer delay than the Mi1 arm. Conversely, if
the inputs were combined with opposing signs, as in the
alternative model, the Barlow–Levick EMD model, which
was suggested from the experimental findings of direction-
selective ganglion cells in the rabbit retina [60], then the pre-
diction would be reversed. A recent study of whole-cell
patch recording has shown significant temporal offset
between Mi1 and Tm3, with Mi1 exhibiting a delayed
response compared with Tm3 [61]. Considering these
response properties of Mi1 and Tm3, T4’s directional select-
ivity is likely accomplished by combining Mi1 and Tm3
inputs with opposite signs, as in the Barlow–Levick model.
Morphological feature differentiating T4
subtypes
There is an additional observation on T4’s anatomy. Close
examination of the T4 dendritic processes in the proximal
medulla (stratum M10) revealed that their branching arbors
are oriented primarily in one direction [41]. Furthermore,
the branch orientation of each T4 clusters around one of
four directions, when measured from the branch tip to
the axon main trunk. This observation reconfirmed the clas-
sification of each T4 into direction preference subtypes
which was determined by its terminal arbor depth in the
lobula plate. This characterization was then used to infer a
direction preference for the remaining three T4 cells, for
which tracing into the lobula plate failed. At the place of
signal convergence on T4 dendrites, there might be an
additional or alternative implementation of relative delay:
since T4 cells express both ionotropic and metabotropic
cholinoceptors as shown by single-cell transcript profiling,
these could mediate a fast and slow signal, respectively [34].
Comprehensive and detailed map of synaptic
connections
Here, in the connectomics study in Drosophila,asmallchunk
of neuropil was initially targeted and automatically segmented
for a dense, volumetric reconstruction. Subsequently, the
reconstruction volume was expanded to examine additional
neurons and synapses of particular interest, which were select-
ively and sparsely traced. The advantage of the dense recon-
struction is that it enables us a comprehensive analysis of
neural circuits and to argue both the presence and absence of
strong connections between any two cell types [41]. On the
other hand, the sparse reconstruction permits us to determine
the inputs to a particular neuron or a brain region more
quickly by skeletonizing neurons [62–65]. The medulla recon-
struction demonstrated that these two approaches are not
mutually exclusive.
Conclusions and outlook
The connectome reconstruction in the Drosophila optic
medulla removed a longstanding block to understanding the
exact mechanism of motion detection, by identifying the
underlying concrete neural elements and circuits. The con-
nectomic approach proved the importance of mapping the
entire synaptic connections in the region of interest,
showing that it can tell which types of cells are involved and
which types are not involved in a particular neural compu-
tation. Further, more generally, the study illustrates that
connectomes can, by identifying the underlying circuits,
provide key insight into neuronal computation.
When driving a car, a global positioning system with full
map information is appreciated when trying to get to the des-
tination via the shortest route. Likewise, it is essential to have a
full connectome map of the brain, to analyze functions or
identify the cause of diseases brought about by abnormal
neural circuits. This is precisely what connectomics aims at.
One obstacle is scaling up, however. How can we make the
jump from the small-scale circuits analyzed thus far to the
whole brain? This will require a lot of both work and patience,
but the effort practiced in the Drosophila connectomics should
keep evolving as the reconstruction strategy, analysis tools and
software, and imaging and automation technologies advance.
These developments and improvements will also be applicable
to larger-scale connectomics, including mammalian brains.
One area that is currently being greatly improved is EM
imaging. Imaging with better resolution than the minimum
required should provide enormous benefit for automation
processes. FIB-SEM offers the best z-resolution and is the
only technique that allows isotropic image datasets. In pre-
liminary results, the automated segmentation on FIB-SEM
data exhibits near-human performance. Moreover, although
the limited field of view of FIB-SEM has been a concern, a
42 Microscopy, 2015, Vol. 64, No. 1
at Janelia Farm Research Campus - Howard Hughes Medical Institute on February 6, 2015http://jmicro.oxfordjournals.org/Downloaded from
new strategy of lossless thick sectioning of embedded brain
tissue by hot-knife technology is being established, where a
large sample block can be losslessly cut into optimally sized
pieces for parallel FIB-SEM imaging (K. J. Hayworth et al.,
submitted for publication). FIB-SEM thus could provide an
avenue for scalable high-resolution imaging efforts. A com-
plete brain connectome of Drosophila is now being contem-
plated, generating a whole brain dataset.
Acknowledgements
I am indebted to all the members in the Janelia FlyEM project team for
their contributions and support and also grateful to Lou and Lucynda
Scheffer for reading the manuscript and for helpful comments. The
project was supported by the Howard Hughes Medical Institute.
