Reconstructing protein networks of epithelial differentiation from histological sections
-
Upload
independent -
Category
Documents
-
view
2 -
download
0
Transcript of Reconstructing protein networks of epithelial differentiation from histological sections
Vol. 23 no. 23 2007, pages 3200–3208BIOINFORMATICS ORIGINAL PAPER doi:10.1093/bioinformatics/btm504
Systems biology
Reconstructing protein networks of epithelial differentiation
from histological sectionsNiels Grabe1,2,*, Thora Pommerencke1,2, Thorsten Steinberg1,3, Hartmut Dickhaus1,2 andPascal Tomakidi1,31Hamamatsu Tissue Imaging and Analysis (TIGA) Center, BIOQUANT, University Heidelberg, BQ0010,Im Neuenheimer Feld 267, 2Institute for Medical Biometry and Informatics, University Hospital Heidelberg,Im Neuenheimer Feld 305 and 3Department of Orthodontics and Dentofacial Orthopedics, Dental School Universityof Heidelberg, Im Neuenheimer Feld 400, 69120 Heidelberg, Germany
Received on July 9, 2007; revised on September 14, 2007; accepted on October 3, 2007
Associate Editor: Limsoon Wong
ABSTRACT
Motivation: For systems biology of complex stratified epithelia like
human epidermis, it will be of particular importance to reconstruct
the spatiotemporal gene and protein networks regulating keratino-
cyte differentiation and homeostasis.
Results: Inside the epidermis, the differentiation state of individual
keratinocytes is correlated with their respective distance from the
connective tissue. We here present a novel method to profile this
correlation for multiple epithelial protein biomarkers in the form of
quantitative spatial profiles. Profiles were computed by applying
image processing algorithms to histological sections stained with
tri-color indirect immunofluorescence. From the quantitative spatial
profiles, reflecting the spatiotemporal changes of protein expression
during cellular differentiation, graphs of protein networks were
reconstructed.
Conclusion: Spatiotemporal networks can be used as a means for
comparing and interpreting quantitative spatial protein expression
profiles obtained from different tissue samples. In combination with
automated microscopes, our new method supports the large-scale
systems biological analysis of stratified epithelial tissues.
Contact: [email protected]
1 INTRODUCTION
Systems biology aims at the analysis and in silico modeling and
simulation of large and complex biological systems. Currently,
most publications in this field are focused on the analysis and
modeling of intracellular networks (Hahn and Weinberg, 2002;
Ideker et al., 2001; Kitano, 2002), or are based on static tissue
structures (Noble, 2004). For systems biology of complex
stratified epithelia, it will be of particular importance to
reconstruct the spatiotemporal gene and protein networks
regulating cellular differentiation, a pivotal cornerstone of
epithelial homeostasis (Grabe and Neuber, 2005, 2007). Hence,
the expression changes of multiple epithelial biomarkers have
to be profiled in parallel during cell migration through the
tissue and furthermore, networks of directly or indirectly
interacting proteins (protein networks) have to be deduced
from these profiles. The prominent omics-technologies like
microarrays (Cole et al., 2001; Koria et al., 2003) and 2D/1MS
proteomics (Huang et al., 2003) have been applied for example
on skin, but employ homogenized, structurally destroyed tissue.
Therefore, these methods are only hardly applicable for
this purpose. Tissue profiling, using mass spectrometry
(Caldwell and Caprioli, 2005) or laser capture microdissection
(Emmert-Buck et al., 1996) represents considerable solutions.
However, these methods are cost-intensive, and can currently
only provide a limited spatial resolution in comparison to
immunohistology. Moreover, they may fail in developing a
profile for a certain biomarker of interest. Instead, antibody-
based proteomics aims at the systematic generation and use of
protein-specific antibodies to functionally explore the proteome
(Uhlen and Ponten, 2005). Based on this spirit, we set out to
develop a method for the quantitative spatial analysis of
immunofluorescence-stained stratified epithelial tissue sections.
With the advent of fully automatic microscopic scanning robots,
it will be possible to scan fluorescence-stained tissue sections in
high throughput. A method which, based on such microscopic
high-throughput data of tissue slides can reconstruct spatio-
temporal networks of differentiation from histological sections
could therefore bring important insights into the dynamics of
differentiation and homeostasis of epithelial tissues.At the example of the human epidermis, we here show how
such a method can be established in principle. We quantita-
tively study the differentiation of five key proteins of epidermal
homeostasis: integrin �6, desmoplakin, involucrin, filaggrin and
keratin K1/10. The functional of individual proteins in
epidermal homeostasis is explained in the following.The epidermis as a squamous, keratinized epithelium consists
of basal, spinous, granular and cornified cell layers (Candi
et al., 2005). In this strata model, each layer is defined by
position, morphology, polarity and stage of differentiation of
keratinocytes. Being basal lamina-adhesion sites, hemidesmo-
somes are linked to the intermediate filament system, and
contain a specific set of proteins including �6�4 integrin.
Suprabasal spinous cells are more oval than the flattened*To whom correspondence should be addressed.
