NeuroImage 224 (2021) 117395

Contents lists available at ScienceDirect

NeuroImage
journal homepage: www.elsevier.com/locate/neuroimage

Scaling of gene transcriptional gradients with brain size across mouse

development
Hoi Yan Gladys Lau a,b,c, Alex Fornito a, Ben D. Fulcher b,∗
a
The Turner Institute for Brain and Mental Health, School of Psychological Sciences, and Monash Biomedical Imaging, Monash University, Victoria, Australia
b
School of Physics, The University of Sydney, NSW 2006, Australia
c
Li Ka Shing Faculty of Medicine, The University of Hong Kong, Pokfulam, Hong Kong

a r t i c l e i n f o a b s t r a c t

Keywords: The structure of the adult brain is the result of complex physical mechanisms acting in three-dimensional space
Gene expression through development. Consequently, the brain’s spatial embedding plays a key role in its organization, including
Transcription the gradient-like patterning of gene expression that encodes the molecular underpinning of functional special-
Gradients
ization. However, we do not yet understand how changes in brain shape and size that occur across development
Brain development
influence the brain’s transcriptional architecture. Here we investigate the spatial embedding of transcriptional
patterns of over 1800 genes across seven time points through mouse-brain development using data from the Allen
Developing Mouse Brain Atlas. We find that transcriptional similarity decreases exponentially with separation
distance across all developmental time points, with a correlation length scale that follows a power-law scaling
relationship with a linear dimension of brain size. This scaling suggests that the mouse brain achieves a character-
istic balance between local molecular similarity (homogeneous gene expression within a specialized brain area)
and longer-range diversity (between functionally specialized brain areas) throughout its development. Extrapo-
lating this mouse developmental scaling relationship to the human cortex yields a prediction consistent with the
value measured from microarray data. We introduce a simple model of brain growth as spatially autocorrelated
gene-expression gradients that expand through development, which captures key features of the mouse develop-
mental data. Complementing the well-known exponential distance rule for structural connectivity, our findings
characterize an analogous exponential distance rule for transcriptional gradients that scales across mouse brain
development, providing new understanding of spatial constraints on the brain’s molecular patterning.

1. Introduction bilities, and the existence of a network core (Ercsey-Ravasz et al., 2013;
Henderson and Robinson, 2011; 2014; Roberts et al., 2016; Song et al.,
Brain structure is the result of physical mechanisms playing out 2014; Stiso and Bassett, 2018; Robinson, 2019). Spatial proximity also
through development, shaped by genetic and environmental factors. In plays a strong role in shaping gradients of gene expression, with nearby
particular, the embedding of neural tissue in physical space places major areas exhibiting more similar gene-expression profiles than more distant
constraints on neural organization. For example, the connection prob- areas in the head neurons of C. elegans (Arnatkeviciūtė et al., 2018), the
ability between pairs of neural elements decays with their separation mouse brain (French and Pavlidis, 2011; Fulcher and Fornito, 2016),
distance in C. elegans (Arnatkeviciūtė et al., 2018), mouse (Fulcher and and the human cortex (Arnatkeviciūtė et al., 2019; Fornito et al., 2019;
Fornito, 2016; Gămănut et al., 2018), rat (Noori et al., 2017), zebrafish Hawrylycz et al., 2010; Krienen et al., 2016; Pantazatos and Li, 2017;
(Kunst et al., 2019), non-human primate (Markov et al., 2013), and hu- Richiardi et al., 2017).
man (Roberts et al., 2016), and is frequently characterized as an expo- The spatial embedding of gene-expression patterns provides impor-
nential distance rule (Ercsey-Ravasz et al., 2013; Horvát et al., 2016; tant clues about how the brain’s functional specialization is organized
Theodoni et al., 2020). Indeed, very simple spatial rules can explain in physical space. Macroscopic functional gradients across the human
much of the high-order statistics of connectome topology, including its cortex maximize the spatial distance between areas functionally in-
modularity, long-tailed degree distribution, local network motif proba- volved in sensory perception and those involved in integrative cognition

∗
Corresponding author.
E-mail address: [email protected] (B.D. Fulcher).

https://doi.org/10.1016/j.neuroimage.2020.117395
Received 31 March 2020; Received in revised form 17 September 2020; Accepted 18 September 2020
Available online 24 September 2020
1053-8119/© 2020 The Authors. Published by Elsevier Inc. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/)
H.Y.G. Lau, A. Fornito and B.D. Fulcher NeuroImage 224 (2021) 117395

