Principles of Bioimage Informatics: Focus on
Machine Learning of Cell Patterns
Luis Pedro Coelho1,2,3, Estelle Glory-Afshar3,6, Joshua Kangas1,2,3,
Shannon Quinn2,3,4, Aabid Shariff1,2,3, and Robert F. Murphy1,2,3,4,5,6
1 Joint Carnegie Mellon University–University of Pittsburgh Ph.D.
Program in Computational Biology
2 Lane Center for Computational Biology, Carnegie Mellon University
3 Center for Bioimage Informatics, Carnegie Mellon University
4 Department of Biological Sciences, Carnegie Mellon University
5 Machine Learning Department, Carnegie Mellon University
6 Department of Biomedical Engineering, Carnegie Mellon University
Abstract. The field of bioimage informatics concerns the development
and use of methods for computational analysis of biological images. Tra-
ditionally, analysis of such images has been done manually. Manual an-
notation is, however, slow, expensive, and often highly variable from one
expert to another. Furthermore, with modern automated microscopes,
hundreds to thousands of images can be collected per hour, making man-
ual analysis infeasible.
This field borrows from the pattern recognition and computer vi-
sion literature (which contain many techniques for image processing and
recognition), but has its own unique challenges and tradeoffs.
Fluorescence microscopy images represent perhaps the largest class
of biological images for which automation is needed. For this modality,
typical problems include cell segmentation, classification of phenotypical
response, or decisions regarding differentiated responses (treatment vs.
control setting). This overview focuses on the problem of subcellular lo-
cation determination as a running example, but the techniques discussed
are often applicable to other problems.
1 Introduction
Bioimage informatics employs computational and statistical techniques to ana-
lyze images and related metadata. Bioimage informatics approaches are useful in
a number of applications, such as measuring the effects of drugs on cells [1], local-
izing cellular proteomes [2], tracking of cellular motion and activity [3], mapping
of gene expression in developing embryos [4,5] and adult brains [6], and many
C. Blaschke and H. Shatkay (Eds.): ISBM/ECCB 2009, LNBI 6004, pp. 8–18, 2010.
c
© Springer-Verlag Berlin Heidelberg 2010

Principles of Bioimage Informatics: Focus on Machine Learning 9
others. Traditionally, bioimage analysis has been done by visual inspection, but
this is tedious and error-prone. Results from visual analysis are not easily com-
pared between papers or groups. Furthermore, as bioimage data is increasingly
used to understand gene function on a genome scale, datasets of subtle pheno-
type changes are becoming too large for manual analysis [7,8,9]. For example, it
is estimated that having a single image for every combination of cell type, pro-
tein, and timescale would require on the order of 100 billion images [10]. Over
the past fourteen years, the traditional visual, knowledge-capture approach has
begun to be replaced with automated, data-driven approaches [11,12,13,14,15].
Approaches to quantitatively associate image feature information with additional
anatomical and ontological knowledge to generate digital atlases have also been
described [16].
This brief review will focus on bioimage informatics approaches to analyz-
ing protein subcellular patterns, with the goal of illustrating many of the prin-
ciples that are also relevant to other areas of bioimage informatics. The goal
of work in this area is to devise a generalizable, verifiable, mechanistic model
of cellular organization and behavior that is automatically derived from
images [17].
The most commonly used method for determining subcellular location is fluo-
rescence microscopy. Images are collected by tagging a protein or other molecule
so that it becomes visible under the fluorescence microscope.
2 Cell Segmentation
A first step in many analysis pipelines is segmentation, which can occur at several
levels (e.g., separating nuclei, cells, tissues). This task has been an active field
of research in image processing over the last 30 years, and various methods have
been proposed and analysed depending on the modality, quality, and resolution
of the microscopy images to analyze [18]. We only discuss two commonly used
approaches, Voronoi and seeded watershed.
The two first methods require seed regions to be defined. These can be simply
locally bright regions or can be defined by a more complex procedure. For ex-
ample, nuclear segmentation, which is a difficult problem by itself [19], is often
used to provide seeds for cell-level segmentation.
In Voronoi segmentation [20], a pixel is assigned to the closest seed. This is a
very fast method and works well for sparse regions, but does not take into account
the location of the cells and makes serious mistakes if the field is crowded.