3200 � The Author 2007. Published by Oxford University Press. All rights reserved. For Permissions, please email: [email protected]
by guest on January 19, 2016http://bioinform
atics.oxfordjournals.org/D
ownloaded from
spinous cells located adjacent to the granular cell zone, which
are joined by the name-giving desmosomes (Kottke et al.,
2006). Tethering of the desmosomes to the cells’ cytoplasmic
intermediate filaments is provided by desmoplakin (DPK).
Moreover, the spinous differentiation stage of keratinocytes
following the basal stage is characterized by an induction of the
keratin K1/K10 couple (Lechler and Fuchs, 2005). Involucrin
(Rice and Green, 1979) is cross-linked early in cornified
envelope formation, and forms a scaffold for incorporation of
other precursors. Its expression is initiated in the early spinous
layer, and maintained in the granular layer (Eckert et al., 2002).
During terminal differentiation of keratinocytes, profilaggrin is
cleaved to yield filaggrin, which temporally coincides with the
transition of the nucleated granular cell to the anucleated
cornified cell. The final layer of the epidermis is the stratum
corneum consisting of terminally differentiated keratinocytes
embedded in an extracellular lipid matrix.
With respect to squamous epithelia, the existence of a spatial
and functional heterogeneity of biomarker expression between
the horizontally arranged layers is of particular importance.
However, up till now, biomarker expression during keratino-
cyte differentiation has been rather described in a qualitative
fashion. For example with the human protein atlas (Uhlen and
Ponten, 2005; Uhlen et al., 2005), such an endeavor is pursued
in a large scale. Also binary vectors of protein expression have
been successfully measured (Schubert, 2006; Schubert et al.,
2006). Unfortunately, the traditional qualitative mode of
description is lacking a quantification of the spatial expression
of a biomarker and a correlation of this quantification with the
time-dependent dynamics of progressive keratinocyte differ-
entiation as well. For applying the visionary concepts of
systems biology onto this spatiotemporal expression of multiple
biomarkers, which as a whole constitute epithelial organization,
an extension of the qualitative description of keratinocyte
biomarkers into a quantitative mode is indispensable.
Observing the change of protein expression in relation to the
distance to the connective tissue yields profiles (or gradients) of
biomarker expressions. The present study describes a new
method to generate such profiles in a quantitative way
(Quantitative Spatial Profiles or QSPs) for certain epithelial
biomarkers. We conceive such profiles as time-series data of
cellular differentiation from which we then can deduce the
topology of a protein network. The resulting outline of a
network should not be understood as directly describing
functional regulation like it occurs when a transcription factor
interacts with a promoter. Instead, it is a general description of
the spatiotemporal coregulation of the studied protein biomar-
kers during differentiation. Such spatiotemporal networks,
describing the changes of protein expressions during progressive
keratinocyte maturation, render a framework for the future in
silico modeling and simulation of epithelial differentiation.
2 METHODS
2.1 Generation of triple-stained tissue sections by
indirect immunofluorescence (IIF)
Three human epidermal tissue samples derived from different head and
neck regions were obtained from healthy patients with their informed
consent according to the Helsinki Declaration, and the protocol was
approved by the institutional ethics committee. For the generation of
the triple stains, 10�m serial frozen sections were mounted on glass
slides (Histobond, Marienfeld, Germany), preceding freezing of the
epidermal tissue in liquid nitrogen vapor. About five sections were used
for each of the five protein biomarkers in each of the three tissues,
thereby leading to a total of 75 sections. To facilitate computational
image analysis, a triple staining, based on double IIF was applied to
each tissue section. For each biomarker studied in each tissue sample,
an exemplary triple-stained tissue section is given in Figure 1.
Triple staining comprises of staining the according marker together
with extracellular matrix collagen type-I, as well as a counterstain of cell
nuclei (DAPI). Collagen type-I marks the dermal connective tissue for
the later applied computational image analysis. IIF was performed
according to previous protocols (Tomakidi et al., 2003) with foregoing
optimization of the working dilutions (wd) of primary and secondary
antibodies applied to the skin specimens (antibody working dilution
and manufacturer are described below). Briefly, after fixation with
methanol and acetone, slides were incubated with respective primary
antibodies overnight at room temperature in a humid chamber. Then,
slides were rinsed in PBS for three times and incubated with
fluorochrome-conjugated secondary antibodies for 1 h, followed by
three washes in PBS.
For cell nuclei labeling, specimens were incubated with 1�g/ml DAPI
for 20min and supplied by three final PBS washes before embedding in
mounting medium (Linaris, Wertheim, Germany). For the biomarker
stainings, primary monoclonal antibodies were the following: integrin
�6 (wd 1:12.5, Chemicon, Hofheim/Taunus, Germany), desmoplakin
(wd 1:10, Progen, Heidelberg, Germany), keratin 1/10 (wd 1:40,
Progen), involucrin (wd 1:50, abcam, Hiddenhausen, Germany) and
filaggrin (wd 1:20, TeBu-Bio, Frankfurt, Germany). For double IIF,
each of these markers was incubated together with a rabbit polyclonal
collagen type-I antibody (1:50, Biodesign/Dunn, Asbach, Germany).
Secondary antibodies were also applied to the sections as a cocktail,
consisting of goat anti-mouse Alexa Fluor� 488 (1:50) and goat anti-
rabbit Alexa Fluor� 594 (1:100, both MoBiTec, Gottingen, Germany).
