In situ transcription profiling of single cells reveals spatial
organization of cells in the mouse hippocampus
Sheel Shah
2,3,5
,
Eric Lubeck
2,3,5
,
Wen Zhou
1
, and
Long Cai
1,4
1
Division of Chemistry and Chemical Engineering, California Institute of Technology, Pasadena,
CA, USA 91125
2
Division of Biology and Biological Engineering, California Institute of Technology, Pasadena, CA
91125, USA
3
UCLA-Caltech Medical Scientist Training Program, David Geffen School of Medicine, University
of California at Los Angeles, Los Angeles, CA, USA 90095
Summary
Identifying the spatial organization of tissues at cellular resolution from single cell gene
expression profiles is essential to understanding biological systems. Using an in situ 3D
multiplexed imaging method, seqFISH, we identify unique transcriptional states by quantifying
and clustering up to 249 genes in 16,958 cells to examine whether the hippocampus is organized
into transcriptionally distinct subregions. We identified distinct layers in the dentate gyrus
corresponding to the granule cell layer and the subgranular zone and contrary to previous reports,
discovered that distinct subregions within the CA1 and CA3 are composed of unique combinations
of cells in different transcriptional states. In addition, we found that the dorsal CA1 is relatively
homogenous at the single cell level, while ventral CA1 is highly heterogeneous. These structures
and patterns are observed using different mice and different sets of genes. Together, these results
demonstrate the power of seqFISH in transcriptional profiling of complex tissues.
Introduction
The mouse brain contains ~10
8
cells arranged into distinct anatomical structures. While cells
in these complex structures have been traditionally classified by morphology and
electrophysiology, their characterization has been recently aided by gene expression studies.
In particular, the Allen Brain Atlas (ABA) provides a systematic gene expression database
using in situ hybridization (ISH) of the entire mouse brain one gene at a time (
Dong et al.,
2009
;
Fanselow and Dong, 2010
;
Thompson et al., 2008
). This comprehensive reference
*
Lead Contact: lcai@caltech.edu.
4
Lead Author
5
Co-First Authors
Publisher's Disclaimer:
This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our
customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of
the resulting proof before it is published in its final citable form. Please note that during the production process errors may be
discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
Author Contributions:
L.C., S.S., E.L. designed the experiments. W.Z., S.S., E.L. performed the experiments. L.C., S.S., E.L.
analyzed the data and wrote the manuscript.
HHS Public Access
Author manuscript
Neuron
. Author manuscript; available in PMC 2017 October 19.
Published in final edited form as:
Neuron
. 2016 October 19; 92(2): 342–357. doi:10.1016/j.neuron.2016.10.001.
Author Manuscript
Author Manuscript
Author Manuscript
Author Manuscript
provides regional gene expression information, but lacks the ability to correlate the
expression of different genes in the same cell. More recently, single cell RNA sequencing
(RNA-seq) has identified many cell types based on gene expression profiles (
Darmanis et al.,
2015
;
Tasic et al., 2016
;
Zeisel et al., 2015
). However, while single cell RNA-seq provides
useful information on multiple genes in individual cells, it has relatively low detection
efficiencies and requires cells to be removed from their native environment resulting in the
loss of spatial information. These different approaches can lead to contradictory descriptions
of cellular organization in the brain and other biological systems.
In the hippocampus, recent RNA-seq data suggests that the CA1 region is composed of cells
with a continuum of expression states (
Cembrowski et al., 2016
,
Zeisel et al 2015
), while
ABA analysis indicates that sub-regions within the CA1 have distinct expression profiles
(
Thompson et al, 2008
). To resolve the two conflicting descriptions of hippocampal
organization, a method to profile transcription
in situ
in the hippocampus with single cell
resolution is needed. Here, we demonstrate a general technique that enables the mapping of
cells and their transcription profiles with single molecule resolution in tissue, allowing an
unprecedented resolution of cellular transcription states for molecular neuroscience (Fig
1A).
A great deal of progress has been made recently in developing highly quantitative methods
to profile the transcriptome of single cells. Building upon single molecule fluorescence
in
situ
hybridization (smFISH) (
Femino et al., 1998
;
Raj et al., 2006
;), Lubeck et al. devised a
general method to highly multiplex single molecule
in situ
mRNA imaging irrespective of
transcript density using super-resolution microscopy (
Betzig et al., 2006
;
Rust et al., 2006
;
Lubeck and Cai, 2012
;). However, the spectral barcoding methods used in these previous
works is difficult to scale up beyond 20–30 genes because of the limited number of
fluorophores (
Fan et al., 2001
;
Lubeck and Cai, 2012
).
To overcome the scalability problem, a temporal barcoding scheme was developed that uses
a limited set of fluorophores and scales exponentially with time (
Lubeck et al., 2014
).
Specifically, sequential probe hybridizations on the mRNAs in fixed cells impart a unique
pre-defined temporal sequence of colors, generating an
in situ
mRNA barcodes. The
multiplex capacity scales as F
N
, where F is the number of fluorophores and N is the number
of rounds of hybridization. Thus, one can increase the multiplex capacity by increasing the
number of rounds of hybridization with a limited pool of fluorophores. We called this
approach Sequential barcoded Fluorescence i
n situ
Hybridization (seqFISH) (
Lubeck et al.,
2014
). In parallel,
in situ
sequencing methods were developed to directly sequence
transcripts in tissue sections, but these methods suffer from low detection efficiency (<1%)
(
Ke et al., 2013
;
Lee et al., 2014
). Recently, Chen et al. expanded the error correction
method in the original seqFISH demonstration by using a Hamming distance 2 based error
correcting barcode system, called merFISH. However, this implementation requires larger
transcripts (>6kb) and many more rounds of hybridization than the method described here
(
Chen et al., 2015b
). Furthermore, seqFISH and its variants have only been applied in cell
culture systems due to the difficulty of smFISH detection in tissue. Here, we demonstrate an
improved version of seqFISH in complex tissues by including signal amplification and a
Shah et al.
Page 2
Neuron
. Author manuscript; available in PMC 2017 October 19.
Author Manuscript
Author Manuscript
Author Manuscript
Author Manuscript
time-efficient error correction scheme (Fig 1A–D, Table S1), allowing us to resolve the
structural organization of the hippocampus with single cell resolution.
Results
Signal amplification and error correction enable robust detection of mRNAs in tissues
To overcome the autofluorescence and scattering inherent to brain tissues, we used an
amplified version of smFISH, called single molecule Hybridization Chain Reaction
(smHCR) (Fig 1E) (
Choi et al., 2014
,
Shah et al., 2016
). Single molecule HCR amplified
signal 22.1 ± 11.5 (mean ± s.d., n=1338, Fig S1B) fold compared to smFISH, enabling
robust and rapid detection of individual mRNA molecules in tissues and facile alignment of
spots between hybridizations (Fig 2A). Single transcripts can be detected and localized in
3D with just 24 probes in tissues, enabling detection of transcripts <1kb in size, with a
fidelity comparable to the smFISH gold standard (Fig S1C–D) but with signals 20-fold
brighter (
Shah et al., 2016
). Single molecule HCR DNA polymers can also be digested by
DNAse and re-hybridized in brain slices, allowing HCR-seqFISH to be robustly
implemented (Fig 2A). We note the smHCR enables true 3D imaging in tissues, whereas the
previous sequential FISH demonstrations (
Lubeck et al., 2014
,
Chen et al., 2015
) were
performed only in flat cell cultures.
Furthermore, we improved upon our existing barcode system by implementing a time-
efficient error correction scheme. The major source of error in seqFISH is the loss of signal
due to mis-hybridization, which increases with the numbers of hybridization. We introduced
an extra round of hybridization to correct loss of signal during any round of hybridization
(Fig 1D) (Supplementary Text). By minimizing the number of hybridizations, this error
correction scheme is efficient to implement. For example, using 5 fluorophores and 4 rounds
(instead of 3 rounds) of hybridization to code for 125 genes, we can still uniquely assign
barcodes to genes even when signal from any single round of hybridization is missing.