(Huntenburg et al., 2017; Margulies et al., 2016). These hierarchical between pairs of voxels were computed as Euclidean distances in three-
functional gradients are underpinned by a corresponding variation in dimensional space using the spatial resolution of the coordinate grid at
structural microarchitecture (Burt et al., 2018; Fulcher et al., 2019; each time point: E11.5 (0.08 mm), E13.5 (0.1 mm); E15.5 (0.12 mm),
Paquola et al., 2019). But given the constraint of a fixed brain size, E18.5 (0.14 mm), P4 (0.16 mm), P14 (0.2 mm), and P28 (0.2 mm).
how does the brain strike the right balance between local molecular
similarity (similar gene-expression patterns supporting regional special- 2.2. Filtering and normalization
ization of function), with inter-regional molecular diversity, which sup-
ports functional heterogeneity. Furthermore, how does the spatial em- At each developmental time point, we first filtered out voxels that
bedding of transcriptional gradients emerge through development? Do were labeled as spinal cord or were unannotated, and labeled voxels
different spatial constraints dominate early in development compared according to four anatomical masks: ‘forebrain’, ‘midbrain’, ‘hindbrain’,
to later? Although macroscopic brain organization in adult is often at- ‘dorsal pallium’. At each time point, we retained genes with valid ex-
tributed to gradients set up during development (e.g., in transcriptional pression (i.e., not a missing value) for at least 70% of voxels, resulting
factors) (Greig et al., 2013), to our knowledge, no study has analyzed in different numbers of good-quality genes across time: E11.5 (2038
the spatial embedding of gene expression through development. genes), E13pt5 (2085), E15pt5 (1861), E18pt5 (2053), P4 (2056), P14
The Allen Developing Mouse Brain Atlas (ADMBA) (Thompson et al., (2015), and P28 (2063). These small differences in the number of genes
2014) is a comprehensive database of gene-expression across mouse- across time did not affect our results: we obtained similar findings when
brain development. The atlas contains in situ hybridization measure- restricting our analysis to a consistent subset of 1861 genes that had
ments for approximately 2100 genes at seven developmental time good-quality data across all time points. We then retained only voxels
points, available in a three-dimensional reference space registered to with valid expression data for at least 70% of these genes. The num-
an anatomical reference atlas at each developmental time point. Many ber of voxels varied across time as 5034 (E11.5), 9473 (E13.5), 11 314
analyses of gene transcription through mouse-brain development have (E15.5), 11 313 (E18.5), 19 754 (P4), 21 579 (P14), and 24 822 (P28).
shed light on how the spatiotemporal variation of specific genes through We generated a voxel × gene expression matrix containing voxels and
development shape different aspects of brain structure (Dillman et al., genes that survived these filtering steps.
2013; Fertuzinhos et al., 2014; Mody et al., 2001; Thompson et al., Raw expression values measured using in situ hybridization are not
2014). However, these developmental atlas data have not previously comparable between genes (Lee et al., 2008). We accounted for this
been analyzed from the viewpoint of spatially embedded gradients of at each time point by normalizing expression values of each gene us-
transcriptional similarity. By characterizing gradients, we can address ing an outlier-robust sigmoidal transformation, 𝑆(𝑥) = [1 + exp(−(𝑥 −
global questions about transcriptional patterning through development 𝑥med )∕1.35𝑥iqr )]−1 , where xmed is the median of the expression vector,
in a parcellation-free way that does not involve the difficult task of track- x, for a given gene, and xiqr is its interquartile range (Arnatkeviciūtė
ing anatomically defined areas over time. et al., 2019; Fulcher et al., 2013). For visualization, this normalization
Here we study how correlated gene expression (CGE), defined as the was followed by linear rescaling to the unit interval. The resulting nor-
similarity in the gene-expression signatures of two points in the brain, malized voxel × gene matrix, Gt , at a given time point, t, was used for
is spatially embedded across mouse-brain development. Matching the further analysis.
well-known exponential distance rule for connectivity, we find an ex-
ponential decay in CGE at each developmental time point, with a spa- 2.3. Quantifying the spatial embedding of CGE
tial correlation length that scales with a linear dimension of brain size.
Extrapolating the relationship fitted on developing mouse-brain data al- To understand the spatial embedding of gene-expression patterns, we
lows us to obtain an estimate consistent with the measured correlation analyzed the distance-dependence of correlated gene expression (CGE)
length of CGE in the human cortex. Using a simple model, we show that at each developmental time point. CGE is a measure of gene-expression
the patterns are consistent with brain growth in which spatially auto- similarity between a pair of voxels and was computed as the Pearson cor-
correlated gene-expression gradients expand with brain size over time. relation coefficient between normalized gene-expression vectors (rows
of a given voxel × gene matrix, Gt , defined above).
2. Methods At each time point, we computed CGE for every pair of voxels. Be-
cause the distribution of pairwise distances is concentrated near lower
2.1. Data distances, we wanted to prevent the exponential fitting from being bi-
ased towards fitting these shorter length scales, but to instead fit well
We analyze data from the ADMBA, which contains in situ hy- across the full range of distances. To do this, we first performed an
bridization (ISH) expression estimates for approximately 2100 genes in equiprobable distance binning of the data using twenty bins, and com-
C57Bl/6J mice across seven developmental stages, spanning both em- puted the mean CGE value in each bin (and summarizing each bin as its
bryonic (E11.5, E13.5, E15.5, and E18.5) and postnatal (P4, P14, and center). To this binned data, we then fitted a three-parameter exponen-
P28) stages (Thompson et al., 2014). These genes were selected to en- tial function, CGE(𝑑) = 𝐴 exp(−𝑑∕𝜆) + 𝑓0 , using nonlinear least squares
compass relevant transcription factors, neurotransmitters and their re- and identified the fitted parameters as: spatial correlation length, 𝜆,
ceptors, cell-type or regional marker genes, genes relevant to develop- strength A, and offset, f0 . Adjusted R2 was used as a goodness-of-fit
mental signaling pathways, and genes corresponding to common drug statistic. Note that for self-correlation, at 𝑑 = 0, CGE = 1, and thus fitting
targets or implicated in neurodevelopmental disorders (Thompson et al., data that includes self-correlations should have a fixed strength param-
2014). The raw imaging data, encompassing 434 946 brain sections eter 𝐴 = 1. Here we excluded self-correlations and allowed the strength
of high-resolution ISH, are registered to a three-dimensional mouse- parameter, A, to vary. When characterizing distances relative to brain
brain atlas at each time point. We accessed this registered data at the size, as 𝑑rel = 𝑑∕𝑑max , we defined dmax as the maximum extent along the
voxel level via the Allen Software Development Kit (SDK, 2015). Our anterior–posterior axis of the brain. For computational efficiency, we
code for retrieving data (python) and for data processing and analysis restricted our analysis to a random subset of 1000 voxels at each devel-
(Matlab) is available at https://github.com/NeuralSystemsAndSignals/ opmental time point. Our results do not strongly depend on this spatial
DevelopingMouse. At each time point, we retrieved expression data for downsampling (e.g., 𝜆 estimates stabilized for samples of approximately
all available genes. We used sagittal section data for its comprehensive- 500 voxels, cf. Fig. S4).
ness, which had coverage across the left hemisphere at each time point. Note that our definition of 𝜆 as a length scale (m) is consistent with its
Expression data for each gene was obtained as ‘expression energy’ (ISH conventional use to denote wavelength in physics (units of length), but
signal intensity) in each voxel in the three-dimensional grid. Distances differs by an inverse relationship from its use as a decay rate with units
H.Y.G. Lau, A. Fornito and B.D. Fulcher NeuroImage 224 (2021) 117395

of inverse length (m−1 ) used to characterize the exponential distance (rows of the matrix in Fig. 1B). We fitted an exponential function to the
rule for structural connectivity (Ercsey-Ravasz et al., 2013; Horvát et al., decay of CGE(d) as a function of separation distance, d, shown in Fig. 1C,
2016; Noori et al., 2017; Theodoni et al., 2020). to quantify how transcriptional similarity is spatially embedded. Varia-
To estimate and visualize the dominant expression pattern across all tion in the parameters of this fit, particularly that of the characteristic
genes in matrices, Gt , we used principal components analysis, account- length scale over which genes are correlated, 𝜆, tell us how the proper-
ing for missing values using alternating least squares. ties of gene-expression gradients vary across mouse-brain development.