All antibodies were adjusted to their final working dilution in PBS
containing 0.5% Tween 20, and as further additives 0.5% BSA and
0.02% sodium acid.
2.2 Imaging of triple-stained tissue sections
Triple-stained multifluorescence slides were scanned using a Nikon
Eclipse 90i upright automated microscope (Nikon GmbH, Dusseldorf)
equipped with a Nikon DS-1QM high-resolution CCD camera.
By applying the ‘large image’ function of the NIC-elements software
controlling the microscope, several large stitched images of individual
microscopic images with 10� objective were acquired from each tissue
section. For each image, distinct monochromatic pictures of the red,
blue, green and phase contrast channel were taken. For further image
analysis, own software was developed using MATLAB 7.1 including the
image processing toolbox. In all large images, the epithelium was
segmented computationally using nuclear DAPI in conjunction with the
collagen staining and the phase contrast. Each large image was
then split computationally into a set of subimages where each subimage
had approximately equal horizontal width. Figure 2a shows a
composite example of the three monochromatic large images taken
from an epidermal tissue section. In Figure 2b, the according phase
contrast image shows the upper tissue boundary. By the special
fluorescent staining used, the individual components of the skin
(connective tissue, epithelium and stratum corneum) are labeled in
different colors and could thus be captured in different monochromatic
channels.
Reconstructing protein networks
3201
by guest on January 19, 2016http://bioinform
atics.oxfordjournals.org/D
ownloaded from
2.3 Computation of quantitative spatial profiles (QSPs)
from each subimage
The following algorithm was applied to each subimage to determine the
differentiation profiles. The individual concepts of the algorithm are
depicted in Figure 3a, while the calculation of the distance-based
quantifications of the degree of differentiation of a cell inside the
epithelium is explained in Figure 3b. A graphical overview of the QSP
computation is given in Figure 5a.
Detecting the epithelium E: the upper tissue border (B), characterized
by the stratum corneum, was detected using the phase-contrast image
PC. The estimated basal lamina (L) of the tissue sections was found by
the dense band of basal keratinocytes visible in the blue DAPI channel.
As an estimate of the epithelial tissue, an initial mask (E) was
determined as the region between L and B and having a low collagen
staining (indicating connective tissue CT) in the red channel near L.
To remove the stratum corneum from E, regions not carrying nuclear
DAPI stain were further excluded. Due to the 3D epidermal–dermal
interdigitations, vertical sections of epidermal specimen occasionally
yield collagen islands (Fig. 2a). These cut papillae are characterized by a
strong collagen stain inside E. Such regions in E were excluded by
removing regions of strong red stain inside E. Finally, the tissue area E
so far detected is slightly extended by X towards the connective tissue
CT. Biomarkers like integrin �6 directly interact with the connective
tissue and are thus located in its close proximity. Therefore, estimating
Fig. 1. Tri-color stains of the studied differentiation markers. Clippings of large of images of indirect immunofluorescences (IIF) of human epidermis
(Staining: red¼ collagen, blue¼ cell nuclei (DAPI) and green¼ respective biomarker protein). The red collagen stain is indicating the connective
tissue, while the DAPI stain above the collagen stain indicates the epithelium.
Fig. 2. Exemplary fluorescent three color image of human epidermis
with the according phase-contrast image. (a) SC¼ stratum corneum,
CT¼ connective tissue with collagen stained with in fluorescent red and
the epithelium E. The green biomarker BM shows filaggrin expression.
BL¼ strongly stained basal layer of E used to estimate the basal lamina
towards CT. Due to 3D epidermal–dermal interdigitations, collagen
islands (CI) may appear in E. (b) Phase contrast image used for
determining the upper tissue outline at the stratum corneum SC.
N.Grabe et al.
3202
by guest on January 19, 2016http://bioinform
atics.oxfordjournals.org/D
ownloaded from
the position of the basal lamina only by the dense basal band of cell
nuclei as described above would not be sufficient. Instead the dermal–
epidermal interaction zone has to be included in E also and thus E was
slightly extended by X.
QSP computation: from the final mask of the epithelial region E,
a normalized distance image (N) was created with a distance of d¼ 0%
at the basal lamina including the above-described extended region, and
d¼ 100% at the border of E to the stratum corneum. A detailed
explanation of the computation is given in Figure 3b. From N and E the
quantitative spatial profiles (QSPs) were computed: for all sliding
distance intervals corresponding regions in E were chosen and the
respective average marker intensity was determined from the green
channel. The signal intensities were corrected for background signal
averaged from the mask of the connective tissue CT. The collected
marker intensities were normalized to a maximum of 100% fluorescence
intensity to account for variations in staining efficiency.
2.4 Computation of protein networks
From the QSPs, an individual spatiotemporal protein network was
computed for each of the three epidermal tissue samples as follows
(Figure 4). For illustration purposes, a graphical representation of the
process flow in protein network computation is given in Figure 5b.
Correlating QSPs: simplified, such a network shows an observed
coregulation of a set of proteins during epidermal differentiation.