Although merFISH can tolerate 2 errors in the barcodes, it requires 16 rounds of
hybridization to code 140 genes (
Chen et al. 2015
). As increasing the number of
hybridizations can potentially lead to more experimental error and analysis complexity, our
simple error correction method corrects for the most common error, dropped signal. Also,
the fewer rounds of hybridizations decrease the total imaging and experimental time, which
is rate-limiting for tissue experiments. HCR-seqFISH with simpler error-correction scheme
allows efficient and accurate quantification of transcription profiles in tissues.
Using this HCR-seqFISH method, we surveyed the regional and sub-regional transcriptional
heterogeneity within the temporal and parietal cortex and hippocampus of the mouse brain
by imaging similar coronal sections collected from 3 different animals. Two similar sections
from separate mice were profiled with probes for 125 genes, while one additional brain slice
was imaged for 249 genes. In each of the coronal slices, between 60–80 fields of view were
imaged, each 216μm × 216μm × 15μm, in the cortex and hippocampus (Fig 1A and S1E).
For the 125 gene set, 56 of the genes (Fig 1D, Table S1) were selected because they showed
spatially heterogeneous expression based on the ABA (
Lein et al., 2007
), another 44 were
selected from a list of transcription factors, and 25 marker genes were selected from single
cell RNA-seq datasets (
Zeisel et al., 2015
). One hundred of these genes were barcoded by 4
Shah et al.
Page 3
Neuron
. Author manuscript; available in PMC 2017 October 19.
Author Manuscript
Author Manuscript
Author Manuscript
Author Manuscript
rounds of hybridization (Fig 1B). The remaining 25 high abundance genes were measured
individually using 5-color smHCR in 5 serial rounds of hybridizations (Fig 1C). This hybrid
approach of measuring medium expression genes with barcoding seqFISH and high copy
number genes serially in subsequent hybridizations allows a large dynamic range of
transcripts to be profiled in the same cell.
seqFISH is an accurate and efficient method to multiplex RNA in situ
To determine the accuracy of the seqFISH method in quantifying mRNA levels in single
cells in tissue, we compared the copy number of 5 of the 100 target genes measured by
barcoding to the copy number found by smHCR detection in the same cell (Fig 2B, S2A) in
15μm brain sections. We found that the copy number of the RNAs per cell as measured by
barcoding and smHCR agreed with an R-value of 0.85 and a slope of 0.84 (N=3851). As
smHCR matches smFISH transcript quantitation (
Shah et al., 2016
), the barcoded seqFISH
method can quantify mRNA molecules in single cells with 84% efficiency compared to the
gold standard of smFISH. In comparison, single cell RNA-seq measurements are 5–20%
efficient based on spike-in controls and
in situ
sequencing is less than 1% efficient
(
Darmanis et al., 2015
;
Klein et al., 2015
;
Lee et al., 2014
;
Macosko et al., 2015
;
Tasic et al.,
2016
;
Zeisel et al., 2015
;
Ståhl et al., 2016
). This high efficiency of detection results from a
low transcript drop rate and a high barcode recovery rate due to the error correction round of
hybridization. In our experiment, 78.9% of barcodes (N=2,115,477 barcodes) were found in
all 4 hybridization rounds and 21.1% were identified in 3 out of the 4 hybridizations (Fig
2C), indicating that the probability of detecting a given mRNA molecule is 94% in each
round of hybridization (Fig S2B).
To quantify the amount of false positive signal due to misalignment of barcodes and
nonspecific binding of probes, we measured the amount of off-target barcodes that were
detected. With four rounds of hybridizations and 5 fluorophores, there were 5
4
=625 unique
codes. We assigned 100 of these barcodes to measure mRNAs detected at 914.8 ± 570.5
counts per cell (mean ± s.d., N=3439). In comparison, the 525 remaining off-target barcodes
that were not used were detected at 4.6 ± 4.7 (mean ± s.d., N =3439) counts per cell (Fig
2D). False positives, due to chance alignment of nonspecifically bound spots, contributed
minimally to the barcode readouts because of this three order of magnitude difference in
detected barcodes (on target vs. off target). The false positives we observe fall only on
barcodes hamming distance one away from on-target barcodes, yet minimally contribute to
undercounting on-target barcodes (Fig 2E). Furthermore, even the most frequent off-target
barcode was observed 65.57 times less frequently than the most infrequent mRNA coding
barcode (Fig 2E, S2). Even though during each round of hybridization, 24.8 ± 0.4% (mean ±
s.e., N=4 rounds of hybridization) of the spots were nonspecifically bound probes, barcode
miss-assignments did not occur frequently because non-specifically bound probes do not
reappear in the same location after digestion with DNAse and re-hybridization (Fig 2A).
Together the quantifications of false positive and false negative barcodes demonstrate that
this method is highly efficient and accurate at detecting RNAs
in situ
in single cells within
tissues.
Shah et al.
Page 4
Neuron
. Author manuscript; available in PMC 2017 October 19.
Author Manuscript
Author Manuscript
Author Manuscript
Author Manuscript
Cell clusters are based on combinatorial expression profiles
We imaged the expression of 125 genes in coronal sections from two mice for a total of
14,908 cells (Fig S1E). Cortical and hippocampal cells were segmented based on DAPI and
Nissl staining. A tessellation algorithm was developed to accurately segment densely packed
cells in the hippocampus. To avoid capturing mRNA from neighboring cells, we contracted
by 10% the borders of cells determined by the segmentation algorithm.
To group the single cell data into distinct transcriptional states, we Z-score normalized the
copy number of each transcript in every cell (Fig 3A) and hierarchically clustered the cells to
identify cells with similar expression patterns (Table S2, Fig S3). While these clusters do not
necessarily represent canonical cell types, many of these clusters contain clear
transcriptional markers of known cell types previously identified by single cell RNA-seq
(Fig 3B) (
Zeisel et al., 2015
,
Tasic et al 2016
). Cell clusters 12 and 13 contained clear
expression of
Gja1
which marks out astrocytes (
Zeisel et al., 2015
,
Tasic et al 2016
). Cluster
12 also expresses
Mfge8
while cluster 13 did not, indicating two distinct subpopulations of
astrocytes (Fig 3B). There are further subclusters within each of the astrocyte populations
with different spatial localization patterns (Fig S3C–E). Cluster 11 cells expressed
Laptm5
, a
known microglia marker (
Zeisel et al., 2015
,
Tasic et al 2016
). Cluster 3 expressed
interneuron genes while cluster 1–2 and 4–5 expressed genes associated with pyramidal
neurons (
Zeisel et al., 2015
,
Tasic et al 2016
). The major clusters were robust to down-
sampling the number of cells used in clustering (Fig S4), with some of the hippocampal
pyramidal and glial clusters robustly defined even with 400 cells. Similarly, principal
component analysis (PCA) visualization of the data (Fig S3H) recapitulated the major
clusters that correspond to astrocyte, microglia, cortical pyramidal, hippocampal pyramidal,
dentate gyrus (DG) granule, and interneuron cells.
As the cluster distance between different cells is proportional to the number of differentially
expressed genes in the target list, an unbiased clustering of the 125 gene data without
weighting specific genes should not be interpreted directly as canonical “cell types,”, but
rather as grouping cells with different patterns of genes expression based on the current
target list. We will refer to some of these clusters as pyramidal neurons or astrocytes for ease
of notation, but strictly speaking, they are cells clusters with similar expression patterns as
neurons or astrocytes.