Seeded watershed segmentation is a region growing approach [21] in which
the image can be considered as a landscape with the pixel intensities as el-
evation. From the seeds, the basins of the landscape are flooded. When two
basins are about to merge, a dam is built that represents the boundary between
the two cells. This method works well if the seeds are carefully defined (see
Figure 1).

10 L.P. Coelho et al.
Fig. 1. Seeded Watershed. The nuclear borders (shown in grey) were first identified by
a model-based algorithm [22] followed by watershed on the protein channel (resulting
in the borders shown in white). Images have been contrast stretched for publication.
3 Supervised Classification
Many applications can be posed as a pattern recognition problem, i.e., given a
set of examples of different classes, attempt to group other datapoints into these
classes.
Given that even a small image has thousands of pixels, a direct pixel com-
parison is impossible. Furthermore, two images that differ only in a rigid body
motion can have no common pixel values, but represent exactly the same cell
state. Thus, the standard approach is to describe the image by a much smaller
set of features, where a feature is a numeric function of the image. Once images
have been summarized by this smaller set of features, machine learning methods
can be applied to learn a classifier.
Features can be computed at several levels: directly from a field that may
contain multiple cells, from cell regions (once the image has been segmented as
described in the previous section), or from individual subcellular objects in the
image. Certain feature classes, such as texture features, are applicable in other
vision problems, but features designed specifically for this problem have also
been presented (such as those that relate location of the protein of interest to
the position of the nucleus) [13].
In some cases, a classifier outperforms a human expert in classification of pro-
tein location patterns. The results of one such experiment are shown in Figure 2,
where the computer achieves 92% accuracy classifying ten location classes [23],
while the human interpreter can only achieve 83% [24]. In another classification
study, on the problem of cell detection, it was observed that computers perform

Principles of Bioimage Informatics: Focus on Machine Learning 11
0 20 40 60 80 100
computer accuracy (%)
0
20
40
60
80
100
h
u
m
a
n
a
c
c
u
r
a
c
y
(
%
)
Fig. 2. A comparison of the accuracy in classification indicates that a supervised clas-
sification algorithm can perform as well or better than a human expert in recognizing
subcellular patterns of ten proteins. Each circle represents one protein.
comparably to a medium-quality expert, but were still outperformed by an expe-
rienced expert. However, high variation between experts was observed [25].
4 Shape Analysis
Shapes of cells, nuclei, and organelles are critical for their function. For example,
the shape of cells contribute to the overall tissue architecture that have charac-
teristic tissue-specific functions [26]. In pathology, shape is also a useful indicator
of deviations from the wild type phenotype. A classical example of a condition
that can be identified by the shape of cells is sickle cell anemia, which results in
deformed red blood cells. Automated analysis of nuclear shape has received the
most attention given the importance of the nucleus in diagnosis.
One approach is to compute a small number of numerical features from the
shape image that measure basic properties such as size or the convexity of the
shape1 [27,28].
This feature-based approach is simply an instance of the supervised classifi-
cation framework as described in Section 3, but approaches that are specifically
designed for shape have also been proposed. In particular, diffeomorphic methods
have been applied in this area because the shape space of nuclei is non-linear [29].
By capturing how much deformation is needed to morph one shape into another,
a distance function in the space of functions is defined. It is then possible to in-
terpolate in the space of shapes [30] to generate intermediate shapes or place
a collection of shapes in a low dimensional space in a way that preserves their
distance relationships as well as possible. Figure 3 shows deformation distance
as a starting nucleus shape is deformed to a target shape.
1 In many applications, it is important to remove the influence of size by normalisation,
so it depends on the application whether size (or, analogously, orientation) should
be captured or normalized out.

12 L.P. Coelho et al.
source 25% 50% 75% target
Fig. 3. The figure shows the transformation using a mapping function from one shape
to another. The deformation distance of each image from the starting shape is shown
as a percentage of total distance.
Analysis on cell and organelle shapes can be done in a manner similar to
nuclear shape. A vast majority of methods use parametric representation of cell
shapes [31,32,33].
