A Texture-based Classification Method for Proteins in
Two-Dimensional Electrophoresis Gel Images
A Feature Selection Method using Support Vector Machines
and Genetic Algorithms
Carlos Fernandez-Lozano, Jose A. Seoane, Marcos Gestal, Daniel Rivero,
Julian Dorado and Alejandro Pazos
Information and Comunnications Technologies Department. Faculty of Computer Science, University of A Coruña,
Campus Elviña s/n, 15071, A Coruña, Spain
Keywords: Texture Analysis, Feature Selection, Electrophoresis, Support Vector Machines, Genetic Algorithm.
Abstract: In this paper, the influence of textural information is studied in two-dimensional electrophoresis gel images.
A Genetic Algorithm-based feature selection technique is used in order to select the most representative
textural features and reduced the original set (296 feat.) to a more efficient subset. Such a method makes use
of a Support Vector Machines classifier. Different experiments have been performed, the pattern set has
been divided into two parts (training and validation) extracting a total of 30%, 20% and 0% of the training
data, and a 10-fold cross validation is used for validation. In case of extracting 0% means that training set is
used for validation. For each division 10 different trials have been done. Experiments have been carried out
in order to measure the behaviour of the system and to achieve the most representative textural features for
the classification of proteins in two-dimensional gel electrophoresis images. This information can be useful
for a protein segmentation process.
1 INTRODUCTION
Proteomics is the study of protein properties in a cell
or tissue aimed at obtaining a global integrated view
of disease, physiological and biochemical processes
of cells and regulatory networks. One of the most
powerful techniques, widely used to analyze
complex protein mixtures extracted from cells,
tissues, or other biological samples, is two-
dimensional polyacrylamide gel electrophoresis
(2D-PAGE). In this method, proteins are classified
by molecular weight (MWt) and iso-electric point
(pI) using a controlled laboratory process and digital
imaging equipment. Among others separation of
proteins of a sample could also be done with several
different techniques such as chromatography or
mass spectrometry.
The main advantages of this approach are its
robustness, its parallelism and its unique ability to
analyze complete proteins at high resolution,
keeping them intact and being able to isolate them
entirely, however this method has also several
drawbacks (Rabilloud, Chevallet et al., 2010).
In this work the most representative group of
textural features are selected using Genetic
Algorithms.
2 THEORETICAL
BACKGROUND
The method proposed in this work intends to assist
in 2D-PAGE image analysis by studying the textural
information present within them. To do so, a novel
combination of Genetic Algorithms (Holland, 1975)
and Support Vector Machines (Vapnik, 1979) is
presented. In this section, the main techniques used
are briefly introduced and explained.
One of the most important characteristics used
for identifying objects or regions of interest in an
image is texture, related with the spatial (statistical)
distribution of the grey levels within an image
(Haralick, Shanmugam et al., 1973). Texture is a
surface’s property and can be regarded as the regular
spatial organization of complex patterns, always
401
Fernandez-Lozano C., Seoane J., Gestal M., Rivero D., Dorado J. and Pazos A..
A Texture-based Classification Method for Proteins in Two-Dimensional Electrophoresis Gel Images - A Feature Selection Method using Support Vector
Machines and Genetic Algorithms.
DOI: 10.5220/0004208704010404
In Proceedings of the International Conference on Computer Vision Theory and Applications (VISAPP-2013), pages 401-404
ISBN: 978-989-8565-47-1
Copyright
c
2013 SCITEPRESS (Science and Technology Publications, Lda.)
present even if they could exist as a non-dominant
feature.
Genetic Algorithms (GAs) are search techniques
inspired by Darwinian Evolution and developed by
Holland in the 1970s (Holland, 1975). In a GA, an
initial population of individuals, i.e. possible
solutions defined within the domain of a fitness
function to be optimized, is evolved by means of
genetic operators: selection, crossover and mutation.