References
1. Keene A C, Waddell S (2007) Drosophila olfactory memory:
single genes to complex neural circuits. Nat. Rev. Neurosci. 8:
341–354.
2. Dickson B J (2008) Wired for sex: the neurobiology of Drosoph-
ila mating decisions. Science 322: 904–909.
3. Olsen S R, Wilson R I (2008) Cracking neural circuits in a tiny
brain: new approaches for understanding the neural circuitry of
Drosophila.Trends Neurosci. 31: 512–520.
4. Venken K J T, Simpson J H, Bellen H J (2011) Genetic manipula-
tion of genes and cells in the nervous system of the fruit fly.
Neuron 72: 202–230.
5. Silies M, Gohl D M, Clandinin T R (2014) Motion-detecting cir-
cuits in flies: Coming into view. Annu. Rev. Neurosci. 37: 307–327.
6. Borst A (2014) Fly visual course control: behaviour, algorithms
and circuits. Nat. Rev. Neurosci. 15: 590–599.
7. Hassenstein B, Reichardt W (1956) Systemtheoretische Analyse
der Zeit-, Reihenfolgen- und Vorzeichenauswertung bei der Bewe-
gungsperzeption des Rüsselkäfers Chlorophanus.Z. Naturforsch.
B11: 513–524.
8. Reichardt W (1961) Autocorrelation, a principle for the evalu-
ation of sensory information by the central nervous system. In:
Rosenblith W A (ed.), Sensory Communication, pp. 303–317,
(MIT Press, New York).
9. Borst A, Egelhaaf M (1989) Principles of visual motion detection.
Trends Neurosci. 12: 297–306.
10. Poggio T, Reichardt W (1976) Visual control of orientation
behaviour in the fly: part II. Towards the underlying neural inter-
actions. Q. Rev. Biophys. 9: 377–438.
11. Adelson E H, Bergen J R (1985) Spatiotemporal energy models
for the perception of motion. J. Opt. Soc. Am. A 2: 284–299.
12. Hildreth E C, Koch C (1987) The analysis of visual motion: from
computational theory to neuronal mechanisms. Annu. Rev.
Neurosci. 10: 477–533.
13. Potters M, Bialek W (1994) Statistical mechanics and visual
signal processing. J. Phys. I (France) 4: 1755–1775.
14. Buchner E (1976) Elementary movement detectors in an insect
visual system. Biol. Cybern. 24: 85–101.
15. Buchner E (1984) Behavioural analysis of spatial vision in insects.
In: Ali M A (ed.), Photoreception and Vision in Invertebrates, pp.
561–621, (Plenum Press, New York).
16. Borst A, Egelhaaf M (1987) Temporal modulation of luminance
adapts time constant of fly movement detectors. Biol. Cybern. 56:
209–215.
17. Egelhaaf M, Reichardt W (1987) Dynamic-response properties of
movement detectors: theoretical-analysis and electrophysiological
investigation in the visual-system of the fly. Biol. Cybern. 56: 69–87.
18. Egelhaaf M, Borst A (1989) Transient and steady-state response
properties of movement detectors. J. Opt. Soc. Am. A 6: 116–127.
19. Egelhaaf M, Borst A, Reichardt W (1989) Computational struc-
ture of a biological motion-detection system as revealed by local
detector analysis in the fly’s nervous system. J. Opt. Soc. Am. A
6: 1070–1087.
20. Schuling F H, Mastebroek H A K, Bult R, Lenting B P M (1989)
Properties of elementary movement detectors in the flyCalliphora
erythrocephala.J. Comp. Physiol. A 165: 179–192.
21. Zanker J M, Srinivasan M V, Egelhaaf M (1999) Speed tuning in
elementary motion detectors of the correlation type. Biol.
Cybern. 80: 109–116.
22. Borst A, Flanagin V L, Sompolinsky H (2005) Adaptation
without parameter change: dynamic gain control in motion detec-
tion. Proc. Natl. Acad. Sci. USA 102: 6172–6176.
23. Clark D A, Bursztyn L, Horowitz M A, Schnitzer M J,
Clandinin T R (2011) Defining the computational structure of
the motion detector in Drosophila. Neuron 70: 1165–1177.
24. Eichner H, Joesch M, Schnell B, Reiff D F, Borst A (2011)
Internal structure of the fly elementary motion detector. Neuron
70: 1155–1164.
25. Laughlin S B, Hardie R C (1978) Common strategies for light
adaptation in the peripheral visual systems of fly and dragonfly.