5 Subcellular Pattern Unmixing
We have described methods to classify fluorescence images according to the
depicted subcellular pattern using supervised classification approaches. These
methods perform well for defined patterns, but cannot handle patterns that are
composed of a mixture. For example, a protein which is partially in the plasma
membrane and partially in the nucleus will exhibit a mixed pattern. A system
whose model is of discrete assignments of patterns to classes will not be able to
even represent the relationship between the plasma membrane and nuclear pat-
terns and the mixed intermediates. In some situations, defining a mixed class as
an extra class might be an acceptable, if inelegant, solution. However, not only
does the number of classes grow combinatorially, but if a researcher is interested
in quantifying the fraction of fluorescence in each compartment (for example, to
study translocation as a function of time), then this solution will not be applicable.
Pattern unmixing directly models this situation. In the supervised case, the
problem is as follows: the system is given examples of pure patterns and attempts
to unmix mixed inputs (i.e., assign mixture fractions for each input condition).
In the unsupervised form, the system is simply given a collection of images and
must identify the fundamental patterns of which mixtures are made.
As one possible approach to this problem, we have described an object-based
method for unmixing of subcellular patterns [34]. These methods work on the
basis of identifying discrete objects in the image. An object is a contiguous set
of pixels that differs in some way from its surroundings (in the simplest case,
defined as being above a global threshold value). Patterns are now described by
the properties of objects that they show. Mixed patterns will show objects that
are characteristic of the basic patterns that compose them. Several variations of
these methods have been proposed and this is still an active area of research.

Principles of Bioimage Informatics: Focus on Machine Learning 13
To evaluate the quality of the proposed methods, the correlation between the
input mixture coefficients and the underlying fraction of fluorescence in each
compartment needs to be computed. In order to know the underlying fractions,
Peng et al. [35] built a test set where two dyes that locate differently, but fluo-
resce similarly were imaged at different concentrations. For supervised unmixing,
an 83% correlation with the underlying concentrations was obtained. For the un-
supervised case, preliminary results of 72% correlation have been obtained [36].
6 Publicly Available Databases and Analysis of Large
Datasets
In previous sections, we described the computational ways in which data is com-
monly processed. In this section, we give an overview of the sources of high-
throughput image data that have been made available by researchers. Some of
these have been the subject of automatic analysis, while others have not yet
been processed in this way.
Besides the collection of images, these databases also provide annotations,
typically manually assigned, describing the protein distribution within cells (see
Table 1). The experimental techniques can be grouped into two major fami-
lies of approaches: fusing the protein of interest to a fluorescence tag, or using
fluorescent antibodies that bind to the protein of interest.
Table 1. Examples of microscopy image databases dedicated to subcellular protein
location within cells and tissues
Database Organism Summary
PSLID-HeLa H. Sapiens 10 protein distributions tagged by immunofluores-
cence, 100x magnification
LOCATE-HeLa H. Sapiens 10 protein distributions tagged by immunofluores-
cence, 60x magnification
PSLID-RandTag M. Musculus ca. 2000 proteins with cd egfp tagging, 40x magni-
fication
Lifedb C. Aethiops ca. 1000 proteins tagged with n- and c-terminal
eyfp and ecfp using 1500 cdna whose vector is ex-
pressed in Vero cells, 63x magnification
YPL.db S. Cerevisiae 371 gfp fusion proteins or vital staining (in 2005)
Yeastgfp S. Cerevisiae 4,156 strains with detectable GFP signal among the
6,029 GFP positive strains, proteins tagged with
gfp fusion at their n- and c-terminals, imaged at
100x magnification
HPA-IHC H. Sapiens 5,000 proteins in 48 normal tissues, 20 cancer tissues,
and 47 different cell lines labeled immunohistochem-
ically (dab and hematoxylin), imaged at 20x
HPA-IF H. Sapiens ca. 5,000 proteins in 3 different cell lines (A-431,
U-2 OS, and U-251MG) tagged with immunofluo-
rescence, imaged at 63x

14 L.P. Coelho et al.
In 2002, the first database dedicated to location proteomics, the Protein Sub-
cellular location database (pslid), was released. Initially composed of 2d single
cell images, it has since incorporated various 2d and multicell image datasets.
The most recently added dataset represents a collection of images generated by
the cd-tagging protocol applied on NIH 3T3 cells [37]. Pslid and its associated
software have a unique interface2 which allows feature calculation, statistical
comparison, clustering, classification and the creation of generative models [13].