Cell clusters show distinct regional localization
Many neuronal clusters mapped to distinct regions in the brain (Fig 3B). Several classes of
pyramidal cells (cluster 1–2) showed exclusive localization to the hippocampus, while other
classes (4–5) showed predominantly cortical localization. There were also a class of cells
(cluster 7) that were almost exclusively present in the DG. Interestingly, these clusters
segregated based solely on gene expression profiles without adding any spatial information
into the clustering algorithm. These differences in transcriptional states of neurons could be
due to intrinsic differences in the cells or due to different local environment and activity
patterns.
Shah et al.
Page 5
Neuron
. Author manuscript; available in PMC 2017 October 19.
Author Manuscript
Author Manuscript
Author Manuscript
Author Manuscript
In contrast, astrocyte, microglia and other non-neuronal cell clusters were generally
uniformly present in all areas of the brain (Fig 3B). However, subclusters of astrocytes did
localize to different regions of the brain preferentially (Fig S3E), with subcluster 12.3
localized preferentially to the cortex, while 12.1 subcluster was uniformly distributed.
Similarly, cluster 9 cells contain subclusters (9.3, 9.5 and 9.6) that localize exclusively to the
DG, while other subcluster (9.1) localize almost exclusively to the cortex. The regional
localization of neurons are especially pronounced with cluster 1 and 2 localized almost
exclusively to the hippocampus, with some of the subclusters localized predominantly to the
CA3. Furthermore, while pyramidal cell clusters 4 and 5 are preferentially cortically
localized, the few hippocampal cells in these clusters form their own subclusters (4.4 and
5.4) (Fig S3E). In cluster 6 cells, many subclusters with distinct expression profiles are
localized almost exclusively in the CA1, CA3 or the DG (Fig S3C). In contrast, cluster 7
cells show a relatively homogenous regionalization pattern, but further subdivide based on
combinatorial expression patterns (Fig S3D). Subclusters of cluster 9 also show significant
regionalization where subclusters 9.1, 9.3, 9.5, and 9.6 show localization to the SGZ (Fig
S3E). Overall, cell clusters with similar expression profiles exhibited similar spatial
localizations across the brain with a correlation coefficient of 0.67 (Fig S3G), indicating the
existence of archetypal regional expression patterns and potential spatial markers in the
brain. These results show that the tissue-optimized HCR seqFISH approach can directly
identify a variety of transcriptional states and quantify broad spatial patterns of expression.
Combinatorial expression patterns define fine clusters
While certain cell clusters contain strong expression of marker genes, not all clusters are
defined based on a few genes. How much power do individual genes or groups of genes have
in explaining the observed cell clusters? To understand this, we examined whether subsets of
genes can recapitulate the observed clusters (Fig 3C–D). We found that any set of 25 genes
recovers about half of the correlation structure in the cell-to-cell correlation map (Fig 3C,
S3I, S4, N=10 bootstrap replicates). The fact that the selection of any 25 genes can explain
the gross patterns in the data is likely due to the high correlations amongst the expression
patterns of genes, as shown in the gene-to-gene correlation map (Fig S3J). Thus, a small
subset of the measured genes can provide sufficient information to infer the gross
transcriptional states of the cells. Interestingly, this may be the same reason why low-
coverage single cell sequencing methods such as drop-seq and inDrop (
Klein et al., 2015
;
Macosko et al., 2015
) can capture the large distinction of cell types, because many highly
expressed genes are correlated to other genes that collectively define cell types.
At the same time, the finer correlation structure in the data, required to define the cell
clusters accurately, can only be captured with accurate quantitation of many genes (Fig 3C–
D). Consistent with this, using a “random-forest” machine learning algorithm (
Breiman,
2001
) to classify cell clusters, we found that 75 genes are needed to classify cells with 50%
accuracy, indicating that correct cluster assignment requires more detailed information from
many genes (Fig 3C). Supporting this view, the first 10 principal components (PC) explained
59.5% of the variation in the data, while the rest of the variation required the remaining 115
PCs (Fig 3D, S3F). The “random forest” algorithm required 10 PCs to predict the cell cluster
assignments with 50% accuracy (Fig 3D), but accuracy steadily increased with more PCs.
Shah et al.
Page 6
Neuron
. Author manuscript; available in PMC 2017 October 19.
Author Manuscript
Author Manuscript
Author Manuscript
Author Manuscript
These observations indicated two levels of information in the data: a coarse level, where
large distinctions in cell clusters are observable by a few genes, and a fine level, where
subtle distinctions require many more genes.
These results suggest two points experimentally. First, multiplexing at the level of 20 genes
by seqFISH can give broad cell cluster identification that is not available with 2–3 gene
smFISH experiments. Although single marker genes are useful for inference, we find that
they frequently are not sufficient for cell classification. For example, all DG specific granule
cells (clusters 7) have
Gpc4
and
Vps13c
as their enriched marker genes (Fig 3B); yet,
Gpc4
and
Vps13c
are also strongly expressed in other hippocampal cells outside of the DG, as
seen in both our experiments and the ABA. Thus, smFISH against
Gpc4
and
Vps13c
alone
would not be sufficient to uniquely identify the DG granule cells. Furthermore, even the
strongly bimodal markers that are known to define cell types (i.e.
Mgfe8, Gja1
, etc.) are
correlated enough to overall expression profiles that cells fall into the appropriate cluster
even when these genes are excluded. This point suggests that while marker genes can be
essential in assigning a cell to a known cell type, they are not necessary to identify unique
clusters in the dataset provided enough measurements are made.
Second, accurate measurement of combinatorial expression of many genes enabled by
seqFISH can allow for more specific cell cluster identification. As a comparison, in single
cell RNAseq data, CA1 pyramidal cells are clustered into a single cluster (
Zeisel et. al, 2015
;
Habib et. al 2016
) potentially because of the relatively lower detection efficiency. In our
seqFISH experiments, measuring hundreds of genes quantitatively, we can resolve several
clusters and subclusters with robust regionalization within the CA1 (Fig 3B, S3C–E).
Cells are patterned in the dentate gyrus
To further visualize the spatial organization of cells, we mapped cluster definitions of cells
back into the images. In the DG, we observed a striking lamina layering of cell classes. The
two blades of the DG (Fig 4A–B) showed mirror arrangements of cells, with cluster 9 cells,
forming the subgranular zone (SGZ), leading into a granule cell layer (GCL) dominated by a
single cluster of granule cells (cluster 7) (Fig 3B). In the 125 gene data set, the cells of the
GCL were found to be dominated by expression of Gpc4 and Vps13c matching ISH data
from the ABA (Fig S8B). Cluster 7 was found to be further subdivided into 6 subclusters
(Fig S3D). These subclusters were found to have varying levels of calbindin D-28K (
Calb1
)
expression which is known to increase with granule cell maturation (
Yang et al., 2015
). On
the other hand, the cells of the SGZ were found to be significantly enriched in astrocyte
markers such as
Mfge8
and
Mertk
, which has been also been observed previously (
Miller et
al, 2013
) and in the ABA data (Fig S8A). However, these cells do not cluster with typical
astrocytes (cluster 12 and 13) because their combinatorial expression patterns are different
from astrocytes, consistent with their classification as a completely different population of
cells.
In the fork region of the DG, the layer of cluster 9 cells appeared on the interior surface of
the fork, followed by a layer of granule cells (cluster 7) (Fig 4C). A different layering
pattern is seen at the crest of the DG, where astrocytes, microglia, and some other glial cells
line the exterior of the crest ensheathing the GCL (Fig 4D). In both brains of the 125 gene
Shah et al.
Page 7
Neuron
. Author manuscript; available in PMC 2017 October 19.