2.4. Modeling 3.1. Low-dimensional spatial gradients

To better understand patterns found in the empirical data, we sim- To better understand the dominant patterns of gene expression in
ulated simple spatial models of gene expression patterning and brain the mouse brain, we first visualized voxel × gene expression matrices
growth. At each developmental time point, we first simulated gene- for each time point, and projected their leading principal components
expression patterns for each of the 1861 genes. We first defined the into physical space. Examples for two selected time points are shown in
geometry by approximating the brain as a rectangular prism with Fig. 2 (see Fig. S1 for all time points). Gene-expression matrices have
dimensions (anterior–posterior × superior–inferior × left–right) been reordered to put genes with similar expression close to each other
equal to that of the developmental reference atlas at that time point: (columns of the expression matrices in Fig. 2A and C) and, within three
3.52 × 4.72 × 1.20 mm (E11.5), 5.70 × 3.30 × 1.20 mm (E13.5), main divisions—forebrain, midbrain, and hindbrain—voxels with simi-
6.60 × 3.72 × 1.68 mm (E15.5), 7.84 × 4.06 × 2.10 mm (E18.5), lar expression signatures close to each other. Visually, we find interest-
10.56 × 5.44 × 3.04 mm (P4), 12.80 × 6.60 × 4.20 mm (P14), and ing gene-expression patterns that clearly differ between the three major
13.60 × 7.40 × 4.20 mm (P28). We formed a grid over this idealized brain divisions. The similarity-based ordering of columns in these ex-
geometry by dividing each dimension into 50 equally spaced points. We pression matrices makes the low-dimensionality of the expression data
then generated a random spatially autocorrelated map independently visually clearer: we see how large numbers of genes display highly cor-
for each of 1861 genes using a spatial-lag model (Anselin, 2013; Burt related expression patterns. The leading principal component of the ex-
et al., 2018; Burt et al., 2020). The model defines a dependence be- pression data (computed using alternating least squares to account for a
tween samples using an exponential kernel as 𝑊𝑖𝑗 = exp(−𝑑𝑖𝑗 ∕𝑑0 ), for small proportion of missing data) is annotated to the left of Fig. 2A and
pairwise distances, dij , and a characteristic spatial scale, d0 . From this C, and visually follows the types of dominant gene-expression patterns
weighting matrix, Wij , an autocorrelated spatial map can be generated that capture the most variance across all genes.
as 𝑥𝑖 = (𝐼 + 𝜌𝑊𝑖𝑗 )𝑢𝑗 , where I is the identity and 𝑢𝑗 ∼  (0, 1) is i.i.d. Gaus- To determine whether these low-dimensional components are spa-
sian noise. This defines the second model parameter: the spatial auto- tially embedded as gradients, we plotted them in Fig. 2B and D for E15.5
correlation strength, 𝜌. By setting d0 as a fixed fraction of brain size, and P28, respectively. We find clear spatial autocorrelation in these low-
𝑑0 = 𝑑max ∕𝑑0scale , we simulated a brain-size scaling rule consistent with a dimensional spatial gradients, an effect that is also seen across the other
linear rescaling of space (spatial expansion through development). Val- developmental time points (Fig. S1), and dummyTXdummy-(for lower-
ues for 𝜌 and 𝑑0scale were set through a gradient descent optimization order principal components (not shown). Having built some intuition for
procedure to best fit the empirical values of 𝜆 and A, yielding 𝜌 = 0.416 the low-dimensional spatial gradients of gene-expression in the develop-
and 𝑑0scale = 15.6. Some illustrative examples of the expression patterns ing mouse brain, in the remainder of this study we aim to quantitatively
generated are plotted in Fig. 6A. characterize these gradients across development using CGE(d) curves.
Note that, at small distances, all grid points are included in the CGE
calculation. But due to the grid’s finite size, at sufficiently large dis- 3.2. Distance dependence of correlated gene expression
tances, only a subset of grid points are included in the CGE calculation
(e.g., in the extreme case of the maximum distance, only the corners of Binned CGE(d) data with exponential fits are shown for each of the
the prism are included). This finite-size spatial sampling effect can bias seven developmental time points in Fig. 3A–G. The three-parameter ex-
the CGE(d) curve at large distances. Distance bins that are affected by ponential form captured the data well at all time points (all adjusted
this effect are labeled pink in Fig. 6B. R2 > 0.96). This exponential distance-dependence of gene-expression
For each simulated gene-expression dataset, we performed the same similarity provides a quantitative way to characterize the gradient-like
processing and analysis methodology as was applied to the empirical expression patterns shown in Fig. 2B and D.
data: (i) a random subsample of 1000 points was taken; (ii) voxel We next investigated whether the CGE(d) curves shown in
× gene expression data were normalized using a scaled sigmoid; (iii) Fig. 3, computed using all voxels, vary as a function of anatom-
CGE was computed; (iv) CGE(d) data were binned using 20 equiprob- ical division, focusing on the forebrain, midbrain, and hind-
able bins; (v) exponential decay was fit. We repeated the process 50 brain. We recomputed CGE(d) curves for each type of voxel
times to estimate error bars on each parameter estimate under the pair: forebrain–forebrain, forebrain–midbrain, midbrain–midbrain,
sources of stochasticity in the model: the generation of spatially auto- midbrain–hindbrain, or hindbrain–hindbrain, shown in Fig. S2. Most
correlated gene-expression maps, and the random spatial subsampling. of these anatomically distinguished CGE(d) curves decayed with dis-
We were then able to compare the parameter estimates from this model tance, but with different length scales and strengths. There were some
with the empirical data. Code for the modeling is available at https: deviations, such as an increase in midbrain–midbrain CGE at ~ 3 mm
//github.com/NeuralSystemsAndSignals/DevelopmentalExpression at E18.5, which causes a slight ‘hump-like’ deviation from an expo-
Modeling. nential form for the bulk CGE(d) curve (Fig. 3D). Consistent with an
increased gene-expression similarity within each anatomical division,
3. Results within-division CGE was higher than between-division CGE at a given
separation distance, d, across all developmental time points. Indeed,
Our approach to investigating the spatial embedding of transcrip- within-division CGE was almost always positive, even at the longest
tional gradients across mouse-brain development is shown schemat- separation distances, whereas between-division CGE often became neg-
ically in Fig. 1. At each of the seven time points in the ADMBA ative (i.e., gene-expression signatures are anti-correlated), even at in-
(Thompson et al., 2014), shown in Fig. 1A, we obtained the expression termediate separation distances. This between-division anti-correlation
of 1861 genes across voxels of the mouse brain (Fig. 1B). We then com- contributes to the overall negative baseline, f0 , of the exponential fits to
puted the correlated gene expression (CGE) for each pair of voxels as bulk CGE(d) in Fig. 3. The increase in CGE towards ~ 0 at the longest
the Pearson correlation coefficient between their expression signatures distance bin—seen for the postnatal times P4, P14, and P28—can mostly
H.Y.G. Lau, A. Fornito and B.D. Fulcher NeuroImage 224 (2021) 117395

Fig. 1. Schematic analysis pipeline. A We analyze gene-expression atlases across seven developmental time points. B At each time point, we created a voxel
× gene expression matrix (data shown for E11.5). We then compared the correlated gene expression (CGE) between each pair of voxels as a function of their physical
separation distance, d. C Plotting CGE(d) across equiprobable distance bins allowed us to analyze the spatial embedding of gene expression at each time point. We
fitted an exponential form CGE(𝑑) = 𝐴 exp(−𝑑∕𝜆) + 𝑓0 . Here we focus on the correlation length, 𝜆, and how it scales with brain development.