The selection operator ensures the survival of the
fittest, while the crossover represents the mating
between individuals, and the mutation operator
introduces random modifications. GAs possesses
effective exploration and exploitation capabilities to
explore the search space in parallel, exploiting the
information about the quality of the individuals
evaluated so far (Goldberg, 1989).
Vapnik introduces Support Vector Machines
(SVMs) in the late 1970s on the foundation of
statistical learning theory (Vapnik, 1979). The basic
implementation deals with two-class problems in
which data are separated by a hyperplane defined by
a number of support vectors. This hyperplane
separates the positive from the negative examples, to
orient it such that the distance between the boundary
and the nearest data point in each class is maximal;
the nearest data points are used to define the
margins, known as support vectors (Burges, 1998).
These classifiers have also proven to be
exceptionally efficient in classification problems of
higher dimensionality (Chapelle, Haffner et al.,
1999; Moulin, Alves Da Silva et al., 2004), because
of their ability to generalize in high-dimensional
spaces, such as the ones spanned by texture patterns.
3 MATERIALS
In order to generate the dataset, ten 2D-PAGE
images of different types of tissues and different
experimental conditions were used. These images
are similar to the ones used by G.-Z. Yang (Imperial
College of Science, Technology and Medicine,
London). It is important to notice that Hunt et al.
(Hunt, Thomas et al. 2005) determined that 7-8 is
the minimum acceptable number of samples for a
proteomic study.
For each image, 50 regions of interest (ROIs)
representing proteins and 50 representing no-
proteins (noise, black non-protein regions, and
background) were selected to build a training set
with 1000 samples in a double-blind process in the
way that two clinicians select as many ROIs as they
considered and after that, within the common ROIs
clinicians selected proteins which are representatives
(isolated, overlapped, big, small, darker, etc.).
4 PROPOSED METHOD
The first step in texture analysis is texture feature
extraction from the ROIs. With a specialized
software called Mazda (Szczypiski et al., 2009), 296
texture features are computed for each element in
the training set. These features are based on the
image histogram, co-ocurrence matrix, run-length
matrix, image gradients, autoregressive models and
wavelet analysis. Histogram-related measures
conform the first-order statistics proposed by
Haralick (Haralick, Shanmugam et al., 1973) but
second-order statistics are those derived from the
Spatial Distribution Grey-Level Matrices (SDGM).
All these feature sets were included in the
dataset. The normalization method applied was the
one set by default in Mazda: image intensities were
normalized in the range from 1 to Ng=2
k
, where k is
the number of bits per pixel used to encode the
image under analysis.
In this work, GA is aimed at finding the smallest
feature subset able to yield a fitness value above a
threshold. Besides optimizing the complexity of the
classifier, feature selection may also improve the
classifiers quality. In fact, classification accuracy
could even improve if noisy or dependent features
are removed.
GAs for feature selection were first proposed by
Siedlecki and Skalansky (Siedlecki and Sklansky,
1989). Many studies have been done on GA for
feature selection since then (Kudo and Sklansky
1998), concluding that GA is suitable for finding
optimal solutions to large problems with more than
40 features to select from.
GA for feature selection could be used in
combination with a classifier such SVM, KNN or
ANN, optimizing it. In our method, based on both
GA and SVM, there is no a fixed number of
variables. As the GA continuously reduces the
number of variables that characterize the samples, a
pruned search is implemented. The fitness function
(1) considers not only the classification results but
also the number of variables used for such a
classification, so it is defined as the sum of two
factors, one related to the classification results and
another to the number of variables selected.
Regarding classification results, it apparently gives
better results taking into account the F-measure than
only using the accuracy obtained with image
features (Müller, Demuth et al., 2008; Tamboli and
VISAPP2013-InternationalConferenceonComputerVisionTheoryandApplications
402
Shah, 2011). F-measure is a function made up of the
recall (true positives rate or sensitivity: proportion of
actual positives which are correctly identified as
such) and precision (or positive predictive value:
proportion of positive test results that are true
positives) measurements.