Author Manuscript
Author Manuscript
Author Manuscript
Author Manuscript
experiments, the same cell clusters and spatial arrangements are observed. Furthermore,
because the mRNAs are imaged in 3D in the 10–15um brain slices, we can obtain a 3D view
of the expression profiles, shown in the fork regions of the DG (Fig 4F).
Distinct regions of CA1 and CA3 are composed of different combination of cell clusters
While each region of the DG contains similar compositions of cells, distinct subregions
within the CA1 and CA3 contained different combinations of cell classes (Fig 5, S6F). In the
CA1, there were 3 distinct regions defined by their individual cellular compositions. In the
dorsal region of CA1 (CA1d), neuron cluster 6 (enriched in
Nell1
, a protein kinase C
binding protein) (Table S3) was the major cell type in the pyramidal layer, with astrocyte,
microglia and other cells (clusters 10–13) intercalating into the stratum pyramidale (SP) (Fig
5A–C). Transitioning into the CA1 intermediate region (CA1i) (Fig 5D), pyramidal cell
cluster 4 displaced cell cluster 6 as the dominant cell, with the co-appearance of cluster 1
and 2 pyramidal cells.
As the middle of the CA1i region was reached, a small amount of cluster 4 pyramidal cells
remain, while cluster 1 and 2 pyramidal cells dominate (Fig 5E–F). Cluster 1 and 2 are
enriched in
Nell1
(EGF like protein),
Npy2r
(neuropeptide Y receptor),
Slc4a8
(sodium
bicarbonate transporter) and
B3gat2
(glucuronosyltransferase) (Table S2). The CA1i region
displayed a characteristic spatial organization where glial cells line the outermost regions,
while pyramidal cell cluster 1 and 2 longitudinally partitioned the pyramidal layer. This
separation of the inner versus the outer layers of CA1 matches those observed in previously
(Dong et al., 2008). Furthermore, interneurons (cluster 3) were found to preferentially line
the inner edge of the pyramidal layer in the CA1i region (Fig 5E–F). This patterning of
interneurons, particularly subcluster 3.1 cells which were enriched in Slc5a7, a choline
transporter, was consistent with the patterning of cholinergic interneurons observed with
ChAT-GFP labeling (
Yi et al., 2015
). Finally, the largest amount of heterogeneity in the CA1
was seen in the ventral CA1 region (CA1v), where cell clusters 3, 5, and 10 began to mix in
with clusters 1 and 2 (Fig 5G–I).
Similarly, the CA3 was found to have four transcriptionally distinct regions with different
pyramidal cell compositions and abrupt transitions. The ventral most region of CA3
contained a high level of heterogeneity of pyramidal cell clusters (Fig 5J–K), while the
intermediate region of CA3 contain a mixture of cell clusters 1 and 2 (Fig 5L–M). As the
CA3 progressed towards the hilus of the DG, the cell types transitioned first to primarily
cluster 4 neurons (enriched in
dcx
, doublecortin, and
Col5a1
, a collagen), and then to almost
exclusively cluster 6 neurons in the region most proximal to the DG hilus (Fig 5O–P). It is
interesting to note that while cluster 6 cells appear in both the CA1 (subcluster 6.8) and CA3
(subclusters 6.1 and 6.4), sub-clusters of 6 show distant regional localization (Fig S3E),
suggesting that the gene expression differences in CA1 and CA3 cells are captured in the
seqFISH data. We note that similar patterns of homogeneous dorsal and heterogeneous
ventral cell populations are observed when only hippocampal cells are clustered (Figure S5).
The regionalized expression patterns we observed in the hippocampus match closely to those
observed in previous literature (
Thompson et al Neuron 2008
and
Dong et al PNAS 2009
).
For example, CA1d, CA1i, CA1v boundaries correspond to the boundaries shown in Dong
Shah et al.
Page 8
Neuron
. Author manuscript; available in PMC 2017 October 19.
Author Manuscript
Author Manuscript
Author Manuscript
Author Manuscript
et al Fig 2B. In CA3, the subregions observed in our experiment match the CA3 subregion
4–7 in Thompson et al. (
Thompson et al., 2008
).
Lastly, we note that the two slices from two different mice in the 125 gene experiment show
not only the same subregional structure (Fig 4–6), but also the same clusters of cells (Fig 5
and 6) in the different subregions of the hippocampus (Fig S6). In both brains, the CA1d
consists of relatively homogenous population of cluster 6 cells, which transition to a mixture
of 1 and 2 cells in CA1i, and finally to a mixture of 1–6 and 10 cells in the CA1v (Fig S6F).
These results together show that the sub-regions of the hippocampus are a robust feature in
the organization of CA1 and CA3, consisting of cells classes with distinct expression
profiles. The stereotypical nature of the spatial arrangement of these structures suggest
further experiments with seqFISH and other functional assays to probe the distinct functions
of the different cell clusters in the CA1 and CA3.
249 gene multiplex experiments show the same hippocampal subregions
To further show that the sub-regional structure of the hippocampus is independent of the
target genes, we performed a 249 gene seqFISH experiment on a third coronal section. Of
these 249 genes, only 22 genes overlapped with the 125 gene experiment set. For this set of
genes, 214 were selected from a list of transcription factors and signaling pathway
components and the remaining 35 were selected from cell identity markers from another
single cell RNAseq dataset (
Tasic et al, 2016
). The 214 genes were barcoded by 5 rounds of
hybridization, while the remaining genes were imaged in 7 rounds of non-barcoding serial
hybridization. To quantify the efficiency of this experiment, 4 genes in the barcoding set
(
Smarca4, Sin3a, Npas3
, and
Neurod4
) were re-probed with smHCR. The barcoding
efficiency of the 249 gene probe set was found to be 71% with and R value of 0.80 (Fig
S6D). In single cells, we detect on average 2807±1660 (mean±s.d., N=2050 cells) total
barcoded barcodes.
The same arrangement in the DG was observed in the 249 gene experiment, despite different
genes used, indicating robust identification of the layering in the DG by seqFISH (Fig 7S–
T). In particular, the cells in the SGZ are clustered independently from cells in the GCL,
similar to the layers observed in the 125 gene experiment. In the SGZ cells, we observed
enrichment of Sox11, a key transcription factor in neurogenesis (
Miller et al, 2013
). Other
transcription factors involved in neurogenesis, NFIA and Tbr1 are also enriched in the SGZ
cells as seen in our data and the ABA images (Fig S8A). The observations of this distinct
layer in both the 249 and 125 gene experiment and the combined gene enrichment pattern
(increased Sox11, Sox9, NFIA, and Tbr1 in the 249 gene experiment and increased Mertk
and Mfge8 in the 125 gene experiment) suggests that many cells in this layer are involved in
adult neurogenesis in the SGZ. Supplementary figure 7B shows distinctive marker gene
expression in the GCL of the dentate gyrus.
In addition, the same regionalized cellular patterns are observed in CA1d, CA1i, and CA1v,
where different subregions utilize different cell classes in characteristic ratios (Fig S6F). As
seen with the 125 gene experiment, while the CA1d uses only a few cell classes and is
relatively homogeneous, while the CA1v region is made up of many different cell classes
resulting in a high level of cellular heterogeneity. Furthermore, the distinction between CA1
Shah et al.
Page 9
Neuron
. Author manuscript; available in PMC 2017 October 19.
Author Manuscript
Author Manuscript
Author Manuscript
Author Manuscript
and CA3 cell clusters are more clear in the 249 gene experiment suggesting more resolving
power of spatial patterns (Fig 7A–K). The 249 gene experiment also suggests that the CA3
may be composed of 3–4 subregions based on cell cluster composition (Fig 7L–R). The
cellular heterogeneity of the CA3 is again shown to mirror that of the CA1, where the
cellular heterogeneity increases along the dorsal to ventral axis. Cells with distinctive marker
gene expression in the hippocampus are shown in Supplementary figure 7A.