Given its importance as a model organism, several large collection of yeast im-
ages are available. The Yeast Protein Localization database was released in 2002
and 2005 [38,39] (YPL.db2)3. The lack of information on image properties, such
as the pixel resolution, limit the potential of automatic analysis of protein dis-
tributions in YPL.db2. Concurrently, a collection of lines producing gfpfusion
proteins for 75% of the open reading frames (orfs) in S. Cerevisiae was re-
leased in 20034 [40]. The assignment of each protein into one or a combination
of 22 different locations was visually performed. Some fluorescent colocaliza-
tion experiments with fluorescent markers were made to disambiguate uncertain
cases. The same dataset was analyzed by a supervised machine learning approach
that showed 81% to 95% accuracy in predicting the correct protein distribution
[41]. The accuracy is calculated as the agreement between the labels found by the
computer and the hand labeling. However, a closer observation of the mismatches
revealed that errors were found in both approaches, human and computer.
The Database for Localization, Interaction, Functional assays and Expression
of Proteins (LIFEdb) was created to collect microscopy images of fluorescent fu-
sion proteins produced from human full-length cdnas expressed in mammalian
cell lines [42,43]5. The dataset was generated by creating fluorescent fusion pro-
teins in Vero cells [44].
The Human Protein Atlas (hpa) project studies the human proteome in situ
using antibodies. They collect images of proteins in stained tissues and cell lines
in brightfield and fluorescence microscopies [45,46]. They aim to cover the entire
human proteome (estimated to consist of ca. 20,000 non-redundant proteins) by
2014. The current release of the hpa database (version 5.0) contains images for
more than 8000 antibodies targeting 7000 proteins [46]. The largest collection of
images is produced by immunohistochemically (ihc) stained tissue microarrays
from normal and cancerous biopsies. The database was extended with 47 differ-
ent human cell lines commonly used in research and 12 primary blood cells. In
addition, a collection of confocal microscopy images of 3 different human cell lines
were tagged using immunofluorescence (if) [47]. The protein distributions were as-
signed by visual inspection to three different compartments—nuclear, cytoplasm,
membranous—for the IHC images. The higher resolution of if images allows a
finer distinction of the protein subcellular distribution into the major cellular or-
ganelles. Two machine learning based systems showed good results in generating
2 Available at http://pslid.cbi.cmu.edu/
3 Available at http://ypl.uni-graz.at/
4 Available at http://yeastgfp.yeastgenome.org/
5 Available at http://www.lifedb.de/

Principles of Bioimage Informatics: Focus on Machine Learning 15
labels on both image collections. The classification of the if images gave 90% accu-
racy to distinguish 9 subcellular compartments [48] and 81% to distinguish 8 sub-
cellar distribution from non-segmented ihc tissue images [49]. Preliminary results
on identifying potential cancer biomarkers by automatically comparing protein
distributions in normal and cancer images have also been presented [50].
7 Discussion
This review presented an overview of bioimage informatics, focusing on the prob-
lems in analysing fluorescence microscope images.
As a result of the continued progress of bioimage analysis, we increasingly
observe domains where automated methods either outperform human analysis
or are, at least, comparable with it. Thus, areas of disagreement between the
two approaches cannot simply be marked as “classification error on the part of
the algorithm.” Although humans still outperform computational approaches in
general vision tasks by a large margin, the same is not necessarily true in the
case of non-natural images of objects which humans do not encounter in their
everyday lives (such as fluorescence microscopy images).
Bioimages also generate their own particular problems such as pattern un-
mixing or learning generative models. These are unique problems that still do
not have a definite answer. Biomage informatics still has many active research
questions, in the development of entirely new methods to capture information
in images, adaptation of existing ones from sister fields, or simply in solving the
challenges of applying them to very large collections of data in real time with
minimal user intervention. The need for such developments is underscored by
the increase in publicly available data which is yet to be fully explored.
