Spatial protein analysis in developing

tissues: a sampling-based image

royalsocietypublishing.org/journal/rstb processing approach
Karolis Leonavicius, Christophe Royer, Antonio M. A. Miranda,
Richard C. V. Tyser, Annemarie Kip and Shankar Srinivas†
Research
Department of Physiology, Anatomy and Genetics, University of Oxford, Oxford OX1 3QX, UK
Cite this article: Leonavicius K, Royer C,
AMAM, 0000-0003-4281-614X; SS, 0000-0001-5726-7791
Miranda AMA, Tyser RCV, Kip A, Srinivas S.
2020 Spatial protein analysis in developing Advances in fluorescence microscopy approaches have made it relatively
tissues: a sampling-based image processing easy to generate multi-dimensional image volumes and have highlighted
approach. Phil. Trans. R. Soc. B 375: 20190560. the need for flexible image analysis tools for the extraction of quantitative
information from such data. Here we demonstrate that by focusing on sim-
http://dx.doi.org/10.1098/rstb.2019.0560
plified feature-based nuclear segmentation and probabilistic cytoplasmic
detection we can create a tool that is able to extract geometry-based infor-
Accepted: 22 May 2020 mation from diverse mammalian tissue images. Our open-source image
analysis platform, called ‘SilentMark’, can cope with three-dimensional
noisy images and with crowded fields of cells to quantify signal intensity
One contribution of 15 to a discussion meeting
in different cellular compartments. Additionally, it provides tissue geometry
issue ‘Contemporary morphogenesis’.
related information, which allows one to quantify protein distribution with
respect to marked regions of interest. The lightweight SilentMark algorithms
Subject Areas: have the advantage of not requiring multiple processors, graphics cards or
cellular biology, computational biology, training datasets and can be run even with just several hundred megabytes
of memory. This makes it possible to use the method as a Web application,
developmental biology
effectively eliminating setup hurdles and compatibility issues with operating
systems. We test this platform on mouse pre-implantation embryos, embryo-
Keywords: nic stem cell-derived embryoid bodies and mouse embryonic heart, and
image analysis, immunofluorescence, protein relate protein localization to tissue geometry.
quantification, segmentation, gene expression, This article is part of a discussion meeting issue ‘Contemporary
morphogenesis’.
cell and developmental biology

Author for correspondence:

Shankar Srinivas 1. Introduction
e-mail: [email protected] One of the main areas of focus in developmental biology is how tissue complexity
is generated and how diverse cell types are consistently organized into a complete
organism. A common challenge in the field is the three-dimensional spatial analy-
sis of protein distribution within tissues and cells [1,2]. The function of a protein
can vary depending on the cellular compartment it is localized to and differential
expression of proteins across tissues during development is important in patter-
ing those tissues. Cutting edge optical microscopy and image analysis tools are
now routinely used for spatial analysis of developing tissues from various organ-
isms [3–9]. Given the variability of different tissues, existing software often must
be optimized to work for specific cases. For example, full cell segmentation tools
like MorphoGraphX [3] or Packing Analyzer [10] work extremely well with samples
such as plants and Drosophila, where high-quality fluorescence signals and low
noise make it possible to segment complete cell shapes.
As the signal-to-noise ratio of the image drops or features become indistinct, it
† becomes more difficult to accurately segment and analyse individual cells. One of
Present address: Droplet Genomics, D. Poškos
g. 55-4, 08114, Vilnius, Lithuania. the most reliable solutions that is still widely used is manual segmentation [11,12],
because humans are adept at feature recognition. However, manual outlining
is time consuming, which is particularly a problem for volume or time-lapse
Electronic supplementary material is available data. To automate image segmentation, there has been a shift towards machine
online at https://doi.org/10.6084/m9.figshare.
c.5056727.
© 2020 The Authors. Published by the Royal Society under the terms of the Creative Commons Attribution
License http://creativecommons.org/licenses/by/4.0/, which permits unrestricted use, provided the original
author and source are credited.
learning segmentation approaches. A notable example is RACE aspect is that it allows one to process routine lower-quality 2
[9], which uses machine learning for tissue segmentation into images, investigate multiple cell compartments and quantify

royalsocietypublishing.org/journal/rstb
individual cells. Arguably, this is one of the most successful the spatial distribution of fluorescence within cells and tissues.
approaches yet for automated segmentation, especially from Here we demonstrate the use of our algorithm on mouse pre-
membrane signals. However, it is still often impossible to cor- and post-implantation embryos, as well as mouse embryonic
rectly segment cells based on their three-dimensional outlines stem cell-derived embryoid bodies.
alone, because of limitations in image quality.
Dealing with multiple spatial and fluorescence variables
raises additional challenges in terms of signal normalization, 2. Results and discussion
which is crucial in correcting for variations due to image acqui-
To quantify protein levels in different cellular compartments
sition or optical heterogeneity within tissues. A commonly
and determine the relative proportion of the plasma membrane
used approach involves taking ratios of nearby cellular
in different domains (apical, junctional and basolateral), we
compartments, such as nuclei and cytoplasm. This has been
developed a sampling-based algorithm named ‘SilentMark’,
previously used for example in investigating the Hippo path-
implemented in Matlab and Python. The overall strategy relies

