Resource

A Complete Electron Microscopy Volume of the Brain

of Adult Drosophila melanogaster
Graphical Abstract Authors
Zhihao Zheng, J. Scott Lauritzen,
Eric Perlman, ..., Stephan Saalfeld,
Richard D. Fetter, Davi D. Bock

Correspondence
[email protected]

In Brief
Electron microscopy imaging of the entire
adult fruit fly brain at synapse resolution
reveals circuitry spanning multiple
regions and connectivity between known
and previously unknown cell types.

Highlights
d A complete adult Drosophila brain was imaged with EM and
has been made publicly available

d The imaged volume enables brain-spanning mapping of

circuits at synaptic resolution

d All mushroom body (MB) calyx inputs were mapped,

revealing a new cell type, MB-CP2

d Previously unidentified synaptic partners form recurrent

microcircuits in MB calyx

Zheng et al., 2018, Cell 174, 730–743

July 26, 2018 ª 2018 The Authors. Published by Elsevier Inc.
https://doi.org/10.1016/j.cell.2018.06.019
 Resource

A Complete Electron Microscopy Volume

of the Brain of Adult Drosophila melanogaster
Zhihao Zheng,1,7 J. Scott Lauritzen,1,7 Eric Perlman,1,8 Camenzind G. Robinson,1 Matthew Nichols,1 Daniel Milkie,2,9
Omar Torrens,2 John Price,3 Corey B. Fisher,1 Nadiya Sharifi,1 Steven A. Calle-Schuler,1 Lucia Kmecova,1 Iqbal J. Ali,1
Bill Karsh,1 Eric T. Trautman,1 John A. Bogovic,1 Philipp Hanslovsky,1 Gregory S.X.E. Jefferis,4,5 Michael Kazhdan,6
Khaled Khairy,1 Stephan Saalfeld,1 Richard D. Fetter,1,10 and Davi D. Bock1,11,*
1Janelia Research Campus, Howard Hughes Medical Institute, Ashburn, VA 20147, USA
2Coleman Technologies, Newtown Square, PA 19073, USA
3Hudson Price Designs, Hingham, MA 02043, USA
4Division of Neurobiology, MRC Laboratory of Molecular Biology, Cambridge CB2 0QH, UK
5Department of Zoology, University of Cambridge, Cambridge CB2 3EJ, UK
6Department of Computer Science, Johns Hopkins University, Baltimore, MD 21218, USA
7These authors contributed equally
8Present address: Center for Imaging Science, Johns Hopkins University, Baltimore, MD 21218, USA
9Present address: Janelia Research Campus, Howard Hughes Medical Institute, Ashburn, VA 20147, USA
10Present address: Department of Biochemistry and Biophysics, University of California, San Francisco, San Francisco, CA 94158, USA
11Lead Contact

*Correspondence: [email protected]
https://doi.org/10.1016/j.cell.2018.06.019

SUMMARY are brain-spanning, and access to whole-brain connectomes

has been limited to a few small organisms, such as the nematode
Drosophila melanogaster has a rich repertoire of C. elegans, the larva of the fruit (or vinegar) fly, Drosophila mela-
innate and learned behaviors. Its 100,000-neuron nogaster, and the tadpole larva of the tunicate Ciona intestinalis
brain is a large but tractable target for comprehen- (Ohyama et al., 2015; Ryan et al., 2016; White et al., 1986).
sive neural circuit mapping. Only electron micro- The adult fruit fly has emerged as a key genetic model system
scopy (EM) enables complete, unbiased mapping for interrogating the neuronal substrates of sophisticated behav-
iors, such as place learning, flight control, courtship, grooming,
of synaptic connectivity; however, the fly brain is
and memory-driven action selection (Dickinson and Muijres,
too large for conventional EM. We developed a
2016; Hampel et al., 2015; Ofstad et al., 2011; Owald and Wad-
custom high-throughput EM platform and imaged dell, 2015; Pavlou and Goodwin, 2013). Given the morphological
the entire brain of an adult female fly at synaptic res- and physiological stereotypy and genetic accessibility of
olution. To validate the dataset, we traced brain- neuronal cell types in the fly brain, connectomes of circuits un-
spanning circuitry involving the mushroom body derlying these behaviors should translate well across individuals
(MB), which has been extensively studied for its and significantly accelerate the dissection of the neuronal
role in learning. All inputs to Kenyon cells (KCs), the basis for behavior (Gruntman et al., 2018; Jovanic et al., 2016;
intrinsic neurons of the MB, were mapped, revealing Ohyama et al., 2015; Takemura et al., 2017a, 2017b). However,
a previously unknown cell type, postsynaptic part- at
8 3 107 mm3 and
100,000 neurons (Simpson, 2009), the
ners of KC dendrites, and unexpected clustering of brain of an adult fly is two orders of magnitude larger than that
of the fruit fly larva, the next-largest brain imaged at synaptic
olfactory projection neurons. These reconstructions
resolution (Ohyama et al., 2015). This combination of scale and
show that this freely available EM volume supports
resolution has heretofore been unattainable by volume electron
mapping of brain-spanning circuits, which will signif- microscopy (EM), the only method capable of simultaneously
icantly accelerate Drosophila neuroscience. resolving all neuronal branches and synapses in a given volume
of brain tissue (Helmstaedter et al., 2008). Therefore, we built
INTRODUCTION new hardware and software for high-speed acquisition and pro-
cessing of serial section transmission EM (TEM) images and
How brain circuits allow animals to implement complex behavior used this infrastructure to image a whole-fly brain at synaptic
remains a central mystery of neurobiology. Where available, resolution (Figure 1).
neuronal ‘‘wiring diagrams’’ or ‘‘connectomes’’ (Lichtman and We took multiple approaches to validate our whole-brain vol-
Sanes, 2008)—maps of the synaptic connectivity between the ume for tracing synaptic connectivity of brain-spanning
neurons in a circuit—have proved extremely useful for under- neuronal circuits. Our efforts focused on the mushroom body
standing circuit function (Ding et al., 2016; Jarrell et al., 2012; (MB), which has been intensively studied for its role in associa-
Kasthuri et al., 2015; Ohyama et al., 2015; Takemura et al., tive memory formation and recall (Guven-Ozkan and Davis,
2017b; Wanner et al., 2016). However, many neuronal circuits 2014; Keene and Waddell, 2007). Olfactory projection neurons

730 Cell 174, 730–743, July 26, 2018 ª 2018 The Authors. Published by Elsevier Inc.
This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
Figure 1. Target Volume and EM Acquisition Infrastructure
(A) Oblique view of a surface model of the Drosophila brain (gray mesh) with specific neuropil compartments highlighted: AL (orange), MB (pink), and LH (green).
(B) Schematic of olfactory pathway. Approximately 150 PNs, divided into
50 subtypes based on their anatomically defined glomeruli of origin in the AL, project to
the MB calyx and LH (Grabe et al., 2016; Masse et al., 2009). Lateral horn is thought to mediate innate olfactory behaviors, whereas the MB is involved in learned
behaviors (Keene and Waddell, 2007). In MB calyx, PN collaterals terminate in boutons and synapse on KCs (Yasuyama et al., 2002). Most PN types project to the
MB calyx via the mALT, but several travel in secondary tracts (data not shown), and a few bypass calyx and project only to LH (Frank et al., 2015; Stocker et al.,
1990; Tanaka et al., 2012).
(C) Workflow for the generation of the whole-brain dataset. Blocks of brain tissue are incubated in heavy metals to label cell membranes, embedded in a resin
polymer, and screened with X-ray tomography. Blocks were then serially sectioned with a diamond knife (for an introduction to serial sectioning techniques, see
Harris et al., 2006). Groups of three serial sections are placed on metal slot grids for imaging in one of the two custom high-throughput TEM systems (TEMCA2 or
ATPS). The imaged sections were assembled into an aligned volume with the custom software pipeline described here. Reconstruction and analyses of neural
circuits in the volume were conducted with the CATMAID tracing environment (Saalfeld et al., 2009).
D, dorsal; V, ventral; A, anterior; P, posterior; M, medial; L, lateral; LH, lateral horn; mALT, medial antennal lobe tract. See also Figures S1, S2, and S3 and
Videos S1, S2, S3, and S4.

(PNs) provide the main sensory input to the MB; their connec- Independent tracing of Kenyon cell dendrites, which have some
tions with the MB intrinsic neurons, Kenyon cells (KCs), form a of the finest neurites in the fly brain (Yasuyama et al., 2002), pro-
crucial stage of a fan-out fan-in network analogous to brain vided a sensitive test of the consistency of neural reconstructions.
structures including the mammalian cerebellum (Farris, 2011; Retrograde tracing from KC dendrites provided a complete
Stevens, 2015). Furthermore the logic of PN to KC connectivity enumeration of olfactory PN input to the MB and yielded an
ratios has been the subject of detailed experimental and improved map of local circuitry in the calyx, the initial site for sam-
theoretical analysis as a model for the construction of high pling and processing of sensory information in the MB. This re-
dimensional sensory representations (e.g., Caron et al., 2013; vealed principles of coordinated organization that were invisible
Litwin-Kumar et al., 2017). KCs and PNs are brain spanning, in previous work using light level data assembled from many
morphologically stereotyped, and anatomically extremely well different brains; for example we found a high degree of clustering
described at the light level (Aso et al., 2014; Jefferis et al., of PN inputs, which may generate biases in PN-to-KC connectiv-
2007; Tanaka et al., 2012), making them well suited for vali- ity and therefore shape how olfactory PN input to the MB is
dating the accuracy of neural reconstructions in the volume. sampled. Although the MB has been intensively studied, we
We have developed software tools to enable co-visualization, also discovered a previously unknown, brain-spanning neuron
quantitative analysis, and rapid cell type identification by merg- that provides input to KCs and likely relays non-olfactory, multi-
ing EM reconstructions with existing large-scale light micro- modal information to the calyx. Finally, we show that KC dendrites
scopy (LM) databases of neuronal morphology (Chiang et al., make output synapses onto a small subset of available cell types,
2011; Costa et al., 2016; Milyaev et al., 2012). defining a specific local recurrent microcircuit in the calyx.

Cell 174, 730–743, July 26, 2018 731

 In conclusion, we describe the largest synaptic-resolution, EM Image Acquisition and Volume Reconstruction of a
whole-brain EM dataset obtained to date. We show that it en- Complete Adult Fly Brain
ables efficient mapping and identification of both known The above-described infrastructure provided an order-of-
and unknown neurons in adult Drosophila, a key model sys- magnitude increase in EM imaging capacity, sufficient to image
tem for circuit neuroscience and, crucially, that it enables reli- the complete brain of a female adult fruit fly at synaptic resolution
able and efficient determination of synaptic connectivity. We in a reasonable amount of calendar time. To span the depth
have made these data and supporting software freely avail- (
250 mm) of the entire fly brain, 7,062 serial
40-nm thin sec-
able for download and immediate use by the scientific tions were cut from a sample optimized for both high membrane
community. contrast and high-quality ultrastructure (Figures 1A, 1C, and
S3A). In total, 7,050 (99.8%) sections were successfully imaged,
RESULTS resulting in a
106 TB dataset comprising
21 million camera
images.
New Tools for Volume EM Data Acquisition The quality of acquired image data was high (Figures 2, S4,
To meet the challenge of acquiring a whole fly brain, we used a and S5; Video S4). In general, image quality improves with in-
variation of classical serial section TEM (ssTEM), in which im- creases to both image resolution and signal-to-noise ratio
ages are acquired at high-speed with a TEM camera array (SNR). Whether a given EM volume has sufficient resolution to
(TEMCA) (Bock et al., 2011). Although sample handling for reliably detect synapses and trace fine neuronal processes can
TEM is challenging (Figures S3A–S3D), in comparison to scan- only be evaluated empirically. We found that the SNR of images
ning EM-based methods, the intrinsically parallel nature of the in the whole-brain volume equals or exceeds that of other pub-
electron optical image formed in TEM makes it relatively straight- licly available datasets (Figures 2G and S4). If greater resolution
forward to achieve high-quality EM images at high-speed (re- is needed in the future, targeted ROIs can be reimaged at higher
viewed in Briggman and Bock, 2012). magnification (Figures S5A–S5C).
We built two second-generation TEMCA (TEMCA2) systems A custom software pipeline was developed to assemble the
(Figures S2A and S3E), using high-speed sCMOS cameras millions of individual camera images into a coherent volume,
and made a custom piezo-driven Fast Stage (Figures S2B, a process known as ‘‘volume reconstruction.’’ A distributed
S3F, and S3G) (Price and Bock, 2016a). Image mosaics were compute cluster was used to stitch images from each section
needed because at typical magnification, each camera in the into a mosaic and then to register mosaics across all sections
array had an
8 mm field of view (FOV), whereas each whole- into an aligned volume (Figures 1C and 2). Artifacts and distor-
brain thin section was
750-mm wide 3
350-mm tall (Fig- tions introduced in serial sectioning, pickup, staining, and imag-
ure 1A). At 4 nm/pixel, this resulted in a large (
187,500 3 ing of samples (Bock et al., 2011; Saalfeld et al., 2012), were
87,500 pixel,
16 GB) stitched image mosaic for each section. either accommodated or corrected, resulting in a high-quality
Each FOV was typically acquired using four 35 ms frames, volume reconstruction of the whole brain (Figures 2 and S5;
versus
1 s in conventional TEM imaging systems. The Fast Video S4). Overall,
3% of camera images could not be incorpo-
Stage moved one FOV in 30–50 ms (including settle time; Fig- rated, due to data loss during sample preparation or imaging,
ures S2C and S2D; Video S1), versus the
4 s in conventional and
0.5% had poor intensity correction due to folds or EM-
systems. The combination of high-speed imaging acquisition dense precipitates (Figures S5D, S5E, and S5K). Fewer sections
and sample translation allowed a single whole-brain thin sec- were lost than in the complete Drosophila larval central nervous
tion to be imaged in less than 7 min, for a per-section system volume, a dataset that has supported multiple significant
throughput of
50 MPix/s (
5 times faster than the first-gener- circuit mapping efforts (Jovanic et al., 2016; Ohyama et al., 2015;
ation TEMCA (Bock et al., 2011), and
40x faster than conven- Schlegel et al., 2016; Schneider-Mizell et al., 2016).
tional TEM systems (Takemura et al., 2013). Since each sample
grid (Figure S3H) typically supported three whole-brain sections Independent Tracing of Fine Dendrites Shows
(Figure 1C), and it takes
10 min to exchange grids and define Reproducibility of Reconstruction
target regions of interest (ROIs), per-grid throughput, which in- Three teams independently reconstructed the dendritic arbor
cludes sample exchange and ROI definition by microscopists, and proximal axon of a KC to completion (STAR Methods).
was
27 MPix/s. This task was representative of challenging, fine-scale recon-
To decrease overhead and to allow unattended multi-day im- struction targets, since KC dendritic claws, which ensheath PN
aging, we built a robotic Automated Transport and Positioning boutons in a characteristic structure called the microglomerulus,
System (ATPS) (Figures S2E, S2F, and S3I–S3N) (Price and are known to have some of the finest neurites (
40 nm) in the fly
Bock, 2015, 2016a, 2016b). It enabled autonomous sample ex- brain (Yasuyama et al., 2002) (Figure S1A). Our tracing method
change in
5 min and autonomous imaging of TEM grids using extended the efficient ‘‘iterative’’ approach of Schneider-Mizell
predefined ROIs (Figure 1C; Videos S2 and S3). We mounted et al. (2016). Every neuron was initially traced and then proofread
the ATPS on a single-camera TEM, resulting in per-section by a different team member. The dendrite claw morphologies
throughput
1/4 that of a TEMCA2 system. Nonetheless, the and synaptic connections were highly consistent between the
ATPS had greatly improved reliability, capacity, and throughput three independent reconstructions (Figure S6), and were qualita-
compared to previous TEM-based platforms for automated tively similar to previous reconstructions of KC dendrites in the
sample exchange and imaging (Lefman et al., 2007; Potter MB calyx (Butcher et al., 2012; Yasuyama et al., 2002). In the
et al., 2004). fly brain, a useful dichotomy can be drawn between larger,

732 Cell 174, 730–743, July 26, 2018

Figure 2. Reconstructed Image Volume
(A–F) Renderings of brain-spanning EM in the sectioning plane (x-y axes) at successive zoom levels. All panels rendered using the ELM viewer (STAR Methods),
which averages several adjacent sections to improve contrast at low magnifications. Red dotted lines in left column indicate orthogonal (y-z axes) section plane
through the whole-brain volume, rendered in right column.
(G) Image signal-to-noise ratio (SNR) versus per-section acquisition rates for the current dataset (TEMCA2) and publicly available volume EM datasets acquired
via comparable techniques: FIB-scanning EM (Takemura et al., 2015), SBEM (Briggman et al., 2011), ATUM-scanning EM (Kasthuri et al., 2015), ssTEM
(Takemura et al., 2013), and TEMCA (Bock et al., 2011). Error bars indicate SD.
(H–K) Serial thin sections succeeding the one in (F). Fine processes can be followed across serial sections and section-to-section image registration is accurate
enough to provide a consistent FOV.
Axis labels are the same as those used in Figure 1. Scale bars, 200 (A), 100 (B), 25 (C), 10 (D), 2 (E), and 0.4 mm (F and H–K).
See also Figures S4, S5, and S6.

