Jayashree Dev et al.

/ International Journal on Computer Science and Engineering (IJCSE)

A Classification Technique for Microarray

Gene Expression Data using PSO-FLANN
Jayashree Dev1, Sanjit Kumar Dash2, Sweta Dash 3, Madhusmita Swain 4
1, 2, 4
College of Engineering and Technology
Biju Patanaik University of Technology
Bhubaneswar, Odisha, India
3
Synergy Institute of Engineering and Technology
Biju Patanaik University of Technology
Dhenkanal, Odisha, India

Abstract— Despite of an increased global effort to end breast cancer, it continues to be most common
cancer deaths in women. This problem reminds that new therapeutic approaches are desperately needed
to improve patient survival rate. This requires proper diagnosis of disease and classification of tumor
type based on genomic information according to which proper treatment can be provided to the patient.
There exists a no. of classification techniques to classify the tumor types. In this paper we have focused on
three different classification techniques: BPN, FLANN and PSO-FLANN and found that the integrated
approach of Functional Link Artificial Neural Network (FLANN) and Particle Swarm Optimization
(PSO) can better predict the disease as compared to other method.

Keywords- Backpropagation (BPN), FLANN, PSO, PSO-FLANN

I. INTRODUCTION
Breast cancer is a heterogeneous disease with respect to molecular alteration, cellular composition, and
clinical outcome. This diversity creates a challenge in developing tumor classifications that are clinically useful
with respect to prognosis or prediction. There are two types of breast cancer: ductal carcinomas and lobular
carcinomas. Cancers originating from ducts are known as ductal carcinomas and cancers originating from
lobules are known as lobular carcinomas. The high incidence of breast cancer is emerging as a public health
problem in the country in past decades. While there is no single reason for the escalating incidence of breast
cancer, obesity, dietary habits, physical inactivity and the over-use of hormonal pills are the main causes.
Prognosis and survival rates for breast cancer vary greatly depending on the cancer type, stage, treatment and
geographic location of the patient. Breast cancer is the most commonly diagnosed form of cancer in women
accounting for about 30% of all cases. Some studies, which were based on low capability and poorly calibrated
equipment, infrared imaging has been shown to be well suited for task of detecting breast cancer, in particular
when the tumor is in its early stages or in dense tissue [1].

Microarrays are a powerful tool for biologists as they enable the simultaneous measurement of the
expression levels of thousands of genes per tissue sample [10]. Microarray analysis is a widely used technology
for studying gene expression on a global scale. Gene expression profiling by DNA microarray has become an
important tool for studying the transcriptome of cancer cells and has been successfully used in many studies of
tumor classification and identification of marker genes associated with cancer. With an increasing number of
microarray data becoming available, the comparative of study on normal tissue versus tumor tissue has gained
high importance.

Microarray breast cancer event prediction, however, has proven to be difficult, as few classification rules are
able to obtain a balanced accuracy rate of over 70%, when properly validated. These performance indicators are
also often associated with wide confidence intervals. Signature composition strongly depends on the subset of
patient samples used for feature selection. In recent years many different signatures have been proposed, mostly
derived using different patient populations and/or array technologies. Although the overall performance of these
signatures is comparable, there is often a high level of inconsistency between class assignments obtained using
different signatures. One of the challenging aspects of microarray data is that they are subject to various sources
of technical variation, arising from the many experimental laboratory steps needed to get from a tissue sample to
an array scan. The noises can be removed from microarray data using some preprocessing methods.

The goal of this study is to investigate the benefit of performing supervised classification analyses on
microarray data [6]. Methods of supervised classification analysis render it possible to automatically build
classifiers that distinguish among specimens on the basis of predefined class label information (phenotypes) and

ISSN : 0975-3397 Vol. 4 No. 09 Sep 2012 1534

 Jayashree Dev et al. / International Journal on Computer Science and Engineering (IJCSE)

in many cancer research studies; the application of these methods has shown promising results of improved
tumor diagnosis and prognosis. In this paper, an integrated approach of functional link artificial intelligence
(FLANN) [2] and particle swarm optimization (PSO) [3] is used to build a more reliable Classifier. This
approach incorporates gene features and FLANN parameters into one common solution code. The problem is
solved by selecting gene features and optimizing the parameters of the FLANN classifier.
II. CLASSIFICATION WITH BPN MODEL
Neural networks with back propagation technique can be used for classification task such as character
recognition, voice recognition, medical application etc [7, 8, 9]. Neurons in neural network have weighted
inputs, threshold values, activation function and an output where activation function=f (∑ (inputs * weights).
Threshold values play an important role in deciding the output. Backpropagation technique can be applied to
multi-layered neural network. Neural networks learn by example. That means, during training we have to apply
input-output pair from the training set to the neural network. Training continues till the network is well-trained.
After that the performance of the network can be checked by applying the patterns belonging to testing set. The
increasing no. of hidden layers results in the computational complexity of the network. The time taken for
convergence and to minimize the error may be very high due to this reason.