J. Comp. Physiol. A 128: 319–340.
26. Laughlin S B (1994) Matching coding, circuits, cells, and mole-
cules to signals: general principles of retinal design in the fly’seye.
Prog. Retin. Eye Res. 13: 165–196.
27. Zheng L, de Polavieja G G, Wolfram V, Asyali M H, Hardie R C,
Juusola M (2006) Feedback network controls photoreceptor
output at the layer of first visual synapses in Drosophila. J. Gen.
Physiol. 127: 495–510.
28. Zheng L, Nikolaev A, Wardill T J, O’Kane C J, de Polavieja G G,
Juusola M (2009) Network adaptation improves temporal
representation of naturalistic stimuli in Drosophila eye: I Dynam-
ics. PLoS ONE 4: e4307.
29. Joesch M, Plett J, Borst A, Reiff D F (2008) Response properties
of motion-sensitive visual interneurons in the lobula plate of
Drosophila melanogaster. Curr. Biol. 18: 368–374.
30. Krapp H G, Hengstenberg R (1996) Estimation of self-motion by
optic flow processing in single visual interneurons. Nature 384:
463466.
31. Meinertzhagen I A, Sorra K E (2001) Synaptic organisation in the
fly’s optic lamina: few cells, many synapses and divergent micro-
circuits. Prog. Brain Res. 131: 53–69.
32. Takemura S Y, Lu Z, Meinertzhagen I A (2008) Synaptic circuits
of the Drosophila optic lobe: the input terminals to the medulla.
J. Comp. Neurol. 509: 493–513.
33. Rivera-Alba M, Vitaladevuni S, Mishchenko Y, Lu Z,
Takemura S Y, Scheffer L K, Meinertzhagen I A, Chklovskii D B,
Microscopy, 2015, Vol. 64, No. 1 43
at Janelia Farm Research Campus - Howard Hughes Medical Institute on February 6, 2015http://jmicro.oxfordjournals.org/Downloaded from
de Polavieja G G (2011) Wiring economy and volume exclusion
determine neuronal placement in the Drosophila brain. Curr.
Biol. 21: 2000–2005.
34. Shinomiya K, Karuppudurai T, Lin T Y, Lu Z, Lee C H,
Meinertzhagen I A (2014) Candidate neural substrates for off-edge
motion detection in Drosophila.Curr. Biol. 24: 1062–1070.
35. Rister J, Pauls D, Schnell B, Ting C Y, Lee C H, Sinakevitch I,
Morante J, Strausfeld N J, Ito K, Heisenberg M (2007) Dissection
of the peripheral motion channel in the visual system of Drosoph-
ila melanogaster. Neuron 56: 155–170.
36. Joesch M, Schnell B, Raghu S V, Reiff D F, Borst A (2010) ON
and OFF pathways in Drosophila motion vision. Nature 468:
300–304.
37. Meinertzhagen I A, O’Neil S D (1991) Synaptic organization of
columnar elements in the lamina of the wild type in Drosophila
melanogaster. J. Comp. Neurol. 305: 232–263.
38. Douglass J K, Strausfeld N J (2003) Anatomical organization of
retinotopic motion-sensitive pathways in the optic lobes of flies.
Microsc. Res. Tech. 62: 132–150.
39. Borst A, Haag J, Reiff D F (2010) Fly motion vision. Annu. Rev.
Neurosci. 33: 49–70.
40. Fischbach K-F, Dittrich A P M (1989) The optic lobe of Drosoph-
ila melanogaster. I. A Golgi analysis of wild-type structure. Cell
Tissue Res. 258: 441–475.
41. Takemura S Y, Bharioke A, Lu Z, Nern A, Vitaladevuni S,
Rivlin P K, Katz W T, Olbris D J, Plaza S M, Winston P, et al.
(2013) A visual motion detection circuit suggested by Drosophila
connectomics. Nature 500: 175–181.
42. White J G, Southgate E, Thomson J N, Brenner S (1986) The
structure of the nervous system of the nematode Caenorhabditis
elegans.Phil. Trans. R. Soc. Lond. B 314: 1–340.
43. Denk W, Heinz H (2004) Serial block-face scanning electron
microscopy to reconstruct three-dimensional tissue nanostruc-
ture. PLoS Biol. 2: e329.
44. Knott G, Marchman H, Wall D, Lich B (2008) Serial section
scanning electron microscopy of adult brain tissue using focused
ion beam milling. J. Neurosci. 28: 2959–2964.