Cell 174, 730–743, July 26, 2018 733

microtubule-containing neurites, termed ‘‘backbones,’’ and fine, in NBLAST scores of 0.64–0.71 (scale of 1 to 1, where 1 indi-
microtubule-free neurites that extend over short distances, cates identical neuron shape and position), and the EM-recon-
termed ‘‘twigs’’ (Schneider-Mizell et al., 2016). In the indepen- structed VM2 PN returned an NBLAST score of 0.64 when
dently reconstructed KC, backbones contained no errors, and compared to the same PNs. Qualitative assessment of similarity
the few discrepancies in twigs were due only to errors of omis- agreed with the quantitative scoring (Figures 3E–3G). Since the
sion, not commission. Our tracing approach (STAR Methods) is LM-imaged template brains have also been aligned to one
biased to produce this category of error, which has been shown another (Manton et al., 2014), this tool chain can be used to
to have minimal impact on network connectivity maps in fly quantitatively compare any aligned LM- and EM-reconstructed
(Schneider-Mizell et al., 2016). The error rate was 60.8 mm/error neurons.
(454 mm total cable length; 37.9 mm/error in twigs alone; 287 mm Of the 578 boutons providing input to KC claws, 497 arose
cable length). This rate is comparable to that reported for from olfactory PNs (86%, from 114 PNs). The PN dendrites in
Drosophila larva (27.2 mm/error, across 6 neurons totaling the antennal lobe (AL) permitted identification of all 50 known
2,666 mm in cable length; 16.2 mm/error in twigs alone, olfactory glomeruli (Grabe et al., 2015) (Figures 4A–4C), although
1,539 mm cable length), and in mouse retina (83.4 mm/error two could not be disambiguated (VC5 and VC3l; STAR Methods).
across 1 neuron totaling 600 mm cable length; Helmstaedter All classified PNs are included in a downloadable tracing envi-
et al., 2011; Schneider-Mizell et al., 2016, respectively). Thus, ronment for use in future efforts mapping olfactory circuits in
the independent KC reconstructions demonstrate that the EM the fly brain (Key Resources Table; STAR Methods). Seventeen
data support accurate tracing of challenging, fine-scale neuronal boutons (3%) arose from a previously unknown neuron that we
morphology and connectivity. named ‘‘MB-CP2’’ (‘‘Mushroom Body Calyx Pedunculus #2,’’
per the naming convention of Tanaka et al. (2008) and further
A Complete Survey of KC Inputs Reveals Tight describe below. The remaining 11% of boutons come from other
Clustering of Homotypic PN Arbors unidentified (presumably non-olfactory) PN types and other cell
To validate that brain-spanning neuronal connectivity can be types from outside the AL. Consistent with previous LM data,
retrogradely reconstructed, the neurons providing input to all the arbors of specific PN subtypes formed concentric clusters
KC claw microglomeruli in the MB calyx on the right side of the in MB calyx (Figure 4C; compare to Figure 4 in Tanaka et al.,
brain were sufficiently traced to identify cell type. The mean 2004). Unsupervised clustering based on NBLAST scores
reconstruction rate for all identified PNs (including tracing and grouped homotypic PNs (Figure 4D). The number of PNs arising
proofreading of skeletons and synapses) was 120.1 mm/hr from each glomerulus (Figure 4E) was consistent with recent LM
(11.2 hr per cell), and 70.1 mm/hr (11.1 hr per cell) for 15 KCs. data (Grabe et al., 2016).
These rates compare favorably to the
73 mm/h for proofread Previously, LM data pooled across multiple animals showed
skeletons (including synapses) reported by Schneider-Mizell that homotypic PN collaterals fasciculate tightly in lateral horn,
et al. (2016) and are comparable to 169.5 mm/hr (excluding syn- but only loosely (or not at all) in MB calyx (Jefferis et al., 2007).
apses) in mammalian retina (Helmstaedter et al., 2011). In contrast, our reconstructions in an individual brain showed
As expected (Caron et al., 2013; Yagi et al., 2016), most KC that in MB calyx, homotypic PNs often cluster tightly (Figures
inputs turned out to be olfactory PNs, which are divided into 5 and S7). The PN cluster at the center of the concentrically ar-
genetically identified subtypes. Each PN subtype has a ste- ranged arbors shown in Figure 4C was also tighter in the EM
reotyped and well-described morphology that spans the fly data than in LM data pooled across multiple animals (Figure 5A,
brain, and which is reproducible across individuals with a pre- bottom row). Quantification of the average distance between
cision of
5 mm (Costa et al., 2016; Jefferis et al., 2007; Lin homotypic PNs revealed that intra-animal arbors are signifi-
et al., 2007), allowing for quantitative assessment of the EM- cantly more clustered than arbors from multi-animal LM data
based reconstructions. Morphological reconstructions of (Student’s t test, p < 1.3 3 109) (Figures 5B and 5C). A similar
LM-imaged PNs from most subtypes are available in online result was obtained based on NBLAST score differences (Stu-
databases, and have been aligned to a small number of LM- dent’s t test, p < 2.2 3 1012) (Figures S7B and S7C). The tight
imaged template brains (Chiang et al., 2011; Manton et al., clustering of EM-traced PNs suggests developmental co-
2014; Milyaev et al., 2012). Alignment of an LM-imaged tem- fasciculation of homotypic inputs to the MB calyx, and may
plate brain (Aso et al., 2014) to the EM volume (STAR bias the KC sampling of olfactory input, in contrast to a purely
Methods) allowed LM-imaged neurons from a volume previ- random model based on data pooled across multiple individ-
ously aligned to the template brain to be overlaid on the EM uals (Caron et al., 2013; see the Discussion). These results
dataset (Figures 3A–3D). indicate that the whole-brain EM dataset supports accurate
Quantitative comparisons of EM- and LM-reconstructed PNs and efficient tracing of PNs, a representative brain-spanning
were made using NBLAST (Costa et al., 2016), which measures cell type.
similarity between neuronal arbors. Similarity scores between
EM- and LM-reconstructed PNs of the same subtype were A Previously Unknown Cell Type, MB-CP2, Provides
high. In nearly all cases, the top NBLAST hit agreed with expert Input to KC Claws
identification of PN subtype (Table S1). EM-to-LM similarity To better characterize MB-CP2, we traced the entire backbone
scores were comparable to LM-to-LM scores. For example, of MB-CP2, which proved to be extensive throughout the ipsilat-
the six VM2 PNs in an LM morphological database (FlyCircuit) eral (right) hemisphere. A second MB-CP2 neuron was located in
(Chiang et al., 2011), when compared with one another, resulted the left hemisphere. Reconstruction of its complete backbone

734 Cell 174, 730–743, July 26, 2018

Figure 3. Validation of Tracing by EM-LM Registration and NBLAST-Based Geometry Matching
(A) An oblique cut plane through the EM volume, selected to reveal the projection from the AL to the MB calyx and LH via the mALT. The AL, mALT, MB calyx, and
LH are false colored to show compartment boundaries.
(B) The LM template brain is labeled with nc82 (magenta), a synapse-specific antibody commonly used to reveal neuropil compartment boundaries (Wagh et al.,
2006). After alignment to the EM volume, the same cut plane reveals corresponding neuropil compartments in both (A) and here. Nc82 labeling is absent in the
mALT, a largely synapse-free PN projection tract.
(C) A subset of LM-imaged PNs labeled with random fluorophore combinations (Y. Aso, personal communication) using MultiColor FlpOut (Nern et al., 2015) were
registered to the template brain, and the transformation defined in (B) was used to project the PNs into the coordinate space of the EM volume. The cut plane used
in (A) reveals the PN dendrites in the AL, their axonal projections in the mALT, and their axonal arborizations in the MB calyx and LH.
(D) Overlaid data from the EM dataset (A), the template brain (B), and the LM-imaged PNs (C) show good co-registration between the respective whole-brain
image volumes.
(E) An EM-reconstructed VM2 PN (black) is projected to a template brain (gray surface mesh) using the inverse of the transformation previously defined in (B) to
align the LM template brain to the EM dataset.
(F) An NBLAST search of the FlyCircuit database for matches to the EM-reconstructed VM2 PN (black) returned an LM-reconstructed VM2 PN (red) as the top hit.
(G) An overlay of the EM- and LM-reconstructed VM2 PNs demonstrates high qualitative similarity.
Axis labels as in Figure 1. Scale bars,
100 (A–D) and
50 mm (E–G).

revealed symmetrical arborization in the equivalent contralateral CP2 receives input from g KCs (Figure 6B) and gd KCs, a sub-
neuropil compartments, which is typical of most cell types in the type originating in the ventral accessory calyx known to receive
fly brain (Aso et al., 2014; Chiang et al., 2011; Jenett et al., 2012; visual inputs (Yagi et al., 2016; data not shown). Many areas
Ohyama et al., 2015). This confirmed MB-CP2 as a bona fide where MB-CP2 is purely postsynaptic are innervated by multi-
morphological cell type (Figure 6; Video S5). modal sensory and motor neurons (e.g., SMP and PLP; Hsu
MB-CP2 boutons are presynaptic to all five known olfactory and Bhandawat, 2016; Namiki et al., 2017). Therefore, MB-CP2
KC subtypes (g, abc, abs, a0 b0 m, a0 b0 ap) in MB main calyx (Aso neurons likely relay multimodal, non-olfactory input to KCs,
et al., 2014). In the dorsal accessory calyx (dAC), MB-CP2 bou- and provide recurrent feedback from KC axons in the MB pedun-
tons also provide input to abp KCs (data not shown). In all brain culus to KC dendrites in the MB main calyx, adding to the set of
regions where MB-CP2 makes output synapses, its neurites also known recurrent circuits in the MB (Aso et al., 2014; Owald and
receive input synapses (Figure 6F). In the MB pedunculus, MB- Waddell, 2015).

Cell 174, 730–743, July 26, 2018 735

 Figure 4. Survey of Olfactory PNs Providing
Driving Input to Microglomeruli in the MB
Calyx Agrees with LM Data
(A) EM-reconstructed uniglomerular olfactory PNs
in the right hemisphere recapitulate known olfac-
tory pathways (summarized in Figure 1B).
(B) A frontal view of EM-reconstructed PNs (top) and
glomerular surface models (bottom) in AL shows
agreement with previous glomeruli reconstructions
(Couto et al., 2005; Grabe et al., 2015).
(C) A frontal-dorsal view of EM-reconstructed
boutons for three groups of PNs in MB calyx re-
veals concentric organization, consistent with LM
data (Tanaka et al., 2004). Bouton skeletons (top)
were used to generate Gaussian-smoothed bou-
ton volumes (bottom; STAR Methods) for each of
the three groups. The PN groups are DM1, VA4,
VC1, and VM2 (green); DL1 and VA6 (blue); and
DA1, DC3, and VA1d (red).
(D) Unsupervised clustering based on morpho-
logical similarity (NBLAST score) produces a
dendrogram in which olfactory PNs are grouped
by glomerular subtype.
(E) Comparison of the number of reconstructed
PNs per glomerulus from EM and LM data (Grabe
et al., 2016).
Colors in (A), (B), and (D) like in Couto et al. (2005).
PNs receive input from olfactory receptor neurons
(ORNs). The dendrite of each ORN innervates an
antennal protuberance called a sensillum. Each
PN is colored by the class of sensillum its input
ORNs innervate. Error bars indicate SD.
Axis and anatomical labels are the same as those
used in Figure 1; lALT, lateral antennal lobe tract;
LB, large basiconic; TB, thin basiconic; SB,
small basiconic; T1, T2, T3, trichoid sensilla; PB,
maxillary palp basiconic; AC, antennal coelo-
conic; AI, antennal intermediate.
Scale bars,
10 mm (A–C). See also Table S1.

Identification of Cell Types Postsynaptic to KCs in the aptic targets (Figure 7; Table S2). Intriguingly, a0 b0 KCs (which are
MB Calyx dispensable for memory retrieval) (Krashes et al., 2007) were
In addition to their canonical input from PN boutons, KC den- even more selective, only synapsing onto APL and other KCs.
drites make synaptic outputs to unknown targets in the MB ca- This selectivity could arise because either the neglected cell
lyx, forming a poorly understood local microcircuit (Butcher types or a0 b0 KCs arborize sparsely in the MB calyx; alternatively,
et al., 2012; Christiansen et al., 2011). To validate use of the they could be specifically neglected despite extensive arbor
whole-brain dataset for anterograde mapping of synaptic con- overlap and high availability of potential synapses (Stepanyants
nectivity, we reconstructed and identified the postsynaptic tar- and Chklovskii, 2005). To distinguish between these alternatives,
gets of 15 KCs in the MB calyx. All dendritic presynaptic release future work should sample more KCs and reconstruct to comple-
sites were annotated on three KCs from each of the five subtypes tion available postsynaptic cell types.
in MB main calyx. Consistent with immunohistochemical data
(Christiansen et al., 2011), most (82%) presynaptic release sites DISCUSSION
arise from abc-, abs-, or g KCs, and 87% of all release sites are
distributed along KC dendrites outside the claws (Figure 7; Here, we contribute a complete EM volume of an adult female
Table S2). Fourteen percent of fine postsynaptic neurites were Drosophila brain for free use by the research community. Valida-
too difficult to trace to a parent backbone with high confidence. tion reconstructions demonstrated the reliable tracing of local
All KC presynaptic release sites divergently target multiple post- and long-range connectivity, revealing both known and new ele-
synaptic processes, which in many cases arise from identified ments in the MB circuitry underlying associative learning and
MB intrinsic and extrinsic neurons and form local recurrent mi- recall. We conclude this volume is suitable for tracing brain-
crocircuits (Figures 7E and 7F). spanning neuronal circuits at synaptic resolution. Reconstruc-
We found that KCs preferentially target a subset of potential tions of the neurons reported here are included in a preconfig-
partner cell types in the MB calyx, with only four of the 14 avail- ured, downloadable tracing environment to serve as entry points
able cell types (STAR Methods) making up 75% of the postsyn- for further investigation.

736 Cell 174, 730–743, July 26, 2018

 Figure 5. PN Arbors in MB Calyx Cluster
More Tightly than Previously Seen with LM
across Individuals
(A) Comparison of EM- versus LM-reconstructed
PNs. EM-reconstructed PNs are shown against a
surface model of MB calyx (gray) in the left col-
umn. Calyx arbors for EM- and LM-reconstructed
PNs are shown in the middle and right columns,
respectively. Data for LM-reconstructed PNs
(right column) are from the FlyCircuit database
(Chiang et al., 2011), as registered to a common
template brain (Costa et al., 2016; see also STAR
Methods).
(B) Pairwise distances between homotypic PN
collaterals in the MB calyx. Each data point rep-
resents the distance between one pair of EM-
(red) or LM-reconstructed (blue) PNs from the
same subtype. Data points are bucketed ac-
cording to PN subtype; subtypes are ordered on
the x axis by how much more clustered EM-re-
constructed PNs are than LM-reconstructed PNs
(STAR Methods).
(C) Histogram of all data points in (B). The mean of
pairwise distances for all EM-reconstructed PN
subtypes was significantly lower than that for all
LM-reconstructed PN subtypes (3.40 ± 1.53 mm
vs. 5.49 ± 2.73 mm, respectively; Student’s t test,
p < 1.3 3 109).
Axis labels as in Figure 1. Scale bars,
20 (A, left
column) and
10 mm (A, middle and right col-
umns). See also Figure S7.

The alignment of template LM brains to

the whole-brain EM volume allows EM-
reconstructed neurons to be searched
for in large on-line databases of morpho-
logically stereotyped cell types. Genetic
driver lines for cell types (Aso et al.,
2014; Grabe et al., 2015; Jenett et al.,
2012; Kvon et al., 2014) discovered to
participate in a given EM-reconstructed
circuit can therefore be readily identified.
This will facilitate testing of physiological
Generation of the whole-brain dataset was enabled by new and behavioral hypotheses generated by newly mapped EM
high-throughput hardware and scalable software for acquiring wiring diagrams.
and aligning large-scale EM data. Although improvements to Our finding that olfactory PN arbors in the MB calyx cluster
TEMCA throughput are readily achievable, alternative volume much more tightly than predicted by LM data pooled across an-
EM imaging approaches such as multibeam-scanning EM and imals is consistent with observations for pairs of neurons from
FIB-scanning EM also show great promise (Eberle et al., 2015; one glomerulus by Kazama and Wilson (2009). This wide-scale
Xu et al., 2017). In larger animals, brain-spanning connectomics clustering may generate non-random PN-to-KC connectivity,
may be achieved by low resolution EM imaging, followed by high since a KC dendritic arbor centered on a cluster of boutons
resolution reimaging of synaptic connectivity in selected sub- arising from a given PN subtype may oversample that subtype.
volumes (Hildebrand et al., 2017). Although the alignment quality The above-chance rate of multi-claw input to KCs from a tightly
shown here sufficed for manual tracing, further improvements, bundled subset of three PN subtypes observed in adult
such as better fine-scale registration near artifacts, may be Drosophila (Gruntman and Turner, 2013) is consistent with this
needed to fully leverage emerging methods for automatic possibility. Our EM-based reconstruction demonstrated tight
segmentation (Arganda-Carreras et al., 2017; Beier et al., clustering of these same subtypes (Figure 5A). However, avail-
2017; Januszewski et al., 2016). Early segmentation results on able EM data in larva and LM data in adult Drosophila (Caron
subsets of this whole-brain dataset are nonetheless promising et al., 2013; Eichler et al., 2017) indicate random sampling of ol-
(Funke et al., 2016). factory input, which has become an assumption in many models

Cell 174, 730–743, July 26, 2018 737

 Figure 6. MB-CP2, a New Cell Type
Providing Microglomerular Input to KC
Claws
(A) Reconstruction of the pair of MB-CP2 neurons
with surface meshes of the whole brain (gray) and
MB (green).
(B–E) Synaptic connectivity between MB-CP2 and
KCs in MB pedunculus and main calyx. Since
many KC claws ensheath a given MB-CP2 bouton
and KCs are traced only sparsely in this study,
most postsynaptic profiles are untraced (STAR
Methods). Arrowheads, presynaptic release sites.
(B) A representative MB-CP2 neurite (orange) in
MB pedunculus, postsynaptic to a KC axon
(green).
(C) A representative MB-CP2 neurite (orange) in
left hemisphere MB pedunculus, with comparable
synaptic arrangements to (B). KCs were not traced
in the left hemisphere, so cell identity is putative.
However, most bundled neurites parallel to the MB
pedunculus long axis arise from KCs (Leitch and
Laurent, 1996; Schürmann, 2016; data not shown).
(D) A cross-section through a representative MB-
CP2 bouton (orange) in MB main calyx at the
center of a canonical microglomerulus. Several
postsynaptic profiles arise from KC claws (green).
(E) A representative MB-CP2 bouton in left hemi-
sphere MB main calyx, with comparable synaptic
arrangements to (D). KCs were not traced in left
hemisphere, so cell identity is putative; however,
most postsynaptic elements at MB calyx micro-
glomeruli arise from KCs (Butcher et al., 2012; data
not shown).
(F) Summary schematic of MB-CP2 input and
output brain regions with synapse counts discov-
ered following partial reconstruction (STAR
Methods). In six brain regions, MB-CP2 dendrites
are purely postsynaptic. In four other regions, MB-
CP2 neurites are both pre- and postsynaptic.
Axis and anatomical labels as in Figure 1; Ped, MB
pedunculus; dAC, dorsal accessory calyx; ATL,
antler; SC, superior clamp; PLP, posterior lateral
protocerebrum; SMP, superior medial proto-
cerebrum; SIP, superior intermediate proto-
cerebrum; SLP, superior lateral protocerebrum.
Scale bars, 100 mm (A, dorsal view), 500 nm
(B and C), and 2 mm (D and E). See also Video S5.

of odor representation (Dasgupta et al., 2017; Litwin-Kumar Schürmann, 2016; Takemura et al., 2017a; Tobin et al., 2017;
et al., 2017; Stevens, 2015; but see Koulakov et al., 2011; Pehle- data not shown), and may turn out to be pervasive in the fly brain.
van et al., 2017). Although in the larval dataset KCs were Mixed I/O arrangements can be physiologically and behaviorally
comprehensively reconstructed at the EM level, adults have important; for example, in the antennal lobe, inhibitory inputs into

20x more KCs as well as additional KC subtypes (Aso et al., olfactory receptor neuron axons are known to mediate gain
2014; Eichler et al., 2017; Lin et al., 2007), and may therefore control and temporal sharpening (Olsen and Wilson, 2008; Rac-
have a different network structure. The adult LM data were cuglia et al., 2016; Root et al., 2008). The whole-brain EM dataset
pooled across many animals and, therefore, could be supports the identification of the cell types present at mixed I/O
confounded by inter-individual variability. More comprehensive synapses, an otherwise challenging class of information to
mapping of intra-animal PN-to-KC connectivity in the whole- obtain.
brain dataset will help resolve whether or not olfactory input to Drosophila exhibits a wide range of complex sensory- and
KCs in adult Drosophila is random. memory-guided behaviors. The algorithms underlying behavior
Kenyon cell dendrites in the MB calyx and MB-CP2 axons are implemented by neuronal circuits, and neuronal circuits are
contain a mixture of synaptic input and output. Mixed input/ defined, in large part (though not entirely; Bargmann and Marder,
output (I/O) neurites have previously been found in multiple re- 2013), by the synaptic connectivity between neurons. Therefore,
gions (Meinertzhagen and O’Neil, 1991; Rybak et al., 2016; connectome maps are necessary to analyze neuronal circuits