Figure 4 illustrates how such networks were computed from QSPs at the
example of two hypothetical tissue samples (Fig. 4a,b). All proteins were
compared pairwise by computing the linear correlation coefficient of
their QSPs, i.e. the QSPs of two proteins are considered to be 100%
correlated during a certain period of time, if their QSPs can be mapped
onto each other by a linear scaling factor. The first peak of two QSPs to
be compared (in the example denoted M1, M2) determines the time
period of mutual increase, while the second peak determines the period
of mutual decrease (Fig. 4a,b). Through normalization, we generally
observed exactly one peak in the QSP of all our tissue samples. This one-
peak-property is a prerequisite for our method as it ensures that all
graphs must intersect at one point as long as the profiles do not have
mutually exclusive flanks. For each phase of mutual increase or decrease,
we required a minimum length of 10% of the total QSP. Also, only
protein abundances above 20% were considered, eliminating back-
ground signals. Finally, the linear correlation coefficient was computed
and stored for each such mutual phase of two protein biomarkers.
Determination of spatiotemporal ‘leadership’ patterns between two
QSPs: for each pair of proteins, we evaluated if one protein
spatiotemporally leads during the whole phase of mutual increase
(Fig. 4c: lead pattern, unidirectional green arrow) or if both proteins
alternate in their leadership (Fig. 4d: alternating pattern, bidirectional
green arrow). This was repeated for the mutual decrease phase yielding
the red arrows. For accepting a ‘lead’, the algorithm required a
leadership of a minimum length of 10% of the total QSP length.
Consensus network: from all three individual networks a consensus
network was derived. At least two arrows with sufficient correlation
(450%) in individual networks substantiated an arrow in the consensus
network. For this calculation, bidirectional arrows were treated as two
single arrows (Fig. 4e).
3 RESULTS
We analyzed the spatiotemporal expression patterns of fiveprotein biomarkers of epidermal differentiation: integrin �6,keratin 1/10, involucrin and filaggrin. To give the reader animpression of the staining patterns obtained, clippings of some
of the large images of the triple-stained tissue sections are givenin Figure 1. As mentioned before, quantitative spatial profiles
(QSPs) were computed not from large images but, from
individual subimages. As the QSPs were derived from a totalof 50 subimages in the average, the shown clippings can only
give a punctual impression of the average profiles we obtainedcomputationally. The average QSPs of all those profiles for
each biomarker for the three epidermal samples are given inFigure 6a. The profiles reflect the normalized quantity of
expression of a biomarker in relation to the quantified degree of
differentiation. Thus, we quantified cellular differentiation bymeasuring the distance of any biomarker-derived pixel in any
location of the epithelial compartment to the epithelial–dermalinterface. Averaging over the profiles of all subimages resulted
in the depicted spatial expression profiles of each biomarkerwith the given SDs.
Figure 6a shows the panel of all QSPs for the three tissuesamples. The x axis describes the relative distance of an
epithelial pixel to the collagen type-I-stained connective tissue.A distance of 0% refers to the marker intensity measured
closest to the connective tissue interface. A distance of 100%
refers to the marker intensity at the stratum corneum. They axis shows the normalized fluorescence intensity for the
respective marker, relative to the background signal intensitiesmeasured in the connective tissue. The SD of the mean value is
given for each data-point. The maximum value of eachbiomarker profile is normalized to 100%. Since we correct
only for background signals measured in the connective tissue,
we still encounter an epithelium-specific fluorescence back-ground noise. From the unspecific background signal of
integrin �6 in the suprabasal layers, this can be estimated to
Fig. 3. Central concepts for generating the quantitative profiles.
(a) Schematic tissue section: the microscopic slide image I, geometric
borders of the analyzed tissue sections B, dense band of the basal
compartment D, connective tissue CT, the basal lamina L, the extended
region X and the resulting epithelium E. During differentiation, cells
migrate according to the arrows from d¼ 0% up to d¼ 100% below the
stratum corneum SC. (b) Generation of the normalized distance
d¼ dAbsD/(dAbsUþ dAbsD), where dAbsU is the absolute distance upwards
of a cell to the stratum corneum, and dAbsD is the absolute distance of a
cell downwards to the epithelial–connective tissue interface. For any cell
in the tissue, dAbsDþ dAbsU is equal to 100% of the spatial distance the
cell migrates from the basal lamina to the stratum corneum.
Reconstructing protein networks
3203
by guest on January 19, 2016http://bioinform
atics.oxfordjournals.org/D
ownloaded from
lie below 20%. For biomarkers having signal intensities above
this threshold, the proteins can be considered as expressed.
3.1 Results of individual epithelial biomarkers
Integrin �6 (Figs 1 and 6a): in relation to the analyzedbiomarkers, integrin �6 was expressed first in the epithelial
compartment. It was visible as a narrow band at theepithelium–connective tissue interface. To measure the intensityof this band, the image processing algorithm extended the
computationally determined epithelium to include the zone ofepithelial–connective tissue interaction. Thus, we were able
to measure integrin �6 as a peak at the beginning ofdifferentiation.Desmoplakin (Figs 1 and 6a): in all analyzed tissues,
desmoplakin (DPK) was already baso-laterally expressed asreported in literature (Wan et al., 2004). Desmoplakin isconsidered to become embedded in the highly cross-linked
cornified envelope structures during the process of keratinocyteterminal differentiation like other cytoskeletal and desmosomal
components (Haftek et al., 1991). In the QSPs, the samplesderived from epidermal tissues exhibited an initially lowamount of DPK, which continually increased, forming
a steep gradient. The observed profile was bell-shaped withits peak at �40% differentiation. With respect to the tested
biomarkers, DPK was the next to be expressed followingintegrin �6.Keratins K1/10 (Figs 1 and 6a): K1/10 forms a set of
intermediate filaments which are among the first proteins in
cornification of keratinized epithelia, therefore indicating their
early differentiation (Candi et al., 2005). It is suprabasally
expressed (Chu and Weiss, 2002) with a strong expression in the
stratum corneum (Stoler et al., 1988). K1/10 showed a medium
expression directly after the onset of DPK which in all samples
then increased to reach its highest expression level in the
stratum corneum or just below.
