PROTEIN CLASSIFICATION BY ANALYSIS OF 
CONFOCAL MICROSCOPIC IMAGES OF SINGLE CELLS 
Tanja Steckling?, Olaf Hellwich?®, Stephanie Wiilter®, Erich Wanker" 
? Technical University Berlin, Computer Vision and Remote Sensing, Sekr. FR 3-1, Franklinstr. 28/29, 10587 Berlin, 
Germany, Phone: +49-30-314-22796, Fax: +49-30-314-21104, e-mail: hellwich@fpk.tu-berlin.de 
? Max-Delbrück-Centrum für Molekulare Medizin (MDC), Berlin-Buch, Robert-Róssle-Str. 10, 13092 Berlin, 
Germany 
Commission WG V/3 
KEY WORDS: image analysis, feature selection, classification, medical image processing, microscopic imagery 
ABSTRACT: 
Proteins being present in a living cell fulfil a certain task in the cell. As a consequence of its functionality a protein is located in 
certain parts of the cell. If it is made visible the resulting patterns can help to identify the protein, as the spatial distribution of the 
visible structures depends on the functionality of the protein inside of the cell and, therefore, characterises the protein. The cells used 
for the experiments were COS-1 cells typically allowing easy microscopic data takes as the cells are much larger than their nuclei. 
With the help of a suitable parameterisation the proteins can be automatically identified. In order to derive such a parameterisation, 
features describing the spatial structure of the protein are extracted. The stochastic behaviour of the features is of major importance 
for the performance of the method. 
1. INTRODUCTION 
A protein present in a cell can be made visible by a chemical 
treatment with antibodies. The spatial distribution of the visible 
structures depends on the functionality of a protein inside of the 
“cell and characterises the protein. Therefore, it allows or at least 
helps to identify the protein. In this work a method to 
automatically classify proteins on the basis of single cell images 
is described. 
The imagery of COS-1 cells used here has been acquired by 
fluorescence confocal microscopy. From a data take, i.e. a focus 
series of images, the image optimally showing the spatial 
distribution of the protein has been selected. A single cell 
extracted from such an image constitutes the input to the 
algorithm described. 
In order to derive a parameterisation identifying proteins, 
features describing the spatial structure of the protein have to be 
extracted. An interactive classification of proteins by a human 
operator has shown that a classification accuracy of 95 to 100 
9$ is possible. Similar classification accuracy can be achieved 
by an automatic analysis when suitable features are selected. As 
the consecutively following steps of the procedure and the facts 
being their basis, such as probability density distributions, their 
derivation from training data, the choice of a classification 
method, and the derivation of a classification decision, are well 
known, feature selection or feature reduction is the crucial step 
of the procedure. The importance of feature reduction 
corresponds to the fact that in human vision, particularly in 
deriving decisions from visual information, the large amount of 
data/information in images based on high spatial and 
radiometric resolution is first severely reduced before being 
extended again by associating knowledge, e.g. about objects 
and context, in order to derive new knowledge or decisions in a 
process of thinking (BECKER-CARUS, 1981). 
Using our method, previously unknown proteins can be 
identified as long as the protein shows an individual spatial 
structure inside of the cell. With an automatic procedure, from a 
specific spatial structure conclusions with respect to the 
chemical role of the protein could be drawn, as the molecules 
appear where they are chemically active. This means that image 
analysis can provide a new method to proteomics research, 
possibly of efficiency previously unknown. It is our long term 
goal to derive and test such a method. 
2. PREVIOUS WORK 
BOLAND et al. (1997) describe a method to classify cellular 
protein localization patterns based on their appearance in 
fluorescence light microscope images. Numeric features were 
used as input values to either a classification tree or a neural 
network (BOLAND et al, 1998). MARKEY et al. (1999) 
developed methods for objectively choosing a typical image 
from a set of images, emphasizing cell biology. The methods 
include calculation of numerical features to describe protein 
patterns, calculation of similarity between patterns as a distance 
in feature space, and ranking of patterns by distance from the 
center of the distribution in feature space. The images chosen as 
most typical were in good agreement with the conventional 
understanding of organelle morphologies. MURPHY et al. (2000) 
describe an approach to quantitatively describe protein 
localization patterns and to develop classifiers able to recognize 
all major subcellular structures in fluorescence microscope 
images. Since fluorescence microscope images are a primary 
source of information about the location of proteins within 
cells, MURPHY et al. (2001) strive to build a knowledge-based 
system which can interpret such images in online journals. They 
developed a robot searching online journals to find fluorescence 
microscope images of individual cells. BOLAND & MURPHY 
(2001) used images of ten different subcellular pattems to train 
a neural network classifier. The classifier was able to correctly 
recognize an average of 83 % of the patterns. Fluorescence 
microscopy is the most common method used to determine 
  
  
  
   
   
  
  
   
  
  
  
  
  
  
   
  
  
   
  
    
  
  
  
   
   
   
  
  
   
   
   
   
    
    
   
  
  
   
  
   
   
  
   
  
   
  
   
   
  
   
  
   
  
  
  
   
  
   
  
  
   
   
  
  
Inter 
  
subc: 
previ 
subc« 
They 
on th 
cell € 
imag 
analy 
two 
visua 
appli 
sets 
repoi 
are fi 
resol 
netw 
test 
netw 
perf 
indi 
In th 
to v 
The 
mak: 
judg 
reco 
direc 
reco 
are 
iden 
from 
a pr 
be u 
insic 
Prio 
incl 
acqu 
foun 
orga 
ima; 
had 
part: 
chos 
allo 
Gol, 
On 
app: 
inva 
defi 
poir 
is si 
1). 
gol