738 Cell 174, 730–743, July 26, 2018

 Figure 7. KC Presynaptic Release Sites in
the MB Calyx Mostly Target a Small Subset
of Available Partners
(A–D) Morphological comparison of LM-imaged
(left panels) and EM-reconstructed (right panels)
neurons of the same class. LM data from Aso et al.
(2014). Neurite densities are lower in the EM re-
constructions, since these cells were traced to
classification, not completion (STAR Methods).
Spheres in the EM-reconstructions indicate the
location of cell bodies.
(A) abc- (green), abs- (yellow), and g- (cyan) KCs.
One representative EM-reconstructed KC from
each class is shown (right panel); their morphol-
ogies and trajectories match those of the LM-
imaged KCs of the same class. Small red square:
location in MB calyx of a g KC sub-arbor shown in
the inset (large red square). Inset: representative
location of a presynaptic release site (black
arrowhead), on a twig arising from the KC back-
bone, outside the claw sub-arbor. A micrograph of
this site is shown in (E).
(B) APL, a wide-field inhibitory neuron that in-
nervates the entire MB and sparsifies KC activity
(Lin et al., 2014; Liu and Davis, 2009).
(C) MB-CP1, a MB output neuron (MBON) with a
dendritic arbor innervating the MB calyx and pe-
dunculus (Tanaka et al., 2008).
(D) MB-C1, a putative inhibitory interneuron that innervates the MB calyx and LH (Tanaka et al., 2008). Two MB-C1 neurons were found in the EM-based survey of
KC postsynaptic targets, in contrast to the single neuron reported by Tanaka et al. (2008). Small red square: location in MB calyx of a g KC sub-arbor shown in the
inset (large red square). Inset: representative location of a presynaptic release site (black arrowhead), on a twig arising from the KC backbone, outside the claw
sub-arbor. A micrograph of this site is shown in (F). For clarity, this KC is not shown in (D, right panel).
(E and F) Micrographs at the synapse locations shown in (A) and (D) insets. Arrowheads, selected presynaptic release sites.
(E) The g KC in (A, inset) and two other g KCs (light and dark purple) that are presynaptic to APL (green), MB-CP1 (red), and each other at the same synaptic cleft.
The APL is also presynaptic to a PN (brown).
(F) The g KC from (D, inset) is presynaptic to MB-C1 (pink), APL (green), and several additional unidentified partners. The APL postsynaptic density is two sections
away (not visible in this section plane).
Scale bars,
25 (A–D) and 1 mm (E and F). See also Table S2.

at the implementation level (Marr, 1982) and could aid in the B Analysis of Neuronal Geometry
inference of underlying algorithms as well. The dataset we share d DATA AND SOFTWARE AVAILABILITY
here should help establish a structural scaffold for future models d ADDITIONAL RESOURCES
of circuit function in the fly and enable comparisons of circuit ar-
chitecture across species. SUPPLEMENTAL INFORMATION

STAR+METHODS Supplemental Information includes seven figures, two tables, and five videos
and can be found with this article online at https://doi.org/10.1016/j.cell.
2018.06.019.
Detailed methods are provided in the online version of this paper
A video abstract is available at https://doi.org/10.1016/j.cell.2018.06.
and include the following: 019#mmc8.

d KEY RESOURCES TABLE

d CONTACT FOR REAGENT AND RESOURCE SHARING ACKNOWLEDGMENTS
d EXPERIMENTAL MODEL AND SUBJECT DETAILS
We thank Jeff Jordan, Bruce Bowers, and Jon Arnold (Janelia Instrumentation
d METHOD DETAILS Design & Fabrication team); Dana Dunkelberger (Grant Scientific); and Jim
B Sample Preparation Mancuso (AMT Imaging) for hardware design & engineering services. We
B Sample Supports, Ultramicrotomy, and Post-staining also thank Brett Mensh, Scott Waddell, and Vivek Jayaraman for critical im-
B Electron Microscopy provements to the manuscript; Wei-Ping Li (Janelia EM Facility) for coating
B Volume Reconstruction of sample grids and poststaining of serial sections; Najla Masoodpanah, Jo-
B Neuron Tracing seph Hsu, Benjamin Gorko, Emily Moore, Arynne Boyes, Jacob Ratliff,
Adam John, Bailey Harrison, Adeleso Adesina, and Claire Managan for neuron
B Neuronal Informatics
tracing; Noah Nelson for graph visualization tools; Arlo Sheridan and Ben
d QUANTIFICATION AND STATISTICAL ANALYSIS Gorko for 3D animations; Philipp Ranft and Gaia Tavosanis for literature review
B Comparison of SNR between Volume EM Datasets of cell types innervating MB calyx; Adam Heath, Marta Costa, and Philip Schle-
B Quantification of Artifacts gel for glomerulus meshes, NBLAST analyses, and PN classification; David

Cell 174, 730–743, July 26, 2018 739

Peale for optimization of carbon coating and film casting protocol; Jon Arnold Briggman, K.L., Helmstaedter, M., and Denk, W. (2011). Wiring specificity in
for technical figures; Tom Kazimiers and Andrew Champion for CATMAID the direction-selectivity circuit of the retina. Nature 471, 183–188.
development; Albert Cardona for CATMAID workflow advice, CATMAID devel- Burke, C.J., Huetteroth, W., Owald, D., Perisse, E., Krashes, M.J., Das, G.,
opment, and helpful discussions; Casey Schneider-Mizell for analysis code Gohl, D., Silies, M., Certel, S., and Waddell, S. (2012). Layered reward signal-
and helpful discussions; Tom Dolafi and Cristian Goina (Janelia Software En- ling through octopamine and dopamine in Drosophila. Nature 492, 433–437.
gineering) for volume reconstruction support; Rob Lines, Ken Carlisle, and
Goran Ceric (Janelia Scientific Computing Systems) for data and cluster man- Busch, S., Selcho, M., Ito, K., and Tanimoto, H. (2009). A map of octopaminer-
agement; Yoshinori Aso for pre-publication image volumes of MultiColor gic neurons in the Drosophila brain. J. Comp. Neurol. 513, 643–667.
FlpOut PN lines; and Eyal Gruntman for helpful discussions. We thank the Ho- Butcher, N.J., Friedrich, A.B., Lu, Z., Tanimoto, H., and Meinertzhagen, I.A.
ward Hughes Medical Institute for funding all aspects of this work. We thank (2012). Different classes of input and output neurons reveal new features in mi-
Wellcome Trust (collaborative award 203261/Z/16/Z) for funding the additional croglomeruli of the adult Drosophila mushroom body calyx. J. Comp. Neurol.
tracing effort. 520, 2185–2201.
Cardona, A., Saalfeld, S., Schindelin, J., Arganda-Carreras, I., Preibisch, S.,
AUTHOR CONTRIBUTIONS Longair, M., Tomancak, P., Hartenstein, V., and Douglas, R.J. (2012). TrakEM2
software for neural circuit reconstruction. PLoS ONE 7, e38011.
Conceptualization, D.D.B., R.D.F., J.S.L., and Z.Z.; Methodology, D.D.B., Caron, S.J., Ruta, V., Abbott, L.F., and Axel, R. (2013). Random convergence
R.D.F., J.S.L., D.M., J.P., C.G.R., O.T., and Z.Z.; Software, J.A.B., P.H., of olfactory inputs in the Drosophila mushroom body. Nature 497, 113–117.
G.S.X.E.J., B.K., K.K., M.K., M.N., D.M., E.P., E.T.T., O.T., and S.S.; Validation,
Chen, C.C., Wu, J.K., Lin, H.W., Pai, T.P., Fu, T.F., Wu, C.L., Tully, T., and
J.S.L. and Z.Z.; Formal Analysis, D.D.B., G.S.X.E.J., J.S.L., M.N., and Z.Z.;
Chiang, A.S. (2012). Visualizing long-term memory formation in two neurons
Investigation, I.J.A., D.D.B., S.A.C.-S., C.B.F., L.K., J.S.L., C.G.R., N.S., and
of the Drosophila brain. Science 335, 678–685.
Z.Z. ; Data Curation, J.A.B., R.D.F., J.S.L., D.M., M.N., E.P., C.G.R., S.S.,
and Z.Z.; Writing—Original Draft, D.D.B., J.S.L., M.N., E.P., C.G.R., and Chiang, A.S., Lin, C.Y., Chuang, C.C., Chang, H.M., Hsieh, C.H., Yeh, C.W.,
Z.Z.; Writing—Review and Editing, I.J.A., D.D.B., J.A.B., S.A.C.-S., C.B.F., Shih, C.T., Wu, J.J., Wang, G.T., Chen, Y.C., et al. (2011). Three-dimensional
R.D.F., G.S.X.E.J., B.K., K.K., J.S.L., M.N., E.P., N.S., S.S., and Z.Z.; Visualiza- reconstruction of brain-wide wiring networks in Drosophila at single-cell reso-
tion, D.D.B., S.A.C.-S., J.S.L., and Z.Z.; Supervision, D.D.B. and J.S.L.; Project lution. Curr. Biol. 21, 1–11.
Administration, D.D.B.; Funding Acquisition, D.D.B. Chou, Y.H., Spletter, M.L., Yaksi, E., Leong, J.C., Wilson, R.I., and Luo, L.
(2010). Diversity and wiring variability of olfactory local interneurons in the
DECLARATION OF INTERESTS Drosophila antennal lobe. Nat. Neurosci. 13, 439–449.
Christiansen, F., Zube, C., Andlauer, T.F., Wichmann, C., Fouquet, W., Owald,
If the patents described in this manuscript were to be licensed, D.D.B. could D., Mertel, S., Leiss, F., Tavosanis, G., Luna, A.J., et al. (2011). Presynapses in
benefit financially. Kenyon cell dendrites in the mushroom body calyx of Drosophila. J. Neurosci.
31, 9696–9707.
Received: May 22, 2017
Costa, M., Manton, J.D., Ostrovsky, A.D., Prohaska, S., and Jefferis, G.S.
Revised: February 28, 2018
(2016). NBLAST: rapid, sensitive comparison of neuronal structure and con-
Accepted: June 10, 2018
struction of neuron family databases. Neuron 91, 293–311.
Published: July 19, 2018
Couto, A., Alenius, M., and Dickson, B.J. (2005). Molecular, anatomical, and
functional organization of the Drosophila olfactory system. Curr. Biol. 15,
REFERENCES
1535–1547.
Arganda-Carreras, I., Kaynig, V., Rueden, C., Eliceiri, K.W., Schindelin, J., Car- Dasgupta, S., Stevens, C.F., and Navlakha, S. (2017). A neural algorithm for a
dona, A., and Sebastian Seung, H. (2017). Trainable Weka segmentation: a fundamental computing problem. Science 358, 793–796.
machine learning tool for microscopy pixel classification. Bioinformatics 33, de Haro, M., Al-Ramahi, I., Benito-Sipos, J., López-Arias, B., Dorado, B.,
2424–2426. Veenstra, J.A., and Herrero, P. (2010). Detailed analysis of leucokinin-express-
Ashburner, M., Golic, K.G., and Hawley, R.S. (2005). Drosophila: A Laboratory ing neurons and their candidate functions in the Drosophila nervous system.
Handbook (Cold Spring Harbor Laboratory Press). Cell Tissue Res. 339, 321–336.
Aso, Y., Hattori, D., Yu, Y., Johnston, R.M., Iyer, N.A., Ngo, T.T., Dionne, H., Dickinson, M.H., and Muijres, F.T. (2016). The aerodynamics and control of
Abbott, L.F., Axel, R., Tanimoto, H., and Rubin, G.M. (2014). The neuronal ar- free flight manoeuvres in Drosophila. Philos. Trans. R. Soc. Lond. B Biol.
chitecture of the mushroom body provides a logic for associative learning. eL- Sci. 371, 20150388.
ife 3, e04577. Ding, H., Smith, R.G., Poleg-Polsky, A., Diamond, J.S., and Briggman, K.L.
Bargmann, C.I., and Marder, E. (2013). From the connectome to brain function. (2016). Species-specific wiring for direction selectivity in the mammalian
Nat. Methods 10, 483–490. retina. Nature 535, 105–110.
Bay, H., Ess, A., Tuytelaars, T., and Van Gool, L. (2008). Speeded-up robust Eberle, A.L., Mikula, S., Schalek, R., Lichtman, J., Knothe Tate, M.L., and Zei-
features (SURF). Comput. Vis. Image Underst. 110, 346–359. dler, D. (2015). High-resolution, high-throughput imaging with a multibeam
Beier, T., Pape, C., Rahaman, N., Prange, T., Berg, S., Bock, D.D., Cardona, A., scanning electron microscope. J. Microsc. 259, 114–120.
Knott, G.W., Plaza, S.M., Scheffer, L.K., et al. (2017). Multicut brings automated Eichler, K., Li, F., Litwin-Kumar, A., Park, Y., Andrade, I., Schneider-Mizell,
neurite segmentation closer to human performance. Nat. Methods 14, 101–102. C.M., Saumweber, T., Huser, A., Eschbach, C., Gerber, B., et al. (2017). The
Bock, D.D., Lee, W.C., Kerlin, A.M., Andermann, M.L., Hood, G., Wetzel, A.W., complete connectome of a learning and memory centre in an insect brain. Na-
Yurgenson, S., Soucy, E.R., Kim, H.S., and Reid, R.C. (2011). Network anat- ture 548, 175–182.
omy and in vivo physiology of visual cortical neurons. Nature 471, 177–182. Enjin, A., Zaharieva, E.E., Frank, D.D., Mansourian, S., Suh, G.S., Gallio, M.,
Bogovic, J.A., Hanslovsky, P., Wong, A., and Saalfeld, S. (2016). Robust regis- and Stensmyr, M.C. (2016). Humidity sensing in Drosophila. Curr. Biol. 26,
tration of calcium images by learned contrast synthesis. Biomedical Imaging 1352–1358.
(ISBI), 2016 IEEE 13th International Symposium on IEEE, pp. 1123–1126. Erdogmus, D., Larsson, E.G., Yan, R., Principe, J.C., and Fitzsimmons, J.R.
Briggman, K.L., and Bock, D.D. (2012). Volume electron microscopy for (2004). Measuring the signal-to-noise ratio in magnetic resonance imaging: a
neuronal circuit reconstruction. Curr. Opin. Neurobiol. 22, 154–161. caveat. Signal Processing 84, 1035–1040.

740 Cell 174, 730–743, July 26, 2018

Farris, S.M. (2011). Are mushroom bodies cerebellum-like structures? Jenett, A., Rubin, G.M., Ngo, T.T., Shepherd, D., Murphy, C., Dionne, H.,
Arthropod Struct. Dev. 40, 368–379. Pfeiffer, B.D., Cavallaro, A., Hall, D., Jeter, J., et al. (2012). A GAL4-driver
Fischler, M.A., and Bolles, R.C. (1981). Random sample consensus: a para- line resource for Drosophila neurobiology. Cell Rep. 2, 991–1001.
digm for model fitting with applications to image analysis and automated Jiang, H., and Sun, S.X. (2013). Cellular pressure and volume regulation and
cartography. Commun. ACM 24, 381–395. implications for cell mechanics. Biophys. J. 105, 609–619.
Frank, D.D., Jouandet, G.C., Kearney, P.J., Macpherson, L.J., and Gallio, M. Jovanic, T., Schneider-Mizell, C.M., Shao, M., Masson, J.B., Denisov, G.,
(2015). Temperature representation in the Drosophila brain. Nature 519, Fetter, R.D., Mensh, B.D., Truman, J.W., Cardona, A., and Zlatic, M. (2016).
358–361. Competitive disinhibition mediates behavioral choice and sequences in
Drosophila. Cell 167, 858–870.e19.
Funke, J., Saalfeld, S., Bock, D.D., Turaga, S.C., and Perlman, E. (2016).
MICCAI Challenge on Circuit Reconstruction from Electron Microscopy Im- Kasthuri, N., Hayworth, K.J., Berger, D.R., Schalek, R.L., Conchello, J.A.,
ages (CREMI). Knowles-Barley, S., Lee, D., Vázquez-Reina, A., Kaynig, V., Jones, T.R.,
et al. (2015). Saturated reconstruction of a volume of neocortex. Cell 162,
Grabe, V., Strutz, A., Baschwitz, A., Hansson, B.S., and Sachse, S. (2015). Dig-
648–661.
ital in vivo 3D atlas of the antennal lobe of Drosophila melanogaster. J. Comp.
Neurol. 523, 530–544. Kaynig, V., Fischer, B., Müller, E., and Buhmann, J.M. (2010). Fully automatic
stitching and distortion correction of transmission electron microscope im-
Grabe, V., Baschwitz, A., Dweck, H.K.M., Lavista-Llanos, S., Hansson, B.S.,
ages. J. Struct. Biol. 171, 163–173.
and Sachse, S. (2016). Elucidating the neuronal architecture of olfactory
glomeruli in the Drosophila antennal lobe. Cell Rep. 16, 3401–3413. Kazama, H., and Wilson, R.I. (2009). Origins of correlated activity in an olfac-
tory circuit. Nat. Neurosci. 12, 1136–1144.
Gruntman, E., and Turner, G.C. (2013). Integration of the olfactory code across
dendritic claws of single mushroom body neurons. Nat. Neurosci. 16, Kazhdan, M., Surendran, D., and Hoppe, H. (2010). Distributed gradient-
1821–1829. domain processing of planar and spherical images. ACM Trans. Graph. 29, 14.