Involucrin (Figs 1 and 6a): together with the switch from
K5/K14 to K1/K10, the expression of involucrin as a precursor
of the cornified envelope is a hallmark of terminal keratinocyte
differentiation (Watt et al., 1987). In epidermis, literature
describes the protein to be expressed from three to four layers
above the basal layer at a zone of enlarged cell size (Banks-
Schlegel and Green, 1981; Watt et al., 1987). We detected the
beginning of the presence of involucrin at approximately half of
the width of the epidermis. Consistent with literature (Walts
et al., 1985), involucrin then increasingly accumulated in the
epidermis until it reached its highest expression level below the
stratum corneum.
Filaggrin (Figs 1 and 6a): Filaggrin has the task of finally
bundling and thus collapsing the keratin intermediate filaments
(Candi et al., 2005). In the epidermis, the filaggrin protein
appeared later than involucrin, therefore following involucrin
in its expression profile at a small distance (Figs 1 and 6a).
Filaggrin protein expression is the terminal step of keratinocyte
differentiation that we were able to observe with our markers,
leading to the final flat cell shape characterizing the stratum
corneum.
Differentiation
Mutualincrease
Mutualdecrease
Relative concentration
M1
M2
M1 leads M2 trails
M1 leads M2 trails
Intersection
Peaks
QSPs of hypothetical tissue 1(a)
t1
Mutualincrease
Relative concentration
Mutualdecrease
Alternatinglead
M2
M1
M1 leadsM2 trails
Differentiation
Intersection
Peaks
QSPs of hypothetical tissue 2
t2
t3
M1 M2
Spatiotemporal network of hypothetical tissue 1
M1 M2
Spatiotemporal network of hypothetical tissue 2
Confidence (M1�M2) = t1 = 100%
Confidence (M1�M2) = t2 =70%Confidence (M2�M1) = t3 =30%
M1 M2
Consensusnetwork of both tissues
(b)(c)
(d)
(e)
Fig. 4. Regulatory networks generalize spatiotemporal multi-biomarker profiles. The QSPs of two proteins of two hypothetical tissue samples (a, b)
are shown. (a) In the hypothetical tissue 1, the protein biomarker M1 precedes the protein biomarker M2 during the phase of mutual increase and
mutual decease. The mutual increase phase ends at the first peak of the protein biomarkers (here the peak of M1). The mutual decrease phase starts at
the second peak (here the peak of M2). The two profiles necessarily intersect at one point. (c) The resulting network illustrates that M1 leads during
both phases of increase and decrease. (b) In the hypothetical tissue sample 2, the protein biomarkers M1 and M2 alternate in their leadership before
the first peak is reached. Therefore, two intersections points occur. From the second peak on, the mutual decrease phase starts, which is lead by M1.
(d) In the resulting network, the alternating lead during the mutual increase phase is depicted by a bidirectional green arrow. (e) From the individual
spatiotemporal network, the consensus network is derived. Only the green arrow from M1 to M2 and not the reverse from M2 to M1 has been
detected as well in (c) as in (d) and is therefore the present. The confidence, which is here only illustrated for the increase is defined as the relative
amount of time one protein leads during a phase of mutual increase or decrease.
N.Grabe et al.
3204
by guest on January 19, 2016http://bioinform
atics.oxfordjournals.org/D
ownloaded from
3.2 Reconstructed consensus network
From all QSPs of all protein biomarkers we computed a
spatiotemporal consensus network, as described in the Methods
section. This consensus network (Fig. 6b) describes the
central spatiotemporal expression patterns common to all
protein biomarkers of the three epidermal tissue samples.
The increase of DPK is correlated with INV and K1/10. The
rising flanks of INV/FIL/K1/10 constitute a subnetwork.
A well-known sequence of keratinocyte differentiation is
DPK! INV!FIL, which is clearly revealed by the green
and red arrows. K1/10 forms a feedback-loop to INV and
weakly to FIL. Integrin �6 is expressed only subbasally and
therefore, lacks shared flanks with other proteins, and is
excluded from the remaining network; it here serves as a
negative example. The high correlation values computed for all
shown arrows suggest a robust network structure.