Figure 1: Basic structure of BPN Classifier

The pseudo-code for backpropagation classification is as follows:

Step 1: Initialize weights to small random values.

Step 2: while stopping condition is false, do steps 3-10.
Step 3: For each training pair do steps 4-9
Step 4: Each input unit receives the input signal xi and transmits the signal to all units in the hidden layer.
Step 5: Each hidden unit (Zh, h=1...p) sums its input signals
Zinj=Voj+∑ (xiVij)
By applying activation function Zj=f(Zinj) and send this signal to all units in output layer.
Step 6: Each output unit (Yk, k=1...m) sums its weighted input signals.
Yink=Wok+ ∑ (ZjWij)
And supplies its activation function to calculate the output signal
Yk= f (Yink)
Step 7: Each output unit (Yk, k=1...m) receives a target pattern corresponding to input pattern, error information
term is calculated as
δk= (tk-Yk) f (Yink) and δinj = ∑δjWjk
Step 8: Each hidden unit (Zj, j=1...n) sums its delta inputs from unit layers above. The error information term is
calculated as:
δj =δinjf(Zinj)
Step 9: Each output unit (Yk, k=1… m) updates its weight (j=0… n)
The weight correction term is given by ΔWjk = αδkZj
And bias correction term is given by ΔWok = αδk
Therefore Wjk(new) = Wjk(old)+ ΔWjk and
Wok (new) = Wok (old) + ΔWok
Each hidden unit (Zj, j=1… p) updates its weights (i=0…n)
The weight correction term is ΔVjk = αδjXj

ISSN : 0975-3397 Vol. 4 No. 09 Sep 2012 1535

 Jayashree Dev et al. / International Journal on Computer Science and Engineering (IJCSE)

And bias correction term is ΔVoj = αδj

Therefore Vij(new) = Vij(old)+ ΔVij and
Voi(new) = Voi (old) + ΔVoi
III. CLASSIFICATION WITH FLANN MODEL
Single layer neural networks are suitable for solving linear problem and also they are simple and easy to
implement. FLANN is a mathematical model or computational model that is inspired by structural and/or
functional aspects of biological neural networks [4, 5]. It consists of an interconnected group of artificial
neurons and it processes information using a model. FLANN can be treated as a single layer neural network but
with the ability to solve non-linear problems. Here, additional input data are generated off-line using non-linear
transformations. In a FLANN, the input layer is expanded to a layer of functional units, which consists of higher
order combinations of the input units. Each functional unit is fully connected to next layer. The addition of
higher order combinations of inputs artificially increases the dimension of input space and hence the hyper
planes generated by the FLANN provide greater discrimination capability in the input pattern space. Both non-
hidden layer and enhanced input pattern space make this network high precision of function approximation.
There are three approaches of input layer expansion:-functional expansion, combined convolution and the
combination of both. Here, we employ the functional expansion. There are three different polynomials for
functional expansion of input pattern in FLANN. They are Chebyschev, Legendre and Power series. It is
important to choose suitable functional expansion.