be attributed to a decrease in anti-correlation of expression profiles be- with an overall increase in transcriptional diversity within the brain
tween forebrain and hindbrain voxels. Postnatal time points were also across development.
distinguished by high expression similarity within forebrain voxels, even
at the longest separation distances (Fig. S2E–G). While the focus of this
study is to investigate CGE(d) curves using voxels from across the whole 3.3.1. Predicting the transcriptional correlation length in human cortex
brain, we note that this global gradient-like spatial embedding is not We next aimed to test the power-law form for the scaling of tran-
isotropic and non-specific, but indeed varies with anatomy. scriptional correlation length, 𝜆, which suggests that the mouse brain
We next investigated whether the CGE(d) curves vary across subsets maintains a characteristic spatial correlation length of transcriptional
of neuron-enriched, astrocyte-enriched, and oligodendrocyte-enriched similarity relative to its size. We also wanted to investigate whether the
genes (defined as being more than 1.5-fold enriched using data from relationship found across mouse-brain development could be extrapo-
Cahoy et al. (2008)). Compared to using all genes, CGE(d) curves and lated to make a prediction of the transcriptional correlation length, 𝜆, in
relationships between fitted CGE decay parameters and brain size re- human cortex. We used existing data for human cortex, dmax ≈ 148 mm
mained similar for all three gene sets, indicating that the spatial embed- (maximal anterior–posterior distance in the Glasser et al. (2016) parcel-
ding of gene expression does not vary strongly with genes enriched for lation) and 𝜆meas = 61.4 mm (95% confidence interval: 54.1–71.0 mm
human
each of these cell types. (Arnatkeviciūtė et al., 2019)).
The extrapolation of the power-law relationship fitted to the mouse
development data, 0.073𝑑max 1.22 , and setting 𝑑
3.3. Scaling of transcriptional correlation length with brain size max = 148 mm yields a pre-
pred
dicted 𝜆human = 33.4 mm (95% CI: 12.6 mm–88.7 mm). This extrapola-
When plotting the fitted CGE(d) exponential curves together (Fig. tion is shown in Fig. 5A. The prediction is consistent with the measured
S3A) and dummyTXdummy-(rescaling distances by brain size, dmax value, as shown in Fig. 5B (for all voxels, labeled ‘brain’). Thus, despite
(maximal anterior–posterior distance), as 𝑑̂ = 𝑑∕𝑑max (Fig. S3B), we differences in data type (ISH for < 2000 genes in mouse, microarray
found that CGE(d) approximately collapsed onto a common relative for > 20 000 genes in human) and spatial discretization (at the level
correlation length. To quantify this relationship, we plotted the fitted of voxels in mouse and a multimodal parcellation in human), we were
transcriptional correlation length, 𝜆, as a function of brain size, dmax , able to use mouse developmental data to predict a key spatial property
in Fig. 4A. We found that 𝜆 increases with dmax through development of human gene-expression gradients.
(Pearson’s 𝑟 = 0.98, 𝑝 = 6 × 10−5 ). Having only seven points over less We next investigated how this prediction varied when we restricted
than a decade of brain sizes, dmax , prevents rigorous inference of power- the fitting to a subset of voxels (dorsal pallium, forebrain, midbrain, and
law scaling (Clauset et al., 2009), but we have physical reasons to expect hindbrain), and when using a proportional linear scaling instead of the
a scaling relationship across brain expansion (Theodoni et al., 2020). To power-law form (Fig. 5B). We expected the closest estimate when us-
1.22 , shown
explore this possibility, we fitted a power-law, 𝜆 = 0.0735 × 𝑑max ing dorsal pallium voxels, as this is the structure that corresponds most
in Fig. 4B closely to the cortex. Indeed, dorsal pallium voxels yielded the closest
pred
The spatial embedding strength, A, controls the strength of the ex- prediction, 𝜆human = 47.3 mm (despite the dorsal pallium not being an-
ponential relationship in CGE relative to other effects such as noise. As notated to the reference atlas at E11.5). The worst human prediction
pred
shown in Fig. 4C, this strength is similar across development, A ≈ 0.75 was obtained using midbrain voxels, 𝜆human = 6.96 mm.
(Pearson’s 𝑟 = 0.65, 𝑝 = 0.1). The offset, f0 , corresponds to the CGE in Proportional linear scaling of 𝜆 with brain size, 𝜆∝dmax across mouse
the limit of large separation distances, d. As plotted in Fig. 4D, f0 de- development consistently yielded a poorer prediction, supporting a
creases linearly, becoming more negative across development (Pear- power-law form, 𝜆 ∝ 𝑑max𝜈 , with exponent 𝜈 > 1. With just seven points

son’s 𝑟 = −0.99, 𝑝 = 3 × 10−5 ). These increasingly anti-correlated gene- across mouse-brain development, our predictions have wide error bars;
expression patterns at the longest separation distances are consistent it will be important for future work to more closely test this scaling law
H.Y.G. Lau, A. Fornito and B.D. Fulcher NeuroImage 224 (2021) 117395

Fig. 2. Developmental mouse gene-expression patterns display low-dimensional spatial gradients. Here we plot results for E15.5 and P28 (see Fig. S1 for all
time points). We show voxel × gene expression matrices as heat maps for E15.5 and P28 in A and C, respectively. All brain voxels and genes passing quality-control
criteria are shown (11 314 × 1851 for E15.5, and 24 797 × 1846 for P28). Each gene has been normalized according to an outlier-robust sigmoidal transform,
followed by rescaling to the unit interval, from relative low expression for a given gene (blue) to relative high values (red). To visualize the structure in these
data, genes have been reordered by their correlation similarity, and within each labeled subdivision (forebrain, midbrain, hindbrain), voxels have been similarly
reordered. Horizontal black lines distinguish voxels from these three major subdivisions. The leading principal component of each matrix is labeled as ‘PC1’, and
visually follows the dominant expression pattern across genes. These leading principal components are plotted in three-dimensional brain space in B and D for E15.5
and P28, respectively. To aid visualization of the three-dimensional structure, a random subset of 5000 voxels is plotted with transparency. There is clear spatial
autocorrelation in these low-dimensional expression gradients in the developing mouse brain, which we quantify in this work.