Gruntman, E., Romani, S., and Reiser, M.B. (2018). Simple integration of fast Kazhdan, M., Lillaney, K., Roncal, W., Bock, D., Vogelstein, J., and Burns, R.
excitation and offset, delayed inhibition computes directional selectivity in (2015). Gradient-domain fusion for color correction in large EM image stacks.
Drosophila. Nat. Neurosci. 21, 250–257. arXiv, arXiv:150602079. https://arxiv.org/abs/1506.02079.
Keene, A.C., and Waddell, S. (2007). Drosophila olfactory memory: single
Guven-Ozkan, T., and Davis, R.L. (2014). Functional neuroanatomy of
genes to complex neural circuits. Nat. Rev. Neurosci. 8, 341–354.
Drosophila olfactory memory formation. Learn. Mem. 21, 519–526.
Kopek, B.G., Paez-Segala, M.G., Shtengel, G., Sochacki, K.A., Sun, M.G.,
Hampel, S., Franconville, R., Simpson, J.H., and Seeds, A.M. (2015). A neural
Wang, Y., Xu, C.S., van Engelenburg, S.B., Taraska, J.W., Looger, L.L., and
command circuit for grooming movement control. eLife 4, e08758.
Hess, H.F. (2017). Diverse protocols for correlative super-resolution fluores-
Hanslovsky, P., Bogovic, J.A., and Saalfeld, S. (2017). Image-based correction cence imaging and electron microscopy of chemically fixed samples. Nat. Pro-
of continuous and discontinuous non-planar axial distortion in serial section toc. 12, 916–946.
microscopy. Bioinformatics 33, 1379–1386.
Koulakov, A.A., Kolterman, B.E., Enikolopov, A.G., and Rinberg, D. (2011). In
Harris, K.M., Perry, E., Bourne, J., Feinberg, M., Ostroff, L., and Hurlburt, J. search of the structure of human olfactory space. Front. Syst. Neurosci. 5, 65.
(2006). Uniform serial sectioning for transmission electron microscopy.
Krashes, M.J., Keene, A.C., Leung, B., Armstrong, J.D., and Waddell, S.
J. Neurosci. 26, 12101–12103.
(2007). Sequential use of mushroom body neuron subsets during Drosophila
Helmstaedter, M., Briggman, K.L., and Denk, W. (2008). 3D structural imaging odor memory processing. Neuron 53, 103–115.
of the brain with photons and electrons. Curr. Opin. Neurobiol. 18, 633–641.
Kvon, E.Z., Kazmar, T., Stampfel, G., Yáñez-Cuna, J.O., Pagani, M., Schern-
Helmstaedter, M., Briggman, K.L., and Denk, W. (2011). High-accuracy neurite huber, K., Dickson, B.J., and Stark, A. (2014). Genome-scale functional char-
reconstruction for high-throughput neuroanatomy. Nat. Neurosci. 14, acterization of Drosophila developmental enhancers in vivo. Nature
1081–1088. 512, 91–95.
Hildebrand, D.G.C., Cicconet, M., Torres, R.M., Choi, W., Quan, T.M., Moon, Lambert, T.J., and Waters, J.C. (2017). Navigating challenges in the applica-
J., Wetzel, A.W., Scott Champion, A., Graham, B.J., Randlett, O., et al. tion of superresolution microscopy. J. Cell Biol. 216, 53–63.
(2017). Whole-brain serial-section electron microscopy in larval zebrafish. Na- Lefman, J., Morrison, R., and Subramaniam, S. (2007). Automated 100-posi-
ture 545, 345–349. tion specimen loader and image acquisition system for transmission electron
Hiraoka, J.I. (1972). A holder for mass treatment of grids, adapted especially to microscopy. J. Struct. Biol. 158, 318–326.
electron staining and autoradiography. Stain Technol. 47, 297–301. Leitch, B., and Laurent, G. (1996). GABAergic synapses in the antennal lobe
Hsu, C.T., and Bhandawat, V. (2016). Organization of descending neurons in and mushroom body of the locust olfactory system. J. Comp. Neurol. 372,
Drosophila melanogaster. Sci. Rep. 6, 20259. 487–514.
Ito, K., Shinomiya, K., Ito, M., Armstrong, J.D., Boyan, G., Hartenstein, V., Li, D., Shao, L., Chen, B.C., Zhang, X., Zhang, M., Moses, B., Milkie, D.E.,
Harzsch, S., Heisenberg, M., Homberg, U., Jenett, A., et al.; Insect Brain Beach, J.R., Hammer, J.A., 3rd, Pasham, M., et al. (2015). Advanced imaging.
Name Working Group (2014). A systematic nomenclature for the insect brain. Extended-resolution structured illumination imaging of endocytic and cyto-
Neuron 81, 755–765. skeletal dynamics. Science 349, aab3500.

Januszewski, M., Maitin-Shepard, J., Li, P., Kornfeld, J., Denk, W., and Jain, V. Lichtman, J.W., and Sanes, J.R. (2008). Ome sweet ome: what can the
(2016). Flood-filling networks. arXiv, arXiv:161100421. https://arxiv.org/abs/ genome tell us about the connectome? Curr. Opin. Neurobiol. 18, 346–353.
1611.00421. Lin, H.H., Lai, J.S., Chin, A.L., Chen, Y.C., and Chiang, A.S. (2007). A map of
Jarrell, T.A., Wang, Y., Bloniarz, A.E., Brittin, C.A., Xu, M., Thomson, J.N., Al- olfactory representation in the Drosophila mushroom body. Cell 128,
bertson, D.G., Hall, D.H., and Emmons, S.W. (2012). The connectome of a de- 1205–1217.
cision-making neural network. Science 337, 437–444. Lin, A.C., Bygrave, A.M., de Calignon, A., Lee, T., and Miesenböck, G. (2014).
Jefferis, G.S., Potter, C.J., Chan, A.M., Marin, E.C., Rohlfing, T., Maurer, C.R., Sparse, decorrelated odor coding in the mushroom body enhances learned
Jr., and Luo, L. (2007). Comprehensive maps of Drosophila higher olfactory odor discrimination. Nat. Neurosci. 17, 559–568.
centers: spatially segregated fruit and pheromone representation. Cell 128, Litwin-Kumar, A., Harris, K.D., Axel, R., Sompolinsky, H., and Abbott, L.F.
1187–1203. (2017). Optimal degrees of synaptic connectivity. Neuron 93, 1153–1164.e7.

Cell 174, 730–743, July 26, 2018 741

Liu, X., and Davis, R.L. (2009). The GABAergic anterior paired lateral neuron Raccuglia, D., McCurdy, L.Y., Demir, M., Gorur-Shandilya, S., Kunst, M., Emo-
suppresses and is suppressed by olfactory learning. Nat. Neurosci. 12, 53–59. net, T., and Nitabach, M.N. (2016). Presynaptic GABA receptors mediate tem-
Manton, J.D., Ostrovsky, A.D., Goetz, L., Costa, M., Rohlfing, T., and Jefferis, poral contrast enhancement in Drosophila olfactory sensory neurons and
G.S. (2014). Combining genome-scale Drosophila 3D neuroanatomical data modulate odor-driven behavioral kinetics. eNeuro 3, pii: ENEURO.0080-
by bridging template brains. bioRxiv. https://doi.org/10.1101/006353. 16.2016.

Mao, Z., and Davis, R.L. (2009). Eight different types of dopaminergic neurons Robinson, C., Price, J., Milkie, D., Torrens, O., Perlman, E., Zheng, Z., Fetter,
innervate the Drosophila mushroom body neuropil: anatomical and physiolog- R., and Bock, D. (2016). Automated infrastructure for high-throughput acquisi-
ical heterogeneity. Front. Neural Circuits 3, 5. tion of serial section TEM image volumes. Microsc. Microanal. 22, 1150–1151.

Marr, D. (1982). Vision: A Computational Investigation into the Human Repre- Root, C.M., Masuyama, K., Green, D.S., Enell, L.E., Nässel, D.R., Lee, C.H.,
sentation and Processing of Visual Information (MIT Press). and Wang, J.W. (2008). A presynaptic gain control mechanism fine-tunes ol-
factory behavior. Neuron 59, 311–321.
Masse, N.Y., Turner, G.C., and Jefferis, G.S. (2009). Olfactory information pro-
cessing in Drosophila. Curr. Biol. 19, R700–R713. Roy, B., Singh, A.P., Shetty, C., Chaudhary, V., North, A., Landgraf, M., Vijayr-
aghavan, K., and Rodrigues, V. (2007). Metamorphosis of an identified seroto-
Meinertzhagen, I.A., and O’Neil, S.D. (1991). Synaptic organization of
nergic neuron in the Drosophila olfactory system. Neural Dev. 2, 20.
columnar elements in the lamina of the wild type in Drosophila melanogaster.
J. Comp. Neurol. 305, 232–263. Ryan, K., Lu, Z., and Meinertzhagen, I.A. (2016). The CNS connectome of a
tadpole larva of Ciona intestinalis (L.) highlights sidedness in the brain of a
Milyaev, N., Osumi-Sutherland, D., Reeve, S., Burton, N., Baldock, R.A., and
chordate sibling. eLife 5, e16962.
Armstrong, J.D. (2012). The Virtual Fly Brain browser and query interface. Bio-
informatics 28, 411–415. Rybak, J., Talarico, G., Ruiz, S., Arnold, C., Cantera, R., and Hansson, B.S.
(2016). Synaptic circuitry of identified neurons in the antennal lobe of
Namiki, S., Dickinson, M.H., Wong, A.M., Korff, W., and Card, G.M. (2017). The
Drosophila melanogaster. J. Comp. Neurol. 524, 1920–1956.
functional organization of descending sensory-motor pathways in Drosophila.
eLife 7, 34272. Saalfeld, S., Cardona, A., Hartenstein, V., and Tomancak, P. (2009). CATMAID:
collaborative annotation toolkit for massive amounts of image data. Bioinfor-
Nern, A., Pfeiffer, B.D., and Rubin, G.M. (2015). Optimized tools for multicolor
matics 25, 1984–1986.
stochastic labeling reveal diverse stereotyped cell arrangements in the fly
visual system. Proc. Natl. Acad. Sci. USA 112, E2967–E2976. ák, P. (2010). As-rigid-
Saalfeld, S., Cardona, A., Hartenstein, V., and Tomanc
Ofstad, T.A., Zuker, C.S., and Reiser, M.B. (2011). Visual place learning in as-possible mosaicking and serial section registration of large ssTEM data-
Drosophila melanogaster. Nature 474, 204–207. sets. Bioinformatics 26, i57–i63.

Ohyama, T., Schneider-Mizell, C.M., Fetter, R.D., Aleman, J.V., Franconville, Saalfeld, S., Fetter, R., Cardona, A., and Tomancak, P. (2012). Elastic volume
R., Rivera-Alba, M., Mensh, B.D., Branson, K.M., Simpson, J.H., Truman, reconstruction from series of ultra-thin microscopy sections. Nat. Methods 9,
J.W., et al. (2015). A multilevel multimodal circuit enhances action selection 717–720.
in Drosophila. Nature 520, 633–639. Sato, T. (1968). A modified method for lead staining of thin sections. J. Electron
Olsen, S.R., and Wilson, R.I. (2008). Lateral presynaptic inhibition mediates Microsc. (Tokyo) 17, 158–159.
gain control in an olfactory circuit. Nature 452, 956–960. Schindelin, J., Arganda-Carreras, I., Frise, E., Kaynig, V., Longair, M., Pietzsch,
Owald, D., and Waddell, S. (2015). Olfactory learning skews mushroom body T., Preibisch, S., Rueden, C., Saalfeld, S., Schmid, B., et al. (2012). Fiji: an
output pathways to steer behavioral choice in Drosophila. Curr. Opin. Neuro- open-source platform for biological-image analysis. Nat. Methods 9, 676–682.
biol. 35, 178–184. Schlegel, P., Texada, M.J., Miroschnikow, A., Schoofs, A., Hückesfeld, S., Pe-
Pavlou, H.J., and Goodwin, S.F. (2013). Courtship behavior in Drosophila mel- ters, M., Schneider-Mizell, C.M., Lacin, H., Li, F., Fetter, R.D., et al. (2016). Syn-
anogaster: towards a ‘courtship connectome’. Curr. Opin. Neurobiol. aptic transmission parallels neuromodulation in a central food-intake circuit.
23, 76–83. eLife 5, e16799.

Pehlevan, C., Genkin, A., and Chklovskii, D.B. (2017). A clustering neural Schneider-Mizell, C.M., Gerhard, S., Longair, M., Kazimiers, T., Li, F., Zwart,
network model of insect olfaction, Proceedings of the 51st Asilomar Confer- M.F., Champion, A., Midgley, F.M., Fetter, R.D., Saalfeld, S., and Cardona,
ence on Systems, Signals, and Computers. A. (2016). Quantitative neuroanatomy for connectomics in Drosophila. eLife
5, e12059.
Pietzsch, T., Saalfeld, S., Preibisch, S., and Tomancak, P. (2015).
BigDataViewer: visualization and processing for large image data sets. Nat. Schürmann, F.W. (2016). Fine structure of synaptic sites and circuits in mush-
Methods 12, 481–483. room bodies of insect brains. Arthropod Struct. Dev. 45, 399–421.
Potter, C.S., Pulokas, J., Smith, P., Suloway, C., and Carragher, B. (2004). Ro- Silbering, A.F., Rytz, R., Grosjean, Y., Abuin, L., Ramdya, P., Jefferis, G.S., and
botic grid loading system for a transmission electron microscope. J. Struct. Benton, R. (2011). Complementary function and integrated wiring of the
Biol. 146, 431–440. evolutionarily distinct Drosophila olfactory subsystems. J. Neurosci. 31,
13357–13375.
Price, J., and Bock, D.D. (2015). Workpiece holder for workpiece transport
apparatus. US patent application publication 14538327, filed November 11, Simpson, J.H. (2009). Mapping and manipulating neural circuits in the fly brain.
2014, and published June 18, 2015. Adv. Genet. 65, 79–143.
Price, J., and Bock, D.D. (2016a). Workpiece transport and positioning appa- Stepanyants, A., and Chklovskii, D.B. (2005). Neurogeometry and potential
ratus. US patent application publication US9449785B2, filed November 11, synaptic connectivity. Trends Neurosci. 28, 387–394.
2014, and granted September 20, 2016. Stevens, C.F. (2015). What the fly’s nose tells the fly’s brain. Proc. Natl. Acad.
Price, J., and Bock, D.D. (2016b). Workpiece transport and positioning appa- Sci. USA 112, 9460–9465.
ratus.US patent application publication 15254557, filed September 1, 2016, Stocker, R.F., Lienhard, M.C., Borst, A., and Fischbach, K.F. (1990). Neuronal
and published December 22, 2016. architecture of the antennal lobe in Drosophila melanogaster. Cell Tissue Res.
Price, J., and Bock, D.D. (2017). Specimen sample holder for workpiece trans- 262, 9–34.
port apparatus. US patent application publication US9601305B2, filed Takemura, S.Y., Bharioke, A., Lu, Z., Nern, A., Vitaladevuni, S., Rivlin, P.K.,
November 11, 2014, and granted March 21, 2017. Katz, W.T., Olbris, D.J., Plaza, S.M., Winston, P., et al. (2013). A visual motion
Prokop, A., and Meinertzhagen, I.A. (2006). Development and structure of syn- detection circuit suggested by Drosophila connectomics. Nature 500,
aptic contacts in Drosophila. Semin. Cell Dev. Biol. 17, 20–30. 175–181.

742 Cell 174, 730–743, July 26, 2018

Takemura, S.Y., Xu, C.S., Lu, Z., Rivlin, P.K., Parag, T., Olbris, D.J., Plaza, S., Wagh, D.A., Rasse, T.M., Asan, E., Hofbauer, A., Schwenkert, I., Dürrbeck, H.,
Zhao, T., Katz, W.T., Umayam, L., et al. (2015). Synaptic circuits and their var- Buchner, S., Dabauvalle, M.C., Schmidt, M., Qin, G., et al. (2006). Bruchpilot, a
iations within different columns in the visual system of Drosophila. Proc. Natl. protein with homology to ELKS/CAST, is required for structural integrity and
Acad. Sci. USA 112, 13711–13716. function of synaptic active zones in Drosophila. Neuron 49, 833–844.
Takemura, S.Y., Aso, Y., Hige, T., Wong, A., Lu, Z., Xu, C.S., Rivlin, P.K., Hess, Wanner, A.A., Genoud, C., Masudi, T., Siksou, L., and Friedrich, R.W. (2016).
H., Zhao, T., Parag, T., et al. (2017a). A connectome of a learning and memory Dense EM-based reconstruction of the interglomerular projectome in the ze-
center in the adult Drosophila brain. eLife 6, e26975. brafish olfactory bulb. Nat. Neurosci. 19, 816–825.
Takemura, S.Y., Nern, A., Chklovskii, D.B., Scheffer, L.K., Rubin, G.M., and
White, J.G., Southgate, E., Thomson, J.N., and Brenner, S. (1986). The struc-
Meinertzhagen, I.A. (2017b). The comprehensive connectome of a neural sub-
ture of the nervous system of the nematode Caenorhabditis elegans. Philos.
strate for ‘ON’ motion detection in Drosophila. eLife 6, e24394.
Trans. R. Soc. Lond. B Biol. Sci. 314, 1–340.
Tanaka, N.K., Awasaki, T., Shimada, T., and Ito, K. (2004). Integration of che-
Xu, C.S., Hayworth, K.J., Lu, Z., Grob, P., Hassan, A.M., Garcı́a-Cerdán, J.G.,
mosensory pathways in the Drosophila second-order olfactory centers. Curr.
Niyogi, K.K., Nogales, E., Weinberg, R.J., and Hess, H.F. (2017). Enhanced
Biol. 14, 449–457.
FIB-SEM systems for large-volume 3D imaging. eLife 6, e25916.
Tanaka, N.K., Tanimoto, H., and Ito, K. (2008). Neuronal assemblies of the
Drosophila mushroom body. J. Comp. Neurol. 508, 711–755. Yagi, R., Mabuchi, Y., Mizunami, M., and Tanaka, N.K. (2016). Convergence of
multimodal sensory pathways to the mushroom body calyx in Drosophila mel-
Tanaka, N.K., Endo, K., and Ito, K. (2012). Organization of antennal lobe-asso-
anogaster. Sci. Rep. 6, 29481.
ciated neurons in adult Drosophila melanogaster brain. J. Comp. Neurol. 520,
4067–4130. Yasuyama, K., Meinertzhagen, I.A., and Schürmann, F.W. (2002). Synaptic or-
Tobin, W.F., Wilson, R.I., and Lee, W.A. (2017). Wiring variations that enable ganization of the mushroom body calyx in Drosophila melanogaster. J. Comp.
and constrain neural computation in a sensory microcircuit. eLife 6, e24838. Neurol. 445, 211–226.
Van Emburg, P., and De Bruijn, W. (1984). Enhanced cellular membrane Yu, H.H., Kao, C.F., He, Y., Ding, P., Kao, J.C., and Lee, T. (2010). A complete
contrast in a marine alga by osmium-azole complexes. Protoplasma developmental sequence of a Drosophila neuronal lineage as revealed by twin-
119, 48–54. spot MARCM. PLoS Biol. 8, 1000461.