3.3 Calculated confidences
To assess in how far the derived network arrows can be trusted,
we calculated a confidence value for each green (Fig. 6c) or red
(Fig. 6d) arrow in each tissue sample. This confidence is
calculated as the relative amount of time a protein leads during
a phase of mutual increase or decrease. For example Figure 6cshows that in epidermis 1, DPK is mutually increasing togetherwith K1/10, and during the whole period DPK is in the lead,
yielding a confidence of 100%. This confidence value holds truefor each tissue sample and is accordingly marked in dark green.Only slight differences in the quantitative amounts of the
confidence values can be observed between the tissue samples inFigure 6c. The confidence values for the falling flanks (Fig. 6d)appear a little less stable, although this is compensated by
averaging over all tissue samples. This higher variability has tobe attributed to the fact that the falling flanks constitute a muchshorter part of each QSP than the rising flanks. Generally, the
computed confidences strongly support the derived consensusnetwork structure.
4 DISCUSSION
For systems biology of complex stratified epithelia, it is of
particular importance to reconstruct the spatiotemporal geneand protein networks regulating cellular differentiation, apivotal cornerstone of epithelial homeostasis. To facilitate this
reconstruction, we here have presented a new method based onthe quantitative spatial analysis of immunofluorescence stainedstratified epithelial tissue sections.
The expression of biomarkers in squamous epithelia, includ-ing those previously mentioned has been described on aqualitative basis for decades. For example, alterations in
protein expression profiles have been described when studyingthe effects of retinoic acid on epidermal morphogenesis
(Asselineau et al., 1989). However, up till now, no quantitativespace- and time-resolving measurements of known epithelialbiomarkers have been published. In current literature, qualita-
tive textual descriptions of individual markers like ‘the studiedmarker is expressed in the suprabasal layers’ still prevail inconjunction with immunohistological images. We argue here,
that such qualitative descriptions do not appear sufficient forreconstructing networks of differentiation and thus answeringthe question of how epidermal homeostasis is achieved by the
collective behavior of the individual cells of the epithelial tissue.However, even after having acquired quantitative spatialprofiles of protein expression it has to be discussed how
network patterns could be derived from such data. Generally,although the epidermal samples exhibited a varying thickness,i.e. total number of epithelial cell layers, the biomarker profiles
were well comparable. This high level of comparability was dueto the normalization of the distance, reflecting differentiation
(Fig. 3b). Therefore, it can be expected that a robust regulatorynetwork controls this differentiation behavior. As there is anobvious relation between cellular differentiation and cell
distance to the connective tissue in one direction and to thestratum corneum in the other direction, this ‘distance’ isobviously incorporated indirectly into such a regulatory
network. Therefore, we try a first approach of interpretingthe obtained QSPs in the context of the upper and lower spatialboundary of the epithelium. Integrin �6 displayed a common
starting point for all profiles, since its expression starts at theepithelium–connective tissue interface. Therefore, the onset ofthe spatial expression of integrin �6 is naturally linked to the
basal lamina. Due to its restriction to the basal pole of basal cell
Fig. 5. Overview of both workflows for (a) computation of quantitative
spatial profiles (QSPs) from each subimage as described in Section 2.3
and (b) computation of protein networks described in Section 2.4.
(a) The mask of the epithelium E is calculated by combining the distinct
images of the phase-contrast image (PC) of the tissue section, the red
collagen and the blue cell nuclei. The epithelial mask E is converted into
a distance image having in each pixel p an intensity equal to the
minimum distance of p to the outline of E. The normalized distance d of
each pixel in E relative to the lower border of E is calculated as
described in Figure 2. The resulting quantitative spatial profile QSP
then gives for each distance d the average biomarker intensity (imaged
in green). By using distinct filters, all fluorescent images (red, green and
blue) are captured in different spectral channels. (b) For all five
biomarkers studied individual QSPs are determined and then linearly
correlated. Pairs of QSPs with sufficient correlation (450%) are
analyzed for mutual or unidirectional leadership. Accordingly, the
final directed graph describing the desired protein network results.
Reconstructing protein networks
3205
by guest on January 19, 2016http://bioinform
atics.oxfordjournals.org/D
ownloaded from
Fig. 6. Reconstructing the regulatory network from quantitative spatial profiles. (a) QSPs of the three epidermal biopsies (Epidermis 1–3). Symbols
denote: INT¼ integrin �6, K1/10¼ keratin 1/10, DPK¼desmoplakin, INV¼ involucrin, FIL¼ filaggrin. In each multidimensional profile, the nor-
malized protein abundance is given in relation to the normalized distance between the by the image processing estimated basal lamina, and the stratum
corneum. The shown SD for each biomarker is computed by averaging the profiles of each of the approximately 50 subimages. (b) Spatiotemporal
regulatory consensus network reconstructed from the profiles of the three epidermal samples. Green arrows indicate a correlated mutual increase in
abundance during which the protein at the start of the arrow leads, i.e. it has a higher relative abundance than the other protein. Bidirectional arrows
indicate that both proteins alternate in taking the lead. Red arrows indicate a correlatedmutual decrease. The small numbers annotating all arrows give
the undirected mutual correlation values. Only relative protein abundances above 20% are considered. For example, the increase of DPK is correlated
with INV and K1/10. The high correlation values computed for all shown arrows indicate a robust network structure. ‘NL’ (No Lead) instead of a
correlation value is given, when in the respective tissue sample the leader protein did not precede the trailing protein for a long enough period of time
(resulting in an insufficient confidence). (c) Percent confidencematrices for the epidermal samples 1–3 for all mutually increasing (green) and decreasing
(red) flanks. Confidence is calculated as the relative amount of time a protein leads during a phase of mutual increase or decrease. For example in
epidermis 1, DPK is mutually increasing with K1/10, and during the whole period DPK is in the lead, yielding a confidence of 100%.