Phil. Trans. R. Soc. B 375: 20190560

way, where YAP localization could be quantified by taking
on detecting well-resolved tissue elements, such as nuclei or
the ratio of nuclear and cytoplasmic fluorescence while at the
membranes, and then analysing their closest environment to
same time normalizing the raw signals [13].
take samples of subcellular region fluorescence. Depending on
Even in cases where image quality prevents reliable
which cell marker is available, the first step is to detect well-
segmentation, image data often contain a great deal of useful
resolved objects such as nuclei or membranes, after which
quantitative information. To extract such information, we have
their closest environment is analysed. Care has been taken
developed a sampling-based approach to measure nuclear,
to implement efficient algorithms for three-dimensional convo-
cytoplasmic and plasma membrane fluorescence levels. Impor-
lution and searching, which enables work with images
tantly, the localization of these samples within the image
containing millions of voxels.
volume is retained, which can then be used for detailed statisti-
cal analysis of the spatial distribution of the detected proteins.
This approach works across the scale of individual cells to (a) Nucleus-based fluorescence sampling
entire tissues. At the level of the cell, it allows us to quantify Available nuclear detection techniques differ in their accuracy,
relative protein levels in different cellular compartments, throughput and noise tolerance. One of the fastest and most
while at the level of the tissue it allows us to identify cell popu- robust methods for spherical object detection is based on the
lations based on relative protein expression, statistically test Hough transform [14], which is a tensor voting technique for
the difference between these populations and analyse how the most probable circle positions within an image. The original
their location within the tissue correlates with protein levels. algorithm is developed for two-dimensional image analysis.
Our approach employs well-defined features such as nuclei However, it can be extended for three-dimensional measure-
to perform partial but robust segmentation using well- ments by slicing the image into sections, analysing them as
established techniques [6,14]. Alternatively, well-resolved mem- separate images and then deconvolving into three-dimensional
brane signals or manually defined regions can also be used. If objects, as done previously [9]. In this case, confocal image
nuclei are used as points of reference, nearby voxels are used stacks were analysed as separate images to detect circular
to sample nearby nuclear and cytoplasmic regions, identified nuclear cross-sections, the coordinates of which were then
based on nuclear stain intensity. Simultaneous sampling of merged to generate a three-dimensional point cloud of nuclear
different cell compartments allows one to normalize the data centres of mass (figure 1a). Using these nuclear coordinates as
with respect to experimental variation and tissue optical hetero- reference points, a region from the confocal image is cropped
geneity. Regions of membrane sampling are located by using a around each detected point and classified according to
combination of detected nuclei positions and intensity of a the intensity of the nuclear stain as either a cytoplasmic or
membrane stain. Membrane signal quantification can be further nuclear region. The fluorescence from both regions is then
improved by using commercially available software such as averaged, with cytoplasm pixels being weighted according to
IMARIS to manually outline membrane regions, which are then their distance from the detection point (figures 1b and 4a,b).
automatically partitioned by our software into contacting (baso- This was done to represent the probability that pixels closer
lateral), exposed (apical) and junctional domains. Since the to the nucleus are more likely to belong to the same cell as the
spatial distribution of proteins is of particular interest to devel- nucleus. The fluorescence levels along with the nuclear coordi-
opmental biologists, we have also incorporated the ability to nates were recorded and this formed the basis of segmentation
define points or regions of interest in the image volume, so and quantification.
that the relative positions of these sampled regions with respect If a membrane stain was available in addition to the nuclear
to these points of interest can be recovered. stain, it was possible to add another fluorescence reading
We have implemented this approach as a software package for exposed and contacting membrane regions, which is
called ‘SilentMark’, designed for general use. The package is a particularly useful for looking at membrane-localized proteins.
standalone GUI-based application written in Matlab, available To detect the outside surface, the membrane fluorescence
at https://process.innovation.ox.ac.uk/software/p/13299a/ marker was thresholded and the outer pixels were assigned
silentmark-academic/1. The core algorithm has also been to the closest nucleus as an exposed membrane sample. To
implemented in Python as a Web application and is hosted detect the contacting membrane samples, a line was drawn
in the analysis section at https://dropletgenomics.com. Some between two neighbouring nuclei, and the highest intensity
limitations of our sampling-based approach are that it does region for membrane fluorescence along the line was desig-
not provide information on cell shape, or the total sum of flu- nated as a contacting membrane sample and was assigned
orescence, or facilitate cell tracking. However, the unique to both nuclei.
 (a) confocal nuclear 3
image stack coordinates

royalsocietypublishing.org/journal/rstb
x,y,z
1 x,y,z
2
3
x,y,z
x,y,z
z

x,y

Phil. Trans. R. Soc. B 375: 20190560

layer histogram noise filtering Hough circle deconvolution
separation equalization detection

(b) nuclear nuclear

coordinates fluorescence

x,y,z 1
2

+ 3
z

region pixel confocal fluorescence

isolation classification mask creation image stack analysis results

nucl x,y,z cyto nuclear

nuclei 1

.........
.........
.........
x,y,zz cyto 2
cytoplasm + 3
z

Figure 1. Schematic illustrating strategy for detecting the coordinates of the nuclear centre of mass (a) and for extracting nuclear and cytoplasmic fluorescence
based on nuclear coordinates (b).