Figure 2: Basic Structure of FLANN Classifier

Training process constitutes following steps:
Step 1: Read the data D= {D1, D2… DN} where each Di is a vector of length M and the class label of order N * 1
Step 2: Normalize the Data.
Step 3: Take 80 % of data for training in DTr and 20% of data for testing as DT.
Step 4: Randomly assign the wt for each FLANN
Step 5: for i=1: maxIter do
Step 6: δw = wt;
Step 7: for j=1: N do
Step 8: evaluate the output of flann v= DTr (j, :)* δw
Step 9: evaluate error e (j) = v – expected_out_put
Step 10: Update δw= δw – 2 * μ * e (j);
Step 11: end //end of step 5
Step 12: er (i) = Mean (e);
Step 13: wt = wt - δw;
Step 14: end //end of step 7
After the training is over, testing process starts. During testing, the prediction capability of the network is tested.
IV. CLASSIFICATION WITH PSO-FLANN MODEL
The proposed hybrid FLANN is based on evolutionary algorithm Particle Swarm Optimization (PSO) [9].
FLANN suffers from following problems:
1. FLANN is more sensitive to error in training pattern.
2. Failure to converge due to insufficient structure, lockup, tracking to local optima, limited cycles
3. Shows good performance on training set and poor performance in testing set due to the problem of training
set not representing variations in patterns and too strict in tolerance.
To avoid such problems we have proposed an evolutionary technique based model which is able to classify the
patterns in a better way.

ISSN : 0975-3397 Vol. 4 No. 09 Sep 2012 1536

 Jayashree Dev et al. / International Journal on Computer Science and Engineering (IJCSE)

Figure 3. Basic Structure of FLANN with PSO Classifier

We can harness the power of Particle Swarm Optimization [9] for training the FLANN to reduce the local
optimality and speed up the convergence. PSO algorithm can be applied to three main attributes of neural
networks: connection based, network architecture and network learning algorithm. Here we have focused on
connection based. PSO simulates the behavior of bird flocking. It learns from the scenario and uses it to solve
the optimization problem. IN PSO, each single solution is a “bird” in the search space. We call it as “particle”.
All particles have fitness values which are evaluated by fitness function to be optimized and have velocities
which direct the flying of particles. The particles fly through the problem space by following the current
optimum particles.
Training process constitutes following steps:
Step 1: Read the data D= {D1, D2… DN} where Di is a vector of length M and the class label of order N * 1
Step 2: Normalize the Data.
Step 3: Take 80 % of data for training in DTr and 20% of data for testing as DT.
Step 4: Calculate the fitness for each swarm for every DTi
Step 5: Randomly assign the wt for each FLANN
Step 6: initialize the parameter pbest, c1 and c2 for swarm, Velocity Vid and position Updation Xid.
Step 7: for i=1: maxIter do
Step 8: δw = wt;
Step 9: for j=1: N do
Step 10: evaluate the output of flann v= Xid (j)* Mean (DTr (j, :))*DTr (j, :)* δw
Step 11: evaluate error and pbest of each swarm
e(j)= v – expected_out_put
pbest (j)=max(pbest(j), v)
Step 12: Update δw= δw – 2 * μ * e(j);
Step 13: update new velocity and new position for swarm
Vid j+1 = wVidj + C1 * rand1 () * (pbest – Xid) + C2 * rand2 () * (gbest – Xid)
Where Xid = Xidj - Vidj
j+1

Step 14: end //end of step 9

Step 15: er(i) = Mean(e);
Step 16: wt = wt - δw;
Step 17: calculate gbest=max (v)
Step 17: end //end of step 7
The above FLANN with PSO algorithm is used to train and test the breast cancer dataset. Here the datasets
are taken and first trained with FLANN with PSO model by providing random weights, position value, the value
of random constants c1 and c2 and pbest, gbest value. Then error will be checked by expected output and
resulted output, according to that weight and position will be modified.
V. PERFORMANE ANALYSIS
The proposed technique for bio-medical data classification has been implemented in the working platform of
MATLAB (version 7.8). For evaluating the proposed technique, we have utilized the microarray gene samples
of human breast cancer dataset. The data set is collected from UCI machine learning repository
(http://archive.ics.uci.edu). The BPN, FLANN and PSO-FLANN have been trained by breast cancer dataset
having class label two. Table 1, 2 and 3 shows the performance of different classification techniques used by us.
From the Table 3, it can be seen that the PSO-FLANN has provided more accuracy and less error rate rather

ISSN : 0975-3397 Vol. 4 No. 09 Sep 2012 1537

 Jayashree Dev et al. / International Journal on Computer Science and Engineering (IJCSE)

than the BPN and FLANN-based gene classification techniques. More accuracy and less error rate leads to
effective classification of the given microarray gene data to the actual class of the gene. From the above training
and testing phases of BPN, FLANN and PSO-FLANN, we observed that FLANN with PSO provides good
results than BPN and FLANN.