and produce more constrained predictions using more comprehensive Fig. 6B for E11.5. Apart from a slight decrease of CGE < 0 at moderate
transcriptomic data, including across species. distances, and an increase back to CGE ≈ 0 at the highest distances, we
find a good fit to an exponential form. Note that the distance bins labeled
3.4. Modeling brain growth pink in Fig. 6B are sufficiently large relative to brain size that they in-
volve restricted sampling of brain space—only voxels towards the edge
To understand potential mechanisms underlying the scaling of tran- of the brain are able to contribute to the CGE calculation at these large
scriptional gradients through brain growth, we developed a simple spa- distances (whereas all voxels contribute at shorter distances). The quali-
tial model of transcriptional patterning and tested its predictions against tative similarity between the simulated CGE(d) curve and the measured
the empirical data characterized above. As described in Methods, we CGE(d) curves from the ADMBA data, suggest that an ensemble of genes
generated a random spatially autocorrelated map independently for with independent random, spatially autocorrelated transcriptional pat-
each gene’s expression pattern, evaluated in an idealized prism geom- terns can provide a good approximation for understanding how gene-
etry that matches the brain’s three-dimensional spatial extent at each expression patterns change through development.
developmental time point. To test whether the observed patterns are After fitting the two parameters of our model, 𝜌 and 𝑑0scale , to the
consistent with a simple rescaling of space in these expression maps, empirical data (see Methods), the model approximately reproduces the
we rescaled the spatial autocorrelation length, d0 , as a function of brain scaling of the correlation length, 𝜆, with brain size, dmax , seen in empir-
size, dmax . The model is therefore similar to uniform, isotropic stretching ical data (Fig. 6C), and the strength, A (Fig. 6D).
of three-dimensional gene-expression patterns that exist at the earliest
developmental time point (it reduces to this case when all brains have 4. Discussion
the same relative dimensions). The types of spatially autocorrelated pat-
terns that were generated are illustrated in Fig. 6A. Here we used data from the ADMBA to characterize how gene tran-
We first verified that our model reproduces an approximately expo- scriptional gradients are spatially embedded across mouse brain de-
nential dependence of CGE with distance at a given time point, shown in velopment. Complementing the well-known exponential distance rules
H.Y.G. Lau, A. Fornito and B.D. Fulcher NeuroImage 224 (2021) 117395

Fig. 3. Correlated gene expression (CGE) exhibits an approximately exponential decay with distance at all developmental time points. A–G: Across twenty
equiprobable distance bins, mean CGE of all pairs of voxels in each bin is plotted as circles, with horizontal lines showing the extent of each bin. Dashed lines indicate
one standard deviation either side of the mean of each bin. Fitted exponential curves, CGE(𝑑) = 𝐴 exp(−𝑑∕𝜆) + 𝑓0 , are shown solid.

Fig. 4. The correlation length of transcriptional coupling, 𝜆, increases with brain size. The three fitted parameters of the exponential form CGE(𝑑) =
𝐴 exp(−𝑑∕𝜆) + 𝑓0 , vary through brain development, plotted here as a function of brain size, dmax (computed along the anterior–posterior dimension): A transcriptional
correlation length, 𝜆 (mm) (with proportional linear fit dashed); B dmax –𝜆 as in A but plotted on logarithmic axes with the best power-law fit, 𝜆 = 0.0735𝑑 1.22 , shown
dashed; C spatial embedding strength, A; and D offset, f0 (with linear fit shown dotted). Each data point represents the fitted parameter value from a given time
point (labeled in A and B). Error bars, as symmetric 95% confidence intervals, are plotted on the three linear-axis plots, A, C, and D.

for structural connectivity in mammals (Ercsey-Ravasz et al., 2013; ing characterizing whether specific classes of genes drive similarities or
Gămănut et al., 2018; Horvát et al., 2016; Noori et al., 2017) that are differences in the spatial organization of gene expression across species.
connected by a scaling rule (Theodoni et al., 2020), we show that cor- A benefit of our method of characterizing spatial statistics of voxel-
related gene expression is also well characterized by an exponential dis- wise expression maps, as CGE(d) curves, is that it does not rely on any
tance rule at each point in mouse-brain development. Furthermore, the assumptions about pre-defined anatomical parcellations or areal bound-
characteristic correlation length of this embedding, 𝜆, was found to scale aries, and can thus be computed and compared across brains of differ-
with a linear dimension of brain size, dmax , as 𝜆 ∝ 𝑑max1.22 (Fig. 4). More ent shapes and sizes. Our visualizations of the voxel-wise expression, as
functionally specialized brains may be expected to encode more molec- clustered voxel × gene expression matrices accompanied by their lead-
ularly distinctive brain areas within the constraints of its brain size, cor- ing principal component (Fig. 2A and C), show that there are low-order
responding to a shorter relative 𝜆. This scaling relationship suggests that expression patterns that explain much of the variance across all mea-
the brain maintains a spatial trade-off between local molecular coordi- sured genes. When plotting this leading principal component in physi-
nation (uniform gene expression within a specialized brain area) and cal space (Fig. 2B and D), we see strong gradient-like spatial autocor-
longer-range molecular differentiation (between molecularly distinct, relation that is quantified here (across all genes) using CGE(d) curves.
functionally specialized areas) across development. We were able to use While these analyses provide clear evidence for gene-expression gra-
this scaling law to predict the transcriptional correlation length in the dients in the brain, they are not simple, non-specific spatial patterns,
adult human cortex (Fig. 5). Given that the brain-related genes exhibit but vary markedly across anatomical divisions. This is seen most clearly
strong consistency in their hierarchical patterning between mouse and when distinguishing CGE(d) by pairs of voxels that lie within or be-
human (Fulcher et al., 2019), it will be interesting for future work to tween each of the three major anatomical subdivisions: forebrain, mid-
further investigate these types of cross-species correspondences, includ- brain, and hindbrain (Fig. S2). Consistent with characteristic expression
H.Y.G. Lau, A. Fornito and B.D. Fulcher NeuroImage 224 (2021) 117395

Fig. 5. The transcriptional correlation length in human cortex, 𝜆human , can be predicted by extrapolating the scaling relationship between 𝜆 and brain
size through mouse-brain development. A Estimates of 𝜆 plotted as a function of brain size, dmax , using all voxels at each of the mouse developmental stages.
1.22
Extrapolation to the human cortex (𝑑max = 148 mm) is shown using a fitted power-law scaling relationship, 𝜆 = 0.0735𝑑max (blue), and a proportional linear relationship,
𝜆 = 0.13𝑑max (brown dashed). B We plot predictions trained from fitting to: all brain voxels, dorsal pallium (isocortex) voxels, forebrain voxels, midbrain voxels, and
hindbrain voxels. We also compare the predictions of power-law scaling (blue) to those from a proportional linear scaling (brown), with 95% confidence intervals
annotated.