N.Grabe et al.
3206
by guest on January 19, 2016http://bioinform
atics.oxfordjournals.org/D
ownloaded from
keratinocytes in epidermis, integrin �6 lacks a leading positionin relation to the other biomarkers, neither during decrease norduring increase. Therefore, it is missing in the consensus
network, and not marked in the corresponding confidencediagrams. Desmoplakin was the first protein we observed,which displayed a maximum expression in all of the analyzed
tissues. Its expression profiles were tightly connected to thebasal lamina, although its peak was measured at �40%differentiation, thereby indicating also a correlation to the
stratum corneum. K1/10 expression profiles started also veryearly, but remained then on a medium level of expression toreach its peak in the stratum corneum. Thus, it also seems to be
regulated in dependence of the basal layer as well as independence of the stratum corneum. Involucrin expressionappeared at a half-way distance, therefore strongly suggesting a
correlation more to the stratum corneum. The filaggrin profilesshowed a peak expression close to the maximum distance fromthe epithelium–connective tissue interface. It therefore seems tobe regulated in dependence of the stratum corneum.
However, discussing the regulation of spatial proteinexpression solely in the context of the basal lamina and thestratum corneum appears rather vague, and further, cannot
explain how the expression of the studied set of proteins iscoordinated to achieve controlled differentiation. For such ananalysis, the spatiotemporal patterns of protein coregulation
have to be characterized in more detail. This can better beaccomplished by the creation of a spatiotemporal proteinnetwork illustrating the interdependence of the individual
biomarkers. Such a network serves two purposes: first, it canuncover regulatory relationships between the proteins understudy. Second, the reconstructed single networks can also serve
as generalized descriptions of QSPs. Clearly, due to multipleinfluences, the absolute values of two QSPs may not be easilycomparable, e.g. due to different sites from the head and neck
region from which the epidermal biopsies were taken, differentages, etc. Therefore, reconstructing networks from QSPs can beseen as a potential tool of comparing QSPs obtained from
different tissue samples. The protein consensus network weobtained showed that indeed all the QSPs we obtained for thethree epidermal samples were very well comparable, despite
locoregional differences. High and similar correlation valueswere obtained for most arrows (Fig. 6b). The conductedconfidence analysis further supported the obtained network
topology (Fig. 6c,d). Especially for mutual increases, highconfidences for the increasing flanks could be observed(Fig. 6c). For mutual decreases, the results were not as robust
as for increases because the time periods of decrease showed tobe much shorter than those of increase. Moreover, unspecificstaining in the stratum corneum could have influenced the
QSPs. Here, potential improvements could be made for thenetwork reconstruction algorithm, e.g. through prior splineinterpolation of the QSPs.
An arrow in a protein networks computed by our methoddoes not necessarily imply a direct functional interactionbetween proteins, like it occurs between a transcription factor
and a gene promoter. Rather, it aims at analyzing whichproteins act concertedly or are coregulated in the same way.Therefore, we suggest that based on a highly informative
consensus network, in a second step, further functional studies
concerning the biomarkers could be performed. Clearly, at thecurrent stage we analyzed a selective set of epithelium-specific
molecules, leading to the presented consensus network ofproteins. But the goal of our study was to show that the general
approach is feasible and should be applied further to morecomplicated networks. In addition, the ability to reconstruct a
common spatiotemporal consensus network from three epider-
mal samples derived from different head and neck regionspoints to the potential of the new method to be more generally
applied to squamous epithelial tissues.Taken together, we presented a novel method for obtaining
quantitative spatial profiles of biomarkers reflecting differentstages of epithelial differentiation, and for reconstructing
spatiotemporal networks of epithelial differentiation fromthese profiles. By applying our method to samples of different
patients, we have shown that the reconstructed spatiotemporal
networks may serve as a helpful means for comparing andgeneralizing quantitative spatial profiles of biomarkers. From
our point of view, our approach may therefore serve as avaluable quantitative extension to current qualitative conven-
tional approaches of mapping the proteome expression data.Our method is currently based on images of immunofluor-
escent tissue sections but may also be applied to other sourcesof images. If in the future other imaging techniques like
MALDI-imaging improve substantially in their spatial resolu-
tion, our approach could also be adapted to such sources ofdata. But independent of the used imaging technique, the
spatial resolution of multiple biomarkers and the reconstruc-tion of protein networks from this data, will be a prerequisite
for the prospective in silico modeling and systems biological
analysis of epithelial differentiation and tissue homeostasis.
ACKNOWLEDGEMENTS
The works have been conducted with support from theHAMAMATSU Tissue Imaging and Analysis (TIGA) Center
and the NIKON Imaging Center at the BIOQUANT Facilityof the University Heidelberg. We also wish to thank Dr Claudia
Haag for support with tissue. For assistance and criticalreading of the manuscript we are grateful to Dr Michael Grabe,
Prof. Dr Karsten Neuber (University Hospital Hamburg-
Eppendorf) and Dr Hauke Busch (German Cancer ResearchCenter). BfR/Federal Institute of Risk Assessment, Berlin
(WK 3-1328-192) to N.G. and Dietmar-Hopp-Stiftung GmbH,Fachbereich Medizin, St. Leon Rot to T.S.