(b) Membrane-based fluorescence sampling manually outline membrane regions, which SilentMark then
In cases where nuclear markers are not available as a reference, automatically partitions into contacting, exposed and junc-
membrane fluorescence, if available, can be used to create tional domains (figure 4d), where proteins are likely to play
the points of reference for subsequent analysis. To detect different roles.
membrane regions in space, a three-dimensional stack was
processed using contrast-limited adaptive histogram normali-
zation and Gaussian blur to reduce noise. Following this, a (c) Development of SilentMark software
three-dimensional Canny edge detector [15] was applied to We have developed these sampling-based algorithms into a
find membrane edges, to which spheres were fitted using an software package designed for general use. The package
outlier-tolerant RANSAC algorithm. The detected membranes is a standalone application, based on a GUI designed on
were represented by a location of the fitted sphere and a vector Matlab. The algorithm has also been implemented in Python
representing membrane curvature and direction (figure 2a). as a Web application. It accepts three-dimensional images as
This information would then be the starting point for sampling stacks of TIFF files or the Zeiss confocal microscopy ‘.lsm’
membrane and cytoplasmic fluorescence intensities. The format. The user is required to enter two intuitive and robust
advantage of this method is that it looks for fragments of parameters—an estimate of the diameter of the nuclei and an
membranes, which is more tolerant to noise than other estimate of nuclear channel brightness for thresholding. These
approaches that rely on total membrane segmentation. These two minimal parameters allow one to work with different
detected membrane regions can now be used as sampling types of tissues and image qualities. All the remaining par-
masks for membrane and cytoplasmic fluorescence. The ameters needed for the analysis are derived and optimized by
mask considers a narrow region in space around the detected the software.
point to sample membrane fluorescence and draws a normal The software will output a list (CSV format) of detected
to the membrane surface to sample cytoplasm on either side. objects (nuclei or membrane fragments), each of which will
These masks are used on three-dimensional image stacks to have a measure of nuclear and cytoplasmic fluorescence as
analyse membrane and cytoplasmic fluorescence distribution well as membrane fluorescence, if it is labelled. The data will
in space (figure 2b). Membrane signal quantification can be also contain object coordinates to describe spatial protein distri-
improved by using commercially available software IMARIS to bution, automatically calculated distance to exposed surface
 (a) confocal membrane 4
image stack coordinates

royalsocietypublishing.org/journal/rstb
i,j,k
1 x,y,z
2
3
i,j,k
z
x,y,z

Phil. Trans. R. Soc. B 375: 20190560

histogram noise filtering 3D Canny RANSAC
equalization edge detection sphere fitting

(b) membrane sampling confocal fluorescence

coordinates masks image stack analysis results

i,j,k cyto
x,y,z cyto membrane
1
x,y,z

.........
.........
.........
2
i,j,k
mem
+ 3

z
x,y,z
cyto

Figure 2. Schematic illustrating the membrane fragment detection algorithm (a) and analysis of membrane and cytoplasmic fluorescence based on membrane
fragment detection (b).

(a) nucleus
cytoplasm exposed
nuclear cytoplasmic membrane
membrane algorithm use scenario surface
stain error stain error stain error
error
need high
accuracy or
manual cell outlines
exposed
3.2% –6.6% 0.7% 6.4%
surface

66 mm only a
membrane membrane
n.a. n.a. –28% –15.7%
(b) fluorescence stain is
available
only nuclear
nuclear
automatic fluorescence
stain is n.a. –7.8% –8.6% n.a.
available
membrane and nuclear and
nuclear membrane
n.a. –7.8% –8.6% –4.2%
fluorescence stains are
combination available

Figure 3. Accuracy of sampling-based image segmentation. (a) A section of a computer-generated image for testing image analysis methods. (b) A three-dimen-
sional rendering of computer-generated objects for testing image analysis methods. Method accuracy was assessed by comparing measured fluorescence levels with
known intensity values in computer-generated images. Four models, each consisting of four to six cells, were used in this test.