TABLE I. CONFUSION MATRIX FOR BREAST CANCER DATA SET USING BPN

Predicted
Class 1 Class 2
Actual Class 1 62 15
Class 2 46 16

TABLE II. CONFUSION MATRIX FOR BREAST CANCER DATA SET USING FLANN
Predicted
Class 1 Class 2
Actual Class 1 56 21
Class 2 30 32

TABLE III. CONFUSION MATRIX FOR BREAST CANCER DATA SET USING PSO-FLANN
Predicted
Class 1 Class 2
Actual Class 1 70 7
Class 2 4 58

Figure 4. Error Graph using BPN

Figure 5. Error Graph using FLANN

ISSN : 0975-3397 Vol. 4 No. 09 Sep 2012 1538

 Jayashree Dev et al. / International Journal on Computer Science and Engineering (IJCSE)

Figure 6. Error Graph using PSO-FLANN

TABLE IV. COMPARISON OF CLASSIFICATION RESULTS
Percentage of Accuracy
Name of data
set BPN FLANN PSO_FLANN
Breast Cancer 56.12% 63.34% 92.36%
VI. CONCLUSION
The performance of PSO-FLANN is better for classification as compared to BPN and FLANN. The
performance has been checked on breast cancer data set which is obtained from the UCI machine learning
repository website. This method overcomes the nonlinearity of the classification problem. The FLANN with
PSO architecture, because of its simple architecture and computational efficiency may be conveniently
employed in other tasks of data mining and knowledge discovery in databases such as clustering, feature
selection, feature extraction, association rule mining, regression, and so on. However, a necessary property of
algorithms which are capable of handling large and growing datasets is their scalability or linear complexity
with respect to the data size. The extra calculation generated by the higher order units can be eliminated,
provided that these polynomial terms are stored in memory instead of being recalculated each time the FLANN
trained.
REFERENCES
[1] A M. Sarhan; Cancer Classification Based on Microarray Gene Expression Data using DCT and ANN; Journal of Theoretical and
Applied Information Technology, pp. 208-216, 2009.
[2] Patra J. C., Nguyen C. Thanh and Meher Pramod K.; Computationally Efficient FLANN-based Intelligent Stock Price Prediction
System; Proceedings of International Joint Conference on Neural Networks, Atlanta, Georgia, USA, June 14-19, 2009.
[3] Enrique Alba, JoseGarcıa-Nieto, Laetitia Jourdan and El-Ghazali Talbi; Gene Selection in Cancer Classification using PSO/SVM and
GA/SVM Hybrid Algorithms.; Author manuscript, published in "Congress on Evolutionary Computation, Singapor : Singapore
(2007)..
[4] Dehuri S., Cho S.B.; A Comprehensive Survey on Functional Link Neural Networks and an Adaptive PSO-BP Learning for CFLNN;
Department of Computer Science, Yonsei University, 262 Seongsanno, Seodaemun-gu, Seoul 120-749, Korea.
[5] Leandro dos Santos Coelho, Cezar Augusto Sierakowski ; A software Tool for Teaching of Particle Swarm Optimization
Fundamentals; Advances in Engineering Software, Volume 39, Issue 11, November 2008, Pages 877-887 .
[6] Frank Rapaport, Marie Dutreix, Emmanuel Barillot, Jean-Philippe Vert; Classification of Micro Array Data using Gene Networks;
February 1, 2007; BMC Bioinformatics.
[7] F. Pauline; Classification of Breast Cancer by Comparing Backpropagation Training Algorithms; IJCSE; Vol-3, No.1; Jan 2011; ISSN:
0975-3397.
[8] A. Gupta, M. Shreevastava; Medical Diagnosis using Backpropagation Algorithm; International Journal of Emerging Technology and
Advanced Engineering; Vol.1, Issue 1, November 2011, Page 55-58 ; ISSN: 2250-2459.
[9] Chow, M.-Y., Sharpe, R. N., & Hung, J. C. (1993). On the Application and Design of Artificial Neural Networks for Motor Fault
Detection. IEEE Transactions on Industrial Electronics, 40(2), 181-188.
[10] S.K. Gruvbergek, H.E. Cuniliffe, K.M. Carr and I.A. Hedenfalk; Microarrays in Breast Cancer Research and Clinical practice-The
Future Lies Ahead; An Official Journal of Society for Endocrinology and European Society of Endocrinology; Dec 1, 2006.

ISSN : 0975-3397 Vol. 4 No. 09 Sep 2012 1539