Discussion
Single cell data resolves cellular organizations in the sub-regions of the CA1 and CA3
Two conflicting views of the cell types in the hippocampus have been proposed based on the
analysis of the Allen Brain Atlas data (
Thompson 2008
) as well as recent RNA-seq data
(
Cembrowski et al., 2016
,
Zeisel et al 2015
). Analysis of the ABA in situ data showed that
distinct subregions of the hippocampus expressed different molecular markers, indicating
that the CA1 and CA3 are “regionalized” into distinct sub-structures (
Fanselow and Dong,
2010
;
Thompson et al., 2008
). However, recent bulk RNA-seq experiments on the CA1
found that gene expression patterns changed gradually along the dorsal to ventral axis,
contradicting the sharp boundaries observed in the ABA analysis (
Cembrowski et al., 2016
).
Further supporting this “continuous” cell type view of the hippocampus, analysis of the
single cell RNA-seq data (
Zeisel et al, 2015
) identified a single continuous population of
cells in the CA1 region.
Our data provides a single cell resolution picture of the spatial organization of cells in the
hippocampus and reconciles both the RNA-seq and the ABA data. While our data mostly
supports a regionalized view of the hippocampus, we observe that a single cell class does not
in general define CA1 and CA3 sub-regions. Instead, we observed that different subregions
of CA1 and CA3 are composed of distinct combinations of cell clusters (Fig 5–7). For
example, CA1d consists primarily of cluster 6 pyramidal cells (Fig 5A–C), in addition to the
cluster 1,2, 10, and 12 cells, while CA1v consists of a large set of cell classes including
cluster 1–6 and 10 cells, but at different relative abundances (Fig 5–6, Fig S6 F–G). Due to
this intermixing of cell classes in each sub-region, a bulk measurement of transcription
profiles would find a lack of regionalization, but single cell analysis with spatial resolution
would identify these distinct regions based on their unique cell class compositions. Indeed,
when we averaged the single cell expression profile within each sub-region of the CA1, we
can reproduce the continuous correlation profiles found by bulk RNA-seq between CA1v,
CA1i, and CA1d (Fig 8) (
Cembrowski et al., 2016
). The bulk RNA-seq observation that
CA1i lacked specific marker genes can also be explained. This is in fact consistent with our
findings that CA1i contained cell classes present in both CA1d and CA1v (Fig 5–7). This
organization of cell classes is observed in both the 125 gene experiments as well as in the
249 gene experiment.
It is worth noting that the complexity of cell populations observed in the CA1d versus the
CA1v matches the functional differences in CA1. CA1d is responsible for spatial learning
and navigation and contains a higher concentration of place cells and send projections to
dorsal subiculum and cortical retrosplenial area (
Cenquizca and Swanson, 2007
;
Jung et al.,
1994
; Risold et al, 1997;
O’Keefe and Dostrovsky, 1971
). We observed that CA1d is
Shah et al.
Page 10
Neuron
. Author manuscript; available in PMC 2017 October 19.
Author Manuscript
Author Manuscript
Author Manuscript
Author Manuscript
composed of a relatively homogeneous population of cells, predominantly of cluster 6 cells.
In contrast, the ventral region is involved in a variety of cognitive tasks, such as stress
response, emotional and social behavior (
Cenquizca and Swanson, 2007
;
Jung et al., 1994
;
Fanselow and Dong, 2010
;
Kishi et al., 2006
;
Muller et al., 1996
;
Petrovich et al., 2001
;
Pitkänen et al., 2000
;
Saunders et al., 1988
;
Witter and Amaral, 1991
;
Yi et al., 2015
).
Correspondingly, we observed a large set of cell classes in the CA1v regions. It is intriguing
to hypothesize that the different cell classes identified based on molecular profiles may
correspond to neurons with distinct connectivity and functional patterns. This hypothesis can
be investigated in future experiments combining anterograde tracing as well as
electrophysiological recording followed by seqFISH.
SeqFISH cell classes versus single cell RNA-seq cell types
While the accurate measurement of 100–200 genes can provide distinctions between the
large functional classes found by RNA-seq, the clusters found by seqFISH, in general,
should not be interpreted as cell types. RNA-seq measurements at the whole transcriptome
level defines cell types based on highly variable genes. On the other hand, seqFISH provides
highly accurate measurements of fewer genes, but uses the combinatorial expression patterns
to group cells into clusters. However, because only 100–200 genes are targeted in the
seqFISH experiments, not all of the “cell types” are equally represented in the gene list and
seqFISH cannot catalogue “cell types” in the same fashion that single cell RNAseq can. For
example, in our 125 gene experiments, we cannot resolve the distinct subpopulation of
interneurons because we lacked marker genes such as
Vip
and
Sst
. seqFISH and RNA-seq
provide two different, yet complementary, levels of resolution into the transcriptional
profiles of cells. RNA-seq measures the transcription levels of thousands of genes but at a
lower quantitative accuracy, while seqFISH measures only 100’s of genes but with much
greater quantitative power. The differing nature of the two sets of data informs how the data
should be analyzed and interpreted. Thus, seqFISH and single cell RNAseq have
complementary roles in elucidating distinct cell subpopulations in tissues. SeqFISH could be
applied to find finer distinctions within cell types found by RNA-seq or to look at the spatial
patterning of cell types found by RNA-seq.
seqFISH provides a generalized method to multiplex mRNA imaging in tissues
seqFISH with amplification and error correction provides a highly quantitative method to
profile hundreds of mRNA species directly in single cells within their native anatomical
context. Our method of stripping the probes from the RNA has many advantages. DNAse
digestion of probes allows false positives to be rejected as nonspecifically bound probes do
not colocalize between different rounds of hybridization (Fig 2A). In addition, the same
region of the transcript can be hybridized in every round, allowing seqFISH to efficiently
target mRNAs shorter than 1kb, enabling targeting of most genes. Lastly, seqFISH allows
exponential scaling of barcode numbers, thus 4–5 rounds of hybridization can code for
hundreds of transcripts with a simple error correction scheme. Theoretically, the entire
transcriptome can be coded for with error correction by using 8–9 rounds of hybridization
with seqFISH. These advantages of HCR seqFISH allows robust multiplexed RNA detection
in tissues, shown here in the mouse brain.
Shah et al.
Page 11
Neuron
. Author manuscript; available in PMC 2017 October 19.
Author Manuscript
Author Manuscript
Author Manuscript
Author Manuscript
Ultimately, the multiplexing capability of seqFISH is limited by the amount of optical space
within a cell, and not by the coding capacity of the method (supplementary text). We showed
previously that super-resolution microscopy can significantly increase the optical space
available in the cell for transcription profile imaging, but super-resolution microscopy
experiments proved difficult to image in samples thicker than 1μm, and were experimentally
cumbersome and time consuming to image (
Lubeck and Cai, 2012
). A recent development
in expansion microscopy as well as correlation methods (Coskun et al., 2016) however offers
promise for multiplexing to levels of high transcript density (
Chen et al., 2015a
;
Treweek et
al., 2015
,
Chen et al., 2016
). In addition, by labeling subcellular components (i.e. dendrites
and axons) with antibodies, the local transcriptome in compartments of the cell can be
measured.
We observed that, because expression patterns amongst genes are highly correlated, the
distinction between large classes of cells can be determined from 10–20 genes, while a finer
classification of cell clusters depends on the quantitative measurement of the combinatorial
expression patterns of many genes (Fig 3C–D). This correlation amongst genes can be used
to “stitch” our seqFISH data with single cell RNAseq data, similar to the approach explored
with single cell RNAseq and ISH in Satija et al (
Satija et al., 2015
). By correlating seqFISH
data to single cell RNA-seq expression data, cells types identified based on RNA-seq can be
“mapped” back into our seqFISH data.