Conflict of Interest: A related patent application has been filed.
REFERENCES
Asselineau,D. et al. (1989) Retinoic acid improves epidermal morphogenesis.
Dev. Biol., 133, 322–335.
Banks-Schlegel,S. and Green,H. (1981) Involucrin synthesis and tissue assembly
by keratinocytes in natural and cultured human epithelia. J. Cell Biol., 90,
732–737.
Caldwell,R.L. and Caprioli,R.M. (2005) Tissue profiling by mass spectrometry: a
review of methodology and applications. Mol. Cell. Proteomics, 4, 394–401.
Candi,E. et al. (2005) The cornified envelope: a model of cell death in the skin.
Nat. Rev. Mol. Cell Biol., 6, 328–340.
Chu,P.G. and Weiss,L.M. (2002) Keratin expression in human tissues and
neoplasms. Histopathology, 40, 403–439.
Reconstructing protein networks
3207
by guest on January 19, 2016http://bioinform
atics.oxfordjournals.org/D
ownloaded from
Cole,J. et al. (2001) Early gene expression profile of human skin to injury using
high-density cDNA microarrays. Wound Repair Regen., 9, 360–370.
Eckert,R.L. et al. (2002) Keratinocyte survival, differentiation, and death:
many roads lead to mitogen activated protein kinase. J. Investig. Dermatol.
Symp. Proc., 7, 36–40.
Emmert-Buck,M.R. et al. (1996) Laser capture microdissection. Science, 274,
998–1001.
Grabe,N. and Neuber,K. (2005) A multicellular systems biology model predicts
epidermal morphology, kinetics and Ca2þ flow. Bioinformatics, 21,
3541–3547.
Grabe,N. and Neuber,K. (2007) Simulating psoriasis by altering transit
amplifying cells. Bioinformatics, 23, 1309–1312.
Haftek,M. et al. (1991) Immunocytochemical evidence for a possible role of cross-
linked keratinocyte envelopes in stratum corneum cohesion. J. Histochem.
Cytochem., 39, 1531–1538.
Hahn,W. and Weinberg,R.A. (2002) Modelling the molecular circuitry of cancer.
Nat. Rev. Cancer, 2, 331–341.
Huang,C.M. et al. (2003) Comparative proteomic profiling of murine skin.
J. Invest. Dermatol., 121, 51–64.
Ideker,T. et al. (2001) A new approach to decoding life: systems biology.
Annu. Rev. Genomics Hum. Genet., 2, 343–372.
Kitano,H. (2002) Computational systems biology. Nature, 420, 206–210.
Koria,P. et al. (2003) Gene expression profile of tissue engineered skin subjected
to acute barrier disruption. J. Invest. Dermatol., 121, 368–382.
Kottke,M.D. et al. (2006) The desmosome: cell science lessons from human
diseases. J. Cell Sci., 119, 797–806.
Lechler,T. and Fuchs,E. (2005) Asymmetric cell divisions promote stratification
and differentiation of mammalian skin. Nature, 437, 275–280.
Noble,D. (2004) Modeling the heart. Physiology (Bethesda), 19, 191–197.
Rice,R.H. and Green,H. (1979) Presence in human epidermal cells of a soluble
protein precursor of the cross-linked envelope: activation of the cross-linking
by calcium ions. Cell, 18, 681–694.
Schubert,W. (2006) Exploring molecular networks directly in the cell.
Cytometry A., 69, 109–112.
Schubert,W. et al. (2006) Analyzing proteome topology and function by
automated multidimensional fluorescence microscopy. Nat. Biotechnol., 24,
1270–1278.
Stoler,A. et al. (1988) Use of monospecific antisera and cRNA probes to localize
the major changes in keratin expression during normal and abnormal
epidermal differentiation. J. Cell Biol., 107, 427–446.
Tomakidi,P. et al. (2003) Discriminating expression of differentiation markers
evolves in transplants of benign and malignant human skin keratinocytes
through stromal interactions. J. Pathol., 200, 298–307.
Uhlen,M. and Ponten,F. (2005) Antibody-based proteomics for human tissue
profiling. Mol. Cell. Proteomics, 4, 384–393.
Uhlen,M. et al. (2005) A human protein atlas for normal and cancer tissues based
on antibody proteomics. Mol. Cell. Proteomics, 4, 1920–1932.
Walts,A.E. et al. (1985) Involucrin, a marker of squamous and urothelial
differentiation. An immunohistochemical study on its distribution in normal
and neoplastic tissues. J. Pathol., 145, 329–340.
Wan,H. et al. (2004) Striate palmoplantar keratoderma arising from desmoplakin
and desmoglein 1 mutations is associated with contrasting perturbations of
desmosomes and the keratin filament network. Br. J. Dermatol., 150, 878–891.
Watt,F.M. et al. (1987) Effect of growth environment on spatial expression of
involucrin by human epidermal keratinocytes. Arch. Dermatol. Res., 279,
335–340.
N.Grabe et al.
3208
by guest on January 19, 2016http://bioinform
atics.oxfordjournals.org/D
ownloaded from