1
(1)
Therefore individuals with less active genes are
favored.
5 EXPERIMENTAL RESULTS
The method proposed in this work requires the
division of the pattern set into two halves. To avoid
overfitting, this work proposes to split the training
dataset into training and validation sets to perform a
validation of the obtained results. Once the GA
finishes, the best individual found (the one with
lowest fitness value) is tested, using a 10-fold cross
validation (10-fold CV), to calculate the error of the
proposed model with the validation set and using
only the features in the best individual chromosome.
This involves dividing the validation set into 10
complementary subsets, performing the analysis on
1 subset, retained as the validation data and the
remaining k-1 subsamples are used as training data.
This second partitioning provokes that either
validation could be carried out with a very reduced
number of data points. In this case, either training or
validation sets, will surely not be representative of
the search space that is being explored. Different
experiments have been performed to verify this, in
these experiments the pattern set has been divided
into two parts (training and validation) extracting a
total of 30 %, 20% and 0% of the training data to the
validation set. In case of extracting 0%, which
means that no validation is performed training set is
used for validation using cross validation. For each
division 10 different trials have been done; a
different seed is used to divide randomly the
elements of the dataset each time.
Parameters domains of the feature selection
method were initially adjusted based on the literature
in the way that are ranging in the case of the
population size from 100 to 250 individuals, elitism
from 0% to 2%, crossover probability from 80% to
98% and mutation probability from 1% to 5%. One-
point, two-point, scattered, arithmetic and heuristic
crossover functions were probed. Regarding with
selection function, uniform, roulette and tournament
functions were evaluated with uniform and Gaussian
mutation functions.
Final combination set population size to 250
individuals, no elite, 95% crossover probability, 2%
mutation probability, crossover scattered,
tournament selection and mutation uniform.
SVM parameters domains are set for the kernel
function as lineal, quadratic, polynomial (order
ranging from 3 to 10) and Gaussian radial basis,
with sigma parameter ranging from 0,1 to 10 and C
parameter from 1 to 100. The RBF(2) kernel
function is selected as the most accurate for solving
this problem.
Experiments results shown in Table 1 separated
in training error and the validation error calculated
using 10-fold CV and the validation set for each
division. Results are in mean of error of the 10 trials
and standard deviation is in brackets. Best results in
validation are achieved for the 0% division.
In these 10 trials, some features seem to be more
relevant and appear recurrently as the solution of
each trial. Skewness, S(0,5) InvDfMom, S(2,2)
Correlat and S(0,4) InvDfMom appears at least in 5
solutions. Skewness is a measure of the degree of
asymmetry of the image histogram distribution.
Correlation analyzes the linear dependency of gray
levels of neighboring pixels. When the scale of local
texture is larger than the distance this measure is
typically high. And inverse difference moment is the
inverse of the contrast of the occurrence matrix so it
is a measure of the amount of local uniformity
present in the image.
The co-ocurrence matrix in Mazda (Szczypiski et
al., 2009) is symmetric and the image is normalized.
Co-ocurrence based parameters are computed up to
20 times, for (d,0), (0,d), (d,d) and (d,-d) where
distance d is ranging from 1 to 5.
6 SUMMARY AND
CONCLUSIONS
In this work we present a method for classification
of proteins in two-dimensional electrophoresis gels
using textural information. The proposed method is
based on a feature selection process using GAs
(Holland, 1975) and SVMs (Vapnik, 1979).
A dataset with 10 images, 100 ROIs for each one
and 296 features per ROI is created. Two different
clinicians have performed this manual protein
detection. This is a high variability process; a
refinement step based on the correlation of the
results of the two clinicians was performed, in order
ATexture-basedClassificationMethodforProteinsinTwo-DimensionalElectrophoresisGelImages-AFeatureSelection
MethodusingSupportVectorMachinesandGeneticAlgorithms
403
to select enough representative proteins.
Different experiments have been performed and
in these experiments the pattern set has been divided
into two parts (training and validation) extracting a
total of 30 %, 20% and 0% of the training data to the
validation set.