References
1. Perlman, Z.E., Slack, M.D., Feng, Y., Mitchison, T.J., Wu, L.F., Altschuler, S.J.:
Multidimensional Drug Profiling By Automated Microscopy. Science 306(5699),
1194–1198 (2004)
2. Boland, M.V., Murphy, R.F.: A Neural Network Classifier Capable of Recognizing
the Patterns of all Major Subcellular Structures in Fluorescence Microscope Images
of HeLa Cells. Bioinformatics 17, 1213–1223 (2001)
3. Meijering, E., Smal, I., Danuser, G.: Tracking in molecular bioimaging. IEEE Signal
Processing Magazine 23(3), 46–53 (2006)
4. Peng, H., Myers, E.W.: Comparing in situ mRNA expression patterns of drosophila
embryos. In: 8th Intl. Conf. on Computational molecular biology, pp. 157–166
(2004)
5. Zhou, J., Peng, H.: Automatic recognition and annotation of gene expression pat-
terns of fly embryos. Bioinformatics 23(5), 589–596 (2007)
6. Le´cuyer, E., Tomancak, P.: Mapping the gene expression universe. Current Opinion
in Genetics & Development 18(6), 506–512 (2008)
7. Boland, M.V., Murphy, R.F.: After Sequencing: Quantitative Analysis of Protein
Localization. IEEE Engineering in Medicine and Biology Magazine 18(5), 115–119
(1999)

16 L.P. Coelho et al.
8. Chen, X., Murphy, R.F.: Objective Clustering of Proteins Based on Subcellular
Location Patterns. Journal Biomedical Biotechnology 2005(2), 87–95 (2005)
9. Roques, E., Murphy, R.: Objective evaluation of differences in protein subcellular
distribution. Traffic 3, 61–65 (2002)
10. Murphy, R.F.: Putting proteins on the map. Nature Biotechnology 24, 1223–1224
(2006)
11. Conrad, C., Erfle, H., Warnat, P., Daigle, N., Lo¨rch, T., Ellenberg, J., Pepperkok,
R., Eils, R.: Automatic Identification of Subcellular Phenotypes on Human Cell
Arrays. Genome Research 14, 1130–1136 (2004)
12. Gasparri, F., Mariani, M., Sola, F., Galvani, A.: Quantification of the Proliferation
Index of Human Dermal Fibroblast Cultures with the ArrayScan High-Content
Screening Reader. Journal of Biomolecular Screening 9(3), 232–243 (2004)
13. Glory, E., Murphy, R.F.: Automated Subcellular Location Determination and High
Throughput Microscopy. Developmental Cell 12(1), 7–16 (2007)
14. Hamilton, N.A., Pantelic, R.S., Hanson, K., Teasdale, R.D.: Fast automated cell
phenotype image classification. BMC Bioinformatics 8, 110 (2007)
15. Huang, K., Lin, J., Gajnak, J., Murphy, R.F.: Image Content-based Retrieval
and Automated Interpretation of Fluorescence Microscope Images via the Protein
Subcellular Location Image Database. In: IEEE Intl. Symp. Biomedical Imaging,
pp. 325–328 (2002)
16. Lein, E., Hawrylycz, M., Ao, N.: Genome-wide atlas of gene expression in the adult
mouse brain. Nature 445, 168–176 (2006)
17. Murphy, R.F.: Systematic description of subcellular location for integration with
proteomics databases and systems biology modeling. In: IEEE Intl. Symp. Biomed-
ical Imaging, pp. 1052–1055 (2007)
18. Nattkemper, T.W.: Automatic segmentation of digital micrographs: A survey.
Studies in health technology and informatics 107(2), 847–851 (2004)
19. Coelho, L.P., Shariff, A., Murphy, R.F.: Nuclear segmentation in microsope cell
images: A hand-segmented dataset and comparison of algorithms. In: IEEE Intl.
Symp. Biomedical Imaging, pp. 518–521 (2009)
20. Jones, T.R., Carpenter, A.E., Golland, P.: Voronoi-based segmentation of cells on
image manifolds. In: Liu, Y., Jiang, T.-Z., Zhang, C. (eds.) CVBIA 2005. LNCS,
vol. 3765, pp. 535–543. Springer, Heidelberg (2005)
21. Beucher, S.: Watersheds of functions and picture segmentation. In: IEEE Intl Conf.
on Acoustics, Speech and Signal Processing, Paris, pp. 1928–1931 (1982)
22. Lin, G., Adiga, U., Olson, K., Guzowski, J.F., Barnes, C.A., Roysam, B.: A hybrid
3D watershed algorithm incorporating gradient cues and object models for auto-
matic segmentation of nuclei in confocal image stacks. Cytometry Part A 56A(1),
23–36 (2003)
23. Huang, K., Murphy, R.F.: Automated Classification of Subcellular Patterns in
Multicell images without Segmentation into Single Cells. In: IEEE Intl. Symp.