Cell 174, 730–743, July 26, 2018 743

STAR+METHODS

KEY RESOURCES TABLE

REAGENT or RESOURCE SOURCE IDENTIFIER

Deposited Data
The full adult fly brain electron microscopy This paper http://www.temca2data.org/bigtiles.html
image dataset
The full adult fly brain electron microscopy This paper https://catmaid-fafb.virtualflybrain.org
image dataset at virtualflybrain.org
Experimental Models: Organisms/Strains
Drosophila [iso] Canton-S G1 females Janelia Research Campus, N/A
Howard Hughes Medical
Institute (HHMI)
Software and Algorithms
TEMCA2 control software This paper N/A
ATPS control software This paper N/A
CATMAID: source code Saalfeld et al., 2009; Schneider- https://github.com/catmaid/CATMAID
Mizell et al., 2016
CATMAID: user documentation Saalfeld et al., 2009; Schneider- http://catmaid.readthedocs.io/en/stable/user.html
Mizell et al., 2016
CATMAID: administrator documentation Saalfeld et al., 2009; Schneider- http://catmaid.readthedocs.io/en/stable/
Mizell et al., 2016 administrator.html
CATMAID: developer documentation Saalfeld et al., 2009; Schneider- http://catmaid.readthedocs.io/en/stable/
Mizell et al., 2016 developer.html
TrakEM2 Cardona et al., 2012 https://imagej.net/TrakEM2; https://github.com/
trakem2/TrakEM2
mpicbg transformation library and feature- Saalfeld et al., 2010 https://github.com/axtimwalde/mpicbg
based point match generation
Section order correction Hanslovsky et al., 2017 https://github.com/saalfeldlab/section-sort
Cross-correlation-based point match generation This paper https://github.com/billkarsh/Alignment_Projects/
and slab alignment blob/master/00_DOC/method_overview.md
Global volume alignment This paper https://github.com/khaledkhairy/EM_aligner
Renderer This paper https://github.com/saalfeldlab/render
Image tile intensity correction (Distributed Kazhdan et al., 2010; this paper http://www.cs.jhu.edu/
misha/Code/DMG/
Gradient-Domain Processing)
Signal-to-noise ratio quantification This paper https://github.com/bocklab/temca2data/tree/
master/SNR
Neuron skeleton analysis code This paper https://github.com/bocklab/temca2data/tree/
master/geometry_analysis
Electron-light microscopy tools (ELM) Bogovic et al., 2016 https://github.com/saalfeldlab/elm
elmr This paper https://github.com/jefferis/elmr
NBLAST Costa et al., 2016 https://github.com/jefferislab/nat.nblast
NBLAST online Costa et al., 2016 http://jefferislab.org/si/nblast
R neuroanatomy toolbox Jefferis et al., 2007; https://github.com/jefferis/nat
Costa et al., 2016
R flycircuit Costa et al., 2016 https://github.com/jefferis/flycircuit
nat.flybrains Manton et al., 2014 https://github.com/jefferislab/nat.flybrains
(Continued on next page)

e1 Cell 174, 730–743.e1–e14, July 26, 2018

Continued
REAGENT or RESOURCE SOURCE IDENTIFIER
rcatmaid This paper https://github.com/jefferis/rcatmaid
CATMAID-to-Blender Schlegel et al., 2016 https://github.com/schlegelp/CATMAID-
to-Blender
Other
CATMAID-in-a-box: installation documentation This paper http://www.temca2data.org/data.html
CATMAID-in-a-box: workstation Dell Dell Precision 5720 with 16 GB of RAM,
i5-7600 processor
CATMAID-in-a-box: operating system Dell, pre-installed Ubuntu 16.04
CATMAID-in-a-box: CATMAID with Docker Saalfeld et al., 2009 http://catmaid.readthedocs.io/en/stable/
docker.html
CATMAID-in-a-box: installation code This paper https://www.github.com/bocklab/
temca2-catmaid
CATMAID-in-a-box: RAID storage device Oyen Digital Part#3R5-EB3-M
CATMAID-in-a-box: Hard disk drive Seagate ST4000NM0035
Precision piezo motor Physik Instruments Cat#N301K151
Vacuum chamber extension to FEI T-12 TEM This paper N/A
Sample support grids This paper N/A
TEMCA2 unit camera Fairchild SciMOS 2051 Model F2
TEMCA2 scintillator, 10 mg/cm2 P43 on Grant Scientific Custom part
5 mm Mylar
Fast Stage This paper N/A
Automated Transport and Positioning This paper N/A
System (ATPS)

CONTACT FOR REAGENT AND RESOURCE SHARING

Further information and requests for resources and reagents should be directed to and will be fulfilled by the Lead Contact, Davi D.
Bock ([email protected]). The following are available to nonprofit organizations upon request: control software for TEMCA2
and ATPS; CAD model files for the Automated Transport and Positioning System (ATPS), vacuum chamber extension to the FEI
T-12 TEM, sample support grids, and the Fast Stage.

EXPERIMENTAL MODEL AND SUBJECT DETAILS

Multiple brains of 7-day-old [iso] w1118 x [iso] Canton-S G1 adult female flies were screened and one was picked for EM imaging. By
this age, development has completed and flies are sexually mature (Ashburner et al., 2005). Females and males were maintained in
the same vials.

METHOD DETAILS

Sample Preparation
Brains from 7-day-old adult [iso] w1118 x [iso] Canton-S G1 flies were dissected in cold fly saline. The dissected brains were fixed with
2% glutaraldehyde in 0.1M sodium cacodylate for 1 h at 4 C, followed by 1 h at room temperature (RT). Following aldehyde fixation,
the brains were rinsed 6 3 5 min with sodium cacodylate buffer at RT, 3 3 10 min incubations in 0.02M 3-amino-1,2,4-triazole (A-TRA)
(Sigma-Aldrich) in sodium cacodylate, the last on ice, followed by post-fixation with 1% OsO4 in sodium cacodylate containing 0.1M
A-TRA for 90 min on ice (Van Emburg and De Bruijn, 1984). The brains were then rinsed with cold sodium cacodylate buffer, allowed
to warm to RT followed by deionized or Milli-Q water at RT before being stained en bloc with 7.5% uranyl acetate (Figure S3A) in water
overnight at 4 C. Following en bloc staining, brains were rinsed with water at RT and then dehydrated in an ascending ethanol series
to 100% ethanol, followed by 100% propylene oxide. Samples were infiltrated with EmBed 812 resin using propylene oxide to resin
ratios of 2:1 and 1:2 for 30 min each followed by 2 3 1 h incubations in 100% resin and a third 100% resin incubation overnight. Sam-
ples were flat embedded between Teflon-coated glass slides and allowed to harden for 24 h at 65-70 C. Samples were subsequently

Cell 174, 730–743.e1–e14, July 26, 2018 e2

screened for whole-brain sectioning by X-ray tomography (Figure S3B) using an Xradia XRM-510 X-ray microscope (subsequently
acquired by Zeiss). Samples without obvious surface defects due to dissection, or internal defects were re-embedded in silicon rub-
ber molds for sectioning.

Sample Supports, Ultramicrotomy, and Post-staining

Custom patented 3 mm bar-coded grids made from 100 mm thick copper beryllium with a 2 3 1 mm slot, a unique serial identifier in
human readable and 2D barcode form and with fiducial markers were used to collect sections (Figure S3H) (Price and Bock, 2016a).
Schematics and vendor information for the custom grids are available to non-profit research organizations upon request. Grids were
prepared for picking up sections by first applying a silver/gold-color film of Pioloform (Pioloform FN, Ted Pella catalog #19244) fol-
lowed by a
8 nm layer of carbon. The Pioloform film was made thicker than normal to provide enhanced sample stability under the
higher beam current necessary for rapid imaging. To prepare the Pioloform film, a 600 mL aliquot of 2.05% Pioloform in dichloro-
ethane was applied to an ethanol and hydrofluoric acid cleaned glass microscope slide (Gold Seal, Ted Pella catalog #260210)
via spin coating using a Laurell WS400B-6NPP/Lite spin coater. After applying the Pioloform solution, the slide was spun for 1.4 s
with a target speed of 8,000 rpm and an acceleration index of 255. The film was released from the slide by scribing the edges of
the slide with a diamond scribe and slowly submerging the glass slide at a shallow angle into a large dish of water. The film remains
floating on the surface of the water and cleaned grids were then carefully placed, bar code side down, onto the film. The film and grids
were subsequently picked up from above on a 1 3 3 inch slotted and anodized aluminum slide. The anodized surface provided a
stable and reusable surface from which the grids could be cut from the surrounding support film using a heated tungsten filament.
Grids were loaded onto custom stainless steel plates for carbon coating.
Carbon coating was carried out in a Denton Explorer 14 high vacuum evaporator equipped with oil diffusion pump, liquid nitrogen
cold trap, and a film thickness monitor using carbon rods (Ted Pella catalog #62-132). The carbon rods were de-gassed at sub-evap-
oration currents (8-14 A) prior to and immediately following sample loading. The coating plate was held at a 90-degree angle to the
source at a distance of 10 cm during evaporation. Following a vacuum recovery period, the carbon rods were de-gassed and warmed
at sub- to near-evaporation currents (8-16 A). To avoid overheating the films, carbon was evaporated in a series of cycles (in our
hands, each cycle was stopped when the deposition rate reached 0.5 Å/sec and resumed when the deposition rate returned
to 0 Å/sec). Vacuum levels prior to evaporation were
5 3 108 torr or better. Evaporation was carried out at 22 ± 1 A. Carbon evap-
oration was halted at an indicated thickness of 70 to 80 Å and final thickness assessed after a 5 min cool down period. Unsuccessfully
prepared films displayed a relaxation of the film tension when held close to water (Figure S3C), whereas successfully prepared grid
films remained perfectly flat and taut when held within
1 mm of a water surface (Figure S3D). Grid batches in which the tested coat-
ings did not remain flat were rejected.
Serial sections of the brain were cut with a Leica UC-6 ultramicrotome at a thickness of 35-40 nm, with periodic retrimming of the
block face. In total, 7,062 sections were necessary to encompass the whole brain. Total sectioning time was
3 weeks. Typically,
three serial sections were collected on each custom bar-coded grid, thus
2400 grids were used for the 7,062 sections.
Following sectioning, grids were stained in 3% aqueous uranyl acetate for 20 min followed by Sato’s lead (Sato, 1968) for 5 min,
with ddH2O washes after each staining step. To facilitate the staining of
2,400 grids, a custom Plexiglass staining device with slots
to hold 100 grids at a time, loosely based on the Hiraoka (1972) device, was used.

Electron Microscopy
Two types of acquisition systems were used to image the whole fly brain series, both of which used FEI Tecnai Spirit BioTWIN TEMs.
The first, the TEMCA2 system (Figure S2A), was equipped with a custom single-axis Fast Stage (Figure S2B; patent pending) (Price
and Bock, 2016a, 2016b), vacuum extension, scintillator (5 mm Mylar on a support ring 9 5/8 inches in diameter, coated with 10 mg
fine-grained P43/cm2; Grant Scientific), and four Fairchild SciMOS 2051 Model F2 5.5 megapixel cameras (2560 3 2160 pixel sensor
size) configured in a 2 3 2 array (Figure S3E). The second type of system was an Automated Transport and Positioning System (ATPS)
(Figures S2E, S2F, and S3I; Videos S2 and S3) (Price and Bock, 2016a, 2016b), which was equipped with a custom scintillator (6 mg
fine-grain P43/cm2; Grant Scientific), and a single Fairchild SciMOS camera. In both systems, 4:1 minifying C-lenses (AMT Imaging)
were mounted on the SciMOS cameras. These systems were previously described in abstract form (Robinson et al., 2016). Sche-
matics and model files for the Fast Stage and ATPS are available to non-profit research organizations upon request.
TEMCA2 Fast Stage
The Fast Stage (Figure S2B; Video S1) of the TEMCA2 system has a single high-speed axis of motion. It interfaces seamlessly with the
commercial FEI CompuStage goniometer, which provides the other degrees of freedom necessary to position a sample in the TEM.
The Fast Stage sample holder is connected to a drive rod, which passes through a custom rolling-element bearing, vacuum sealing
bellows, and a rolling-element damper. The drive rod is connected to a slide-mounted encoder which provides nanometer-resolution
positional feedback. It is moved linearly by a precision piezo motor (Physik Instrument cat N301K151). The custom rolling-element tip
bearing provides rigid lateral support to the drive rod within the outer drive rod tube, while minimizing axial friction required to move
the driven mass of the system. The custom rolling-element dampers kill vibrations of the drive rod induced by the pulsed motion of the
piezo motor during moves, reducing settle time. Without these dampers, the drive rod would vibrate for hundreds of milliseconds
under the pulsed motion of a move, rendering the system unusable. With the dampers, 8-24 mm moves are reliably achieved
where all vibrations are damped to less than 5 nm in less than 50 ms (Figure S2C). The miniature vacuum bellows isolates the

e3 Cell 174, 730–743.e1–e14, July 26, 2018

specimen-holding region of the device from atmospheric pressure of the operating environment. By locating the vacuum bellows just
behind the O-ring in an FEI style holder, the volume needed to be evacuated after sample insertion is minimized, allowing samples to
be exchanged in the same amount of time as a conventional holder (Price and Bock, 2016b).
Image acquisition
For TEMCA2-imaged samples, a 16.2 nm/pixel pre-bake mosaic was acquired at 60 ms exposure time to pre-irradiate the sample
and reduce specimen warping and shrinkage under high-dose acquisition. The 16.2 nm/pixel mosaics were used to specify ROIs for
4 nm/pixel mosaic acquisition. The 4 nm/pixel mosaics were acquired at 35 ms exposure times. Frames were analyzed for drift in real
time and four frames with less than 16 nm frame-to-frame drift were translated into pixel-level alignment, summed, intensity cor-
rected, and saved. Mosaics were acquired in a boustrophedonic fashion column by column running down the long axis of the
2 3 1 mm slot across the three sections such that use of the fast, piezo-driven stage axis was maximized during acquisition while
slower CompuStage moves were minimized (Figure S2D). Due to non-overlapping FOVs on TEMCA2, a two-step approach was
used where a small stage displacement (
1900 pixels, or 7.6 mm) filled the gap between the FOVs, and was followed by a large
displacement (
5500 pixels, or 22.0 mm) moving to a completely fresh FOV; this schema was used on both x and y axes with
x and y steps being slightly different (5550/1950 and 5450/1850, respectively, big step/small step, in pixels). Accurate calibration
of pixels per micron is essential for converting pixel distances into physical distances and allows for pixel distances to be kept con-
stant while the conversion factor was varied depending on the indicated magnification of the microscope. Image acquisition on the
TEMCA2 system was performed at an indicated scope magnification of 2900x, while the single camera ATPS equipped system oper-
ated at 4800x indicated magnification. The longer vacuum extension of the TEMCA2 system enlarged the projected image by
1.7x,
resulting in
4 nm/pixel for both systems.
ATPS grid handling
To better support automatic sample handling, we made custom 100 mm-thick beryllium-copper sample support grids, each etched
with unique identifier numbers and spatial fiducial marks (Figure S3H) to guide machine vision-based pick-and-place software for
grid exchange (Price and Bock, 2017).
The ATPS grid positioning system (Figures S2E, S2F, S3K, and S3M) mounts to an accessory port on the TEM and is a complete
replacement for the FEI CompuStage goniometer and specimen holder, providing all required degrees of freedom to position a spec-
imen within the TEM column, and incorporating a vacuum system allowing for sample exchange under vacuum. High-speed single-
axis motion is supported by the same drive mechanisms used in the Fast Stage. Other axes of motion are provided by piezo-driven
linear and electromagnetic rotary motors (Figure S2F). The rotational angle of the sample can be changed by placing the sample grid
on a rotary pre-aligner, rotating to the desired angle, and picking the sample back up again in the gripper (Video S3). The machine
vision system enabling automated handling of samples in the ATPS recorded continuous video while operating, providing visual
confirmation of proper operation and an invaluable debugging tool in the event of handling errors. To enable a continuous video
stream as well as high dynamic range images suitable for image processing, the acquisition stream automatically adjusts image
gain and exposure time for the required regime. These changes can be seen in Video S3.
Samples are organized in the ATPS as follows. The ATPS holds a magazine (Price and Bock, 2015) containing eight cassettes, and
each cassette holds 64 sample grids for a total magazine capacity of 512 sample grids (Figure S3J). The ATPS affords random access
to the individual grids, which can be retrieved, oriented, loaded into the TEM, imaged, and reliably returned to their proper address in
the ATPS. In the event of issues during grid exchange, the exchange is automatically halted, and the grids re-stowed. Re-stowed
grids were removed from the ATPS and imaged manually with a Fast Stage on a TEMCA2. The ATPS imaged samples in a two-
pass routine where grids were returned to cassettes between acquiring pre-bake mosaics and 4 nm/pixel mosaics. The interval be-
tween imaging steps allows for the designation of ROIs for 4 nm/pixel imaging. To ensure that ROIs were accurately acquired, the
ATPS found the center of the grid slot every time a grid was loaded into the TEM column. This center point was used to align
ROIs and correct for small differences in grid orientation resulting from the two-pass workflow. The ATPS system employed a single
point autofocus routine at the center of each section to determine focus for each ROI acquired. High-speed generation of mosaics
necessitates high electron dose rate at the sample (typically
180x the dose rate required for a 2 s exposure on Kodak 4489 film at
120 kV) to saturate the sensor wells within the short interval (35 ms in our case, versus
1 - 2 s typical integration time). Pre-irradiation
images of the grids were used to subdivide the samples into three ROI classes: (1) included areas sufficiently free of substrate dam-
age and contaminants to sustain imaging at the highest beam currents; (2) excluded areas to be masked out of the dataset entirely; (3)
borderline areas of usable but lower quality to be imaged at 1/10 intensity.
TEMCA2 and ATPS system control
Software control of the TEMCA2 and ATPS systems was written in LabVIEW (National Instruments). Wrapper software to interface the
Fairchild SciMOS cameras with LabVIEW was written in C. Hardware triggers were used to interleave stage motion with camera
frame buffer acquisition. Each camera was read out by a dedicated analysis workstation (Dell), or ‘acquisition node’, connected
via 10 Gb Ethernet to a central ‘control node’ which managed hardware triggering, stage control, ROI specification, mosaic preview,
and user interface for hardware control. Low-latency TEM hardware control (such as beam blanking, valve operation, CompuStage
control, magnification and focus adjustments, and electron beam diameter) was achieved by direct communication between Lab-
VIEW software and the FEI dynamic-link library (DLL) files supporting the FEI Tecnai scripting environment, through the DLLs’ compo-
nent object model (COM) interfaces.

Cell 174, 730–743.e1–e14, July 26, 2018 e4

 Acquisition nodes measured translational drift between successive image frames in near real-time, using the National Instruments
image analysis package. If image frame drift exceeded a user-specified threshold, the image frame was discarded and additional
frames were acquired until the requested number was acquired or until a user-specified timeout was exceeded. Each acquisition
node allocated three tiers of memory buffer to the image processing pipeline, to allow real-time acquisition to continue unimpeded,
regardless of variations in CPU load, operating system memory management, disk performance, or network throughput. In the first
tier, raw image frames were processed for drift estimation. In the second tier, sets of image frames were translated (to correct for
small translations by the sample stage), summed, normalized to a background image of the scintillator, and histogram-adjusted.
In the third tier, the summed and normalized images were written to disk. As images exited each of these buffers, memory was re-
cycled so new images could be acquired and processed. Due to the rapid rate of data acquisition, multiple storage servers, each
connected via 10 Gb Ethernet, were written to in round-robin fashion. Each server contained two RAID-6 volumes, and up to four
servers were deployed in parallel during data acquisition. If a RAID-6 volume or a server went offline, images were written to other
volumes in the available set. Solid state drives were installed in each acquisition node to allow an acquisition to complete in the event
of total network failure during acquisition. This infrastructure was capable of supporting sustained output from the two TEMCAs and
the ATPS. No data were lost due to storage or network issues during acquisition of the whole-brain EM volume.
ATPS control software was substantially similar to the TEMCA2 software except that it also controlled the ATPS hardware. ATPS-
specific functionality included machine vision-guided pick-and-place and pre-alignment of sample grids, automatic focus of the
TEM, and ROI relocation across grid picks. We also developed a user interface to let the operator define the sequence of imaging
steps to be performed as well as accompanying microscope parameters for each step. All software for control of the SciMOS cam-
eras, TEMCA2 systems, and the ATPS is available to non-profit research organizations upon request.
Section loss
Nearly all (99.83%) targeted serial section data were successfully acquired. In total 12 sections were lost prior to full resolution im-
aging (Figure S5K). Four sections (4192, 4353, 4951, 6462)) were lost during sectioning; and two grids, each containing three serial
sections (3595-3597 and 6883-6885)), were found to have ruptured support films after post-sectioning staining but prior to EM im-
aging. Two sections (4474, 3715)) were blown when electrons hit artifacts prior to full resolution imaging of the section. Sections with
debris or cracks in the support film were imaged in two rounds: a high-dose, high-throughput round, excluding potentially fragile
areas of a section; and a subsequent low-dose, slow exposure round, of the fragile region only. Twenty-five sections in nine grids
ruptured toward the end of the second-round imaging when the low-dose electron beam hit artifacts. However, a majority (if not
the entirety) of the section was already successfully imaged. In this case, although the sections were successfully imaged, the sup-
port film rupture precludes future re-imaging of these sections.
The largest serial section gaps were of three adjacent sections (sections 3595-3597 and 6883-6885; the result of two support film
ruptures while grids were in storage, awaiting imaging). No other adjacent sections were lost. Image acquisition of the whole-brain
began before the ATPS was complete, so most (
83%) of the sections were acquired by the TEMCA2 systems, and the remainder
was acquired by the ATPS. No sections were lost or damaged due to ATPS or Fast Stage malfunction.
Acquiring the whole brain
The data were acquired over a period of
16 calendar months. The TEMCA2 systems were staffed by a single microscopist, working
normal business hours; a second microscopist staffed the ATPS once it came online. Eighty-three percent of imaged sections were
acquired with a TEMCA2 system (4.3 million Fast Stage moves), while the ATPS was still in development, and 17% of imaged sec-
tions were acquired by the ATPS (3.5 million grid positioning system moves;
6,800 machine vision-guided steps to pick, pre-align,
and re-stow each grid). Eighty-two percent of ATPS grid exchanges were successful; 14% were automatically halted and the grids re-
stowed, usually due to variations in the manual placement of grids in the ATPS cassettes or inhomogeneities in the support film; and
4% required manual control of the ATPS for re-stowing. Re-stowed grids were removed from the ATPS and imaged manually with a
Fast Stage on a TEMCA2.
Acquisition rates
The TEMCA2 per-section acquisition rate is 50 MPix/s, which does not include overhead such as sample exchange time and defi-
nition of ROIs. Since the ATPS only has one camera instead of a 2 3 2 array, its per-section acquisition rate is
1/4 that of a TEMCA2,
or 12.5 MPix/s. With overhead included, our effective per-grid throughput for TEMCA2 was 27 MPix/s. The TEMCA2 overhead
was
10 min per grid, whereas the ATPS overhead was
5 min per grid due to autonomous sample exchange and predefined
ROIs. Therefore the per-grid throughput for the ATPS was 8 MPix/s, slightly faster than 1/4 of the TEMCA2 per-grid throughput. If
an ATPS was mounted on a TEMCA2 system it could achieve another 4-fold throughput increase, for
32 MPix/s per-grid
throughput. Microscopist supervision was needed for most of the whole-brain acquisition. Assuming a 2000-hour work year
(50 weeks at 40 hours per week) for
16 months, it took
2,666 microscopist hours to acquire the
106 TB whole-brain dataset.
Estimated operator throughput, derived from the total microscopist hours, is therefore
11 MPix/s.
The ATPS grid exchange failure rate (18%) had minimal impact on its daily productivity for two reasons. For most exchange failures,
the problematic grid was re-stowed automatically and imaging of the next grid began. Grids that were not successfully acquired by
the ATPS were unloaded from the ATPS cassettes and imaged manually on a TEMCA2.
For comparison of acquisition rates between the TEMCA2 and other techniques, per-section throughput was used for ssTEM and
TEMCA2. For ssTEM, TEMCA, and TEMCA2, step and settle times for sample motion during mosaic acquisition were included. Data
rates for ssTEM were calculated based on image acquisition parameters used by one of us (RF) during acquisition of the dataset