Fig. 6. A simple spatial model for gene expression through development reproduces key features of the empirical data. A Illustrative ensemble of independent,
spatially autocorrelated gene-expression maps evaluated on three-dimensional prisms matching the dimensions of the brain at each point in development. Examples
are shown for two genes in a three-dimensional volume (expression levels shown as color). B Each ensemble exhibits an approximately exponential decay of correlated
gene expression (CGE) with distance, d, as CGE(𝑑) = 𝐴 exp(−𝑑∕𝜆) + 𝑓0 . Binned, model-simulated data (black) and fit (dotted blue) are plotted here for E11.5. Distance
bins for which not all grid points were sampled (due to finite-size effects) are filled pink. Exponential fit parameters of the model-simulated data (shown with dark
colors) and empirical data (shown with light colors) are plotted across development for: C spatial correlation length, 𝜆; D strength, A; and E offset, f0 . To aid visibility,
empirical values are slightly offset horizontally.

patterns for different divisions, seen visually in Fig. 2A and C, we found Characterizing CGE(d) curves allowed us to extract interesting pat-
that within-division CGE is higher than between-division CGE at a given terns using information combined across thousands of genes, even
separation distance. Future work could extend these types of analyses though data for each individual gene may be noisy. However, a limi-
to combine the benefits of a parcellation-free characterization that fa- tation of this bulk approach is the lack of specificity it provides about
cilitates comparisons across development, while also investigating how the patterning of specific genes of interest. Here, we applied a similar
properties of these spatial gradients vary as a function of anatomy. CGE(d) analysis to subsets of genes enriched for neurons, astrocytes,
H.Y.G. Lau, A. Fornito and B.D. Fulcher NeuroImage 224 (2021) 117395

or oligodendrocytes (Cahoy et al., 2008), and found similar results in Credit authorship contribution statement
each case. Zooming in further to smaller sets of genes, or even individ-
ual genes, would be valuable, especially in leveraging lists of cell-type Hoi Yan Gladys Lau: Software, Formal analysis, Data curation,
marker genes from single-cell RNA sequencing experiments (Mancarci Writing - original draft. Alex Fornito: Conceptualization, Writing -
et al., 2017; Tasic et al., 2018). However, this requires detailed assess- review & editing, Supervision. Ben D. Fulcher: Conceptualization,
ment of the reliability of expression data at the level of individual genes, Methodology, Software, Formal analysis, Data curation, Writing - re-
such as assessing the consistency across multiple repeated experiments view & editing, Supervision.
(Fulcher et al., 2019). We also note that the current dataset contains
an incomplete set of 2100 genes, and may therefore not include data Acknowledgements
for many relevant cell-type marker genes. Better coverage and quality
of these developmental expression data would allow us to comprehen- The authors would like to thank Aurina Arnatkevic̆iūtė for assisting
sively score individual genes to better understand whether specific types with human expression data. GL was supported by the Wong Ching Yee
of genes vary in, for example, their strength and spatial scale of gradient- Medical Elective Travel Grants from the University of Hong Kong. AF
like patterning, and how these properties change across development. was supported by the Sylvia and Charles Viertel Charitable Foundation.
Furthermore, combining validated expression data of individual genes
Supplementary material
with the spatial statistics introduced here could be used to construct
new, parcellation-free definitions of ‘differentially expressed genes’ to
Supplementary material associated with this article can be found in
test important biological hypotheses, such as whether the hour-glass pat-
the online version at doi:10.1016/j.neuroimage.2020.117395
tern of transcriptional variation, characterized in human (Pletikos et al.,
2014), holds similarly in mouse. References
Our analysis was performed at the voxel level using in situ hybridiza-
tion data from the ADMBA. This allowed us to explore CGE(d) at a much Anselin, L., 2013. Spatial econometrics: methods and models. Springer Science & Business
Media. 4
finer spatial resolution than previous analyses of correlated gene ex-
Arnatkeviciūtė, A., Fulcher, B.D., Fornito, A., 2019. A practical guide to linking brain-wide
pression, performed at the level of parcellations in adult mouse brain gene expression and neuroimaging data. NeuroImage 189, 353–367.
(Fulcher and Fornito, 2016) and human cortex (Arnatkeviciūtė et al., Arnatkeviciūtė, A., Fulcher, B.D., Pocock, R., Fornito, A., 2018. Hub connectivity, neu-
ronal diversity, and gene expression in the caenorhabditis elegans connectome. PLoS
2019). Nevertheless, each spatial element in our analysis still contains
Comput. Biol. 14 (2), e1005989.
transcriptional contributions from a large number of molecularly dis- Avalos, C.B., Maier, G.L., Bruggmann, R., Sprecher, S.G., 2019. Single cell transcriptome
tinct brain cells. The advent of three-dimensional in situ single-cell tran- atlas of the Drosophila larval brain. Elife 8, e50354.
scriptome profiling (Codeluppi et al., 2018; Moffitt et al., 2018; Wang Ball, G., Seidlitz, J., O’Muircheartaigh, J., Dimitrova, R., Fenchel, D., Makropoulos, A.,
Christiaens, D., Schuh, A., Passerat-Palmbach, J., Cordero-Grande, L., et al.,. Cortical
et al., 2018; Xia et al., 2019), which preserves spatial and transcrip- morphology at birth reflects spatio-temporal patterns of gene expression in the fetal
tomic information at the cellular level, will allow future analyses of cor- brain. bioRxiv.