and distance to regions of interest, if they are designated and, in addition to quantifying fluorescence intensity, could
manually in three dimensions. This information can then be also be used to estimate the proportion of ‘exposed’ cell surface,
processed with statistical software tools for further analysis. that is, cell surface not in contact with surrounding cells. In gen-
eral, intensity measurement errors were below 10%, except for
feature detection relying on membrane fragments. The most
(d) Method validation with CAD images likely source of error for membrane-based feature detection
To validate this probabilistic sampling-based approach, we
came from the detection algorithm while fitting spheres,
first used computer-generated three-dimensional images to
which is responsible for identifying both membrane segment
compare the accuracy of our automated sampling-based
position and orientation.
approaches against standard manual outlining (using IMARIS;
figure 3). Mock-up tissues composed of several spherical
objects representing cells were created on Google SketchUp (e) Method validation with pre-implantation embryos
and converted to a stack of images, which were then filled The mouse pre-implantation embryo is a model system where
with pixels of known brightness. Unsurprisingly, manual out- cell fate is strongly influenced by tissue geometry [16,17].
lining was the most accurate method of image segmentation The underlying basis for cell-type specification in the
 (a) (c) 5
(e) manual/automatic quantification error

royalsocietypublishing.org/journal/rstb
0.020 nuclear
cytoplasmic

normalized frequency
0.015

(b) (d) 0.010

0.005

0
–40 –20 0 20 40 60
error (%)

Phil. Trans. R. Soc. B 375: 20190560

(f ) transcription levels of key proteins (g) the first cell fate decision
1.00 Cdx2
1.0

fraction of exposed cell surface

normalized transcript number

Eomes
Nanog
0.8
0.75 Oct3/4

0.6
0.50
0.4

0.25 YAP, N/C

0.2 0.7
1.0
0 0 2.5

2 8 16 32 2-cell 8-cell 16-cell 32-cell 64-cell

embryo stage embryo stage

(h) spatial YAP patterning (i) CDX2 and YAP relationship

64 cells 3.0
32 cells
2.5
CDX2 ratio, ln(N/C)

2 16 cells
YAP ratio, ln(N/C)

8 cells
2 cells 2.0

1 1.5
1.0
0 0.5
0.0 32 cells
16 cells
–1 –0.5 8 cells

0 1 2 3 4 –0.5 0 0.5 1.0 1.5 2.0

distance from surface normalized by cell radius YAP ratio, ln(N/C)

Figure 4. Quantifying lineage determinants during mouse pre-implantation development. (a) Sampling algorithm for detecting three subcellular regions; (b) three-
dimensional rendering of different sampling regions. Red and green are exposed and contacting membranes, respectively; grey is cytoplasm; and cyan is nucleus.
(c) Three-dimensional rendering of a mouse embryo (grey, actin; red, YAP; cyan, DAPI); (d ) 8-cell embryo in which cell membranes have been manually segmented
and then membrane regions automatically sub-segmented into ‘exposed’ (blue), ‘junctional’ (white) and basolateral (red). (e) Error distribution of automated
sampling-based quantitation compared with manual segmentation-based quantitation. ( f ) Levels of key transcription factors during development. (g) YAP distri-
bution in the early pre-implantation embryo. (h) Spatial YAP patterning during pre-implantation development. (i) YAP and CDX2 relationship in the morula and early
blastocyst. Scale bars in all panels represent 50 µm. N/C, ratio of nuclear to cytoplasmic fluorescence.

pre-implantation embryo has been studied in great depth and excluded from the nucleus and these cells go on to become
this large amount of existing information makes it ideal for the pluripotent inner cell mass [18–20].
validating our analytical approach. To estimate the errors resulting from the sampling-
During the process of the first cell fate decision in mam- based approach (figure 4a), we used manual cell outlining
mals, the protein YAP shuttles between the cytoplasm (figure 4d) to create a ground-truth dataset of pre-
(where it is inactive) and the nucleus (active), in a tightly implantation embryos stained for YAP (figure 4b,c), (n = 67
regulated manner. The differential localization of YAP embryos, comprising 20 2-cell, 15 8-cell, 17 16-cell, 12
protein in outside and inside cells ultimately determines 32-cell and three 64-cell embryos; examples in figure 5 and
cell fate in this context, and the ratio of nuclear to cytoplasmic electronic supplementary material, movie S1). For nuclear
YAP reflects the position of cells. In outside cells, YAP is in and cytoplasmic fluorescence measurements, mean error
the nucleus, where it can interact with nuclear TEAD4 to was lower than 5% (figure 4e), which indicates a low bias
drive the expression of tissue-specific genes such as Cdx2 during the automated sampling. Most of the errors result
to give rise to the trophectoderm. In inside cells, YAP is from poor contrast of nuclear stain (which particularly
 6
(a) (b) DAPI

royalsocietypublishing.org/journal/rstb
Phil. Trans. R. Soc. B 375: 20190560
(c) YAP (d) CDX2

Figure 5. (a) An example of a manually segmented 16-cell mouse embryo, with cell outlines in green and nuclei in yellow. (b–d ) Example of a 16-cell mouse
embryo stained for DAPI, YAP and CDX2. In all images, cell outlines are in yellow, stained with phalloidin (F-actin). The scale bar represents 50 µm.