e5 Cell 174, 730–743.e1–e14, July 26, 2018

described in Takemura et al. (2013). Data rates for FIB-SEM, SBEM, and ATUM-SEM were calculated from published values (Brigg-
man and Bock, 2012; Kasthuri et al., 2015).

Volume Reconstruction
Volume reconstruction includes stitching images from a single thin section into a coherent mosaic and aligning mosaics across all
sections. Stitching and alignment of ssTEM images into a traceable volume must overcome artifacts and distortions introduced in
sample preparation, serial sectioning, pickup, staining, and imaging of samples. During sectioning, ultrathin sections can be torn,
misfolded, or lost completely. EM-dense debris may be introduced during sectioning and/or post-staining of thin sections, obscuring
the sample during imaging (Figures S5D and S5E). This debris can charge and heat in the electron beam, putting the support film at
risk of rupturing. The sample can also deform during high-intensity imaging. These challenges have largely been addressed for
smaller EM volumes (Saalfeld et al., 2012), but the scale of the whole fly brain required the development of a cluster-backed software
alignment pipeline to successfully assemble the dataset. Importantly, the new alignment pipeline also enabled projection of traced
neuron reconstructions across different versions of the aligned dataset. This allowed tracing to begin before acquisition of the entire
series was complete, and preserving tracing done in previous versions of the aligned volume.
Data migration
Camera images were written in a round-robin fashion across multiple high-speed storage servers. Mosaics selected for inclusion into
the final reconstructed volume were copied to a centrally managed distributed file system at Janelia Research Campus offering high-
throughput connectivity to the computational cluster as well as off-site backups. All images were checksum verified after file copy
operations.
Calibration mosaics
In our TEMCA2 system, we operated with a wider FOV than a conventional TEM which comes at the cost of individual images
showing significant non-linear distortion. This distortion is the accumulation of camera lens-distortion, variation in camera mounting,
and warping in the electron beam path. We compensated for this distortion using the lens-correction method available in TrakEM2
(Cardona et al., 2012; Kaynig et al., 2010) followed by affine normalization between all distortion models. For each individual camera,
we imaged a 3 3 3 mosaic of redundantly (60%) overlapping tiles of a neuropil compartment in one of our sample grids. This mosaic
was then used to estimate a non-linear distortion correction model in TrakEM2. To compensate for the remaining affine distortion
(scale and shear) of each of these camera models, we imaged a large reference mosaic in the neuropil compartment of three refer-
ence sections (to account for accidental section loss) that we then jointly aligned with TrakEM2. This way, we obtained a globally
consistent camera calibration model for each individual camera. We repeated the calibration step each time an imaging system
was adjusted.
Stitching and alignment
The image acquisition process provides partially overlapping images that are assumed to cover each entire section. Image mosaics
need to be stitched within each z section plane, as well as aligned across z to produce a seamless volume. The stitching and align-
ment process includes: (1) matching point-pairs within mosaics, (2) matching point-pairs across mosaics, (3) registering high-dose/
low-dose images, (4) section order correction, and (5) solving the volume.
Details of the methods and documentation of actively used code are available at:
https://github.com/billkarsh/Alignment_Projects/blob/master/00_DOC/method_overview.md
https://github.com/billkarsh/Alignment_Projects/blob/master/00_DOC/ptest_reference.md
https://github.com/khaledkhairy/EM_aligner
Matching point-pairs within mosaics
The first step of the stitching process is to match putatively identical content between pairs of overlapping images; those matched
point-pairs are stored in a table. Matching was done within each of the serial sample sections (z-layers), considered independently of
any other sections. Two neighboring images would match essentially perfectly except for very slight differential beam heating.
TEM stage coordinates provide useful guesses about which pairs of images have overlaps worth characterizing, as well as the
expected relative transform between pair members that we can use to constrain content matching. For each prospective pair of im-
ages we first perform coarse matching using normalized fast fourier transform (FFT)-based cross-correlation to obtain a best rigid
transform between them: relative rotation and xy-translation. The expected constraint transform enters as a mask describing a
disc of preferred xy-translations within the correlation image.
The coarse transform between image A and B is then refined using a deformable mesh as follows. Within the overlap region of A
and B, the A-pixels remain at fixed coordinates. For the B-image pixels, we erect a mesh of triangles and each of the B-pixel coor-
dinates within are translated into barycentric coordinates (functions of the triangle vertices) which are variables. The normalized
cross-correlation between A and B can now be expressed as a function of mesh vertex coordinates. A gradient descent process
is used to find vertex positions that optimize correlation.
The reported point-pairs linking A to B are derived from the triangles of the mesh. Image-point A is defined as the centroid of a given
mesh triangle prior to optimization. Its corresponding B-image point is obtained by calculating the affine transform that takes the tri-
angle to its optimized counterpart, and applying that to the A-centroid.

Cell 174, 730–743.e1–e14, July 26, 2018 e6

Matching point-pairs across mosaics
Since the layers are nominally 40 nm thick and neural processes propagate through tissue at all possible angles, content in adjacent
layers is grossly similar but isn’t a precise match. Nevertheless, content-based matching as described above for same-layer image
pairs (FFTs followed by deformable mesh optimization) works very well if combined again with expected pair-pair transforms for
which we have high confidence.
First we match whole layers to each other: For each layer, individually, we collect the reported in-layer point matches and solve for
its set of affine transform parameters that register that layer’s 2D images to form a so-called montage. These data are used to render
the layer at a reduced scale (
20X) to an image that we call the ‘montage scape’. Scale reduction allows the problem to fit comfort-
ably in RAM, reduces computation time, and most importantly, emphasizes larger size tissue features such as large neurites running
parallel to the z axis, which vary slowly as a function of Z. Each pair of montage scapes is matched by FFT cross-correlation at a series
of angles and the best correspondence is determined. This is followed by manual inspection using TrakEM2 (Cardona et al., 2012) to
verify this rough alignment.
Registering high-dose/low-dose images
High-dose/low-dose imaging (see Electron Microscopy) of robust and fragile areas of a section, respectively, could be stitched
together seamlessly (Figures S5D–S5F). The majority of high-dose/low-dose sections are acquired during a single session, without
the sample being removed from the microscope. Therefore, a reliable first guess for relative positions of these layer patches is usually
provided. Generally, high-dose/low-dose sections are registered in a process that takes advantage of components of the general
registration pipeline above. Montages of individual acquisitions are generated and their point-matches stored. All montages sharing
the same z-value (i.e., the high-dose/low-dose group of sections), together with reference neighbor ‘sandwich’ sections are treated
as a set of sections that are roughly aligned to each other as if they were all separate sections. This rough alignment is used to deter-
mine potential overlap of high-dose and low-dose areas. Tile-pairs are determined and their point-matches calculated and stored.
Finally, all point-matches (within-layer, across high-dose/low-dose patches, and cross-layer to neighboring reference sections) are
used to solve a linear system to determine transformation parameters for a seamless registration.
Section order correction
We implemented a fully automated procedure for whole-layer matching. Image features are extracted from section montages (Saal-
feld et al., 2012), and point-correspondences are determined for all pairs of sections within a range of expected ordering mistakes
(in our case within 100 sections). During alignment of the volume,
250 sections were found to be misordered. We then used the
number of point-correspondences between two sections as a surrogate for their inverse relative distance and identified the shortest
possible path to visit all sections, resulting in an ordered series (Hanslovsky et al., 2017). Then, a regularized linear system is solved to
calculate an affine transformation for each section that roughly aligns the volume.
With all of the layers now coarsely aligned, we subdivide each layer into an array of ‘blocks’ (
10 3 10 neighborhoods of camera
images). We again step angles and calculate FFT cross-correlation, this time on pairs of corresponding blocks to find the best block-
block cross-layer transforms. As a result we know which images within the blocks pair with each other and what their relative
transform ought to be. Again, we subdivide each image into local regions, estimate point correspondences using FFT-based
cross-correlation, and collect these correspondences in a database.
Solving the volume
With the full set of point-pairs tabulated, each image is typically connected to several of its neighbors. We then construct a system of
equations requiring that, under the sought affine parameter set that defines each image transformation, point-pairs should map to the
same global point in the reconstructed volume. To avoid spurious deformation and volume shrinkage, the equation system is regu-
larized to a roughly aligned volume. This roughly aligned volume depends on individual montages that were in turn regularized to a
rigid model approximation that is independently estimated. The full system constitutes a large linear sparse matrix problem, whose
solution provides the globally optimal transformation for all images simultaneously.
Sources of error
Wrong (low-quality) point-pairs. These may occur due to the self-similarity of nominally good quality neural EM images. Errors are
even more likely in tissue regions that are substantially devoid of neurons or texture, such as the lumen of the esophagus, or along
the outer boundary of the sample where tissue is sparse or even absent from several camera images. To address this error we employ
(a) auxiliary contextual information about the likely transform between any two images that constrains matching derived from local
image content alone, and (b) we impose a strict point-matching filter using Random Sample Consensus (RANSAC) (Fischler and
Bolles, 1981) to separate true correspondences that behave consistently with respect to an affine transform up to a maximal corre-
spondence displacement (Saalfeld et al., 2012).
Missing point-matches. In some cases tissue damage, contamination or folds within a section lead to a lack of point-matches in a
smaller region within the volume. This is most prominent when searching for point-matches across z. We address this issue by ex-
tending the point-match search beyond immediate neighbor sections.
Image Intensity Correction
Variations in section thickness, electron beam etching, or deposition of contaminants from post-staining or section pickup create
intra-mosaic variations in camera image intensity. For ease of tracing, these variations were corrected (Figures S5G–S5J) using a
scalable implementation of an existing algorithm (Distributed Gradient-Domain Processing) (Kazhdan et al., 2010).

e7 Cell 174, 730–743.e1–e14, July 26, 2018

 During iterative volume reconstruction, gradient-domain processing is used to remove seams in two dimensions. A target gradient
field is constructed by computing the gradient field of the input mosaic and zeroing out seam-crossing gradients. Then, a least-
squares system is solved to find the new image whose gradients best fit the target field. In addition, low-frequency modulation is
removed by computing the windowed average of adjacent mosaics and replacing the low-frequency components of an input mosaic
with the low-frequency components of the average. We anticipate that future work will allow 3D processing of the whole-brain image
volume (Kazhdan et al., 2015), reducing or eliminating section-to-section variations in intensity.
Projection of tracing across alignments
With each new alignment, the CATMAID PostgreSQL database containing all neuronal skeleton coordinates (Schneider-Mizell et al.,
2016) is dumped to retrieve their ‘world’ coordinates (coordinates representing their physical location in the brain). Each of these
world coordinates is then inversely transformed using the Renderer service (see Stack management and relational database) to a
set of ‘local’ coordinates detailing the source tile visible at that location and the relative location within. The local coordinates are
projected back into world coordinates using the new alignment’s transformations. The updated coordinates are then applied to a
new copy of the database.
Stack management and relational database
We created a relational database for storing and querying metadata associated with the thousands of image mosaics and millions of
acquired images. We use SQL Server 2012 for our production system and SQLite for development. Metadata required for down-
stream processing included: paths to image data (with checksums), stage coordinates, ROIs associated with nominal section
numbers, ordering of sections and microscope configurations with associated calibrations. The input for the alignment process –
a stack – can be generated with a single SQL query joining the majority of tables. The result is a list of images with their layer (z), stage
coordinates (x,y), and camera configuration (for associating the correct lens correction model).
The alignment process of the approximately 21 million images and associated projection of already-traced skeletons between
alignment iterations is computationally expensive. To manage this we developed the Renderer toolkit (https://github.com/
saalfeldlab/render), a set of image stack management tools and RESTful HTTP web services now in use in multiple additional pro-
jects. Renderer was designed in order to handle large -scale (hundreds of millions) individual records efficiently while supporting
large-scale concurrent access for the stitching, section order analysis, skeleton mapping and intensity correction. Briefly, Renderer
is able to quickly materialize (i.e., render) modified images for a set of transformation parameters using the mpicbg transformation
library (https://github.com/axtimwalde/mpicbg). The use of the mpicbg library allows simple conversion between the Renderer data-
base (a MongoDB instance) and TrakEM2 projects. For large-scale rendering and coordinate mapping, we used Java stand-alone
and Spark framework clients to allow it to be processed in bulk on a cluster.
Assessment of alignment quality
Reconstructed neuron skeletons typically recapitulate the natural tortuosity of neuronal arbors, but can be distorted artificially by
section misalignments. Large abrupt changes in the coordinates of skeleton nodes often indicate section misalignments. To evaluate
the alignment quality of the data, smoothness of reconstructed neurons was assessed. The reconstructed skeletons of all PNs, KCs,
and the APL were convolved with a Gaussian kernel (s = 12 mm). The s was chosen to eliminate large abrupt changes in the skeletons.
The distance between each node of the original skeletons and the smoothed skeletons was computed, and the mean distance for all
skeleton nodes in each section was then calculated. The absolute difference between the mean distance in each section and that in
the preceding section was computed, which we call the displacement (Figure S5L). The median and the 95% percentile of the dis-
placements are 0.09 mm and 0.57 mm, respectively.

Neuron Tracing
Neuron reconstructions are based on manual skeleton tracing of neuronal arbors and annotation of synapses from image stacks in
CATMAID (http://www.catmaid.org) as described in Schneider-Mizell et al. (2016). All neurons included in analyses are reconstructed
by at least two team members, an initial tracer and a subsequent proofreader who corroborates the tracer’s work. In the event that
any tracer or proofreader encounters ambiguous features (neural processes or synapses that are not identifiable with high confi-
dence), they consult other tracers and proofreaders to determine the validity of said features, climbing the experience ladder up
to expert tracers as needed. If any feature remains ambiguous after scrutiny by an expert tracer, then said feature is not included
in the neural reconstruction and/or flagged to be excluded from analyses. During the proofreading phase, the proofreader and tracer
iteratively consult each other until each neuron is deemed complete per the specific tracing protocol to which it belongs. An assign-
ment of completion does not necessarily entail that an entire neuron’s processes and synapses have been reconstructed (see
Tracing to classification and Tracing to completion). We traced 114 PNs, the APL, two MB-C1s, MB-CP1, and two MB-CP2 neurons
to classification (120 neurons in total). We also traced the calyx sub-arbors of the 15 KCs to completion, and their remaining sub-
arbors to morphological, but not synaptic, completion. The total cable length of the neurons above is 206.6 mm.
The criteria to identify a chemical synapse include at least three of the four following features, with the first as an absolute require-
ment: 1) an active zone with vesicles; 2) presynaptic specializations such as a ribbon or T-bar with or without a platform; 3) synaptic
clefts; 4) postsynaptic membrane specializations such as postsynaptic densities (PSDs). In flies, PSDs are variable, clearer at
postsynaptic sites of KCs in a microglomerulus but often subtle, unclear, or absent in other atypical synaptic contacts (Prokop