related gene expression to be performed at even finer spatial scales. For Burt, J.B., Demirtas, M., Eckner, W.J., Navejar, N.M., Ji, J.L., Martin, W.J., Bernacchia, A.,
Anticevic, A., Murray, J.D., 2018. Hierarchy of transcriptomic specialization across
example, this could allow testing of whether the bulk scaling rule for human cortex captured by structural neuroimaging topography. Nat. Neurosci. 27,
CGE characterized here applies at the finer scale of neuronal circuits. 889.
As transcriptomic atlas data become available for other species, such as Burt, J. B., Helmer, M., Shinn, M., Anticevic, A., Murray, J. D., 2020. Genera-
tive modeling of brain maps with spatial autocorrelation. NeuroImage. bioRxiv.
Drosophila melanogaster (Avalos et al., 2019; Davie et al., 2018) and ze- https://doi.org/10.1016/j.neuroimage.2020.117038
brafish (Raj et al., 2018), we will also be able to more comprehensively Cahoy, J.D., Emery, B., Kaushal, A., Foo, L.C., Zamanian, J.L., Christopherson, K.S.,
compare the spatial organization of the brain’s molecular diversity, and Xing, Y., Lubischer, J.L., Krieg, P.A., Krupenko, S.A., Thompson, W.J., Barres, B.A.,
2008. A transcriptome database for astrocytes, neurons, and oligodendrocytes: a new
assess the cross-species validity of the scaling rule proposed here. Inves-
resource for understanding brain development and function. J. Neurosci. 28 (1),
tigating whether changes across mouse-brain development are consis- 264–278.
tent with the development of other species will also be important, and Clauset, A., Shalizi, C.R., Newman, M.E., 2009. Power-law distributions in empirical data.
SIAM Rev. 51 (4), 661–703.
could be tested in human, for example, using gene-expression assays
Codeluppi, S., Borm, L.E., Zeisel, A., La Manno, G., van Lunteren, J.A., Svensson, C.I.,
across human brain development (Ball et al.; Li et al., 2018). Linnarsson, S., 2018. Spatial organization of the somatosensory cortex revealed by
A simple model of independent, spatially autocorrelated transcrip- osmfish. Nat. Methods 15 (11), 932–935.
tional maps that scale with brain size (similar to the isotropic expansion Davie, K., Janssens, J., Koldere, D., De Waegeneer, M., Pech, U., Kreft, L., Aibar, S.,
Makhzami, S., Christiaens, V., González-Blas, C.B., 2018. A single-cell transcriptome
of transcriptional patterns set up early in development) reproduced key atlas of the aging drosophila brain. Cell 174 (4), 982–998.
features of CGE(d) scaling observed empirically. We kept the spatial au- Dillman, A.A., Hauser, D.N., Gibbs, J.R., Nalls, M.A., McCoy, M.K., Rudenko, I.N., Gal-
tocorrelation scale for our simulated gene-expression maps at a fixed ter, D., Cookson, M.R., 2013. MRNA expression, splicing and editing in the embryonic
and adult mouse cerebral cortex. Nat. Neurosci. 16 (4), 499.
proportion of brain size, consistent with linear, isotropic brain growth Ercsey-Ravasz, M., Markov, N.T., Lamy, C., Essen, D.C.V., Knoblauch, K., Toroczkai, Z.,
(but allowing us to use realistic brain dimensions from empirical mea- Kennedy, H., 2013. A predictive network model of cerebral cortical connectivity based
surements). Our results are thus consistent with molecular specialization on a distance rule. Neuron 80 (1), 184–197.
Fertuzinhos, S., Li, M., Kawasawa, Y.I., Ivic, V., Franjic, D., Singh, D., Crair, M., Šestan, N.,
occurring very early in development, with brain expansion stretching 2014. Laminar and temporal expression dynamics of coding and noncoding rnas in
the correlation length through development. While our simple isotropic the mouse neocortex. Cell Rep. 6 (5), 938–950.
model reproduces much of the CGE(d) relationship, our model is only Fornito, A., Arnatkeviciūtė, A., Fulcher, B.D., 2019. Bridging the gap between connectome
and transcriptome. Trends Cognit. Sci. 23 (1), 34–50.
evaluated against these coarse-grained properties, having blurred the
French, L., Pavlidis, P., 2011. Relationships between gene expression and brain wiring in
contribution of the many individual genes that display precise devel- the adult rodent brain. PLoS Comput. Biol. 7 (1), e1001049.
opmental spatiotemporal programs, and the variation in these CGE(d) Fulcher, B.D., Fornito, A., 2016. A transcriptional signature of hub connectivity in the
mouse connectome. Proc. Natl. Acad. Sci. USA 113 (5), 1435–1440.
curves as a function of anatomy (Fig. S2). It will thus be important for
Fulcher, B.D., Little, M.A., Jones, N.S., 2013. Highly comparative time-series analysis: the
future work to precisely investigate and model these important prop- empirical structure of time series and their methods. J. Roy. Soc. Interface 10 (83)
erties of brain development not considered here. By characterizing a 20130048–20130048.
generic variation in CGE, the current work may have a role to play in Fulcher, B.D., Murray, J.D., Zerbi, V., Wang, X.J., 2019. Multimodal gradients across
mouse cortex. Proc. Natl. Acad. Sci. USA 116 (10), 4689–4695.
this research effort. For example, genes that most strongly counter the Glasser, M.F., Coalson, T.S., Robinson, E.C., Hacker, C.D., Harwell, J., Yacoub, E., Ugur-
generic trends in CGE(d) could be considered as candidates for playing bil, K., Andersson, J., Beckmann, C., Jenkinson, M., Smith, S.M., Essen, D.C.V., 2016.
a more specific, targeted role in shaping brain development, facilitating A multi-modal parcellation of human cerebral cortex. Nature 536 (7615), 171–178.
Greig, L.C., Woodworth, M.B., Galazo, M.J., Padmanabhan, H., Macklis, J.D., 2013. Molec-
data-driven identification of genes that play important developmental ular logic of neocortical projection neuron specification, development and diversity.
roles. Nat. Rev. Neurosci. 14 (11), 755–769.
H.Y.G. Lau, A. Fornito and B.D. Fulcher NeuroImage 224 (2021) 117395