Biomedical Imaging, pp. 1139–1142 (2004)
24. Murphy, R., Velliste, M., Porreca, G.: Robust Numerical Features for Description
and Classification of Subcellular Location Patterns in Fluorescence Microscope
Images. Journal of VLSI Signal Processing-Systems for Signal, Image, and Video
Technology 35, 311–321 (2003)
25. Nattkemper, T.W., Twellmann, T., Schubert, W., Ritter, H.J.: Human vs. machine:
Evaluation of fluorescence micrographs. Computers in Biology and Medicine 33(1),
31–43 (2003)
26. Allen, T.D., Potten, C.S.: Significance of cell shape in tissue architecture. Na-
ture 264(5586), 545–547 (1976)

Principles of Bioimage Informatics: Focus on Machine Learning 17
27. Olson, A.C., Larson, N.M., Heckman, C.A.: Classification of cultured mammalian
cells by shape analysis and pattern recognition. Proceedings of the National
Academy of Sciences (USA) 77(3), 1516–1520 (1980)
28. Pincus, Z., Theriot, J.A.: Comparison of quantitative methods for cell-shape anal-
ysis. Journal of microscopy 227, 140–156 (2007)
29. Rohde, G.K., Ribeiro, A.J.S., Dahl, K.N., Murphy, R.F.: Deformation-based nu-
clear morphometry: capturing nuclear shape variation in hela cells. Cytometry Part
A 73A(4), 341–350 (2008)
30. Peng, T., Wang, W., Rohde, G.K., Murphy, R.F.: Instance-based generative biolog-
ical shape modeling. In: IEEE Intl. Symp. Biomedical Imaging, vol. 1, pp. 690–693
(2009)
31. Cootes, T.F., Taylor, C.J., Cooper, D.H., Graham, J.: Active shape models—their
training and application. Computer Vision and Image Understanding 61(1), 38–59
(1995)
32. Albertini, M.C., Teodori, L., Piatti, E., Piacentini, M.P., Accorsi, A., Rocchi,
M.B.L.: Automated analysis of morphometric parameters for accurate definition
of erythrocyte cell shape. Cytometry Part A 52A(1), 12–18 (2003)
33. Lehmussola, A., Ruusuvuori, P., Selinummi, J., Huttunen, H., Yli-Harja, O.: Com-
putational framework for simulating fluorescence microscope images with cell pop-
ulations. IEEE Trans. Medical Imaging 26(7), 1010–1016 (2007)
34. Zhao, T., Velliste, M., Boland, M., Murphy, R.F.: Object type recognition for
automated analysis of protein subcellular location. IEEE Trans. on Image Process-
ing 14(9), 1351–1359 (2005)
35. Peng, T., Bonamy, G.M., Glory, E., Daniel Rines, S.K.C., Murphy, R.F.: Auto-
mated unmixing of subcellular patterns: Determining the distribution of probes
between different subcellular locations. Proceedings of the National Academy of
Sciences, USA (2009) (in press)
36. Coelho, L.P., Murphy, R.F.: Unsupervised unmixing of subcellular location pat-
terns. In: Proceedings of ICML-UAI-COLT 2009 Workshop on Automated Inter-
pretation and Modeling of Cell Images (Cell Image Learning), Montreal, Canada
(2009)
37. Garc´ıa Osuna, E., Hua, J., Bateman, N.W., Zhao, T., Berget, P.B., Murphy, R.F.:
Large-scale automated analysis of location patterns in randomly tagged 3T3 cells.