occurs deeper into the tissue), as this makes it harder to (f ) Method validation with embryonic stem cell-derived
correctly classify voxels as nuclear or cytoplasmic. Another
source of errors is sampling outside the tissue, as the software embryoid bodies
does not calculate bounding tissue surfaces. Similar to mouse pre-implantation embryos, mouse embryonic
We used our automated sampling-based approach to stem cell aggregates also display inside–outside patterning
explore the relationship between YAP and CDX2 levels and during embryoid body (EB) formation in vitro [24]. The outer
the extent to which this correlates with a blastomere’s position, layer of cells forms an endoderm-like layer around an inner
as measured by exposure to the outside or distance of a cell from core that is epiblast-like. These aggregates tend to be composed
the surface. We used embryos stained for YAP, CDX2 and of crowded cells (figure 6a,b) and represent a more challenging
DAPI. Published single-cell RNA-seq data on mouse task for both manual and automated segmentation or
pre-implantation embryos [21] (GEO accession: GSE45719) sampling approaches. To test its capabilities, we applied our
indicate that Cdx2 becomes upregulated at the 8-cell stage sampling-based image approach to investigate endoderm for-
(figure 4f, N = four 2-cell, four 8-cell, four 16-cell and three mation during EB differentiation. We stained EB for the
32-cell embryos). Our analysis shows that at the 8-cell stage, endoderm marker GATA6 (figure 6b), which specifically
blastomeres begin to show signs of segregating into two popu- stains outside cells, and quantified the relative levels of nuclear
lations, with high or low nuclear to cytoplasmic YAP ratios GATA6 expression with respect to the surface of EB (n = 6).
(figure 4g and example in figure 5; N = 20 2-cell, 15 8-cell, Quantitative statistical analysis (figure 6c) shows preferential
17 16-cell, 12 32-cell and three 64-cell embryos). More robust GATA6 expression at the outer layer of the EB (distance to
segregation of these two populations occurs at the 16-cell outer surface is smaller than two cell radii).
stage, approximately 12 h after Cdx2 transcript appearance at
the 8-cell stage. Analytical detail makes it possible to demon-
strate that populations are separated gradually, as YAP (g) Method validation with embryonic hearts
translocates to the cytoplasm in inside cells (figure 4h and elec- While mouse pre-implantation embryos present an ideal case
tronic supplementary material, figure S1). It is also evident, that for testing automatic image segmentation, because it is
the CDX2–YAP correlation is stronger at the 32-cell stage than possible to also manually outline the cells for comparison,
the 16-cell stage (figure 4i and example in figure 5; N = seven they do not necessarily represent a typical mammalian
8-cell, nine 16-cell and six 32-cell embryos). This is consistent tissue. A more representative mammalian tissue is the
with the presence of additional stabilizing mechanisms for embryonic heart (figure 6d–g), with a larger number of cells
robust lineage specification [22,23]. and closely spaced nuclei.
 (c) 2.5 7

(a) (b) 2.0

GATA6 ratio, ln(N/C)

royalsocietypublishing.org/journal/rstb
1.5
1.0
0.5
0
–0.5

0 2.0 4.0 6.0 8.0 10.0

normalized distance from cell surface
(cell radii)

(h) –2 –1 0 1 2 3

Phil. Trans. R. Soc. B 375: 20190560

head
(d) (e) heart
2.0 head region 2.0
heart region
1.5 1.5

SRF ratio, ln(N/C)

F-actin F-actin
DAPI NKX2-5 1.0 1.0
0.5 0.5
(f) (g)
0 0
–0.5 –0.5
F-actin F-actin –1.0 –1.0
SRF NKX2-5

heart
head
–2 –1 0 1 2 3
NKX2-5 ratio, ln(N/C)

Figure 6. Automated quantitative analysis of differentiated mouse stem cell-derived embryoid bodies and mouse embryonic heart. (a,b) Optical sections of a differ-
entiating mouse stem cell EB stained with DAPI (a, cell nuclei) and GATA6 (b, marking outer endoderm layer). (c) GATA6 localization in the EB as a function of distance
from exposed surface (n = 6 EBs). The approximately single-cell-wide outer layer is distinct from the remaining EB. (d–f ) Optical section of an embryonic heart stained
for F-actin (magenta in all panels, showing cell outlines), DAPI (d, grey, all cell nuclei), NKX2-5 (e, grey nuclei in cardiac crescent) and SRF (grey nuclei in cardiac
crescent). (g) Maximum intensity projection of image volume of embryonic heart, with F-actin in magenta and NKX2-5 in grey. The green dots and line are regions
of interest marking the head folds and cardiac crescent, respectively. (h) Localization and correlation of SRF and NKX2-5 transcription factors, with boxplots comparing
proteins in different embryo regions. The blue and red dots were clustered by the Clara algorithm based on SRF and NKX2-5 levels and their distance from the heart and
head folds. p-values are less than 0.001 (t-test). Scale bar in panel (a) represents 50 µm and in panel (f ) 100 µm. N/C, ratio of nuclear to cytoplasmic fluorescence.