45. Chklovskii D B, Vitaladevuni S, Scheffer L K (2010) Semi-
automated reconstruction of neural circuits using electron micros-
copy. Curr. Opin. Neurobiol. 20: 667–675.
46. Sosinsky G E, Crum J, Jones Y Z, Lanman J, Smarr B, Terada M,
Martone M E, Deerinck T J, Johnson J E, Ellisman M H (2008)
The combination of chemical fixation procedures with high pres-
sure freezing and freeze substitution preserves highly labile tissue
ultrastructure for electron tomography applications. J. Struct.
Biol. 161: 359–371.
47. Scheffer L K, Karsh B, Vitaladevuni S (2013) Automated align-
ment of imperfect EM images for neural reconstruction. Preprint
at http://arXive.org/abs/1304.6034.
48. Canny J (1986) A computational approach to edge detection.
IEEE Trans. Pattern Anal. Mach. Intell. 8: 679–698.
49. Vincent L, Soille P (1991) Watersheds in digital spaces: an effi-
cient algorithm based on immersion simulations. IEEE Trans.
Pattern Anal. Mach. Intell. 13: 583–598.
50. Soille P (2003) Morphological Image Analysis: Principles and
Applications, 2nd ed., (Springer-Verlag, New York).
51. Martin D R, Fowlkes C C, Malik J (2004) Learning to detect
natural image boundaries using local brightness, color, and
texture cues. IEEE Trans. Pattern Anal. Mach. Intell. 26:
530–549.
52. Dollar P, Tu Z, Belongie S (2006) Supervised learning of edges and
object boundaries. IEEE Comp. Soc. Conf. Comp. Vis. Pattern
Rec. 2: 1964–1971.
53. Mishchenko Y (2009) Automation of 3D reconstruction of neural
tissue from large volume of conventional serial section transmis-
sion electron micrographs. J. Neurosci. Methods 176: 276–289.
54. Van Eck N J, Waltman L (2007) VOS: a new method for visualiz-
ing similarities between objects. Adv. Data Anal. 299–306.
55. Buchner E, Buchner S, Bülthoff I (1984) Deoxyglucose mapping
of nervous activity induced in Drosophila brain by visual move-
ment. J. Comp. Physiol. A 155: 471–483.
56. Strausfeld N J, Lee J K (1991) Neuronal basis for parallel visual
processing in the fly. Vis. Neurosci. 7: 13–33.
57. Bausenwein B, Fischbach K F (1992) Activity labelling patterns in
the medulla of Drosophila melanogaster caused by motion
stimuli. Cell Tissue Res. 270: 25–35.
58. Schnell B, Raghu S V, Nern A, Borst A (2012) Columnar cells
necessary for motion responses of wide-field visual interneurons
in Drosophila.J. Comp. Physiol. A 198: 389–395.
59. Maisak M S, Haag J, Ammer G, Serbe E, Meier M, Leonhardt A,
Schilling T, Bahl A, Rubin G M, Nern A, et al. (2013) A direc-
tional tuning map of Drosophila elementary motion detectors.
Nature 500: 212–216.
60. Barlow H B, Levick W R (1965) The mechanism of directionally
selective units in rabbit’s retina. J. Physiol. 178: 477–504.
61. Behnia R, Clark D A, Carter A G, Clandinin T R, Desplan C
(2014) Processing properties of ON and OFF pathways for Dros-
ophila motion detection. Nature 512: 427–430.
62. Saalfeld S, Cardona A, Hartenstein V, Tomančák P (2009)
CATMAID: collaborative annotation toolkit for massive
amounts of image data. Bioinformatics 25: 1984–1986.
63. JeongWK,BeyerJ,HadwigerM,BlueR,LawC,Vazquez-ReinaA,
Reid R C, Lichtman J, Pfister H. (2010) Ssecrett and NeuroTrace:
interactive visualization and analysis tools for large-scale neurosci-
ence data sets. IEEE Comput. Graph Appl. 30: 58–70.
64. Helmstaedter M, Briggman K L, Denk W (2011) High-accuracy
neurite reconstruction for high-throughput neuroanatomy. Nat.
Neurosci. 14: 1081–1088.
65. Xu M, Jarrell T A, Wang Y, Cook S J, Hall D H, Emmons S W
(2013) Computer assisted assembly of connectomes from electron
micrographs: application to Caenorhabditis elegans.PLoS ONE
8: e54050.
44 Microscopy, 2015, Vol. 64, No. 1
at Janelia Farm Research Campus - Howard Hughes Medical Institute on February 6, 2015http://jmicro.oxfordjournals.org/Downloaded from