As shown here, seqFISH with hundreds of genes in tissues can become a general and widely
used tool to answer a wide range of fundamental questions in biology and medicine. For
neuroscience, by combining the insights into the spatial organization of transcription
provided by seqFISH with connectomics and electrophysiological measurements, we can
obtain a comprehensive understanding of the molecular basis of the neuroanatomy of the
brain.
Experimental Procedure
Probe Design
Genes were selected from the Allen Brain Atlas database. We identified genes that are
heterogeneously expressed in coronal sections containing the hippocampus at Bregma
coordinates −2.68 mm anterior. We selected 100 genes that had high variances across these
distinct regions and that also had low-medium expression levels. Probe sequences were
designed using software developed in house. Full details are described in Supplemental
Experimental Procedures.
Probe Generation
All oligoarray pools were purchased as 92k synthesis from Customarray Inc. Probes were
amplified from array-synthesized oligo pool as previously described (
Chen et al., 2015b
).
Full details are described in Supplemental Experimental Procedures.
Shah et al.
Page 12
Neuron
. Author manuscript; available in PMC 2017 October 19.
Author Manuscript
Author Manuscript
Author Manuscript
Author Manuscript
Brain extraction and sample mounting
C57BL/6 with Ai6 Cre-reporter (uncrossed) (Jackson Labs, SN: 007906) female mice aged
50–80 days were anesthetized with isoflurane according to institute protocols (protocol
#1701-14) (
Madisen et al., 2012
). Mice were perfused with 4% PFA and the brain was
dissected out and placed in a 4% PFA buffer for 2 hours at room temperature. The brain was
then immersed in 4C 30% RNAse-free Sucrose\1× PBS until the brain sank. Once sunk, the
brain was embedded in OCT and sectioned. Full details are described in Supplemental
Experimental Procedures.
Sample permeabilization, hybridization, and Imaging
Sections were permeabilized in 4C 70% EtOH for 12–18 hours. Brains were further
permeabilized by the addition of rnase-free 8% SDS. A hybridization chamber was adhered
around the brain section. RNA integrity test probes were hybridized overnight at 37 in
hybridization buffer (Table S3). Samples were washed in 30% wash buffer (WB) for 30
minutes. Probes were amplified. Following amplification, samples were washed in the same
30% WB for at least 10 minutes to remove excess hairpins. Samples were stained with DAPI
and submerged in pyranose oxidase antibleaching buffer (
Lubeck et al., 2014
). If the RNA
was deemed to be intact, DAPI data was collected in this hybridization. Samples were
digested with DNAse I for 4 hours at room temperature on the scope. Following DNAse I
the sample was washed several times with 30% WB and the probes were hybridized
overnight (Table S4 and S5). Samples were again washed and amplified. Repeating this
cycle with the appropriate probes for each hybridization developed barcode digits.
Fluorescent Nissl stain was collected at the end of the experiment along with images of
multispectral beads to aid chromatic aberration corrections. Full details are described in
Supplemental Experimental Procedures.
Image Processing
The images were first corrected for to remove the uneven illumination profiles in each
channel and to remove the effects of chromatic aberration. The background intensity in the
images was then subtracted. A 150-pixel border region around the image was ignored in all
analysis to avoid errors from edge effects of illumination. Full details are described in
Supplemental Experimental Procedures.
Image Registration
The processed images were then registered by first taking a maximum intensity projection
along the z direction in each channel. All of the maximum projections of the channels of a
single hybridization were then collapsed resulting in 4 composite images containing all the
points in a particular round of hybridization. Each of these composite images of
hybridization 1–3 was then cross-correlated individually with the composite image of
hybridization 4 and the position of the maxima of the cross-correlation was used as the
translation factor to align hybridizations 1–3 to hybridization 4.
Shah et al.
Page 13
Neuron
. Author manuscript; available in PMC 2017 October 19.
Author Manuscript
Author Manuscript
Author Manuscript
Author Manuscript
Cell Segmentation
For cells in the cortex, the cells were segmented manually using the DAPI images taken in
the first round of hybridization and the fluorescent nissl stain taken at the end of the
experiment. Furthermore, the density of the point cloud surrounding a cell was taken into
account when forming cell boundaries, especially in cells that did not stain with the nissl
stain. For the hippocampus, the cells were segmented by first manually selecting the centroid
in 3D of each DAPI signal of every cell. Transcripts were first assigned based on nearest
centroids. These point clouds were then used to refine the centroid estimate and create a 3D
voronoi tessellation with a 10% boundary-shrinking factor to eliminate ambiguous mRNA
assignments from neighboring cells. Regional segmentation was performed manually using
the ImageJ ROI tool.
Barcode calling
The potential mRNA signals were then found by LOG filtering the registered images and
finding points of local maxima above a specified threshold value. Once all potential points in
all channels of all hybridizations were obtained, dots were matched to potential barcode
partners in all other channels of all other hybridizations using a 1-pixel search radius to find
symmetric nearest neighbors. This procedure was repeated using each hybridization as a
seed for barcode finding and only barcodes that were called similarly in at least 3 out of 4
rounds were used in the analysis. The number of each barcode was then counted in each of
the assigned cell volumes and transcript numbers were assigned based on the number of on-
target barcodes present in the cell volume. All image processing and image analysis code
can be obtained upon request. Full details are described in Supplemental Experimental
Procedures.
Clustering
To cluster the dataset with two brain measured with 125 genes, we first Z-score normalized
each of the slices based on gene expression (Table S6). Once the single cell gene expression
data is converted into z-scores, we compute a matrix of cell-to-cell correlations using
Pearson correlation coefficients for all of the cells in the two brains. Then hierarchical
clustering with Ward linkage is performed on the cell-to-cell correlation data using cells
taken from the center of the field of view. To analyze the robustness of individual clusters, a
random forest model was trained using varying subsets of the data and used to predict the
cluster assignment of the remaining cells (
Breiman, 2001
). For Figure 4–6, the entire field of
cells was classified using the clustered cells as the training set. A bootstrap analysis by
dropping different sets of cells was performed in increments (Fig S5). To determine the
effect of dropping out genes on the accuracy of the clustering analysis, we used a random
forest decision tree to learn the cluster definition based on the 125 gene data. Then we ask
the decision tree to re-compute the cluster assignment on cell-to-cell correlation matrices
with fewer and fewer genes (Fig 3C–D, green line). Bootstrap resampling was also
performed with this analysis (Fig 3C–D, blue lines). The PCA and stone analysis were
performed using the same cell-to-cell z-scored Pearson correlation matrix. The cell-to-cell
correlation in Fig S3EI was calculated with increasing number of principal components
dropped (have their eigenvalues set to zero). The cluster assignment accuracy is again
Shah et al.
Page 14
Neuron
. Author manuscript; available in PMC 2017 October 19.
Author Manuscript
Author Manuscript
Author Manuscript
Author Manuscript
computed through the random forest decision tree. The 249 gene experiment was clustered
independently with Z-score normalized data.
Supplementary Material
Refer to Web version on PubMed Central for supplementary material.
Acknowledgments
We thank Michael Elowitz, Henry Lester, Hongwei Dong, and Bosiljka Tasic for advice on the manuscript. We also
would like to thank Niles Pierce, Viviana Gradinaru, Thanos Siapas, Mary Kennedy, and Carlos Lois for productive
discussions. Raw data can be found in the Supplementary Materials of this paper. The National Institute of Health,
the McKnight Foundation, and the Allen Foundation supported this work.