Cell 174, 730–743.e1–e14, July 26, 2018 e8

and Meinertzhagen, 2006). In the absence of clear PSDs, we marked all cells with membranes that have unobstructed access to a
clearly visible synaptic cleft as postsynaptic. We did not attempt to identify electrical synapses (gap junctions), since they are unlikely
to be resolved at the resolution of this dataset.
Very rarely, aberrant neurons were found in the dataset. For example, two PNs with a piece of fused cell membrane were discov-
ered in our tracing. It is unknown what factors might cause this, but cell membrane pathologies resultant from EM fixation protocols
have been observed (Kopek et al., 2017). Overall, however, the ultrastructural quality of the whole brain was excellent.
Tracing to classification
Often only reconstruction of backbone (e.g., microtubule-containing backbone neurites) (Schneider-Mizell et al., 2016) or gross
morphology is needed to classify a neuron based on expert identification or NBLAST-based neuron searching against an existing
LM dataset. If either approach fails to find a match (as in the case of MB-CP2 in our study), the neuron may be deemed a new
cell type. Neurons traced to classification are at a minimum skeletonized, with or without synapses, to the point at which their gross
morphologies (or backbone skeletons) unambiguously recapitulate that observed by LM for a given cell class, or are unambiguously
deemed as a new cell type not previously observed in LM databases from NBLAST neuron morphology search and/or multiple
experts.
Tracing to completion
All steps for tracing to classification were completed. Additionally, for morphological completion every identifiable neurite is traced,
and for pre- or postsynaptic completion every identifiable presynaptic release site or postsynaptic density, respectively, on the
neuron is annotated. Some neurons were traced to morphological or synaptic completion only in certain brain regions (e.g., PNs
were traced to morphological, but not synaptic, completion in the MB calyx).
Quantification of tracing speed
We reconstructed 114 olfactory PNs to classification and 15 KCs to morphological, but not synaptic completion, for a total cable
length of 153.4 mm and 11.6 mm, respectively. The mean PN cable length was 1.35 mm and the mean KC cable length was
0.78 mm. The PN total construction time was 1,272 h and the KC total construction time was 166 h, with mean per cell construction
times of 11.2 h and 11.1 h. To calculate tracing speed we divided the mean cable length produced by the sum of the mean time
required for both manual tracing and proofreading for all 114 PNs and all 15 KCs reported in this manuscript. Tracing and proof-
reading timestamps for all additions, deletions, modifications, and proofreading of skeleton node events were recorded into a data-
base and can be queried via a history widget in the CATMAID tracing environment. Construction time was measured as active bouts
of tracing and proofreading. Active bouts were defined as temporal windows during which tracing and proofreading timestamps were
continuously applied to a neuron by each user with no larger than a 3 min pause. Any 3 min pause or longer introduced a gap between
active bouts, and the next active bout began at the next timestamped event of that neuron. The duration of active bouts for all users
for each neuron was then summed to determine total tracing and proofreading time.
Validation of tracing accuracy
To test the reproducibility of tracing in the whole-brain EM dataset, three independent teams each comprising two members, one
tracer and one proofreader, each reconstructed one KC sub-arbor to completion in tracing environments blinded to each other (Fig-
ure S6). The KC sub-arbor included the soma, all arbors in the MB calyx, and a proximal portion of the MB pedunculus (Figure S6A). In
the fly brain, microtubule-free neurites (twigs) as fine as 40 nm in diameter tend to extend for short distances before joining larger,
microtubule-containing backbone neurites (Schneider-Mizell et al., 2016). KC claws are good examples of twigs, whereas their den-
dritic trunks and the axonal neurite leaving the MB calyx are larger-diameter backbones. In the tracing phase, the tracer had access to
the proofreader for consult and verification. During the proofreading phase the proofreader had access to the tracer for consult and
verification. The neuronal arbors and associated synapses reconstructed by each team were essentially identical for both twigs and
backbones. An expert validated all sites of disagreement between the three teams to determine a ‘gold-standard’ skeleton of the KC.
PN-to-KC claw inputs with high synapse counts were detected in all three reconstructions (Figure S6B). Consistent with a tracing
approach biased toward false negatives rather than false positives, only one input with low synapse count was missed by one of
the tracing teams (Figure S6, red asterisks).
We re-implemented a skeleton-to-skeleton agreement measurement (Helmstaedter et al., 2011) to locate the sites of discrepancy
between the skeletons of each of the three teams and the gold-standard skeleton, respectively. All skeletons were resampled with an
interval of 80 nm (Schneider-Mizell et al., 2016), and nodes from each team skeleton were compared with nodes from the gold-stan-
dard skeleton. Each team node was considered to be consistent with a gold-standard node if the team node was within 800 nm of a
gold-standard node. Thresholds smaller than 800 nm resulted in correct reconstructions of lower order neurites to be inconsistent.
We visually inspected the inconsistent nodes to determine the number of errors. All errors were false negatives. The length of missed
skeletons for each error was measured (Figure S6A). Error rates differ significantly between microtubule-containing backbones and
microtubule-free twigs (Schneider-Mizell et al., 2016). In our tracings, the shortest path from soma to MB pedunculus of the KC
(spine), and the first order branches of the spine (arms) that give rise to the claws were classified as backbone as these neurites
contain microtubules. The remaining higher order branches (e.g., claws) are considered twigs. We found no errors for 167 mm of
backbone skeletons, and the twig (287 mm) mean error rate was 37.9 mm/error. The total mean error rate for the gold standard KC
skeleton was 60.8 mm/error (mean number of disagreements was 7.3 across a neuron 454 mm in extent). Our error rates were
compared with those from two previous studies. In mammalian retina (Helmstaedter et al., 2011), the total mean error rate for
five-fold independent tracing of a single neuron was 83.4 mm/error (mean number of disagreements was 7.2 across a neuron

e9 Cell 174, 730–743.e1–e14, July 26, 2018

600 mm in extent); backbone and twig error rates were not reported separately. In Drosophila larva, the mean error rate for an iter-
atively traced skeleton was 375.8 mm for backbone (three errors) and 16.2 mm for twigs (95 errors) (Schneider-Mizell et al., 2016).
In total 1127 mm of backbone cable length and 1,539 mm of twig cable length were traced. Therefore, with 98 errors across a neuron
2,666 mm in extent, the total mean error rate was 27.2 mm/error in the larva.
Tracing of PNs
Three protocols were used to reconstruct PNs on the right side of the brain. First, putative PN boutons presynaptic to all claws
of
200 KCs traced from a separate ongoing effort were seeded and traced to classification. Second, a seed section at the postero-
lateral bend of the mALT, proximal to MB calyx, was selected and all neurons not found via the first protocol were traced directly
toward the calyx. Neurons that innervated MB calyx were traced to classification, whereas those that bypassed calyx were halted.
Third, a thorough visual survey of the MB calyx was conducted to ensure that all microglomerular structures had been identified and
the untraced boutons within these microglomeruli were seeded with single skeleton nodes then traced to classification.
We focused our classification efforts on uniglomerular olfactory PNs that provide input to the MB main calyx. Additional PNs were
discovered but excluded from PN counts per subtype (Figure 4E) if they bypass calyx. The expert identification of PN subtypes was
based on gross morphology. Key diagnostic features included the somatic locations, projection tracts, and distribution of arbors in
the AL (the glomerulus innervated by each PN), MB calyx, and LH. Classification of olfactory glomeruli in AL followed that of Grabe
et al. (2015), with a few exceptions: we did not attempt to classify VP2, VP3, and VP4 because they are either non-olfactory or multi-
modal (Enjin et al., 2016; Grabe et al., 2015); our reconstructions indicated clear independent domains for VC3l and VC3m glomeruli
(consistent with Chou et al., 2010; Silbering et al., 2011); and VC3l and VC5 could not be disambiguated. PNs arising from these two
glomeruli were therefore left unclassified, but future tracing of the neurites from neurons in surrounding glomeruli may clarify their
identities. Following Grabe et al. (2015) and Yu et al. (2010), VM6 and VP1 were combined into a single glomerulus due to morpho-
logical ambiguities, which we label as VM6 in this work. Reconstruction of ORNs, multiglomerular/multimodal PNs, and tracing of
PN dendrites to completion, will likely clarify remaining ambiguities in glomerular classification.
Axonal boutons of PNs in the MB calyx were identified by varicosities containing arrays of presynaptic active zones, each of which
was apposed to many postsynaptic processes (Figure S1A). Skeleton nodes at the varicosity/intervaricosity borders were tagged as
‘bouton borders’ such that they contained all synapses inside each varicosity.
Of the 578 microglomerular boutons in MB main calyx, 497 arose from olfactory PNs (86%, from 114 PNs). Of these, 16 boutons
(3%) arose from six olfactory multiglomerular PNs. The other inputs include 17 boutons (3%) arising from MB-CP2; nine boutons from
other PNs (2%, arising from four neurons), traveling either via tracts alternative to the mALT (eight boutons from three PNs) or from the
subesophageal region (one bouton from one putative PN; data not shown); 55 boutons (9%) from the other 10 unidentified (presum-
ably non-olfactory, e.g., thermosensory, hygrosensory) multiglomerular or uniglomerular PNs.
Tracing MB-CP2 to classification
The two MB-CP2 neurons were traced to classification. In this process, many second order branches within each neuropil compart-
ment were traced. To obtain a representative sample of the synapses in the right hemisphere, we selected several third order
branches in each brain region, traced them to their terminal processes (the highest order of branching within that region), and anno-
tated all synapses on these neurites. We identified its synaptic partners in cases where those neurons had already been traced for
other purposes (e.g., previously reconstructed KCs). Most cell types in the fly brain have symmetric contralateral equivalents (Aso
et al., 2014; Chiang et al., 2011; Jenett et al., 2012). Therefore we also reconstructed the contralateral MB-CP2 neuron and its
synapses to assess the symmetry of its morphology and connectivity, a previously used technique to validate the reconstruction ac-
curacy of known and previously unidentified cell types in the Drosophila larval EM volume (Eichler et al., 2017; Ohyama et al., 2015;
Schneider-Mizell et al., 2016) (Figure 6A). Fewer processes and synapses were traced and annotated for the MB-CP2 neuron in the
left hemisphere than for the MB-CP2 neuron in the right hemisphere.
Kenyon cells and their postsynaptic partners in the MB calyx
Three KCs from each of the five KC classes that innervate MB main calyx (g, abc, abs, a’b’m, and a’b’ap) were randomly selected
from a larger set of several hundred KCs already traced to classification as part of a separate, ongoing study. Only presynaptic sites in
the neuropil region of the MB calyx were included. Synapses in MB pedunculus and lobes were not reconstructed. All postsynaptic
partners at each presynaptic site were enumerated and traced to classification.
The 14 previously known cell types that could potentially be postsynaptic partners of KCs in MB main calyx were enumerated by a
literature search. They are: the octopaminergic ventral paired median neurons (OA-VPM3 and OA-VPM5), the octopaminergic ventral
unpaired median neuron (OA-VUMa2); the putatively GABAergic glutamic acid decarboxylase interneurons (GAD1-GAD4), MB-C1,
MB-C2; the contralaterally projecting serotonin-immunoreactive deutocerebral (CSD) neuron; the dopaminergic posterior lateral pro-
tocerebum neuron (PPL2ab); the MB calyx MBON MB-CP1; the lateral horn leucokinergic (LHLK) neuron; and the APL (Aso et al.,
2014; Burke et al., 2012; Busch et al., 2009; Chen et al., 2012; de Haro et al., 2010; Mao and Davis, 2009; Roy et al., 2007; Tanaka
et al., 2008).

Neuronal Informatics
Electron-Light Microscopy registration tools (ELM)
The ELM tool provides a user interface to interactively define a thin plate spline 3D warp field between an LM dataset and the
whole-brain EM dataset by manually specifying corresponding landmark points. It was built on top of the BigWarp Fiji plugin

Cell 174, 730–743.e1–e14, July 26, 2018 e10

(Bogovic et al., 2016), which in turn was built on top of the BigDataViewer plugin (Pietzsch et al., 2015) for Fiji (Schindelin et al., 2012).
The ELM software is aware of standard compartment boundary models available for the template fly brains and provides hotkeys to
view the labels for these compartments; to go between coordinates in ELM and the EM dataset as viewed in CATMAID; and to go
from a CATMAID URL to the corresponding point in ELM. Code for ELM is available at https://github.com/saalfeldlab/elm. This soft-
ware can be used to map individual LM-imaged neurons that have been registered to a template brain into the EM whole brain.
An LM image volume of a whole fly brain in which random fluorophore combinations are expressed in a subset of PNs (MultiColor
FlpOut) (Nern et al., 2015) was kindly provided by Yoshino Aso and colleagues (Y. Aso, personal communication). The image volume
was registered to the JFRC2013 template brain. Using the JFRC2013-to-EM volume transformation defined above the PNs were then
projected into the coordinate space of the EM whole brain.
Transforming data between EM and LM templates: elmr
The elmr software tool is a package (https://github.com/jefferis/elmr) written in R (http://www.r-project.org) to facilitate bidirectional
transfer of 3D data between the EM whole-brain dataset and LM template brains. This package reproduces the thin plate spline trans-
formation that maps between the whole-brain EM volume and the JFRC2013 template brain (Aso et al., 2014) from the landmark pairs
exported by the ELM tool described in the previous section. Using an approach described by Manton et al. (2014), 3D data can then
be transformed into other light level template brains used by the fly neurobiology community, including the template (JFRC2, a.k.a.
JFRC2010) (Jenett et al., 2012) used for data on the virtualflybrain.org site. This process depends on an additional R package (nat.fly-
brains, https://github.com/jefferislab/nat.flybrains) that provides non-rigid registrations defined with the CMTK image registration
toolkit (https://www.nitrc.org/projects/cmtk). This and all additional R packages used for subsequent analysis are automatically
installed as dependencies of elmr. Inverse transformation is also supported, enabling data registered to a light level template to
be projected into the space of the whole-brain EM volume.
Surface models of neuropil compartments
Previously defined surface models of the whole fly brain and several brain regions (Ito et al., 2014; Jenett et al., 2012; Manton et al.,
2014), based on the same template brain as the virtualfybrain.org project (https://github.com/VirtualFlyBrain/
DrosAdultBRAINdomains), were transformed (Manton et al., 2014) to the EM volume using elmr. The AL glomerulus meshes were
generated in Blender (https://www.blender.org/) from EM-reconstructed skeletons of PN dendrites and ORN termini retrieved
with the CATMAID-Blender interface (Schlegel et al., 2016).
Remote deployment for tracing in the whole brain (‘‘CATMAID-in-a-box’’)
We provide a preconfigured tracing environment including the CATMAID tracing software, the reconstructed neurons presented in
this manuscript, and scripts to download the entire EM volume from the temca2data.org website (http://www.temca2data.org/). The
CATMAID software has built-in functionality for importing and exporting neuron skeletons, which facilitates sharing of traced neurons
across laboratories.
Groups wishing to use CATMAID-in-a-box will need a mid-range Linux or Mac workstation capable of running Docker Community
Edition and 12 TB of disk space to store the EM volume. We tested a Dell Precision 5720 (16 GB of RAM with a quad-core 3.5 GHz
Intel i5-7600 processor) and a commodity external RAID-5 enclosure configured with five 4 TB hard disk drives. Full details for sug-
gested configurations are maintained on the temca2data.org website.
CATMAID (and its backing PostgreSQL database) are distributed as Docker containers to facilitate installation without complicated
dependencies. The full 12 TB image volume is downloaded over the Internet. The provided script to download data will checkpoint its
status to allow resumption in the event of network interruption. Full step-by-step installation instructions are provided on the
temca2data.org website.

QUANTIFICATION AND STATISTICAL ANALYSIS

Comparison of SNR between Volume EM Datasets

Determining the SNR of biological images is in general a subjective task, due to its variance under non-linear transformations (Erdog-
mus et al., 2004). As users of this data will likely care about biological structures, the determination of SNR should account for this,
considering only the level of signal of these structures and not of things such as staining or cutting artifacts. The problem of SNR
determination has been thoroughly treated in the case of super-resolution imaging where these ambiguities don’t exist (Lambert
and Waters, 2017; see also Supplementary Note 1 in Li et al., 2015), but as yet there are no universally accepted, automated tech-
niques to calculate the SNR in individual images where signal is dense in both spatial and frequency spaces, such as EM data of brain.
We present two measures of SNR here, an automated measure which avoids user biases, but can include some signal in noise and
background calculations (feature-based SNR), and a simple technique which gives more precise SNRs but is prone to bias (the cell
membrane technique) which we use to verify the feature-based SNR calculation. We apply these techniques to a range of publicly
available data in order to evaluate the TEMCA2 method. Sample images from each dataset are shown in Figure S4A.
We assume all methods are shot-noise limited. For comparison purposes SNR values are normalized to the TEMCA2 voxel size
(4 3 4 x 40 nm) by the square root of the source data’s voxel size. In both the feature-based SNR and the cell membrane SNR mea-
sures, we assume that noise is additive (and is independent of the magnitude of the signal) and symmetric. Such an assumption is
likely false (e.g., electron shot noise is Poissonian and not symmetric at low numbers), however, such impacts are likely small
based on manual examination of images and we assume the impacts of such an assumption are the same for all techniques.

e11 Cell 174, 730–743.e1–e14, July 26, 2018

Such assumptions may fail, however, at very low signals where Charge-coupled device (CCD) and shot noise dominates, or at high
signals, where processes such as non-linearity in CCD absorption become important.
Feature-based SNRs
Fundamentally, an SNR calculation of an image involves a calculation of the background level, the variation in this background level
(which is assumed to be due to noise) and the calculation of the difference between the ROI and this background. Detecting what
these regions are provides a challenge in EM data where images may not have clear background regions and where noise is contrib-
uted to through sample preparation.
In order to measure the SNR we assumed that in any given image, the structures of interest provide the majority of features above
the noise. That is, most structures present are biological in nature, rather than artifacts of sample preparation. Therefore with this
assumption, it can be further assumed that key-points detected by feature detection algorithms will disproportionately fall on the ROI.
Given that cells and structures therein tend to be ‘blobby’ due to hydrostatic processes (Jiang and Sun, 2013), we used a blob-
detection algorithm (which compares areas of interest, cf. edge or corner detection) to identify areas of interest. We used the
SURF algorithm (Bay et al., 2008).
Following the above, the variation in intensity of an image, I, in the local region of many feature points is likely to be mostly due to
signal, and the variation in intensity nearby few (or no) feature points will be dominated by noise. The determination of such regions
was done by generating an array of equal size to the original image and for each element, setting it to one if there is a feature in the
corresponding element of the image. This array was then convolved with a Gaussian of width n, where n was chosen to maximize the
SNR in a random selection of five images from each sample in an effort to avoid bias between samples.
To select a region dominated by noise we then shuffled this array before sorting it (to avoid biases in sorting algorithms) and took
the lowest point. We then blocked out a region 2n square and resorted the array nine more times (for a total of 10 selected regions),
forming the set of points plow. We likewise performed a selection for the points of maximum variation (phigh). See Figure S4B for an
illustration of the entire process.
To determine the level of noise, we first generated a copy of the image to which a three pixel median filter had been applied. We
then subtracted this median image from the original to generate a noise dominated image, I’. At each of the minimum feature points
(where noise is most dominant); the standard deviation of this image was taken over a three pixel square neighborhood. The level of
noise, N, was then calculated as the mean of these standard deviations, i.e.,

0 
N = s I ðplow ± 1Þ : (1)

The background level of the image, B, was determined by taking the mean of these noise-dominated regions (again taking a mean
over the three pixel neighborhood), following on from the assumption of symmetric noise, giving
B = hhIðplow ± 1Þii: (2)

The level of signal was then taken to be the mean of the (absolute) difference of the mean of these three pixel neighborhoods around
phigh, and the background. This resulted in the SNR being given by



 I phigh ± 1  B 
S N= : (3)
N

As most images lack large areas that consist of only resin, this simple background selection is not perfect. Therefore, the SNRs gener-
ated should be considered lower limits in most cases. We show the SNR as a function of the acquisition rate for a variety of EM tech-
niques in Figure 2G.
To validate the feature-based method, we examined the effect of varying resolutions and confirmed that SNR changed in the ex-
pected manner for the additive Gaussian noise model. The feature-based method works reasonably well when combining voxels pro-
ducing SNRs within 20% (1.5 dB) of the expected based on additive Gaussian noise (Figure S4C), although the ATUM-SEM data of
Kasthuri et al. (2015) increases by more than others, a possible sign of their voxels (3 3 3 nm in x-y) under-sampling biological fea-
tures. This method produces the expected increase when scaling down images producing equivalent normalized SNRs (Figure S4C).
Increasing the size of images also increases the SNR, but this is due to the generation of new pixels with similar values to old ones
inside the regions considered for noise due to the fact that creation of these new pixels functions as a pseudo-low pass filter. As
expected this measure reports larger SNR values when Gaussian blurring is applied (as noise is disproportionately removed when
a low pass filter is applied) (Figure S4C). In images generated by super resolution techniques therefore, this method may be inappro-
priate and should be modified to, for example, use distance based regions rather than pixel based regions.
Cell membrane SNR
Although the feature-based SNR measure avoids many human biases in the selection of regions used to calculate background and
signal levels, it unfortunately can often incorporate biological structure (our signal of interest) into these calculations.
We therefore introduced a complementary measure to compare the SNR of biological EM data and verified that the feature-based
SNR calculation is valid. At its heart, this is simply a comparison between the signal level at a cell membrane and the background
nearby, taken at multiple points within an image.