Gămănut, R., Kennedy, H., Toroczkai, Z., Ercsey-Ravasz, M., Essen, D.C.V., Knoblauch, K., Pantazatos, S.P., Li, X., 2017. Commentary: brain networks correlated gene expression
Burkhalter, A., 2018. The mouse cortical connectome, characterized by an ultra-dense supports synchronous activity in brain networks. Sci. Front. Neurosci. 11, 412.
cortical graph, maintains specificity by distinct connectivity profiles. Neuron 97 (3), Paquola, C., De Wael, R.V., Wagstyl, K., Bethlehem, R.A.I., Hong, S.-J., Seidlitz, J., Bull-
698–715. more, E.T., Evans, A.C., Mišić, B., Margulies, D.S., Smallwood, J., Bernhardt, B.C.,
Hawrylycz, M.J., Bernard, A., Lau, C., Sunkin, S.M., Chakravarty, M.M., Lein, E., 2019. Microstructural and functional gradients are increasingly dissociated in trans-
Jones, A.R., Ng, L., 2010. Areal and laminar differentiation in the mouse neocortex modal cortices. PLoS Biol. 17 (5), e3000284.
using large scale gene expression data. Methods 50 (2), 113–121. Pletikos, M., Sousa, A.M.M., Sedmak, G., Meyer, K.A., Zhu, Y., Cheng, F., Li, M., Kawa-
Henderson, J.A., Robinson, P.A., 2011. Geometric effects on complex network structure sawa, Y.I., Šestan, N., 2014. Temporal specification and bilaterality of human neocor-
in the cortex. Phys. Rev. Lett. 107, 018102. tical topographic gene expression. Neuron 81 (2), 321–332.
Henderson, J.A., Robinson, P.A., 2014. Relations between the geometry of cortical gyrifi- Raj, B., Wagner, D.E., McKenna, A., Pandey, S., Klein, A.M., Shendure, J., Gagnon, J.A.,
cation and white-matter network architecture. Brain Conn. 4 (2), 112–130. Schier, A.F., 2018. Simultaneous single-cell profiling of lineages and cell types in the
Horvát, S., Gămănut, R., Ercsey-Ravasz, M., Magrou, L., Gămănut, B., Essen, D.C.V., vertebrate brain. Nat. Biotech. 36 (5), 442.
Burkhalter, A., Knoblauch, K., Toroczkai, Z., Kennedy, H., 2016. Spatial embedding Richiardi, J., Altmann, A., Greicius, M., 2017. Distance is not everything in imaging ge-
and wiring cost constrain the functional layout of the cortical network of rodents and nomics of functional networks: Reply to a commentary on correlated gene expression
primates. PLoS Biol. 14 (7), e1002512. supports synchronous activity in brain networks. bioRxiv 132746.
Huntenburg, J.M., Bazin, P.-L., Bazin, P.L., Margulies, D.S., 2017. Large-scale gradients in Roberts, J.A., Perry, A., Lord, A.R., Roberts, G., Mitchell, P.B., Smith, R.E., Calamante, F.,
human cortical organization. Trends Cognit. Sci. 22 (1), 21–31. Breakspear, M.J., 2016. The contribution of geometry to the human connectome. Neu-
Krienen, F.M., Yeo, B.T.T., Ge, T., Buckner, R.L., Sherwood, C.C., 2016. Transcriptional roImage 124, 379–393.
profiles of supragranular-enriched genes associate with corticocortical network archi- Robinson, P.A., 2019. Physical brain connectomics. Phys. Rev. E 99 (1), 012421.
tecture in the human brain. Proc. Natl. Acad. Sci. USA 113 (4), E469–E478. Song, H.F., Kennedy, H., Wang, X.J., 2014. Spatial embedding of structural similarity in
Kunst, M., Laurell, E., Mokayes, N., Kramer, A., Kubo, F., Fernandes, A. M., Förster, D., the cerebral cortex. Proc. Natl. Acad. Sci. USA 111 (46), 16580–16585.
Maschio, M. D., Baier, H., 2019. A cellular-resolution atlas of the larval zebrafish Stiso, J., Bassett, D.S., 2018. Spatial embedding imposes constraints on neuronal network
brain. Neuron 103 (1), 21–38. architectures. Trends Cognit. Sci. 22 (12), 1127–1142.
Lee, C.-K., Sunkin, S.M., Kuan, C., Thompson, C.L., Pathak, S., Ng, L., Lau, C., Fischer, S., Theodoni, P., Majka, P., Reser, D. H., Wjcik, D. K., Rosa, M. G. P., Wang, X. J., 2020.
Mortrud, M., Slaughterbeck, C., Jones, A., Lein, E., Hawrylycz, M.J., 2008. Quantita- Structural attributes and principles of the neocortical connectome in the marmoset
tive methods for genome-scale analysis of in situ hybridization and correlation with monkey. bioRxiv. 10.1101/2020.02.28.969824
microarray data. Genome Biol. 9 (1), R23. Tasic, B., Yao, Z., Graybuck, L.T., Smith, K.A., Nguyen, T.N., Bertagnolli, D., Goldy, J., Gar-
Li, M., Santpere, G., Kawasawa, Y. I., Evgrafov, O. V., Gulden, F. O., Pochareddy, S., ren, E., Economo, M.N., Viswanathan, S., Penn, O., Bakken, T., Menon, V., Miller, J.,
Sunkin, S. M., Li, Z., Shin, Y., Zhu, Y., et al., 2018. Integrative functional genomic Fong, O., Hirokawa, K.E., Lathia, K., Rimorin, C., Tieu, M., Larsen, R., Casper, T.,
analysis of human brain development and neuropsychiatric risks. Science 362, 6420. Barkan, E., Kroll, M., Parry, S., Shapovalova, N.V., Hirschstein, D., Pendergraft, J.,
Mancarci, B.O., Toker, L., Tripathy, S., Li, B., Rocco, B., Sibille, E., Pavlidis, P., 2017. Sullivan, H.A., Kim, T.K., Szafer, A., Dee, N., Groblewski, P., Wickersham, I., Cetin, A.,
Cross-laboratory analysis of brain cell type transcriptomes with applications to inter- Harris, J.A., Levi, B.P., Sunkin, S.M., Madisen, L., Daigle, T.L., Looger, L., Bernard, A.,
pretation of bulk tissue data. eNeuro 4 (6), e0212. Phillips, J., Lein, E., Hawrylycz, M.J., Svoboda, K., Jones, A.R., Koch, C., Zeng, H.,
Margulies, D.S., Ghosh, S.S., Goulas, A., Falkiewicz, M., Huntenburg, J.M., Langs, G., Bez- 2018. Shared and distinct transcriptomic cell types across neocortical areas. Nature
gin, G., Eickhoff, S.B., Castellanos, F.X., Petrides, M., Jefferies, E., Smallwood, J., 563 (7729), 72–78.
2016. Situating the default-mode network along a principal gradient of macroscale Thompson, C.L., Ng, L., Menon, V., Martinez, S., Lee, C.-K., Glattfelder, K., Sunkin, S.M.,
cortical organization. Proc. Natl. Acad. Sci. USA 113 (44), 12574–12579. Henry, A., Lau, C., Dang, C., Garcia-Lopez, R., Martinez-Ferre, A., Pombero, A., Ruben-
Markov, N.T., Ercsey-Ravasz, M., Van Essen, D.C., Knoblauch, K., Toroczkai, Z., stein, J.L.R., Wakeman, W.B., Hohmann, J., Dee, N., Sodt, A.J., Young, R.C., Smith, K.,
Kennedy, H., 2013. Cortical high-density counterstream architectures. Science 342 Nguyen, T.N., Kidney, J., Kuan, L., Jeromin, A., Kaykas, A., Miller, J., Page, D.,
(6158). Orta, G., Bernard, A., Riley, Z., Smith, S., Wohnoutka, P., Hawrylycz, M.J., Puelles, L.,
Mody, M., Cao, Y., Cui, Z., Tay, K.-Y., Shyong, A., Shimizu, E., Pham, K., Schultz, P., Jones, A.R., 2014. A high-resolution spatiotemporal atlas of gene expression of the de-
Welsh, D., Tsien, J.Z., 2001. Genome-wide gene expression profiles of the developing veloping mouse brain. Neuron 83 (2), 309–323.
mouse hippocampus. Proc. Natl. Acad. Sci. USA 98 (15), 8862–8867. Wang, X., Allen, W.E., Wright, M.A., Sylwestrak, E.L., Samusik, N., Vesuna, S., Evans, K.,
Moffitt, J.R., Bambah-Mukku, D., Eichhorn, S.W., Vaughn, E., Shekhar, K., Perez, J.D., Ru- Liu, C., Ramakrishnan, C., Liu, J., 2018. Three-dimensional intact-tissue sequencing
binstein, N.D., Hao, J., Regev, A., Dulac, C., 2018. Molecular, spatial, and functional of single-cell transcriptional states. Science 361 (6400).
single-cell profiling of the hypothalamic preoptic region. Science 362 (6416). Xia, C., Fan, J., Emanuel, G., Hao, J., Zhuang, X., 2019. Spatial transcriptome profiling
Noori, H.R., Schöttler, J., Ercsey-Ravasz, M., Cosa-Linan, A., Varga, M., Toroczkai, Z., by MERFISH reveals subcellular RNA compartmentalization and cell cycle-dependent
Spanagel, R., 2017. A multiscale cerebral neurochemical connectome of the rat brain. gene expression. Proc. Natl. Acad. Sci. USA 116 (39), 19490–19499.
PLoS Biol. 15 (7), e2002612.