Annals of Biomedical Engineering 35, 1081–1087 (2007)
38. Habeler, G., Natter, K., Thallinger, G.G., Crawford, M.E., Kohlwein, S.D., Tra-
janoski, Z.: YPL.db: the Yeast Protein Localization database. Nucleic Acids Re-
search 30(1), 80–83 (2002)
39. Kals, M., Natter, K., Thallinger, G.G., Trajanoski, Z., Kohlwein, S.D.: Ypl.db2:
the yeast protein localization database, version 2.0. Yeast 22(3), 213–218 (2005)
40. Huh, W.K., Falvo, J.V., Gerke, L.C., Carroll, A.S., Howson, R.W., Weissman,
J.S., O’Shea, E.K.: Global analysis of protein localization in budding yeast. Na-
ture 425(6959), 686–691 (2003)
41. Chen, S.C., Zhao, T., Gordon, G., Murphy, R.: Automated image analysis of protein
localization in budding yeast. Bioinformatics 23(13), 66–71 (2007)
42. Bannasch, D., Mehrle, A., Glatting, K.H., Pepperkok, R., Poustka, A., Wiemann,
S.: LIFEdb: a database for functional genomics experiments integrating informa-
tion from external sources, and serving as a sample tracking system. Nucleic Acids
Research 32, D505–D508 (2004)
43. del Val, C., Mehrle, A., Falkenhahn, M., Seiler, M., Glatting, K.H., Poustka, A.,
Suhai, S., Wiemann, S.: High-throughput protein analysis integrating bioinformat-
ics and experimental assays. Nucleic Acids Research 32(2), 742–748 (2004)

18 L.P. Coelho et al.
44. Simpson, J., Wellenreuther, R., Poustka, A., Pepperkok, R., Wiemann, S.: Sys-
tematic subcellular localization of novel proteins identified by large-scale cDNA
sequencing. EMBO reports 1(3), 287–292 (2000)
45. Uhlen, M., Bjorling, E., Agaton, C., Szigyarto, C.A.K., Amini, B., Andersen,
E., Andersson, A.C., Angelidou, P., Asplund, A., Asplund, C., Berglund, L.,
Bergstrom, K., Brumer, H., Cerjan, D., Ekstrom, M., Elobeid, A., Eriksson, C.,
Fagerberg, L., Falk, R., Fall, J., Forsberg, M., Bjorklund, M.G., Gumbel, K.,
Halimi, A., Hallin, I., Hamsten, C., Hansson, M., Hedhammar, M., Hercules, G.,
Kampf, C., Larsson, K., Lindskog, M., Lodewyckx, W., Lund, J., Lundeberg, J.,
Magnusson, K., Malm, E., Nilsson, P., Odling, J., Oksvold, P., Olsson, I., Oster,
E., Ottosson, J., Paavilainen, L., Persson, A., Rimini, R., Rockberg, J., Runeson,
M., Sivertsson, A., Skollermo, A., Steen, J., Stenvall, M., Sterky, F., Stromberg, S.,
Sundberg, M., Tegel, H., Tourle, S., Wahlund, E., Walden, A., Wan, J., Wernerus,
H., Westberg, J., Wester, K., Wrethagen, U., Xu, L.L., Hober, S., Ponten, F.: A Hu-
man Protein Atlas for Normal and Cancer Tissues Based on Antibody Proteomics.
Molecular & Cellular Proteomics 4(12), 1920–1932 (2005)
46. Berglund, L., Bjo¨rling, E., Oksvold, P., Fagerberg, L., Asplund, A., Szigyarto,
C.A.K., Persson, A., Ottosson, J., Werne´rus, H., Nilsson, P., Lundberg, E., Siverts-
son, A., Navani, S., Wester, K., Kampf, C., Hober, S., Ponte´n, F., Uhle´n, M.:
A genecentric Human Protein Atlas for expression profiles based on antibodies.
Molecular & cellular proteomics 7(10), 2019–2027 (2008)
47. Lundberg, E., Sundberg, M., Gra¨slund, T., Uhle´n, M., Svahn, H.A.: A novel method
for reproducible fluorescent labeling of small amounts of antibodies on solid phase.
Journal of Immunological Methods 322(1-2), 40–49 (2007)
48. Newberg, J., Li, J., Rao, A., Ponten, F., Uhlen, M., Lundberg, E., Murphy, R.F.:
Automated analysis of human protein atlas immunofluorescence images. In: IEEE
Intl. Symp. Biomedical Imaging, pp. 1023–1026 (2009)
49. Newberg, J., Hua, J., Murphy, R.F.: Location Proteomics: Systematic Determina-
tion of Protein Subcellular Location. In: Systems Biology, vol. 500, pp. 313–332.
Humana Press (2009)
50. Glory, E., Newberg, J., Murphy, R.F.: Automated comparison of protein subcellular
location patterns between images of normal and cancerous tissues. In: IEEE Intl.
Symp. Biomedical Imaging, pp. 304–307 (2008)