To test the ability of our image analysis approach for this analyses enabled by our program will enable high-throughput
problem, we examined the localization of NKX2-5 and SRF, quantification and statistical analysis of spatial protein organiz-
two transcription factors important for cardiomyocyte differen- ation. As three-dimensional microscopy is versatile and
tiation [25] in 8.0 days post coitum embryos (figure 6e,f ). In order widespread, we expect that our publicly available open-source
to represent protein levels of these two transcription factors, we software will be useful not only for developmental biology
used a ratio of nuclear to cytoplasmic fluorescence (N/C), as the but also more broadly, in the context of cell biology.
latter is expected to represent background signal for internal
normalization. To reduce the dynamic range of the measure-
ments and normalize the distribution of the values, we used a
natural logarithm of the ratio. In addition to these normalized 3. Methods
fluorescence measurements, we also measured cell distance to
(a) Mouse strains, husbandry and embryo collection
two regions, the head folds and cardiac crescent, manually All animal experiments were carried out according to UK Home
designated using our software (green dots and line in figure 6g). Office project licence PPL 30/3155 and 30/2887 compliant with
The four variables formed a dataset, which was clustered the UK Animals (Scientific Procedures) Act 1986 and approved by
using partitioning around medoids (PAM, R ‘Clara’ package) the local Biological Services Ethical Review Process. All mice were
[26]. The method was chosen over other clustering approaches maintained in a 12 h L : 12 h D cycle. Noon of the day of finding a
[27] owing to computational efficiency in dealing with large vaginal plug was designated 0.5 dpc (days post coitum). To obtain
datasets. The analysis verified as expected that both proteins embryos, C57BL/6 males were crossed with CD1 females (Charles
were preferentially expressed in the heart region and there River). Embryos of the appropriate stage were dissected in M2
medium (Sigma-Aldrich) at room temperature. Pre-implantation
was a region-specific correlation between nuclear levels of
embryos were collected by oviduct and uterus flushing.
NKX2-5 and SRF (figure 6h).
In summary, we have presented a sampling-based three-
dimensional image analysis approach, designed for relative
quantification of protein levels in microscopy images of com-
(b) Embryonic stem cell culture
Embryonic stem cells were cultured in T-25 flasks coated with 0.1%
plex tissues. Our approach is the first to simultaneously w/v gelatin and used at passage no. 50–60. Cell culture medium
quantify nuclear, membrane and cytoplasmic fluorescence, consisted of 90% DMEM medium (Gibco), 10% fetal calf serum
while also marking regions of interest to extract spatial infor- (FCIII, Hyclone), 1 mM sodium pyruvate, 1 mM non-essential
mation relating to the quantified signals. Automation of amino acid mix, 0.1 mM beta-mercaptoethanol, 2 mM L-glutamine
and 1000 U ml−1 leukaemia inducible factor (LIF). Cells were pas- (f ) Software and statistics 8
saged every 2 days and seeded at approximately 400 000 cells per The software and graphical user interface were written in
T-25 flask. Embryoid bodies were formed by aggregating cells in

royalsocietypublishing.org/journal/rstb
MATLAB (v. 2015b, Mathworks) and compiled using the in-built
hanging drops and culturing for 2 days before fixation. compiler for Mac OSX and Windows operating systems (https://
process.innovation.ox.ac.uk/software/p/13299a/silentmark-aca-
(c) Antibodies demic/1). The algorithm has also been implemented in Python as
Vectashield with DAPI (H-1200) was purchased from Vector a Web application (hosted by Droplet Genomics at http://droplet-
Laboratories, phalloidin atto647 was purchased from Sigma genomics.com). Source code for SilentMark is available upon
(65906-10MMOL). Mouse anti-YAP monoclonal (sc-101199) and request. Three-dimensional image visualization was done using
goat anti-Nkx2-5(N-19) (sc-8697) polyclonal antibodies were VOLOCITY software (Perkin Elmer), and IMARIS software (v 6.1) was
purchased from Santa Cruz Biotechnology Rabbit anti-CDX2 used for manual cell outlining and analysis. R statistics package
(no. 3977S) monoclonal and rabbit anti-pERM(T567) (no. 3149P) alongside ‘cluster’ library was used for automatic normalized
monoclonal antibodies were purchased from Cell Signaling. Rat cluster detection (PAM algorithm) and cell population analysis.
anti-uvomorulin/E-cadherin (U3254-100UL) monoclonal anti- The first step of the algorithm is nuclei or membrane detection.
body was purchased from Sigma. Rabbit anti-SRF (PA5-27307) This initial two-dimensional segmentation step can be done using