The proposed method has been successfully
applied to different real images, including images
with high complexity, which means larger number
of proteins and larger deformation between images.
Furthermore, the method presented have important
implications for the analysis of two-dimensional
electrophoresis gel images in the sense that this
classification step can be very useful in order to
discard over-segmented areas after a protein
segmentation or identification process.
ACKNOWLEDGEMENTS
This work is supported by the General Directorate of
Culture, Education and University Management of
the Xunta de Galicia (Ref. 10SIN105004PR).
REFERENCES
Burges, C. J. C. (1998). A tutorial on support vector
machines for pattern recognition. Data Mining and
Knowledge Discovery 2(2): 121-167.
Chapelle, O., P. Haffner, et al. (1999). Support vector
machines for histogram-based image classification.
IEEE Transactions on Neural Networks 10(5): 1055-
1064.
Goldberg, D. (1989). Genetic Algorithms in Search,
Optimization, and Machine Learning, Addison-Wesley
Professional.
Haralick, R. M., K. Shanmugam, et al. (1973). Textural
features for image classification. IEEE Transactions
on Systems, Man and Cybernetics smc 3(6): 610-621.
Holland, J. H. (1975). Adaptation in natural and artificial
systems: an introductory analysis with applications to
biology, control, and artificial intelligence, University
of Michigan Press.
Hunt, S. M. N., M. R. Thomas, et al. (2005). Optimal
Replication and the Importance of Experimental
Design for Gel-Based Quantitative Proteomics.
Journal of Proteome Research 4(3): 809-819.
Kudo, M. and J. Sklansky (1998). A comparative
evaluation of medium- and large-scale feature
selectors for pattern classifiers. Kybernetika 34(4):
429-434.
Moulin, L. S., A. P. Alves Da Silva, et al. (2004). Support
vector machines for transient stability analysis of
large-scale power systems. IEEE Transactions on
Power Systems 19(2): 818-825.
Müller, M., B. Demuth, et al. (2008). An evolutionary
approach for learning motion class patterns. 5096
LNCS: 365-374.
Rabilloud, T., M. Chevallet, et al. (2010). Two-
dimensional gel electrophoresis in proteomics: Past,
present and future. Journal of Proteomics 73(11):
2064-2077.
Siedlecki, W. and J. Sklansky (1989). A note on genetic
algorithms for large-scale feature selection. Pattern
Recognition Letters 10(5): 335-347.
Szczypiski, P. M., M. Strzelecki, et al. (2009). MaZda-A
software package for image texture analysis.
Computer Methods and Programs in Biomedicine
94(1): 66-76.
Tamboli, A. S. and M. A. Shah (2011). A Generic
Structure of Object Classification Using Genetic
Programming. Communication Systems and Network
Technologies (CSNT), 2011 International Conference
on.
Vapnik, V. N. (1979). Estimation of dependences based on
empirical data [in Russian]. Nauka, English translation
Springer Verlang, 1982.
APPENDIX
Table 1: Results separated in training and validation.
Training
%
division
Accuracy Sensitivity Specificity F
70-30 0.9128(0.0103) 0.9450(0.0180) 0.8816(0.0131) 0.9141(0.0107)
80-20 0.9183(0.0093) 0.9504(0.0118) 0.8866(0.0086) 0.9202(0.0097)
100-0 0.9236(0.0035) 0.9593(0.0050) 0.8880(0.0072) 0.9262(0.0032)
Validation
70-30 0.9073(0.0058) 0.9442(0.0112) 0.8699(0.0118) 0.9110(0.0057)
80-20 0.9165(0.0090) 0.9518(0.0107) 0.8812(0.0135) 0.9192(0.0085)
100-0 0.9254(0.0040) 0.9580(0.0037) 0.8928(0.0068)
0.9277(0.0037)
VISAPP2013-InternationalConferenceonComputerVisionTheoryandApplications
404