References and Notes
Beliveau BJ, Joyce EF, Apostolopoulos N, Yilmaz F, Fonseka CY, McCole RB, Chang Y, Li JB,
Senaratne TN, Williams BR, et al. Versatile design and synthesis platform for visualizing genomes
with Oligopaint FISH probes. Proc. Natl. Acad. Sci. U.S.A. 2012; 109:21301–21306. [PubMed:
23236188]
Betzig E, Patterson GH, Sougrat R, Lindwasser OW, Olenych S, Bonifacino JS, Davidson MW,
Lippincott-Schwartz J, Hess HF. Imaging Intracellular Fluorescent Proteins at Nanometer
Resolution. Science. 2006; 313:1642–1645. [PubMed: 16902090]
Breiman L. Random Forests. Mach. Learn. 2001; 45:5–32.
Cajigas IJ, Tushev G, Will TJ, Dieck S. tom, Fuerst N, Schuman EM. The Local Transcriptome in the
Synaptic Neuropil Revealed by Deep Sequencing and High-Resolution Imaging. Neuron. 2012;
74:453–466. [PubMed: 22578497]
Cembrowski MS, Bachman JL, Wang L, Sugino K, Shields BC, Spruston N. Spatial Gene-Expression
Gradients Underlie Prominent Heterogeneity of CA1 Pyramidal Neurons. Neuron. 2016; 89:351–
368. [PubMed: 26777276]
Cenquizca LA, Swanson LW. Spatial organization of direct hippocampal field CA1 axonal projections
to the rest of the cerebral cortex. Brain Res. Rev. 2007; 56:1–26. [PubMed: 17559940]
Chen F, Tillberg PW, Boyden ES. Expansion microscopy. Science. 2015a; 347:543–548. [PubMed:
25592419]
Chen F, Wassie AT, Cote AJ, Sinha A, Alon S, Asano S, Daugharthy ER, Chang J-B, Marblestone A,
Church GM, Raj A, Boyden ES. Nanoscale imaging of RNA with expansion microscopy. Nat Meth
advance online publication. 2016
Chen KH, Boettiger AN, Moffitt JR, Wang S, Zhuang X. Spatially resolved, highly multiplexed RNA
profiling in single cells. Science. 2015b; 348:aaa6090. [PubMed: 25858977]
Choi HMT, Beck VA, Pierce NA. Next-Generation in Situ Hybridization Chain Reaction: Higher Gain,
Lower Cost, Greater Durability. ACS Nano. 2014; 8:4284–4294. [PubMed: 24712299]
Darmanis S, Sloan SA, Zhang Y, Enge M, Caneda C, Shuer LM, Gephart MGH, Barres BA, Quake
SR. A survey of human brain transcriptome diversity at the single cell level. Proc. Natl. Acad. Sci.
2015; 112:7285–7290. [PubMed: 26060301]
Dong H-W, Swanson LW, Chen L, Fanselow MS, Toga AW. Genomic–anatomic evidence for distinct
functional domains in hippocampal field CA1. Proc. Natl. Acad. Sci. 2009; 106:11794–11799.
[PubMed: 19561297]
Fan Y, Braut SA, Lin Q, Singer RH, Skoultchi AI. Determination of transgenic loci by expression
FISH. Genomics. 2001; 71:66–69. [PubMed: 11161798]
Fanselow MS, Dong H-W. Are the dorsal and ventral hippocampus functionally distinct structures?
Neuron. 2010; 65:7–19. [PubMed: 20152109]
Femino AM, Fay FS, Fogarty K, Singer RH. Visualization of Single RNA Transcripts in Situ. Science.
1998; 280:585–590. [PubMed: 9554849]
Shah et al.
Page 15
Neuron
. Author manuscript; available in PMC 2017 October 19.
Author Manuscript
Author Manuscript
Author Manuscript
Author Manuscript
Habib N, Li Y, Heidenreich M, Swiech L, Trombetta JJ, Zhang F, Regev A. Div-Seq: A single nucleus
RNA-Seq method reveals dynamics of rare adult newborn neurons in the CNS. bioRxiv.
2016:045989.
Jung MW, Wiener SI, McNaughton BL. Comparison of spatial firing characteristics of units in dorsal
and ventral hippocampus of the rat. J. Neurosci. 1994; 14:7347–7356. [PubMed: 7996180]
Ke R, Mignardi M, Pacureanu A, Svedlund J, Botling J, Wählby C, Nilsson M. In situ sequencing for
RNA analysis in preserved tissue and cells. Nat. Methods. 2013; 10:857–860. [PubMed:
23852452]
Kishi T, Tsumori T, Yokota S, Yasui Y. Topographical projection from the hippocampal formation to
the amygdala: A combined anterograde and retrograde tracing study in the rat. J. Comp. Neurol.
2006; 496:349–368. [PubMed: 16566004]
Klein AM, Mazutis L, Akartuna I, Tallapragada N, Veres A, Li V, Peshkin L, Weitz DA, Kirschner
MW. Droplet Barcoding for Single-Cell Transcriptomics Applied to Embryonic Stem Cells. Cell.
2015; 161:1187–1201. [PubMed: 26000487]
Lee JH, Daugharthy ER, Scheiman J, Kalhor R, Yang JL, Ferrante TC, Terry R, Jeanty SSF, Li C,
Amamoto R, et al. Highly Multiplexed Subcellular RNA Sequencing in Situ. Science. 2014;
343:1360–1363. [PubMed: 24578530]
Lein ES, Hawrylycz MJ, Ao N, Ayres M, Bensinger A, Bernard A, Boe AF, Boguski MS, Brockway
KS, Byrnes EJ, et al. Genome-wide atlas of gene expression in the adult mouse brain. Nature.
2007; 445:168–176. [PubMed: 17151600]
Lubeck E, Cai L. Single-cell systems biology by super-resolution imaging and combinatorial labeling.
Nat. Methods. 2012; 9:743–748. [PubMed: 22660740]
Lubeck E, Coskun AF, Zhiyentayev T, Ahmad M, Cai L. Single-cell in situ RNA profiling by
sequential hybridization. Nat. Methods. 2014; 11:360–361. [PubMed: 24681720]
Macosko EZ, Basu A, Satija R, Nemesh J, Shekhar K, Goldman M, Tirosh I, Bialas AR, Kamitaki N,
Martersteck EM, et al. Highly Parallel Genome-wide Expression Profiling of Individual Cells
Using Nanoliter Droplets. Cell. 2015; 161:1202–1214. [PubMed: 26000488]
Madisen L, Zwingman TA, Sunkin SM, Oh SW, Zariwala HA, Gu H, Ng LL, Palmiter RD, Hawrylycz
MJ, Jones AR, et al. A robust and high-throughput Cre reporting and characterization system for
the whole mouse brain. Nat. Neurosci. 2010; 13:133–140. [PubMed: 20023653]
Madisen L, Mao T, Koch H, Zhuo J, Berenyi A, Fujisawa S, Hsu Y-WA, Iii AJG, Gu X, Zanella S, et
al. A toolbox of Cre-dependent optogenetic transgenic mice for light-induced activation and
silencing. Nat. Neurosci. 2012; 15:793–802. [PubMed: 22446880]
Miller JA, Nathanson Jason, Franjic Daniel, Shim Sungbo, Dalley Rachel A, Shapouri Sheila, Smith
Kimberly A, Sunkin Susan M. Bernard Amy, Bennett Jeffrey L, Lee Chang-Kyu, Hawrylycz
Michael J, Jones Allan R, Amaral David G, Sestan Nenad, Gage Fred H, Lein Ed S. Conserved
molecular signatures of neurogenesis in the hippocampal subgranular zone of rodents and
primates. Development. 2013; 140(22):4633–4644. [PubMed: 24154525]
Muller R, Stead M, Pach J. The hippocampus as a cognitive graph. J. Gen. Physiol. 1996; 107:663–
694. [PubMed: 8783070]
O’Keefe J, Dostrovsky J. The hippocampus as a spatial map. Preliminary evidence from unit activity in
the freely-moving rat. Brain Res. 1971; 34:171–175. [PubMed: 5124915]
Petrovich GD, Canteras NS, Swanson LW. Combinatorial amygdalar inputs to hippocampal domains
and hypothalamic behavior systems. Brain Res. Brain Res. Rev. 2001; 38:247–289. [PubMed:
11750934]
Pitkänen A, Pikkarainen M, Nurminen N, Ylinen A. Reciprocal Connections between the Amygdala
and the Hippocampal Formation, Perirhinal Cortex, and Postrhinal Cortex in Rat: A Review. Ann.