Phil. Trans. R. Soc. B 375: 20190560

polyclonal antibody was purchased from Thermo Fisher. We a variety of approaches (including convolutional neural nets), but
used the following secondary antibodies from Life Technologies: we found Hough circle detection to be the most efficient and
goat anti-rabbit AlexaFluor 488 (A11008), donkey anti-goat Alexa- robust. Detected nuclear or membrane feature coordinates on
Fluor 488 (A11055), donkey anti-mouse AlexaFluor 555 (A31570), two-dimensional stacks are then convolved into three-dimensional
donkey anti-rabbit AlexaFluor 647 (A31573). space to build an interconnected graph. The second step then relies
on efficient spatial searching, which segments stack voxels based
on this graph. Delaunay triangulation was used for this search
(d) Immunofluorescence on pre-implantation embryos algorithm. Once the voxels are assigned to nearest nuclear or mem-
brane centres, normalized intensity of nuclear or membrane stain is
and embryoid bodies used to classify them as cytoplasmic. Then the algorithm measures
Embryos and EBs were fixed for 20 min at room temperature in all fluorescence markers in the voxels designated by type (nuclear,
4% paraformaldehyde (PFA). Samples were then washed three cytoplasmic or membrane).
times in PBT-0.1% ( phosphate-buffered saline with 0.1%
Tween) for 10 min, permeabilized in PBT-0.25% for 40 min and
washed again in PBT-0.1%. The tissues were transferred to a
blocking solution for 1 h at 4°C. Primary antibodies were then Ethics. All animal experiments were carried out according to UK
Home Office project licence PPL 30/3155 and 30/2887 compliant
added to the solution and incubated overnight at 4°C. The
with the UK Animals (Scientific Procedures) Act 1986 and approved
embryos were washed in PBT-0.1% and incubated for 1 h at 4°
by the local Biological Services Ethical Review Process.
C in PBT-0.1% with the secondary antibodies, then subsequently
Data accessibility. The software and graphical user interface were written
washed two times in PBT-0.1% for 15 min and mounted in
in MATLAB (v. 2015b, Mathworks) and compiled using the in-built
Vectashield with DAPI at least 6 h prior to imaging. compiler for Mac OSX and Windows operating systems (available
at https://process.innovation.ox.ac.uk/software/p/13299a/silent-
mark-academic/1).
(e) Immunofluorescence on post-implantation embryos Authors’ contributions. K.L. developed the analysis method in discussion
Dissected embryos were fixed for 1 h at room temperature in
with S.S. K.L., C.R. and A.K. performed early embryo and stem cell
4% PFA. The embryos were then washed three times in PBT-0.1% experiments. R.C.V.T. performed the embryonic heart experiments.
for 15 min, permeabilized in PBT-0.25% for 40 min and washed K.L. and A.M. A.M. designed statistical analyses. S.S., K.L. and
again three times in PBT-0.1%. The embryos were transferred to C.R. wrote the manuscript.
blocking solution overnight at 4°C. Primary antibodies were then Competing interests. We declare we have no competing interests.
added to the solution and incubated at 4°C. The embryos were Funding. K.L. was supported by a Biotechnology and Biological
washed three times in PBT-0.1% and incubated overnight at 4°C Sciences Research Council (BBSRC) Doctoral Training Programme
in PBT-0.1% with the secondary antibodies, then subsequently grant to the University of Oxford (BB/J014427/1). This work was
washed three times in PBT-0.1% for 15 min and mounted in supported through a BBSRC equipment grant (BB/F011512/1) and
Vectashield with DAPI at least 24 h prior to imaging. a Wellcome Senior Investigator Award to S.S. (ref. 103788/Z/14/Z).

References
1. Luengo-Oroz MA, Ledesma-Carbayo MJ, Peyriéras N, 4. Chickarmane V, Roeder AHK, Tarr PT, Cunha A, Tobin Source Code Biol. Med. 8, 16. (doi:10.1186/1751-
Santos A. 2011 Image analysis for understanding C, Meyerowitz EM. 2010 Computational 0473-8-16)
embryo development: a bridge from microscopy to morphodynamics: a modeling framework to 7. Lou X, Kang M, Xenopoulos P, Muñoz-Descalzo S,
biological insights. Curr. Opin. Genet. Dev. 21, understand plant growth. arXiv 61, 65–87. Hadjantonakis A-K. 2014 A rapid and efficient 2D/3D
630–637. (doi:10.1016/j.gde.2011.08.001) 5. Fernandez R, Das P, Mirabet V, Moscardi E, Traas J, nuclear segmentation method for analysis of early
2. Schindelin J et al. 2012 Fiji: an open-source Verdeil J-L, Malandain G, Godin C. 2010 Imaging mouse embryo and stem cell image data. Stem Cell
platform for biological-image analysis. plant growth in 4D: robust tissue reconstruction and Rep. 2, 382–397. (doi:10.1016/j.stemcr.2014.01.010)
Nat. Methods 9, 676–682. (doi:10.1038/ lineaging at cell resolution. Nat. Methods 7, 8. Peng H, Ruan Z, Long F, Simpson JH, Myers EW.
nmeth.2019) 547–553. (doi:10.1038/nmeth.1472) 2010 V3D enables real-time 3D visualization and
3. Barbier de Reuille P et al. 2015 MorphoGraphX: a 6. Hodneland E, Kögel T, Frei DM, Gerdes H-H, quantitative analysis of large-scale biological image
platform for quantifying morphogenesis in 4D. eLife Lundervold A. 2013 CellSegm - a MATLAB data sets. Nat. Biotechnol. 28, 348–353. (doi:10.
4, e05864. (doi:10.7554/eLife.05864) toolbox for high-throughput 3D cell segmentation. 1038/nbt.1612)
9. Stegmaier J, Amat F, Lemon WC, McDole K, Wan Y, 15. Bähnisch C, Stelldinger P, Köthe U. 2009 Fast and 21. Deng Q, Ramsköld D, Reinius B, Sandberg R. 2014 9
Teodoro G, Mikut R, Keller PJ. 2016 Real-time three- accurate 3D edge detection for surface Single-cell RNA-seq reveals dynamic, random