Cell 174, 730–743.e1–e14, July 26, 2018 e12

 This was achieved by a user creating a line inside a random 100 3 100 pixel region which contains only resin and, ideally nearby, a
line which covers only a stained cell membrane. Pixels along these lines were considered to be background or signal, respectively.
After selection of a background and signal line within each region, another 100 3 100 pixel region was chosen, until 20 lines in total
(10 background lines, 10 signal lines) were selected, skipping a region if there was not a suitable location to select both.
The noise was considered to be the standard deviation of the pixel intensities across all background intensities, and the back-
ground level the mean. The signal value was considered to be a mean of the signal pixels.
We show an example of the selection process in Figure S4D. Signal-to-noise ratios found by this method (Figure S4E), were within
10% of those found via the feature-based method, suggesting the former may be used for a fast, bias-free comparison between
methods.

Quantification of Artifacts
In order to examine the quality of the dataset, an assessment of missing and degraded tiles was undertaken. Using the losslessly
compressed tiles, we examined each 1024 3 1024 pixel region from which the CATMAID image tile pyramid is built, and checked
if this region was first, present, and second, if it was degraded due to an artifact in the data. We plot the total number of tiles and
number of degraded tiles per section in Figure S5K. A section was considered to have missing tiles, if its tile count was 5% below
that of the median of the neighboring five sections on each side. Qualitatively, this was the threshold at which biology is noticeably
impacted and not arising due to a tighter ROI around the brain. Within the core region of the brain (sections 2000-6000), 3% of sec-
tions had tile loss above this threshold, with 1% of all sections having extensive tile loss exceeding 20%. In addition to these missing
tiles, a tile was considered degraded if the standard deviation of its grayscale histogram either exceeded 40, or was below 20 and had
a mean grayscale value below 50. These thresholds included small-scale precipitate, folds and intensity correction artifacts in the
former and large-scale precipitate errors in the latter. These degraded tiles comprised 0.5% of the entire dataset with 1% of sections
having over 5% of degraded tiles.

Analysis of Neuronal Geometry

Analysis of neuronal geometry was conducted using custom packages developed in R. We imported skeletons of EM-reconstructed
neurons from the CATMAID tracing environment using rcatmaid (https://github.com/jefferis/rcatmaid). Section thickness in
CATMAID was specified to be 35 nm; all analyses of skeleton geometry therefore use this value. For qualitative and quantitative com-
parison with skeletons of LM-reconstructed neurons, the skeletons of EM-reconstructed neurons were transformed into coordinate
spaces of various LM template brains using elmr based on landmark pairs defined with ELM (see Electron-Light Microscopy regis-
tration tools). The R NeuroAnatomy Toolbox package (nat, https://github.com/jefferis/nat) was used for geometric computations, 3D
visualization of both reconstructed neurons and surface models.
NBLAST search for PNs
The geometric search tool NBLAST (Costa et al., 2016) was used to aid identifications of PN subtypes (Table S1) by comparing
morphological similarity. NBLAST breaks neuron skeletons into segments and computes an NBLAST score based on a geometric
comparison of the segments of a given pair of neurons (a query neuron and a target neuron). NBLAST scores range from 1
(completely anticorrelated) to 1 (identical), but for practical purposes, 0 is a natural cutoff. A score of 0 means that the similarity level
for segment pairs are, on average, as likely to be from neurons of the same type as they are to be from a random pair of neurons in the
database. The EM-reconstructed PNs were transformed into the FlyCircuit template brain space for NBLAST neuron search against
the
400 LM-reconstructed PNs whose subtypes were previously classified (Chiang et al., 2011; Costa et al., 2016). This is enabled
by a single elmr function nblast_fafb. The search functionality is built on the nat.nblast package (https://github.com/jefferislab/nat.
nblast) and uses data distributed with the FlyCircuit package (https://github.com/jefferis/flycircuit), both of which are installed
with elmr. Only EM-reconstructed PNs whose candidate subtypes exist in the FlyCircuit dataset were used for NBLAST search.
The EM-reconstructed skeletons, which typically had many additional fine processes as compared to LM-reconstructed skeletons,
were used as the target neuron rather than the query neuron for geometric comparison, in reverse to the conventional NBLAST
option. NBLAST search of each EM-reconstructed PN returned the top five hits of LM-reconstructed neurons and their NBLAST
scores, which were tabulated to aid expert glomerular identification of PNs. For the VM2 example shown in Figures 3E–3G, the
top NBLAST score for the EM-reconstructed VM2 PN was compared with the top NBLAST scores from searching all LM-recon-
structed VM2 PNs against each other. Further details of the NBLAST neuron search, the associated LM data, and an online web-
app for on-the-fly NBLAST queries are available at http://jefferislab.org/si/nblast.
NBLAST clustering and PN dendrogram generation
Pairwise NBLAST scores were computed for all uniglomerular PNs (the nblast_allbyall function in the nat.nblast package at https://
github.com/jefferislab/nat.nblast) after transformation into the JFRC2 template brain (Jenett et al., 2012) space using elmr. We used
unsupervised hierarchical clustering with Ward’s method based on the NBLAST scores (nat function nhclust).
Renderings and analysis of PN arbors in MB calyx
Reconstructed PNs from two LM datasets, one from the FlyCircuit database (Chiang et al., 2011) and the other from Jefferis et al.
(2007), were previously registered to a common template brain (Costa et al., 2016), and were used for comparisons with EM-recon-
structed PNs reported here. We first determined which PN subtypes had multiple EM- and LM-reconstructed neurons available. For a
few PN subtypes, we analyzed EM-reconstructed PNs although no LM reconstructions were available from the dataset, enabling

e13 Cell 174, 730–743.e1–e14, July 26, 2018

comparison of a more complete set of PN subtypes within the EM modality. The DA3 subtype was excluded because the DA3 PNs
from the LM data do not extend collaterals into the MB calyx (Jefferis et al., 2007). For analysis, we then selected a random set of LM
skeletons so that we had the same number of LM and EM skeletons for each PN subtype.
For visualizations, LM-reconstructed PNs from the FlyCircuit database were used for Figure 5A, and LM-reconstructed PNs from
Jefferis et al. (2007) were used for Figure S7A. Linear interpolation of neighboring skeleton nodes was applied to EM-reconstructed
PNs to smooth artifactual spikes in neuron tracings due to registration errors. EM- and LM-reconstructed PNs were resampled with a
1 mm interval to ensure uniform representation of skeletons. For analysis, LM-reconstructed PNs from both the FlyCircuit database
and Jefferis et al. (2007) datasets were used. For both visualization and analysis, EM- and LM-reconstructed PNs were transformed
onto a common template brain (Costa et al., 2016; Jenett et al., 2012; Manton et al., 2014).
To obtain the subset of PN arbors ramifying in MB calyx for both visualization and analysis, a bounding box defined by the calyx
surface mesh (Ito et al., 2014; Manton et al., 2014; Milyaev et al., 2012) in the JFRC2 template brain was used to trim the skeletons of
EM- and LM-reconstructed PNs. For visualization in Figure 4C, the boutons of PNs (as defined in Tracing of PNs) in MB calyx were
isolated and used to generate bouton volumes by applying a Gaussian smoothing. For analysis of PN arbors in the MB calyx (Figures
5B, 5C, S7B, and S7C), the backbone of each PN was removed so that only the collaterals entering the calyx were included in geo-
metric measurements.
We computed geometric measurements (pairwise distance and NBLAST scores) for only the MB calyx arbors of EM- versus LM-
reconstructed homotypic PNs. For each PN subtype, pairwise distances quantified physical co-location of arbors and NBLAST
scores quantified morphological similarity for pairs of neurons. To calculate the pairwise distance, we iterated over each node of
a query neuron to find the nearest node in a target neuron, measured the Euclidean distances between each pair of nodes, and calcu-
lated the mean nearest distances for all nodes in the query neurons (forward distance). When a query neuron has significantly longer
arbors than a target neuron, large nearest distances can be introduced by end points in the query neuron’s arbors that are far away
from closest nodes of the target neuron’s arbors. To address this issue, we calculated the same mean nearest distance with the query
neuron and target neuron in reverse (reverse distance) and picked only the smaller of the forward and reverse distances, which was
the pairwise distance. To quantify morphological similarity, NBLAST scores were computed for PN arbors in MB calyx in a similar
pairwise manner. The distribution of all pairwise distances and NBLAST scores of homotypic PNs were plotted, and for both mea-
surements the difference between EM and LM population data were analyzed with a Student’s t test.

DATA AND SOFTWARE AVAILABILITY

All files and videos are available through the following website: http://www.temca2data.org.

ADDITIONAL RESOURCES

Access to the full adult fly brain dataset is available at: http://www.temca2data.org.
Analysis code is available at: https://github.com/bocklab/temca2data.

Cell 174, 730–743.e1–e14, July 26, 2018 e14

Supplemental Figures

Figure S1. Neuronal Architecture of the MB Calyx, Related to Figure 1

(A) Micrograph of a microglomerulus in MB calyx. A canonical olfactory PN bouton (pink) is presynaptic to several fine KC dendrites, forming a synaptic complex
referred to as a microglomerulus (Yasuyama et al., 2002). Arrows: presynaptic release sites.
(B) Schematic of microglomerular inputs to KCs in MB calyx of Drosophila. The PN axons extend collaterals from the mALT into the calyx and provide bouton
inputs to KCs. The Drosophila MB has
2,000 KCs on each side of the brain (Aso et al., 2014). Each KC projects a highly variable dendritic arbor into the calyx,
which terminates in elaborations known as claws. Claws from many KCs wrap individual PN boutons to form each microglomerulus (Yasuyama et al., 2002), and
each KC receives input from multiple PNs of diverse types across its claws (Caron et al., 2013; Gruntman and Turner, 2013). The complete composition of cell
types that provide driving inputs via microglomeruli in the calyx is unknown. KCs have been shown to form presynaptic release sites in the calyx mostly outside of
claws (Butcher et al., 2012; Christiansen et al., 2011), but the complete set of postsynaptic partners is unknown.
Scale Bar: 1 mm in (A).
Figure S2. TEMCA2, Fast Stage, and ATPS for High-Throughput EM Imaging, Related to Figure 1
(A) A TEMCA2 equipped with a Fast Stage. Arrowheads: Fast Stage (black); elongated vacuum chamber (white); 2 3 2 camera array (red).
(B) Fast Stage Schematic. Upper, driven mass (left) and exterior view (right). Lower, cutaway of Fast Stage showing the locations of dampers, bearings, and
vacuum bellows. Arrows: rolling element damper locations (black arrows); rolling-element ‘tip’ bearing (white arrowhead); vacuum bellows (black arrowhead).
(C) Plot of Fast Stage motion over time following an 8 mm move. Top trace, vibration perpendicular to Fast Stage motion axis. Bottom trace, vibration along the
Fast Stage motion axis.
(D) Schematic of Fast Stage stepping pattern. Left, small-step/big-step schematic. Numbers indicate camera identity within the array. Right, CompuStage and
Fast Stage scanning axes. Red point is origin of scanning.
(E) ATPS (white arrowhead) mounted to an accessory port on an FEI Tecnai Spirit BioTWIN TEM.
(F) Schematic of the ATPS system diagraming motor positions and movement axes as well as vacuum and pneumatics subsystems.
Figure S3. Sample Preparation and EM Acquisition Infrastructure, Related to Figure 1
(A) Drosophila brain following en bloc staining.
(B) Frontal view of a 3D volumetric rendering of an X-ray tomogram from the embedded Drosophila brain.
(C) Sample support film test showing a failed result with wrinkling of the support film on a 3 mm grid with 2 3 1 mm slot.
(D) Sample support film test showing a successful result with no wrinkling or relaxation of the support film.
(E) Left, schematic of TEMCA2 vacuum extension, scintillator, and camera array. Right, camera diagram showing the non-overlapping FOV of each camera on the
scintillator.
(F) An FEI CompuStage-compatible single-axis Fast Stage.
(G) A Fast Stage grid holder.
(H) A 3 mm grid with a 2 3 1 mm slot, custom-etched fiducial marks, 2D barcodes, and unique serial identifier.
(I) Cassettes, magazine, and four-axis stage inside the ATPS vacuum.
(J) ATPS cassettes and magazines.
(K) Grid holder and end-effector of the ATPS grid positioning system. Arrows: prism and LED assembly (red); sample grid (black); lever of the grip assembly (white)
which actuates grid release.
(L) ATPS end-effector with LED lights for machine vision-guided pick-and-place.
(M) Four-axis ATPS stage. Arrows: transverse precision piezo-driven axis (blue); pitch-axis pivot point (red); grid positioning system shuttle piezo motor (black);
end effector and vision-system camera (white).
(N) Rotational aligner integrated into the ATPS cassette shuttle orients the grids for imaging.
Scale Bars: 250 mm in (A) and (B).
Figure S4. Comparison of SNR between EM Imaging Methods, Related to Figure 2
(A) Sample images from a variety of EM datasets acquired via different techniques. The data sources are as in Figure 2G. The top row shows images of side length
3 mm while the lower row shows 100 3 100 pixel subimages of each. Red squares indicate the areas of the subimages.
(B) From left to right, a TEMCA2 image, the key-points detected in the image, convolution of the key-points illustrating dense and sparse feature regions (purple –
low, yellow – high), the region of sparse features selected from the TEMCA2 image showing a resin filled area suitable for noise calculation.
(C) The normalized SNR versus acquisition rates of a variety of EM techniques are shown for different SNR methods. Color code, points and data sources are as in
Figure 2G. From left to right, the feature-based method is as described in (B); for the stacked voxels method, voxels are combined across a layer (SBEM not
shown due to unclear alignments) and across 50 random images; for the reduced resolution method, voxels correspond to a larger physical size across 100
random images; for the scaled up resolution method, voxels correspond to a smaller physical size across 100 random images; for the Gaussian Blur method,
voxels have been blurred with a Gaussian filter across 100 random images. Error bars indicate SD.
(D) Cell membrane SNR method. Left, a representative image used to select two lines of pixels for quantifying signal (red line) and noise (green line), respectively.
The pixels used for signal quantification were selected from cell membranes, and pixels used for noise quantification were selected from areas that contained only
resin. Right, the grayscale values for signal (red) and noise (green) pixels selected in each region.
(E) Normalized SNR versus acquisition rates as determined via the cell membrane method across five random images from each technique, each of which had 10
regions of background/noise and signal determined. Color code, points, and data sources are as in Figure 2G. Error bars indicate SD.
Figure S5. Re-imaging Synapses in MB Pedunculus, Montaging, 2D Intensity Correction, and Assessment of Volume Quality, Related to
Figure 2
(A–C) Matching FOVs in section 3887 from the whole-brain volume (A) and re-imaged at higher resolution in (B-C). Resolutions in (A) and (B) and (C) are 4 nm/pixel
and 0.5 nm/pixel, respectively.
(D, G, and I) Whole-brain sections.
(D–F) Registration of images acquired with high-dose and low-dose current beams. Debris present on a section (red rectangle) necessitated collection of a small
subset of tiles at lower dose than the remainder of the mosaic. Red rectangle indicates the subregion displayed in (E).
(E) A higher magnification image of the debris and border of the low-dose mosaic indicated in (D).
(F) The boundary (arrowheads) in the joined high-dose and low-dose montage.
(G–J) Mosaic of the same section prior to (G-H) and after (I-J) 2D intensity correction. Red squares in (G) and (I) indicate the subregions shown in (H) and
(J), respectively. Intensity differences visible in (H) are greatly diminished in (J).
(K) Most sections have few lost or degraded tiles. Red line, the running median of the total number of tiles per section for an 11-section window (five either side).
For sections with lost tiles, only those with tile loss more than 5% below the median are shown. Triangles indicate complete loss of three consecutive sections. For
sections containing degraded tiles, only those with 100 or more tiles contaminated by artifacts are shown. The dense data points and the fluctuations of running
medians toward both ends of the series (sections before
1500 or after
5800) are due to tiles that contain extraneous tissue or resin outside the neuropil
compartment. The tile count per section (running medians) across the series is proportional to the cross-sectional area of the brain normal to the cutting direction
(z axis).

(legend continued on next page)

(L) Volume alignment quality is sufficient for neural reconstruction. Alignment quality was assessed by analyzing the section-by-section displacements of in-
dividual Gaussian-smoothed skeletons of reconstructed neurons (STAR Methods). Large displacements are generally indicative of section misalignments. Data
outside of a core region of the brain (sections <
500 and >
5500, shaded regions) are not informative since: 1) these regions mostly contain somata which
typically have larger diameters than neurites, resulting in increased variability of tracing node placements, and 2) not enough tracing exists outside this range. The
median and the 95% percentile of the displacements are 0.09 mm and the 0.57 mm, respectively.
Scale Bars: 200 nm in (A)–(C), 100 mm in (D), (G), and (I), 25 mm in (E), 1 mm in (F), 2 mm in (H) and (J).
Figure S6. Reproducibility of Tracing, Related to Figure 2
(A and B) Three teams (indicated by colors), each comprising one tracer and one proofreader, reconstructed the same KC, with each team blinded to the others.
(A) Morphologies are comparable across teams. Upper panel: asterisks indicate locations of a KC claw neurite postsynaptic to a PN input discovered by two out
of three teams. Zoom-ins show the discrepancy of the fine claw neurites. Lower panel: cable length of missed branches for each of the three teams compared to
an expert-generated gold-standard skeleton (STAR Methods). Consistent with Schneider-Mizell et al. (2016), the vast majority of our missed branches have a
cable length of fewer than 5 mm. The reconstructed KC dendrite spanned serial sections 3451-4899. This range included one 3-section loss (3595-3597), and four
single-section losses (3715, 4192, 4353, 4474), demonstrating the robustness of reproducible traceability to occasional data loss.
(B) Synaptic annotations are comparable across teams. The KC (gray circle) receives input (arrows) from seven PN boutons (pale orange circles). The number of
input synapses is indicated for each bouton, with the same team colors as in (A). All PN bouton inputs, except the rightmost one, were recapitulated by the three
teams. Red asterisks mark the discrepant inputs caused by the missed branches in (A).
Scale bars:
20 mm in (A), 250 nm in (A) inset.
Figure S7. Comparison of MB Calyx Collaterals Reveals Greater Similarity between EM-Reconstructed PNs than LM-Reconstructed PNs,
Related to Figure 5
(A) Pairs of EM-reconstructed PNs are qualitatively more similar than pairs of LM-reconstructed PNs in the MB calyx. Subtypes are ordered by how much more
clustered EM-reconstructed PNs are than LM-reconstructed PNs (STAR Methods). Middle column shows the pair of LM-reconstructed PNs with the median
pairwise distance across all pairs.
(B) Pairwise NBLAST scores between homotypic PN collaterals in the MB calyx. Each data point represents the NBLAST scores between one pair of EM- (red) or
LM-reconstructed (blue) PNs. Data points are bucketed according to PN subtype; subtypes are ordered on the x axis by how much more similar EM-re-
constructed PNs are than LM-reconstructed PNs (STAR Methods).
(C) Histogram of all data points in (B). The mean of pairwise NBLAST scores for all EM-reconstructed PN subtypes was significantly higher than that for all LM-
reconstructed PN subtypes (0.56 ± 0.18 versus 0.35 ± 0.21, respectively; Student’s t test, p < 2.2 3 1012), indicating that EM-reconstructed PN subtypes are
morphologically more similar to each other than LM-reconstructed PN subtypes.
Scale bar:
10 mm in (A).