N. Y. Acad. Sci. 2000; 911:369–391. [PubMed: 10911886]
Raj A, Peskin CS, Tranchina D, Vargas DY, Tyagi S. Stochastic mRNA Synthesis in Mammalian Cells.
PLoS Biol. 2006; 4:e309. [PubMed: 17048983]
Risold PY, Swanson LW. Structural evidence for functional domains in the rat hippocampus. Science.
1996; 272:1484–1486. [PubMed: 8633241]
Rust MJ, Bates M, Zhuang X. Sub-diffraction-limit imaging by stochastic optical reconstruction
microscopy (STORM). Nat Meth. 2006; 3:793–796.
Shah et al.
Page 16
Neuron
. Author manuscript; available in PMC 2017 October 19.
Author Manuscript
Author Manuscript
Author Manuscript
Author Manuscript
Satija R, Farrell JA, Gennert D, Schier AF, Regev A. Spatial reconstruction of single-cell gene
expression data. Nat Biotech. 2015; 33:495–502.
Saunders RC, Rosene DL, Van Hoesen GW. Comparison of the efferents of the amygdala and the
hippocampal formation in the rhesus monkey: II. Reciprocal and nonreciprocal connections. J.
Comp. Neurol. 1988; 271:185–207. [PubMed: 2454247]
Shah S, Lubeck E, Schwarzkopf M, He T, Greenbaum A, Sohn C. ho, Lignell A, Choi HMT,
Gradinaru V, Pierce NA, Cai L. Single-molecule RNA detection at depth via hybridization chain
reaction and tissue hydrogel embedding and clearing. Development dev.138560. 2016
Ståhl PL, Salmén F, Vickovic S, Lundmark A, Navarro JF, Magnusson J, Giacomello S, Asp M,
Westholm JO, Huss M, Mollbrink A, Linnarsson S, Codeluppi S, Borg Å, Pontén F, Costea PI,
Sahlén P, Mulder J, Bergmann O, Lundeberg J, Frisén J. Visualization and analysis of gene
expression in tissue sections by spatial transcriptomics. Science. 2016; 353:78–82. [PubMed:
27365449]
Tasic B, Menon V, Nguyen TN, Kim TK, Jarsky T, Yao Z, Levi B, Gray LT, Sorensen SA, Dolbeare T,
et al. Adult mouse cortical cell taxonomy revealed by single cell transcriptomics. Nat. Neurosci.
2016
advance online publication
.
Thompson CL, Pathak SD, Jeromin A, Ng LL, MacPherson CR, Mortrud MT, Cusick A, Riley ZL,
Sunkin SM, Bernard A, et al. Genomic Anatomy of the Hippocampus. Neuron. 2008; 60:1010–
1021. [PubMed: 19109908]
Treweek JB, Chan KY, Flytzanis NC, Yang B, Deverman BE, Greenbaum A, Lignell A, Xiao C, Cai L,
Ladinsky MS, et al. Whole-body tissue stabilization and selective extractions via tissue-hydrogel
hybrids for high-resolution intact circuit mapping and phenotyping. Nat. Protoc. 2015; 10:1860–
1896. [PubMed: 26492141]
Van der Maaten L, Hinton G. Visualizing data using t-SNE. J. Mach. Learn. Res. 2008; 9:85.
Witter MP. Organization of the entorhinal—hippocampal system: A review of current anatomical data.
Hippocampus. 1993; 3:33–44. [PubMed: 8287110]
Witter MP, Amaral DG. Entorhinal cortex of the monkey: V. Projections to the dentate gyrus,
hippocampus, and subicular complex. J. Comp. Neurol. 1991; 307:437–459. [PubMed: 1713237]
Yang B, Treweek JB, Kulkarni RP, Deverman BE, Chen C-K, Lubeck E, Shah S, Cai L, Gradinaru V.
Single-Cell Phenotyping within Transparent Intact Tissue through Whole-Body Clearing. Cell.
2014
Yang SM, Alvarez DD, Schinder AF. Reliable Genetic Labeling of Adult-Born Dentate Granule Cells
Using Ascl1 CreERT2 and Glast CreERT2 Murine Lines. J Neurosci. 2015; 35(46):15379–15390.
[PubMed: 26586824]
Yi F, Catudio-Garrett E, Gábriel R, Wilhelm M, Erdelyi F, Szabo G, Deisseroth K, Lawrence J.
Hippocampal “cholinergic interneurons” visualized with the choline acetyltransferase promoter:
anatomical distribution, intrinsic membrane properties, neurochemical characteristics, and capacity
for cholinergic modulation. Front. Synaptic Neurosci. 2015; 7
Zeisel A, Manchado ABM, Codeluppi S, Lönnerberg P, Manno GL, Juréus A, Marques S, Munguba H,
He L, Betsholtz C, et al. Cell types in the mouse cortex and hippocampus revealed by single-cell
RNA-seq. Science. 2015 aaa1934.
Shah et al.
Page 17
Neuron
. Author manuscript; available in PMC 2017 October 19.
Author Manuscript
Author Manuscript
Author Manuscript
Author Manuscript
Highlights
•
Amplified seqFISH enables in situ detection of 100’s genes in single
cells in tissues.
•
Combinatorial expression patterns of genes define cell classes in the
mouse brain.
•
Subregions of the hippocampus are composed of distinct combinations
of cell classes.
•
Heterogeneity in cell class compositions increases along the dorsal to
ventral axis
Shah et al.
Page 18
Neuron
. Author manuscript; available in PMC 2017 October 19.
Author Manuscript
Author Manuscript
Author Manuscript
Author Manuscript
Fig. 1. Overview of the Sequential barcode FISH (seqFISH) in brain slices
A.
A coronal section from a mouse brain was mounted on a slide and imaged in all boxed
areas. Each image was taken at 60x magnification.
B.
Example of barcoding hybridizations
from one cell in field from A. The same points are re-probed through a sequence of 4
hybridizations (numbered). The sequence of colors at a given location provides a barcode
readout for that mRNA (“barcode composite”). These barcodes are identified through
referencing a lookup table abbreviated in
D
and quantified to obtain single cell expression.
In principle, the maximum number of transcripts that can be identified with this approach
scales to F
N
, where F is the number of fluorophores and N is the number of hybridizations.
Error correction adds another round of hybridization.
C.
Serial smHCR is an alternative
detection method where 5 genes are quantified in each hybridization and repeated N times.
Serial hybridization scales as F*N.
D.
Schematic for multiplexing 125 genes in single cells.
100 genes are multiplexed in 4 hybridizations by seqFISH barcoding. This barcode scheme
is tolerant to loss of any round of hybridization in the experiment. 25 genes are serially
hybridized 5 genes at a time by 5 rounds of hybridization. Each number represents a color
channel in single molecule HCR. As a control, 5 genes are measured both by double rounds
of smHCR as well as barcoding in the same cell.
E.
SmHCR amplifies signal from
individual mRNAs. After imaging, DNAse strips the smHCR probes from the mRNA,
Shah et al.
Page 19
Neuron
. Author manuscript; available in PMC 2017 October 19.
Author Manuscript
Author Manuscript
Author Manuscript
Author Manuscript