royalsocietypublishing.org/journal/rstb
dimensional cell segmentation in large-scale reconstruction. In Pattern recognition DAGM 2009. monoallelic gene expression in mammalian cells.
microscopy data of developing embryos. Dev. Cell Lecture Notes in Computer Science, vol. 5748 (eds Science 343, 193–196. (doi:10.1126/science.1245316)
36, 225–240. (doi:10.1016/j.devcel.2015.12.028) J Denzler, G Notni, H Süße), pp. 111–120. Berlin, 22. Rayon T et al. 2014 Notch and Hippo converge on
10. Aigouy B, Farhadifar R, Staple DB, Sagner A, Röper Germany: Springer. (doi:10.1007/978-3-642-03798- Cdx2 to specify the trophectoderm lineage in the
J-C, Jülicher F, Eaton S. 2010 Cell flow reorients the 6_12) mouse blastocyst. Dev. Cell 30, 410–422. (doi:10.
axis of planar polarity in the wing epithelium of 16. Sasaki H. 2010 Mechanisms of trophectoderm fate 1016/j.devcel.2014.06.019)
Drosophila. Cell 142, 773–786. (doi:10.1016/j.cell. specification in preimplantation mouse 23. Ralston A et al. 2010 Gata3 regulates trophoblast
2010.07.042) development. Dev. Growth Differ. 52, 263–273. development downstream of Tead4 and in parallel
11. Trichas G, Joyce B, Crompton LA, Wilkins V, (doi:10.1111/j.1440-169X.2009.01158.x) to Cdx2. Development 137, 395–403. (doi:10.1242/
Clements M, Tada M, Rodriguez TA, Srinivas S. 17. Wennekamp S, Mesecke S, Nédélec F, Hiiragi T. 2013 dev.038828)
2011 Nodal dependent differential localisation A self-organization framework for symmetry 24. Doetschman TC, Eistetter H, Katz M, Schmidt W,
of Dishevelled-2 demarcates regions of differing breaking in the mammalian embryo. Nat. Rev. Mol. Kemler R. 1985 The in vitro development of

Phil. Trans. R. Soc. B 375: 20190560

cell behaviour in the visceral endoderm. Cell Biol. 14, 452–459. (doi:10.1038/nrm3602) blastocyst-derived embryonic stem cell lines:
PLoS Biol. 9, e1001019. (doi:10.1371/journal. 18. Nishioka N et al. 2009 The Hippo signaling pathway formation of visceral yolk sac, blood islands and
pbio.1001019) components Lats and Yap pattern Tead4 activity to myocardium. J. Embryol. Exp. Morphol. 87, 27–45.
12. Watanabe T, Biggins JS, Tannan NB, Srinivas S. 2014 distinguish mouse trophectoderm from inner cell 25. Schlesinger J et al. 2011 The cardiac transcription
Limited predictive value of blastomere angle of mass. Dev. Cell 16, 398–410. (doi:10.1016/j.devcel. network modulated by Gata4, Mef2a, Nkx2.5, Srf,
division in trophectoderm and inner cell mass 2009.02.003) histone modifications, and microRNAs. PLoS Genet.
specification. Development 141, 2279–2288. 19. Niwa H, Toyooka Y, Shimosato D, Strumpf D, 7, e1001313. (doi:10.1371/journal.pgen.1001313)
(doi:10.1242/dev.103267) Takahashi K, Yagi R, Rossant J. 2005 Interaction 26. van der Laan M, Pollard K, Bryan J. 2002 A new
13. Halder G, Dupont S, Piccolo S. 2012 Transduction of between Oct3/4 and Cdx2 determines partitioning around medoids algorithm. J. Stat.
mechanical and cytoskeletal cues by YAP and TAZ. trophectoderm differentiation. Cell 123, 917–929. Comput. Simul. 73, 575–584. (doi:10.1080/
Nat. Rev. Mol. Cell Biol. 13, 591–600. (doi:10.1038/ (doi:10.1016/j.cell.2005.08.040) 0094965031000136012)
nrm3416) 20. Sasaki H. 2017 Roles and regulations of Hippo 27. Wiwie C, Baumbach J, Röttger R. 2015 Comparing
14. Illingworth J, Kittler J. 1987 The adaptive Hough signaling during preimplantation mouse the performance of biomedical clustering methods.
transform. IEEE Trans. Pattern Anal. Mach. Intell. PAMI- development. Dev. Growth Differ. 59, 12–20. Nat. Methods 12, 1033–1038. (doi:10.1038/
9, 690–698. (doi:10.1109/TPAMI.1987.4767964) (doi:10.1111/dgd.12335) nmeth.3583)